CN118037362B - Sequence recommendation method and system based on user multi-intention comparison - Google Patents
Sequence recommendation method and system based on user multi-intention comparison Download PDFInfo
- Publication number
- CN118037362B CN118037362B CN202410437551.8A CN202410437551A CN118037362B CN 118037362 B CN118037362 B CN 118037362B CN 202410437551 A CN202410437551 A CN 202410437551A CN 118037362 B CN118037362 B CN 118037362B
- Authority
- CN
- China
- Prior art keywords
- vector
- intention
- sequence
- original
- user
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 60
- 239000013598 vector Substances 0.000 claims abstract description 242
- 239000002775 capsule Substances 0.000 claims abstract description 49
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 41
- 238000010168 coupling process Methods 0.000 claims abstract description 30
- 238000005859 coupling reaction Methods 0.000 claims abstract description 30
- 230000008878 coupling Effects 0.000 claims abstract description 29
- 230000006399 behavior Effects 0.000 claims abstract description 25
- 238000003064 k means clustering Methods 0.000 claims abstract description 16
- 238000012545 processing Methods 0.000 claims abstract description 8
- 239000011159 matrix material Substances 0.000 claims description 28
- 230000006870 function Effects 0.000 claims description 19
- 230000008569 process Effects 0.000 claims description 17
- 230000007246 mechanism Effects 0.000 claims description 12
- 238000010606 normalization Methods 0.000 claims description 10
- 238000012546 transfer Methods 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 6
- 230000009471 action Effects 0.000 claims description 4
- 239000000945 filler Substances 0.000 claims description 4
- 230000000873 masking effect Effects 0.000 claims description 4
- 230000009466 transformation Effects 0.000 claims description 4
- 230000002779 inactivation Effects 0.000 claims description 3
- 239000000758 substrate Substances 0.000 claims description 3
- 239000012633 leachable Substances 0.000 claims description 2
- 238000013135 deep learning Methods 0.000 abstract description 7
- 230000003993 interaction Effects 0.000 description 14
- 238000012549 training Methods 0.000 description 13
- 230000000694 effects Effects 0.000 description 8
- 239000003795 chemical substances by application Substances 0.000 description 7
- 238000010586 diagram Methods 0.000 description 5
- 210000002569 neuron Anatomy 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000012512 characterization method Methods 0.000 description 3
- 230000002452 interceptive effect Effects 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 230000003044 adaptive effect Effects 0.000 description 2
- 230000001186 cumulative effect Effects 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000012821 model calculation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 230000000087 stabilizing effect Effects 0.000 description 1
- 239000004575 stone Substances 0.000 description 1
- 238000000844 transformation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0251—Targeted advertisements
- G06Q30/0269—Targeted advertisements based on user profile or attribute
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Accounting & Taxation (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Strategic Management (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Development Economics (AREA)
- Finance (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Probability & Statistics with Applications (AREA)
- Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a sequence recommendation method and a system based on multi-intention comparison of users, wherein the method comprises the following steps: processing an original sequence and an enhanced sequence of user behavior data through a preset sequence encoder respectively to obtain an original coding vector and an enhanced coding vector; the original intention vector is obtained through an EM algorithm and a k-means clustering algorithm, and the enhanced intention vector is obtained through a capsule network; the similarity between the original intention vector and the enhanced intention vector is obtained through self-supervision contrast learning respectively; acquiring a user multi-intention preference representation according to the similarity between the original intention vector and the enhanced intention vector and the similarity between the initial code vector and the enhanced code vector; and acquiring a recommendation list and recommending to the user according to the multi-intention preference representation and the item preference representation of the user. The method and the device can solve the problem that the prior deep learning-based sequence recommendation aims to explore single-purpose and multi-purpose coupling.
Description
Technical Field
The invention relates to the technical field of personalized recommendation, in particular to a sequence recommendation method and system based on multi-intention comparison of users.
Background
The personalized recommendation system is widely applied to platforms such as electronic commerce, movies, music, news, online advertisements and the like at present, wherein two main recommendation systems are respectively conventional general recommendation and sequence recommendation. The sequence recommendation predicts articles with which the user interacts in the future by mining time sequence information accessed by the user, so that the sequence behavior of the user can be modeled explicitly, and meanwhile, the dynamic interest change of the user can be captured. Although current deep learning based sequence recommendation models have achieved significant success in model performance, there are two important issues that are rarely noted with sequence recommendation: the intent explores the problem of singleness and intent coupling. First, the intent to explore a single problem remains serious because sequential recommendations often predict the user's next interaction item, the simplest and efficient way is to capture the user's single intent to recommend, but this ignores the fact that in reality the user's behavior is driven by multiple intents. Meanwhile, most of sequence recommendation algorithms based on intention often use a clustering method to capture multiple intention of a user, but neglect contribution degrees of different intention to user behaviors, so that the captured intention of the user has a very serious coupling problem, and a model achieves a suboptimal result.
There are two challenges to solving the above problems; first, it is difficult to accurately model multiple intents from user history interactions because of the lack of intent tag data. Most existing user intent modeling efforts require side information (e.g., user clicks, purchases, etc.) to capture user intent, but these are not always available in actual recommendation systems. And fusing the intent information into the sequence recommendation model requires that the intent representation be orthogonal to the sequence embedding, otherwise resulting in redundant information. Secondly, the user's schematic coupling cannot be successfully completed due to lack of supervisory signals. User behavior data is used as a unique supervision signal of intention, the same behavior may reflect different intention, and decoupling of the intention is a problem to be solved urgently.
Disclosure of Invention
In view of the above problems, the present invention aims to provide a sequence recommendation method and system based on user multi-intention comparison, so as to solve the problem of intention exploration single and multi-intention coupling existing in the existing sequence recommendation based on deep learning.
The invention provides a sequence recommendation method based on multi-intention comparison of users, which comprises the following steps:
carrying out data enhancement on user behavior data to obtain an enhancement sequence;
Processing the original sequence and the enhancement sequence of the user behavior data through a preset sequence encoder respectively to obtain an original coding vector and an enhancement coding vector;
performing multi-intention learning on the original coding vector through an EM algorithm and a k-means clustering algorithm to obtain an original intention vector; and
Multi-intention learning is carried out on the enhanced coding vector through a capsule network, and an enhanced intention vector is obtained;
And respectively obtaining the similarity between the original intention vector and the enhanced intention vector and the similarity between the original code vector and the enhanced code vector, wherein,
Obtaining the similarity between the original intention vector and the enhanced intention vector by performing self-supervised contrast learning on the original intention vector and the enhanced intention vector, and obtaining the similarity between the original code vector and the enhanced code vector by performing self-supervised contrast learning on the original code vector and the enhanced code vector;
Acquiring a user multi-intention preference representation according to the similarity between the original intention vector and the enhanced intention vector and the similarity between the initial coding vector and the enhanced coding vector;
And acquiring a recommendation list and recommending to the user according to the multi-intention preference representation and the item preference representation of the user.
Furthermore, it is preferable that the data enhancement means include masking, clipping, and reordering, wherein a sequence is given for a given sequenceAnd a predefined set of data transfer functionsCreation ofIs shown as follows:
，
Wherein, Indicating a particular value of the characteristic,Which represents the user's representation of the user,The time is represented by the time period of the day,AndRepresenting slaveRandomly sampled transfer functions to create a sequenceIs shown in the figure).
In addition, the method is preferable in that the original sequence and the enhancement sequence of the user behavior data are respectively processed through a preset sequence encoder to obtain an original coding vector and an enhancement coding vector; wherein,
The preset sequence encoder adopts a transducer model and comprises an embedded part, a self-attention part and a stacked-attention part, wherein,
The matrix formula of the embedded part is as follows:
，
Wherein, Representing the maximum length of the sequence of user actions,Representing a leachable location embedding;
representing an item embedding matrix, wherein, In order to obtain the number of items,Is a potential dimension;
Embedding matrix Wherein, the method comprises the steps of, wherein,Embedding using a constant zero vector as a filler;
The self-attention portion includes a self-attention mechanism and a point-by-element feed forward network, wherein the formula of the scalable dot product attention is as follows:
，
Wherein, Representing a query; Representing a bond; Representing values, each row representing an item; Representing scale factors;
the self-attention mechanism uses the same objects as queries, keys and values to embed a matrix As input, it is converted into three matrices by linear projection and input into a scalable dot product attention, the formula of which is:
，
wherein the projection matrix ；
The formula of the point element-by-element feed forward network is as follows:
，
Wherein, AndIs thatIs a matrix of the (c) in the matrix,AndIs thatA dimension vector;
the stacked attention portion includes Attenion blocks and FFN blocks, the th The formula of the block is:
，
meanwhile, block 1 is represented as And; Residual error linking, layer normalization and random inactivation are introduced to alleviate overfitting, and the adopted formula is as follows:
，
Wherein, Representing self-attention layers or feed forward networks, layers of each block in the self-attention portionIn the prior pair inputApplication layer normalization followed by alignment ofThe output of (1) is Dropout, the inputAdded to the final output;
，
Wherein, Is a product of the element levels,AndIs thatIs a function of the mean and variance of (a),AndIs a learned scale factor and bias term.
Furthermore, it is preferable that in the process of obtaining the original intention vector by multi-intention learning of the original code vector by an EM algorithm and a k-means clustering algorithm, wherein,
The EM algorithm includes E-step and M-step, where E-step represents the desired steps for each iteration and M-step represents the maximum steps;
calculated in E-step using k-means clustering algorithm In time stepsThe original multi-purpose pattern is composed of a plurality of patterns,Represent the firstEstimated from small batch gradient dips in M-stepOptimizing parameters。
In addition, preferably, the multi-purpose learning is performed on the enhancement coding vector through a capsule network, and the enhancement intention vector is obtained;
the length of the capsule network output vector is the probability of the entity represented by the capsule in the current input, and is assumed that Is a capsuleIs obtained by dynamic path calculationIs the next layer vector of (a);
firstly, calculating a prediction vector, and adopting the following formula:
，
Wherein, Is a transformation matrix;
At this time, capsule Is input as all prediction vectorsIs given by the formula:
，
Wherein, For coupling coefficient in iterative dynamic routing process, capsuleAnd the sum of the coupling coefficients between all the capsules of the next layer is 1, wherein,The formula is adopted as follows:
，
Wherein, Is a capsuleAnd capsuleLogarithmic prior probability of coupling, hence capsuleThe vector formula of (2) is:
，
Wherein, I.e. the user intent vector obtained in one iteration,。
In addition, preferably, in the process of obtaining the similarity between the original encoding vector and the enhanced encoding vector by performing self-supervised contrast learning on the original encoding vector and the enhanced encoding vector,
And adopting InfoNCE loss functions to perform contrast learning, wherein the formula is as follows:
，
Wherein, Respectively representing the enhanced intent vector,Representing the original intent vector of the user,
、Respectively representing the minimization of the learning task,Representing dot product.
In addition, preferably, in the process of obtaining the similarity between the original encoding vector and the enhanced encoding vector by performing self-supervised contrast learning on the original encoding vector and the enhanced encoding vector,
Further optimizing parameters of a sequence encoder by InfoNCE lossThe formula is:
，
Wherein, Representing an enhancement encoding direction; representing dot product.
In addition, preferably, in the process of acquiring a recommendation list and recommending to a user according to the multi-intention preference representation of the user and the item preference representation,
Obtaining an item embedding matrix by multi-layer perceptionWherein the item embedding matrix is the item preference representation, the user multi-intent preference representation employing；
Making a next prediction suggestion for a time step by computing a next probability distribution for the entire set of itemsThe next item is a candidate itemThe formula of the recommended probability of (2) is:
，
Wherein, Is thatAt a time stepIs provided with an embedding in the substrate,Is thatIs embedded in the mold;
Will be Users as model outputsFor projectsAnd sorting in descending order according to the interest degree value, and recommending Top-K to the user.
Furthermore, it is preferable that the predetermined sequence encoder is an MSCR model sequence encoder：
，
Wherein,AndThe intensities of schematic coupling tasks and sequence-level self-supervision and comparison learning tasks are controlled respectively.
The invention also provides a sequence recommendation system based on user multi-intention comparison, which comprises the following steps:
The enhancement sequence acquisition unit is used for carrying out data enhancement on the user behavior data to acquire an enhancement sequence;
The coding vector acquisition unit is used for respectively processing the enhancement sequence and the original sequence through the sequence encoder to acquire an original coding vector and an enhancement coding vector;
the original intention vector acquisition unit is used for acquiring an original intention vector by multi-intention learning on the original code vector through an EM algorithm and a k-means clustering algorithm;
the enhancement intention vector acquisition unit is used for carrying out multi-intention learning on the enhancement coding vector through a capsule network to acquire an enhancement intention vector;
the self-supervision contrast learning unit is used for obtaining the similarity between the original intention vector and the enhanced intention vector by carrying out self-supervision contrast learning on the original intention vector and the enhanced intention vector;
The sequence comparison learning unit is used for obtaining the similarity between the original coding vector and the enhanced coding vector through self-supervision comparison learning of the original coding vector and the enhanced coding vector;
a user multi-intention preference expression acquisition unit configured to acquire a user multi-intention preference expression according to a similarity between the original intention vector and the enhanced intention vector, and a similarity between the original code vector and the enhanced code vector;
and the recommendation list acquisition unit is used for acquiring a recommendation list and recommending the recommendation list to the user according to the user multi-intention preference representation and the project preference representation.
According to the technical scheme, the sequence recommendation method and the system based on the multi-intention comparison of the user, provided by the invention, learn the original intention distribution of the user from the user behavior sequence through a clustering algorithm, generate the agent intention as a supervision signal by using a capsule network, introduce the user intention into a sequence recommendation model by using self-supervision comparison learning, and capture the multi-dimensional potential intention in the user sequence to the maximum extent. And the problem of single-intention and multi-intention coupling of intention exploration in the prior deep learning-based sequence recommendation is solved by introducing latent variables and an EM framework to solve the problem of intention exploration of a user and learning more accurate multi-intention characterization of the user by comparing a learning mechanism.
To the accomplishment of the foregoing and related ends, one or more aspects of the invention comprise the features hereinafter fully described. The following description and the annexed drawings set forth in detail certain illustrative aspects of the invention. These aspects are indicative, however, of but a few of the various ways in which the principles of the invention may be employed. Furthermore, the invention is intended to include all such aspects and their equivalents.
Drawings
Other objects and attainments together with a more complete understanding of the invention will become apparent and appreciated by referring to the following description taken in conjunction with the accompanying drawings. In the drawings:
FIG. 1 is a flowchart of a sequence recommendation method based on user multi-intention comparison according to an embodiment of the invention;
FIG. 2 is a schematic diagram of a sequence recommendation framework based on user multi-intent comparison according to an embodiment of the present invention;
fig. 3 is a schematic logic structure diagram of a sequence recommendation system based on user multi-intention comparison according to an embodiment of the invention.
The same reference numerals will be used throughout the drawings to refer to similar or corresponding features or functions.
Detailed Description
In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of one or more embodiments. It may be evident, however, that such embodiment(s) may be practiced without these specific details. In other instances, well-known structures and devices are shown in block diagram form in order to facilitate describing one or more embodiments.
Aiming at the problem that the intention explores single and multi-intention coupling in the prior deep learning-based sequence recommendation, the invention provides a sequence recommendation method and system based on user multi-intention comparison.
Specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
In order to illustrate the sequence recommendation method based on the user multi-intention comparison provided by the invention, fig. 1 shows a sequence recommendation method flow based on the user multi-intention comparison according to an embodiment of the invention.
As shown in fig. 1, the sequence recommendation method based on multi-intention comparison of users provided by the invention comprises the following steps:
S1: carrying out data enhancement on user behavior data to obtain an enhancement sequence;
S2: respectively obtaining an original coding vector and an enhanced coding vector for the original sequence and the enhanced sequence of the user behavior data through a preset sequence encoder, and
S3: performing multi-intention learning on the original coding vector through an EM algorithm and a k-means clustering algorithm to obtain an original intention vector;
respectively obtaining the similarity between the original intention vector and the enhanced intention vector and the similarity between the original coding vector and the enhanced coding vector, wherein S4: multi-intention learning is carried out on the enhanced coding vector through a capsule network, and an enhanced intention vector is obtained; and
S5: obtaining the similarity between the original intention vector and the enhanced intention vector by performing self-supervision contrast learning on the original intention vector and the enhanced intention vector;
s6: obtaining the similarity between the original coding vector and the enhanced coding vector by performing self-supervision contrast learning on the original coding vector and the enhanced coding vector;
S7: acquiring a user multi-intention preference representation according to the similarity between the original intention vector and the enhanced intention vector and the similarity between the initial coding vector and the enhanced coding vector;
S8: and acquiring a recommendation list and recommending to the user according to the multi-intention preference representation and the item preference representation of the user.
In the embodiment of the invention, starting from the two problems of single intention exploration and intention coupling, a plurality of modules and self-supervision comparison auxiliary tasks are designed, so that the multi-intention of a user is fully modeled, and the intention diagram coupling work is completed. Aiming at the problem of 'intention exploration single', clustering and capsule networks are introduced into a recommendation model to capture multiple intention of a user, and aiming at the problem of 'intention coupling', the invention introduces the decoupling of intention in comparison learning.
Specifically, fig. 1 and fig. 2 show detailed procedures of the sequence recommendation method based on user multi-intention comparison of the present invention. The invention firstly carries out data enhancement operation on the original sequence to obtain an enhancement sequence, and then inputs the original sequence and the enhancement sequence into the sequence encoder together to obtain an original coding vector and an enhancement coding vector. Capturing original multiple intentions in an original coding vector through K-Means clustering, capturing enhancement intentions in an enhancement coding vector through a capsule network, taking the enhancement intentions as supervision agency signals of the original vector, and finally comparing the original multiple intentions with the enhancement intentions by using a comparison learning method to realize meaning graphic coupling, finally completing a multi-intention comparison sequence recommendation model, and helping the model to accurately capture the multiple intentions of a user. Among other things, the present invention will use self-supervised learning to enhance user intent recognition in a manner that helps to improve model performance and robustness.
Aiming at the problems of single intention exploration and multi-intention coupling existing in the prior deep learning-based sequence recommendation, the invention introduces a clustering algorithm and comparison learning into a recommendation model, builds a sequence recommendation model based on intention comparison, solves the two problems to a certain extent, and improves the recommendation effect. The integral frame of the model is shown in fig. 1 and 2, and mainly comprises three parts: the first part comprises step S1 and step S2, the second part comprises steps S3, S4 and S5, and the third part is step S6.
The first part is a basic stone of the whole model, mainly embeds and codes the sequence, obtains optimal encoder parameters through training, and captures personalized preferences of users. Specifically, first, data enhancement is performed on original data to obtain an enhancement sequence, the original sequence and the enhancement sequence are respectively input into an encoder for learning, and the encoder outputs an original vector and an enhancement vector. The invention uses a sequence encoder based on a transducer model to perform sequence representation learning, wherein a multi-head attention mechanism can capture the relationship of different interaction items in the same sequence. The calculated penalty will be the next predicted master task to the sequence recommended task.
The second part mainly realizes multi-purpose exploration and decoupling, is an innovative part of the whole model and mainly comprises three algorithms of clustering, enhanced clustering and contrast learning. Specifically, the partial input is the output of the first part, namely the original coded vector and the enhanced coded vector, then the user multi-intentions (original intent vector) in the original sequence are captured using the k-means clustering algorithm, the proxy multi-intentions (enhanced intent vector) in the enhanced sequence are captured using the capsule network, and this is taken as a supervisory signal. And finally, capturing the correlation between the user multi-intention and the agent multi-intention by utilizing contrast learning, realizing the purpose of schematic diagram coupling, and improving the expression capacity of the model. To train a highly expressive sequence encoder and obtain more accurate user multi-intent characterization, latent variables and EM frameworks are introduced to solve the user intent exploration problem.
And thirdly, in order to enhance the expression capability and generalization capability of the recommendation model, the input enhanced coding vector is subjected to contrast learning, and the correlation between different views of a sequence is fused according to the principle of Mutual Information Maximization (MIM), so that the generalization capability of the recommendation model is improved.
The invention uses capital letters) Lower case letter @) And calligraphic font letters # -、) The variables, specific values and vectors in the personalized sequence recommendation problem are represented respectively, and the symbols and definitions are shown in table 1. Assuming a recommendation system having user sets and item sets, respectivelyAndThe expression is shown in (1) and (2).
（1）
（2）
Wherein the method comprises the steps ofWhich represents the presence of a certain user,Representing the total number of users,A certain item is indicated as such,Representing the total number of items. Each userHaving unique chronologically ordered interactive items, defined as a sequence of user actionsThe formula is shown as (3).
(3)
Wherein,Representing a userAt the time ofThe item of interaction is a set of items,Is the number of user interaction items.
Table 1 symbols and definitions
In a recommendation scenario, the behavior decisions of users are often driven concurrently by multiple intents. Such as: in the media arts, a user's intent to view a spanish movie may be to time, learn spanish, or make travel attacks, etc. Recommendation model definition userThe intent in the historical interaction sequence isWherein, the method comprises the steps of, wherein,Is that the user is atTime step ofAnd potential intent vectors.
Without losing generality, the model defines a sequence encoderFor user behavior sequenceEncoding and outputting a representation of the user's preferences. Thus, the next predicted target for the depth sequence algorithm can be expressed as: for each userIn a given user behavior sequenceUnder the condition of (a) searching for optimal encoder parametersMaximizing the log likelihood function of the next term for all time steps, the formula is shown as (4):
（4）
this is equivalent to minimizing the adaptive binary cross entropy loss, as shown in equations (5) and (6):
（5）
（6）
Wherein the method comprises the steps of For the sigmoid activation function,The number of small batches in the model is recommended for the sequence,For the user preference vector at the previous time step,The embedded vector representing the target interaction item (without italics),(Without italics) indicates that all are notAn embedded vector of interactive items. Sequence encoderSequencing user behaviorEach history item in (1) is projected asIn the hidden space of the dimension, equation (4) can be rewritten at this time, as shown in (7):
（7）
By measuring the user's position The intent of the step and the similarity of all candidate items in vector space, predict the user at the firstProbability of interacting with the item in the step.
The recommended tasks to be solved by the model are as follows: inputting a sequence of user behaviorsAnd item collectionOutputting a prediction function for predicting the userNext interactive itemProbability of (2)。
In the first part, in step S1, the input sequence of user actions is mainly encoded. The input sequence is an original sequence and an enhanced sequence subjected to data enhancement, wherein the data enhancement modes are masking, cutting and reordering. Given a sequenceAnd a predefined set of data transfer functionsCan createIs shown in (8):
（8）
Wherein, AndIs fromRandomly sampled transfer functions to create a sequenceIs shown in the figure). Typically, views created by the same sequence are considered to be front view pairs, and views created by any different sequence are considered to be negative view pairs.
In step S2, the original sequence is subjected to data enhancement to generate two enhancement sequences, which are jointly used as a sequence encoderInput of (a) and output of original coded vectorEnhancing coding vectors(Vector at time step by default) as input to the multi-schematic coupler. The invention uses a transducer model as a sequence encoder, which mainly consists of an embedded part, a self-attention part and a stacked-attention part, and is described in detail below.
In the embedding section, the input sequence is first of allTo a fixed length sequence. Defining the maximum length of a sequence of user actions asIf the sequence itself is longer thanConsider the most recent interaction of the userItems; if the sequence itself is not of sufficient lengthThen zero is used as a filler item to be added to the left end of the user behavior sequence so that the sequence length satisfies. In addition, an item embedding matrix is createdWhereinIn order to obtain the number of items,For potential dimensions and retrieving input embedding matricesWhereinThe constant zero vector is used as an embedding of the filler. Since the self-care mechanism does not include any loop and convolution operations, it is necessary to add a learnable position embedding in the embedding partThe final input embedding matrix is shown in equation (9):
（9）
the Self-Attention part mainly comprises a Self-Attention mechanism (Self-Attention) and a Point-Wise Feed-Forward Network (FFN), and a scalable dot-product Attention (scaled dot-product) is shown in formula (10):
（10）
Wherein, The query is represented by a representation of the query,The key is represented by a number of keys,Representing values (each row representing an item). The attention mechanism calculates a weighted sum of all values, where the querySum valueWeights and queries betweenAnd keyThe interactions between them are related. Scale factorIn order to avoid that the higher dimension is the value of the inner product too large. Wherein the self-attention mechanism uses the same objects as queries, keys and values, i.e. in an embedded matrixAs input, it is converted into three matrices by linear projection and input into a scalable dot product attention, as shown in (11):
（11）
Wherein the projection matrix Such a projection approach makes the model more flexible, e.g. the model can learn asymmetric interactions (i.e. < query-Keys (C)Sum < queryKeys (C)There may be different interactions).
While the self-attention mechanism can aggregate the embedding of all items before with adaptive weights, it is still a linear model in the end. To impart non-linearity to the model and to account for interactions between different potential dimensions. The invention uses a two-layer FFN network to realize parameter sharing, and the specific operation is shown in a formula (12):
（12）
Wherein, AndIs thatIs a matrix of the (c) in the matrix,AndIs thatAnd (5) a dimension vector.
After passing through the first self-attention block,Essentially aggregating all item embeddings in a user sequence, but learning more complex item transformations by multiple self-attention blocks may improve model capabilities, in particular, to enable stacking of self-attention blocks, i.e., attenion blocks and FFN blocks, item oneThe block may be as shown in equation (13):
（13）
Meanwhile, block 1 is represented as And. In addition, the invention introduces residual link (Residual Connections), layer normalization (Layer Normalization) and random inactivation (Dropout) to alleviate the problems of overfitting, gradient disappearance, etc., as shown in formula (14):
（14）
Wherein, Representing self-attention layers or feed forward networks, in particular layers of each block in self-attentionIn the prior pair inputApplication layer normalization followed by alignment ofThe output of (1) is Dropout, the inputAdded to the final output.
Since the core idea of residuals is to propagate the underlying features to the higher layers through the remaining connections, the model can easily propagate them to the last layer if the lower layer features are useful. Similarly, it is also useful to assume that the remaining connections are also useful; however, after a few self-attention blocks, the embedding of the last accessed item is entangled with all previous items; adding the remaining connections propagates the embedding of the last accessed item to the last layer, making it easier for the model to utilize the underlying information. Layer normalization is used to normalize the input across features (i.e., zero mean and unit variance), which is beneficial to stabilizing and accelerating neural network training. Assuming that the input is a vectorIt contains all the features of one sample, and the layer normalization definition is as shown in equation (15):
（15）
Wherein, Is an element-wise product,AndIs thatIs a function of the mean and variance of (a),AndIs a learned scale factor and bias term. To alleviate the over-fitting problem in deep neural networks, dropout regularization techniques have proven to be effective in a variety of neural network architectures. Dropout is the probability of randomly "turning off" neurons in training and using all neurons at the time of testing. And Dropout can be seen as a form of ensemble learning that considers a large number of models that share parameters (the number of neurons and input features is exponential).
The second part, the column encoder extracts as inputs the original user preference vector (original encoding vector) and the enhancement encoding vector. In order to capture the multi-intention of a user, the invention adopts an EM algorithm and a k-means clustering algorithm to learn the multi-intention of the preference vector of the original user, adopts a capsule network algorithm to learn the multi-intention in the enhancement vector, sequentially serves as a proxy signal, and finally utilizes contrast learning to complete multi-intention graphic coupling. The invention defines the first clustering through k-meansThe original intention vectors areThe first obtained through capsule networkThe enhanced intent vector is。
In step S3, the EM algorithm (iteration-Maximization Algorithm) is an iterative optimization strategy, and each iteration is divided into a hopeful step (E-step) and a maximum step (M-step) for solving the parameter estimation problem in the case of data missing. E-step estimates parameters primarily by observing the data and the existing model, and then uses this estimated parameter value to calculate the expected value of the likelihood function; m-step finds the corresponding parameters when the likelihood function is maximized. The function eventually converges because the algorithm guarantees that the likelihood function increases after each iteration. The user's interaction behavior is driven by multiple intentions, which are often unobserved, i.e., lack of tag data, and thus the EM algorithm of the present invention estimates intent variablesAnd combining the model parameters of the formula (7)And (5) optimizing. Specifically, the k-means clustering algorithm is used in step E to calculateIn time stepsA number of purposes of the present invention,Represent the firstThe personal intent vector estimated from small batch gradient descent in M stepsTo optimize parameters. In the invention, the original coding vector output in the sequence encoder is used for coding the sequenceAs input of k-means algorithm, calculateThe cluster centroids (cluster centroids) are the user's original intent vectors.
In step S4, multi-purpose extraction of the user enhancement sequence is achieved using a dynamic routing method in the capsule network (CapsNet), the enhancement encoding vectors of the users output by the sequence encoder are regarded as main capsules, the intentions of a plurality of users are regarded as intention capsules, and the dynamic routing of vector inputs and outputs of the computation capsules will be briefly described below. A capsule is a group of neurons whose activity vector represents an instantiation parameter of a particular type of entity, such as an object or an object portion. The length of the capsule output vector is the probability of the entity represented by the capsule in the current input. Assume thatIs a capsuleIs obtained by dynamic path calculationIs a vector of the next layer. First, a predictive vector is calculated, and the formula is shown as (16):
（16）
Wherein, Is a transformation matrix. At this time, capsuleThe input of (a) is all the prediction vectorsAs shown in equation (17):
（17）
Wherein, Is the coupling coefficient in the iterative dynamic routing process, capsuleAnd the sum of the coupling coefficients between all the capsules of the next layer is 1,The calculation formula is shown as (18):
（18）
Wherein, Is a capsuleAnd capsuleLog prior probability of coupling. Final capsuleThe vector calculation of (2) is shown in formula (19):
（19）
Here, the I.e. the user intention vector obtained in one iteration, i.e.。
In step S5, the present invention provides enhanced intent vectors via a capsule network for multi-purpose graphical couplingAs raw intent vectors through k-means clusteringIs implemented with a supervisory sequence encoder. In particular, contrast learning is performed between the original intent vector and the enhanced intent vector, which requires that the learned original multi-intent be infinitely close to the corresponding agent, rather than the opposite agent. The task of contrast learning can be expressed as minimizationAndWhereinIs dot product. Through the comparison learning task of the similarity between the original intention and the agent, the invention increases self-supervision on multi-intention modeling, and compared with an unsupervised method, stronger disentanglement is realized. The invention uses InfoNCE loss functions to realize contrast learning in multi-purpose graphic coupling, and the formulas are shown as (20) and (21):
（20）
（21）
In summary, the intent decoupling process learns the user's multi-intent representation on a sequence representation basis, computes multi-intent agents from the enhanced data, and directs the sequence encoder to accomplish multi-intent graphical coupling using a loss function in a self-supervised manner.
In the third part, in step S6, in order to enhance generalization ability and capture ability to multi-intentions of the model, the present invention proposes a sequence contrast learning process, taking enhancement encoding vectors as input, to enhance expressive ability of the model by maximizing mutual information between two enhancement sequences. In the present invention, for simplicity, the time step is used to connect the user's preference representation, after which we further optimize the parameters of the sequence encoder by InfoNCE lossesThe formula is shown as (22) (23):
（22）
（23）
Wherein, Is a sequence ofIs a negative view of (c).
In steps S7 and S8, the method is used for predicting and training the recommendation model. Through steps S3, 4,5, 6, a user multi-intent preference representation can be reachedItem preference representation is obtained by Multi-Layer Perceptron (MLP), i.e. item embedding matrix. For multi-intent sequence recommendation, the last time step in the interaction sequence is taken in the prediction part to make the recommendation, and the next prediction suggestion is made by calculating the next probability distribution of the whole item set. For time stepsThe next item is a candidate itemThe recommendation probability calculation of (2) is shown in the formula (24):
（24）
Wherein, Is thatAt a time stepIs provided with an embedding in the substrate,Is thatIs embedded in the memory. Finally, willUsers as model outputsFor projectsAnd sorting in descending order according to the value, and recommending the Top-K to the user.
The invention uses a multi-task training strategy training sequence recommendation model, a MSCR model is jointly optimized through a formula (25), a main next project prediction task (5) which is performed through the formula, an intention-level decoupling comparison learning task (20) and a sequence-level self-supervision comparison learning task (22) which is performed through the formula. Formally, the present invention co-trains an MSCR model sequence encoder：
（25）
Wherein,AndThe intensities of schematic coupling tasks and sequence-level self-supervision and comparison learning tasks are controlled respectively.
The specific optimization algorithm of parameters in the recommended model training process uses an Adam algorithm, an effective random optimization method which can be completed only by one step, and the algorithm dynamically calculates and adjusts individual self-adaptive learning rates of different parameters by using first moment and second moment estimation of a first-order gradient.
In order to prevent the occurrence of the overfitting phenomenon in the training process, the invention controls the training round number by using an early stop method. First, the total training wheel number is setTo prevent the model training time from being too long, the number of stop is set. Then, verifying the recommended effect of the model on the verification set by using the corresponding index in each round, storing the model parameters with the best effect when the model effect is gradually improved, and continuously verifying the index on the verification set when the model is continuously usedThe wheels descend or reach the total training wheel numberAnd stopping training, and applying the optimal parameters to the test set to obtain a test set recommendation result of the model.
In addition, in the embodiment of the invention, in order to comprehensively evaluate the MSCR model sequence encoder, experiments are performed in a Top-K recommendation scene, K items with the highest matching degree with the user are obtained through model calculation and are used as recommendation lists to recommend to the user. Therefore, experimental results analysis will be performed using the relevant evaluation index under the Top-K recommendation. The specific indexes are divided into hit ratio HitRate@K (HR@K) indexes and recall ratio Recall@K indexes for measuring the accuracy of the model recommendation result, normalized damage accumulated gain NDCG@K indexes for measuring the sequence relevance of recommended items in a model recommendation list, and the K value is set to be {5,10,20}.
Hit rate (Hit rate@k): the hit rate represents the proportion of the number of correctly recommended items of the model and is mainly used for measuring how many items liked by the user are hit in the recommendation list. The higher the hit rate, the better the recommendation effect of the model. The specific definition of hit rate is shown in formula (26):
（26）
Recall (recall @ k): the recall rate represents the ratio of the number of the correctly recommended articles of the model to the number of the true interested articles of the user, and is mainly used for measuring how many articles liked by the user are successfully recalled and recommended. The higher the recall, the better the model recommendation effect. The specific definition of recall is shown in equation (27):
（27）
Wherein, Represents the K items of most interest to the user recommended by the recommendation model,Representing a true positive sample set of interest to the user in the test set.
NDCG (Normalized Discounted Cumulative Gain): NDCG is obtained by normalizing IDCG with the loss cumulative gain DCG. The NDCG is mainly used for measuring the ordering of the articles in the recommendation list, the more interested articles of the user are more front in score, the higher the NDCG score is, and the better the recommendation effect is. NDCG is defined as shown in equation (28):
（28）
Wherein the molecule Representing the G gain in NDCG, i.e., a prediction correct score of 1 and a prediction error score of 0.Represents the accumulation of C in NDCG and denominatorTo break the scores of the goods after ranking D, finally passingNormalization processing with score of all 1 is performed.
Corresponding to the method, the invention also provides a sequence recommendation system based on user multi-intention comparison, and fig. 3 shows a logic structure of the sequence recommendation system based on user multi-intention comparison according to an embodiment of the invention.
As shown in fig. 3, the sequence recommendation system based on multi-intention comparison of users provided by the invention comprises: an enhanced sequence obtaining unit 310, configured to perform data enhancement on user behavior data to obtain an enhanced sequence; a coded vector obtaining unit 320, configured to process the enhanced sequence and the original sequence by using a sequence encoder, to obtain an original coded vector and an enhanced coded vector; an original intention vector obtaining unit 330, configured to obtain an original intention vector by performing multi-intention learning on the original encoded vector through an EM algorithm and a k-means clustering algorithm; an enhanced intention vector obtaining unit 340, configured to perform multi-intention learning on the enhanced code vector through a capsule network, to obtain an enhanced intention vector; a self-supervised contrast learning unit 350, configured to obtain similarity between the original intent vector and the enhanced intent vector by performing self-supervised contrast learning on the original intent vector and the enhanced intent vector; a sequence comparison learning unit 360, configured to obtain similarity between the original encoding vector and the enhancement encoding vector by performing comparison learning processing on the original encoding vector and the enhancement encoding vector; a user multi-intention preference expression obtaining unit 370, configured to obtain a user multi-intention preference expression according to the similarity between the original intention vector and the enhanced intention vector, and the similarity between the original code vector and the enhanced code vector; a recommendation list obtaining unit 380, configured to obtain a recommendation list and recommend the recommendation list to a user according to the multi-intention preference representation and the item preference representation of the user.
For the embodiment of the sequence recommendation system based on user multi-intention comparison provided by the invention, because the embodiment is basically similar to the embodiment of the sequence recommendation method based on user multi-intention comparison, the relevant parts are referred to in the part of the description of the method embodiment, and the description is omitted here.
According to the technical scheme, the sequence recommendation method and the system based on the multi-intention comparison of the user, provided by the invention, learn the original intention distribution of the user from the user behavior sequence through a clustering algorithm, generate the agent intention as a supervision signal by using a capsule network, introduce the user intention into a sequence recommendation model by using self-supervision comparison learning, and capture the multi-dimensional potential intention in the user sequence to the maximum extent. And the problem of single-intention and multi-intention coupling of intention exploration in the prior deep learning-based sequence recommendation is solved by introducing latent variables and an EM framework to solve the problem of intention exploration of a user and learning more accurate multi-intention characterization of the user by comparing a learning mechanism.
The sequence recommendation method and system based on user multi-intention comparison according to the present invention are described above by way of example with reference to fig. 1 to 3. It will be appreciated by those skilled in the art that various modifications may be made to the above-described user multi-intent comparison-based sequence recommendation method and system of the present invention without departing from the scope of the present invention. Accordingly, the scope of the invention should be determined from the following claims.
Claims (8)
1. The sequence recommendation method based on the multi-intention comparison of the user is characterized by comprising the following steps of:
carrying out data enhancement on user behavior data to obtain an enhancement sequence;
Processing the original sequence and the enhancement sequence of the user behavior data through a preset sequence encoder respectively to obtain an original coding vector and an enhancement coding vector;
performing multi-intention learning on the original coding vector through an EM algorithm and a k-means clustering algorithm to obtain an original intention vector, and performing multi-intention learning on the enhanced coding vector through a capsule network to obtain an enhanced intention vector;
And respectively obtaining the similarity between the original intention vector and the enhanced intention vector and the similarity between the original code vector and the enhanced code vector, wherein,
Obtaining similarity between the original intention vector and the enhanced intention vector by performing self-supervision contrast learning on the original intention vector and the enhanced intention vector, and
Obtaining the similarity between the original coding vector and the enhanced coding vector by performing self-supervision contrast learning on the original coding vector and the enhanced coding vector;
Acquiring a user multi-intention preference representation according to the similarity between the original intention vector and the enhanced intention vector and the similarity between the original coding vector and the enhanced coding vector;
acquiring a recommendation list and recommending to a user according to the multi-intention preference representation and the item preference representation of the user;
the data enhancement modes include masking, clipping and reordering, wherein a sequence is given And a predefined set of data transfer functionsCreation ofIs shown as follows:
Wherein, Indicating a particular value of the characteristic,Which represents the user's representation of the user,The time is represented by the time period of the day,AndRepresenting slaveRandomly sampled transfer functions to create a sequenceIs shown in the figure; in the process of obtaining the similarity between the original encoding vector and the enhancement encoding vector by performing self-supervised contrast learning on the original encoding vector and the enhancement encoding vector,
Further optimizing parameters of a sequence encoder by InfoNCE lossThe formula is as follows: Wherein, Representing an enhancement encoding vector; representing a dot product;
Representing sequences Is a negative view of (c).
2. The sequence recommendation method based on multi-intention comparison of users according to claim 1, wherein an original coding vector and an enhanced coding vector are obtained by processing an original sequence and an enhanced sequence of the user behavior data respectively through a preset sequence encoder; wherein,
The preset sequence encoder adopts a transducer model and comprises an embedded part, a self-attention part and a stacked-attention part, wherein,
The matrix formula of the embedded part is as follows:
Wherein, Representing the maximum length of the sequence of user actions,Representing a leachable location embedding; representing an item embedding matrix, wherein, In order to obtain the number of items,Is a potential dimension;
Embedding matrix Wherein, the method comprises the steps of, wherein,Embedding using a constant zero vector as a filler;
The self-attention portion includes a self-attention mechanism and a point-by-element feed forward network, wherein the formula of the scalable dot product attention is as follows:
Wherein, Representing a query; Representing a bond; Representing values, each row representing an item; Representing scale factors;
the self-attention mechanism uses the same objects as queries, keys and values to embed a matrix As input, it is converted into three matrices by linear projection and input into a scalable dot product attention, the formula:
wherein the projection matrix ；
The formula of the point element-by-element feed forward network is as follows:
Wherein, AndIs thatIs a matrix of the (c) in the matrix,AndIs thatA dimension vector;
the stacked attention portion includes Attenion blocks and FFN blocks, the th The formula of the block is:
meanwhile, block 1 is represented as And; Residual error linking, layer normalization and random inactivation are introduced to alleviate overfitting, and the adopted formula is as follows:
Wherein, Representing self-attention layers or feed forward networks, layers of each block in the self-attention portionIn the prior pair inputApplication layer normalization followed by alignment ofThe output of (1) is Dropout, the inputAdded to the final output;
Wherein, Is a product of the element levels,AndIs thatIs a function of the mean and variance of (a),AndIs a learned scale factor and bias term.
3. The method for sequence recommendation based on user multi-intent comparison according to claim 1,
In the process of obtaining the original intention vector by multi-intention learning of the original code vector through an EM algorithm and a k-means clustering algorithm, wherein,
The EM algorithm includes E-step and M-step, where E-step represents the desired steps for each iteration and M-step represents the maximum steps;
calculated in E-step using k-means clustering algorithm In time stepsThe number of original multi-purpose vectors,Represent the firstEstimated from small batch gradient dips in M-stepOptimizing the sequence encoder parameters.
4. The method for sequence recommendation based on user multi-purpose comparison according to claim 3,
The multi-intention learning is carried out on the enhancement coding vector through a capsule network, and the enhancement intention vector is obtained;
the length of the capsule network output vector is the probability of the entity represented by the capsule in the current input, and is assumed that Is a capsuleIs obtained by dynamic path calculationIs the next layer vector of (a);
firstly, calculating a prediction vector, and adopting the following formula:
Wherein, Is a transformation matrix;
At this time, capsule Is input as all prediction vectorsIs a weighted sum of the following formula:
Wherein, For coupling coefficient in iterative dynamic routing process, capsuleAnd the sum of the coupling coefficients between all the capsules of the next layer is 1, wherein,The formula is adopted as follows:
Wherein, Is a capsuleAnd capsuleLogarithmic prior probability of coupling, hence capsuleThe vector formula of (2) is:
Wherein, I.e. the user intent vector obtained in one iteration,。
5. The method for sequence recommendation based on user multi-intent comparison according to claim 4,
In the process of obtaining the similarity between the original intention vector and the enhanced intention vector by performing self-supervised contrast learning on the original intention vector and the enhanced intention vector,
And adopting InfoNCE loss functions to perform contrast learning, wherein the formula is as follows:
Wherein, Respectively representing the enhanced intent vector,Representing the original intent vector of the user,
、Respectively representing the minimization of the learning task,Representing dot product.
6. The method for sequence recommendation based on user multi-intent comparison according to claim 1,
In the process of acquiring a recommendation list and recommending to a user according to the multi-intention preference representation of the user and the item preference representation,
Obtaining an item embedding matrix by multi-layer perceptionWherein the item embedding matrix is the item preference representation, the user multi-intent preference representation employing；
Making a next prediction suggestion for a time step by computing a next probability distribution for the entire set of itemsThe next item is a candidate itemThe formula of the recommended probability of (2) is:
Wherein, Is thatAt a time stepIs provided with an embedding in the substrate,Is thatIs embedded in the mold;
Will be Users as model outputsFor projectsAnd sorting in descending order according to the interest degree value, and recommending Top-K to the user.
7. The method for sequence recommendation based on user multi-intent comparison according to claim 1,
The adopted preset sequence encoder is an MSCR model sequence encoder：
Wherein,AndThe intensities of schematic coupling tasks and sequence-level self-supervision and comparison learning tasks are controlled respectively.
8. A user multi-intent comparison-based sequence recommendation system, comprising:
The enhancement sequence acquisition unit is used for carrying out data enhancement on the user behavior data to acquire an enhancement sequence;
The coding vector acquisition unit is used for respectively processing the original sequence and the enhancement sequence of the user behavior data through a preset sequence encoder to acquire an original coding vector and an enhancement coding vector;
the original intention vector acquisition unit is used for acquiring an original intention vector by multi-intention learning on the original code vector through an EM algorithm and a k-means clustering algorithm;
the enhancement intention vector acquisition unit is used for carrying out multi-intention learning on the enhancement coding vector through a capsule network to acquire an enhancement intention vector;
the self-supervision contrast learning unit is used for obtaining the similarity between the original intention vector and the enhanced intention vector by carrying out self-supervision contrast learning on the original intention vector and the enhanced intention vector;
The sequence comparison learning unit is used for obtaining the similarity between the original coding vector and the enhanced coding vector through self-supervision comparison learning of the original coding vector and the enhanced coding vector;
A user multi-intention preference expression acquisition unit, configured to acquire a user multi-intention preference expression according to similarity between the original intention vector and the enhanced intention vector, and similarity between the original encoding vector and the enhanced encoding vector;
A recommendation list acquisition unit, configured to acquire a recommendation list and recommend the recommendation list to a user according to the user multi-intention preference representation and the item preference representation; the data enhancement modes include masking, clipping and reordering, wherein a sequence is given And a predefined set of data transfer functionsCreation ofIs shown as follows:
Wherein, Indicating a particular value of the characteristic,Which represents the user's representation of the user,The time is represented by the time period of the day,AndRepresenting slaveRandomly sampled transfer functions to create a sequenceIs shown in the figure; in the process of obtaining the similarity between the original encoding vector and the enhancement encoding vector by performing self-supervised contrast learning on the original encoding vector and the enhancement encoding vector,
Further optimizing parameters of a sequence encoder by InfoNCE lossThe formula is as follows:
Wherein, Representing an enhancement encoding vector; representing a dot product; Representing sequences Is a negative view of (c).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410437551.8A CN118037362B (en) | 2024-04-12 | 2024-04-12 | Sequence recommendation method and system based on user multi-intention comparison |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410437551.8A CN118037362B (en) | 2024-04-12 | 2024-04-12 | Sequence recommendation method and system based on user multi-intention comparison |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN118037362A CN118037362A (en) | 2024-05-14 |
| CN118037362B true CN118037362B (en) | 2024-07-05 |
Family
ID=90991781
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410437551.8A Active CN118037362B (en) | 2024-04-12 | 2024-04-12 | Sequence recommendation method and system based on user multi-intention comparison |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN118037362B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113204952A (en) * | 2021-03-26 | 2021-08-03 | 南京邮电大学 | Multi-intention and semantic slot joint identification method based on clustering pre-analysis |
| CN116304984A (en) * | 2023-03-14 | 2023-06-23 | 烟台大学 | Multi-modal intention recognition method and system based on contrast learning |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019050474A1 (en) * | 2017-09-11 | 2019-03-14 | Hitachi, Ltd. | Multiple intent interpreter and recommender |
| CN111984780A (en) * | 2020-09-11 | 2020-11-24 | 深圳市北科瑞声科技股份有限公司 | Multi-intention recognition model training method, multi-intention recognition method and related device |
| CN114840747B (en) * | 2022-04-18 | 2024-08-23 | 清华大学深圳国际研究生院 | News recommendation method based on contrast learning |
| CN116401542A (en) * | 2023-03-20 | 2023-07-07 | 华南理工大学 | Multi-intention multi-behavior decoupling recommendation method and device |
-
2024
- 2024-04-12 CN CN202410437551.8A patent/CN118037362B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113204952A (en) * | 2021-03-26 | 2021-08-03 | 南京邮电大学 | Multi-intention and semantic slot joint identification method based on clustering pre-analysis |
| CN116304984A (en) * | 2023-03-14 | 2023-06-23 | 烟台大学 | Multi-modal intention recognition method and system based on contrast learning |
Non-Patent Citations (1)
| Title |
|---|
| 胶囊网络用于短文本多意图识别的研究;刘娇等;计算机科学与探索;20201231;第14卷(第10期);1735-1743 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN118037362A (en) | 2024-05-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112598462B (en) | Personalized recommendation method and system based on collaborative filtering and deep learning | |
| Wu et al. | Session-based recommendation with graph neural networks | |
| US20210004682A1 (en) | Adapting a sequence model for use in predicting future device interactions with a computing system | |
| Nilashi et al. | A multi-criteria collaborative filtering recommender system for the tourism domain using Expectation Maximization (EM) and PCA–ANFIS | |
| Chen et al. | Efficient ant colony optimization for image feature selection | |
| Zhao et al. | Variational self-attention network for sequential recommendation | |
| Wang | Bankruptcy prediction using machine learning | |
| CN110955826B (en) | Recommendation system based on improved cyclic neural network unit | |
| CA2634020A1 (en) | System and method for multi-level online learning | |
| Yu et al. | Control chart recognition based on the parallel model of CNN and LSTM with GA optimization | |
| CN110866145A (en) | Co-preference assisted deep single-class collaborative filtering recommendation method | |
| KR102070049B1 (en) | Apparatus and method for collaborative filtering using auxiliary information based on conditional variational autoencoder | |
| US20230061517A1 (en) | Verification of the Authenticity of Images Using a Decoding Neural Network | |
| Le et al. | A multi-criteria collaborative filtering approach using deep learning and Dempster-Shafer theory for hotel recommendations | |
| CN115760271A (en) | Electromechanical commodity personalized recommendation method and system based on graph neural network | |
| CN115760270A (en) | Novelty enhancement recommendation method based on knowledge graph | |
| CN118037362B (en) | Sequence recommendation method and system based on user multi-intention comparison | |
| Jha et al. | A trustworthy model of recommender system using hyper-tuned restricted boltzmann machine | |
| Omidvar et al. | Tackling cold-start with deep personalized transfer of user preferences for cross-domain recommendation | |
| CN117252665A (en) | Service recommendation method and device, electronic equipment and storage medium | |
| CN114912984B (en) | Self-attention-based time scoring perception context recommendation method and system | |
| CN116757747A (en) | Click rate prediction method based on behavior sequence and feature importance | |
| Tran et al. | Improvement graph convolution collaborative filtering with weighted addition input | |
| CN114565436A (en) | Vehicle model recommendation system, method, device and storage medium based on time sequence modeling | |
| Behara et al. | Integrating metadata into deep autoencoder for handling prediction task of collaborative recommender system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |