CN118656550B - Location-aware collaborative recommendation method and device - Google Patents

Abstract

The invention discloses a location-aware collaborative recommendation method and a device, wherein the method comprises the following steps: acquiring a user and article node position code, converting the user and article node position code dimension into a node embedding dimension and injecting the node embedding dimension into the node initial embedding; feeding the node with the embedded position codes to each transducer layer of Graph Transformer for calculation to respectively obtain a local interaction signal and a global collaboration signal of the target node; the local interaction signals and the global collaboration signals are fused and output to obtain node representations of each layer of Graph Transformer; and inputting the L-layer representation of the node into a reading function to obtain the final representation of the node, and obtaining the target recommendation of the user. The invention utilizes Graph Transformer technology to alleviate the generalized deviation limitation faced by the graph recommendation method. The modeling capability of the attention calculation module on the node relation is effectively improved by adopting the node position coding technology, the capturing capability of the model on the global node relation is enhanced, the problem of induction deviation of the graph model is further solved, and the generalization effect of the graph neural network is improved.

Description

Location-aware collaborative recommendation method and device

Technical Field

The application relates to the technical field of recommendation systems, in particular to a location-aware collaborative recommendation method and device.

Background

Recommendation systems have gained increasing importance as an integral tool in a variety of real world applications, helping users to explore and find items of interest. Recently, a graph recommendation model has attracted considerable attention in the field of recommendation systems because it can capture complex interactions between users and items. By exploiting the modeling capabilities of the graph neural network, these methods generally surpass traditional recommendation techniques. Although graph recommendation models achieve advanced performance, they still suffer from an important drawback: the problem of serious generalized bias (Inductive Bias) is faced due to the excessive dependence on the local neighbor aggregation technology, and the capability of capturing global cooperative signals is insufficient, so that the generalization performance of the model is weak. While the graph recommendation model may capture long-range/global synergy signals through multi-round neighbor aggregation, it suffers from Over-smoothing and Over-compression (Over-squashing), resulting in suboptimal node characterization.

Currently, many techniques are directed to solving the problem of local information restriction faced by graph recommendation models. One class of methods is based on enhanced interaction graph training to additionally explore and learn user preferences. The EGLN approach implements this idea by capturing potential interactions using trained user and item characterizations, augmenting the original interaction graph, and then retraining the graph recommendation model. Still further, graphDA also builds on user-user and item-to-item links to capture implicit relationships between more nodes, exploring more possible interactions. Yet another class of methods implicitly utilizes more data based on self-supervised learning graph models, enabling exploration and utilization of potential interactions. GFormer method provides a generated self-supervision learning automation data enhancement, extracts a user-object interaction mode with information quantity, realizes the selective enhancement of data in a parameterized mode, reserves a globally perceived user-object relationship, and helps to alleviate the problem of induction deviation faced by a graph recommendation model.

The prior art aims at expanding the learning capability of a graph recommendation model by enhancing an interaction graph or enhancing the generalization performance of the model by implicitly encoding global information into node characterization of users and articles, and the methods firstly cannot learn the relationship between node pairs; secondly, the node position information is lacked, but reliable global information is difficult to mine, so that the problem of induction deviation of the graph recommendation model is difficult to effectively relieve.

Disclosure of Invention

The embodiment provides a location-aware collaborative recommendation method and device, which are used for solving the problems that in the related technology, the relation cannot be learned from node to node, node location information is lacked, reliable global information is difficult to mine, and the induction deviation of a graph recommendation model is difficult to effectively relieve.

In a first aspect, an embodiment of the present invention provides a location-aware collaborative recommendation method, where a user and an item node location code are obtained, and the user and item node location code dimension is converted into a node embedding dimension and is injected into node initial embedding;

Feeding the node with the embedded position codes to each transducer layer of Graph Transformer for calculation to respectively obtain a local interaction signal and a global collaboration signal of the target node;

the local interaction signals and the global collaboration signals are fused and output to obtain Graph Transformer node representations of each layer;

and inputting the L-layer representation of the node into a reading function to obtain a final representation of the node, calculating the vector inner product of the final representation of the node of the user and the article, and sequencing the articles to obtain the target recommendation of the user.

In an alternative embodiment, the user and item node position code is one of a spectrum code, a degree code, a PageRank code, and a type code.

In an alternative embodiment, the Graph Transformer includes L Transformer layers, each having a graph convolution module for capturing local interaction signals of the target node and a NodeFormer convolution module for capturing global collaboration signals of the target node.

In an alternative embodiment, feeding the node with embedded position codes to the respective transducer layer computations of Graph Transformer to obtain the local interaction signal and the global collaboration signal of the target node, respectively, includes:

feeding the node to Graph Transformer;

the local interaction signal is captured by the graph convolution module, and the global collaboration signal is captured by the NodeFormer convolution module.

In an alternative embodiment, the obtaining Graph Transformer node representations of each layer after the local interaction signal and the global collaboration signal are fused and output includes:

fusing the local interaction signals and the global collaboration signals according to a certain fusion proportion and outputting the local interaction signals and the global collaboration signals to obtain target node representation of the current transducer layer;

And calculating by the graph convolution module and the Nodeformer convolution modules of the L Transformer layers to obtain node representation of each layer.

In an alternative embodiment, inputting the L-level representation of the node into the read function to obtain a final representation of the node, calculating and ordering the items via the user and item node final representation vector inner products, the obtaining the target recommendation of the user comprising:

finally representing the vector inner product by the user and the object node to obtain a plurality of values with highest inner product results;

and obtaining target recommendation of the user.

In an alternative embodiment, the read-out function may be replaced by an arbitrary graph recommendation model.

Compared with the prior art, the position sensing collaborative recommendation method has the following beneficial effects:

The invention utilizes Graph Transformer technology to alleviate the generalized deviation limitation faced by the graph recommendation method. The modeling capability of the attention calculation module on the node relation is effectively improved by adopting the node position coding technology, the capturing capability of the model on the global node relation is enhanced, the problem of induction deviation of the graph model is further solved, and the generalization effect of the graph neural network is improved.

In a second aspect, the present invention provides a location-aware collaborative recommendation device, including:

the preprocessing injection module is used for acquiring the node position codes of the user and the article, converting the node position code dimensions of the user and the article into node embedding dimensions and injecting the node embedding dimensions into the node initial embedding;

The node calculation module is used for feeding the node with the embedded position codes to each transducer layer of Graph Transformer for calculation to respectively obtain a local interaction signal and a global collaboration signal of the target node;

The node updating module is used for obtaining Graph Transformer node representations of each layer after the local interaction signals and the global collaboration signals are fused and output;

And the recommendation output module is used for inputting the L-layer representation of the node into a reading function to obtain a final representation of the node, calculating the vector inner product of the final representation of the node of the user and the article, and sequencing the articles to obtain the target recommendation of the user.

In a third aspect, the present invention provides an electronic device comprising a processor, a communication interface, a memory and a bus, wherein the processor, the communication interface, the memory are in communication with each other via the bus, the processor being operable to invoke logic instructions in the memory to perform the steps of the method as provided in the first aspect.

In a fourth aspect, the present invention provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the location-aware collaborative recommendation method according to the first aspect.

Compared with the prior art, the location-aware collaborative recommendation device, the electronic equipment and the storage medium have the same beneficial effects as the location-aware collaborative recommendation method in the first aspect, so that the description is omitted here.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flowchart of a location-aware collaborative recommendation method in an embodiment of the present invention;

FIG. 2 is a schematic diagram of a PGTR overall frame in an embodiment of the present invention;

FIG. 3 is a graph showing comparison of performance of PGTR under different sparsity levels in an embodiment of the present invention;

FIG. 4 is a graph showing comparison of performance of PGTR under different noise levels in an embodiment of the present invention;

FIG. 5 is a graph showing performance under different position codes according to an embodiment of the present invention;

FIG. 6 is a block diagram illustrating a collaborative recommendation device with location awareness according to an embodiment of the present invention;

Fig. 7 is a block diagram of an electronic device according to an embodiment of the present invention.

Detailed Description

The application is described and illustrated below with reference to the drawings and examples for a clearer understanding of the objects, technical solutions and advantages of the application.

Unless defined otherwise, technical or scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terms "a," "an," "the," "these" and similar terms in this application are not intended to be limiting in number, but may be singular or plural. The terms "comprising," "including," "having," and any variations thereof, as used herein, are intended to encompass non-exclusive inclusion; for example, a process, method, and system, article, or apparatus that comprises a list of steps or modules (units) is not limited to the list of steps or modules (units), but may include other steps or modules (units) not listed or inherent to such process, method, article, or apparatus. The terms "connected," "coupled," and the like in this disclosure are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The term "plurality" as used herein means two or more. "and/or" describes an association relationship of an association object, meaning that there may be three relationships, e.g., "a and/or B" may mean: a exists alone, A and B exist together, and B exists alone. Typically, the character "/" indicates that the associated object is an "or" relationship. The terms "first," "second," "third," and the like, as referred to in this disclosure, merely distinguish similar objects and do not represent a particular ordering for objects.

In the embodiment of the invention, a Position-aware collaborative recommendation method (Position-AWARE GRAPH Transformer for Recommendation, PGTR) is provided, which is a Position-aware Graph Transformer recommendation method, and Graph Transformer technology is applied to a collaborative recommendation scene to model the relationship between a user and an article and capture global collaborative signals. The PGTR method is centered on designing a node position code for characterizing and learning the user's relationship to the item. Specifically, four interactive characteristics of a user and an object in a recommended scene are carefully considered, and four corresponding node position codes are constructed. Using these codes can effectively capture and learn the relationship between the user and the item, exploring more sophisticated user preferences.

In accordance with the problem definition, and the proposed PGTR method of the present invention, it can be understood on another level that a model PGTR is constructed, which model PGTR needs to be described in connection with specific location coding, problem definition, and coding design.

First, the basic symbols that need to be used are described.Representing a slave datasetA user-item interaction matrix constructed in (1), whereinAndIndicating the number of users and items, respectively.AndRepresenting a user set and an item set, respectively.Each item of (2)Are binary values of 0 or 1 indicating whether the user interacted with the item. Further, based on the interaction matrixCreating a user-object bipartite graphWhereinConsisting of a collection of user nodes and item nodesRepresenting a collection of edges. If it isThere is an edge between the user and the item.Representing a first order neighbor item (user). Degree) Is the degree of the node on the graph.Is an adjacency matrix. The present invention is directed to predicting K items that a user may interact with given a user item interaction record. Thus, at a known pointOr (b)Under the condition of (a), the invention needs to construct a prediction model f (PGTR model) to realize that each user is in an article setAn efficient prediction of potential interactions in the system.

Fig. 1 is a flowchart of a location-aware collaborative recommendation method according to the present invention, as shown in fig. 1, the flowchart includes the following steps:

s100, acquiring a user and article node position code, converting the user and article node position code dimension into a node embedding dimension and injecting the node embedding dimension into node initial embedding;

In this embodiment, four kinds of position codes are exemplified, which are one of spectrum codes, degree codes, pageRank codes, and type codes. In particular as described below. Location identification and awareness in node representations, in particular for a given node (user or item), by converting a location encoding dimension into a node embedding dimension and injecting into the node initial embedding The method comprises the following steps:

,

Wherein, Is a nodeIs used to determine the initial embedded representation of (c),Control the intensity of position information andIs a nodeIs a position representation of (2);

specifically, a node The position characterization of (c) is as follows:

，

In the formula, AndTo learn parameters for transformations that are different for an item than for a userAndIs a learnable transformation matrix; transforming the corresponding position coding dimension into a node embedding dimension; Is a specific position code used to characterize node position information from different angles.

S200, feeding the node with the embedded position codes to each transducer layer of Graph Transformer for calculation to respectively obtain a local interaction signal and a global collaboration signal of the target node;

It should be noted that Graph Transformer includes L convertors layers, each having a graph convolution module and a NodeFormer convolution module, where the graph convolution module is used to capture a local interaction signal of a target node, and the NodeFormer convolution module is used to capture a global collaboration signal of the target node.

Specifically, feeding the node with the embedded position code to each transducer layer of Graph Transformer for calculation, and respectively obtaining a local interaction signal and a global collaboration signal of the target node includes:

feeding the node to Graph Transformer;

The local interaction signal is captured by the graph convolution module and the global interaction signal is captured by the NodeFormer convolution module.

Illustratively, at any%Transformer layer, nodeInput representation of (a)Fed into a graph convolution module (Graph Transformer, a single graph convolution operation process) and a Nodeformer convolution module (a transform operation process applied to Graph Transformer), respectively, to realize target nodesCapturing of local interaction signals and global collaboration signals. Specifically, there are the following steps:

，

Wherein, AndThe combination function and the aggregation function in the graph convolution module are respectively realized in various modes, and any mode of realizing a traditional graph recommendation model can be directly adopted.

This step is directed to the target node by the graph convolution moduleAnd (3) capturing local interaction signals, and accurately describing the interaction mode of the target node in the local. Thereafter, for the acquisition nodeModeling a potential global interaction pattern of a target node by adopting a Nodeformer convolution module, wherein the specific implementation mode is defined as follows:

，

Wherein, . This means that the position information is re-injected into the output representation of the graph convolution module, while also controlling the injection strength.Is a low-dimensional feature mapping function with random transformation for arbitrary inputsThe method comprises the following steps:

，

， Is an h-dimensional unit array.

S300, the local interaction signals and the global collaboration signals are fused and output to obtain node representations of each layer of Graph Transformer;

further, the step of obtaining Graph Transformer node representations of each layer after the local interaction signals and the global collaboration signals are fused and output includes:

fusing the local interaction signals and the global collaboration signals according to a certain fusion proportion and outputting the local interaction signals and the global collaboration signals to obtain target node representation of the current transducer layer;

And calculating by the graph convolution module and the Nodeformer convolution modules of the L Transformer layers to obtain node representation of each layer.

Finally, through the fusion nodeThe local interaction signal and the global interaction signal of the system can be output to obtain the current first timeThe target node j of the transducer layer represents:

，

Wherein the method comprises the steps of The mixing ratio between the local interaction signal and the global interaction signal is controlled. After the calculation operations of the L Tansformer-layer graph convolution modules and the Nodeformer convolution modules, the PGTR method obtains the final node representation.

S400, inputting the L-layer representation of the node into a reading function to obtain a final representation of the node, calculating the vector inner product of the final representation of the node of the user and the article, and sequencing the articles to obtain the target recommendation of the user.

Specifically, inputting the L-layer representation of the node into a reading function to obtain a final node representation, calculating a vector inner product of the final node representation of the user and the object node, and sequencing the objects, wherein obtaining the target recommendation of the user comprises the following steps:

finally representing the vector inner product by the user and the object node to obtain a plurality of values with highest inner product results;

and obtaining target recommendation of the user.

It should be noted that the readout function may be replaced by an arbitrary graph recommendation model.

It should be further noted that the final node represents:

。

Wherein, To characterize the read-out function, it can also be implemented using the read-out function of the arbitrary graph recommendation model. The final node representation is input into the read Function (Readout Function). The role of the read-out function is to convert the node level features into a feature representation of the whole graph or sub-graph. In a recommendation system, this typically means converting the representations of user nodes and item nodes into a form that can be used to predict user preferences for items.

The representation vectors of users and items are typically multi-dimensional real vectors, which can be regarded as points in a high-dimensional space. Inner product (also known as dot product) computation is one type of vector operation that can measure the direction and length of two vectors. The inner product values may be ordered by obtaining the inner product values of all items by the user. The goal of the ranking is typically to find the top N items that are most interesting to the user, i.e., those items with the largest inner products. This process can be achieved simply by selecting the top few values in the inner product result.

The invention designs four position coding schemes altogether, characterizes the positions of nodes in a graph network from different angles, namely spectrum coding, degree coding, pageRank coding and type coding, combines the problem definition to carry out coding design, and essentially builds a PGTR model by combining the PGTR method, specifically builds a PGTR model, firstly determines the problem definition, for example, builds a user-object interaction matrix, defines a user and object set and an interaction relation; creating a user-object bipartite graph, wherein nodes comprise users and objects, and edges represent interaction relations; the goal is to predict K items that a user is likely to interact with after a given user item interaction record.

Next is a position coding design, which designs four position coding schemes, including spectral coding, degree coding, pageRank coding, and type coding, to characterize the position of nodes in the graph network from different angles. The spectrum coding is based on the Laplace feature vector and is used for coding the relative position information among the nodes. The degree codes are coded according to popularity of the articles and liveness of users, and reflect the quality of the articles and the behavior patterns of the users. PageRank codes reflect the importance and influence of nodes in the graph based on PageRank values. The type code distinguishes between different attributes and relationship patterns of the user and the item.

Thirdly, constructing PGTR an overall framework, and acquiring a final node representation to complete PGTR overall framework construction by combining the position-aware collaborative recommendation method;

it should be noted that the PGTR model was finally trained by Sampling Softmax (SSM) loss, namely:

，

Wherein the method comprises the steps of ，AndUser node characterizationAnd item node characterizationIs a normalized representation of the number (f),Is the temperature coefficient of the temperature of the material,Is a negative sample of user u. Here, the negative sampling technique in the batch is adopted, and positive samples of other users in the same training sample batch are taken as negative samples, so that parallel calculation is convenient. Training PGTR is accomplished by optimizing the above equation.

Four position coding schemes are described one by one.

The first is spectral encoding, which is commonly used in the atlas-domain in the task of atlas learning. Such position coding may encode relative position information between nodes, and neighboring nodes will have similar coding. The present invention uses this method to encode the location information for each user and item in the user-item interaction map. Specifically for interactive adjacency matrixIts laplace feature vector is calculated in advance.

,

Wherein, Is thatIs used for the degree matrix of the (c),Is an identity matrix of the unit cell,AndThe eigenvalues and eigenvectors, respectively. In general, the number of the devices used in the system,The eigenvector of the smallest non-trivial eigenvalue is used as the laplace position code. Thus from the feature vectorObtaining Laplace position codesWhereinAndLaplace position codes representing the user and the item, respectively.

In an actual recommendation scenario, establishing user-user and item-item relationships is of great significance. Thus, the present invention encodes the location information for each user and item on the user-side and item-side interaction map, respectively. In particular, the present invention constructs a user side adjacency matrixAnd article side adjacency matrixWherein the element is 0 or 1, indicatingThe presence or absence of a second order connection. Then, there are:

Also, the invention selects The feature vector corresponding to the smallest non-trivial feature value serves as a laplace position code for the user and the item on the respective interaction map (user side and item side). I.e. slaveObtained in (1)From the slaveObtained in (1). Further, there are. Finally, the invention combinesAndCreating spectral codes, namely:

，

Wherein, Is a super-parameter for controlling the coding proportion of different Laplace positions,. Each user (item) can be determined by its spectral coding on the user item interaction map and the user side interaction map (item side interaction map), respectively, helping the model to more clearly understand the relationship between the user and the item.

The second is a degree code, in a recommendation system, popularity of the item and liveness of the user are important indexes reflecting the quality of the item and the behavior pattern of the user. From the item perspective, when two items have similar popularity, they tend to appear in similar contexts or scenes, such as in similar recommendation lists, purchase histories, or social circles. As another example, some users have been attracted to popular songs while others focus on the masses or the less well known music. Thus, items of similar popularity are more likely to exhibit connectivity in the user behavior data. From the user's perspective, when the liveness of two users is similar, it means that they exhibit similar patterns of behavior and engagement on the platform. For example, based on similar preferences, they may co-discover and explore similar items, or focus on each other on a social network. Based on this, the present invention designs a degree code for each user and item.

Specifically, for the articles, each articleAll have an associated popularityThus, there may be:

，

the invention sorts the articles according to the ascending order of popularity of the articles and evenly divides the articles into And groups. For item i, the degree code is:

,

Wherein, Is a degree packet encoder of an item,Is the group to which item i belongs. Finally, a degree code is obtained for all items, expressed as，Is the encoding dimension.

Similarly, for users, each userAll have corresponding liveness Degree A kind of electronic device；

The invention sorts the users according to the ascending order of the activity of the users and evenly divides them intoA group. The degree code for user u is:

Wherein the method comprises the steps of Is the user's degree packet encoder,Is the group to which user u belongs. Further, the degree codes of all users can be obtained。

It should be noted that such coding is not entirely effective for strategically located items (users) within a group, as there are some items (users) with similar popularity (liveness) located in different groups. But in general, the model is adjusted byTo enable the degree coding.

The third is the PageRank code, which can be used to describe the importance or impact of items and users on the interaction graph. The PageRank value of an item (or user) is not only related to its own degree, but also depends on the importance of its neighbors in the graph. When two items have similar PageRank values, this means that they are liked by the user. If a user likes one of the items, they are also likely to like the other item. On the other hand, when the PageRank values of two items are significantly different, favoring one of them does not mean favoring the other. For users, when two users have similar PageRank values, their interaction patterns tend to be similar. To this end, the invention proposes PageRank coding.

The invention pre-calculates the interaction diagram of all users and objectsPageRank value above. Then, each item i has a PageRank valueFurther, there may be:

；

The invention sorts the articles in ascending order of their PageRank values and then uniformly divides the articles into Nr groups. For item i, its PageRank position code is:

，

Wherein the method comprises the steps of Is an article PageRank block encoder,Is the group to which item i belongs. PageRank position coding for all items available，Is the encoding dimension. Similarly to the degree code, from the user's point of view, the invention can also obtain PageRank position codes corresponding to all usersAnd will not be described in detail herein.

Implementing encoding by grouping faces similar challenges as degree encoding. However, by adjusting the number of groupings, the present invention can also achieve its effectiveness in building the relationship between the user and the item. For simplicity, the present invention sets the number of groups of both items and users to Nr.

The fourth is type coding, in a recommendation system, the user and the item differ in relation pattern and attribute. For example, items presented on a website or application (e.g., organized by type or popularity) are not the same way that users interact (e.g., click at will). By differentiating the node types in the graph, interactions and connections between users and items can be better understood. To this end, the invention proposes a type coding based on the node type.

In the scheme, type position codes are allocated to the article group and the user group respectively. All items share the same type of position coding, as do all users, namely:

,

Wherein the method comprises the steps of Is a node type encoder. The invention can obtain the type position codes of all the articles and users, which are respectively expressed asAndIs the encoding dimension.

In general, the four position encodings presented herein contemplate relationships between users and items from different perspectives, enabling models to capture and model various relationships. By encoding four corresponding positions of the target node jIn the above equation, PGTR models can learn information related to characterization.

According to the method, local interaction information is learned by using the graph recommendation model, the node representation after fusion of the position codes and the local feature codes is sent to a transducer module, reliable node relation is captured, global node information is aggregated, and further the problem of induction deviation of the graph recommendation model is effectively relieved by adopting a local-global collaborative aggregation technology. The invention proves that the proposed method has remarkable effect in relieving the problem of induction deviation and improving the generalization capability of the graph recommendation model by widely verifying on a plurality of real-world data sets. The data set statistics are shown in table 1. To simulate varying degrees of sparsity, the present invention divides the data set into 20%/40%/60%/80% as the training set. In the training set, 20% was used as the validation set for adjusting the hyper-parameters. The remaining 80%/60%/40%/20% are considered the test set. In the model evaluation stage, the invention adopts a full ordering protocol (All-ranking Protocal) to evaluate the top-K recommendation performance. For each user in the test set, all items that have not interacted in the training set will be candidates, and all items that have actually interacted will be scored as positive samples. The invention adopts evaluation indexes of recall@20 and NDCG@20 to illustrate the beneficial effects of the method. Next, the beneficial effects of the proposed method will be introduced from the aspects of prediction accuracy, generalization and robustness, and method rationality, respectively.

Table 1: data set statistics

First, overall performance comparison:

Table 2 shows the overall performance of the method of the invention and all baselines from which the invention concludes the following observations. First, lightGCN performance is generally better than NGCF, but is weaker than NGCF on Amazon-elec and Douban-book datasets, possibly because the two datasets have more complex node interactions, while NGCF architecture is easier to model and capture. On the remaining data sets LightGCN and UltraGCN achieve an effective boost. UltraGCN approximates infinite layer graph convolution at training, and also uses the item-to-item relationship matrix to enhance the synergistic signal, which allows it to achieve optimal performance over multiple datasets. Secondly, it can be observed that the SGL is trained by adopting a self-supervision learning mode based on LightGCN architecture, the performance of the SGL exceeds LightGCN on Amazon-elec, douban-book and LastFM, and the superiority of self-supervision learning is verified. While EGLN and GFormer are also based on self-supervised learning, although claimed to use global collaborative signals, in extremely sparse scenarios the weakness of building global relationships still exists. This suggests that a practical knowledge of node relationships needs to be proposed to guide the model in capturing the co-ordination signal in this case. Finally, the method provided by the invention achieves significant performance improvement on various reference models. In one aspect, this is because the present invention can guide the model learning process by incorporating a priori knowledge of the user's and item relationships into the model. On the other hand, the transducer architecture enables the model to achieve real global collaborative signal capture, and the limitation of the traditional graph recommendation model is overcome.

The second is generalization and robustness:

In order to evaluate generalization of the invention under different sparsity of data, the present section performs model training on training data of different sparsity and tests model performance. In particular, experiments were performed here on different training set scales (20%, 40%, 60%, 80% of the original data set for training, the remainder for testing), simulating varying degrees of sparsity. Contrast methods LightGCN and GFormer, lightGCN are used to compare the backbone model of the present method, GFormer is the most current technique that uses a technique that samples a set of anchor nodes as global information agents in order to capture global collaboration signals. Figure 3 shows performance results on both Douban-book and LastFM data sets. It can be observed from the figure that the method provided by the invention is always better than LightGCN and GFormer under different sparseness. It is believed that using PGTR to explicitly model user-item relationships and capture global collaborative signals is efficient and necessary in the current sparse data case, PGTR can help to refine the exploration and modeling of user preferences and thus can alleviate the problem of generalized bias faced by graph recommendation models.

To demonstrate the robustness of the proposed method to noise in cases where the data contains noise, this section randomly inserts different proportions (10%, 20%, 30%) of noise interactions in the training set. Noise interactions are defined herein as randomly sampling some items from the items that the user has not interacted with and introducing them into the user's interaction history as virtual interaction behavior. Because the collection of items is typically large, these noisy interacting items do not typically coincide with the items that are actually of interest to the user. As shown in fig. 4, fig. 4 shows the performance comparison result, wherein the relative performance index is adopted, that is, the performance when noise is not added is set to be 1, and the performance after noise is added is recorded according to the relative value. First, it can be observed that LightGCN has the most severe performance degradation on both datasets (Amazon-elec and Douban-book). This suggests that the use of such fixed neighbor aggregation techniques is susceptible to noise resulting in reduced performance and that improvements in the aggregation techniques are necessary. Second, SGLs and GFormer can be observed to have some resistance to noise, which demonstrates the robustness of self-supervised learning against noise. The final result shows that the performance of the method is least reduced on two data sets, which indicates that the proposed method has good noise recognition capability when processing interactive data, and the noise interaction can be well recognized by adopting node relation design and capturing interaction between nodes by adopting a transducer architecture. Thus, the PGTR method can integrate multiple location information to determine the authenticity of the interaction, which is the source of the advancement of the present invention.

Finally, the method is reasonable, and in order to verify the core design rationality of the method PGTR provided by the application, experimental verification is carried out on four position codes in this section. Specifically, the four types of position codes and All position codes are individually removed from the PGTR model, respectively, and the performance after removal is tested, wherein, "-PL" represents the removal spectrum code, "-DG" represents the removal degree code, "-PR" represents the removal PageRank code, "-TP" represents the removal node type code, and "-All" represents the removal of All position codes. The present application demonstrates performance results on Amazon-elec and Douban-book datasets, as shown in FIG. 5. In general, it can be seen that there is a different magnitude of degradation in recommended performance after each position code is removed alone. This means that each position code contributes to the PGTR model. In particular, the performance is greatly reduced when all the position codes are removed, which fully demonstrates the importance of the position codes. The contribution of each position code to the performance is different. This suggests that they play different roles in building user-item relationships in different aspects. For example, on Amazon-elec data, the degree coding is more obvious than the other three, which indicates that on this data, this information is of great importance, since the user may see the sales and heat of electronic products when they are purchased, resulting in some strong link between the products. PageRank code showed a strong effect on Douban-book datasets. Since PageRank coding is an indicator of the importance of a node, it is not only composed of hotness, but also of neighbors. It can be seen that on a cricket website, the interaction pattern of the user may be influenced by different important authors or persons, thus having a strong effect in showing the code. Different codes exhibit different roles on different data, essentially because the data sets are acquired from the platforms, while users of different platforms obviously have different interaction modes, even though the push modes of the items are different. In general, the method provided by the application utilizes four kinds of position coding information, captures different interaction relations between users and articles, and strengthens the recommendation performance of the model in different scenes.

Table 2: overall performance comparison

The invention utilizes Graph Transformer technology to alleviate the generalized deviation limitation faced by the graph recommendation method. The modeling capability of the attention calculation module on the node relation is effectively improved by adopting the node position coding technology, the capturing capability of the model on the global node relation is enhanced, the problem of induction deviation of the graph model is further solved, and the generalization effect of the graph neural network is improved.

The embodiment of the invention also provides a collaborative recommendation device for location awareness, which is used for realizing the method embodiment, and is not described in detail. The terms "module," "unit," "sub-unit," and the like as used below may refer to a combination of software and/or hardware that performs a predetermined function. Although the means described in the following embodiments are preferably implemented in software, implementation of hardware or a combination of software and hardware is also possible and contemplated.

As shown in fig. 6, fig. 6 is a block diagram of a location-aware co-recommender according to the present invention, which comprises:

the preprocessing injection module is used for acquiring the node position codes of the user and the article, converting the node position code dimensions of the user and the article into node embedding dimensions and injecting the node embedding dimensions into the node initial embedding;

The node calculation module is used for feeding the node with the embedded position codes to each transducer layer of Graph Transformer for calculation to respectively obtain a local interaction signal and a global collaboration signal of the target node;

The node updating module is used for obtaining Graph Transformer node representations of each layer after the local interaction signals and the global collaboration signals are fused and output;

And the recommendation output module is used for inputting the L-layer representation of the node into a reading function to obtain a final representation of the node, calculating the vector inner product of the final representation of the node of the user and the article, and sequencing the articles to obtain the target recommendation of the user.

The Graph Transformer technique is utilized to alleviate the generalized bias limitations faced by the graph recommendation method. The modeling capability of the attention calculation module on the node relation is effectively improved by adopting the node position coding technology, the capturing capability of the model on the global node relation is enhanced, the problem of induction deviation of the graph model is further solved, and the generalization effect of the graph neural network is improved.

Fig. 7 is a block diagram of an electronic device according to an embodiment of the present invention, where, as shown in fig. 7, the electronic device may include: processor 610, communication interface (Communications Interface) 620, memory 630, and communication bus 640, wherein processor 610, communication interface 620, memory 630 communicate with each other via communication bus 640. The processor 610 may call logic instructions in the memory 630 to perform the following methods:

acquiring a user and article node position code, converting the user and article node position code dimension into a node embedding dimension and injecting the node embedding dimension into node initial embedding;

Feeding the node with the embedded position codes to each transducer layer of Graph Transformer for calculation to respectively obtain a local interaction signal and a global collaboration signal of the target node;

the local interaction signals and the global collaboration signals are fused and output to obtain Graph Transformer node representations of each layer;

and inputting the L-layer representation of the node into a reading function to obtain a final representation of the node, calculating the vector inner product of the final representation of the node of the user and the article, and sequencing the articles to obtain the target recommendation of the user.

Further, the logic instructions in the memory 630 may be implemented in the form of software functional units and stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a usb disk, a removable hard disk, a Read-only memory (ROM), a random access memory (RAM, randomAccessMemory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

Embodiments of the present invention also provide a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, is implemented to perform the methods provided by the above embodiments, for example, comprising:

acquiring a user and article node position code, converting the user and article node position code dimension into a node embedding dimension and injecting the node embedding dimension into node initial embedding;

Feeding the node with the embedded position codes to each transducer layer of Graph Transformer for calculation to respectively obtain a local interaction signal and a global collaboration signal of the target node;

the local interaction signals and the global collaboration signals are fused and output to obtain Graph Transformer node representations of each layer;

and inputting the L-layer representation of the node into a reading function to obtain a final representation of the node, calculating the vector inner product of the final representation of the node of the user and the article, and sequencing the articles to obtain the target recommendation of the user.

From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on such understanding, the foregoing technical solutions may be embodied essentially or in part in the form of a software product, which may be stored in a computer-readable storage medium, such as a ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the various embodiments or methods of some parts of the embodiments.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. A location-aware collaborative recommendation method, the recommendation method comprising:

acquiring a user and article node position code, converting the user and article node position code dimension into a node embedding dimension and injecting the node embedding dimension into node initial embedding;

Feeding the node with the embedded position codes to each transducer layer of Graph Transformer for calculation to respectively obtain a local interaction signal and a global collaboration signal of the target node;

The Graph Transformer includes L convertor layers, each having a graph convolution module and a NodeFormer convolution module, where the graph convolution module is configured to capture a local interaction signal of a target node, and the NodeFormer convolution module is configured to capture a global collaboration signal of the target node;

the local interaction signals and the global collaboration signals are fused and output to obtain Graph Transformer node representations of each layer;

the step of obtaining Graph Transformer node representations of each layer after the local interaction signals and the global collaboration signals are fused and output comprises the following steps:

fusing the local interaction signals and the global collaboration signals according to a certain fusion proportion and outputting the local interaction signals and the global collaboration signals to obtain target node representation of the current transducer layer;

Calculating by the graph convolution module and the Nodeformer convolution modules of the L Transformer layers to obtain node representations of each layer;

and inputting the L-layer representation of the node into a reading function to obtain a final representation of the node, calculating the vector inner product of the final representation of the node of the user and the article, and sequencing the articles to obtain the target recommendation of the user.

2. The location-aware collaborative recommendation method according to claim 1, wherein the user and item node location codes are one of spectrum codes, degree codes, pageRank codes, and type codes.

3. The location-aware collaborative recommendation method according to claim 1, wherein feeding the node with embedded location codes to respective Transformer layer computations of Graph Transformer to obtain local interaction signals and global collaboration signals of a target node respectively comprises:

feeding the node to Graph Transformer;

the local interaction signal is captured by the graph convolution module, and the global collaboration signal is captured by the NodeFormer convolution module.

4. The location-aware collaborative recommendation method according to claim 1, wherein inputting the L-level representation of the node into a read function to obtain a final representation of the node, computing and ordering items via a vector inner product of the final representation of the user and the item node, obtaining a target recommendation for the user comprises:

finally representing the vector inner product by the user and the object node to obtain a plurality of values with highest inner product results;

and obtaining target recommendation of the user.

5. The location aware collaborative recommendation method according to claim 4, wherein the readout function is replaced with an arbitrary graph recommendation model.

6. A location-aware collaborative recommendation device, comprising:

the preprocessing injection module is used for acquiring the node position codes of the user and the article, converting the node position code dimensions of the user and the article into node embedding dimensions and injecting the node embedding dimensions into the node initial embedding;

the node calculation module is used for feeding the node with the embedded position codes to each transducer layer of Graph Transformer for calculation to respectively obtain a local interaction signal and a global collaboration signal of the target node; the Graph Transformer includes L convertor layers, each having a graph convolution module and a NodeFormer convolution module, where the graph convolution module is configured to capture a local interaction signal of a target node, and the NodeFormer convolution module is configured to capture a global collaboration signal of the target node;

The node updating module is used for obtaining Graph Transformer node representations of each layer after the local interaction signals and the global collaboration signals are fused and output;

the step of obtaining Graph Transformer node representations of each layer after the local interaction signals and the global collaboration signals are fused and output comprises the following steps:

fusing the local interaction signals and the global collaboration signals according to a certain fusion proportion and outputting the local interaction signals and the global collaboration signals to obtain target node representation of the current transducer layer;

Calculating by the graph convolution module and the Nodeformer convolution modules of the L Transformer layers to obtain node representations of each layer;

And the recommendation output module is used for inputting the L-layer representation of the node into a reading function to obtain a final representation of the node, calculating the vector inner product of the final representation of the node of the user and the article, and sequencing the articles to obtain the target recommendation of the user.

7. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the location-aware collaborative recommendation method of any one of claims 1 to 5 when the program is executed by the processor.

8. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the steps of the location aware collaborative recommendation method according to any one of claims 1 to 5.