CN108897823A - Personalized commercial search method and device based on deep learning attention mechanism - Google Patents

Abstract

The invention discloses a personalized product retrieval method and device based on a deep learning attention mechanism. Among them, the method includes: constructing a short-term preference model based on the attention mechanism; constructing a long-term preference model based on the attention mechanism; The long-term preference model of the mechanism and the current query, learn the interactive relationship between the three through a multi-layer fully connected network, obtain a reorganized query representation, and use a distance function to define the degree of relevance between all commodities and the current query; training the After the end of the multi-layer fully connected network, for the new query submitted by each user, the distance value between all the products and the current query is obtained, and then all the distance values are sorted from high to low, and the products corresponding to the first n distance values are returned to the user, where n is a positive integer.

Description

Personalized commodity retrieval method and device based on deep learning attention mechanism

technical field

The invention belongs to the field of data retrieval, and in particular relates to a personalized product retrieval method and device based on a deep learning attention mechanism.

Background technique

With the prevalence of the Internet, e-commerce has become more and more popular with the public. When users in an e-commerce website (such as Tmall) want to buy a commodity, they usually need to find the one they like from millions of commodities in a retrieval manner. For this kind of online product retrieval, a relatively common situation is that the user first submits a query, and then the search engine returns a sorted list of products related to the current query. However, the queries submitted by users are generally only composed of several keywords (for example, men's long-sleeved tops), which also leads to the inability to accurately convey the needs of users, thereby causing users to be dissatisfied with the search results.

In addition, users' shopping preferences can be very broad (due to different backgrounds, such as age, gender, income), or influenced by the current environment (such as seasons, location). Therefore, for the same query from different users, returning the same retrieval result will cause different degrees of economic loss to the e-commerce website. In view of this, it is particularly important to consider the user's shopping intention in different situations to return relevant products to the query submitted by the user, that is, personalized product retrieval, to meet the current shopping needs of the user.

Traditional item retrieval methods are limited to simple matching between queries and items without taking the user's own attributes into consideration. Because these methods ignore the heterogeneity of individual needs of many users, they often lead to limited retrieval performance of search engines and cannot provide users with satisfactory retrieval results.

Ai et al. recently proposed a personalized item retrieval approach, where they model users' long-term shopping preferences via a latent space that jointly learns user, item, and query representations. However, this method also has two defects: (1) it is assumed that the long-term shopping preference of the user is stable, but in fact it will change slowly over time; (2) the short-term shopping preference of the user is not taken into account, and the user The short-term purchase behavior of is likely to reflect the user's recent purchase habits.

Among them, the long-term refers to the user's internal and relatively stable shopping preferences, such as favorite color, suitable size and consumption ability. At the same time, it will be affected by the user's respective background, such as age, marriage, education, income, etc. In contrast, short-term shopping preferences reflect users' shopping intentions in a relatively short period of time, and are affected by unexpected events, such as new product launches, seasonal changes, and special personal festivals (such as birthdays), etc. These can be inferred from the attributes of items recently purchased by the user. Compared with long-term shopping preferences, short-term shopping preferences are updated more frequently and less predictably.

At present, the following problems exist in the personalized commodity retrieval method:

One is that accurately modeling users' long-term and short-term shopping preferences is extremely complex. A user's long-term shopping preference includes multiple aspects, such as spending power or favorite colors and brands, and will change with the user's background (such as income). Correspondingly, users' short-term shopping preferences are usually dynamically changing and are easily affected by emergencies;

The second is that users describe their shopping needs through a text query consisting of only a few keywords, which makes it difficult to accurately locate the aspects of the user's long-term shopping preferences that are related to the current query. For example, the design of a short-sleeved top, rather than the price, will affect users who can afford it more. For the user's short-term preference, multiple recently purchased items will also have different impacts on the user's next purchase behavior;

Third, it is also difficult to combine the user's long-term and short-term shopping preferences with the current query.

Contents of the invention

In order to solve the deficiencies of the prior art, the first object of the present invention is to provide a personalized product retrieval method based on deep learning attention mechanism, which improves the accuracy of retrieval, thereby improving the user's retrieval experience.

A personalized product retrieval method based on deep learning attention mechanism of the present invention, comprising:

Step 1: Construct a short-term preference model based on the attention mechanism; the step 1 specifically includes:

Given the current query, use the neural network to estimate the degree of correlation between the first m queries and the current query, and use different attention weights to represent; where m is a positive integer;

Multiply the obtained attention weight by the corresponding product representation, input it into the RNN model, and output m vectors representing the short-term preference based on the attention mechanism;

Step 2: Construct a long-term preference model based on the attention mechanism; the step 2 specifically includes:

Initialize its long-term preference by using the initially purchased m product representations, and use the same principle as building a short-term preference model based on the attention mechanism to obtain a long-term preference model based on the attention mechanism;

After that, every time m products are passed, the current user's long-term preference is updated, and the updated long-term preference model based on the attention mechanism is obtained;

Step 3: query re-representation; the process is: integrate the short-term preference model based on the attention mechanism, the long-term preference model based on the attention mechanism and the current query, and learn the interactive relationship between the three through a multi-layer fully connected network, and get A restructured query representation, using a distance function to define how relevant all items are to the current query;

Step 4: After training the multi-layer fully connected network, for each new query submitted by the user, get the distance values between all commodities and the current query, and then sort all the distance values from high to low, and sort the top n The product corresponding to the distance value is returned to the user, where n is a positive integer.

Further, the method also includes:

A "pair-based" learning method is employed to train the multi-layer fully connected network.

Further, in the step 1, a two-layer neural network is used to estimate the degree of correlation between the first m queries and the current query.

Further, in the step 3, the distance function is cosine distance, or dot product, or Euclidean distance, or Manhattan distance.

The second object of the present invention is to provide a personalized product retrieval device based on deep learning attention mechanism.

A personalized commodity retrieval device based on the deep learning attention mechanism of the present invention includes a personalized commodity retrieval processor, and the personalized commodity retrieval processor includes:

A short-term preference model building block configured to: build a short-term preference model based on an attention mechanism; the short-term preference model building block specifically includes:

The attention weight acquisition sub-module is configured to: given the current query, use a neural network to estimate the degree of correlation between the first m queries and the current query, and use different attention weights to represent; where m is a positive integer; and

The short-term preference representation sub-module is configured to: multiply the obtained attention weight by the corresponding commodity representation, input it into the RNN model, and output m vectors representing the short-term preference based on the attention mechanism;

A long-term preference model building block configured to: build a long-term preference model based on an attention mechanism; the long-term preference model building block specifically includes:

The long-term preference initialization and construction sub-module is configured to: initialize the long-term preference by using the initially purchased m commodity representations, and use the same principle as building the short-term preference model based on the attention mechanism to obtain the long-term preference model based on the attention mechanism preference model;

The long-term preference update sub-module is configured to update the current user's long-term preference every time m products are passed, and obtain the updated long-term preference model based on the attention mechanism;

The query re-representation module is configured to: integrate the short-term preference model based on the attention mechanism, the long-term preference model based on the attention mechanism and the current query, and learn the interactive relationship between the three through a multi-layer fully connected network to obtain the reorganized query representation, and uses a distance function to define how relevant all items are to the current query;

Commodity return module, which is configured to: after the training of the multi-layer fully connected network is completed, obtain the distance values between all commodities and the current query for the new query submitted by each user, and then sort all the distance values from high to low, And return the products corresponding to the first n distance values to the user, where n is a positive integer.

Further, the personalized commodity retrieval processor also includes:

A model training module configured to: use a "pair-based" learning method to train the multi-layer fully connected network.

Further, in the attention weight acquisition sub-module, a two-layer neural network is used to estimate the degree of correlation between the first m queries and the current query.

Further, in the query representation module, the distance function is cosine distance, or dot product, or Euclidean distance, or Manhattan distance.

Compared with prior art, the beneficial effect of the present invention is:

(1) In order to alleviate many problems faced in the existing methods, the present invention can combine the user's relatively stable long-term shopping preferences and time-sensitive short-term shopping preferences to carry out personalized retrieval of commodities, and finally improve the accuracy of retrieval, Thereby improving the user's retrieval experience.

(2) The present invention can effectively combine the user's long-term and short-term shopping preferences and the query submitted by the current user, and re-express the user's shopping needs.

(3) The present invention can highlight factors relevant to the current query in the two preferences through two separate attention mechanisms in modeling the user's long-term and short-term shopping preferences.

(4) The present invention improves the accuracy of personalized commodity retrieval, thereby retaining more users and increasing income for e-commerce websites to a certain extent.

Description of drawings

The accompanying drawings constituting a part of the present application are used to provide further understanding of the present application, and the schematic embodiments and descriptions of the present application are used to explain the present application, and do not constitute improper limitations to the present application.

Fig. 1 is a flowchart of a personalized product retrieval method based on deep learning attention mechanism of the present invention.

Fig. 2 is a schematic structural diagram of the personalized commodity retrieval processor of the present invention.

Detailed ways

It should be pointed out that the following detailed description is exemplary and intended to provide further explanation to the present application. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

It should be noted that the terminology used here is only for describing specific implementations, and is not intended to limit the exemplary implementations according to the present application. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural, and it should also be understood that when the terms "comprising" and/or "comprising" are used in this specification, they mean There are features, steps, operations, means, components and/or combinations thereof.

A personalized product retrieval method based on deep learning attention mechanism of the present invention will be described in detail below in combination with implementation examples and accompanying drawings.

As shown in Figure 1, a kind of personalized product retrieval method based on deep learning attention mechanism of the present invention comprises the following steps:

Step 1: Build an attention-based short-term preference model.

Specifically, the process of building a short-term preference model based on the attention mechanism includes:

Step 1.1: Given the current query, use the neural network to estimate the degree of correlation between the first m queries and the current query, and use different attention weights to represent; where m is a positive integer;

Since the previous queries (or purchased products) are not all closely related to the current query (or the target product), for example, compared with the previously submitted query "mobile phone", "mouse" and "keyboard" are obviously related to the current query " display" closer. Due to the difference in the degree of correlation between the previous query and the current query, the present invention uses a two-layer neural network to estimate the degree of correlation between them, which is represented by different attention weights here.

Step 1.2: Multiply the obtained attention weight by the corresponding product representation, input it into the RNN model, and output m vectors representing the short-term preference based on the attention mechanism.

Corresponding to the query, the purchased products corresponding to the previous queries also have different impacts from the user's current purchase intention. Therefore, the correlation between the previous query and the current query is used to determine the correlation between the purchased product of the previous corresponding query and the current purchase intention. The specific representation is to multiply the attention weight obtained before by the corresponding product representation to obtain the attention-based mechanism short-term preferences.

Before constructing the short-term preference model based on the attention mechanism, it also includes:

All the comment information of the current product and a text query learn the representation of the product and the query through a unified PV-DM model, and then project this representation into the same subspace.

Step 2: Build a long-term preference model based on attention mechanism.

Users' long-term shopping preferences are more stable than short-term, but they will change slowly. The long-term preference is represented by all the purchased items of the current user, and it is updated slowly.

Specifically, the process of building a long-term preference model based on the attention mechanism includes:

Step 2.1: Initialize the long-term preference by using the initially purchased m commodity representations, and use the same principle as building the short-term preference model based on the attention mechanism to obtain the long-term preference model based on the attention mechanism;

Step 2.1: After that, update the long-term preference of the current user every time m products are passed, and obtain the updated long-term preference model based on the attention mechanism.

Step 3: Query Re-Representation.

Specifically, the process of query re-expression is:

Integrate the short-term preference model based on the attention mechanism, the long-term preference model based on the attention mechanism and the current query, learn the interactive relationship between the three through a multi-layer fully connected network, obtain a reorganized query representation, and use a distance function to Define how relevant all items are to the current query.

Wherein, the distance function is cosine distance, or dot product, or Euclidean distance, or Manhattan distance.

Step 4: After training the multi-layer fully connected network, for each new query submitted by the user, get the distance values between all commodities and the current query, and then sort all the distance values from high to low, and sort the top n The product corresponding to the distance value is returned to the user, where n is a positive integer.

For example: n is 10 or 20.

In another embodiment, the method also includes:

A "pair-based" learning method is employed to train the multi-layer fully connected network.

This can improve the robustness of the learned multi-layer fully connected network, where the BPR (Bayesian Personalized Ranking) loss function is used to obtain the final loss, and SGD (Stochastic Gradient Descent) with momentum is used optimization method.

In order to alleviate many problems faced in the existing methods, the present invention can combine the user's relatively stable long-term shopping preferences and time-sensitive short-term shopping preferences to carry out personalized retrieval of commodities, and finally improve the accuracy of retrieval, thereby improving the User search experience.

The invention can effectively combine the user's long-term and short-term shopping preference and the query submitted by the current user, and re-express the user's shopping demand.

Through two separate attention mechanisms in the long-term and short-term modeling of the user's shopping preference, the present invention can highlight factors relevant to the current query in both preferences.

The invention improves the accuracy of personalized commodity retrieval, thereby retaining more users and increasing income for e-commerce websites to a certain extent.

The present invention also provides a personalized commodity retrieval device based on deep learning attention mechanism.

A personalized commodity retrieval device based on deep learning attention mechanism of the present invention includes a personalized commodity retrieval processor, as shown in Figure 2, said personalized commodity retrieval processor includes:

(1) A short-term preference model building module, which is configured to: build a short-term preference model based on an attention mechanism.

The building block of the short-term preference model specifically includes:

(1.1) Attention weight acquisition sub-module, which is configured to: given the current query, use a neural network to estimate the degree of correlation between the first m queries and the current query, and use different attention weights to represent; where , m is a positive integer.

Since the previous queries (or purchased products) are not all closely related to the current query (or the target product), for example, compared with the previously submitted query "mobile phone", "mouse" and "keyboard" are obviously related to the current query " display" closer. Due to the difference in the degree of correlation between the previous query and the current query, the present invention uses a two-layer neural network to estimate the degree of correlation between them, which is represented by different attention weights here.

(1.2) The short-term preference representation sub-module is configured to: multiply the obtained attention weight by the corresponding commodity representation, input it into the RNN model, and output m vectors representing the short-term preference based on the attention mechanism.

Corresponding to the query, the purchased products corresponding to the previous queries also have different impacts from the user's current purchase intention. Therefore, the correlation between the previous query and the current query is used to determine the correlation between the purchased product of the previous corresponding query and the current purchase intention. The specific representation is to multiply the attention weight obtained before by the corresponding product representation to obtain the attention-based mechanism short-term preferences.

Before constructing the short-term preference model based on the attention mechanism, all the comment information of the current product and a text query learn the representation of the product and the query through a unified PV-DM model, and then project this representation to the same child in space.

(2) A long-term preference model building module, which is configured to: build a long-term preference model based on an attention mechanism.

Users' long-term shopping preferences are more stable than short-term, but they will change slowly. The long-term preference is represented by all the purchased items of the current user, and it is updated slowly.

The long-term preference model construction module specifically includes:

(2.1) Long-term preference initialization and construction sub-module, which is configured as follows: Initialize the long-term preference by using the initially purchased m commodity representations, and use the same principle as building the short-term preference model based on the attention mechanism to obtain the attention-based The long-term preference model of the mechanism;

(2.2) The long-term preference update sub-module, which is configured to update the long-term preference of the current user every time after passing through m items, and obtain the updated long-term preference model based on the attention mechanism;

(3) The query re-representation module, which is configured to: integrate the short-term preference model based on the attention mechanism, the long-term preference model based on the attention mechanism and the current query, and learn the interactive relationship between the three through a multi-layer fully connected network, Get the restructured query representation, and use a distance function to define the relevance of all items to the current query;

In the query re-expression module, the distance function is cosine distance, or dot product, or Euclidean distance, or Manhattan distance.

(4) Commodity return module, which is configured as follows: After the training of the multi-layer fully connected network is completed, for the new query submitted by each user, the distance values between all commodities and the current query are obtained, and then all distance values are ranked from high to low Sort and return the items corresponding to the first n distance values to the user, where n is a positive integer.

For example: n is 10 or 20.

In another embodiment, the personalized commodity retrieval processor further includes:

A model training module configured to: use a "pair-based" learning method to train the multi-layer fully connected network.

This can improve the robustness of the learned multi-layer fully connected network, where the BPR (Bayesian Personalized Ranking) loss function is used to obtain the final loss, and SGD (Stochastic Gradient Descent) with momentum is used optimization method.

In order to alleviate many problems faced in the existing methods, the present invention can combine the user's relatively stable long-term shopping preferences and time-sensitive short-term shopping preferences to carry out personalized retrieval of commodities, and finally improve the accuracy of retrieval, thereby improving the User search experience.

The invention can effectively combine the user's long-term and short-term shopping preference and the query submitted by the current user, and re-express the user's shopping demand.

Through two separate attention mechanisms in the long-term and short-term modeling of the user's shopping preference, the present invention can highlight factors relevant to the current query in both preferences.

The invention improves the accuracy of personalized commodity retrieval, thereby retaining more users and increasing income for e-commerce websites to a certain extent.

Those skilled in the art should understand that the embodiments of the present invention may be provided as methods, systems, or computer program products. Accordingly, the present invention can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each procedure and/or block in the flowchart and/or block diagram, and a combination of procedures and/or blocks in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions may be provided to a general purpose computer, special purpose computer, embedded processor, or processor of other programmable data processing equipment to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing equipment produce a An apparatus for realizing the functions specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions The device realizes the function specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby The instructions provide steps for implementing the functions specified in the flow chart or blocks of the flowchart and/or the block or blocks of the block diagrams.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented through computer programs to instruct related hardware, and the programs can be stored in a computer-readable storage medium. During execution, it may include the processes of the embodiments of the above-mentioned methods. Wherein, the storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM) or a random access memory (Random Access Memory, RAM) and the like.

Although the specific implementation of the present invention has been described above in conjunction with the accompanying drawings, it does not limit the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, those skilled in the art do not need to pay creative work Various modifications or variations that can be made are still within the protection scope of the present invention.

Claims (8)

1. A personalized commodity retrieval method based on deep learning attention mechanism, characterized in that, comprising: Step 1: Construct a short-term preference model based on the attention mechanism; the step 1 specifically includes: Given the current query, use the neural network to estimate the degree of correlation between the first m queries and the current query, and use different attention weights to represent; where m is a positive integer; Multiply the obtained attention weight by the corresponding product representation, input it into the RNN model, and output m vectors representing the short-term preference based on the attention mechanism; Step 2: Construct a long-term preference model based on the attention mechanism; the step 2 specifically includes: Initialize its long-term preference by using the initially purchased m product representations, and use the same principle as building a short-term preference model based on the attention mechanism to obtain a long-term preference model based on the attention mechanism; After that, every time m products are passed, the current user's long-term preference is updated, and the updated long-term preference model based on the attention mechanism is obtained; Step 3: query re-representation; the process is: integrate the short-term preference model based on the attention mechanism, the long-term preference model based on the attention mechanism and the current query, and learn the interactive relationship between the three through a multi-layer fully connected network, and get A restructured query representation, using a distance function to define how relevant all items are to the current query; Step 4: After training the multi-layer fully connected network, for each new query submitted by the user, get the distance values between all commodities and the current query, and then sort all the distance values from high to low, and sort the top n The product corresponding to the distance value is returned to the user, where n is a positive integer. 2. A kind of personalized commodity retrieval method based on deep learning attention mechanism as claimed in claim 1, it is characterized in that, this method also comprises: A "pair-based" learning method is employed to train the multi-layer fully connected network. 3. a kind of personalized commodity retrieval method based on deep learning attention mechanism as claimed in claim 1, it is characterized in that, in described step 1, adopt a two-layer neural network to estimate the preceding m queries and current The degree of correlation between queries. 4. A kind of personalized commodity retrieval method based on deep learning attention mechanism as claimed in claim 1, is characterized in that, in described step 3, described distance function is cosine distance, or dot product, or Euclidean distance , or the Manhattan distance. 5. A personalized commodity retrieval device based on a deep learning attention mechanism, characterized in that it includes a personalized commodity retrieval processor, and the personalized commodity retrieval processor includes: A short-term preference model building block configured to: build a short-term preference model based on an attention mechanism; the short-term preference model building block specifically includes: The attention weight acquisition sub-module is configured to: given the current query, use a neural network to estimate the degree of correlation between the first m queries and the current query, and use different attention weights to represent; where m is a positive integer; and The short-term preference representation sub-module is configured to: multiply the obtained attention weight by the corresponding commodity representation, input it into the RNN model, and output m vectors representing the short-term preference based on the attention mechanism; A long-term preference model building block configured to: build a long-term preference model based on an attention mechanism; the long-term preference model building block specifically includes: The long-term preference initialization and construction sub-module is configured to: initialize the long-term preference by using the initially purchased m commodity representations, and use the same principle as building the short-term preference model based on the attention mechanism to obtain the long-term preference model based on the attention mechanism preference model; The long-term preference update sub-module is configured to update the current user's long-term preference every time m products are passed, and obtain the updated long-term preference model based on the attention mechanism; The query re-representation module is configured to: integrate the short-term preference model based on the attention mechanism, the long-term preference model based on the attention mechanism and the current query, and learn the interactive relationship between the three through a multi-layer fully connected network to obtain the reorganized query representation, and uses a distance function to define how relevant all items are to the current query; Commodity return module, which is configured to: after the training of the multi-layer fully connected network is completed, obtain the distance values between all commodities and the current query for the new query submitted by each user, and then sort all the distance values from high to low, And return the products corresponding to the first n distance values to the user, where n is a positive integer. 6. A kind of personalized commodity retrieval device based on deep learning attention mechanism as claimed in claim 5, it is characterized in that, described personalized commodity retrieval processor also comprises: A model training module configured to: use a "pair-based" learning method to train the multi-layer fully connected network. 7. A kind of personalized commodity retrieval device based on deep learning attention mechanism as claimed in claim 5, is characterized in that, in described attention weight acquisition submodule, adopts a two-layer neural network to estimate the former m The degree of relevance between a query and the current query. 8. A kind of personalized commodity retrieval device based on deep learning attention mechanism as claimed in claim 5, it is characterized in that, in said query re-expression module, said distance function is cosine distance, or dot product, or Euclidean distance, or Manhattan distance.