Recommendation method based on dynamic difference map

Abstract

A recommendation method based on dynamic differential graph, based on the interaction record of user and item, the user-item interaction graph and the user-item category interaction graph at each moment are constructed, and the user new interest dynamic differential graph and the user loss interest dynamic differential graph are further constructed; based on the user-item interaction graph and the user-item category interaction graph, the long-term / short-term interest feature expression of the user is extracted through the long-term / short-term interest module; based on the user new interest dynamic differential graph and the loss interest dynamic differential graph, the user interest change trend feature expression is extracted through the interest change trend module; based on the user-item interaction graph and the user-item category interaction graph, the feature expression containing the item category information is extracted through the item module; based on the user long-term / short-term interest feature expression, the user interest change trend feature expression and the feature expression containing the item category information, the user and the item are scored through the recommendation module, and the item with high score is selected as the item recommended to the user.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of Internet information artificial intelligence, and particularly relates to a recommendation method based on a dynamic difference graph. BACKGROUND

[0002] A graph-based recommendation model constructs a graph data structure according to user and item interaction information, models the user and the item by using the graph data, and performs recommendation. A common graph structure is a bipartite graph of users and items, that is, user nodes and item nodes that have interacted with each other are connected by edges. Through the graph structure, multi-order information between the user and the item can be obtained, so that the similarity between users or between items can be captured. At present, the graph-based recommendation model has two defects. First, most of the graph-based recommendation algorithms do not model the time dimension information, but only consider whether the user and the item have interacted with each other. Second, most of the graph-based recommendation models ignore the behavior characteristics of the user, and in actual application scenarios, the behavior characteristics of the user contain a lot of important information, such as the dynamic interest change trend of the user. By modeling the behavior characteristics of the user, the performance of the recommendation system can be further optimized. SUMMARY

[0003] The application provides a recommendation method based on a dynamic difference graph, which is proposed in view of the above defects in the prior art. The recommendation method constructs an interest dynamic difference graph of a user according to interaction information of the user and an item category, and designs a recommendation model based on the dynamic difference graph. The recommendation model comprises a long-term / short-term interest modeling module, an interest change trend modeling module, an item modeling module and a recommendation module. The long-term / short-term interest modeling module models long-term interest and short-term interest feature expressions of the user according to historical interaction information of the user and the item. The interest change trend modeling module models interest change trend feature expressions of the user by using a dynamic difference graph neural network layer and according to the interest dynamic difference graph. The item modeling module models feature expressions of the item according to the item and category information of the item. The recommendation module integrates long-term and short-term interests and interest change trend features of the user to obtain complete user feature expressions, and finally performs item recommendation in combination with the item feature expressions.

[0004] The application is implemented by the following technical solutions.

[0005] The application relates to a recommendation method based on a dynamic differential graph, which constructs a user-item interaction graph and a user-item category interaction graph at each moment based on user and item interaction records, and further constructs a user new interest dynamic differential graph and a user lost interest dynamic differential graph; based on the user-item interaction graph and the user-item category interaction graph, long-term / short-term interest feature expressions of a user are extracted through a long-term / short-term interest module; based on the user new interest dynamic differential graph and the user lost interest dynamic differential graph, user interest change trend feature expressions are extracted through an interest change trend module; based on the user-item interaction graph and the user-item category interaction graph, feature expressions containing item category information are extracted through an item module; based on the long-term / short-term interest feature expressions of the user, the user interest change trend feature expressions and the feature expressions containing item category information, a recommendation module is used to score the user and the item, and an item with a high score is selected as an item recommended to the user.

[0006] Technical effects

[0007] The application considers the interest change of a user at different timestamps by defining and constructing an interest dynamic differential graph and a recommendation model designed based on the dynamic differential graph, constructs the interest dynamic differential graph, solves the problem that a graph-based method lacks the use of time information, and introduces a new dimension of user features, i.e., a user interest change trend feature. The corresponding dynamic differential graph neural network layer is applied to the constructed interest dynamic differential graph, models the user interest change trend feature, and finally combines the long-term / short-term interest features of the user to perform item recommendation. The application achieves better recommendation effect than existing methods on public data sets such as a movie data set MovieLens and a retail data set Retailrockets. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 (a) is a user new interest analysis diagram in the embodiment, and (b) is a lost interest analysis diagram;

[0009] Figure 2 It is a basic graph structure definition diagram in the embodiment;

[0010] Figure 3 It is a recommendation model based on a dynamic differential graph diagram;

[0011] Figure 4 It is a dynamic differential graph neural network layer diagram in the embodiment. DETAILED DESCRIPTION

[0012] The embodiment relates to a recommendation method based on a dynamic differential graph, and the method comprises the following steps:

[0013] Step 1, analysis example of user interest change: user interest change is divided into new interest, that is, new interest in a certain type of goods, and loss of interest, that is, loss of interest in a certain type of goods.

[0014] As Figure 1 shown, the change of new interest and loss of interest of users on different types of movies on the public movie data set MovieLens is disclosed, wherein: the horizontal axis is the type of movie, and the vertical axis is the proportion of users with interest change. This embodiment is divided into 1 month as the time division granularity. For example, a user interacts with a type of movie on a certain day, if the user has not interacted with the movie of this type at least one month before this day, then it is considered that the user has new interest in the movie of this type on this day. On the contrary, if the user has not interacted with the movie of this type at least one month after this day, then it is considered that the user loses interest in the movie of this type after one month.

[0015] As Figure 2 shown, the definition of the basic interaction sequence and the interaction graph involved in this embodiment is shown, wherein: the user set is U={u1,u2,....,u n}, the item set is I={i1,i2,....,i m}, and the category set to which the item i belongs is C i ={c3,c7,....c j}. The interaction data of user u and item i is in the format of <u, i, t>, wherein: t is the time of interaction. After the user and item interaction data is combined together, a specific timestamp t is specified, which can form a user and item interaction sequence S u，t , after the user-item interaction sequence is determined, the user-item category interaction sequence can also be determined. The user-item interaction graph at t represents the interaction information between users and items, wherein: V U and V I represent the node set of users and items respectively, is the edge data of user and item interaction generated in the time period between t-1 and t, that is, when the user u and the item f interact between t-1 and t, there is an edge e between u and i in u，i =1. Similarly, the user-category record graph at t represents the interaction information between users and item categories.

[0016] Step 2, constructing an interest dynamic difference graph, specifically including:

[0017] 2.1) Constructing the dynamic differential graph of new interest: For a user u, the interaction with a category c at time t should at least meet the following two requirements to be considered as the new interest of the user: first, the user u interacts with the category c at time t; second, the user u has not interacted with the category c in the time interval τ before time t. The final constructed dynamic differential graph of new interest is wherein: is the user-category interaction graph at time t, τ is the length of the time interval hyperparameter, Θ is a zero matrix with the same size as max(A, B) represents the larger value between A and B at each position of the matrix.

[0018] 2.2) Constructing the dynamic differential graph of lost interest: Similar to the new interest, for a user u, the interaction with a category c at time t should at least meet the following two requirements to be considered as the lost interest of the user: first, the user has not interacted with the category c in the time interval before time t; second, the user has interacted with the category c at an earlier time. The final constructed dynamic differential graph of lost interest is

[0019] Step 3, as shown in Figure 3 , a recommendation system model is constructed, which includes: a long-term / short-term interest modeling module for extracting long-term / short-term interest feature expression of users, an item modeling module for extracting feature expression of items, an interest change trend modeling module for extracting interest change trend feature expression of users, and a recommendation module for recommending suitable items that the user may interact with at the next time. In the user-item interaction graph and the user-category interaction graph, all nodes are initially represented by one-hot encoding vectors, and the index position corresponding to the value of 1 in the vector is the number of the user / item / category. Since the one-hot vector is generally a sparse vector, it contains less information and requires a large storage space, so all modeling modules of the recommendation system model use an embedding expression layer to process the one-hot encoding vector, and then the one-hot encoding vector is converted into a low-dimensional vector. The processing method of different types of nodes in the embedding expression layer is similar. Taking user u as an example, u ∈R m represents the one-hot encoding column vector of user u, then the embedding vector e u ∈R d is calculated by , E u ∈R m×d is the parameter matrix randomly initialized by the embedding expression layer. Similarly, the embedding vectors e i and e c of item i and category c can also be obtained.

[0020] A user's short-term interests can primarily be extracted from their last interaction with an item. Based on this, for a given user, the short-term interest representation layer selects the user-item interaction graph corresponding to the moment of the user's last interaction with an item, and then performs a graph convolution operation on the selected graph. The short-term interest representation layer in the long-term / short-term interest modeling module is based on the formula... Calculate the short-term interest feature representation S of all users at time t. t ,in: It is a user-item interaction diagram The adjacency matrix representation of W S It is a randomly generated parameter matrix, E u，i This represents the matrix resulting from concatenating the user's embedding representation with the item's embedding representation, where σ is the sigmoid non-linear activation function. After the graph convolution operation, the user u... j Features at time t That is, S t In the j-th row of the graph; unlike short-term interests, a user's long-term interests can be obtained from the user's interactions with items over different time periods. Therefore, in the long-term interest representation layer, all user u's interaction information up to the current moment is used to construct the user's long-term interest representation layer. Specifically, the long-term interest representation layer first applies the same graph convolution operation as the short-term interest representation layer to... From the dynamic graph at each moment, we obtain S. 1 S 2 ,...S τ After a series of short-term user interest features are expressed, for a given user, the user interest features expressed at different times should be given different weights. Therefore, the long-term interest expression layer uses an attention network to determine the importance of the user's expression at each time point, as shown in the formula. As shown. After obtaining the weight of each timestamp user's expression, then according to the formula... Calculate users' long-term interest expressions Where WL is a randomly initialized and trainable parameter matrix.

[0021] like Figure 4 As shown, the interest change trend modeling module processes two kinds of interest dynamic difference maps through a dynamic difference graph neural network layer with two built-in dynamic graph attention convolutional networks to generate two kinds of interest feature expressions of the user. The specific structure of the dynamic difference graph neural network layer receives the two types of dynamic graphs and enters their respective dynamic graph attention convolutional networks, i.e., heterogeneous graph attention networks, to obtain the user's new interest expressions and lost interest expressions at each time stamp, and then performs the subsequent integration operation.

[0022] The heterogeneous graph attention network extracts a category feature expression, i.e., calculates first-order correlation of the heterogeneous graph, and extracts a user interest change trend feature expression, i.e., calculates second-order correlation of the heterogeneous graph.

[0023] In processing the added and lost interest dynamic graph, the neural network structure adopted is the same, but the parameters in the network are initialized respectively. The following takes the added user interest dynamic graph as an example to explain the specific structure of the graph attention neural network: first, the first-order correlation in the graph is calculated by formula The first-order correlation in the graph is calculated, where: represents the neighbor user node set of category c at time t, so contains information and features of all users who have interacted with category c at time t. For a user, the proportion of different categories is also different, so the weight of the item category interacted by the user is calculated through the attention mechanism, specifically: Then the second-order correlation in the graph is calculated, specifically: Where: represents the neighbor item category node set of user u at time t, so is the user added interest feature expression integrating neighbor user and neighbor category information, and all the above are parameter matrices that are randomly initialized and can be trained, and are shared for different times t.

[0024] Similarly, the user lost interest feature expression After obtaining the user added interest feature expression and the lost interest feature expression, according to different users, the two kinds of feature expressions are combined by giving different weights through a layer of attention mechanism to obtain the user interest change trend feature That is Where: weight Finally, the feature expressions of different times take different weights, i.e., the attention mechanism is applied in the time dimension to obtain the final user interest change trend feature vector, i.e. Where:

[0025] The item feature representation layer needs to integrate the category information of the item into the feature expression of the item, specifically: as shown, where: ⊙ represents vector splicing, is a randomly initialized parameter matrix, and C i represents the category set to which the item i belongs.

[0026] The integration layer in the recommendation module combines the final user expression, i.e., the short-term interest expression of the user Long-term interest expression of a user and interest trend expression of the user and the final expression of the item, i.e., the embedding vector e of the item i and the category feature expression λ of the item i combined, specifically comprising: wherein: the embedding expression e of the item itself i dimension, resulting in the score of the user u on the item i, Γ(X) = W2tanh(W1X + b1) + b2 is a classic multi-layer perception; then, the category information integrated item feature λ i dimension, obtaining the category interest degree score r of the item u on the item i c (u, i). Then, the two scores are combined using a self-attention mechanism, i.e., to calculate the final score information at time t+1.

[0027] Step 4, training the recommendation system model constructed in step 3: considering that different categories of items are also different, the method uses a negative sampling method to ensure that for each category, wherein: a item of the category that the user has not interacted with is selected as a negative sample; after negative sampling, the ratio of positive and negative samples for each user is close to 1:10, and the objective function is as follows: wherein: T and S are the negative sample set and the positive sample set, respectively, w + is the positive and negative sample balance weight function, and in this embodiment, 0.91.

[0028] In this embodiment, when constructing the user-item dynamic interaction graph and the user-category dynamic interaction graph, the time segmentation granularity is: 1 month for MovieLens adjacent timestamps, and 5 days for Retailrockets adjacent timestamps. In addition, the selection of interest time interval τ when constructing the user interest difference graph is τ = 2 for MovieLens and τ = 1 for Retailrockets.

[0029] The parameter settings involved in the DDGR model proposed by the method in the training process of the embodiment are: the number of training rounds ∈ = 6, the batch size b = 2046, the learning rate lr = 0.02, the dropout rate β = 0.5, and the embedding vector dimension size k = 128.

[0030] The experiment recommends K items for each user, calculates the hit rate (Hit Ratio, HR) and the ranking of the recommended results (Normalized Discounted Cumulative Gain, NDCG), that is, HR@K and NDCG@K. The comparison methods include MF, NCF, GCN+NCF, NGCF, GCMC, and GraphRec, and the following experimental data can be obtained.

[0031] Compared with the existing methods, the method has obvious recommendation effect improvement.

[0032]

[0033]

[0034]

[0035] The above specific embodiments can be adjusted in different ways by those skilled in the art without departing from the principles and purposes of the present application. The protection scope of the present application is subject to the claims and is not limited by the above specific embodiments. Each implementation within the scope is subject to the constraints of the present application.

Claims

1. A recommendation method based on dynamic difference map, characterized in that, Based on the interaction record of the user and the item, a user-item interaction graph and a user-category interaction graph at each moment are constructed, and further a newly added interest dynamic difference graph and a lost interest dynamic difference graph are constructed; based on the user-item interaction graph and the user-category interaction graph, a long-term / short-term interest modeling module is used to extract a long-term interest feature expression and a short-term interest feature expression of the user; Based on the newly added interest dynamic difference graph and the lost interest dynamic difference graph, an interest change trend modeling module is used to extract an interest change trend feature expression; Based on the user-item interaction graph and the user-category interaction graph, an item modeling module is used to extract a feature expression containing the category of the item; based on the long-term interest feature expression and the short-term interest feature expression of the user, the interest change trend feature expression, and the feature expression containing the category information of the item, a recommendation module is used to score the item, and an item with a high score is selected as the item recommended to the user; The newly added interest dynamic difference graph is wherein: is a user-category interaction graph at a moment, is a hyperparameter of time interval length, is a full zero matrix, and the size of the full zero matrix is the same as , represents taking the larger value between each position of the matrix and ; the loss-of-interest dynamic difference graph is ; The scoring includes: user scoring of items and scoring of interest level in categories of items .

2. The dynamic differential map-based recommendation method of claim 1, wherein the recommendation is made based on the dynamic differential map. It comprises: Step 1, collecting analysis samples of user interest changes; Step 2, constructing an interest dynamic difference graph, specifically including: 2.1) constructing a newly added interest dynamic difference graph; 2.2) constructing a lost interest dynamic difference graph; Step 3, constructing a recommendation system model including a long-term / short-term interest modeling module for extracting long-term / short-term interest feature expressions of the user, an item modeling module for extracting feature expressions of the item, an interest change trend modeling module for extracting interest change trend feature expressions of the user, and a recommendation module for recommending suitable items that the user is likely to interact with at the next moment; Step 4, training the recommendation system model constructed in step 3.

3. The dynamic differential map-based recommendation method according to claim 1 or 2, characterized in that, User-Item Interaction Graph This represents the interaction information between the user and the item, where: and These represent the sets of nodes for users and items, respectively. Is Time and Side data of user-item interactions generated over time intervals, i.e., when a user... With items exist and If there has been interaction between moments, then middle and There is an edge between them ;exist User-Category Interaction Graph This indicates the interaction information between the user and the item category.

4. The dynamic differential map-based recommendation method of claim 2, wherein, All modeling modules of the recommendation system model adopt an embedding expression layer to process one-hot encoding vectors, and then the one-hot encoding vectors are converted into low-dimensional vectors. Different types of nodes are processed in the same way in the embedding expression layer. The embedding vector of a user is obtained by calculating the one-hot encoding column vector of the user . The embedding vector of a user is obtained by calculating the one-hot encoding column vector of the user . is a parameter matrix randomly initialized by the embedding expression layer, is the transpose of . Similarly, the embedding vectors of an item and an item category are obtained. and .

5. The dynamic differential map-based recommendation method according to claim 1 or 2, characterized by, The short-term interest representation layer in the long-term / short-term interest modeling module is based on the formula... Calculation in Short-term interest characteristics of all users at any given time ,in: It is a user-item interaction diagram The adjacency matrix representation, It is a randomly generated parameter matrix. This represents the matrix resulting from concatenating the embedded representations of users and items. It is the sigmoid non-linear activation function; after the graph convolution operation, the user exist Characteristics of a moment That is The first in OK; The long-term interest representation layer in the long-term and short-term interest modeling module first applies the same graph convolution operation as the short-term interest representation layer to On the dynamic graph at each moment, the user short-term interest feature expression is obtained Then, a layer of attention network is used to determine the importance of the expression of the user at each timestamp, specifically as follows: After obtaining the weight of the expression of the user at each timestamp, the long-term interest feature expression of the user is calculated according to formula , wherein: is a randomly initialized and trainable parameter matrix.​ 6. The dynamic differential map-based recommendation method according to claim 1 or 2, characterized by, The interest change trend modeling module processes the two interest dynamic difference graphs through a dynamic difference graph neural network layer with two built-in dynamic graph attention convolution networks, to generate two interest feature expressions of the user. The specific structure of the dynamic difference graph neural network layer receives two types of dynamic graphs into the respective dynamic graph attention convolution networks inside it, i.e. a heterogeneous graph attention network, to obtain the newly added interest expression and the lost interest expression of the user at each timestamp, and then perform integration operations to obtain the interest change trend feature expression.

7. The dynamic differential map-based recommendation method of claim 6, wherein, The heterogeneous graph attention network extracts the category feature expression of the item, i.e. calculates the first-order correlation of the heterogeneous graph, and extracts the interest change trend feature expression of the user, i.e. calculates the second-order correlation of the heterogeneous graph.

8. The dynamic differential map-based recommendation method of claim 6, wherein, The integration layer in the recommendation module combines the user's final expression with the item's final expression, specifically including: the embedded expression of the item itself. ,pass To obtain user For items The scores, among which It is a multilayer perceptron; then, the category feature representations of the items are integrated. ,pass Get users items The categories are scored based on interest level, and then the two scores are combined using a self-attention mechanism. , The final calculation is obtained The scoring information at any given moment; The user's final expression is expressed as a short-term interest feature expression of the user , a long-term interest feature expression of the user , and an interest change trend expression of the user ; The item is finally expressed as an embedding vector of the item and a category feature expression of the item .

Images