A bidirectional lightweight graph convolution collaborative filtering recommendation method

Abstract

The application discloses a bidirectional lightweight graph convolution collaborative filtering recommendation method, comprising the following steps: obtaining rating data of different items interacted by different users; classifying different items according to the rating data; respectively establishing bipartite graphs between different users and first items and second items, and respectively constructing high-order connectivity graphs between different users and the first items and the second items according to the two bipartite graphs, training the two high-order connectivity graphs on a LightGCN model to obtain two recommendation result sets; selecting K items with high rankings in the two recommendation result sets respectively, and further filtering the selected 2K items through a difference set operation to obtain a final recommendation result set. The application solves the problem that only a single scale aggregation is considered in the prior art without considering the multi-scale space aggregation problem, so that the obtained recommendation data has more discriminative recommendation results.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of graph neural networks, and particularly relates to a bidirectional light-weight graph convolution collaborative filtering recommendation method. BACKGROUND

[0002] Due to the rapid development of Internet technology, human society has entered an era of high informationization, and people can more easily and efficiently obtain various information. The quantity of network content also increases exponentially, leading to information explosion and information overload. In the face of massive information, how to quickly and accurately find accurate information needed by users, meet the personalized needs of users, and improve recommendation accuracy is the main challenge currently faced by a recommendation system.

[0003] In recent years, deep learning recommendation models have become the mainstream in the recommendation field, and traditional collaborative filtering algorithms have gradually combined with deep learning. At present, the mainstream is to combine deep learning methods and collaborative filtering models. From the perspective of a graph, user-item interaction can be regarded as a bipartite graph, and nodes include users and items; the edges in the bipartite graph are formed by connecting the user nodes with the item nodes (such as purchasing products, rating movies, sharing posts, etc.) that interact with them. The recommendation prediction problem can be regarded as a prediction task for the existence of edges in the graph. In the user-item interaction bipartite graph, the direct neighbors of a user are the items that interact with the user, and the direct neighbors of an item are the users that interact with the item. For the constructed user-item interaction graph, the graph neural network technology learns the effective embedding of the user or the item by capturing the nonlinear high-order interaction of the user-item relationship, thereby enhancing the expression ability of the model.

[0004] In addition, the recommendation model based on the graph convolutional network (Graph Convolutional Network) has achieved good results in recommendation performance, and the key feature is to use the connectivity of the graph to extract additional information and iteratively aggregate the feature information of the neighbors to improve the embedding representation of the user and the item. The neural graph collaborative filtering (Neural Graph Collaborative Filtering) uses the high-order connectivity of the graph to alleviate the sparsity problem of the recommendation system, and the light-weight graph neural collaborative filtering (LightGCN) model only retains the neighborhood aggregation operation and removes the self-loop in the aggregation operation to alleviate the over-smoothing problem.

[0005] However, most of the existing methods based on the graph convolutional network aggregate information in a single-scale space without considering information aggregation in a multi-scale space. SUMMARY

[0006] The application aims to provide a bidirectional lightweight graph convolution collaborative filtering recommendation method, which filters out items that users do not like through twice graph convolution, so as to ensure that the recommendation result of items that users like is more discriminative.

[0007] The application provides a bidirectional lightweight graph convolution collaborative filtering recommendation method, comprising the following steps:

[0008] Obtaining rating data of different items interacted by different users;

[0009] Classifying different items according to the rating data, wherein the classified items include first items and second items, and the rating score of the first item is greater than that of the second item;

[0010] Establishing a bipartite graph between different users and the first items, and constructing a high-order connectivity graph between different users and the first items according to the bipartite graph, and training the high-order connectivity graph on a LightGCN model to obtain a recommendation result set of the first items;

[0011] Establishing a bipartite graph between different users and the second items, and constructing a high-order connectivity graph between different users and the second items according to the bipartite graph, and training the high-order connectivity graph on a LightGCN model to obtain a recommendation result set of the second items;

[0012] Selecting the top K items in the recommendation result set of the first items, selecting the top K items in the recommendation result set of the second items, and further filtering the selected 2K items through a difference set operation to obtain a final recommendation result set.

[0013] Preferably, the obtaining of the rating data of different items interacted by different users comprises:

[0014] Obtaining the rating data of different items interacted by different users from shopping points data of different users respectively;

[0015] Obtaining the rating data of different items interacted by different users from the product browsing click volume of different users on a shopping website respectively;

[0016] Obtaining the rating data of different items interacted by different users from film review data of different users on a film review website respectively.

[0017] Preferably, the classifying of different items according to the rating data comprises the following steps:

[0018] Setting a threshold value X;

[0019] When the score of the item that the user has interacted with is greater than or equal to the threshold value X, the item is a first item;

[0020] When the score of the item that the user has interacted with is less than the threshold value X, the item is a second item.

[0021] Preferably, the definition expression form of the LightGCN model is as follows:

[0022]

[0023]

[0024] wherein is a preference prediction value of a user u for a first item i;

[0025] is a preference prediction value of a user u for a second item i:

[0026] e u is a set of all users who score high on certain items, and its expression is:

[0027]

[0028] wherein, a k represents the weight of the kth layer, and its value is

[0029] wherein is a set of all users who score high on certain items in the kth layer of the high-order connectivity graph;

[0030] e i is a set of all items scored high by certain users, and its expression is:

[0031]

[0032] wherein is a set of all items scored high by certain users in the kth layer of the high-order connectivity graph;

[0033] is a set of all users who score low on certain items, and its expression is:

[0034]

[0035] wherein is a set of all users who score low on certain items in the kth layer of the high-order connectivity graph;

[0036] is a set of all items scored low by certain users, and its expression is:

[0037]

[0038] wherein, is the set of all items scored low by some users in the kth layer of the high-order connectivity graph between the different users and the first item.

[0039] Preferably, the set of all users scored high on some items in the k+1th layer of the high-order connectivity graph between the different users and the first item is expressed as:

[0040]

[0041] wherein, is the normalization parameter in the LightGCN model;

[0042] N u is the one-hop neighbor of user u in the high-order connectivity graph;

[0043] N i is the one-hop neighbor of item i in the high-order connectivity graph;

[0044] wherein the set of all users scored low on some items in the k+1th layer of the high-order connectivity graph between the different users and the second item is expressed as:

[0045]

[0046] wherein the set of all items scored high by some users in the k+1th layer of the high-order connectivity graph between the different users and the first item is expressed as:

[0047]

[0048] wherein the set of all items scored low by some users in the k+1th layer of the high-order connectivity graph between the different users and the second item is expressed as:

[0049]

[0050] Preferably, the top K values corresponding to the items are selected in the preference prediction value of the user u on the first item i as the recommended result set M of the first item,

[0051] M = TopK(A) (11)

[0052] in the preference prediction value of the user u on the second item i selecting top K values corresponding to the items as the recommended result set N of the second item,

[0053] N = TopK (B) (12).

[0054] Preferably, the recommended result set M of the first item and the recommended result set N of the second item are subjected to difference set filtering, and a final recommended result set R is output.

[0055] R = M \ N (13).

[0056] The application provides a bidirectional lightweight graph convolution collaborative filtering recommendation device, comprising:

[0057] A first module is configured to acquire rating data of different items interacted by different users;

[0058] A second module is configured to classify different items according to the rating data, wherein the classified items include first items and second items, and the rating score of the first item is greater than that of the second item;

[0059] A third module is configured to establish a bipartite graph between different users and the first items and a bipartite graph between different users and the second items;

[0060] A fourth module is configured to construct a high-order connectivity graph between different users and the first items according to the bipartite graph between different users and the first items, and to construct a high-order connectivity graph between different users and the second items according to the bipartite graph between different users and the second items;

[0061] A fifth module is configured to train the high-order connectivity graph between different users and the first items on a LightGCN model to obtain a recommended result set of the first item, and to train the high-order connectivity graph between different users and the second items on the LightGCN model to obtain a recommended result set of the second item;

[0062] A sixth module is configured to select top K items in the recommended result set of the first item and top K items in the recommended result set of the second item, and to further refine and filter the selected 2K items through a difference set operation to obtain a final recommended result set.

[0063] Compared with the prior art, the application has the following beneficial effects:

[0064] This invention proposes a multi-scale algorithm model based on graph convolutional neural networks to address the aforementioned single-scale problem, which only considers aggregation at a single scale and neglects aggregation across multiple scales. To improve the accuracy of recommendation results, this technical solution divides the original user-item interaction graph into two subgraphs: one for user preferences and the other for user aversions. Then, the LightGCN model is applied to each subgraph to predict item preferences. The predicted preferences are then ranked, and the top K preferred options (top-k items) are selected, resulting in two result sets. These two result sets are then filtered using a difference filter to obtain the final recommendation result. Attached Figure Description

[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0066] Figure 1 This is a flowchart of a bidirectional lightweight graph convolutional collaborative filtering recommendation method proposed in this invention;

[0067] Figure 2 This is an example illustration of step 1 in one embodiment of this disclosure;

[0068] Figure 3 This is an example illustration of step 2 in one embodiment of this disclosure;

[0069] Figure 4 This is an example illustration of step 2 in one embodiment of this disclosure;

[0070] Figure 5 This is an illustration of an overall process example in one embodiment of the present disclosure. Detailed Implementation

[0071] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. However, it should be understood that the scope of protection of the present invention is not limited to the specific implementation.

[0072] Example

[0073] like Figures 1-5 As shown, a bidirectional lightweight graph convolutional collaborative filtering recommendation method includes the following steps:

[0074] Step 1: Obtain rating data from different users for the different items they have interacted with, specifically including:

[0075] Obtain rating data for different items that different users have interacted with from their shopping points data;

[0076] We obtain rating data for different items that different users have interacted with from the product browsing and click volume of different users on shopping websites;

[0077] We obtained rating data for different items that different users had interacted with from film review data on film review websites.

[0078] In practice, rating data of users on various projects they have interacted with can be obtained from the URLs used by various users, forming a rating database of different projects by different users.

[0079] Step 2: Classify the different items according to the scoring data. The classified items include the first item and the second item. The score of the first item is greater than the score of the second item.

[0080] The different items are categorized based on the scoring data, including the following steps:

[0081] Set the threshold X;

[0082] When a user's rating of an item they have interacted with is greater than or equal to a threshold X, that item is designated as the first item.

[0083] When a user's rating for an item they have interacted with is less than a threshold X, that item is designated as the second item.

[0084] In this embodiment, as Figure 2 The left side shows the rating data for all items that all users have interacted with, with a threshold of X=3.

[0085] The ratings for the first item, i.e., the items users like, will be re-marked according to the following rules.

[0086]

[0087] Where P ij Indicates user u i Among the projects he has interacted with, his favorite projects i j The rating data yielded the following results: Figure 2 The rating data in the upper right corner.

[0088] The ratings for the second item, i.e., items that users dislike, will be re-marked according to the following rules.

[0089]

[0090] Indicates user u i Disliked items among the projects they have interacted with j The rating data yielded the following results: Figure 2 The rating data in the lower right corner.

[0091] Based on this, we can obtain two sets of data: one set represents data on users' liking of certain items, and the other set represents data on users' dislike of certain items.

[0092] Step 3: Establish a bipartite graph between different users and the first item, and construct a higher-order connectivity graph between different users and the first item based on the bipartite graph. Train the higher-order connectivity graph on the LightGCN model to obtain the recommendation result set of the first item.

[0093] Step 4: Establish a bipartite graph between different users and the second item, and construct a higher-order connectivity graph between different users and the second item based on the bipartite graph. Train the higher-order connectivity graph on the LightGCN model to obtain the recommendation result set of the second item.

[0094] Most existing methods only use descriptive features to construct embedding functions, such as user IDs and attributes, without considering the interaction between users and items. When the embedding cannot adequately capture the collaboration signals, interaction functions are needed to compensate for the shortcomings of the embedding. In this invention, we utilize the higher-order connectivity of user-item interactions to extract collaboration signals between users and items.

[0095] This invention plots higher-order connectivity subgraphs from two sets of data: one representing a higher-order connectivity bipartite graph L of items the user likes, and the other representing a higher-order connectivity bipartite graph D of items the user dislikes, as shown below. Figure 3 , Figure 4 As shown.

[0096] The basic idea of ​​this invention is to obtain the final recommendation result through two LightGCN iterations, and then refine the recommendation result trained by the first LightGCN model using the recommendation result trained by the second LightGCN model. For the two high-order connectivity bipartite graphs obtained above, training is performed on the LightGCN model separately, resulting in two sets of recommendation results: one set representing the recommendations for the first item (items the user likes), and the other set representing the recommendations for the second item (items the user dislikes).

[0097] In the bidirectional lightweight graph convolutional collaborative filtering recommendation method, we only aggregate connected neighbors, without integrating the target node itself (i.e., self-connections). The propagation law of bidirectional lightweight graph convolution is as follows:

[0098] The specific steps involved in training a LightGCN model are as follows:

[0099] Calculate the set of all users who have given high ratings to some items at level k+1 of the higher-order connectivity graph between different users and the first item.

[0100]

[0101] in, These are the normalized parameters in the LightGCN model;

[0102] N u Let u be a one-hop neighbor of user u in a higher-order connectivity graph;

[0103] N i It is a one-hop neighbor of product i in the higher-order connectivity graph;

[0104] in This is the set of all items in the k-th layer of a higher-order connectivity graph that have been highly rated by some users.

[0105] Calculate the set of all users with low ratings for certain items at level k+1 of the higher-order connectivity graph between different users and the second item.

[0106]

[0107] in This is the set of all items in the k-th layer of a higher-order connectivity graph that have been rated low by some users.

[0108] Calculate the set of all items at level k+1 of the higher-order connectivity graph between different users and the first item that are highly rated by some users.

[0109]

[0110] in Let be the set of all users who give high ratings to certain items at the k-th level of the higher-order connectivity graph;

[0111] Calculate the set of all items at level k+1 of the higher-order connectivity graph between different users and the second item that are rated low by some users.

[0112]

[0113] in Let be the set of all users at the k-th level of the higher-order connectivity graph who have given low ratings to certain items.

[0114] In the bidirectional lightweight graph convolutional collaborative filtering recommendation method, the trainable model parameters are the embeddings of layer 0, given... and Then, based on the propagation law of bidirectional graph convolution, higher-level embeddings are calculated. After K layers of propagation, we further combine the embeddings obtained at each layer to form the final representation of a user and an item as follows:

[0115] e u The set of all users who give high ratings to certain items is expressed as:

[0116]

[0117] Among them, a k This represents the weight of the k-th layer, and its value is...

[0118] e i The set of all items highly rated by certain users is expressed as:

[0119]

[0120] The set of all users who gave low ratings to certain items is expressed as:

[0121]

[0122] The set of all items that have been rated low by some users is expressed as:

[0123]

[0124] The LightGCN model predicts preference values ​​as the inner product of the final representations of users and items.

[0125] The LightGCN model is defined as follows:

[0126]

[0127]

[0128] in Let be the predicted preference value of user u for the first item i;

[0129] Here is the predicted preference value for user u for the second item i:

[0130] Step 5: Select the top K items from the recommendation result set of the first project, select the top K items from the recommendation result set of the second project, and further filter the selected 2K items through the difference operation to obtain the final recommendation result set.

[0131] The predicted preference value of user u for the first item i The top K items corresponding to the values ​​are selected as the recommendation result set M for the first item.

[0132] M = TopK(A) (11)

[0133] The predicted preference value of user u for the second item i The top K items corresponding to the values ​​are selected as the recommendation result set N for the second item.

[0134] N = TopK(B) (12)

[0135] Perform difference filtering on the recommendation result set M of the first project and the recommendation result set NN of the second project to output the final recommendation result set R.

[0136] R = M\N (13).

[0137] This invention proposes a bidirectional lightweight graph convolutional collaborative filtering recommendation device, comprising:

[0138] The first module is used to obtain rating data from different users for different items they have interacted with;

[0139] The second module is used to classify different items based on the scoring data. The classified items include the first item and the second item, and the score of the first item is greater than the score of the second item.

[0140] The third module is used to create a bipartite graph between different users and the first project, as well as a bipartite graph between different users and the second project.

[0141] The fourth module is used to construct a higher-order connectivity graph between different users and the first project based on the bipartite graph between different users and the first project, and is also used to construct a higher-order connectivity graph between different users and the second project based on the bipartite graph between different users and the second project.

[0142] The fifth module is used to train the high-order connectivity graph between different users and the first item on the LightGCN model to obtain the recommendation result set of the first item, and to train the high-order connectivity graph between different users and the second item on the LightGCN model to obtain the recommendation result set of the second item.

[0143] The sixth module is used to select the top K items from the recommendation result set of the first project and the top K items from the recommendation result set of the second project. The selected 2K items are further refined and filtered through the difference operation to obtain the final recommendation result set.

[0144] Finally, it should be noted that the above-disclosed embodiment is only one specific embodiment of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A bidirectional lightweight graph convolution collaborative filtering recommendation method, characterized in that, The method comprises the following steps: obtaining rating data of different users on different items interacted by the users; classifying different items according to the rating data, wherein the classified items comprise first items and second items, and a rating score of the first items is greater than a rating score of the second items; establishing a bipartite graph between different users and the first items, and constructing a high-order connectivity graph between the different users and the first items according to the bipartite graph, training the high-order connectivity graph on a LightGCN model to obtain a recommended result set of the first items; establishing a bipartite graph between different users and the second items, and constructing a high-order connectivity graph between the different users and the second items according to the bipartite graph, training the high-order connectivity graph on the LightGCN model to obtain a recommended result set of the second items; selecting K items with high rankings in the recommended result set of the first items, selecting K items with high rankings in the recommended result set of the second items, and further filtering the selected 2K items by a difference set operation to obtain a final recommended result set. The method comprises the following steps: obtaining rating data of different users on different items interacted by the users; obtaining rating data of different users on different items interacted by the users from shopping point data of the different users; obtaining rating data of different users on different items interacted by the users from product browsing click data of the different users on a shopping website; obtaining rating data of different users on different items interacted by the users from film review data of the different users on a film review website.

3. The bidirectional LightGCN collaborative filtering recommendation method according to claim 2, wherein the classifying different items according to the rating data comprises the following steps: setting a threshold X; 2. The bidirectional lightweight graph convolution collaborative filtering recommendation method according to claim 1, characterized in that: when a rating of an item interacted by a user is greater than or equal to the threshold X, the item is a first item; when a rating of an item interacted by a user is less than the threshold X, the item is a second item. The definition expression of the LightGCN model is as follows: a set of different users and a set of different items; all nodes in a k+1 layer of the high-order connectivity graph between the different users and the first items all nodes in a k+1 layer of the high-order connectivity graph between the different users and the second items all nodes in a k+1 layer of the high-order connectivity graph between the different users and the first items all nodes in a k+1 layer of the high-order connectivity graph between the different users and the second items In the present application: performing difference set filtering on the recommended result set M of the first items and the recommended result set M' of the second items to output a final recommended result set R; The method comprises the following steps: a first module for obtaining rating data of different users on different items interacted by the users; a second module for classifying different items according to the rating data, wherein the classified items comprise first items and second items, and a rating score of the first items is greater than a rating score of the second items; a third module for establishing a bipartite graph between different users and the first items, and a bipartite graph between different users and the second items; ​ 4. The bidirectional lightweight graph convolution collaborative filtering recommendation method according to claim 3, characterized in that: ​ （ i =1,2,3...） （1） （ i =1,2,3...） （2） wherein for a user u a preference prediction value for the first item i; for the user u a predicted value of the preference for the second item i: For all the set of users who scored high on the project, the expression is: （3） wherein, represents the weight of the layer, whose value is k ; and ; wherein is the set of all users in the high-order connectivity graph whose score for the k layer is higher than the score of the item; All the items that are rated highly by the user, expressed as: （4） wherein is the highest ranked user in the high-order connectivity graph k all users of the layer that have a high user rating ​ For all the set of users with low scores on the project, the expression is: （5） wherein is the set of all users in the high-order connectivity graph whose score for the item in the layer is low k layer; For all the item set scored low by the user, the expression is: （6） wherein, is the set of all items of the layer k with a low user score in the high-order connectivity graph.

5. The bidirectional lightweight graph convolution collaborative filtering recommendation method according to claim 4, characterized in that: ​ A set of users with high scores on the project Expression: （7） wherein, is a normalization parameter in the LightGCN model; for the user u one-hop neighbors in a high-order connectivity graph; for and merchandise i One-hop neighbors in high-order connectivity graphs; ​ A set of users scoring low on the project Expression: （8） ​ A collection of items that have high user ratings Expression of: （9） ​ A collection of items with low user ratings Expression of: （10）。 6. The bidirectional lightweight graph convolution collaborative filtering recommendation method according to claim 5, characterized in that: In the user u The first item i The preference prediction value The first item （11） In the user u The second item i The preference prediction value The K top values are selected as the recommended result set N of the second item （12）。 7. The bidirectional lightweight graph convolution collaborative filtering recommendation method according to claim 6, characterized in that: ​ ​ （13）。 8. A bidirectional lightweight graph convolutional collaborative filtering recommendation apparatus characterized by comprising: ​ ​ ​ ​ ​ ​ a bipartite graph between the user and the second item; a fourth module configured to construct a high-order connectivity graph between different users and the first item according to bipartite graphs between the different users and the first item, and to construct a high-order connectivity graph between different users and the second item according to bipartite graphs between the different users and the second item; a fifth module configured to train the high-order connectivity graph between the different users and the first item on a LightGCN model to obtain a recommended result set of the first item, and to train the high-order connectivity graph between the different users and the second item on the LightGCN model to obtain a recommended result set of the second item; a sixth module configured to select the top K items in the recommended result set of the first item, and to select the top K items in the recommended result set of the second item, and to further refine and filter the selected 2K items through a difference set operation to obtain a final recommended result set. ​ ​ ​ ​ ​ ​ ​

Images