A news recommendation method based on local graph self-supervised learning

Abstract

The application discloses a news recommendation method based on local graph self-supervised learning. According to the feature expression of news, the news is divided into two groups, and the interaction records of all users and one group of news are sorted into two local graphs respectively. The attribute information of the news is iteratively propagated on the two local graphs by using a deep graph propagation model. At the same time, the user feature vector and the news feature vector used for the recommendation task can be obtained by iteratively propagating on the global graph by using the deep graph propagation model. The learned news features on the two local graphs are used as self-supervised signals, and the self-supervised task and the recommendation task are jointly trained. The over-smoothing problem encountered in deep graph learning is constrained by using the self-supervised signals learned on the local graph, and more accurate user feature expressions and news feature expressions are learned, so that personalized news recommendation is realized. The application uses joint training of the recommendation task and the self-supervised task, and relieves the common over-smoothing problem of deep representation learning.

Description

Technical Field

[0001] This invention relates to the field of news recommendation technology, and specifically to a news recommendation method based on local graph self-supervised learning. Background Technology

[0002] The advent of the information age allows people to access social information more quickly through news. However, the rapid development of society has led to an increasing variety and quantity of news on various platforms, undoubtedly causing confusion for users who struggle to quickly find the news they want to know. News recommendation technology solves this problem by recommending news that users might be interested in based on their preferences, bringing convenience to users.

[0003] Traditional news recommendation methods primarily rely on historical behavioral data to obtain user preferences and then learn user behavioral representations through deep models. However, as model depth increases and behavioral representations become more refined, an oversmoothing problem arises. This leads to extremely high similarity in user feature vectors learned by deep networks, impairing recommendation performance. Therefore, mitigating the oversmoothing problem encountered when deep models learn behavioral representations is particularly important for news recommendation technology. Summary of the Invention

[0004] Therefore, the purpose of this invention is to provide a news recommendation method based on local graph self-supervised learning. News is divided into two groups according to its feature representation. All user interaction records for one group of news are organized into two local graphs. A deep graph propagation model is used to iteratively propagate the news attribute information across the two local graphs. Simultaneously, iterative propagation on the global graph using the deep graph propagation model yields user feature vectors and news feature vectors for the recommendation task. The news features learned from the two local graphs are used as self-supervised signals, enabling joint training of the self-supervised task and the recommendation task. The self-supervised signals learned from the local graphs constrain the oversmoothing problem encountered in deep graph learning, resulting in more accurate user and news feature representations, thus achieving personalized news recommendation.

[0005] The present invention solves the above-mentioned technical problems through the following technical means:

[0006] A news recommendation method based on local graph self-supervised learning includes:

[0007] Collect users' historical news browsing records as raw data, and number the users and news items in the raw data; map users and news items into feature vector space, and obtain the initial feature representation of users and the initial feature representation of news items through Gaussian initialization;

[0008] Using a multilayer perceptron, all users are divided into two groups according to their feature vectors, and two local graphs are constructed based on the two groups of users.

[0009] By using a depth graph propagation model to propagate information across local graphs, two sets of multi-layer local graph feature vectors for news are learned. The multi-layer local graph feature vectors are then pooled to obtain the final multi-layer local graph news feature vector.

[0010] The deep graph propagation model is used to propagate information across the global graph, and multi-layered refined user feature vectors and item feature vectors are learned. The multi-layered refined user feature vectors and item feature vectors are pooled to obtain the final user feature vectors and news feature vectors.

[0011] The local graph feature vectors of the two sets of news are used as self-supervised signals for joint training with the user feature vector and news feature vector of the global graph.

[0012] The final user feature vector and news feature vector are used to calculate the similarity. The similarity metric of each user vector with respect to all news vectors is sorted, and the top K news items are selected according to the size of the metric to generate Top-K news recommendations for the user.

[0013] Furthermore, the user's historical news browsing records are collected as raw data. Users and news items in the raw data are then numbered, specifically by assigning unique IDs to each. A Gaussian distribution is used to initialize the feature representations of users and news items, resulting in initial feature vectors for users and news items.

[0014] Furthermore, a multilayer perceptron is used to cluster all news items according to their feature vectors, dividing all users into two groups and retaining all news items. Two local graphs are constructed based on the two groups of users after classification, where the nodes on each local graph represent some users and all news items, and the edges represent the observable interactions between some users and all news items, i.e., the interaction records that exist in the historical data.

[0015] Furthermore, the depth graph propagation model is a graph convolution model, which uses the depth graph convolution model to propagate information in two local graphs to obtain two sets of multi-layer local graph feature vectors of news; the pooling operation is a mean operation, which takes the mean of the two sets of multi-layer local graph news feature vectors to obtain two sets of local graph news feature vectors.

[0016] Furthermore, the depth graph propagation model is a graph convolution model, which uses the depth graph convolution model to propagate information deep across the global graph, resulting in multi-layered refined user feature vectors and item feature vectors. The pooling operation is a mean operation, which takes the mean of the multi-layered refined user feature vectors and item feature vectors to obtain the final user feature vectors and news feature vectors. The global graph is defined as follows: the nodes on the global graph represent all users and all news items, and the edges represent all observable interactions of all users and all news items, i.e., all interaction records existing in the historical data.

[0017] Furthermore, the joint training involves training the deep network by adding the self-supervised loss and the recommendation loss together. The self-supervised loss uses the local graph feature vectors of the two sets of news as self-supervised signals, the recommendation loss is a pairwise loss function of BPR (Bayesian Pairwise Loss), and the optimizer is a stochastic gradient descent algorithm. This optimization method optimizes the final feature vectors of users and news.

[0018] Furthermore, the spatial distance similarity between the user vector and the candidate news vector is calculated using the inner product; the similarity metrics are then sorted to obtain the Top-K news recommendations for the user.

[0019] The beneficial effects of this invention are:

[0020] This invention uses a multilayer perceptron to divide the original news into two categories. Two local graphs are constructed using the classified partial news and all users respectively. The news representations in different local graphs can be used as self-supervised signals for the self-supervised task in this invention.

[0021] This invention uses joint training of recommendation tasks and self-supervised tasks to alleviate the oversmoothing problem common in deep representation learning; it utilizes efficient inner product operations to calculate the similarity between user vectors and news vectors to achieve personalized news recommendations. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0023] Figure 1 This is a general block diagram of the news recommendation method based on local graph self-supervised learning in this invention;

[0024] Figure 2 Construct a schematic diagram for the local graph;

[0025] Figure 3 A schematic diagram of modeling local graph features;

[0026] Figure 4 A schematic diagram of modeling global graph features;

[0027] Figure 5 This is a schematic diagram of joint training and interactive prediction. Detailed Implementation

[0028] To make the objectives, features, and advantages of the technical solution proposed in this invention more apparent and understandable, the embodiments of the technical solution proposed in this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the proposed technical solution, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0029] like Figure 1 As shown, the present invention provides a news recommendation method based on local graph self-supervised learning, comprising:

[0030] S1: Collect users' historical news browsing records as raw data, number the users and news in the raw data, map the users and news into the feature vector space, and obtain the initial feature representation of users and the initial feature representation of news through Gaussian initialization;

[0031] S2: Use a multilayer perceptron to divide all users into two groups according to their feature vectors, and construct two local graphs based on the two groups of users after grouping.

[0032] S3: Utilize the depth graph propagation model to propagate information across the local graph, learn two sets of multi-layer local graph feature vectors for news, and pool the multi-layer local graph feature vectors to obtain the final multi-layer local graph news feature vector.

[0033] S4: Utilize the depth graph propagation model to propagate information across the global graph, learn multi-layered refined user feature vectors and item feature vectors, and pool the multi-layered refined user feature vectors and item feature vectors to obtain the final user feature vector and news feature vector.

[0034] S5: Use the local graph feature vectors of the two sets of news as self-supervised signals for joint training with the user feature vector and news feature vector of the global graph;

[0035] S6: Calculate the similarity between the final user feature vector and the news feature vector. Sort the similarity metric values ​​of each user vector with respect to all news vectors, select the top K news items according to the size of the metric values, and generate Top-K news recommendations for the user.

[0036] Figure 2 The diagram shown is a partial view of the construction of this invention, and the specific steps include:

[0037] S1: Collect users' historical news browsing records as raw data, and number the users {u|u∈U} and news items {i|i∈I} in the raw data; assign unique IDs to users and news items, starting from 0 and sorted in ascending order. Map users and news items to the feature vector space, and obtain the initial feature representations of users and news items through Gaussian initialization, represented as follows: and

[0038] S2: After obtaining the initial feature representations of users and news from step S1, a multilayer perceptron is used to cluster all users according to their feature vectors, classifying the features. for

[0039]

[0040] in The feature transformation matrix, Let d be the bias matrix. The dimension is N, which is the number of users observed in the interactions. In this invention, d is set to 64. All users are divided into two main categories, U1 and U2, based on the classification feature F. All news items are retained, and two local graphs, G1 and G2, are constructed based on the two groups of users after classification, where:

[0041] Each node in the local graph represents all users and a portion of the news, and the edges represent the observable interactions between all users and a portion of the news, i.e., the interaction records r existing in the historical data. ui That is, G1 = {r ui |u∈U1,i∈I},G2={r ui |u∈U2,i∈I}.

[0042] Figure 3 This diagram illustrates the local graph feature modeling method used in this approach. The specific steps include:

[0043] S3: After obtaining two sets of local graphs G1 and G2 from step S2, a graph convolution model GCN is used. A depth graph convolution model is then employed to propagate information deep across the two local graphs G1 and G2, resulting in multi-layer local graph feature vectors {E} for the two sets of news. p1 0 E p1 1 , ..., E p1 L} and {E p2 0 E p2 1 , ..., E p2L},Right now

[0044] E p1 l+1 =D1 -1 A1 -1 E p1 l (2)

[0045] E p2 l+1 =D2 -1 A2 -1 E p2 l (3)

[0046] Where l is the number of layers in the depth graph propagation model, l∈[0,L]; when l=0,

[0047] L is the depth of the depth graph convolution model, which is set to 3 in this invention. A1 is the adjacency matrix formed by interactions on G1, and D1 is the degree matrix of A1, with the values ​​on the diagonal of the matrix representing the number of interactions between the user and the news. A2 is the adjacency matrix formed by interactions on G2, and D2 is the degree matrix of A2, with the values ​​on the diagonal of the matrix representing the number of interactions between the user and the news.

[0048] The pooling operation is a mean operation, which takes the mean of the two sets of multi-layer local graph news feature vectors to obtain two sets of local graph news feature vectors E. p1 and E p2 ,Right now

[0049] E p1 =1 / L(E p1 0 +E p1 1 +...+E p1 L (4)

[0050] E p2 =1 / L(E p2 0 +E p2 1 +...+E p2 L (5)

[0051] in

[0052] Figure 4 This diagram illustrates the global graph feature modeling process in this method. The specific steps include:

[0053] S4: After obtaining all interaction records from the original data in step S1, a depth graph convolution model is used to propagate information on the global graph G, where the nodes in the global graph represent all users and all news items, and the edges represent the observable interactions of all users and some news items, i.e., the interaction records r existing in the historical data. ui That is, G = {r} ui |u∈U,i∈I}.

[0054] The learning process yields multi-layered refined user feature vectors and item feature vectors {E}. u 0 E u 1 , ..., E u L}, and {E i 0 E i 1 , ..., E i L},Right now

[0055] E u l+1 =D -1 A -1 E u l (6)

[0056] E i l+1 =D -1 A -1 E i l (7)

[0057] Where A is the adjacency matrix formed by the interactions on G, and D is the degree matrix of A, with the values ​​on the diagonal of the matrix being the number of interactions between the user and the news.

[0058] Pooling and refining the user feature vectors and item feature vectors at multiple levels yields the final user feature vector E. u and news feature vector E i ,Right now

[0059] E u =1 / L(E u 0 +E u 1 +...+E u L (8)

[0060] E i =1 / L(E i 0 +E i 1+...+E i L (9)

[0061] Figure 5 This diagram illustrates the joint training and interactive prediction methods used in this approach. The specific steps include:

[0062] S5: Obtain two sets of local image news feature vectors E from step S3. p1 and E p2 The final user feature vector E is obtained from step S4. u and news feature vector E i Next, joint training is conducted. The loss L during joint training. all Two parts: self-monitored loss and recommended loss

[0063] L all =L ssl +L classify (10)

[0064] Self-monitored loss L ssl The formula is:

[0065]

[0066] Where i≠j, log is the logarithmic function, exp is the exponential function, τ is the hyperparameter, and in this experiment, τ is set to 1.5, e i1 For E p1 The vector in the i-th row represents the feature vector obtained from the first set of local images of news i. i2 T For e i2 transpose, e i2 For E p2 The vector in the i-th row represents the feature vector obtained from the second set of local images of news i, e j2 T Similarly.

[0067] Recommended loss L classify The loss function is called BPR (Bayesian Pairwise Loss). In the loss function, positive examples are news articles with user interaction, and negative examples are news articles with no observed interaction. The stochastic gradient descent optimization algorithm is used to optimize the Bayesian pairwise loss.

[0068] S6: Obtain the final feature vector E of users and news from step S4. u and E i The spatial distance similarity between user vectors and candidate news vectors is calculated using the inner product, i.e.

[0069] Y ui =E u T ·Ei (12)

[0070] Where Y ui E represents the predicted scores for user u and news item i. u T Represents vector E u After transposing, the similarity metrics are sorted to obtain the user's Top-K news recommendations.

Claims

1. A news recommendation method based on local graph self-supervised learning, characterized in that, The specific steps are as follows: S1: Collect users' historical news browsing records as raw data, number the users and news in the raw data, map the users and news into the feature vector space, and obtain the initial feature representation of users and the initial feature representation of news through Gaussian initialization; S2: Use a multilayer perceptron to divide all users into two groups according to their feature vectors. Construct two local graphs based on the two groups of users after grouping. The local graphs are the interaction relationship graphs G1 and G2 composed of some users and all news after user grouping. S3: Utilize the depth graph convolution model to propagate information across the local graph, learn two sets of multi-layer local graph feature vectors for news, and pool the multi-layer local graph feature vectors to obtain the final multi-layer local graph news feature vector. S4: Use the depth graph convolution model to propagate information on the global graph. The global graph is a complete interaction record graph G composed of all users and all news. Learn multi-layer refined user feature vectors and news feature vectors. Pool the multi-layer refined user feature vectors and news feature vectors to obtain the final user feature vectors and news feature vectors. S5: The local graph feature vectors of the two sets of news are used as self-supervised signals for joint training with the user feature vector and news feature vector of the global graph. The joint training is to add the self-supervised loss and the recommendation loss and then train them in the deep network. S6: Calculate the similarity between the final user feature vector and the news feature vector. Sort the similarity metric values ​​of each user vector with respect to all news vectors, select the top K news items according to the size of the metric values, and generate Top-K news recommendations for the user.

2. The news recommendation method based on local graph self-supervised learning according to claim 1, wherein, The user's historical news browsing history is collected as raw data. Users and news items in the raw data are then numbered, specifically including: S11: Assign a unified number to users and news items, and set a unique ID for each user and news item; S12: Initialize the feature representations of users and news using a Gaussian distribution to obtain the initial feature vectors of users and news.

3. The news recommendation method based on local graph self-supervised learning according to claim 1, wherein, S2 uses a multilayer perceptron to cluster all users according to their feature vectors, thereby dividing all users into a first group of users U1 and a second group of users U2, retaining all news articles, and constructing two local graphs based on the two groups of users after classification, wherein: Each node in the local graph represents a subset of users and all news items, while the edges represent observable interactions between a subset of users and all news items—that is, interaction records existing in historical data.

4. The news recommendation method based on local graph self-supervised learning according to claim 1, characterized in that: S3 utilizes a depth graph convolution model to propagate information through two local graphs, resulting in two sets of multi-layer local graph feature vectors for the news. The pooling operation is a mean operation, which takes the mean of the two sets of multi-layer local graph news feature vectors to obtain two sets of local graph news feature vectors.

5. The news recommendation method based on local graph self-supervised learning according to claim 1, characterized in that: S4 utilizes a depth graph convolution model to propagate information deep across the global graph, resulting in multi-layered refined user feature vectors and news feature vectors. The pooling operation is a mean operation, which takes the mean of the multi-layered refined user feature vectors and news feature vectors to obtain the final user feature vectors and news feature vectors. The global graph refers to the nodes on the global graph representing all users and all news items, and the edges representing the observable interactions of all users and all news items, i.e., all interaction records existing in the historical data.

6. The news recommendation method based on local graph self-supervised learning according to claim 1, characterized in that, The S5 joint training involves adding the self-supervised loss and the recommendation loss together and then training them collaboratively in the deep network. The specific operation is as follows: S51: Self-supervised loss uses the local graph feature vectors of the two news items as self-supervised signals; S52: The recommendation loss is a Bayesian pairwise loss function; S53: The optimizer is a stochastic gradient descent algorithm, which optimizes the final feature vectors of users and news.

7. The news recommendation method based on local graph self-supervised learning according to claim 1, characterized in that: S61: Calculate the spatial distance similarity between user vectors and candidate news vectors using the inner product; S62: Sort the similarity metrics to obtain the user's Top-K news recommendations.

Images