A service recommendation method based on decoupled representation learning and graph neural networks

Abstract

The application discloses a service recommendation method based on decoupling representation learning and a graph neural network, takes user interaction logs recorded in online applications as a data source, obtains complete historical interaction sequences and corresponding category sequences of a user, first respectively captures long-term stable interest and short-term dynamic interest of the user in the degree of personalization, and then captures herd interest of the user by using service category information. The method aims at the problem that an implicit decoupling representation learning-based method depends on a number of artificially set latent interests, is difficult to adapt to different application scenarios, and does not consider different online application service category factors. The long-term interest, the short-term interest and the herd interest of the user are respectively learned based on complete sequences, recent sequences and category sequences, the disadvantages of artificially setting the number of latent interests are avoided, the influence of different online application category factors is considered, the herd interest of the user is captured as a basis for interpretability of a recommendation result, and application providers are helped to adjust a recommendation strategy in a targeted manner.

Description

A service recommendation method based on decoupled representation learning and graph neural networks Technical Field

[0001] This invention belongs to the field of recommendation system technology, specifically relating to a service recommendation method based on decoupled representation learning and graph neural networks. Background Technology

[0002] In recent years, with the rapid popularization of internet technology, people's demand for online application services (such as online music, video streaming, and e-shopping) in their daily lives has been increasing. Recommendation systems utilize users' historical behavior records to build user profiles, helping online application providers to recommend services that users may be interested in, which has significant potential economic value. For example, NetEase Cloud Music generates user profiles based on users' listening history, and then recommends songs with potential membership badges to attract users to upgrade their memberships; the Tmall shopping platform recommends similar products based on users' purchasing and search behaviors to help users select products, thereby improving purchase conversion rates.

[0003] As research into deep learning technology deepens, many current models of user preferences based on user historical behavior sequences focus on the relationship between users and interactive services. For example, they capture users' short- and long-term interests through sequence patterns, and capture second- and higher-order relationships between users and services by building graph structures, thus learning the overall interest characteristics of users. However, user behavior is the result of multiple factors. In terms of personalization, there are long-term stable interests and short-term random dynamic interests; in terms of service popularity, there are herd-based popularity interests. Different interests will have varying degrees of influence on users' future behavior.

[0004] To capture users' potential interests across different aspects through their historical behavior records, many methods employ implicit decoupling. First, a predetermined number of potential interests is set as a prerequisite for decoupling representation learning. Then, feature learning methods such as graph neural networks are used to learn the potential interests in different aspects, ultimately yielding the user's overall preferences. The shortcomings of existing methods lie in their failure to analyze the overall distribution of potential interests, their difficulty in adjusting the number of potential interests based on actual application recommendation scenarios, and their neglect of the unique characteristics of different online services (such as category factors), impacting the economic benefits for online application providers. Summary of the Invention

[0005] This invention addresses the problems of implicit decoupled representation learning-based methods, which rely on manually set potential interest numbers, struggle to adapt to different service scenarios, and fail to consider factors related to different online application categories. It proposes a service recommendation method based on decoupled representation learning and graph neural networks. First, it captures users' long-term stable interests and short-term dynamic interests to varying degrees of personalization. Then, it utilizes service category information to capture users' conformity interests.

[0006] The specific steps of this invention are as follows:

[0007] Step (1). Using user interaction logs recorded by online applications as the data source, collect the user's complete historical interaction sequence and corresponding category sequence. The complete historical interaction sequence refers to the user's interaction records with the service (such as clicks, favorites, ratings, etc.), and the corresponding category sequence refers to the sequence constructed from the category information corresponding to each service. Define I as the set of all services, with n services; define C as the set of all service categories, with k categories; define the historical interaction sequence of user u as T = t1, t2, ..., t l-1 ,t l ,t l+1 , where t i The record representing the i-th interaction contains a tuple of the service and its corresponding category, i.e., t. i =(I i C i ), l represents the sequence length, t l+1 This indicates the next interaction to be predicted.

[0008] Step (2). Input the service set I and the category set C into the embedding layer to obtain the embedded feature representation of each service and its corresponding category. The embedding layer refers to the multi-dimensional feature vector obtained through a one-hot encoder and a multilayer perceptron to represent each service or category. The embedded feature representation obtained by the embedding layer is defined as two feature matrices: X i ∈R n×d For the service feature matrix, X c ∈R k×d Let be the category feature matrix, where n represents the number of services, k represents the number of categories, and d represents the feature dimension.

[0009] Step (3). In order to introduce category factors to decouple user interest learning, a long-term interest and short-term interest modeling network based on user interaction sequences and a category-based user conformity interest modeling module were designed, including the following three steps:

[0010] Step (3.1). First, a hypergraph is constructed based on the service sequence in the complete historical interaction sequence, and then the user's long-term interests are captured after passing through the hypergraph convolutional layer.

[0011] 1) Define the service hypergraph as GI∈(V i E i V i and E i These are the vertex set and the hyperedge set, respectively, and the hypergraph adjacency matrix H. i ∈{0,1} m×n Hyperedges (a hyperedge is a special edge that can connect multiple vertices) are constructed based on the service sequences in a user's historical behavior sequence. For m users, m hyperedges will be constructed. The hypergraph adjacency matrix H corresponds to the services contained in the service sequences. i The position will be set to 1 otherwise, 0. For example, if the first user interacts with the first service, then H... i (0,0)=1.

[0012] 2) Define the vertex degree matrix D of the hypergraph. iv ∈R n×n and hypermarginality matrix D ie ∈R m×m They are diagonal matrices relating vertex degree and hyperedge degree, respectively, where vertex v∈V i degree Hyperedge e∈E i degree

[0013] 3) Define a hypergraph convolutional neural network for feature learning and information propagation. The computation method between adjacent network layers is as follows: in W represents the service feature representation corresponding to the service sequence. i ∈R m×m Let represent the weight matrix, and σ(·) represent the nonlinear activation function. Then, an average pooling strategy is used to obtain the feature representation of the service:

[0014]

[0015] Where L represents the number of convolutional network layers, and finally, the user's long-term interest h is obtained by aggregating the feature information in the sequence using an averaging strategy. lo :

[0016]

[0017] Where l represents the length of the user's historical interaction sequence.

[0018] Step (3.2). Then, capture the user's short-term interests based on the service sequence in the most recent historical interaction sequence. The specific calculation process is as follows.

[0019] 1) For the complete user interaction sequence T = t1, t2, ..., tl-1 ,t l Select the nearest interactive subsequence T s =t l-a ,…,t l-1 ,t l 'a' represents the number of interaction records selected, which can be set according to different application scenarios. Then, the service feature sequences corresponding to the sub-sequences are passed through a bidirectional LSTM encoding layer to combine temporal information and capture the user's short-term interest representation. in Let H represent the initial service feature representation. The feature matrix of m users' short-term interest representations is denoted by H, and f represents the feature aggregation method.

[0020] 2) Construct a user graph (GU) with users as vertices. The condition for connecting vertices is that users have the same nearest-neighbor service. Use graph representation learning to capture the potential connections between different users. The specific propagation method is as follows:

[0021] H (l+1) =σ(D -1 AH l )

[0022] Where D and A are the degree matrix and identity matrix of the user graph GU, respectively, and σ(·) is the activation function. Similarly, the short-term user interest representation h is obtained using the average pooling strategy. sh ∈H:

[0023]

[0024] Step (3.3). To adapt to the data characteristics of different applications, this method introduces service category information as the data foundation to capture users' herd interest in category popularity. Similar to step (3.1), a category hypergraph GC∈(V) is constructed using the category sequence in the complete historical interaction sequence. c E c V c and E c These are the vertex set and the hyperedge set, respectively, and the hypergraph adjacency matrix H. c ∈{0,1} m×k Define the categorical hypergraph convolution calculation: in D represents the category feature representation corresponding to the category sequence. cv ∈R k×k and D ce ∈R m×m Let W represent the vertex degree matrix and hyperedge degree matrix of the categorical hypergraph, respectively. c ∈R m×m Let σ(·) represent the weight matrix, and let σ(·) represent the non-linear activation function. After passing through the average pooling layer, the user's conformity interest h is obtained.co .

[0025] Step (4). Based on step (3), the long-term user interests h learned are respectively... lo Short-term interest h sh and herd interest co Feature fusion is performed through a multi-interest aggregation layer to obtain the final user interest representation h. The fusion method is as follows:

[0026] h = W[h] lo ||h sh ||h co ]+b

[0027] Where W and b represent the weight matrix and bias term, respectively, and [·||·||·] represents the concatenation of eigenvectors along the dimension.

[0028] Step (5). Based on the user interest h learned in step (4), perform the service recommendation task. For the candidate service set I, the recommendation score z for each service is determined by the service feature vector x∈X. i The inner product of z and the user interest feature vector h is obtained, i.e., z i =x T h, use the softmax function to normalize z to obtain the final service score. Then, the top-N services with the highest ratings are selected as candidate services in the recommendation list. The objective function for the recommendation task is represented by cross-entropy loss:

[0029]

[0030] in It refers to the probability that the target service will generate an interaction given a sequence of interaction records T.

[0031] Compared with the prior art, the present invention has the following beneficial effects: by using the decoupled learning of user interests, the long-term interests, short-term interests and herd interests of users can be learned separately, which avoids the drawbacks of manually setting the number of potential interests to the greatest extent. Furthermore, considering the influence of different online application categories, the herd interests of users are captured as the basis for the interpretability of recommendation results, helping application providers to adjust recommendation strategies in a targeted manner. Attached Figure Description

[0032] Figure 1 is a framework diagram of a service recommendation method based on decoupled representation learning and graph neural network according to an embodiment of the present invention.

[0033] Figure 2 is a diagram of the feature embedding module based on the nearest sequence in an embodiment of the present invention. Detailed Implementation

[0034] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0035] The technical solution of the present invention will be further explained and described below with reference to the accompanying drawings.

[0036] This embodiment provides a service recommendation method based on decoupled representation learning and graph neural networks. The overall execution framework of the method is shown in Figure 1.

[0037] Step (1). Using user interaction logs recorded by online applications as the data source, collect the user's complete historical interaction sequence and corresponding category sequence. The complete historical interaction sequence refers to the user's interaction records with the service (such as clicks, favorites, ratings, etc.), and the corresponding category sequence refers to the sequence constructed from the category information corresponding to each service. Define I as the set of all services, with n services; define C as the set of all service categories, with k categories; define the historical interaction sequence of user u as T = t1, t2, ..., t l-1 ,t l ,t l+1 , where t i The record representing the i-th interaction contains a tuple of the service and its corresponding category, i.e., t. i =(I i C i ), l represents the sequence length, t l+1 This indicates the next interaction to be predicted.

[0038] Step (2). Input the service set I and the category set C into the embedding layer to obtain the embedded feature representation of each service and its corresponding category. The embedding layer refers to the multi-dimensional feature vector obtained through a one-hot encoder and a multilayer perceptron to represent each service or category. The embedded feature representation obtained by the embedding layer is defined as two feature matrices: X i ∈R n×d For the service feature matrix, X c ∈R k×d Let be the category feature matrix, where n and k represent the number of services and categories, respectively, and d represents the feature dimension.

[0039] Step (3). In order to introduce category factors to decouple user interest learning, a long-term interest and short-term interest modeling network based on user interaction sequences and a category-based user conformity interest modeling module were designed, including the following three steps:

[0040] Step (3.1). First, based on the service sequence T in the complete historical interaction sequence. i =I1,I2,…,I l-1 ,I lA hypergraph is constructed, and after passing through a hypergraph convolutional layer, the user's long-term interest is captured and represented as h. lo .

[0041] 1) Define the service hypergraph as GI∈(V i E i V i and E i These are the vertex set and the hyperedge set, respectively, and the hypergraph adjacency matrix H. i ∈{0,1} m×n Hyperedges (special edges that connect multiple vertices) are constructed based on the service sequences in a user's historical behavior sequence. For m users, m hyperedges are constructed. The adjacency matrix Hi corresponding to the services included in the service sequence is set to 1; otherwise, it is 0. If the first user interacts with the first service, then H... i (0,0) = 1. For example, if the service sequences in the historical interaction records of two users u1 and u2 are {I1,I2,I2,I3} and {I2,I4,I2,I5} respectively, then the corresponding service hypergraph adjacency matrix is:

[0042]

[0043] 2) Define the vertex degree matrix D of the hypergraph. iv ∈R n×n and hypermarginality matrix D ie ∈R m×m They are diagonal matrices relating vertex degree and hyperedge degree, respectively, where vertex v∈V i degree Hyperedge e∈E i degree

[0044] 3) Define a hypergraph convolutional neural network for feature learning and information propagation. The computation method between adjacent network layers is as follows: in W represents the service feature representation corresponding to the service sequence. i ∈R m×m Let represent the weight matrix, and σ(·) represent the nonlinear activation function. Then, an average pooling strategy is used to obtain the feature representation of the service:

[0045]

[0046] Where L represents the number of convolutional network layers. This represents the service feature matrix of layer l. Finally, the user's long-term interest h is obtained by aggregating the feature information in the sequence using an averaging strategy. lo :

[0047]

[0048] Where l represents the length of the user's historical interaction sequence, X i,t This represents the feature representation of the t-th service in the service sequence.

[0049] Step (3.2). Then, based on the service sequence in the most recent historical interaction sequence, the user's short-term interests are captured. The session embedding aggregation module is shown in Figure 2. The specific calculation process is as follows.

[0050] 1) For the complete user interaction sequence T = t1, t2, ..., t l-1 ,t l Select the nearest interactive subsequence T s =t l-a ,…,t l-1 ,t l 'a' represents the number of interaction records selected, which can be set according to different application scenarios. Then, the service feature sequences corresponding to the sub-sequences are passed through a bidirectional LSTM encoding layer to combine temporal information and capture the user's short-term interest representation. in This represents the initial service feature representation. The feature matrix of the short-term interest representations of m users is H∈R. m×d f represents the bidirectional LSTM feature aggregation method.

[0051] 2) Construct a user graph (GU) with users as vertices. An edge connects vertices if users share the same nearest-neighbor service. For example, if users u1 and u2 have nearest-neighbor service sequences {I2, I3} and {I2, I5} respectively, they share the same service I2, therefore users u1 and u2 will be connected by an edge. Graph representation learning is used to capture the potential connections between different users. The specific propagation method is as follows:

[0052] H (l+1) =σ(D -1 AH l )

[0053] Where D∈R m×m and A∈R m×m Here, are the degree matrix and identity matrix of the user graph GU, respectively, and σ(·) is the activation function. Similarly, the short-term user interest representation h is obtained using the average pooling strategy. sh ∈H:

[0054]

[0055] Where L represents the number of layers in the user graph convolutional neural network, This represents the short-term interest features of users in the l-th layer.

[0056] Step (3.3). To adapt to the data characteristics of different applications, this method introduces service category information as the data basis to capture users' herd interest h generated by category popularity. co Similar to step (3.1), the category sequence T in the complete historical interaction sequence is used. c =C1,C2,…,C l-1 C l Construct a category hypergraph GC∈(V) c E c V c and E c These are the vertex set and the hyperedge set, respectively, and the hypergraph adjacency matrix H. c ∈{0,1} m×k For example, if the category sequences in the historical interaction records of two users u1 and u2 are {C1,C1,C1,C2} and {C2,C1,C2,C3} respectively, then the corresponding category hypergraph adjacency matrix is:

[0057]

[0058] Define categorical hypergraph convolution calculation: in D represents the category feature representation corresponding to the category sequence. cv ∈R k×k and D ce ∈R m×m Let W represent the vertex degree matrix and hyperedge degree matrix of the categorical hypergraph, respectively. c ∈R m×m Let σ(·) represent the weight matrix, and let σ(·) represent the non-linear activation function. After passing through the average pooling layer, the user's conformity interest h is obtained. co .

[0059] Step (4). Based on step (3), the long-term user interests h learned are respectively... lo Short-term interest h sh and herd interest co Feature fusion is performed through a multi-interest aggregation layer to obtain the final user interest representation h. The fusion method is as follows:

[0060] h = W[h] lo ||h sh ||h co ]+b

[0061] Where W∈R d×3d and b∈R d×1 represents the weight matrix and the bias term, respectively, and [·||·||·] represents the concatenation of eigenvectors along the dimension.

[0062] Step (5). Based on the user interest h learned in step (4), perform the service recommendation task. For the candidate service set I, the recommendation score z for each service is determined by the service feature vector x∈X. i The inner product of z and the user interest feature vector h is obtained, i.e., z i =x T h, use the softmax function to normalize z to obtain the final service score. Then, the top-N services with the highest ratings are selected as candidate services in the recommendation list. The objective function for the recommendation task is represented by cross-entropy loss:

[0063]

[0064] in This refers to the probability that an interaction will occur with the target service given a sequence of interaction records T. Regarding the interpretability of the results, if user u1's service sequence and category sequence are {I1,I2,I3} and {C1,C1,C2} respectively, and user u1's next interaction item is actually I4 (category C1), the result cannot be interpreted solely based on the service history. However, if the category data shows that user u1 frequently interacts with services of category C1, then the interpretation of the recommendation results can focus on the user's herd interests.

[0065] This method uses historical user interaction data as a foundation, and captures users' long-term and short-term interests based on decoupled representation learning. Then, it introduces the service category popularity factor and constructs a category hypergraph convolutional neural network to capture users' herd interest in service categories, thereby improving the recommendation model's ability to adapt to different application scenarios and the interpretability of the results.

Claims

1. A service recommendation method based on decoupled representation learning and graph neural networks, characterized in that, The process includes the following steps: S1. Collecting the user's complete historical interaction sequence and corresponding category sequence. The complete historical interaction sequence refers to the user's interaction records with services, and the corresponding category sequence refers to the sequence constructed from the category information corresponding to each service; S2. Inputting the set of services and the set of categories into an embedding layer to obtain the embedded feature representation of each service and its corresponding category. The embedding layer refers to obtaining multi-dimensional feature vectors to represent each service or category through a one-hot encoder and a multilayer perceptron. The embedded feature representation obtained by the embedding layer is defined as two feature matrices: X i ∈R n×d For the service feature matrix, X c ∈R k×d Let be the category feature matrix, where n represents the number of services, k represents the number of categories, and d represents the feature dimension; S3. Construct a long-term and short-term interest modeling network based on user interaction sequences, and a category-based user conformity interest modeling module. S3-1. Construct a hypergraph based on service sequences in the complete historical interaction sequence, and capture users' long-term interests after passing through hypergraph convolutional layers. S3-2. Capture users' short-term interests based on service sequences in the most recent historical interaction sequence. S3-3. Introduce service category information as a data basis to capture users' conformity interests generated by category popularity. S4. Based on step S3, the learned user long-term interests h are respectively... lo Short-term interest h sh and herd interest co Feature fusion is performed through a multi-interest aggregation layer to obtain the final user interest representation h. The fusion method is as follows: h = W[h lo ||h sh ||h co [+b] where W and b represent the weight matrix and bias term, respectively, and [·||·||·] represents the concatenation of feature vectors along the dimension; S5. Based on the user interest h learned in step S4, perform the service recommendation task. For the candidate service set I, the recommendation score z of each service is determined by the service feature vector x∈X. i The inner product of z and the user interest feature vector h is obtained, i.e., z i =x T h, use the softmax function to normalize z to obtain the final service score. Then, the top-N services with the highest scores are selected as candidate services in the recommendation list. The objective function for the recommendation task is represented by cross-entropy loss: in It refers to the probability that the target service will generate an interaction given a sequence of interaction records T.

2. The service recommendation method based on decoupled representation learning and graph neural networks according to claim 1, characterized in that, The complete historical interaction sequence and corresponding category sequence of the user are based on the user interaction logs recorded by the online application, and the interaction records include clicks, favorites, and ratings.

3. The service recommendation method based on decoupled representation learning and graph neural networks according to claim 1, characterized in that, Let I be the set of all services, with n services; let C be the set of all service categories, with k categories; let the historical interaction sequence of user u be T = t1, t2, ..., t l-1 , t l , t l+1 , where t i The record representing the i-th interaction contains a tuple of the service and its corresponding category, i.e., t. i =(I i C i ), l represents the sequence length, t l+ This indicates the next interaction to be predicted.

4. The service recommendation method based on decoupled representation learning and graph neural networks according to claim 3, characterized in that, The specific steps of S3-1 are as follows: S3-1-1, Define the service hypergraph as GI∈(V i E i V i and E i These are the vertex set and the hyperedge set, respectively, and the hypergraph adjacency matrix H. i ∈{0,1} m×n Hyperedges are constructed based on the service sequences in a user's historical behavior sequence. A hyperedge is a special edge that can connect multiple vertices. For m users, m hyperedges are constructed. The adjacency matrix H of the hypergraph corresponding to the services included in the service sequence is set to 1, otherwise it is 0. S3-1-2, Define the vertex degree matrix D of the hypergraph. iv ∈R n×n and hypermarginality matrix D ie ∈R m×m Vertex degree matrix D iv ∈R n×n and hypermarginality matrix D ie ∈R m×m These are diagonal matrices relating vertex degree and hyperedge degree, respectively, where vertex v∈V. i degree Hyperedge e∈E i degree S3-1-3. Define a hypergraph convolutional neural network for feature learning and information propagation. The computation method between adjacent network layers is as follows: in W represents the service feature representation corresponding to the service sequence. i Let represent the weight matrix, σ(·) represent the nonlinear activation function, and then use the average pooling strategy to obtain the feature representation of the service: Where L represents the number of convolutional network layers, and finally, the user's long-term interest h is obtained by aggregating the feature information in the sequence using an averaging strategy. lo : Where l represents the length of the user's historical interaction sequence.

5. The service recommendation method based on decoupled representation learning and graph neural networks according to claim 4, characterized in that, The specific method of step S3-2 is as follows: S3-2-1, For the complete user interaction sequence T = t1, t2, ..., t l-1 , t l Select the nearest interactive subsequence T s =t l-a , ..., t l-1 , t l 'a' represents the number of interaction records selected. Then, the service feature sequences corresponding to the subsequences are passed through a bidirectional LSTM encoding layer to combine temporal information and capture the user's short-term interest representation. in The initial service feature representation is represented by H, and the feature matrix of m users' short-term interest representations is f, where f represents the feature aggregation method; S3-2-2, a user graph GU is constructed with users as vertices. The condition for generating edges between vertices is that users have the same recent interaction service. Graph representation learning is used to capture the potential connections between different users. The specific propagation method is as follows: H (l+1) =σ(D -1 AH l Where D and A are the degree matrix and identity matrix of the user graph GU, respectively, and σ(·) is the activation function. The short-term user interest representation h is obtained using the average pooling strategy. sh ∈H:

6. The service recommendation method based on decoupled representation learning and graph neural networks according to claim 5, characterized in that, The specific method for step S3-3 is as follows: Construct a category hypergraph GC∈(V) using the category sequence from the complete historical interaction sequence. c E c V c and E c These are the vertex set and the hyperedge set, respectively, and the hypergraph adjacency matrix H. c ∈{0,1} m×k Define the categorical hypergraph convolution calculation: in D represents the category feature representation corresponding to the category sequence. cv ∈R k×k and D ce ∈R m×m Let W represent the vertex degree matrix and hyperedge degree matrix of the categorical hypergraph, respectively. c Let σ(·) represent the weight matrix, and let σ(·) represent the non-linear activation function. After passing through the average pooling layer, the user's conformity interest h is obtained. co .

Images