CN116186385A - Service recommendation method based on decoupling characterization learning and graph neural network - Google Patents

Abstract

The invention discloses a service recommendation method based on decoupling characterization learning and a graphic neural network, which takes a user interaction log recorded by online application as a data source to acquire a complete historical interaction sequence and a corresponding category sequence of a user, firstly, respectively capturing long-term stable interests and short-term dynamic interests of the user in a personalized degree, and then capturing popular interests of the user by utilizing service category information. The method aims at the problems that the implicit decoupling-based method relies on the number of potential interests set manually, is difficult to adapt to different application scenes, and does not consider different online application service class factors, long-term interests, short-term interests and popular interests of users are learned based on complete sequences, nearest sequences and class sequences respectively, the defect that the number of potential interests is set manually is avoided, the influence of different online application class factors is considered, the popular interests of the users are captured as the basis of the interpretation of recommendation results, and application providers are helped to adjust recommendation strategies in a targeted manner.

Description

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

Technical Field

The present invention belongs to the technical field of recommendation systems, and specifically relates to a service recommendation method based on decoupled representation learning and graph neural networks.

Background Art

In recent years, with the rapid popularization of Internet technology, people's daily life has an increasing demand for online application services (such as online music, video streaming, e-shopping, etc.). Recommendation systems use users' historical behavior records to model user portraits, helping online application providers to recommend services that users may be interested in in a targeted manner, which has great potential economic value. For example, NetEase Cloud Music will generate user portraits based on users' listening records, and then recommend potential songs with member logos to users to attract users to upgrade their membership; Tmall shopping platform will recommend similar products based on users' purchase, search and other behaviors to help users choose products, thereby improving purchase conversion rates.

With the deepening of research on deep learning technology, many user preference models based on user historical behavior sequences currently focus on the relationship between users and interactive services, such as capturing users' long-term and short-term interests through sequence patterns, capturing the second-order and high-order connections between users and services through the establishment of graph structures, and learning users' overall interest characteristics. 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 popular interests that conform to the crowd. Different interests will have different degrees of influence on users' future behavior.

In order to capture users' potential interests in different aspects through their historical behavior records, many methods adopt implicit decoupling. First, the number of potential interests is set as the premise of decoupling representation learning, and then feature learning methods such as graph neural networks are used to learn different aspects of potential interests, and finally the overall preference of users is obtained. The shortcomings of existing methods are that they do not analyze the main distribution of potential interests as a whole, making it difficult to adjust the number of potential interests according to actual application recommendation scenarios, and ignoring the unique characteristics of different online services (such as category factors), which affects the economic benefits of online application providers.

Summary of the invention

Aiming at the problem that the implicit decoupled representation learning-based method relies on the number of potential interests set manually, is difficult to adapt to different service scenarios, and does not consider the factors of different online application categories, this paper proposes a service recommendation method based on decoupled representation learning and graph neural network. First, the user's long-term stable interests and short-term dynamic interests are captured in terms of personalization, and then the user's herd interests are captured using service category information.

The specific steps of the present invention are:

Step (1). Using the user interaction log recorded by the online application as the data source, collect the user's complete historical interaction sequence and the corresponding category sequence. The complete historical interaction sequence refers to the interaction record that the user has generated for the service (such as clicks, favorites, ratings, etc.), and the corresponding category sequence refers to the sequence constructed by the category information corresponding to each service. Define I as the set of all services, the number of services is n; define C as the set of categories corresponding to all services, the number of categories is k; define the historical interaction sequence of user u as T = t ₁ , t ₂ , ..., t _l-1 , t _l , t _l+1 , where _ti represents the record of the i-th interaction, including a binary tuple of the service and the corresponding category, that is, _ti = (I _i , C _i ), l represents the sequence length, and t _l+1 represents the next interaction to be predicted.

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 the corresponding category. The embedding layer refers to a multi-dimensional feature vector obtained by a one-hot encoder and a multi-layer perceptron to represent each service or category. The embedded feature representation obtained by the embedding layer is defined as two feature matrices: _Xi∈Rn ^×d is the service feature matrix, and _Xc∈Rk ^×d is the category feature matrix, where n represents the number of services, k represents the number of categories, and d represents the feature dimension.

Step (3). In order to introduce category factors to decouple user interests, a long-term interest and short-term interest modeling network based on user interaction sequences and a category-based user herd interest modeling module are designed, which includes the following three steps:

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

1) Define the service hypergraph as GI∈(V _i ,E _i ), where V _i and E _i are vertex sets and hyperedge sets respectively, and the hypergraph adjacency matrix H _i ∈{0,1} ^m×n . Hyperedges are established with the service sequence in a user's historical behavior sequence (hyperedges are special edges that can connect multiple vertices). m users will establish m hyperedges, and the hypergraph adjacency matrix H _i corresponding to the service included in the service sequence will be set to 1, otherwise it will be 0. For example, if the first user has an interaction record with the first service, then H _i (0,0)＝1.

超边e∈E_i的度

2) Define the vertex degree matrix _Div∈Rn ^×n and the hyperedge degree matrix _Die∈Rm ^×m of the hypergraph, which are diagonal matrices of vertex degree and hyperedge degree respectively, where the degree of vertex _v∈Vi

The degree of hyperedge e∈E _i

其中

表示服务序列对应的服务特征表示，W_i∈R^m×m表示权重矩阵，σ(·)表示非线性激活函数。然后使用平均池化策略得到服务的特征表示：3) Define a hypergraph convolutional neural network for feature learning and information propagation. The calculation method between adjacent network layers is

in

represents the service feature representation corresponding to the service sequence, W _i ∈ R ^m×m represents the weight matrix, and σ(·) represents the nonlinear activation function. Then the average pooling strategy is used to obtain the feature representation of the service:

Where L represents the number of convolutional network layers, and finally the feature information in the sequence is aggregated using the average strategy to obtain the user's long-term interest h _lo :

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

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

其中

表示初始服务特征表示。m个用户短期兴趣表示的特征矩阵为H，f表特征聚合方法。1) For the user's complete interaction sequence T = t ₁ , t ₂ , ..., t _l-1 , t _l , select the most recent interaction subsequence T ^s = t _la , ..., t _l-1 , t _l , where a represents the number of selected interaction records, which can be set according to different application scenarios. Then the service feature sequence corresponding to the subsequence is passed through a bidirectional LSTM encoding layer to combine time information and capture the user's short-term interest representation.

in

represents the initial service feature representation. The feature matrix of m users’ short-term interest representation is H, and f represents the feature aggregation method.

2) A user graph GU is established with users as vertices. The condition for edge connection between vertices is that there are the same recent interactive services between users. 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 average pooling strategy is also used to obtain the user's short-term interest representation h _sh ∈H:

其中

表示类别序列对应的类别特征表示，D_cv∈R^k×k和D_ce∈R^m×m分别表示类别超图的顶点度矩阵和超边度矩阵，W_c∈R^m×m表示权重矩阵，σ(·)表示非线性激活函数，通过平均池化层后得到用户的从众兴趣h_co。Step (3.3). In order to adapt to the data characteristics of different applications, this method introduces service category information as the data basis to capture the user's herd interest in category popularity. Similar to step (3.1), the category sequence in the complete historical interaction sequence is used to construct a category hypergraph GC∈(V _c ,E _c ), where V _c and E _c are the vertex set and hyperedge set respectively, and the hypergraph adjacency matrix H _c ∈{0,1} ^m×k . Define the category hypergraph convolution calculation:

in

represents the category feature representation corresponding to the category sequence, D _cv ∈R ^k×k and D _ce ∈R ^m×m represent the vertex degree matrix and hyperedge degree matrix of the category hypergraph respectively, W _c ∈R ^m×m represents the weight matrix, σ(·) represents the nonlinear activation function, and the user’s herd interest h _co is obtained after the average pooling layer.

Step (4). Based on step (3), the learned user long-term interest h _lo , short-term interest h _sh and herd interest h _co are respectively fused through the 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 in dimension.

然后选择评分最高的top-N个服务作为推荐列表中的候选服务。推荐任务学习的目标函数表示为交叉熵损失：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 of each service is obtained by the inner product operation of the service feature vector x∈X _i and the user interest feature vector h, that is, z _i = x ^T h. The softmax function is used 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 of recommendation task learning is expressed as cross entropy loss:

是指目标服务在给定交互记录序列T情况下会产生交互的概率。in

It refers to the probability that the target service will interact with a given interaction record sequence T.

Compared with the prior art, the present invention has the following beneficial effects: by utilizing the decoupled learning of user interests, the user's long-term interests, short-term interests and herd interests can be learned separately, avoiding the disadvantages of manually setting the number of potential interests to the greatest extent, and considering the influence of different online application category factors, capturing the user's herd interests as the basis for the interpretability of recommendation results, helping application providers to adjust recommendation strategies in a targeted manner.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 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.

FIG. 2 is a diagram of a feature embedding module based on a recent sequence in an embodiment of the present invention.

DETAILED DESCRIPTION

It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

The technical solution of the present invention is further explained and illustrated below in conjunction with the accompanying drawings.

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

Step (1). Using the user interaction log recorded by the online application as the data source, collect the user's complete historical interaction sequence and the corresponding category sequence. The complete historical interaction sequence refers to the interaction record that the user has generated for the service (such as clicks, favorites, ratings, etc.), and the corresponding category sequence refers to the sequence constructed by the category information corresponding to each service. Define I as the set of all services, the number of services is n; define C as the set of categories corresponding to all services, the number of categories is k; define the historical interaction sequence of user u as T = t ₁ , t ₂ , ..., t _l-1 , t _l , t _l+1 , where _ti represents the record of the i-th interaction, including a binary tuple of the service and the corresponding category, that is, _ti = (I _i , C _i ), l represents the sequence length, and t _l+1 represents the next interaction to be predicted.

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 the corresponding category. The embedding layer refers to a multi-dimensional feature vector obtained by a one-hot encoder and a multi-layer perceptron to represent each service or category. The embedded feature representation obtained by the embedding layer is defined as two feature matrices: _Xi∈Rn ^×d is the service feature matrix, and _Xc∈Rk ^×d is the category feature matrix, where n and k represent the number of services and categories respectively, and d represents the feature dimension.

Step (3). In order to introduce category factors to decouple user interests, a long-term interest and short-term interest modeling network based on user interaction sequences and a category-based user herd interest modeling module are designed, which includes the following three steps:

Step (3.1). First, a hypergraph is constructed based on the service sequence T _i =I ₁ ,I ₂ ,…,I _l-1 ,I _l in the complete historical interaction sequence. After the hypergraph convolution layer, the long-term interests of users are captured and represented as h _lo .

1) Define the service hypergraph as GI∈(V _i ,E _i ), where V _i and E _i are vertex sets and hyperedge sets respectively, and the hypergraph adjacency matrix H _i ∈{0,1} ^m×n . Hyperedges are established with the service sequence in a user's historical behavior sequence (hyperedges refer to special edges that can connect multiple vertices). m users will establish m hyperedges, and the hypergraph adjacency matrix Hi corresponding to the service included in the service sequence will be set to 1, otherwise it will be 0. If the first user generates an interaction record with the first service, then H _i (0,0)＝1. For example, if the service sequences in the historical interaction records of two users u ₁ and u ₂ are {I ₁ ,I ₂ ,I ₂ ,I ₃ } and {I ₂ ,I ₄ ,I ₂ ,I ₅ } respectively, then the corresponding service hypergraph adjacency matrix is:

超边e∈E_i的度

2) Define the vertex degree matrix _Div∈Rn ^×n and the hyperedge degree matrix _Die∈Rm ^×m of the hypergraph, which are diagonal matrices of vertex degree and hyperedge degree respectively, where the degree of vertex _v∈Vi

The degree of hyperedge e∈E _i

其中

表示服务序列对应的服务特征表示，W_i∈R^m×m表示权重矩阵，σ(·)表示非线性激活函数。然后使用平均池化策略得到服务的特征表示：3) Define a hypergraph convolutional neural network for feature learning and information propagation. The calculation method between adjacent network layers is

in

represents the service feature representation corresponding to the service sequence, W _i ∈ R ^m×m represents the weight matrix, and σ(·) represents the nonlinear activation function. Then the average pooling strategy is used to obtain the feature representation of the service:

表示第l层的服务特征矩阵，最后使用平均策略聚合序列中的特征信息得到用户长期兴趣h_lo：Where L represents the number of convolutional network layers,

represents the service feature matrix of the lth layer, and finally uses the average strategy to aggregate the feature information in the sequence to obtain the user's long-term interest h _lo :

Where l represents the length of the user's historical interaction sequence, and Xi _,t represents the feature representation of the tth service in the service sequence.

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

其中

表示初始服务特征表示。m个用户短期兴趣表示的特征矩阵为H∈R^m×d，f表示双向LSTM特征聚合方法。1) For the user's complete interaction sequence T = t ₁ , t ₂ , ..., t _l-1 , t _l , select the most recent interaction subsequence T ^s = t _la , ..., t _l-1 , t _l , where a represents the number of selected interaction records, which can be set according to different application scenarios. Then the service feature sequence corresponding to the subsequence is passed through a bidirectional LSTM encoding layer to combine time information and capture the user's short-term interest representation.

in

represents the initial service feature representation. The feature matrix of m users’ short-term interest representation is H∈R ^m×d , and f represents the bidirectional LSTM feature aggregation method.

2) Build a user graph GU with users as vertices. The condition for edge connections between vertices is that there are the same recent interactive services between users. For example, the recent service sequences of users u ₁ and u ₂ are {I ₂ ,I ₃ } and {I ₂ ,I ₅ } respectively. There is the same service I ₂ , so users u ₁ and u ₂ will have edge connections. 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∈R ^m×m and A∈R ^m×m are the degree matrix and identity matrix of the user graph GU respectively, and σ(·) is the activation function. The average pooling strategy is also used to obtain the user short-term interest representation h _sh ∈H:

表示第l层的用户短期兴趣特征表示。Where L represents the number of layers of the user graph convolutional neural network,

Represents the user's short-term interest feature representation at the lth layer.

Step (3.3). In order to adapt to the data characteristics of different applications, this method introduces service category information as the data basis to capture the user's herd interest h _co generated by category popularity. Similar to step (3.1), the category sequence T _c =C ₁ ,C ₂ ,…,C _l-1 ,C _l in the complete historical interaction sequence is used to construct a category hypergraph GC∈(V _c ,E _c ), where V _c and E _c are vertex sets and hyperedge sets respectively, and the hypergraph adjacency matrix H _c ∈{0,1} ^m×k . For example, there are two users u ₁ and u ₂ whose category sequences in their historical interaction records are {C ₁ ,C ₁ ,C ₁ ,C ₂ } and {C ₂ ,C ₁ ,C ₂ ,C ₃ } respectively, then the corresponding category hypergraph adjacency matrix is:

其中

表示类别序列对应的类别特征表示，D_cv∈R^k×k和D_ce∈R^m×m分别表示类别超图的顶点度矩阵和超边度矩阵，W_c∈R^m×m表示权重矩阵，σ(·)表示非线性激活函数，通过平均池化层后得到用户的从众兴趣h_co。Define category hypergraph convolution calculation:

in

represents the category feature representation corresponding to the category sequence, D _cv ∈R ^k×k and D _ce ∈R ^m×m represent the vertex degree matrix and hyperedge degree matrix of the category hypergraph respectively, W _c ∈R ^m×m represents the weight matrix, σ(·) represents the nonlinear activation function, and the user’s herd interest h _co is obtained after the average pooling layer.

Step (4). Based on step (3), the learned user long-term interest h _lo , short-term interest h _sh and herd interest h _co are respectively fused through the 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∈Rd ^×3d and b∈Rd ^×1 represent the weight matrix and bias term respectively, and [·||·||·] represents the concatenation of feature vectors in dimension.

然后选择评分最高的top-N个服务作为推荐列表中的候选服务。推荐任务学习的目标函数表示为交叉熵损失：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 of each service is obtained by the inner product operation of the service feature vector x∈X _i and the user interest feature vector h, that is, z _i = x ^T h. The softmax function is used 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 of recommendation task learning is expressed as cross entropy loss:

是指目标服务在给定交互记录序列T情况下会产生交互的概率。在结果的可解释性方面，如用户u₁的服务序列和类别序列分别为{I₁,I₂,I₃}和{C₁,C₁,C₂}，若用户u₁实际上下一个交互的物品是I₄(类别是C₁)，仅根据服务历史无法对结果进行解释，但是从类别数据上可以看出用户u₁频繁交互类别为C₁的服务，则可以侧重于用户从众兴趣对推荐结果进行解释。in

It refers to the probability that the target service will interact with a given interaction record sequence T. In terms of the interpretability of the results, if the service sequence and category sequence of user u ₁ are {I ₁ , I ₂ , I ₃ } and {C ₁ , C ₁ , C ₂ } respectively, if the next item that user u ₁ actually interacts with is I ₄ (category C ₁ ), the result cannot be explained based on the service history alone. However, from the category data, it can be seen that user u ₁ frequently interacts with services of category C ₁ , so the recommendation results can be explained by focusing on the user's herd interest.

This method is based on the user's historical interaction record data, and captures the user's long-term and short-term interests respectively based on decoupled representation learning. Then, the service category popularity factor is introduced to construct a category hypergraph convolutional neural network to capture the user's 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 (6)

1. A service recommendation method based on decoupled representation learning and graph neural network, characterized by comprising the following steps: S1. Collect the user's complete historical interaction sequence and corresponding category sequence, where the complete historical interaction sequence refers to the user's interaction record with the service, and the corresponding category sequence refers to the sequence constructed by the category information corresponding to each service; S2. Input the set of services and the set of categories into the embedding layer to obtain the embedded feature representation of each service and the corresponding category. The embedding layer refers to obtaining a multi-dimensional feature vector through a one-hot encoder and a multi-layer perceptron to represent each service or category. The embedded feature representation obtained by the embedding layer is defined as two feature matrices: _Xi∈Rn ^×d is the service feature matrix, and _Xc∈Rk ^×d is 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 interest and short-term interest modeling network based on user interaction sequences, as well as a category-based user herd interest modeling module S3-1, build a hypergraph based on the service sequence in the complete historical interaction sequence, and capture the user's long-term interests after the hypergraph convolution layer; S3-2, capturing users’ short-term interests based on service sequences in recent historical interaction sequences; S3-3, introduce service category information as the data basis to capture the user's herd interest in category popularity; S4. Based on step S3, the learned user long-term interest h _lo , short-term interest h _sh and herd interest h _co are respectively fused through the 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 in 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 obtained by the inner product operation of the service feature vector x∈X _i and the user interest feature vector h, that is, 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 of recommendation task learning is expressed as cross entropy loss:

in

It refers to the probability that the target service will interact with a given interaction record sequence T. 2. According to claim 1, the service recommendation method based on decoupled representation learning and graph neural network is characterized in that the user's complete historical interaction sequence and the corresponding category sequence are based on the user interaction log recorded by the online application as the data source, and the interaction record includes clicks, favorites, and ratings. 3. According to the service recommendation method based on decoupled representation learning and graph neural network in claim 1, it is characterized in that I is defined as the set of all services, the number of services is n; C is defined as the set of categories corresponding to all services, the number of categories is k; the historical interaction sequence of user u is defined as T = t ₁ , t ₂ , ..., t _l-1 , t _l , t _l+1 , where _ti represents the record of the i-th interaction, including a tuple of the service and the corresponding category, that is, _ti = (I _i , C _i ), l represents the sequence length, and t _l+ represents the next interaction to be predicted. 4. The service recommendation method based on decoupled representation learning and graph neural network according to claim 3 is characterized in that the step S3-1 is specifically as follows: S3-1-1. Define the service hypergraph as GI∈(V _i ,E _i ), where V _i and E _i are vertex sets and hyperedge sets respectively, and the hypergraph adjacency matrix H _i ∈{0,1} ^m×n . Hyperedges are established with the service sequence in a user's historical behavior sequence. The hyperedge refers to a special edge that can connect multiple vertices. m users will establish m hyperedges. The position of the hypergraph adjacency matrix H corresponding to the service included in the service sequence will be set to 1, otherwise it will be 0. S3-1-2. Define the vertex degree matrix _Div∈Rn ^×n and the hyperedge degree matrix Die∈Rm _× ^m of the hypergraph. The vertex degree matrix _Div∈Rn ^× n and the hyperedge degree matrix _Die∈Rm ^×m are diagonal matrices of vertex degree and hyperedge degree, respectively. The degree of vertex _v∈Vi is

The degree of hyperedge e∈E _i

S3-1-3. Define the hypergraph convolutional neural network for feature learning and information propagation. The calculation method between adjacent network layers is:

in

represents the service feature representation corresponding to the service sequence, _Wi represents the weight matrix, σ(·) represents the nonlinear activation function, and then the average pooling strategy is used to obtain the feature representation of the service:

Where L represents the number of convolutional network layers, and finally the feature information in the sequence is aggregated using the average strategy to obtain the user's long-term interest h _lo :

Where l represents the length of the user's historical interaction sequence. 5. According to the service recommendation method based on decoupled representation learning and graph neural network according to claim 4, it is characterized in that the specific method of step S3-2 is as follows: S3-2-1. For the complete user interaction sequence T = t ₁ , t ₂ , ..., t _l-1 , t _l , select the most recent interaction subsequence T ^s = t _la , ..., t _l-1 , t _l , where a represents the number of selected interaction records. Then, the service feature sequence corresponding to the subsequence is passed through a bidirectional LSTM encoding layer to combine time information and capture the user's short-term interest representation.

in

represents the initial service feature representation, the feature matrix of m users’ short-term interest representation is H, and f represents the feature aggregation method; S3-2-2. Build a user graph GU with users as vertices. The condition for edge connections between vertices is that there are the same recent interactive services between users. Use graph representation learning 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, σ(·) is the activation function, and the average pooling strategy is also used to obtain the user's short-term interest representation h _sh ∈H:

6. The service recommendation method based on decoupled representation learning and graph neural network according to claim 5, characterized in that the specific method of step S3-3 is as follows: Construct a category hypergraph GC∈( _Vc , _Ec ) with the category sequence in the complete historical interaction sequence, where _Vc and _Ec are the vertex set and hyperedge set respectively, and the hypergraph adjacency matrix _Hc∈ {0,1} ^m×k . Define the category hypergraph convolution calculation:

in

represents the category feature representation corresponding to the category sequence, D _cv ∈R ^k×k and D _ce ∈R ^m×m represent the vertex degree matrix and edge degree matrix of the category hypergraph respectively, W _c represents the weight matrix, σ(·) represents the nonlinear activation function, and the user’s herd interest h _co is obtained after the average pooling layer.

Images