CN114186139A - Graph neural network session recommendation method based on time enhancement - Google Patents

Abstract

The invention relates to the technical field of internet big data, in particular to a graph neural network session recommendation method based on time enhancement, which inputs target sessions into a trained time enhancement graph neural network model; the time-enhanced graph neural network model generates user interest drift degrees through time intervals of conversation item conversion, constructs a time-enhanced conversation graph capable of correspondingly processing conversion relations among conversation items according to the user interest drift degrees, learns item embedding and generates new conversation representation based on the time-enhanced conversation graph, and calculates probability distribution of candidate items based on the new conversation representation to complete conversation recommendation. The conversation recommendation method can improve the project embedding quality based on the user interest drift degree, so that the conversation recommendation accuracy can be improved.

Description

A Conversational Recommendation Method Based on Temporal Augmentation in Graph Neural Networks

technical field

The invention relates to the technical field of Internet big data, in particular to a time-enhanced graph neural network session recommendation method.

Background technique

Session-based recommendation is a recommendation mode for anonymous users or users who are not logged in, which plays an important role in today's major e-commerce platforms (Taobao, JD, etc.) or streaming media platforms (Douyin, YouTube, etc.). In actual scenarios, only short-term historical interactions of users can be obtained at certain times, such as new users or users who have not logged in. At this time, recommendation algorithms that rely on long-term historical interactions of users will be limited in their performance in conversational recommendation, such as methods based on collaborative filtering or Markov chains. Therefore, session-based recommendation has become a research hotspot, and its goal is to recommend the next item (or item) of interest to the user based on the user's behavior sequence in the session.

Aiming at the problem that the item recommendation accuracy of the existing conversation recommendation method is not high, the Chinese Patent Publication No. CN112035746A discloses "A Conversation Recommendation Method Based on Spatio-temporal Sequence Graph Convolutional Network", which includes: modeling all conversation sequences It is a directed session graph; build a global graph with the common commodities in the session as links; embed the ARMA filter into the gated graph neural network to extract the time-varying topology graph signal in the graph model, and obtain the graphs involved in the session graph. The feature vector of each node of the node; use the attention mechanism to obtain the global preference information from the user's historical sessions; obtain the user's local preference information from the last session clicked by the user, and combine the global preference information to obtain the user's final preference information; predict each The probability that the next clicked item in the session may appear, and the Top-K recommended item is given.

The conversational recommendation methods in the above existing schemes capture rich conversational representations (contextual relations) from a global graph, learn users' global and local preferences through an attention mechanism, and then provide accurate item predictions. However, the existing conversational recommendation methods have limited efforts in exploring the change of user interest in the transition relationship between items. They generally treat each transition relationship between items equally, ignoring the rich user interest drift information in the transition relationship. In turn, only sequential transitions between consecutive actions can be captured, but complex transitions between non-adjacent actions cannot be effectively modeled, resulting in low quality of item embeddings, resulting in still low accuracy of session recommendation. Therefore, how to design a conversational recommendation method that can improve the quality of item embedding based on the degree of user interest drift is an urgent technical problem to be solved.

SUMMARY OF THE INVENTION

In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is: how to provide a graph neural network session recommendation method based on time enhancement, so as to improve the quality of item embedding based on the degree of user interest drift, thereby improving the accuracy of session recommendation sex.

In order to solve the above-mentioned technical problems, the present invention adopts the following technical solutions:

A time-enhanced graph neural network session recommendation method, wherein a target session is input into a trained time-enhanced graph neural network model; the time-enhanced graph neural network model generates a user interest drift degree through the time interval when the conversation item conversion occurs, And construct a time-enhanced session graph that can handle the conversion relationship between conversation items according to the degree of user interest drift, then learn item embedding based on the time-enhanced session graph and generate a new session representation, and finally calculate the probability distribution of candidate items based on the new session representation to complete Session recommendation.

Preferably, the time-enhanced graph neural network model completes the session recommendation through the following steps:

S1: map the target session S=(v ₁ ,v ₂ ,...,v _n ) to the session embedding sequence H=(h ₁ ,h ₂ ,...,h _n );

S2: Based on the session graph of the target session, the degree of user interest drift is mapped to the weights of the edges of the session graph to generate the corresponding time-enhanced session graph;

S3: Learn item representations of temporally augmented conversational graphs and aggregate high-order neighbor information of items based on temporally augmented conversational graphs and conversational embeddings through a multi-layered temporal graph convolutional network to generate corresponding conversational item representations;

S4: Allocate user interest boxes for each session item in the target session through the temporal interest attention network, and generate a user interest box sequence representation; then aggregate the user interest box sequence representation and the session item representation to generate a new session representation;

S5: Calculate the long-term interest representation and the short-term interest representation based on the new session representation, and fuse to obtain the final session representation; then select candidate items based on the final session representation, and calculate the probability distribution of the candidate items to complete the session recommendation.

Preferably, in step S2, the time-enhanced session graph is defined as: for the target session S=(v ₁ ,v ₂ ,...,v _n ), the time-enhanced session graph is g _s =(V _s ,ε _s ,W _s ), where ε _S represents the set of edges, and W _S represents the weight matrix of the edges;

Each node v _i ∈ V _S and edge (v _i-1 ,v _i )∈ε _s represents the adjacency relationship between two consecutive items v _i-1 and v _i , and its matrix expression is the in-edge matrix A ^I and out- Edge matrix A ^O ; each edge (v _i-1 ,v _i ) corresponds to a weight W _i-1,i ∈W _S ; each node v _i ∈ V _S adds a self-connected edge, and its matrix is expressed as Self-connection matrix A ^S .

Preferably, in step S3, the conversation item representation is generated by the following steps:

S301: Layer l output embedding representation through a multilayer temporal graph convolutional network

S302: Use the embedded representation h _i ^L output by the multi-layer temporal graph convolutional network as the embedded representation of the conversation item v _i ∈ S;

与其初始嵌入

的表示进行合并得到会话项目v_i∈S的项目表示

并生成会话项目表示

S303: Embedding representation of the output of a multi-layer temporal graph convolutional network via a highway network

with its initial embedding

to merge the representations of the conversation items to obtain the item representation of the conversation item v _i ∈ S

and generate the session item representation

计算嵌入表示

Preferably, by formula

Computational embedded representation

计算项目表示

by formula

Calculated item representation

in,

和

分别表示时间增强会话图入边矩阵A^I、入边矩阵A^O和自连接矩阵A^S的第i行；l、L均表示多层时间图卷积网络的层数；

表示可训练参数；σ表示函数Sigmoid。In the above formula:

and

respectively represent the ith row of the time-enhanced session graph input edge matrix A ^I , the input edge matrix A ^O and the self-connection matrix A ^S ; l and L both represent the number of layers of the multi-layer time graph convolutional network;

represents the trainable parameters; σ represents the function Sigmoid.

Preferably, in step S4, a new session representation is generated through the following steps:

S401: Calculate the time interval sequence Q=(q ₁ , q ₂ , the distance between each item and the last item) based on the click timestamp T=(t ₁ , t ₂ , . . . , t _n ) of each session item in the target session S ...,q _n );

S402: Map the time interval sequence Q to an interest-sensitive sequence Γ=(γ ₁ ,γ ₂ ,...,γ _n ), and calculate the adaptive time span μ;

S403: Assign user interest bins bin _i to each session item in the target session S based on the adaptive time span μ through the temporal interest attention network, and generate a user interest bin sequence B=(bin ₁ , bin ₂ ,..., bin _n ) ;

S404: Map each user interest box bin _i in the user interest box sequence B=(bin ₁ , bin ₂ , . . . , bin _n ) to a low-dimensional dense vector e _i , and generate a corresponding user interest box representation E=(e ₁ ,e ₂ ,…, _en );

S405: Perform asymmetric processing on the forward context information and backward context information represented by the user interest box through an asymmetric gated recurrent neural network, and generate a corresponding enhanced representation of the user interest box sequence

和会话项目表示

生成捕获了项目顺序结构信息和用户时间兴趣特征的新会话表示C＝(c₁,c₂,…,c_n)。S406: Aggregate user interest box sequences to enhance representation through attention mechanism

and session item representation

A new session representation C=(c ₁ , c ₂ , . . . , c _n ) is generated that captures item order structure information and user temporal interest features.

计算兴趣敏感序列Γ＝(γ₁,γ₂,…,γ_n)；Preferably, by formula

Calculate the interest-sensitive sequence Γ=(γ ₁ ,γ ₂ ,...,γ _n );

m＝t_n-t₁；in,

m=t _n -t ₁ ;

计算自适应时间跨度μ；by formula

Calculate the adaptive time span μ;

The user interest bin bin _i is calculated by the formula bin _i =k,whereγ _i ∈(μ×(k-1),μ×k];

The asymmetric gated recurrent neural network generates the enhanced representation of the sequence of user interest bins by the following formula

计算新会话表示C＝(c₁,c₂,…,c_n)；by formula

Calculate the new session representation C=(c ₁ ,c ₂ ,..., _cn );

in,

和l_p分别表示衰减常数和左偏移量；D_init和D_final表示两个预先指定的常数设置D_init＝0.98，D_final＝0.01；e表示自然常数；N表示用户兴趣箱的个数；μ表示目标会话S所需要的自适应时间间隔；k∈(1,2,…,N)；

表示可训练参数。In the above formula:

and l _p represent the decay constant and left offset, respectively; D _init and D _final represent two pre-specified constant settings D _init = 0.98, D _final = 0.01; e represents a natural constant; N represents the number of user interest boxes; μ denotes the adaptive time interval required by the target session S; k∈(1,2,…,N);

Represents trainable parameters.

Preferably, in step S5, the probability distribution of the candidate items is generated by the following steps:

S501: Based on the new session representation C=(c ₁ , c ₂ , . . . , c _n ) combined with addition and pooling, generate a long-term interest representation z _long ;

S502: Obtain the short-term interest representation z _short from the new session representation C through the GRU;

S503: Combine the long-term interest representation z _long and the short-term interest representation z _short in combination with the gating mechanism, and generate the final session representation z _final ;

S504: Select the top K candidate items v _i ∈ V from the candidate item set V=(v ₁ ,v ₂ ,...,v _|V| ) based on the session final representation z _final for recommendation, and calculate each candidate item v _i ∈ Fraction of V;

S505: Apply a softmax function based on the scores of each candidate item v _i ∈ V to generate a probability distribution of the candidate items.

计算长期兴趣表示z_long；Preferably, by formula

Calculate the long-term interest representation z _long ;

计算短期兴趣表示z_short，其中，

代表在时间步长i-1处的会话项目表示，将最后一个会话项目表示作为会话短期表示

by formula

Calculate the short-term interest representation z _short , where,

Represents the session item representation at time step i-1, taking the last session item representation as the session short term representation

Calculate the session final representation z _final by the formula z _final =f⊙z _long +(1-f)⊙z _short ;

in,

计算候选项目v_i∈V的分数

by formula

Calculate the score of the candidate item v _i ∈ V

计算候选项目v_i∈V的概率，进而生成候选项目的概率分布；by formula

Calculate the probability of the candidate item v _i ∈ V, and then generate the probability distribution of the candidate item;

表示候选项目v_i∈V的初始化嵌入表示；

表示可训练参数。In the above formula:

represents the initialized embedding representation of candidate items v _i ∈ V;

Represents trainable parameters.

Preferably, a time-enhanced graph neural network model is trained by a time back-propagation algorithm, and cross-entropy is used as a loss function;

Among them, the cross entropy loss function is expressed as

In the above formula: y _i represents the true label.

Compared with the prior art, the session recommendation method of the present invention has the following beneficial effects:

The present invention generates the degree of user interest drift through the time interval when item conversion occurs, and correspondingly processes the conversion relationship between conversation items according to the degree of user interest drift, so that the rich user interest drift information in the conversion relationship can be paid attention to, and the difference between non-adjacent actions can be captured. It generates a new session representation that captures item sequence structure information and user temporal interest features, which can improve the quality of item embedding based on the degree of user interest drift, thereby improving the accuracy of session recommendation. At the same time, the multi-layer temporal graph convolutional network can effectively obtain the structural information of the temporally enhanced conversation graph, so that the embedding of the item can be learned accurately, and the quality of the item embedding can be better improved. Furthermore, the temporal interest attention network is able to capture the complex transition patterns of items with common user interests, enabling the modeling of items with similar user interests in the temporal dimension, which in turn can focus on local commonalities between items, generating a captured The new session representation of item order structure information and user temporal interest features can further improve the accuracy of session recommendation.

Description of drawings

In order to make the purpose, technical solutions and advantages of the invention clearer, the present invention will be described in further detail below in conjunction with the accompanying drawings, wherein:

Fig. 1 is the logical block diagram of the session recommendation method;

Fig. 2 is the network structure diagram of the time-enhanced graph neural network model;

Fig. 3 is the schematic diagram of practical example;

Fig. 4 is the experimental result graph of the change of the number of layers of the time graph convolutional network;

Figure 5 is a graph of the experimental results of changes in the number of temporal interest bins.

Detailed ways

The following is a further detailed description through specific embodiments:

Example:

This embodiment discloses a graph neural network session recommendation method based on time enhancement.

The time-enhanced graph neural network session recommendation method is based on inputting the target session into the trained time-enhanced graph neural network model; the time-enhanced graph neural network model generates the degree of user interest drift through the time interval when the conversation item transition occurs, and constructs A time-enhanced session graph that can handle the conversion relationship between conversation items according to the degree of user interest drift, then learn item embedding based on the time-enhanced session graph and generate a new session representation, and finally calculate the probability distribution of candidate items based on the new session representation to complete the session recommendation. .

As shown in Figure 1 and Figure 2, the time-enhanced graph neural network model completes the session recommendation through the following steps:

S1: map the target session S=(v ₁ ,v ₂ ,...,v _n ) to the session embedding sequence H=(h ₁ ,h ₂ ,...,h _n );

S2: Based on the session graph of the target session, the degree of user interest drift is mapped to the weights of the edges of the session graph to generate the corresponding time-enhanced session graph (TES Graph);

S3: Learn item representations of temporally augmented conversational graphs and aggregate high-order neighbor information of items based on temporally augmented conversational graphs and conversational embeddings through a multi-layered temporal graph convolutional network (T-GCN) to generate corresponding conversational item representations ;

S4: Allocate user interest boxes for each session item in the target session through the Temporal Interest Attention Network (TIAN), and generate a sequence representation of user interest boxes; then aggregate the sequence representation of user interest boxes and session item representations to generate new session representation;

S5: Calculate the long-term interest representation and the short-term interest representation based on the new session representation, and fuse to obtain the final session representation; then select candidate items based on the final session representation, and calculate the probability distribution of the candidate items to complete the session recommendation.

The present invention generates the degree of user interest drift through the time interval when item conversion occurs, and correspondingly processes the conversion relationship between conversation items according to the degree of user interest drift, so that the rich user interest drift information in the conversion relationship can be paid attention to, and the difference between non-adjacent actions can be captured. It generates a new session representation that captures item sequence structure information and user temporal interest features, which can improve the quality of item embedding based on the degree of user interest drift, thereby improving the accuracy of session recommendation. At the same time, the multi-layer temporal graph convolutional network can effectively obtain the structural information of the temporally enhanced conversation graph, so that the embedding of the item can be learned accurately, and the quality of the item embedding can be better improved. Furthermore, the temporal interest attention network is able to capture the complex transition patterns of items with common user interests, enabling the modeling of items with similar user interests in the temporal dimension, which in turn can focus on local commonalities between items, generating a captured The new session representation of item order structure information and user temporal interest features can further improve the accuracy of session recommendation.

In the specific implementation process, the time-enhanced session graph is defined as: for the target session S=(v ₁ ,v ₂ ,...,v _n ), the time-enhanced session graph is g _s =(V _s ,ε _s ,W _s ) , where ε _S represents the set of edges, W _S represents the weight matrix of the edge; each node v _i ∈ V _S and edge (vi _-1 ,vi ) _{∈ε s} represents two consecutive items vi _-1 Adjacency relationship with v _i , its matrix expression is in-edge matrix A ^I and out-edge matrix A ^O ; each edge (v _i-1 ,v _i ) corresponds to a weight Wi _-1,i ∈W _S ; Each node v _i ∈ V _S adds a self-connected edge, and its matrix is denoted as a self-connected matrix A ^S .

User interest drift is closely related to the time interval between two consecutive items in an ongoing session, i.e., the longer the time interval between two consecutive items, the greater the user interest drift, and vice versa. Combined with Figure 3, we can observe: (1) The conversion relationship between "Sweater" and "Iphone" is weaker than that between "Iphone" and "Airpods", because the user interest drift of the former is much larger than that of the latter (2) User interest drift is closely related to the time interval, that is, the larger the time interval between two consecutive items, the higher the degree of user interest drift; (3) Items within a short time interval (such as "Shirt" , "Overcoat" and "Sweater") generally have similar user interests with only a small interest shift.

In Figure 3, the session contains 5 items (shirt 8' coat 7' sweater 40 5 ' mobile phone 15' earphone), and the time interval between sweater and mobile phone is 405 minutes, which is much longer than 15 minutes for mobile phone and earphone At the same time, we can also see that there is a shift in interest from sweaters to mobile phone users. According to this observation, we calculate the weights of the edges as follows: For each session S=(v ₁ ,v ₂ , ...,v _n ), we first obtain the click time T=(t ₁ ,t ₂ ,...,t _n ) of its item, and measure the weight of each edge in the time-enhanced session graph based on the degree of user interest drift, which degree Calculated based on the time interval corresponding to the edge.

The edge weight is calculated using Newton's law of cooling. The weight τ( _i,j) of the edge between two adjacent item sessions v _i ∈ S and v _j ∈ S in the target session S is calculated as:

m=t _n -t ₁ ;

和l_p分别表示衰减常数和左偏移量，l_p用于使τ_(i,j)适应不同的会话。In the formula: D _init and D _final represent two pre-specified constants, which are used for the weights of the initial and final edges of the decay, set D _init =0.98, D _final =0.01; e represents a natural constant; t _i , t _j represent Click timestamps of item sessions v _i ∈ S and v _j ∈ S; t _n , t ₁ show click timestamps of session item v ₁ ∈ S and session item v _n ∈ S;

and _lp denote the decay constant and left offset, respectively, and _lp is used to adapt τ _(i,j) to different sessions.

The present invention maps the degree of user interest drift to the weight of the edge of the conversation graph to generate a time-enhanced conversation graph, which can realize the corresponding processing of the conversion relationship between conversation items according to the degree of user interest drift, so that the rich user interest drift information in the conversion relationship can be paid attention to. , capture the complex transformation relationship between non-adjacent actions, and generate a new session representation that captures item sequence structure information and user temporal interest features, that is, it can improve the quality of item embedding based on the degree of user interest drift, thereby improving the accuracy of session recommendation. sex.

In the specific implementation process, the session item representation is generated through the following steps:

S301: Layer l output embedding representation through a multilayer temporal graph convolutional network

作为会话项目v_i∈S的嵌入表示；S302: Embedding representation of the output of the multi-layer temporal graph convolutional network

as the embedded representation of the conversation item vi _∈ S;

与其初始嵌入

的表示进行合并得到会话项目v_i∈S的项目表示

并生成会话项目表示

S303: Embedding representation of the output of a multi-layer temporal graph convolutional network via a highway network

with its initial embedding

to merge the representations of the conversation items to obtain the item representation of the conversation item v _i ∈ S

and generate the session item representation

A single-layer temporal graph convolutional network (T-GCN) will aggregate information about the item itself and its first-order neighbors. In T-GCN, we abandon the two most commonly used mechanisms (i.e. linear transformation and nonlinear activation) in existing graph neural networks (GCNs). To capture the transfer relationship between distant items, we stack multiple layers of T-GCN to aggregate high-order neighbor information of items.

计算嵌入表示

by formula

Computational embedded representation

计算项目表示

by formula

Calculated item representation

in,

和

分别表示时间增强会话图入边矩阵A^I、入边矩阵A^O和自连接矩阵A^S的第i行；l、L均表示多层时间图卷积网络的层数；

表示可训练参数；σ表示函数Sigmoid。In the above formula:

and

respectively represent the ith row of the time-enhanced session graph input edge matrix A ^I , the input edge matrix A ^O and the self-connection matrix A ^S ; l and L both represent the number of layers of the multi-layer time graph convolutional network;

represents the trainable parameters; σ represents the function Sigmoid.

The present invention can effectively acquire the structure information of the time-enhanced session graph through the multi-layer temporal graph convolution network, so that the embedding of the item can be learned accurately, and the quality of the item embedding can be better improved. At the same time, the problem of over-smoothing of item representations can be solved by expressway networks.

In the specific implementation process, a new session representation is generated through the following steps:

S401: Calculate the time interval sequence Q=(q ₁ , q ₂ , the distance between each item and the last item) based on the click timestamp T=(t ₁ , t ₂ , . . . , t _n ) of each session item in the target session S ...,q _n );

S402: Map the time interval sequence Q to an interest-sensitive sequence Γ=(γ ₁ ,γ ₂ ,...,γ _n ), and calculate the adaptive time span μ;

S403 : Allocate user interest bins bin _i for each session item in the target session S based on the adaptive time span μ through the temporal interest attention network, and generate a user interest bin sequence B=(bin ₁ , bin ₂ , . . . , bin _n ) ;

S404: Map each user interest box bin _i in the user interest box sequence B=(bin ₁ , bin ₂ , . . . , bin _n ) to a low-dimensional dense vector e _i , and generate a corresponding user interest box representation E=(e ₁ ,e ₂ ,…, _en );

S405: Perform asymmetric processing on the forward context information and backward context information represented by the user interest box through an asymmetric gated recurrent neural network (Asym-BiGRU), and generate a corresponding enhanced representation of the user interest box sequence

和会话项目表示

生成捕获了项目顺序结构信息和用户时间兴趣特征的新会话表示C＝(c₁,c₂,…,c_n)。S406: Aggregate user interest box sequences to enhance representation through attention mechanism

and session item representation

A new session representation C=(c ₁ , c ₂ , . . . , c _n ) is generated that captures item order structure information and user temporal interest features.

Specifically, we separate items with shorter time intervals into the same user interest bin. Each bin is then mapped to an embedding and concatenated with its corresponding item embedding. Next, an asymmetric bidirectional gated recurrent neural network is used to analyze the user interest bin sequences in the session and obtain an enhanced user interest bin sequence representation. Finally, the enhanced user interest box representation is used to form the importance coefficient for the item by using the attention network, and the corresponding item is weighted to form a new item representation.

计算兴趣敏感序列Г=(γ₁,γ₂,…,γ_n)；by formula

Calculate the interest-sensitive sequence Г=(γ ₁ ,γ ₂ ,…,γ _n );

m＝t_n-t₁；in,

m=t _n -t ₁ ;

计算自适应时间跨度μ；by formula

Calculate the adaptive time span μ;

The user interest bin bin _i is calculated by the formula bin _i =k,whereγ _i ∈(μ×(k-1),μ×k];

The asymmetric gated recurrent neural network generates the enhanced representation of the sequence of user interest bins by the following formula

计算新会话表示C＝(c₁,c₂,…,c_n)；by formula

Calculate the new session representation C=(c ₁ ,c ₂ ,..., _cn );

in,

和l_p分别表示衰减常数和左偏移量；D_init和D_final表示两个预先指定的常数设置D_init＝0.98，D_final＝0.01；e表示自然常数；N表示用户兴趣箱的个数；μ表示目标会话S所需要的自适应时间间隔；k∈(1,2,…,N)；

表示可训练参数。In the above formula:

and l _p represent the decay constant and left offset, respectively; D _init and D _final represent two pre-specified constant settings D _init = 0.98, D _final = 0.01; e represents a natural constant; N represents the number of user interest boxes; μ denotes the adaptive time interval required by the target session S; k∈(1,2,…,N);

Represents trainable parameters.

The present invention can capture complex conversion patterns of items with common user interests through a temporal interest attention network, so that items with similar user interests in the time dimension can be modeled, and then local commonality between items can be paid attention to, generating captures A new session representation that incorporates item order structure information and user temporal interest features can further improve the accuracy of session recommendation.

In the specific implementation process, the probability distribution of candidate items is generated through the following steps:

S501: Based on the new session representation C=(c ₁ , c ₂ , . . . , c _n ) combined with addition and pooling, generate a long-term interest representation z _long ;

S502: Obtain the short-term interest representation z _short from the new session representation C through the GRU;

S503: Combine the long-term interest representation z _long and the short-term interest representation z _short in combination with the gating mechanism, and generate the final session representation z _final ;

S504: Select the top K candidate items v _i ∈ V from the candidate item set V=(v ₁ ,v ₂ ,...,v _|V| ) based on the session final representation z _final for recommendation, and calculate each candidate item v _i ∈ Fraction of V;

S505: Apply a softmax function based on the scores of each candidate item v _i ∈ V to generate a probability distribution of the candidate items.

计算长期兴趣表示z_long；by formula

Calculate the long-term interest representation z _long ;

计算短期兴趣表示z_short，其中，

代表在时间步长i-1处的会话项目表示，将最后一个会话项目表示作为会话短期表示

by formula

Calculate the short-term interest representation z _short , where,

Represents the session item representation at time step i-1, taking the last session item representation as the session short term representation

Calculate the session final representation z _final by the formula z _final =f⊙z _long +(1-f)⊙z _short ;

in,

计算候选项目v_i∈V的分数

by formula

Calculate the score of the candidate item v _i ∈ V

计算候选项目v_i∈V的概率，进而生成候选项目的概率分布；by formula

Calculate the probability of the candidate item v _i ∈ V, and then generate the probability distribution of the candidate item;

表示候选项目v_i∈V的初始化嵌入表示；

表示可训练参数。In the above formula:

represents the initialized embedding representation of candidate items v _i ∈ V;

Represents trainable parameters.

The invention calculates the final representation of the session based on the long-term interest representation and the short-term interest representation, and calculates and generates the probability distribution of the candidate items based on the final session representation, so that the accuracy of the session recommendation can be effectively guaranteed.

In the specific implementation process, the time-enhanced graph neural network model is trained by the time backpropagation algorithm, and the cross-entropy is used as the loss function; wherein, the cross-entropy loss function is expressed as

In the above formula: y _i represents the true label.

The present invention can effectively train the time-enhanced graph neural network model through the time back-propagation algorithm and the cross-entropy loss function, thereby improving the session recommendation effect of the model.

In order to better illustrate the advantages of the session recommendation method of the present invention, this embodiment discloses the following experiments.

1. Data set

This experiment tests the performance of the time-enhanced graph neural network model of the present invention (hereinafter referred to as TE-GNN) and a series of baseline models on three widely used benchmark datasets (Diginetica, Tmall, Nowplaying, Retailrocket).

Diginetica: The data is from the 2016 CIKM Cup Challenge. Since it contains commodity transaction type data, it is often used for conversational recommendation tasks, and we extract the data of the last week as test data.

Tmall: This data comes from the 2015 IJCAI competition, which consists of the shopping logs of anonymous users on the Tmall online shopping platform. Due to the large number of items, we choose the most recent 1/64 part of the session as the dataset, in which the session of the last day is used as test data and the rest of the sessions are used for training.

Nowplaying: This dataset is constructed by Twitter, which describes users' music listening habits. We divide the data into two parts: training and testing, in which the data of the last two months is used for testing, and the remaining historical data is used as training set.

Retailrocket: This data comes from a 2016 Kaggle competition, which includes 4.5 months of user behavior data on e-commerce sites. We extract the most recent 1/4 of the data as the training set and the last 15 days of data as the test set.

In the three datasets, sessions with session length less than 2 and items with item occurrences less than 5 were filtered.

We also perform data augmentation, for example: for session S=(v ₁ ,v ₂ ,...,v _n ), samples and labels can be generated: ([v ₁ ],v ₂ ),([v ₁ ,v ₂ ] ,v ₃ ),…,([v ₁ ,v ₂ ,…,v _n-1 ],v _n ).

2. Evaluation indicators

We use 2 widely used evaluation metrics P@20 and MRR@20 to evaluate the performance of all models. Higher values of P@K and MRR@K represent better model performance.

P@K (Precision): It measures the number proportion of the target item when it is ranked in the top-K recommendation, and is an indicator for evaluating the unranked results.

其中N是测试集数量，n_hit是目标项目在预测的top-K列表中的样本数量。

where N is the number of test sets and n _hit is the number of samples of the target item in the predicted top-K list.

MRR@K (Mean Reciprocal Rank): It is the average of the reciprocal rank of the target item in the recommendation list. This metric takes into account the position of the correctly recommended item in the ranking list.

其中N是测试集数量，rank_i是第i个目标项目在推荐列表中的位置。若目标项目未在top-K推荐列表中，则MRR@K为0。

where N is the number of test sets and rank _i is the position of the i-th target item in the recommendation list. If the target item is not in the top-K recommendation list, MRR@K is 0.

3. Baseline Model

To comprehensively evaluate the performance of our model, we compare it with 11 baseline methods, which can be roughly divided into three categories, namely: traditional recommendation methods, methods based on recurrent neural networks (RNNs) and attention mechanisms, and graph neural network based methods Methods of Network GNNs. Details of all baselines are briefly described below.

Traditional recommended method:

POP: A commonly used baseline method in recommender systems, it recommends the top N items with the highest frequency in the training set.

Item-KNN: is a collaborative filtering based method that recommends the most similar items to the current session to the user.

FPMC: This method combines matrix factorization and Markov chains, where sequence data is modeled by transition matrices, all of which are user-specific. It introduces a factorization model that gives a low-rank approximation of the transition cube, where each part is the transition matrix of the user's historical clicks under the Markov chain.

Methods based on recurrent neural networks and attention mechanisms:

GRU4REC: This method uses a gated recurrent neural network GRU to simulate the sequential behavior of users and adopts a parallel mini-batch training scheme for model training.

NARM: This method uses a recurrent neural network (RNN) to model the sequential behavior of users combined with an attention mechanism to capture the user's main preferences. At the same time, it combines bilinear matching mechanism to generate recommendation probability for each candidate item.

STAMP: This model alleviates the problem of user preference transfer by capturing users' long-term preferences and short-term interests.

CSRM: This method proposes to utilize collaborative neighborhood information for session-based recommendation. It uses the inner encoder to capture the information of the current session, while it also captures the collaboration information of the neighborhood session using the outer encoder.

SR-IEM: This method utilizes an improved attention mechanism to generate item importance scores and generates session representations based on the user's global preferences and current interests.

Methods based on graph neural networks:

SR-GNN: SR-GNN captures the complex transition relationships of items in a conversation by modeling the conversation sequence as a conversation graph. At the same time, it also combines a gated graph neural network and a self-attention mechanism to generate conversational representations.

TAGNN: This method models the conversation sequence as a conversation graph and obtains the embedding representation of the item through a graph neural network, it also introduces a target-aware module to reveal the correlation of a given target item with all candidate items, thereby improving the quality of the conversation representation .

GCE-GNN: is currently the best performing model, it learns the representation of items through 2 different perspectives, such as: conversational perspective and global perspective. Conversational perspective aims to learn the representation of items through their transition relations within a session, and global perspective aims to learn the representation of items through their transition relations across all sessions.

Fourth, the experimental parameter settings

In all experiments on TE-GNN, we set the training batch size to 256 and the vector dimension of item embeddings to be 256. The number N of user interest boxes is 6, and the number of layers of T-GCN is 2.

The model parameters in the experiment are initialized according to the mean value of 0 and the variance of 0.1. We use the Adam optimizer with a learning rate of 0.001, which decays by a factor of 0.1 every 3 epochs of training, and the random dropout rate for dropout is set to 0.3. _On the other hand, we utilize L2 regularization to avoid overfitting, which is set to 1e-4. TE-GNN is implemented by Pytorch and deployed on GPU GeForce GTX 2080Ti. Meanwhile, we set their corresponding parameters according to the papers of the baseline models.

5. Overall experiment

To verify the effectiveness of TE-GNN, the performance comparison between TE-GNN and state-of-the-art baselines is reported in Table 1.

Table 1 Performance comparison of TE-GNN and baseline models

From Table 1, it can be observed that the traditional recommendation method POP performs the worst on both metrics across all datasets. This is mainly because it only focuses on frequently occurring items and ignores useful sequence information within a session. FPMC shows better performance than POP because it applies first-order Markov chains and matrix factorization to capture user preferences. Among all traditional recommendation methods, Item-KNN achieves the best performance on all datasets. This is because the potential preferences of users have a greater impact on the recommendation effect

Compared with traditional recommendation methods, methods based on recurrent neural network (RNN) and attention mechanism show better performance. Among them, GRU4REC is the first RNN-based conversational recommendation method, which outperforms traditional recommendation methods on most datasets, which shows the practicality of pairing and modeling sequence information through RNN. Both NARM and STAMP outperform GRU4REC because they further incorporate an attention mechanism to dynamically capture the importance of items in a session. CSRM enhances the current session representation by fusing auxiliary information from other sessions and shows better performance than NARM and STAMP on all datasets. SR-IEM is a recently proposed model. This method modifies the self-attention mechanism to more accurately obtain the importance of each item in the session to capture the user's long-term preference. At the same time, the method is based on the user's long-term preference and session. The short-term preference generated by the last item in the recommendation is made.

From the experimental reports of all the baseline models, the graph neural network-based method demonstrates its superiority. The reason for this may be that the graph neural network GNN relaxes the assumption of temporal dependencies between consecutive items and models complex item transition relationships as pairwise relationships (eg: directed graphs). For example: SR-GNN captures more implicit connections between items by modeling conversation sequences as graph structures. Based on SR-GNN, TAGNN further considers user preference by combining target-aware attention mechanism. Among all GNN-based methods, GCE-GNN has the best performance on all datasets. This is because GCE-GNN effectively learns representations of items from the global context (other sessions) and the current session simultaneously.

Our proposed method TE-GNN outperforms all state-of-the-art baseline methods on all datasets. Specifically, TE-GNN outperforms the state-of-the-art baseline models by 1.03%, 16.73%, 6.62%, and 2.73% on P@20 on Diginetica, Tmall, Nowplaying, and Retairocket, respectively. Similar performance improvements can be seen on MRR@20. Experimental results show that it is crucial to model complex user interest drift and user common interests, and TE-GNN can effectively model these two modal information by introducing temporal graph convolutional network and temporal interest attention network .

6. Influence of Temporal Graph Convolutional Network (T-GCN)

To explore the impact of Temporal Graph Convolutional Network (T-GCN) on model performance, we compare it with several variant models, which are briefly described as follows:

TE-GNN-MLP: Utilizes a multilayer perceptron (MLP) to replace the T-GCN module in TE-GNN.

TE-GNN-GGNN: This variant uses a gated graph neural network (GGNN) to replace T-GCN in TE-GNN.

TE-GNN-GAT: Replace T-GCN in TE-GNN with Graph Attention Network (GAT).

Table 2 Experimental results of performance of variant models

Table 2 reports the performance experimental results of all variant models. From Table 2, it can be observed that the performance of the three variant models drops to a large extent on both metrics. More precisely, replacing T-GCN with MLP exhibits the worst performance because it lacks the ability to capture complex item transformation relationships. Both TE-GNN-GGCN and TE-GNN-GAT show better performance than TE-GNN-MLP because they effectively model the rich structural information between items based on graph neural networks. Among them, TE-GNN-GAT outperforms TE-GNN-GGCN because it considers the importance of adjacent items through a graph attention network. Our proposed method TE-GNN achieves higher performance than all three variants, which demonstrates the effectiveness of incorporating temporal graph convolutional networks into our model.

7. The influence of Temporal Interest Attention Network (TIAN)

We compare the performance of the Temporal Interest Attention Network (TIAN) with the self-attention network and location-aware attention network used in SASRec and SGNN, respectively. Different from the self-attention network and the location-aware attention network, the temporal interest attention network (TIAN) of the present invention performs fine-grained mining of temporal information. Furthermore, we further investigate the effectiveness of the component Asymmetric Bidirectional Gated Recurrent Neural Network (Asym-BiGRU) proposed in TIAN by dropping the Aasym-BiGRU component or replacing it with traditional BiGRU. A brief description of all variants is shown below:

w/o TIAN: Remove Temporal Interest Attention Network (TIAN) in TE-GNN.

TE-GNN-SA: Replacing the Temporal Interest Attention Network (TIAN) in TE-GNN with a Self-Attention Network.

TE-GNN-PA: Adopt the location-aware attention network in SGNN instead of the temporal interest attention network (TIAN) in TE-GNN.

TIAN-w/o GRU: We discard Asym-BiGRU in TIAN and do not consider the order relation of user interest bins in the session.

TIAN-BiGRU: We replace the Asym BiGRU with a regular BiGRU, which will handle the context information of user interest bin forward and backward equally.

It can be seen through experiments that equipping the Temporal Interest Attention Network will achieve significant performance improvements, which demonstrates the effectiveness of the Temporal Interest Attention Network. TE-GNN-w/o TIAN shows a significant performance drop because it cannot model users' temporal interest information. Compared with TE-GNN-w/o TIAN, neither TE-GNN-SA nor TE-GNN-PA show significant performance improvement, which is due to the fact that TE-GNN-SA only captures semantic relations between items, while TE-GNN-PA only focuses on the location information of items, so neither can capture the temporal interest information of users. Unlike these variants, TIAN-w/o GRU uses a multi-granularity interest separation strategy to explicitly model the temporal interest information of users, thus achieving excellent performance. TIAN-BiGRU improves the performance by further considering the contextual information of each user's interest bin in the session. However, a major limitation over BiGRU is that it treats the forward and backward contextual information of a point of interest equally. However, in session-based recommendation scenarios, forward context information plays a more important role than backward context information. Compared to all variants, our proposed method TE-GNN achieves the best performance by introducing an asymmetric bidirectional gated recurrent neural network (Asym BiGRU) that differentiates both sides of the user interest bin context information. modeling.

8. The influence of the number of layers of the time graph convolutional network

In order to verify the impact of Exploratory Graph Augmented Attention Network (GEA) on recommendation performance, we design the following three variant models: To examine the impact of the number of layers of T-GNN in TE-GNN, we study the Variation in performance of T-GNNs with 1 to 6 layers. Since the highway network is exploited in our model to alleviate the over-smoothing problem, we also compare the performance of our method with its corresponding variant TE-GNN-w/o HN, which runs on different numbers of T-GNNs The highway network is dropped on the layer. The experimental results on all four datasets are shown in Fig. 4. From Figure 4, we can observe that on the Diginetica dataset, the performance of our method TE-GNN first rises and peaks when the number of T-GNN layers L = 2, and then starts to decline and become larger. On the Tmall dataset, with the increasing number of T-GNN layers,

The performance of TE-GNN gradually improves and peaks when the number of T-GNN layers is L=4. If we further expand the number of layers, it will lead to a large performance drop of TE-GNN. Similar results are observed on the other two datasets (Nowplaying and Retailrocket). This trend shows that when we increase the number of T-GNN layers, our model can effectively capture the translation relationship between distant items. However, if the number of TGNN layers is too large, more irrelevant or noisy high-order neighbors are injected into the model, resulting in suboptimal model performance. We can also observe in Fig. 4 that our model TE-GNN consistently performs better than its corresponding variant TE-GNN-w/o HN. Furthermore, as the number of T-GNN layers increases, the performance of TE-GNN-w/o HN is more degraded compared to our method TE-GNN. The results show that the highway network can effectively alleviate the over-smoothing problem caused by these multilayer GCN structures.

9. The influence of the number of time interest boxes

To explore the effect of the number N of temporal interest bins in TE-GNN, we change the number of temporal interest bins from 1 to 10 on Diginetica and Tmall datasets. The experimental results are shown in Figure 5. From Figure 5, we can see that the number of user interest bins has a great impact on the performance of TE-GNN. More precisely, on the Diginetica dataset, the performance of TE-GNN first improves and peaks when N = 6, and the model performance gradually decreases when N is 7-10. The performance trend of the model on the Tmall dataset is similar to that on the Diginetica dataset. Specifically, as the number of temporal interest bins increases, the performance of TE-GNN first keeps rising until it reaches a peak (equal to 8), and then starts to drop rapidly. This variation trend is reasonable because when it is small, it cannot capture the fine-grained user interest as well as the temporal dimension. When the scale is too large, the common user interests in the session will be ignored, resulting in poor recommendation performance.

It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described with reference to the preferred embodiments of the present invention, those of ordinary skill in the art should Various changes in the above and in the details may be made therein without departing from the spirit and scope of the present invention as defined by the appended claims. Meanwhile, common knowledge such as well-known specific structures and characteristics in the embodiments are not described too much here. Finally, the scope of protection claimed in the present invention should be based on the contents of the claims, and the descriptions of the specific implementation manners in the description can be used to interpret the contents of the claims.

Claims (10)

1. a graph neural network session recommendation method based on time enhancement, it is characterized in that: target session is input in the time enhancement graph neural network model that has been trained; Described time enhancement graph neural network model converts the time interval that occurs by conversation item Generate the degree of user interest drift, and construct a time-enhanced session graph that can handle the conversion relationship between conversation items according to the degree of user interest drift, then learn item embedding based on the time-enhanced session graph and generate a new session representation, and finally calculate candidate items based on the new session representation. The probability distribution of to complete the session recommendation. 2. the graph neural network session recommendation method based on time enhancement as claimed in claim 1 is characterized in that, the time enhancement graph neural network model completes the session recommendation through the following steps: S1: map the target session S=(v ₁ ,v ₂ ,...,v _n ) to the session embedding sequence H=(h ₁ ,h ₂ ,...,h _n ); S2: Based on the session graph of the target session, the degree of user interest drift is mapped to the weights of the edges of the session graph to generate the corresponding time-enhanced session graph; S3: Learn item representations of temporally augmented conversational graphs and aggregate high-order neighbor information of items based on temporally augmented conversational graphs and conversational embeddings through a multi-layered temporal graph convolutional network to generate corresponding conversational item representations; S4: Allocate user interest boxes for each session item in the target session through the temporal interest attention network, and generate a user interest box sequence representation; then aggregate the user interest box sequence representation and the session item representation to generate a new session representation; S5: Calculate the long-term interest representation and the short-term interest representation based on the new session representation, and fuse to obtain the final session representation; then select candidate items based on the final session representation, and calculate the probability distribution of the candidate items to complete the session recommendation. 3. The time-enhanced graph neural network session recommendation method according to claim 2, wherein in step S2, the time-enhanced session graph is defined as: for the target session S=(v ₁ , v ₂ ,..., v _n ), its time-enhanced session graph is g _s =(V _s ,ε _s ,W _s ), where ε _S represents the set of edges, and W _S represents the weight matrix of the edges; Each node v _i ∈ V _S and edge (v _i-1 ,v _i )∈ε _s represents the adjacency relationship between two consecutive items v _i-1 and v _i , and its matrix expression is the in-edge matrix A ^I and out- Edge matrix A ^O ; each edge (v _i-1 ,v _i ) corresponds to a weight W _i-1,i ∈W _S ; each node v _i ∈ V _S adds a self-connected edge, and its matrix is expressed as Self-connection matrix A ^S . 4. The graph neural network session recommendation method based on time enhancement as claimed in claim 3, wherein in step S3, a session item representation is generated by the following steps: S301: Layer l output embedding representation through a multilayer temporal graph convolutional network

S302: Use the embedded representation h _i ^L output by the multi-layer temporal graph convolutional network as the embedded representation of the conversation item v _i ∈ S; S303: Embedding representation of the output of a multi-layer temporal graph convolutional network via a highway network

with its initial embedding

to merge the representations of the conversation items to obtain the item representation of the conversation item v _i ∈ S

and generate the session item representation

5. The time-enhanced graph neural network session recommendation method according to claim 4, characterized in that, by formula

compute the embedded representation h _i ^l ; by formula

Calculated item representation

in,

In the above formula:

and

respectively represent the ith row of the time-enhanced session graph input edge matrix A ^I , the input edge matrix A ^O and the self-connection matrix A ^S ; l and L both represent the number of layers of the multi-layer time graph convolutional network;

represents the trainable parameters; σ represents the function Sigmoid. 6. The time-enhanced graph neural network session recommendation method according to claim 4, wherein in step S4, a new session representation is generated by the following steps: S401: Calculate the time interval sequence Q=(q ₁ , q ₂ , between each item and the last item based on the click timestamp T=(t ₁ , t ₂ , . . . , t _n ) of each session item in the target session S ...,q _n ); S402: Map the time interval sequence Q to an interest-sensitive sequence Γ=(γ ₁ ,γ ₂ ,...,γ _n ), and calculate the adaptive time span μ; S403: Assign user interest bins bin _i to each session item in the target session S based on the adaptive time span μ through the temporal interest attention network, and generate a user interest bin sequence B=(bin ₁ , bin ₂ ,..., bin _n ) ; S404: Map each user interest box bin _i in the user interest box sequence B=(bin ₁ , bin ₂ , . . . , bin _n ) to a low-dimensional dense vector e _i , and generate a corresponding user interest box representation E=(e ₁ ,e ₂ ,…, _en ); S405: Perform asymmetric processing on the forward context information and backward context information represented by the user interest box through an asymmetric gated recurrent neural network, and generate a corresponding enhanced representation of the user interest box sequence

S406: Aggregate user interest box sequences to enhance representation through attention mechanism

and session item representation

A new session representation C=(c ₁ , c ₂ , . . . , c _n ) is generated that captures item order structure information and user temporal interest features. 7. The time-enhanced graph neural network session recommendation method according to claim 6, characterized in that, by formula

Calculate the interest-sensitive sequence Γ=(γ ₁ ,γ ₂ ,...,γ _n ); in,

m=t _n -t ₁ ; by formula

Calculate the adaptive time span μ; The user interest bin bin _i is calculated by the formula bin _i =k,whereγ _i ∈(μ×(k-1),μ×k]; The asymmetric gated recurrent neural network generates the enhanced representation of the sequence of user interest bins by the following formula

by formula

Calculate the new session representation C=(c ₁ ,c ₂ ,..., _cn ); in,

In the above formula:

and l _p represent the decay constant and left offset, respectively; D _init and D _final represent two pre-specified constant settings D _init = 0.98, D _final = 0.01; e represents a natural constant; N represents the number of user interest boxes; μ denotes the adaptive time interval required by the target session S; k∈(1,2,…,N);

Represents trainable parameters. 8. The time-enhanced graph neural network session recommendation method as claimed in claim 6, wherein in step S5, the probability distribution of candidate items is generated by the following steps: S501: Based on the new session representation C=(c ₁ , c ₂ , . . . , c _n ) combined with addition and pooling, generate a long-term interest representation z _long ; S502: Obtain the short-term interest representation z _short from the new session representation C through the GRU; S503: Combine the long-term interest representation z _long and the short-term interest representation z _short in combination with the gating mechanism, and generate the final session representation z _final ; S504: Select the top K candidate items v _i ∈ V from the candidate item set V=(v ₁ ,v ₂ ,...,v _|V| ) based on the session final representation z _final for recommendation, and calculate each candidate item v _i ∈ Fraction of V; S505: Apply a softmax function based on the scores of each candidate item v _i ∈ V to generate a probability distribution of the candidate items. 9. The time-enhanced graph neural network session recommendation method according to claim 8, characterized in that, by formula

Calculate the long-term interest representation z _long ; by formula

Calculate the short-term interest representation z _short , where,

Represents the session item representation at time step i-1, taking the last session item representation as the session short term representation

Calculate the session final representation z _final by the formula z _final =f⊙z _long +(1-f)⊙z _short ; in,

by formula

Calculate the score of the candidate item v _i ∈ V

by formula

Calculate the probability of the candidate item v _i ∈ V, and then generate the probability distribution of the candidate item; In the above formula:

represents the initialized embedding representation of candidate items v _i ∈ V;

Represents trainable parameters. 10. The time-enhanced graph neural network session recommendation method as claimed in claim 8, wherein the time-enhanced graph neural network model is trained by a time back-propagation algorithm, and cross entropy is used as a loss function; Among them, the cross entropy loss function is expressed as

In the above formula: y _i represents the true label.

Images