Session recommendation method based on target attention and attribute decomposition

Through the TADRL-GNN model, graph neural network and independent factor learning technology are used to solve the problem of intersession information utilization and interference in the session recommendation system, achieving more efficient recommendation performance.

CN119067128APending Publication Date: 2024-12-03LIAONING UNIVERSITY OF TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411158308.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-22
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing session recommendation systems are difficult to effectively utilize inter-session information, and cannot effectively eliminate interference in inter-session information, resulting in limited recommendation performance.

Method used

A session recommendation method based on target attention and attribute decomposition is proposed. The TADRL-GNN model is adopted. The graph neural network is used to decouple representation learning, graph initialization, weight update, information dissemination, target attention, contrast learning and information combination modules. Technologies such as (GNN) and independent factor learning (DRL) build embedded representations between sessions and within sessions to eliminate interference information and improve recommended performance.

Benefits of technology

The TADRL-GNN model can effectively utilize inter-session information, eliminate interference, significantly improve recommendation performance, and is better than traditional models, especially when dealing with complex sequence patterns and global relationships.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067128A_ABST
    Figure CN119067128A_ABST
Patent Text Reader

Abstract

The invention discloses a session recommendation method based on target attention and attribute decomposition, relates to the field of target session recommendation, and provides a TADRL-GNN model. Comprising a decoupling representation learning module, a graph initialization module, a weight updating module, an information spreading module, a target attention module, an attention mechanism, a session embedding learning module, a comparison learning module and an information combination module. According to the method, cross-session information is introduced, other session information is also used for assistance when an article is recommended in one session, and a weight mechanism is added to adjust different influences of different sessions on the current session, so that an interest model of a user on a factor level is generated. After the interest model of the user is generated, different interest preferences of the user can also be used when different articles are used as candidate items, so that a target attention module is added to calculate specific interest preferences.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of radio station communication, and in particular to a conversation recommendation method based on target attention and attribute decomposition. Background Art

[0002] Recommendation systems play an increasingly important role in helping people find the information they need. In early SRS, modeling was based on items that appeared in multiple users at the same time, rather than directly modeling sequences. The representative method is matrix factorization, which is mainly used to model the scattered relationships between users and items. However, the matrix factorization model only considers the existence or non-existence of interactions between users and items, so it has limitations in capturing the user's intentions that change over time. To solve this problem, a Markov chain-based method is proposed. The Markov chain-based method is a method that models user sequences based on the Markov assumption, but it has difficulties in modeling the sequential patterns of complex sequences.

[0003] With the booming development of neural networks, many models based on it have emerged to improve the recommendation accuracy of SBR. GRU4Rec first applied RNN to SBR. It used RNN and gated recurrent unit (GRU) to learn session representation by analyzing the sequential relations in the session and then predict the next item that is most likely to be interacted. NARM was subsequently proposed, which used both RNN and attention mechanism. The attention mechanism is responsible for capturing the global relations in the session because they noticed that there are global relations in the session in addition to the sequential relations. STAM also adopts the attention mechanism to achieve better results by placing more emphasis on short-term interests and focusing on the user's recent clicks. However, these models all face a limitation, that is, insufficient information within the session. Many conversational recommendation system models try to use rich inter-session information to assist recommendations. However, inter-session information contains both effective information and interference information, and existing models cannot filter out this interference information.

[0004] Therefore, it is necessary to provide a conversation recommendation method based on target attention and attribute decomposition to solve the above problems. Summary of the invention

[0005] The purpose of this invention is to provide a conversation recommendation method based on target attention and attribute decomposition, and proposes TADRL-GNN which makes full use of inter-conversation information and eliminates interference during learning to improve recommendation performance.

[0006] To achieve the above objectives, the present invention provides a conversation recommendation method based on target attention and attribute decomposition, and proposes a TADRL-GNN model, which includes a decoupled representation learning module, a graph initialization module, a weight update module, an information propagation module, a target attention module, an attention mechanism, a conversation embedding learning module, a contrastive learning module and an information combination module; specifically, the following steps are included:

[0007] S1: In the decoupled learning representation module, the independent factors of the decoupled representation learning learn the DRL-level representation to obtain the different interest characteristics shown by the user in the session;

[0008] S2: constructing the inter-session graph and the intra-session graph in the initialization module;

[0009] S3: Use independent factor learning (DRL) to capture the interest preferences of different factors hidden in the interaction and embed the items into different spaces;

[0010] S4: The inter-session graph involves an inter-session module. The inter-session module uses GCN to learn factor-level cross-session embeddings. The input of the cross-session module includes the initial factor-level embeddings of all sessions and the cross-session graph. The attribute level of the cross-session graph is affected by the edge weights.

[0011] S5: The basis for updating weights in the weight update module is stability and interest units;

[0012] S6: Based on the established factor-level cross-session dependencies, graph propagation is used to update the embedding of vertices to learn cross-session information;

[0013] S7: In the session embedding learning module, the final embedding of each session is learned through the GNN model through the target attention module and the attention mechanism;

[0014] S8: The contrastive learning module introduces a contrastive learning method to learn the basic features of objects through the similarities or differences between objects, and updates the embedded representation by comparing the information learned from two different perspectives;

[0015] S9: The information combination module uses the fusion gating mechanism to aggregate the information learned in step S8 to obtain the final session representation.

[0016] Preferably, in step S1, the original embedding is split into multiple different independent dimensions, and DRL is introduced to improve the interpretability and robustness of the model. The input of the model is the embedding of items or sessions; specifically, it is expressed as:

[0017] Let c∈R d represents a d-dimensional embedding, DRL re-embeds it into the corresponding space, calculates k factor-level embeddings from c, and calculates the attribute-level embedding f on attribute t.t As shown below:

[0018] f t =σ(c T W t )+b t (1≤t≤k)

[0019] In the formula, and represents the weight matrix and offset on attribute t, Represents the dimension of each attribute embedding, R represents the set of original embeddings, and DRL uses the following loss function as the learning objective to generate independent factor-level embeddings:

[0020]

[0021] In the formula, y i represents the true embedding factor label, represents the probability estimate of the TADRL-GNN model for the i-th sample prediction, and represents the predicted probability distribution, and m represents the number of embedded factor samples;

[0022] The formula for using dCor to measure the correlation between different spatial variables during the embedding process is as follows:

[0023]

[0024] In the formula, g and j represent all variables from 1 to k, f g and f j Represent the mapping or feature vector of the g-th and j-th variables in the embedding space respectively.

[0025] Preferably, in step S2, the graph between sessions: S = {s1...s M}, representing the M sessions in the input, and the graph constructed between sessions is an undirected graph: Represents the dependency between sessions in S, the superscript g represents the global, vertex set A set of vertices and edges representing a session Where s i ,s j ∈S, each edge represents the same node between two sessions, and the edge weight is assigned to 1.0;

[0026] The graph constructed within the session is a directed graph The superscript l represents the local graph, the vertex set Represents the set of all points and edges in session s

[0027] is the i-th item in session s, and each edge represents the connection from the previous node to the next node.

[0028] Preferably, in step S3, the following steps are specifically included:

[0029] For each session S = v (s,1) ....v (s,n) Embed the corresponding project;

[0030] Suppose a project v (s,1) ∈S is initialized to c (i,j) , and then use DRL to obtain the interest preferences of different factors hidden in the interaction, embedding each item into c (i,j) Re-embed into a different space as shown below:

[0031]

[0032] in the formula Yes (i,j) Embedding at the factor level, Yes (i,j) The collection of factor-level embeddings;

[0033] The embedding of a session is obtained by averaging the embeddings of each item:

[0034]

[0035] Where n represents the number of initialized embedded items.

[0036] Preferably, in step S4, factor-level data is processed by GCN in the inter-session module in a manner of learning cross-session embeddings;

[0037] S={s1…s M} is a batch input of M sessions. The input of the cross-session module includes the set The collection contains the initial factor level embeddings for all sessions, as well as the cross-session graph The cross-session graph is divided into the following according to the attribute level: The attribute-level cross-session graph is exactly the same as the item-level cross-session graph in terms of vertices and edges, the difference is that the edge weights are affected by the cosine similarity between attributes between sessions;

[0038] let As The adjacency matrix is For i and j The formula for defining the weight in the attribute t weight text is:

[0039]

[0040] The connections established by the above formula are based on similarity.

[0041] Preferably, in step S5, the following steps are specifically included:

[0042] The first i items in S establish the units of interest at the level of the attributes:

[0043] U (s,i) = {u (s,1) ...u (s,i)};

[0044] The formula for the interest unit of session S on attribute k is:

[0045]

[0046] When measuring the stability of interest, the following formula is used:

[0047]

[0048] along with As the interest preference of factor t increases, The weight of is updated according to the following formula:

[0049]

[0050] In the formula, express For any weight in , a session has its own inter-session graph, whose adjacency matrix is

[0051] Preferably, in step S6, graph propagation is used to update the embedding of the vertices to learn cross-session information, and the formula is expressed as:

[0052]

[0053] In the formula, is the adjacency matrix of the factor-level cross-session graph, and is the embedding after l propagations, and the final output is the concatenation of the factor-level session embeddings:

[0054] Preferably, in step S7, the target attention model dynamically matches the user's preferences in different directions according to the different recommended items through the following formula:

[0055]

[0056] Generate a conversation embedding formula with interest preferences through items that may be recommended:

[0057]

[0058] Where w is a parameter matrix, which calculates the score weights of attribute-level items in the session according to the type of target item. Indicates items that need to be recommended. represents the i-th item in the session, and m represents the number of items in the current session;

[0059] The session embedding learning module learns the embedding of the current session by aggregating the information in the item session that interacts with it, and deploys k channels to process the corresponding factor-level embeddings, including the set of factor-level embeddings of all items in s. and A collection of derived factor-level session-internal graphs

[0060]

[0061] A graph gated neural network (GGNN) is used to update the embedding and filter information during the propagation process. The conversation encoder uses soft attention and calculates the weights using the following formula:

[0062]

[0063] The target embedding generated by the selected item is obtained by the following formula:

[0064]

[0065] The final factor-level session embedding is obtained by the following formula:

[0066]

[0067] In the formula, is the learning parameter, yes The factor level embedding of the last item, the final layer output of GGNN is

[0068] Preferably, step S8 specifically includes the following steps:

[0069] The two embedded representations of session S obtained by the cross-session module and the internal session module generate a self-supervisory signal, and the standard binary cross entropy BCE loss function formula is selected as the learning objective of the final model to measure the difference between the two:

[0070]

[0071] In the formula, The random permutation of rows and columns is σ represents the activation function.

[0072] Preferably, in step S9, a fusion gating mechanism is used to aggregate the information learned from the cross-session module and the internal session module, as shown below:

[0073]

[0074] Get the final session representation formula:

[0075]

[0076] In the formula, and are the weight matrices in the fusion gate, respectively. Formula ζ is used to control the overall proportion of information from the session module and the internal session module. The final output is k channels.

[0077] Therefore, the present invention adopts the above-mentioned conversation recommendation method based on target attention and attribute decomposition, and proposes TADRL-GNN to make full use of inter-conversation information and eliminate interference during learning to improve recommendation performance.

[0078] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a flow chart of a conversation recommendation method based on target attention and attribute decomposition of the present invention. DETAILED DESCRIPTION

[0080] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0081] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0082] The words "include" or "comprises" and the like used in the present invention mean that the elements before the word include the elements listed after the word, and do not exclude the possibility of also including other elements. The orientation or position relationship indicated by the terms "inside", "outside", "upper", "lower", etc. is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation of the present invention. When the absolute position of the described object changes, the relative position relationship may also change accordingly. In the present invention, unless otherwise clearly specified and limited, the terms "attachment" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral body; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0083] like Figure 1 As shown, the present invention provides a conversation recommendation method based on target attention and attribute decomposition, and proposes a TADRL-GNN model, including a decoupled representation learning module, a graph initialization module, a weight update module, an information propagation module, a target attention module, an attention mechanism, a conversation embedding learning module, a contrastive learning module and an information combination module; specifically, the following steps are included:

[0084] S1: In the decoupled learning representation module, the independent factors of the decoupled representation learning learn the DRL-level representation to obtain the different interest characteristics shown by the user in the session; in step S1, the original embedding is split into multiple different independent dimensions, and DRL is introduced to improve the interpretability and robustness of the model. The input of the model is the embedding of the item or session; specifically, it is expressed as:

[0085] Let c∈R d represents a d-dimensional embedding, DRL re-embeds it into the corresponding space, calculates k factor-level embeddings from c, and calculates the attribute-level embedding f on attribute t. t As shown below:

[0086] f t =σ(c T W t )+b t (1≤t≤k)

[0087] In the formula, and represents the weight matrix and offset on attribute t, Represents the dimension of each attribute embedding, R represents the set of original embeddings, and DRL uses the following loss function as the learning objective to generate independent factor-level embeddings:

[0088]

[0089] In the formula, y i represents the true embedding factor label, represents the probability estimate of the TADRL-GNN model for the i-th sample prediction, and represents the predicted probability distribution, and m represents the number of embedded factor samples;

[0090] The formula for using dCor to measure the correlation between different spatial variables during the embedding process is as follows:

[0091]

[0092] In the formula, g and j represent all variables from 1 to k, f g and f j Represent the mapping or feature vector of the g-th and j-th variables in the embedding space respectively.

[0093] S2: Initialize the module to build the inter-session graph and the intra-session graph; in step S2, the inter-session graph: S = {s1…s M}, representing the M sessions in the input, and the graph constructed between sessions is an undirected graph: Represents the dependency between sessions in S, the superscript g represents the global, vertex set A set of vertices and edges representing a session Where s i ,s j ∈S, each edge represents the same node between two sessions, and the edge weight is assigned to 1.0;

[0094] The graph constructed within the session is a directed graph The superscript l represents the local graph, the vertex set Represents the set of all points and edges in session s is the i-th item in session s, and each edge represents the connection from the previous node to the next node.

[0095] S3: Use independent factor learning DRL to obtain the interest preferences of different factors hidden in the interaction and embed the items into different spaces; in step S3, the following steps are specifically included:

[0096] For each session S = v (s,1) ....v (s,n)All are embedded in the corresponding projects;

[0097] Suppose a project v (s,1) ∈S is initialized to c (i,j) , and then use DRL to obtain the interest preferences of different factors hidden in the interaction, embedding each item into c (i,j) Re-embed into a different space as shown below:

[0098]

[0099] in the formula Yes (i,j) Embedding at the factor level, Yes (i,j) The collection of factor-level embeddings;

[0100] The embedding of a session is obtained by averaging the embeddings of each item:

[0101]

[0102] Where n represents the number of initialized embedded items.

[0103] S4: The inter-session graph involves an inter-session module, which uses GCN to learn factor-level cross-session embeddings. The input of the cross-session module includes the initial factor-level embeddings of all sessions and the cross-session graph. The attribute level of the cross-session graph is affected by the edge weights. In step S4, the factor-level data is processed in the inter-session module by GCN in a way of learning cross-session embeddings.

[0104] S={s1…s M} is a batch input of M sessions. The input of the cross-session module includes the set The collection contains the initial factor level embeddings for all sessions, as well as the cross-session graph The cross-session graph is divided into the following according to the attribute level: The attribute-level cross-session graph is exactly the same as the item-level cross-session graph in terms of vertices and edges, the difference is that the edge weights are affected by the cosine similarity between attributes between sessions;

[0105] let As The adjacency matrix is For i and j The formula for defining the weight in the attribute t weight text is:

[0106]

[0107] The connections established by the above formula are based on similarity.

[0108] S5: The basis for updating the weight in the weight updating module is stability and interest units; in step S5, the following steps are specifically included:

[0109] The first i items in S establish the units of interest at the level of the attributes:

[0110] U (s,i) = {u (s,1) ...u (s,i)};

[0111] The formula for the interest unit of session S on attribute k is:

[0112]

[0113] When measuring the stability of interest, the following formula is used:

[0114]

[0115] along with As the interest preference of factor t increases, The weight of is updated according to the following formula:

[0116]

[0117] In the formula, express For any weight in , a session has its own inter-session graph, whose adjacency matrix is

[0118] S6: Based on the established factor-level cross-session dependencies, graph propagation is used to update the embedding of vertices to learn cross-session information; In step S6, graph propagation is used to update the embedding of vertices to learn cross-session information, and the formula is expressed as:

[0119]

[0120] In the formula, is the adjacency matrix of the factor-level cross-session graph, and is the embedding after l propagations, and the final output is the concatenation of the factor-level session embeddings:

[0121] S7: In the session embedding learning module, the final embedding of each session is learned through the target attention module and the attention mechanism through the GNN model; in step S7, the target attention model dynamically matches the user's preferences in different directions according to the different recommended items through the following formula:

[0122]

[0123] Generate a conversation embedding formula with interest preferences through items that may be recommended:

[0124]

[0125] Where w is a parameter matrix, which calculates the score weights of attribute-level items in the session according to the type of target item. Indicates items that need to be recommended. represents the i-th item in the session, and m represents the number of items in the current session;

[0126] The session embedding learning module learns the embedding of the current session by aggregating the information in the item session that interacts with it, and deploys k channels to process the corresponding factor-level embeddings, including the set of factor-level embeddings of all items in s. and A collection of derived factor-level session-internal graphs

[0127]

[0128] The graph gated neural network GGNN is used to update the embedding. GGNN is a variant of GNN. It filters information during the propagation process. The conversation encoder uses soft attention and calculates the weights through the following formula:

[0129]

[0130] The target embedding generated by the selected item is obtained by the following formula:

[0131]

[0132] The final factor-level session embedding is obtained by the following formula:

[0133]

[0134] In the formula, is the learning parameter, yes The factor level embedding of the last item, the final layer output of GGNN is

[0135] S8: The contrastive learning module introduces a contrastive learning method to learn the basic features of objects through the similarities or differences between objects, and updates the embedded representation by comparing the information learned from two different perspectives;

[0136] Step S8 specifically includes the following steps:

[0137] The two embedded representations of session S obtained by the cross-session module and the internal session module generate a self-supervisory signal, and the standard binary cross entropy BCE loss function formula is selected as the learning objective of the final model to measure the difference between the two:

[0138]

[0139] In the formula, The random permutation of rows and columns is σ represents the activation function.

[0140] S9: The information combination module uses the fusion gating mechanism to aggregate the information learned in step S8 to obtain the final session representation. In step S9, the fusion gating mechanism is used to aggregate the information learned in the cross-session module and the internal session module, as shown below:

[0141]

[0142] Get the final session representation formula:

[0143]

[0144] In the formula, and are the weight matrices in the fusion gate, respectively. Formula ζ is used to control the overall proportion of information from the session module and the internal session module. The final output is k channels.

[0145] Example

[0146] In order to demonstrate the comparative performance of TADRL-GNN, a comparative experiment was set up and three types of comparative models were selected, namely:

[0147] (1) Non-GNN models: FPMC, NARM, GRU4Rec and STAMP;

[0148] (2) GNN models that only consider intra-session information: SR-GNN, Disen-GNN, and TAGNN;

[0149] (3) GNN models that consider both inter-session and intra-session information: DHCN, I3GN, and COTREC.

[0150] FPMC uses Markov chains for recommendation. It only focuses on the sequential relationship between interactive items and does not construct the user's potential interests.

[0151] GRU4REC applies GRU in NLP to SBR. As an RNN-based model, it only cares about the sequential relationship between items.

[0152] NARM combines the attention mechanism with the Gated Recurrent Unit (GRU) to consider global and sequential relationships for recommendation.

[0153] STAMP emphasizes the impact of short time and designs a special attention mechanism with MLP.

[0154] SR-GNN introduces GGNN to obtain item embeddings through information propagation and adopts soft attention mechanism to obtain session embedding for recommendation.

[0155] Disen-GNN applies DRL to SBR to learn latent factor-level embeddings. Then GGNN is used in each factor channel to learn factor-level session embeddings. Its model structure is very simple, but it has become a state-of-the-art model.

[0156] TAGNN is an improved variant of SR-GNN. It replaces the original conversation encoder of SR-GNN with a target-oriented attention network.

[0157] DHCN noticed that traditional contrastive learning strategies have limited effectiveness in SBR due to the extreme scarcity of data.

[0158] This scheme proposes a new contrastive learning network to achieve data enhancement by learning inter-session information. The data enhancement is compared with the above method. The results are shown in Table 1.

[0159] Table 1

[0160]

[0161]

[0162] In order to verify the effectiveness of TADRL-GNN, three commonly used datasets were used in conversational recommendation systems for experimental demonstration. These datasets are Nowplaying1, Yoochoose1 / 64, and Diginetica. Nowplaying is a music dataset that contains users' listening behavior. Yoochoose is a dataset released by RecsysChallenge, which contains users' click behavior on the e-commerce website Yoochoose. Diginetica contains users' transaction data.

[0163] For the Diginetica and Nowplaying datasets, the last week of sessions is used as test data for detection. For the Yoochoose dataset, the last day of sessions is used as test data. It should be noted that due to the large size of the Yoochoose dataset, only the most recent 1 / 64 part of the data is used in the experiment, and the rest of the data is used as training data. The number of factors k is set to 5 in the Diginetica dataset and the Yoochoose1 / 64 dataset, and is set to 10 in the Nowplaying dataset.

[0164] By comparing the experimental results, we can finally draw the following three observations:

[0165] (1) Compared with RNN-based and attention-based models, the GNN-based model clearly performs better. This demonstrates the powerful ability of GNN in learning more accurate embeddings and modeling conversational data.

[0166] (2) Among the models that only consider intra-session information, Disen-GNN performs best, even better than some models that consider both intra-session and inter-session information. The reason is that the DRL technology is introduced to help it deeply explore the limited intra-session dependency semantics.

[0167] (3) TADRL-GNN outperforms all baseline models in all datasets. Compared with the previous best performing model, TADRL-GNN improves by about 3% on Yoochoose1 / 64 and by about 1% on Diginetica. In particular, through experiments, it is speculated that there are three reasons for this model improvement. First, TADRL-GNN adopts GNN to establish the relationship between items and sessions to better learn embedding. Second, both TADRL-GNN and Disen-GNN build dependencies at the factor level, TADRL-GNN performs better, and TADRL-GNN adopts a target attention mechanism. Finally, TADRL-GNN adopts a cautious strategy to effectively learn inter-session information while avoiding the interference of noise and redundancy contained therein.

[0168] Therefore, the present invention adopts the above-mentioned session recommendation method based on target attention and attribute decomposition, introduces cross-session information, and uses other session information to assist when recommending items in a session. A weight mechanism is added to adjust the different impacts of different sessions on the current session, thereby generating a user's interest model at the factor level, eliminating interference during learning, and improving recommendation performance.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A conversation recommendation method based on target attention and attribute decomposition, characterized by: A TADRL-GNN model is proposed, which includes a decoupled representation learning module, a graph initialization module, a weight update module, an information propagation module, a target attention module, an attention mechanism, a session embedding learning module, a contrastive learning module and an information combination module. Specifically, the following steps are included: S1: In the decoupled learning representation module, the independent factors of the decoupled representation learning learn the DRL-level representation to obtain the different interest characteristics shown by the user in the session; S2: constructing the inter-session graph and the intra-session graph in the initialization module; S3: Use independent factor learning (DRL) to capture the interest preferences of different factors hidden in the interaction and embed the items into different spaces; S4: The inter-session graph involves an inter-session module. The inter-session module uses GCN to learn factor-level cross-session embeddings. The input of the cross-session module includes the initial factor-level embeddings of all sessions and the cross-session graph. The attribute level of the cross-session graph is affected by the edge weights. S5: The basis for updating weights in the weight update module is stability and interest units; S6: Based on the established factor-level cross-session dependencies, graph propagation is used to update the embedding of vertices to learn cross-session information; S7: In the session embedding learning module, the final embedding of each session is learned through the GNN model through the target attention module and the attention mechanism; S8: The contrastive learning module introduces a contrastive learning method to learn the basic features of objects through the similarities or differences between objects, and updates the embedded representation by comparing the information learned from two different perspectives; S9: The information combination module uses the fusion gating mechanism to aggregate the information learned in step S8 to obtain the final session representation.

2. The conversation recommendation method based on target attention and attribute decomposition according to claim 1, characterized in that: In step S1, the original embedding is split into multiple independent dimensions, and DRL is introduced to improve the interpretability and robustness of the model. The input of the model is the embedding of items or sessions; specifically, it is expressed as: Let c∈R d represents a d-dimensional embedding, DRL re-embeds it into the corresponding space, calculates k factor-level embeddings from c, and calculates the attribute-level embedding f on attribute t. t As shown below: f t =σ(c T W t )+b t (1≤t≤k) In the formula, and represents the weight matrix and offset on attribute t, represents the dimension of each attribute embedding, and R represents the set of original embeddings. DRL uses the following loss function as the learning objective to generate independent factor-level embeddings: In the formula, y i represents the true embedding factor label, represents the probability estimate of the TADRL-GNN model for the i-th sample prediction, and represents the predicted probability distribution, and m represents the number of embedded factor samples; The formula for using dCor to measure the correlation between different spatial variables during the embedding process is as follows: In the formula, g and j represent all variables from 1 to k, f g and f j Represent the mapping or feature vector of the g-th and j-th variables in the embedding space respectively.

3. The conversation recommendation method based on target attention and attribute decomposition according to claim 1, characterized in that: In step S2, the graph between sessions: S = {s1…s M }, representing the M sessions in the input, and the graph constructed between sessions is an undirected graph: Represents the dependency between sessions in S, the superscript g represents the global, vertex set A set of vertices and edges representing a session Where s i ,s j ∈S, each edge represents the same node between two sessions, and the edge weight is assigned to 1.0; The graph constructed within the session is a directed graph The superscript L represents the local graph, the vertex set Represents the set of all points and edges in session s is the i-th item in session s, and each edge represents the connection from the previous node to the next node.

4. The conversation recommendation method based on target attention and attribute decomposition according to claim 1, characterized in that: In step S3, the following steps are specifically included: For each session S = v (s , 1) ....υ (s,n) All are embedded in the corresponding projects; Set up a project (s,1) ∈S is initialized to c (i,j) , and then use DRL to obtain the interest preferences of different factors hidden in the interaction, embedding each item into c (i,j) Re-embed into a different space as shown below: in the formula Yes (i,j) Embedding at the factor level, Yes (i,j) The collection of factor-level embeddings; The embedding of a session is obtained by averaging the embeddings of each item: Where n represents the number of initialized embedded items.

5. The conversation recommendation method based on target attention and attribute decomposition according to claim 4, characterized in that: In step S4, factor-level data is processed by GCN in the inter-session module to learn cross-session embeddings; S={s1…s M } is a batch input of M sessions. The input of the cross-session module includes the set The collection contains the initial factor level embeddings for all sessions, as well as the cross-session graph The cross-session graph is divided into the following according to the attribute level: The attribute-level cross-session graph is exactly the same as the item-level cross-session graph in terms of vertices and edges, the difference is that the edge weights are affected by the cosine similarity between attributes between sessions; let As The adjacency matrix is For i and j The formula for defining the weight in the attribute t weight text is: The connections established by the above formula are based on similarity.

6. The conversation recommendation method based on target attention and attribute decomposition according to claim 5, characterized in that: In step S5, the following steps are specifically included: The first i items in S establish the units of interest at the level of the attributes: IN (s,i) ={in (s,1) ...in (s,i) }; The formula for the interest unit of session S on attribute k is: When measuring the stability of interest, the following formula is used: along with As the interest preference of factor t increases, The weight of is updated according to the following formula: In the formula, express For any weight in , a session has its own inter-session graph, whose adjacency matrix is 7. The conversation recommendation method based on target attention and attribute decomposition according to claim 5, characterized in that: In step S6, graph propagation is used to update the embedding of vertices to learn cross-session information, which is expressed as: In the formula, is the adjacency matrix of the factor-level cross-session graph, and is the embedding after l propagations, and the final output is the concatenation of the factor-level session embeddings:

8. The conversation recommendation method based on target attention and attribute decomposition according to claim 5, characterized in that: In step S7, the target attention model dynamically matches the user's preferences in different directions according to the different recommended items through the following formula: Generate a conversation embedding formula with interest preferences through items that may be recommended: Where w is a parameter matrix that calculates the score weights of attribute-level items in a session according to the type of the target item. Indicates items that need to be recommended. represents the i-th item in the session, and m represents the number of items in the current session; The session embedding learning module learns the embedding of the current session by aggregating the information in the item session that interacts with it, and deploys k channels to process the corresponding factor-level embeddings, including the set of factor-level embeddings of all items in s. and A collection of derived factor-level session-internal graphs A graph gated neural network (GGNN) is used to update the embedding and filter information during the propagation process. The conversation encoder uses soft attention and calculates the weights using the following formula: The target embedding generated by the selected item is obtained by the following formula: The final factor-level session embedding is obtained by the following formula: In the formula, is the learning parameter, yes The factor level embedding of the last item, the final layer output of GGNN is 9. The conversation recommendation method based on target attention and attribute decomposition according to claim 5, characterized in that: Step S8 specifically includes the following steps: The two embedded representations of session S obtained by the cross-session module and the internal session module generate a self-supervisory signal, and the standard binary cross entropy BCE loss function formula is selected as the learning objective of the final model to measure the difference between the two: In the formula, The random permutation of rows and columns is σ represents the activation function.

10. The conversation recommendation method based on target attention and attribute decomposition according to claim 9, characterized in that: In step S9, the information learned from the cross-session module and the intra-session module is aggregated using a fusion gating mechanism as follows: Get the final session representation formula: In the formula, and are the weight matrices in the fusion gate, respectively. Formula ζ is used to control the overall proportion of information from the session module and the internal session module. The final output is k channels.