Session-based recommendation method and system for behavioral habit enhancement intention

Abstract

The invention belongs to the technical field of recommendation systems, and provides a session-based recommendation method for behavioral habit enhancement intention, which specifically comprises the following steps of: session intention feature acquisition: constructing a global item conversion graph to extract semantic connection between items, aggregating item behavior sequences, and acquiring a session intention feature; the semantic connection and the behavior sequence are fused through a gating mechanism to capture a fine-grained user intention, and session intention features of fine-grained behavior operation are obtained based on the user intention; session behavior habit feature acquisition: assigning a session item as a hyperedge in a hypergraph, assigning all behaviors related to the item and occurring in the session as vertexes connected with the hyperedge, constructing a behavior habit hypergraph, and learning user-level behavior habit features based on the behavior habit hypergraph; and session feature fusion: fusing the session intention features and the behavioral habit features to learn real intention representation of the user, and predicting items that the user may be interested in through the representation.

Description

Technical Field

[0001] The present invention belongs to the technical field of recommendation systems, and the present invention relates to a conversation-based recommendation method and system for enhancing the intention of behavioral habits. Background Art

[0002] Session-based recommendation algorithms aim to predict the next item in anonymous, sequential sessions. Due to the increasing improvement of user privacy protection in recent years, session-based recommendation systems have gradually become a hot research field. Traditional methods use Markov chains to encode sessions. However, these methods can only predict the next item based on the previous item, which leads to insufficient utilization of the overall sequence information of the session. To solve this problem, some methods use recurrent neural networks (RNNs) to model sessions and capture the temporal relationships in sequences. In recent years, due to the advantages of graph neural networks (GNNs) in capturing structural features, many methods have introduced GNNs into the feature learning process and achieved good results.

[0003] However, the above methods only consider the macroscopic connections between items, but do not utilize the complex behavioral information between users and items, and cannot capture fine-grained user preferences. For example, in a shopping scenario, users are obviously more interested in products that are more relevant to the products they have added to the shopping cart or purchased than those they have only browsed. Therefore, considering behavioral information can better capture the user's intention. Recently, some methods have introduced multi-behavioral information into conversational recommendations. TGT designs a temporal graph Transformer to capture dynamic user-item interaction patterns by exploring the evolving correlations between behaviors. MBHT designs a multi-scale Transformer to learn the behavioral order patterns of sequences and utilizes hypergraphs to learn cross-behavioral dependencies. EMBSR designs multi-graphs to aggregate micro-operations in sequences and utilizes an extended attention mechanism to learn binary behavioral patterns.

[0004] Although existing methods have achieved good results in aggregating sequential representations of behavioral information and mining relationships between behavioral dependencies, there are still some issues to be addressed:

[0005] First, it is impossible to learn user-level behavioral habits. Current multi-behavior-based conversational recommendation methods are limited to learning behavioral dependencies at the behavior or item level, while ignoring complex behavioral dependencies at the user level. Behavior-level dependency learning methods only model a single behavior and have difficulty handling complex multi-behavior scenarios. Although item-level methods can capture cross-behavior dependencies, they can usually only learn binary behavioral dependencies or continuous behaviors at continuous timestamps and cannot break through the limitations of time. This limitation means that they cannot capture all of a user's behavioral habits towards items. Existing methods do not fully exploit user-level dependencies, which is critical to accurately reflect user intent.

[0006] Second, in a session-based recommender system, the task of predicting the items under the target behavior is unreasonable. Current multi-behavior-based session recommender system approaches usually define a target behavior and treat other behaviors as auxiliary behaviors, focusing on predicting the next item under the target behavior. However, in reality, users often perform auxiliary behaviors before performing the target behavior. This approach may cause the system to recommend items that the user has already interacted with instead of new items, which contradicts the purpose of the recommender system and may degrade the user experience. Summary of the invention

[0007] In view of this, the present invention provides a conversation-based recommendation method and system with behavioral habit enhanced intention, which can accurately predict the items of interest to the user by fusing conversation intention features and conversation behavioral habit features to learn the user's true intention representation.

[0008] The technical solution for implementing the present invention is as follows:

[0009] A conversation-based recommendation method for behavioral habit enhancement intention, the specific process is as follows:

[0010] Conversational intent feature acquisition: A global item transition graph is constructed to extract semantic connections between items, and item behavior sequences are aggregated using a gated recurrent unit (GRU). The semantic connection is fused with the behavior sequence through a gating mechanism to capture fine-grained user intent, and the conversational intent features of fine-grained behavior operations are further acquired based on the user intent.

[0011] Acquisition of conversation behavior habit features: specifying conversation items as hyperedges in a hypergraph, and specifying all item-related behaviors that occurred in the conversation as vertices connected by the hyperedge, constructing a behavior habit hypergraph, and learning user-level behavior habit features based on the behavior habit hypergraph;

[0012] Session feature fusion: The session intention feature is fused with the behavior habit feature to learn the user's true intention representation, and the items that the user may be interested in are predicted through the representation.

[0013] Furthermore, the specific process of acquiring the conversation intention feature of the present invention is as follows:

[0014] 1.1 For the items in the session, a global item transformation graph is constructed based on the semantic connections between items to obtain the item representation of global high-order neighborhood information

[0015] 1.2 Item Representation Based on the Global High-Order Neighborhood Information Aggregate the behavior sequence of items and learn the macro-item representation X of the final intent of the session to capture fine-grained user intent;

[0016] 1.3 Based on the macro-item representation X of the final intent of the conversation, generate a conversation intent containing the item order information in the conversation, the global level neighborhood relationship and the fine-grained behavior operation within the item

[0017] Furthermore, the specific process of step 1.1 of the present invention is as follows:

[0018] First, based on the semantic connection between conversation items, we find the i The set of subsequent neighbors Construct a global project conversion graph G; generate a normalized adjacency matrix A based on the global project conversion graph G g ;

[0019] Secondly, based on the normalized adjacency matrix A g , using graph convolutional layers to globally encode items, in L g After layer propagation, we get the item representation that integrates global high-order neighborhood information

[0020] Furthermore, the specific process of step 1.2 of the present invention is as follows:

[0021] First, for each item v in the session i The corresponding behavior sequence h i Send it to the gated recurrent unit (GRU) to learn, and take the hidden state of the last behavior as the item v i The behavior time characteristics of z i , and then we get the behavior time characteristics of all items Z = [z1, z2, …, z n ];

[0022] Secondly, the item representation from the global high-order neighborhood information The sampled items are embedded into representation E, which is connected with the temporal features Z to achieve the aggregation of the behavior temporal features and the target representation;

[0023] Again, the aggregated information is fed into two adjacency matrices A respectively.in and A out Propagate in and obtain a high-order fusion of the two views I in , I out

[0024] Finally, for the high-order fusion of the two views I in , I out , by filtering and updating the item embedding and behavioral intention information using a gating mechanism, the final macro-item representation X in the session is obtained.

[0025] Furthermore, the specific process of step 1.3 of the present invention is as follows:

[0026] Constructed a conversational intent prototype p ;

[0027]

[0028] Among them, e i Indicates the initialized item;

[0029] Based on the conversation intent prototype, the gated attention layer is used to pass the conversation intent prototype into the item representation as shown below:

[0030]

[0031]

[0032] Among them, α i For the macro project v in the session i The attention weight is is a learnable parameter, d is the dimension of the embedding vector, x i is an element of X;

[0033] Use the last macro item v i As the overall intent of the conversation:

[0034]

[0035] in, An intent representation for a session that integrates the sequential information of items in the session, global-level neighborhood relations, and fine-grained behavioral operations within items.

[0036] Furthermore, the specific process of acquiring the conversation behavior habit features of the present invention is as follows:

[0037] 2.1 Assign the conversation items as hyperedges in the hypergraph, and assign all the item-related behaviors that occurred in the conversation as the vertices connected by the hyperedges, and construct a behavior habit hypergraph G for each conversation. h, and residual connections are added when constructing the hypergraph;

[0038] 2.2 Hypergraph G for the behavioral habits h , using the hypergraph convolution method for message passing, integrating behavioral habits into item representation, and obtaining the hypergraph learning network E h ;

[0039] 2.3 Perform average pooling on the hypergraph learning network to obtain the final session representation after integrating the behavioral habits into the project

[0040] Furthermore, the session information fusion described in the present invention is: using a gated attention layer to achieve mutual information fusion with weights.

[0041] Furthermore, the session information fusion of the present invention is as follows:

[0042]

[0043]

[0044] Among them, W s and b s is the learning parameter, e s is the final session representation, σ(·) represents the sigmoid function, Represents the element-wise multiplication operator.

[0045] Furthermore, the present invention also includes optimizing and predicting:

[0046] L2 regularization is performed on the session representation and item embedding, the formula is as follows:

[0047]

[0048]

[0049] Among them, E g is the initialized global project embedding;

[0050] The final prediction results are calculated, and cross entropy loss is used as the optimizer for model training:

[0051]

[0052]

[0053] in, and i Respectively represent v i The predicted probability and true value of the item, Represents the loss function of the model.

[0054] On the other hand, the embodiment of the present application is a conversation-based recommendation system for enhancing intentions through behavioral habits, including an intention learning module, a behavioral habit learning module, and a feature fusion module;

[0055] The intent learning module is used to construct a global item transition graph to extract semantic connections between items, aggregate item behavior sequences, and fuse semantic connections with behavior sequences through a gating mechanism to capture fine-grained user intent, and further obtain conversational intent features of fine-grained behavior operations based on the user intent;

[0056] A behavior habit learning module specifies a conversation item as a hyperedge in a hypergraph, and specifies all item-related behaviors occurring in the conversation as vertices connected by the hyperedge, constructs a behavior habit hypergraph, and learns user-level behavior habit features based on the behavior habit hypergraph;

[0057] Feature fusion is used to fuse the conversation intention feature with the behavior habit feature to learn the user's real intention representation, and predict the items that the user may be interested in through the representation.

[0058] Beneficial Effects

[0059] First, compared with the existing technology that learns behavior dependencies at the behavior and project levels, the present invention learns user-level behavior dependencies by building a hypergraph to obtain behavior habits in the session. The information captured in this way is more comprehensive, interpretable, and can better reflect the user's intentions.

[0060] Second, the present invention aims to discover items that users may be interested in, rather than simply predicting items that may be interacted with under the target behavior. The present invention constructs a global graph to search the entire semantic space, combines behavioral information, and learns complex user behavior habits in order to more accurately capture user intentions and have a better ability to recommend new items to users. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0062] Figure 1 A schematic diagram of a conversation-based recommendation framework for behavioral habit enhancement intention according to the present invention. DETAILED DESCRIPTION

[0063] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0064] It should be noted that the following embodiments and features in the embodiments may be combined with each other in the absence of conflict; and, based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making any creative work are within the scope of protection of the present disclosure.

[0065] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.

[0066] The design concept of the present invention is: Figure 1 As shown in the figure, first, the semantic information at the project level and the behavior sequence pattern are learned to obtain the intention representation of behavior-aware fusion semantics. Specifically, a global directed graph is designed to extract the semantic relationship (neighborhood relationship) in the project space, and GRU is used to capture the behavior sequence relationship (i.e., behavior time feature) of each macro project to obtain the intention representation of behavior-aware fusion semantics; secondly, for the input session, a hypergraph is constructed to capture the behavior habits in the session, and the item representation is updated through the hypergraph neural network to enhance the user intention. Finally, the two types of mutual information are fused to obtain the final session representation.

[0067] like Figure 1 As shown, a conversation-based recommendation method for behavioral habit enhancement intention, the specific process is: the first step and the second step in the following steps are in parallel relationship, and there is no order of implementation. For ease of understanding, this embodiment records the acquisition of conversation intention features as the first step, and the acquisition of conversation behavior habit features as the second step.

[0068] The first step is to obtain the features of conversation intention

[0069] Construct a global item transition graph to extract semantic connections between items, aggregate item behavior sequences, and fuse semantic connections with behavior sequences through a gating mechanism to capture fine-grained user intent and further obtain conversational intent features of fine-grained behavior operations. The specific process of this step is as follows:

[0070] 1.1 For the items in the session, a global item transformation graph is constructed based on the semantic connections between items, and an item representation that integrates global high-order neighborhood information is obtained

[0071] Currently, most multi-action recommendation methods only encode the items in the current session, ignoring the global complex high-order item transformation relationship, which is conducive to learning the item embedding in the session. To utilize this information, this application constructs a global item interaction graph to aggregate global-level neighborhood information.

[0072] Define the project collection as The number of items is N. If in the session set Existence And satisfy v i →v j , which means that the user is in the same project as v i After interaction with item v j Interaction, Project v i With project v j There is a semantic connection between j Defined as v i The successor neighbors of all items can be obtained Therefore, the constructed global graph can be expressed as in Represents the global set of interaction edges. It is project v i The set of subsequent neighbors of . Based on G, a normalized adjacency matrix is ​​generated as follows:

[0073]

[0074] in, and g ij Respectively represent A g and elements in G.

[0075] A simplified graph convolutional layer is designed to globally encode items, the formula is as follows:

[0076]

[0077] In the formula, and denote the item embedding and parameter matrix of the lth layer respectively, d denotes the embedding dimension size, is the degree matrix. g After layer propagation, we get the item representation that integrates global high-order neighborhood information

[0078] 1.2 Item Representation Based on the Global High-Order Neighborhood Information Aggregate the behavior sequence of items and learn the macro-item representation X of the final intent of the session to capture fine-grained user intent;

[0079] For the original session We can get the reconstructed session s = [v1, v2, ..., v n ] and the behavior sequence H = [h1,h2,…,h n ], where h i =[b i,1 ,b i,2 ,…,b i,l ] indicates the macro project v i After obtaining the item embedding representation that aggregates global neighborhood information, a macro-item representation generator is designed to capture the personalized characteristics of each macro-item in the session.

[0080] To extract the macro project v i The behavior sequence characteristics of i It is sent to the gated neural unit (GRU) for learning, and the formula is as follows:

[0081]

[0082] in, and Respectively represent behavior b i,j The hidden state and embedding representation of Φ GRU are all the parameters of GRU. After GRU processing, h i The hidden state z of the last action in i,l Contains aggregate information of the entire behavior sequence. Therefore, it is considered as a macro project v i Behavioral characteristics of i In this way, the behavioral time characteristics of all macro items in the session can be obtained

[0083] Z=[z1,z2,...,z n ] (4)

[0084] Although sessions are anonymous, learning their items and behaviors from a structural and temporal perspective can still capture the unique preferences of each session. To achieve personalized recommendations, for each session, an adjacency matrix is ​​constructed Contains the in-degree and out-degree of the macro-item sequence in chronological order, respectively, where n is the length of the session containing the macro-item. At the same time, the item embedding is aggregated with the behavioral features and the information is propagated through these two adjacency matrices to capture the preference characteristics of the session:

[0085]

[0086] in, It is the aggregation of the behavior time characteristics of macro projects and project representation. It is based on all the macro items in the session from The sampled item embedding representation, is a connection operation. Then, the aggregated information is sent to the two adjacency matrices A respectively. in and A out Propagate through the two views to obtain a high-order fusion representation of the two views:

[0087] I in =A in W in P+b in

[0088] I out =A out W out P+b out (6)

[0089] in, and is a learnable parameter, Represents the high-order fusion representation of macro items in two views. After local neighborhood propagation, the intention representation of the items hidden by the user in his behavior is extracted. However, the extracted information may contain noise. In order to improve the performance and generalization ability of the model, a gating mechanism is introduced to remove some unimportant information and realize the aggregation of multiple information. The calculation formula is as follows:

[0090]

[0091] U=σ(W zi I+W zh E)

[0092] R=σ(W ri I+W rh E)

[0093]

[0094]

[0095] in, Represents the fused intent representation that combines the two views. are all trainable parameters. σ(·) represents the sigmoid function, Represents an element-wise multiplication operator. U represents the update gate, which controls the size of the information update, and R is the reset gate that controls how much information needs to be retained. By using the gating mechanism to filter and update the item embedding and behavior intention information, the final macro item representation in the session is obtained.

[0096] 1.3 Based on the macro-item representation X of the final intent of the conversation, generate a conversation intent containing the item order information in the conversation, the global level neighborhood relationship and the fine-grained behavior operation within the item

[0097] In step 1.2, the behavior-aware semantic representation of each macro-item in the session is obtained, and the intent of the entire session can be learned based on this. Generally speaking, a user is always driven by one type of intent in a session. However, the intent of items within a session may be different, so there will inevitably be noise. In order to more accurately capture the intent representation of the session, the attention gating mechanism is used to determine the weight of each item in the session.

[0098] First, the intent prototype of the session is constructed. In order to aggregate all the item information in the session, an average pooling operation is used:

[0099]

[0100] in, is the intention prototype of the conversation, e i is the initialized item. Then, a gated attention layer is designed to pass the conversation intent prototype into the item representation as follows:

[0101]

[0102]

[0103] Among them, α i For the macro project v in the session i The attention weight is is a learnable parameter. This gives the final fusion item representation. Due to the information transfer mechanism of the attention mechanism, the last item in the session contains information from all previous items. Therefore, using the last macro item v i As the overall intent of the conversation:

[0104]

[0105] in, An intent representation for a session that integrates the sequential information of items in the session, global-level neighborhood relations, and fine-grained behavioral operations within items.

[0106] Step 2: Acquisition of conversational behavior habits

[0107] The conversation items are designated as hyperedges in the hypergraph, and all the behaviors related to the items in the conversation are designated as vertices connected by the hyperedges, so as to construct a habit hypergraph, and learn the user-level behavior habit features based on the habit hypergraph. The specific process of this step is as follows:

[0108] The purpose of this step is to capture the complete behavior process of each item in a session to capture user-level behavior habits, and guide item representation learning based on this information to achieve user intent enhancement. In step 1.2, although GRU is used to learn the behavior intention corresponding to the macro-item, the learned information is still limited to the sequential pattern in the session and cannot capture the user's entire behavior habits. For example, the user first clicks, browses, and adds item v1 to the shopping cart, then finds items v2 and v3 that may be of interest, and browses them separately, and then he buys v1. In this way, the macro-item sequence corresponding to the user is [v1, v2, v3, v1]. The intention learning module in step 1.2 can only learn different behaviors of v1 at two time positions respectively, and cannot learn a complete behavior habit of v1. Therefore, in this step, user-level behavior habits are learned to achieve intent enhancement.

[0109] Since hypergraphs allow a hyperedge to connect more than two vertices, they can represent more complex relationships than general graph structures. For behavioral habits, a single item may correspond to multiple types of behaviors, which makes the hypergraph structure very suitable for capturing behavioral habits. The specific process is:

[0110] 2.1 Construct a behavior habit hypergraph for each session to capture all the item behavior habits in it, specify the session items as hyperedges in the hypergraph, and specify all the item-related behaviors that occurred in the session as vertices connected by the hyperedges. in represents the set of behavior types in the behavior space, ε h Represents the set of macro items in the session. The elements in the correlation matrix Q corresponding to the hypergraph are defined as follows:

[0111]

[0112] in, Indicates the item v in the session i The number of behaviors executed, β is the residual connection operation to prevent gradient explosion. The behavior habit hypergraphs generated by different users are not exactly the same, so the behavior preferences of individual users can be extracted from them.

[0113] 2.2 After obtaining the behavior habit hypergraph of the conversation, the hypergraph convolution method is used for message passing to integrate the behavior habits into the item representation. The hypergraph learning network is designed as follows:

[0114]

[0115] in, represents the macro-item embedding initialized in the session, D v and D e are the normalized degree matrices of vertex dimension and edge dimension respectively. After the hypergraph convolution processing, the item embedding representation E of integrated behavior habit information is obtained. h .

[0116] 2.3 Use the average pooling method to calculate the session representation:

[0117]

[0118] in, It refers to the final conversation representation after integrating behavioral habits into the project. For E h elements.

[0119] The third step is mutual information fusion

[0120] The conversation intention feature is integrated with the behavior habit feature to learn the user's real intention representation, and the items that the user may be interested in are predicted through the representation. The specific process of this step is:

[0121] In the framework, conversation information is encoded from two perspectives, namely, conversation intention and behavior habits. In order to generate the final conversation representation, a gated attention layer is designed to achieve mutual information fusion with weights. The specific implementation process is as follows:

[0122]

[0123]

[0124] in, and are all learnable parameters, and e s is the final session representation. In this way, the mutual information between the user’s intention and behavior habits is successfully integrated in the session, and this feature vector can be used for recommendation prediction.

[0125] Step 4: Model optimization and prediction

[0126] In order to make the prediction process more stable, L2 regularization is performed on the session representation and item embedding respectively, and the formula is as follows:

[0127]

[0128]

[0129] in, is the initialized global item embedding, then the final prediction result is calculated, and the cross entropy loss is used as the optimizer for model training:

[0130]

[0131]

[0132] in, and i Respectively represent v i The predicted probability and true value of the item.

[0133] The embodiment of the present application is a conversation-based recommendation system for enhancing intentions through behavioral habits, including: an intention learning module, a behavioral habit learning module, and a feature fusion module;

[0134] The intent learning module is used to construct a global item transition graph to extract semantic connections between items, aggregate item behavior sequences, and fuse semantic connections with behavior sequences through a gating mechanism to capture fine-grained user intent, and further obtain conversational intent features of fine-grained behavior operations based on the user intent;

[0135] A behavior habit learning module is used to designate a conversation item as a hyperedge in a hypergraph, and designate all item-related behaviors occurring in the conversation as vertices connected by the hyperedge, construct a habit hypergraph, and learn user-level behavior habit features based on the habit hypergraph;

[0136] Feature fusion is used to fuse the conversation intention feature with the behavior habit feature to learn the user's real intention representation, and predict the items that the user may be interested in through the representation.

[0137] This embodiment focuses on implementing the recommendation of the next item that the user may be interested in, and designs an intent learning module for global semantic fusion behavior. This module can effectively integrate the global item transformation relationship and internal behavior timing information, and further learn the behavior-aware semantic relationship between items. Technically, we construct a global item interaction graph to extract the semantic connection between items. Then, the behavior sequence under the macro item is aggregated, and the semantic connection is fused with the behavior information through a gating mechanism to capture the fine-grained user intent. At the same time, the coarse-grained intent representation contained in the macro item sequence in the session is captured, aggregated with the fine-grained intent, and the final intent representation of the user is learned.

[0138] This embodiment uses a hypergraph to capture user behavior habits and learns behavior dependencies at the user level; specifically, macro items are designated as hyperedges in the hypergraph, and all item-related behaviors that occur in the session are constructed as vertices connected by the hyperedge. This technology can extract user behavior habits from the session, transcending time and quantity limitations, thereby enhancing user intentions and reducing noise in behavior sequences. At the same time, in order to offset the challenges of gradient explosion and over-smoothing, residual connections are added when constructing the hypergraph.

[0139] In this embodiment, the attention gating mechanism is used to adaptively fuse the conversation representations learned in the above two modules to learn the user's true intention representation, and predict the items that the user may be interested in through the representation.

[0140] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A conversation-based recommendation method for behavioral habit enhancement intention, characterized in that: The specific process is: Conversational intent feature acquisition: A global item transition graph is constructed to extract semantic connections between items, item behavior sequences are aggregated using gated recurrent units, and semantic connections are fused with behavior sequences through a gating mechanism to capture fine-grained user intent. Conversational intent features of fine-grained behavior operations are further acquired based on the user intent. Acquisition of conversation behavior habit features: specifying conversation items as hyperedges in a hypergraph, and specifying all item-related behaviors that occurred in the conversation as vertices connected by the hyperedge, constructing a behavior habit hypergraph, and learning user-level behavior habit features based on the behavior habit hypergraph; Session feature fusion: The session intention feature is fused with the behavior habit feature to learn the user's true intention representation, and the items that the user may be interested in are predicted through the representation.

2. The conversation-based recommendation method for enhancing behavioral habit intention according to claim 1, characterized in that: The specific process of acquiring the conversation intention feature is as follows: 1.1 For the items in the session, a global item transformation graph is constructed based on the semantic connections between items to obtain the item representation of global high-order neighborhood information 1.2 Item Representation Based on the Global High-Order Neighborhood Information Aggregate the behavior sequence of items and learn the macro-item representation X of the final intent of the session to capture fine-grained user intent; 1.3 Based on the macro-item representation X of the final intent of the conversation, generate a conversation intent containing the item order information in the conversation, the global level neighborhood relationship and the fine-grained behavior operation within the item 3. The conversation-based recommendation method for enhancing behavioral habit intention according to claim 2, characterized in that: The specific process of step 1.1 is as follows: First, based on the semantic connection between conversation items, we find the i The set of subsequent neighbors Construct a global project conversion graph G; generate a normalized adjacency matrix A based on the global project conversion graph G g ; Secondly, based on the normalized adjacency matrix A g , using graph convolutional layers to globally encode items, in L g After layer propagation, we get the item representation that integrates global high-order neighborhood information 4. The conversation-based recommendation method for enhancing behavioral habit intention according to claim 3 is characterized in that: The specific process of step 1.2 is as follows: First, for each item v in the session i The corresponding behavior sequence h i Send it to the gated recurrent unit (GRU) to learn, and take the hidden state of the last behavior as the item v i The behavior time characteristics of z i , and then we get the behavior time characteristics of all items Z = [z1, z2, …, z n ]; Secondly, the item representation from the global high-order neighborhood information The sampled items are embedded into representation E, which is connected with the temporal features Z to achieve the aggregation of the behavior temporal features and the target representation; Again, the aggregated information is fed into two adjacency matrices A respectively. in and A out Propagate in and obtain a high-order fusion of the two views I in , I out Finally, for the high-order fusion of the two views I in , I out , by filtering and updating the item embedding and behavioral intention information using a gating mechanism, the final macro-item representation X in the session is obtained.

5. The conversation-based recommendation method for enhancing behavioral habit intention according to claim 4, characterized in that: The specific process of step 1.3 is as follows: Constructed a conversational intent prototype p ; Among them, e i Indicates the initialization i tem ; Based on the conversation intent prototype, the gated attention layer is used to pass the conversation intent prototype into the item representation as shown below: Among them, α i For the macro project v in the session i The attention weight is is a learnable parameter, d is the dimension of the embedding vector, x i is an element of X; Use the last macro item v i As the overall intent of the conversation: in, An intent representation for a session that integrates the sequential information of items in the session, global-level neighborhood relations, and fine-grained behavioral operations within items.

6. The conversation-based recommendation method for enhancing behavioral habit intention according to claim 2, characterized in that: The specific process of acquiring the conversation behavior habit features is as follows: 2.1 Assign the conversation items as hyperedges in the hypergraph, and assign all the item-related behaviors that occurred in the conversation as the vertices connected by the hyperedges, and construct a behavior habit hypergraph G for each conversation. h , and residual connections are added when constructing the hypergraph; 2.2 Hypergraph G for the behavioral habits h , using the hypergraph convolution method for message passing, integrating behavioral habits into item representation, and obtaining the hypergraph learning network E h ; 2.3 Perform average pooling on the hypergraph learning network to obtain the final session representation after integrating the behavioral habits into the project 7. The conversation-based recommendation method for enhancing behavioral habit intention according to claim 1, characterized in that: The session feature fusion is: using a gated attention layer to achieve mutual information fusion with weights.

8. The conversation-based recommendation method for enhancing behavioral habit intention according to claim 7, characterized in that: The session feature fusion is: Among them, W s and b s is the learning parameter, e s is the final session representation, σ(·) represents the sigmoid function, Represents the element-wise multiplication operator.

9. The conversation-based recommendation method for enhancing behavioral habit intention according to claim 1, characterized in that: Also includes optimization and prediction: L2 regularization is performed on the session representation and item embedding, the formula is as follows: Among them, E g is the initialized global project embedding; The final prediction results are calculated, and cross entropy loss is used as the optimizer for model training: in, and i Respectively represent v i The predicted probability and true value of the item, Represents the loss function of the model.

10. A conversation-based recommendation system with behavioral habit enhancement intention, characterized in that: Including intention learning module, behavior habit learning module and feature fusion module; The intent learning module is used to construct a global item transition graph to extract semantic connections between items, aggregate item behavior sequences, and fuse semantic connections with behavior sequences through a gating mechanism to capture fine-grained user intent, and further obtain conversational intent features of fine-grained behavior operations based on the user intent; A behavior habit learning module specifies a conversation item as a hyperedge in a hypergraph, and specifies all item-related behaviors occurring in the conversation as vertices connected by the hyperedge, constructs a behavior habit hypergraph, and learns user-level behavior habit features based on the behavior habit hypergraph; feature Fusion is used to fuse the conversation intention features with the behavior habit features to learn the user's true intention representation, and predict the items that the user may be interested in through the representation.

Images