CN119180288A - Session-based recommendation method and system for behavior habit enhancement intention - Google Patents

Session-based recommendation method and system for behavior habit enhancement intention Download PDF

Info

Publication number
CN119180288A
CN119180288A CN202411213221.7A CN202411213221A CN119180288A CN 119180288 A CN119180288 A CN 119180288A CN 202411213221 A CN202411213221 A CN 202411213221A CN 119180288 A CN119180288 A CN 119180288A
Authority
CN
China
Prior art keywords
behavior
item
conversation
session
representation
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202411213221.7A
Other languages
Chinese (zh)
Inventor
薛文皓
秦志达
张颢耀
杨诗晓
杜恩俊
田兴博
李爽
黄天羽
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Institute of Technology BIT
Original Assignee
Beijing Institute of Technology BIT
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Institute of Technology BIT filed Critical Beijing Institute of Technology BIT
Priority to CN202411213221.7A priority Critical patent/CN119180288A/en
Publication of CN119180288A publication Critical patent/CN119180288A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G06F40/35—Discourse or dialogue representation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Machine Translation (AREA)

Abstract

The invention belongs to the technical field of recommendation systems and provides a session-based recommendation method for behavior habit enhancement intentions, which comprises the following steps of acquiring session intention characteristics, namely constructing a global project transformation diagram to extract semantic connections among projects, aggregating project behavior sequences, fusing the semantic connections with the behavior sequences through a gating mechanism to capture fine-grained user intentions, and acquiring the session intention characteristics of fine-grained behavior operation based on the user intentions; the method comprises the steps of obtaining conversation behavior habit characteristics, namely designating conversation items as supersides in a supergraph, designating all behaviors related to the items, which occur in the conversation, as vertices connected with the supersides, constructing a behavior habit supergraph, and learning user-level behavior habit characteristics based on the behavior habit supergraph, and fusing the conversation intention characteristics and the behavior habit characteristics to learn real intention representation of a user and predict the items possibly interested by the user through the representation.

Description

Session-based recommendation method and system for behavior habit enhancement intention
Technical Field
The invention belongs to the technical field of recommendation systems, and relates to a session-based recommendation method and system for behavior habit enhancement intentions.
Background
Session-based recommendation algorithms aim to make predictions of next items for anonymous, time-sequential sessions. Due to the increasing sophistication of user privacy protection in recent years, session-based recommendation systems have gradually become a popular research area. Conventional approaches utilize markov chains to encode the session. But these methods can only predict the latter item from the former item, which results in underutilization of the overall sequence information of the session. To address this problem, some methods utilize Recurrent Neural Networks (RNNs) to model sessions and capture temporal relationships in sequences. In recent years, because of the advantages of Graph Neural Networks (GNNs) in capturing structural features, many methods introduce GNNs into the feature learning process, achieving good results.
However, the above method only considers macroscopic association between items, and does not utilize complex behavior 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 products that they have added to a shopping cart or have purchased than products that the user has only browsed. Therefore, considering behavior information can better capture the user's intention. Recently, some methods introduce multiple behavior information into a session recommendation. TGT designed a time graph Transformer to capture dynamic user-project interaction patterns by exploring evolving correlations between behaviors. MBHT a multi-scale transducer is designed to learn the behavior order pattern of the sequence and to learn cross behavior dependencies using hypergraphs. EMBSR devise multiple graphs to aggregate micro-operations in a sequence and learn binary patterns of behavior with extended attention mechanisms.
Although existing approaches achieve good results in terms of the relationship between sequential representation of aggregated behavior information and mining behavior dependencies, there are still some problems to be solved:
First, the behavior habits at the user level cannot be learned. Current multi-behavior based session recommendation methods are limited to learning behavior dependencies at the behavior or project level, but ignore user-level complex behavior dependencies. The behavior-level dependent learning method only models a single behavior, and is difficult to process complex multi-behavior scenes. Item level methods, while capable of capturing cross behavior dependencies, typically only learn binary behavior dependencies or continuous behavior under continuous time stamps, cannot break through the time constraints. This limitation means that they cannot capture the full behavioral habits of the user for the item. Existing methods do not fully exploit the user-level dependencies, which is critical to accurately reflect the user's intent.
Second, in a session-based recommendation system, the task of predicting items under a target behavior is unreasonable. Current multi-behavior based session recommendation system methods typically define a target behavior and consider other behaviors as auxiliary behaviors focusing on predicting the next item under the target behavior. In reality, however, the user often performs auxiliary actions before performing the target action. This approach may cause the system to recommend items that the user has interacted with, rather than new items, which contradicts the purpose of the recommendation system and may reduce the user experience.
Disclosure of Invention
In view of the above, the invention provides a session-based recommendation method and system for behavior habit enhancement intention, which can accurately predict an item of interest to a user by fusing session intention characteristics and session behavior habit characteristics to learn the actual intention representation of the user.
The technical scheme for realizing the invention is as follows:
A session-based recommendation method for behavior habit enhancement intention comprises the following specific processes:
Constructing a global project transformation graph to extract semantic connection among projects, aggregating project behavior sequences by using a gating circulating unit (GRU), and fusing the semantic connection with the behavior sequences through a gating mechanism to capture fine-grained user intention, and further acquiring the conversation intention characteristic of fine-grained behavior operation based on the user intention;
The method comprises the steps of obtaining conversation behavior habit characteristics, namely designating conversation items as supersides in a supergraph, designating all behaviors related to the items, which occur in the conversation, as vertexes connected with the supersides, constructing a behavior habit supergraph, and learning user-level behavior habit characteristics based on the behavior habit supergraph;
Session feature fusion-fusion of the session intention feature with the behavioral habit feature to learn a representation of the user's actual intention and predict items of likely interest to the user from the representation.
Further, the specific process of obtaining the session intention characteristic in the invention is as follows:
1.1, constructing a global project transformation graph aiming at projects in a session based on semantic connection among the projects to obtain project representations of global high-order neighborhood information
1.2 Project representations based on the Global higher order neighborhood informationAggregating the behavior sequence of the items, learning the macroscopic item representation X of the final intention of the session, and capturing the intention of the fine-grained user;
1.3 generating a Session intention including in-Session project order information, global-level neighborhood relationships, and fine-grained behavioral operations within a project from a macroscopic project representation X of the Session final intention
Further, the specific process of step 1.1 of the present invention is as follows:
First, based on semantic connections between session items, find a set of subsequent neighbors for each item v i in the session Generating a normalized adjacency matrix A g according to the global project conversion diagram G;
Secondly, based on the normalized adjacency matrix A g, performing global coding on the project by using a graph convolution layer, and after the L g layer is propagated, obtaining the project representation integrated with global high-order neighborhood information
Further, the specific process of step 1.2 of the present invention is as follows:
Firstly, sending a behavior sequence h i corresponding to each item v i in a session to a gate-controlled circulation unit (GRU) for learning, taking the hidden state of the last behavior as a behavior time characteristic Z i of the item v i, and further obtaining behavior time characteristics z= [ Z 1,z2,…,zn ] of all the items;
second, from the item representation of the global high-order neighborhood information The sampling item is embedded into a representation E, is connected with the time feature Z, and achieves aggregation of the behavior time feature and the target representation;
Thirdly, the aggregated information is respectively sent into two adjacent matrixes A in and A out for propagation, and a high-order fusion I of two views is obtained in,Iout
Finally, for the high-order fusion I in,Iout of the two views, the final macroscopic item representation X in the session is obtained by filtering and updating the item embedding and behavior intention information by using a gating mechanism.
Further, the specific process of step 1.3 of the present invention is as follows:
An intention prototype e p of the session is constructed;
Wherein e i represents initialized item;
based on the conversation intention prototype, the conversation intention prototype is transferred into the item representation by using a gating attention layer as follows:
where α i is the attention weight of macroscopic item v i in the session, an As a learnable parameter, d is the dimension of the embedded vector, and X i is an element of X;
the representation of the last macro item v i is used as the overall intent representation of the session:
Wherein, Representing an intended representation of a conversation that integrates sequence information of items in the conversation, global level neighborhood relationships, and fine-grained behavioral operations within the items.
Further, the specific process of acquiring the session behavior habit characteristics is as follows:
2.1, designating conversation items as hyperedges in hypergraphs, designating all actions related to the items, which occur in the conversation, as vertexes connected with the hyperedges, constructing a behavior habit hypergraph G h for each conversation, and adding residual connection when constructing hypergraphs;
2.2, aiming at the behavior habit hypergraph G h, carrying out message transmission by using a hypergraph convolution method, and integrating the behavior habit into a project representation to obtain a hypergraph learning network E h;
2.3 carrying out average pooling treatment on the hypergraph learning network to obtain a fused final session representation after the behavior habit is integrated into the project
Furthermore, the session information fusion is realized by adopting a gating attention layer to realize mutual information fusion with weight.
Further, the session information fusion of the present invention is:
Where W s and b s are learned parameters, e s is the final session representation, σ (·) represents the sigmoid function, Representing element-level multiplication operators.
Further, the invention also comprises optimizing and predicting:
L2 regularization is performed on the session representation and the item embedding, and the formula is as follows:
Wherein E g is the initialized global item embedding;
the final prediction result is calculated, and cross entropy loss is used as an optimizer for model training:
Wherein, And y i represent the probability and the true value of the v i term being predicted respectively,Representing the loss function of the model.
On the other hand, the session-based recommendation system for behavior habit enhancement intention comprises an intention learning module, a behavior habit learning module and a feature fusion module;
The intention learning module is used for constructing a global project transformation diagram to extract semantic connection among projects, aggregating project behavior sequences, fusing the semantic connection with the behavior sequences through a gating mechanism to capture fine-grained user intention, and further acquiring conversation intention characteristics of fine-grained behavior operation based on the user intention;
The behavior habit learning module designates a conversation item as a hyperedge in the hypergraph, designates all behaviors related to the item, which occur in the conversation, as vertexes connected with the hyperedge, constructs a behavior habit hypergraph, and learns user-level behavior habit characteristics based on the behavior habit hypergraph;
And the feature fusion is used for fusing the conversation intention feature and the behavior habit feature to learn the real intention representation of the user and predicting the items possibly interested by the user through the representation.
Advantageous effects
First, compared with the prior art of learning behavior dependence on the behavior and project level, the invention learns the behavior dependence relationship at the user level by constructing a hypergraph to acquire behavior habits in a session. The captured information is more comprehensive and has interpretability, and the intention of the user can be reflected.
Second, the present invention aims to discover items that may be of interest to a user, rather than simply predicting items that may interact under a target behavior. The invention constructs a global map to search the whole semantic space, and learns complex user behavior habits by combining behavior information so as to more accurately capture the intention of the user, and has better capability of recommending new articles to the user.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and that other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic diagram of a session-based recommendation framework for behavioral habit enhancement intents of the present invention.
Detailed Description
Embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
It is noted that the following embodiments and features of the embodiments may be combined with each other without conflict, and that all other embodiments obtained by persons of ordinary skill in the art without creative efforts based on the embodiments in the present disclosure are within the scope of protection of the present disclosure.
It is noted that various aspects of the embodiments are described below within the scope of the following claims. It should be apparent that the aspects described herein may be embodied in a wide variety of forms and that any specific structure and/or function described herein is merely illustrative. Based on the present disclosure, one skilled in the art will appreciate that one aspect described herein may be implemented independently of any other aspect, and that two or more of these aspects may be combined in various ways. For example, an apparatus may be implemented and/or a method practiced using any number of the aspects set forth herein. In addition, such apparatus may be implemented and/or such methods practiced using other structure and/or functionality in addition to one or more of the aspects set forth herein.
The design concept of the invention is that, as shown in figure 1, firstly, semantic information and a behavior sequence mode of a project level are learned to obtain an intention representation of behavior perception fusion semantics, specifically, a global directed graph is designed to extract semantic relations (neighbor relations) in a project space, GRU is utilized to capture the behavior sequence relation (i.e. behavior time characteristics) of each macroscopic project to obtain the intention representation of the behavior perception fusion semantics, and secondly, a hypergraph is constructed for an input session to capture behavior habits in the session, and the project representation is updated through a hypergraph neural network to enhance the intention of a user. Finally, the two types of mutual information are fused, and a final session representation form is obtained.
As shown in FIG. 1, a session-based recommendation method for behavior habit enhancement intention specifically includes the steps of enabling a first step and a second step to be in parallel relationship, and enabling no implementation sequence to exist.
First step, session intention feature acquisition
Constructing a global project conversion diagram to extract semantic connection among projects, aggregating project behavior sequences, fusing the semantic connection with the behavior sequences through a gating mechanism to capture fine-granularity user intention and further obtain conversation intention characteristics of fine-granularity behavior operation, wherein the specific process of the steps is as follows:
1.1 constructing a global project conversion map aiming at projects in a session based on semantic connection among the projects to obtain project representations integrated with global high-order neighborhood information
At present, most multi-behavior recommendation methods only code items in a current session, and globally complex high-order item conversion relations are ignored, and the high-order relations are beneficial to item embedding in a learning session. To take advantage of this information, the present application constructs a global project interaction graph to aggregate neighborhood information at the global level.
Defining a set of items asThe number of items is N. If in session setIs present in (a)And v i→vj is satisfied, which means that the user interacts with item v j after interacting with item v i, there is a semantic connection between item v i and item v j. V j is defined as the successor neighbor of v i so that a successor neighbor set of all items can be obtainedThe global graph thus constructed can be expressed asWherein the method comprises the steps ofRepresenting a global set of interacting edges.Is the subsequent neighbor set for item v i. Generating a normalized adjacency matrix based on GThe following are provided:
Wherein, And G ij represents the elements in a g and G, respectively.
A simplified picture convolution layer is designed to globally code an item as follows:
In the formula, AndRepresenting the item embedding and parameter matrix of the first layer, respectively, d representing the embedding dimension size,Is a degree matrix. After the L g layer is propagated, the project expression integrated with the global high-order neighborhood information is obtained
1.2 Project representations based on the Global higher order neighborhood informationAggregating the behavior sequence of the items, learning the macroscopic item representation X of the final intention of the session, and capturing the intention of the fine-grained user;
For the original session A reconstructed session s= [ v 1,v2,…,vn ] and a behavior sequence h= [ H 1,h2,…,hn ] can be obtained, where H i＝[bi,1,bi,2,…,bi,l ] represents the behavior sequence of the macroscopic item v i. After obtaining the item embedded representation of the aggregated global neighborhood information, a macro item representation generator is designed to capture the personalized features of each macro item in the session.
In order to extract the behavior sequence characteristics of macroscopic item v i, the corresponding behavior sequence h i is sent into a gating neural unit (GRU) for learning, and the formula is as follows:
Wherein, AndRepresenting the hidden state and embedded representation of behavior b i,j, respectively, Φ GRU is all parameters of the GRU. After the GRU processing, the hidden state z i,l of the last behavior in h i contains the aggregate information of the whole behavior sequence. Thus, it is considered the behavioral characteristic z i of macro item v i. In this way, the behavioral temporal characteristics of all macro items in the session can be derived
Z=[z1,z2,...,zn] (4)
While the sessions are anonymous, learning their items and behavior from a structural and timing perspective can still capture the unique preferences of each session. To achieve personalized recommendations, for each session, an adjacency matrix is constructedThe time sequence of macro items is included in the time sequence of the in-degree and the time sequence of macro items is included out of the time sequence of macro items, wherein n is the length of the session including the macro items. Meanwhile, the item embedding and behavior characteristics are aggregated, and information is propagated through the two adjacency matrices to capture the preference characteristics of the session:
Wherein, Is a representation of aggregated behavior time features of macroscopic items with item representations,Is based on all macroscopic item slaves in the sessionThe sampled item is embedded into the representation,Is a join operation. The aggregated information is then fed into two adjacency matrices a in and a out, respectively, for propagation, resulting in a higher order fusion representation of the two views:
Iin＝AinWinP+bin
Iout＝AoutWoutP+bout (6)
Wherein, AndIs a parameter that can be learned and is,Representing a high-level fused representation of a macroscopic item in two views. After local neighborhood propagation, the intent representation of the item that the user has hidden in its behavior is extracted. However, the extracted information may contain noise. In order to improve the performance and generalization capability of the model, a gating mechanism is introduced to remove some unimportant information, so that aggregation of a plurality of information is realized. The calculation formula is as follows:
U=σ(WziI+WzhE)
R=σ(WriI+WrhE)
Wherein, The representation incorporates a fused intent representation of the two views. Are trainable parameters. Sigma (·) represents a sigmoid function,Representing element-level multiplication operators. U denotes an update gate, controlling the size of the information update, and R is a reset gate controlling how much information needs to be retained. Filtering and updating item embedding and behavior intention information by using a gating mechanism to obtain final macroscopic item representation in a session
1.3 Generating a Session intention including in-Session project order information, global-level neighborhood relationships, and fine-grained behavioral operations within a project from a macroscopic project representation X of the Session final intention
In step 1.2 a behavior-aware semantic representation of each macroscopic item in the session will be obtained and the intent of the whole session can be learned on this basis. Generally, users are always driven by one type of intent in one session. However, the intent of items within a session may be different, so there is inevitably noise. To more accurately capture an intended representation of a conversation, an attention gating mechanism is utilized to determine the weight of each item in the conversation.
First, an intention prototype of a session is constructed. To aggregate all item information in a session, an average pooling operation is employed:
Wherein, Is the intent prototype of the session and e i is the initialized item. Then, a gating attention layer is designed to transfer the conversation intention prototype into the project representation as follows:
where α i is the attention weight of macroscopic item v i in the session, an Is a learnable parameter. This results in a final fusion item representation. The last item in the session contains information from all previous items due to the information transfer mechanism of the attention mechanism. Thus, the representation of the last macro item v i is used as the overall intent representation of the session:
Wherein, Representing an intended representation of a conversation that integrates sequence information of items in the conversation, global level neighborhood relationships, and fine-grained behavioral operations within the items.
Second step, session behavior habit acquisition
The conversation item is designated as a hyperedge in the hypergraph, all the behaviors related to the item, which occur in the conversation, are designated as vertexes connected with the hyperedge, the habit hypergraph is constructed, and the habit characteristics of the user-level behaviors are learned based on the habit hypergraph. The specific process of the step is as follows:
The purpose of this step is to capture the complete course of behavior for each item in a session to capture user-level behavior habits and to guide the item representation learning based on these information to achieve enhancement of the user's intent. In step 1.2, while the GRU is used to learn the behavioral intention corresponding to the macro item, the learned information is still limited to the sequential pattern in the session, and the entire behavioral habit of the user cannot be captured. For example, the user first clicks, browses and adds item v 1 to the shopping cart, then finds items v 2 and v 3 of possible interest and browses them separately, then he purchases v 1. Thus, the user corresponds to a macro item sequence v 1,v2,v3,v1. Step 1.2 the intention learning module can only learn the different behaviors of v 1 at two time positions respectively, and cannot learn a complete behavior habit of v 1. Thus in this step, the behavior habit of the user level is learned to achieve enhancement of the intention.
Because hypergraphs allow one hyperedge to connect more than two vertices, they can represent a more complex relationship than a typical graph structure. Whereas for behavior habits, a single item may correspond to multiple types of behavior, which makes the hypergraph structure well suited for capturing behavior habits. The specific process is as follows:
2.1 building a behavior habit hypergraph for each session to capture all project behavior habits therein, designating a session project as a hyperedge in the hypergraph, and designating all project-related behaviors occurring in the session as vertices to which the hyperedge is connected. For constructed hypergraphs Wherein the method comprises the steps ofRepresenting a set of behavior types in a behavior space, epsilon h representing a set of macroscopic items in a session. The elements in the correlation matrix Q corresponding to the hypergraph are defined as follows:
Wherein, Representing the number of actions performed on item v i in the session, β is the residual join operation to prevent gradient explosion. The behavior habit hypergraphs generated by different users are not identical, so that the behavior preferences of the individual users can be extracted from the hypergraphs.
2.2 After obtaining the behavior habit hypergraph of the session, using a hypergraph convolution method to transfer the information, and integrating the behavior habit into the project representation. The hypergraph learning network is designed as follows:
Wherein, Representing the macroscopic item embedding initialized in the session, D v and D e are respectively normalized degree matrixes of vertex and edge dimensions, and the item embedding representation E h of the integrated behavior habit information is obtained after hypergraph convolution processing.
2.3 Computing session representations using an average pooling approach:
Wherein, Refers to the final session representation that merges after the behavioral habits are integrated into the project,Is an element of E h.
Third, mutual information fusion
The conversational intent features are fused with behavioral habit features to learn a representation of the user's actual intent and predict items that may be of interest to the user from the representation. The specific process of the step is as follows:
In the framework, session information, i.e., session intent and behavioral habits, is encoded from two perspectives. In order to generate the final session representation, a gating attention layer is designed to realize mutual information fusion with weights. The specific implementation process is as follows:
Wherein, AndAre both learnable parameters, and e s is the final session representation. In this way, mutual information between the user's intent and behavior habits is successfully integrated in the session, and recommendation predictions can be made using this feature vector.
Fourth step, model optimization and prediction
In order to make the prediction process more stable, the session representation and the item embedding are respectively subjected to L2 regularization, and the formula is as follows:
Wherein, Is an initialized global project embedding, then, the final prediction result is calculated, and cross entropy loss is adopted as an optimizer for model training:
Wherein, And y i represent the probability and the true value of the v i term being predicted, respectively.
The embodiment of the application relates to a session-based recommendation system for behavior habit enhancement intents, which comprises an intension learning module, a behavior habit learning module and a feature fusion module;
The intention learning module is used for constructing a global project transformation diagram to extract semantic connection among projects, aggregating project behavior sequences, fusing the semantic connection with the behavior sequences through a gating mechanism to capture fine-grained user intention, and further acquiring conversation intention characteristics of fine-grained behavior operation based on the user intention;
The behavior habit learning module is used for designating the conversation item as a hyperedge in the hypergraph, designating all behaviors related to the item, which occur in the conversation, as vertexes connected with the hyperedge, constructing a habit hypergraph, and learning user-level behavior habit characteristics based on the habit hypergraph;
And the feature fusion is used for fusing the conversation intention feature and the behavior habit feature to learn the real intention representation of the user and predicting the items possibly interested by the user through the representation.
The embodiment focuses on realizing recommendation of the next item possibly interested by the user, and designs an intention learning module aiming at global semantic fusion behaviors. The module can effectively integrate the global project conversion relation and the internal behavior time sequence information, and further learn the behavior perception semantic relation among projects. Technically, we construct a global project interaction graph to extract semantic connections between projects. Then, behavior sequences under macroscopic items are aggregated, and semantic connections are fused with behavior information through a gating mechanism to capture fine-grained user intent. At the same time, coarse-grained intent representations included in the macro-item sequence in the session are captured, aggregated with fine-grained intent, and the final intent representation of the user is learned.
The embodiment captures behavior habits of a user by utilizing a hypergraph and learns behavior dependency relationships at a user level, and specifically designates a macroscopic item as a hyperedge in the hypergraph and constructs all the behaviors related to the item, which occur in a session, as vertices to which the hyperedge is connected. The technology can extract the behavior habit of the user from the session and surpass the limit of time and quantity, thereby enhancing the intention of the user and reducing the noise in the behavior sequence. Meanwhile, to offset the challenges of gradient explosion and excessive smoothing, a residual connection is added when constructing hypergraphs.
In this embodiment, the session representations learned in the above two modules are adaptively fused by using an attention gating mechanism, so as to learn the actual intention representation of the user, and predict the items possibly of interest to the user through the representation.
The foregoing is merely illustrative of the present invention, and the present invention is not limited thereto, and any changes or substitutions easily contemplated by those skilled in the art within the scope of the present invention should be included in the present invention. Therefore, the protection scope of the invention is subject to the protection scope of the claims.

Claims (10)

1.一种行为习惯增强意图的基于会话的推荐方法，其特征在于，具体过程为：1. A conversation-based recommendation method for behavior habit enhancement intention, characterized in that the specific process is as follows: 会话意图特征获取：构造一个全局项目转换图来提取项目之间的语义连接，利用门控循环单元聚合项目行为序列，并通过门控机制将语义连接与行为序列融合，以捕获细粒度的用户意图，进一步基于所述用户意图获取细粒度行为操作的会话意图特征；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. 会话行为习惯特征获取：将会话项目指定为超图中的超边，并将包含在会话中发生的所有与项目相关的行为指定为该超边所连接的顶点，构造行为习惯超图，基于所述行为习惯超图学习用户级行为习惯特征；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; 会话特征融合：将所述会话意图特征与行为习惯特征进行融合以学习用户真实的意图表示，并通过该表示来预测用户可能感兴趣的项目。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.根据权利要求1所述行为习惯增强意图的基于会话的推荐方法，其特征在于，所述会话意图特征获取的具体过程为：2. According to the conversation-based recommendation method for enhancing behavioral habit intentions according to claim 1, the specific process of acquiring the conversation intention features is as follows: 1.1针对会话中的项目，基于项目之间语义连接，进行全局项目转换图构建，得到全局高阶邻域信息的项目表示 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基于所述全局高阶邻域信息的项目表示聚合项目的行为序列，学习会话最终意图的宏观项目表示X，实现细粒度用户意图的捕获；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根据所述会话最终意图的宏观项目表示X，生成包含会话中项目顺序信息、全局级邻域关系和项目内的细粒度行为操作的会话意图 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.根据权利要求2所述行为习惯增强意图的基于会话的推荐方法，其特征在于，所述步骤1.1的具体过程为：3. According to the conversation-based recommendation method for behavior habit enhancement intention according to claim 2, it is characterized in that the specific process of step 1.1 is: 首先，基于会话项目之间语义连接，寻找会话中每一项目vi的后续邻居集合构造全局项目转换图G；根据所述全局项目转换图G，生成归一化邻接矩阵Ag；First, based on the semantic connections between session items, we find the set of subsequent neighbors of each item vi in the session. Constructing a global project conversion graph G; generating a normalized adjacency matrix Ag according to the global project conversion graph G; 其次，基于所述归一化邻接矩阵Ag，利用图卷积层对项目进行全局编码，在Lg层传播后，得到了集成了全局高阶邻域信息的项目表示 Secondly, based on the normalized adjacency matrix Ag , the graph convolution layer is used to globally encode the items. After propagation through the Lg layer, the item representation that integrates the global high-order neighborhood information is obtained. 4.根据权利要求3所述行为习惯增强意图的基于会话的推荐方法，其特征在于，所述步骤1.2的具体过程为：4. According to the conversation-based recommendation method for behavior habit enhancement intention according to claim 3, it is characterized in that the specific process of step 1.2 is: 首先，针对会话中每一项目vi对应的行为序列hi送入门控循环单元(GRU)学习，将最后一个行为的隐藏状态作为项目vi的行为时间特征zi，进而得到所有项目的行为时间特征Z＝[z1,z2,…,zn]；First, the behavior sequence h i corresponding to each item vi in the session is sent to the gated recurrent unit (GRU) for learning, and the hidden state of the last behavior is used as the behavior time feature z i of item vi , and then the behavior time feature Z = [z 1 ,z 2 ,…,z n ] of all items is obtained; 其次，从所述全局高阶邻域信息的项目表示采样项目嵌入表示E，连接所述时间特征Z，实现行为时间特征与目标表示的聚合；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; 再次，将聚合的信息分别送入两个邻接矩阵Ain和Aout中进行传播，得到两个视图的高阶融合Iin，Iout Again, the aggregated information is sent to two adjacency matrices A in and A out for propagation, and the high-order fusion of the two views I in , I out is obtained. 最后，针对所述两个视图的高阶融合Iin，Iout，通过使用门控机制对项目嵌入和行为意图信息进行过滤和更新，得到了会话中最终的宏观项目表示X。Finally, for the high-order fusion of the two views I in , I out , the final macro-item representation X in the session is obtained by filtering and updating the item embedding and behavior intention information using a gating mechanism. 5.根据权利要求4所述行为习惯增强意图的基于会话的推荐方法，其特征在于，所述步骤1.3的具体过程为：5. According to the conversation-based recommendation method for behavior habit enhancement intention according to claim 4, it is characterized in that the specific process of step 1.3 is: 构建了会话的意图原型ep；Constructed the conversation intention prototype e p ; 其中，ei表示初始化的item；Among them, e i represents the initialized item ; 基于所述会话意图原型，利用门控注意层，将会话意图原型传递到项目表示中，如下所示：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: 其中，αi为会话中宏观项目vi的注意力权重，而为可学习参数，d为嵌入向量的维度，xi为X的元素；Among them, αi is the attention weight of the macro item vi in the session, and is a learnable parameter, d is the dimension of the embedding vector, and xi is the element of X; 使用最后一个宏观项目vi的表示作为会话的总体意图表示：Use the representation of the last macro-item vi as the overall intent representation of the session: 其中，表示会话的意图表示，它集成了会话中项目的顺序信息、全局级邻域关系和项目内的细粒度行为操作。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.根据权利要求2所述行为习惯增强意图的基于会话的推荐方法，其特征在于，所述会话行为习惯特征获取的具体过程为：6. According to the conversation-based recommendation method for behavior habit enhancement intention according to claim 2, it is characterized in that the specific process of obtaining the conversation behavior habit features is: 2.1将会话项目指定为超图中的超边，并将包含在会话中发生的所有与项目相关的行为指定为该超边所连接的顶点，为每一会话构建一个行为习惯超图Gh，且构造超图时添加了残差连接；2.1 The conversation items are designated as hyperedges in the hypergraph, and all the behaviors related to the items in the conversation are designated as the vertices connected by the hyperedges. A behavior habit hypergraph G h is constructed for each conversation, and residual connections are added when constructing the hypergraph. 2.2针对所述行为习惯超图Gh，利用超图卷积的方法进行消息传递，将行为习惯集成到项目表示中，得到超图学习网络Eh；2.2 For the behavior habit hypergraph G h , a hypergraph convolution method is used to perform message passing, and the behavior habit is integrated into the item representation to obtain a hypergraph learning network E h ; 2.3将所述超图学习网络进行平均池化处理，得到将行为习惯整合到项目中后融合的最终会话表示 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.根据权利要求1所述行为习惯增强意图的基于会话的推荐方法，其特征在于，所述会话特征融合为：采用门控注意层来带权重的实现相互信息融合。7. According to the conversation-based recommendation method for behavioral habit enhancement intention in claim 1, it is characterized in that the conversation feature fusion is: using a gated attention layer to realize mutual information fusion with weights. 8.根据权利要求7所述行为习惯增强意图的基于会话的推荐方法，其特征在于，所述会话特征融合为：8. The conversation-based recommendation method for enhancing behavioral habit intention according to claim 7, characterized in that the conversation feature fusion is: 其中，Ws和bs为学习的参数，es为最终的会话表示，σ(·)表示sigmoid函数，表示元素级乘法运算符。Where Ws and bs are the learned parameters, es is the final session representation, σ(·) represents the sigmoid function, Represents the element-wise multiplication operator. 9.根据权利要求1所述行为习惯增强意图的基于会话的推荐方法，其特征在于，还包括优化与预测：9. The conversation-based recommendation method for enhancing behavioral habit intention according to claim 1, characterized in that it also includes optimizing and predicting: 对会话表示和项目嵌入进行L2正则化，公式如下：L2 regularization is performed on the session representation and item embedding, the formula is as follows: 其中，Eg是初始化的全局项目嵌入；Where E g is the initialized global item embedding; 计算了最终的预测结果，并采用交叉熵损失作为模型训练的优化器：The final prediction results are calculated, and cross entropy loss is used as the optimizer for model training: 其中，和yi分别表示vi项被预测的概率和真实值，表示模型的损失函数。in, and yi represent the predicted probability and true value of item vi, respectively. Represents the loss function of the model. 10.一种行为习惯增强意图的基于会话的推荐系统，其特征在于，包括意图学习模块、行为习惯学习模块及特征融合模块；10. A conversation-based recommendation system for behavior habit enhancement intention, characterized by comprising an intention learning module, a behavior habit learning module and a feature fusion module; 意图学习模块，用于构造一个全局项目转换图来提取项目之间的语义连接，聚合项目行为序列，并通过门控机制将语义连接与行为序列融合，以捕获细粒度的用户意图，并基于所述用户意图进一步获取细粒度行为操作的会话意图特征；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; 行为习惯学习模块，将会话项目指定为超图中的超边，并将包含在会话中发生的所有与项目相关的行为指定为该超边所连接的顶点，构造行为习惯超图，基于所述行为习惯超图学习用户级行为习惯特征；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; 特征融合，用于将所述会话意图特征与行为习惯特征进行融合以学习用户真实的意图表示，并通过该表示来预测用户可能感兴趣的项目。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.
CN202411213221.7A 2024-08-30 2024-08-30 Session-based recommendation method and system for behavior habit enhancement intention Pending CN119180288A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202411213221.7A CN119180288A (en) 2024-08-30 2024-08-30 Session-based recommendation method and system for behavior habit enhancement intention

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202411213221.7A CN119180288A (en) 2024-08-30 2024-08-30 Session-based recommendation method and system for behavior habit enhancement intention

Publications (1)

Publication Number Publication Date
CN119180288A true CN119180288A (en) 2024-12-24

Family

ID=93898769

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202411213221.7A Pending CN119180288A (en) 2024-08-30 2024-08-30 Session-based recommendation method and system for behavior habit enhancement intention

Country Status (1)

Country Link
CN (1) CN119180288A (en)

Similar Documents

Publication Publication Date Title
CN113610610B (en) Conversational recommendation method and system based on graph neural network and review similarity
CN113918832B (en) Graph convolution collaborative filtering recommendation system based on social relationship
CN115292585B (en) A conversational item recommendation method based on context-aware graph neural networks
CN115114526A (en) A Weighted Graph Convolutional Network Score Prediction Recommendation Method with Multi-action Enhanced Information
CN115423016A (en) Training method of multi-task prediction model, multi-task prediction method and device
WO2024067884A1 (en) Data processing method and related apparatus
CN115203568A (en) A content recommendation method, related device and equipment based on a deep learning model
CN116383515B (en) Social recommendation method and system based on lightweight graph convolution network
CN111177460B (en) Method and device for extracting key frame
Zhang et al. GACOforRec: Session-based graph convolutional neural networks recommendation model
CN119478403A (en) A semantic segmentation method for large-scale remote sensing images based on ultra-long context modeling
CN118916731B (en) Live broadcast room crowd portrait generation method and system based on big data
CN116501991B (en) Recommendation method applied to anonymous login scene and based on click sequence time enhancement
Yang et al. SA‐FlowNet: Event‐based self‐attention optical flow estimation with spiking‐analogue neural networks
CN117390267A (en) Personalized multi-task enhanced recommendation model based on knowledge graph
CN115618125B (en) A Multi-behavior Recommendation Method and System Based on Multi-Graph Semantic Fusion
CN113449193B (en) Information recommendation method and device based on multi-classification images
CN119537705B (en) A multimodal recommendation method and system based on spectral polynomial filter
CN120386941A (en) Personalized learning content recommendation method and system based on neighborhood and hypergraph collaboration
CN119862424A (en) Multi-mode data processing method and device, model training method and device
CN119180288A (en) Session-based recommendation method and system for behavior habit enhancement intention
Zhang et al. Hybrid recommendation system with graph neural collaborative filtering and local self-attention mechanism
Chen et al. Improved ctr prediction algorithm based on lstm and attention
CN117077735A (en) Dimension-dependent integrated service quality prediction method based on convolutional neural network
CN116541613A (en) Item recommendation method of double-message propagation mechanism based on scoring weighting

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination