A conversation recommendation method based on cross-category heterogeneous hypergraph multi-intent representation

Abstract

This invention belongs to the field of conversation recommendation technology, and specifically provides a conversation recommendation method based on multi-intent expression using cross-category heterogeneous hypergraphs. The method includes preprocessed data and the following steps: S1: The preprocessed data is input to a category recognition processing module, which uses a Latent Dirichlet Allocation (LDA) model to mine the latent category distribution of items and outputs category-classified items; S2: The preprocessed data and category-classified items are respectively input to a self-loop star graph module and a cross-category heterogeneous hypergraph module to construct the graph structure, outputting a self-loop star graph and a cross-category heterogeneous hypergraph. This invention combines the conversational intra-conversational structure modeling of self-loop star graphs with the semantic association expression of cross-category heterogeneous hypergraphs to construct a multi-intent modeling framework that can simultaneously capture local behavioral dependencies and global interest transfers. It dynamically identifies latent intent patterns and multi-category preference features even without long-term user history information, thereby significantly improving the accuracy, personalization, and robustness of conversation recommendations.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of conversation recommendation system, and particularly relates to a conversation recommendation method based on cross-category heterogeneous hypergraph multi-intention expression. BACKGROUND

[0002] With the rapid development of e-commerce, social media and content recommendation platforms, users' interaction behaviors in different scenarios are increasingly frequent and fragmented. Therefore, the session-based recommendation (SBR) technology emerges as the times require. This technology aims to predict users' next interest preferences based only on their current session sequence information (such as browsing, clicking, collecting, purchasing, etc.), so as to provide personalized recommendations in the absence of long-term historical data or in the case of anonymous users. Traditional session recommendation methods are mostly based on sequence modeling (such as RNN, GRU, Transformer, etc.) or graph neural networks (GNN) for modeling, which can effectively mine the temporal dependencies and local transfer relationships between items within the session. However, these methods still have many limitations when dealing with complex cross-category and multi-intention interaction scenarios.

[0003] Firstly, existing session recommendation models lack the ability to model cross-category semantics. Most methods only model the relationships between items within a single category or a single session, and cannot capture the potential semantic commonalities and synergies between items of different categories. For example, after browsing "phone cases", a user may be interested in "chargers" or "Bluetooth earphones", but traditional models cannot identify this cross-category implicit preference based only on adjacent item sequences. The lack of depiction of cross-category dependency relationships results in insufficient generalization ability of the model in multi-scenario recommendation.

[0004] Secondly, in terms of the depth of user intention expression, traditional methods generally use one-way or single-layer semantic representation, relying only on the surface features of items for embedding, ignoring the potential intention structure of users at different semantic levels (such as functional needs, aesthetic preferences, usage scenarios, etc.). This single representation method cannot distinguish between short-term immediate behavior and long-term interest motivation, thereby limiting the diversity and accuracy of the recommendation results.

[0005] Thirdly, existing research has shown insufficient performance in capturing long-term interests across sessions. Users' interests are usually dynamic and temporal, evolving with time, environment or context. Existing models mostly focus on item transfer within a single session, and do not adequately consider long-term dependencies and migration patterns across sessions, resulting in the model's inability to accurately grasp the changing trends of users' preferences. For example, a user has browsed "running shoes", "sports watches" and "sports backpacks" in multiple independent sessions, and traditional models often treat these behaviors as unrelated, ignoring the underlying overall "sports equipment" interest context.

[0006] To improve the conversation semantic capture ability and recommendation accuracy, existing research attempts to combine high-order graph neural networks, hypergraph modeling and attention mechanism, etc. to enhance the modeling ability of complex item relationships. However, such methods still have deficiencies in multi-intention expression and dynamic preference modeling: on the one hand, existing models are difficult to realize dynamic interaction and hierarchical fusion among multiple intentions, and cannot simultaneously represent the mutual influence of multiple user interests; on the other hand, in the cross-category semantic fusion process, the model often lacks effective information filtering and weight regulation mechanism, resulting in feature redundancy and information noise accumulation, thereby affecting the recommendation performance.

[0007] In summary, the existing conversation recommendation technology still has problems such as insufficient semantic association capture, shallow user intention modeling and insufficient long-term interest expression when dealing with complex cross-category, multi-intention and multi-conversation scenarios. To solve the above problems, a conversation recommendation method that can fuse cross-category relationships, mine multi-layer semantic intentions and has dynamic learning ability is needed to accurately model the user's implicit behavior context and dynamic interest migration rules, thereby improving the accuracy, robustness and intelligence level of the recommendation system. SUMMARY

[0008] (I) Technical problems solved

[0009] The present application provides a conversation recommendation method based on cross-category heterogeneous hypergraph multi-intention expression, which models the high-order dependency relationship of items within the conversation to capture the deep semantic association between items in the conversation. At the same time, by generating, selecting and aggregating node information in different intention subspaces, the intra-conversation information in the star graph is fused with the multi-intention of the cross-category heterogeneous hypergraph, thereby realizing accurate multi-intention modeling.

[0010] (II) Technical solutions

[0011] The present application specifically adopts the following technical solutions to achieve the above purposes:

[0012] A conversation recommendation method based on cross-category heterogeneous hypergraph multi-intention expression, comprising preprocessed data, the method comprising the following steps:

[0013] S1: The preprocessed data is input into a category recognition processing module, and a latent Dirichlet allocation (LDA) model is used to mine the latent category distribution of the items, and a category classified item is output;

[0014] S2: The preprocessed data and the category classified item are respectively input into a self-loop star graph module and a cross-category heterogeneous hypergraph module for graph structure construction, and a self-loop star graph and a cross-category heterogeneous hypergraph are output;

[0015] S3: The self-loop star graph input is input to the intra-session item transfer information learning layer for learning representation, and the intra-session representation learning is output. The node features of each layer are obtained through the stacking calculation of multiple self-loop star graph network layers, and then the expressway network based on the gating mechanism is introduced to realize the adaptive fusion of the initial features and the high-level features.

[0016] S4: The cross-category heterogeneous hypergraph input is input to the inter-session item transfer information learning layer for learning representation, and the inter-session representation learning is output. The node features are updated through the attention network HyperGAT, and then the multi-intent embedding generation, sparse intent selection and multi-intent fusion are performed by using the dynamic intent context weaver.

[0017] S5: The intra-session representation learning and the inter-session representation learning are input to the fusion module for fusion, and the fused feature representation is output.

[0018] S6: The fused feature representation is input to the prediction module, and the final output result is output.

[0019] Further, in S1, the category recognition processing module performs semantic category recognition on the preprocessed data, uses an LDA model to mine the latent category distribution of each item, and outputs the corresponding classified items.

[0020] The preprocessed data is the denoised session data obtained by preprocessing the original session data.

[0021] The LDA model specifically includes: first, vectorizing the preprocessed data to convert text or item sequences into computable feature vectors; second, using the LDA model to model the categories of the vectorized data and learn the probability distribution of each item under the latent category; then, determining the dominant category of each item according to the category distribution output by the LDA model; finally, outputting the category classified item set as the input of the subsequent graph structure construction module.

[0022] Further, in the entire process, the preprocessed data of the session is referred to as , the classified items are referred to as , the self-loop star graph is referred to as , the cross-category heterogeneous hypergraph is referred to as , the intra-session item transfer information learning layer is referred to as , the inter-session item transfer information learning layer is referred to as , the fused feature representation obtained after the fusion of the intra-session and inter-session representation learning is referred to as , and the final output recommendation result is referred to as .

[0023] Further, in the S2, the self-loop star graph module models a single session sequence, effectively capturing the item order relationship within the session, cross-item information interaction, and representation enhancement of the item itself; the cross-category heterogeneous hypergraph module constructs complex multi-element relationships between sessions, captures cross-category relationships, and enhances context association and recommendation effect;

[0024] The self-loop star graph module includes node initialization, star node construction, generation of bidirectional link edges, and generation of self-loop feature edges.

[0025] The cross-category heterogeneous hypergraph module includes session segmentation, category association analysis, hyperedge generation, and multi-window aggregation.

[0026] Further, in the S3, the intra-session item transfer information learning layer is used for hierarchical feature learning and semantic information transmission of the item nodes in the self-loop star graph, to capture the temporal dependence relationship and local semantic association between items within the session.

[0027] The intra-session item transfer information learning layer includes the following steps:

[0028] The stacking of multiple self-loop star graph network layers is calculated as follows:

[0029]

[0030] In the above formula, is the feature representation of the i-th node in the j-th layer of the star graph; is the feature representation of the i-th node in the j-th layer of the star graph; is the aggregated feature of the i-th node after the last layer j-1; is the aggregated feature of the i-th node after the last layer j-1; is the aggregated neighbor node feature representation of the star node in the j-th layer; is the aggregated neighbor node feature representation of the star node in the j-th layer; is the aggregated neighbor node feature representation of the star node in the j-th layer; is the adjacency matrix of the j-th layer, describing the connection relationship between nodes; is the initial layer adjacency matrix; SSGNN(·) represents the star self-attention graph neural network operation; The expressway network based on the gating mechanism is calculated as follows:

[0031] In the above formula, is the initial node embedding;

[0032] is the updated node embedding of the i-th node in the j-th layer of the current session; is the updated node embedding of the i-th node in the j-th layer of the current session; is the updated node embedding of the i-th node in the j-th layer of the current session;

[0033] is the updated node embedding of the i-th node in the j-th layer of the current session; is the updated node embedding of the i-th node in the j-th layer of the current session; is the updated node embedding of the i-th node in the j-th layer of the current session; is the updated node embedding of the i-th node in the j-th layer of the current session; is the updated node embedding of the i-th node in the j-th layer of the current session; This represents the feature representation after fusion.

[0034] Furthermore, in S4, the inter-session item transfer information learning layer is used to perform high-order feature learning and multi-intent modeling on the node relationships in the cross-category heterogeneous hypergraph. The dynamic intent network weaver realizes the combination of multi-intent embedding, sparse intent filtering and multi-intent fusion. It can not only simulate the interest migration pattern of users in cross-category scenarios, but also dynamically adjust the weight distribution of different intents in the global semantic space, thereby effectively improving the interpretability and robustness of the recommendation results.

[0035] The inter-session item transfer information learning layer includes HyperGAT attention node feature update and dynamic intent weaver; the dynamic intent weaver sequentially includes multi-intent embedding generation, sparse intent selection, and multi-intent fusion. Multi-intent embedding generation maps node features to the multi-intent semantic space through linear mapping; sparse intent selection uses the α-entmax sparsity adjustment function to filter key intents and suppress redundant information; multi-intent fusion achieves global information integration through the joint aggregation operation of Max-Pooling and Average-Pooling, outputting a comprehensive multi-intent semantic representation for subsequent feature fusion and recommendation prediction.

[0036] Node feature updates are calculated using the following formula:

[0037]

[0038] In the above formula, , It is a set of superedges; The attention weights between nodes and superedges; These are learnable parameters; For linking hyperedges Node aggregation information; The updated features of node k;

[0039] Multi-intent embedding generation, the calculation formula is as follows:

[0040]

[0041] In the above formula, For the first A vector representing an intention; For hypergraph network nodes; For the first Each intention mapping weight matrix; It is the bias vector; (·) is a non-linear activation function;

[0042] Sparse intent selection is calculated using the following formula:

[0043]

[0044]

[0045] In the above formula, This is the calculated intent score vector; This represents the probability distribution of the output. This is the generalized Tsallis entropy regularization term, used to control the sparsity of the distribution; To control sparsity; Let be the probability value of the j-th intention; Let the probability be t-dimensional;

[0046] Multi-intent fusion, the calculation formula is as follows:

[0047]

[0048] In the above formula, For the first Semantic features corresponding to each intent vector; These are weighting coefficients used to adjust the fusion ratio between max pooling and average pooling. The fused multi-intent semantic vector is then fed into subsequent modules.

[0049] Furthermore, in step S5, the fusion module learns the in-session item transfer information layer. Inter-session item transfer information learning layer Perform fusion processing to obtain the fused feature representation. To obtain a comprehensive feature representation that combines local behavioral dependencies and global intent transfer information, thereby providing a semantically complete user feature vector for the final recommendation prediction;

[0050] Calculate the gating coefficient The calculation formula is as follows:

[0051]

[0052] In the above formula, These are learnable parameters; For bias terms; Indicates feature concatenation operation; In-session item transfer information learning layer; Learning layer for inter-session item transfer information; The sigmoid activation function controls the fusion ratio of semantic information and star node information.

[0053] The final fusion representation is calculated using the following formula:

[0054]

[0055] In the above formula, This is the final fusion feature representation of star nodes and multi-intention preferences.

[0056] Furthermore, the prediction module is used to predict the recommendation results based on the feature representation output by the fusion module. By establishing a bidirectional dependency relationship between the sequences and calculating the recommendation probability of candidate items, it can accurately predict the user's next behavior.

[0057] The prediction module includes an itemset, a transposed location map, a softmax function, and a cross-entropy loss function. The itemset contains all candidate item nodes. Each The vectorized representation of the j-th candidate item is used to capture the bidirectional dependency between items in the sequence. The positional information of the input sequence is transposed and mapped so that the model can learn not only positive contextual dependencies but also negative dependencies. The softmax function performs probabilistic processing on the score vector to obtain the recommendation probability distribution of each candidate item. The cross-entropy loss function is used to measure the difference between the model's predicted probability distribution and the actual user behavior distribution to guide the optimization and update of model parameters, thereby improving the accuracy and robustness of the prediction results.

[0058] The transpose position mapping is calculated using the following formula:

[0059]

[0060] In the above formula, This is the transpose position matrix; This is the feature representation after location mapping;

[0061] The softmax function is calculated using the following formula:

[0062]

[0063] In the above formula, the softmax function is used to normalize the result into a probability distribution in order to determine the predicted score of the candidate item; This refers to the project and its included candidate projects. This is the final recommended output;

[0064] The cross-entropy loss function is calculated using the following formula:

[0065]

[0066] In the above formula, Predict probabilities for the model; This represents the true label distribution; This represents the overall loss value, used for backpropagation to optimize model parameters.

[0067] Furthermore, including:

[0068] The self-looping star graph module is used for preprocessing data. Perform graph structure construction to obtain the ring star graph structure;

[0069] Cross-category heterogeneous hypergraph module, used for classifying items Perform graph structure construction to obtain cross-category heterogeneous hypergraph structures;

[0070] In-session item transfer information learning layer, used for learning self-loop star graphs Perform learning representations and obtain in-session representation learning;

[0071] Inter-session item transfer information learning layer, used for cross-category heterogeneous hypergraphs Perform learning representations and obtain inter-session representation learning;

[0072] The fusion module is used for learning in-session representations. Inter-session representation learning Perform fusion representation to obtain the fused feature representation;

[0073] The prediction module is used to learn and fuse the intra-session and inter-session representations to obtain the fused feature representation. Make predictions and obtain the final output results. .

[0074] (III) Beneficial Effects

[0075] Combining all the above technical solutions, the beneficial effects of this invention are as follows:

[0076] This invention proposes a session recommendation method based on multi-intent representation using cross-category heterogeneous hypergraphs. By constructing a self-looping star graph and a cross-category heterogeneous hypergraph, it achieves unified modeling of local dependencies within a session and global associations between sessions. The self-looping star graph is used to capture the sequential relationships and higher-order dependency features between items in a single session, while the cross-category heterogeneous hypergraph is used to characterize the category relationships and interest migration patterns between different sessions. Simultaneously, a dynamic intent weaver is innovatively designed to perform multi-level identification and fusion of potential user intents. This dynamic intent weaver includes three stages: multi-intent embedding generation, sparse intent selection, and multi-intent fusion. These stages respectively realize the expression from basic behavioral features to deep interest intents, the filtering from redundant features to key intents, and the global aggregation of multi-intent features, thereby effectively improving the model's semantic expression capability and intent recognition accuracy.

[0077] This invention enables dynamic modeling and intelligent recognition of user intent in multi-category, cross-session recommendation scenarios. By combining the local feature capture capability of self-loop star graphs with the global relationship modeling capability of heterogeneous hypergraphs, the model can deeply mine the correlation structure between users' short-term behavior and long-term interests. Simultaneously, the dynamic intent weaver enhances the system's intent perception and semantic fusion capabilities, making recommendation results more accurate, personalized, and interpretable. This invention not only significantly improves the prediction accuracy and generalization performance of recommendation systems but also enhances the model's understanding and response capabilities to complex user behaviors, providing a scalable, high-performance, and interpretable technical solution for intelligent recommendation. Attached Figure Description

[0078] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure;

[0079] Figure 1 This is a schematic diagram of a conversation recommendation method based on cross-category heterogeneous hypergraph multi-intent expression provided by an embodiment of the present invention;

[0080] Figure 2 This is a schematic diagram illustrating the structural principles of the self-loop star graph module and the cross-category heterogeneous hypergraph module provided in this embodiment of the invention;

[0081] Figure 3 This is a schematic diagram illustrating the learning representation structure principle of the in-session item transfer information learning layer provided in this embodiment of the invention;

[0082] Figure 4 This is a schematic diagram illustrating the learning representation structure principle of the inter-session item transfer information learning layer provided in this embodiment of the invention;

[0083] Figure 5 This is a schematic diagram of the fusion module structure provided in an embodiment of the present invention;

[0084] Figure 6 This is a schematic diagram of the prediction module structure provided in an embodiment of the present invention.

[0085] In the diagram: 101, Category Recognition Processing Module; 201, Self-looping Star Graph Module; 202, Cross-category Heterogeneous Hypergraph Module; 301, Intra-session Item Transfer Information Learning Layer; 302, Inter-session Item Transfer Information Learning Layer; 303, Fusion Module; 401, Prediction Module. Detailed Implementation

[0086] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0087] Example

[0088] like Figure 1 As shown, this embodiment of the invention provides a session recommendation method based on cross-category heterogeneous hypergraph multi-intent expression, including preprocessed data. The method is composed of a deep learning network, which includes a category recognition processing module 101, a self-looping star graph module 201, a cross-category heterogeneous hypergraph module 202, an intra-session item transfer information learning layer 301, an inter-session item transfer information learning layer 302, a fusion module 303, and a prediction module 401. The method includes the following steps:

[0089] S1: The preprocessed data is input into the category recognition processing module 101, which mines the latent category distribution of the items through the Latent Dirichlet Allocation (LDA) model and outputs the category classification items;

[0090] The LDA model specifically includes: first, vectorizing the preprocessed data to transform the text or item sequence into a computable feature vector; second, using the LDA model to perform category modeling on the vectorized data to learn the probability distribution of each item under the potential category; then, determining the dominant category of each item based on the category distribution output by the LDA model; and finally, outputting a set of category-classified items as input to the subsequent graph structure construction module.

[0091] S2: The preprocessed data and category classification items are respectively input into the self-loop star graph module 201 and the cross-category heterogeneous hypergraph module 202 to construct the graph structure, and output the self-loop star graph and the cross-category heterogeneous hypergraph;

[0092] In step S2 of this embodiment of the invention, as follows Figure 2 As shown, the self-loop star graph module 201 includes node initialization, star node construction, generation of bidirectional links, and generation of self-loop feature edges. It constructs a graph structure within a single session to capture sequential relationships, information interactions, and self-feature enhancements between items. First, the node initialization layer vectorizes the item nodes in the session sequence, forming a node set: in, For the i-th project node, This serves as the virtual star-shaped central node, used to aggregate global information. Secondly, the star-shaped node construction layer utilizes this virtual star-shaped central node. Connect all project nodes to form a central-radial structure to capture global dependencies between projects. Then, a bidirectional edge generation layer establishes bidirectional edges between projects and star nodes to enable bidirectional information flow; directed edges ( , ) is used to aggregate project node features into star nodes, reverse edges ( , This layer is used to feed back the star-shaped node features to each item node, thereby achieving global feature propagation within the session. Finally, the self-loop feature enhancement layer adds a self-loop edge to each item node. , This preserves the node's own feature information and prevents feature degradation during information propagation. Through the synergistic effect of the above layers, the self-looping star graph module 201 can extract high-order dependency features in a single session, generate a graph structure containing local and global information, and thus generate a star graph representation with complete structure and rich feature expression, providing high-quality input for subsequent session feature learning and cross-session modeling.

[0093] In step S2 of this embodiment of the invention, as follows Figure 2 As shown, the cross-category heterogeneous hypergraph module 202 includes: a session segmentation layer, a category association analysis layer, a hyperedge generation layer, and multi-window aggregation. First, the session segmentation layer uses a sliding window mechanism to segment the input session sequence to ensure that the model models temporal dependencies within a local scope. Each session subsequence within a window serves as a basic unit for hyperedge construction. Second, the category association analysis layer performs category consistency analysis on the item nodes within each window, identifying sets of item nodes belonging to the same category or with similar semantics. Finally, the hyperedge generation layer constructs a hyperedge set within each window based on category consistency. Each hyperedge can connect multiple category-related nodes simultaneously, thus forming a local high-order relational structure. Finally, the multi-window aggregation layer slides the window with a fixed step size and aggregates the local hyperedges generated by each window to form an overall cross-category heterogeneous hypergraph. :

[0094]

[0095] In the above formula, It is a set of superedges; A represents a project node with a category; A is an attribute mapping function used to represent the heterogeneous mapping relationship between categories and projects.

[0096] S3: The self-circular star diagram is input to the in-session item transfer information learning layer 301 for learning representation, and in-session representation learning is obtained;

[0097] In step S3 of this embodiment of the invention, as follows Figure 3 As shown, firstly, a multi-layer self-loop star graph network is computed to obtain the node features of each layer. To avoid feature degradation and gradient shift caused by multi-layer stacking, a highway network based on a gating mechanism is introduced to adaptively fuse the initial features and high-level features. By dynamically adjusting the information flow ratio between layers, effective complementarity and balance between deep and shallow features are achieved. Finally, the fused session feature representation is output to characterize the high-order dependencies and global semantic information between items within the session, providing a basic session semantic representation for subsequent cross-category heterogeneous hypergraph modeling.

[0098] The in-session item transfer information learning layer 301 includes the following steps:

[0099] The stacking of multiple self-loop star graph network layers is calculated using the following formula:

[0100]

[0101] In the above formula, For the first The node in the star diagram is the _ . Layer feature representation; For the first The node is on the upper layer. -1 Features after aggregation; For the star node at the 1st Layer-based aggregated neighbor node feature representation; This represents the star node's neighbor node at the next higher level; For the first The adjacency matrix of a layer describes the connection relationships between nodes; is the initial layer adjacency matrix; SSGNN() represents the star-shaped self-attention graph neural network operation;

[0102] The calculation formula for a highway network based on a gating mechanism is as follows:

[0103]

[0104] In the above formula, Embed for the initial node; For the current session The node embedding after layer aggregation update; This is the element-wise multiplication operator; This represents the feature representation after fusion.

[0105] S4: The cross-category heterogeneous hypergraph is input to the inter-session item transfer information learning layer 302 for learning representation and obtaining inter-session representation learning; the inter-session item transfer information learning layer 302 includes HyperGAT attention node feature update and dynamic intent weaver; the dynamic intent weaver sequentially includes multi-intent embedding generation, sparse intent selection and multi-intent fusion.

[0106] In step S3 of this embodiment of the invention, as follows Figure 3 First, the HyperGAT (Hypergraph Graph Attention Network) is used to perform weighted calculations on the semantic associations between nodes and hyperedges. By learning the attention weights of nodes to hyperedges, dynamic updates of node features in cross-category heterogeneous hypergraphs are achieved, thereby capturing high-order semantic dependencies between sessions.

[0107] Node feature updates are calculated using the following formula:

[0108]

[0109] In the above formula, , It is a set of superedges; The attention weights between nodes and superedges; These are learnable parameters; For linking hyperedges Node aggregation information; Let k be the updated features of node k.

[0110] Furthermore, a dynamic intent weaver is introduced, into which updated node features are input to simulate the multi-layered interest evolution process of users in different conversational scenarios. The dynamic intent weaver sequentially includes three steps: multi-intent embedding generation, sparse intent selection, and multi-intent fusion.

[0111] Multi-intent embedding generation, the calculation formula is as follows

[0112]

[0113] In the above formula, For the first A vector representing an intention; For hypergraph network nodes; For the first Each intention mapping weight matrix; It is the bias vector; (·) is a non-linear activation function;

[0114] Sparse intent selection is calculated using the following formula:

[0115]

[0116]

[0117] In the above formula, This is the calculated intent score vector; This represents the probability distribution of the output. This is the generalized Tsallis entropy regularization term, used to control the sparsity of the distribution; To control sparsity; Let be the probability value of the j-th intention; Let be the t-dimensional probability.

[0118] Multi-intent fusion, the calculation formula is as follows:

[0119]

[0120] In the above formula, For the first Semantic features corresponding to each intent vector; These are weighting coefficients used to adjust the fusion ratio between max pooling and average pooling. The fused multi-intent semantic vector is then fed into subsequent modules.

[0121] Multi-intent embedding generation is based on the features of each node. The model employs linear mapping transformations to embed node features into a multi-intent semantic space. Sparse intent selection uses an α-entmax sparsity adjustment function to filter and aggregate multi-intent vectors, retaining key intent information and suppressing redundant noise. Multi-intent fusion uses a combined Max-Pooling and Average-Pooling operation to globally aggregate the filtered multi-intent representations. To avoid excessively high embedding dimensions leading to feature redundancy or increased computational complexity, the dynamic intent weaver uses a sparsity constraint mechanism, highlighting high-weight intents while suppressing low-weight intents, thus ensuring model efficiency and interpretability. Finally, the fused multi-intent semantic representation vector is output for subsequent feature fusion and recommendation prediction, dynamically capturing user interest migration paths between different categories and improving the accuracy and robustness of recommendation results.

[0122] S5: The intra-session representation learning and inter-session representation learning are input into the fusion module 303 for fusion to obtain the fused feature representation;

[0123] In step S5 of this embodiment, as follows Figure 5 As shown, intra-session representation learning and inter-session representation learning will be passed to the subsequent fusion module 303, when the gating coefficient... When the value is large, the model tends to utilize in-session structural features to reinforce the expression of the user's short-term interests; when... When the interface is smaller, the model focuses more on global intent features across sessions to capture long-term user preference trends. Through this dynamic control mechanism, the model can adaptively adjust the fusion ratio based on the feature complexity of the input sequence, achieving information complementarity and semantic balance between feature levels. This fusion is then combined with the output features from the intra-session information layer to form a complete user feature representation, supporting the final recommendation prediction. The final fused representation S is input as a unified user semantic vector to the prediction module 401 to calculate the recommendation probability of candidate items.

[0124] Calculate the gating coefficient The calculation formula is as follows:

[0125]

[0126] In the above formula, These are learnable parameters; For bias terms; Indicates feature concatenation operation; In-session item transfer information learning layer; Learning layer for inter-session item transfer information; The sigmoid activation function controls the fusion ratio of semantic information and star node information.

[0127] The final fusion representation is calculated using the following formula:

[0128]

[0129] In the above formula, This is the final fusion feature representation of star nodes and multi-intention preferences.

[0130] S6: The fused feature representation is input to the prediction module 401, and the final output is the result. The prediction module 401 includes an itemset, a transpose position mapping, a softmax function, and a cross-entropy loss function.

[0131] In step S6 of this embodiment, as follows Figure 6 As shown, an item set is created to construct a candidate set, which contains all candidate item nodes. Each The vectorized representation of the j-th candidate item is used. Each item node vector describes the semantic features and historical behavior relevance of the candidate item, providing a foundation for subsequent score calculation. Subsequently, a transposed position mapping mechanism is used to bidirectionally encode the position information of the input sequence. By transposing the position matrix, the bidirectional dependencies between items in the sequence are captured, enabling the model to learn not only positive contextual dependencies (from front to back) but also inverse dependencies (from back to front), thereby enhancing the overall modeling ability of the sequence semantics. Next, the matching relationship between the candidate item vectors and the fused feature representations is used to calculate the candidate item score vectors. Then, the softmax function is used to normalize the item score vectors, mapping any real value to a probability distribution, and calculating the probability of each candidate item being recommended, thus obtaining the item recommendation probability distribution. To avoid the prediction results deviating from the real user behavior distribution, a cross-entropy loss function is further introduced to calculate the error between the predicted probability output by the model and the real behavior label. By minimizing this loss function, the model parameters are adaptively optimized, improving the accuracy and robustness of the prediction results. Finally, the optimized recommendation probability result is output as the model's prediction output, used to generate the session recommendation list.

[0132] The transpose position mapping is calculated using the following formula:

[0133]

[0134] In the above formula, This is the transpose position matrix; This is the feature representation after location mapping;

[0135] The softmax function is calculated using the following formula:

[0136]

[0137] In the above formula, the softmax function is used to normalize the result into a probability distribution in order to determine the predicted score of the candidate item; This refers to the project and its included candidate projects. This is the final recommended output;

[0138] The cross-entropy loss function is calculated using the following formula:

[0139]

[0140] In the above formula, Predict probabilities for the model; This represents the true label distribution; This represents the overall loss value, used for backpropagation to optimize model parameters.

[0141] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A conversation recommendation method based on cross-category heterogeneous hypergraph multi-intention expression, comprising pre-processed data, characterized in that, The method comprises the following steps: S1: the preprocessed data is input into a category identification processing module (101), a latent Dirichlet allocation (LDA) model is used to mine the latent category distribution of the items, and a category classified item is output; S2: the preprocessed data and the category classified item are input into a self-loop star graph module (201) and a cross-category heterogeneous hypergraph module (202) respectively to construct the structure of the graph, and a self-loop star graph and a cross-category heterogeneous hypergraph are output; In S2, the self-loop star graph module (201) models a single session sequence, effectively capturing the order relationship of items within the session, cross-item information interaction, and the representation enhancement of the items; the cross-category heterogeneous hypergraph module (202) constructs complex multi-element relationships between sessions, captures cross-category relationships, and enhances context association and recommendation effect; The self-loop star graph module (201) comprises node initialization, star node construction, generation of bidirectional link edges, and generation of self-loop feature edges; The cross-category heterogeneous hypergraph module (202) comprises session segmentation, category association analysis, generation of hyperedges, and multi-window aggregation; S3: the self-loop star graph is input into an intra-session item transfer information learning layer (301) for learning representation, and intra-session representation learning is output; through the stacking calculation of multiple self-loop star graph network layers, the node features of each layer are obtained, and then a highway network based on a gating mechanism is introduced to realize the adaptive fusion of initial features and high-level features; S4: the cross-category heterogeneous hypergraph is input into an inter-session item transfer information learning layer (302) for learning representation, and inter-session representation learning is output; the node features are updated through an attention network HyperGAT, and then a dynamic intention context weaver is used for multi-intention embedding generation, sparse intention selection, and multi-intention fusion; S5: the intra-session representation learning and the inter-session representation learning are input into a fusion module (303) for fusion, and a feature representation after fusion is output; S6: the feature representation after fusion is input into a prediction module (401), and finally an output result is output.

2. The conversation recommendation method based on cross-category heterogeneous hypergraph multi-intention representation according to claim 1, characterized in that, In S1, the category identification processing module (101) performs semantic category identification on the preprocessed data, uses an LDA model to mine the latent category distribution of each item, and outputs a corresponding classified item; The preprocessed data is obtained by preprocessing the original session data to obtain denoised session data; The LDA model specifically comprises: first, the preprocessed data is vectorized, and a text or item sequence is converted into a calculable feature vector; Secondly, the LDA model is used to model the categories of the vectorized data, and the probability distribution of each item under the latent category is learned; then, the dominant category of each item is determined according to the category distribution output by the LDA model; finally, a set of category classified items is output as the input of the subsequent graph structure construction module.

3. The conversation recommendation method based on cross-category heterogeneous hypergraph multi-intention expression according to claim 1, characterized in that, Throughout the flow, the data after session pre-processing is called ; the classification items are called ; and the self-loop star graph is called ; The cross-category heterogeneous hypergraph is referred to as The intra-session item transfer information learning layer (301) is referred to as The inter-session item transfer information learning layer (302) is referred to as The fused feature representation obtained after the intra-session and inter-session representation learning fusion is referred to as The final output recommendation result is referred to as .

4. The conversation recommendation method based on cross-category heterogeneous hypergraph multi-intention representation according to claim 1, characterized in that, In S3, the intra-session item transfer information learning layer (301) is used for hierarchical feature learning and semantic information transmission of the item nodes in the self-loop star graph, so as to capture the time sequence dependency relationship and local semantic association between the items within the session; The intra-session item transfer information learning layer (301) includes the following steps: A stack of multiple self-loop star-shaped graph network layers, with the following calculation formula: ; In the above formula, is the feature representation of the th node in the layer of the star graph; is the aggregated feature of the th node in the previous layer -1; is the aggregated neighbor node feature representation of the star node in the layer; is the representation of the neighbor nodes of the star node in the previous layer; is the adjacency matrix of the th layer, describing the connection relationship between nodes; is the initial layer adjacency matrix; SSGNN(·) represents the star self-attention graph neural network operation; A highway network based on a gating mechanism, with the following calculation formula: ; In the above formula, is the initial node embedding; is the first layer aggregated updated node embedding; is an element-wise multiplication operator; denotes the fused feature representation.

5. The conversation recommendation method based on cross-category heterogeneous hypergraph multi-intention representation according to claim 1, characterized in that, In the S4, the inter-session item transfer information learning layer (302) is configured to perform high-order feature learning and multi-intent modeling on the node relationship in the cross-category heterogeneous hypergraph, and the dynamic intent context weaver realizes the combination of multi-intent embedding, sparse intent filtering and multi-intent fusion, which not only simulates the interest migration law of the user in the cross-category scene, but also dynamically adjusts the weight distribution of different intents in the global semantic space, thereby effectively improving the interpretability and robustness of the recommendation result. The inter-session item transfer information learning layer (302) includes HyperGAT attention node feature updating and a dynamic intent context weaver; the dynamic intent context weaver sequentially includes multi-intent embedding generation, sparse intent selection and multi-intent fusion, and the multi-intent embedding generation maps the node features to a multi-intent semantic space through linear mapping; The sparse intent selection uses an alpha-entmax sparsity adjustment function to filter key intents and suppress redundant information; the multi-intent fusion realizes global information integration through Max-Pooling and Average-Pooling joint aggregation operations, and outputs a comprehensive multi-intent semantic representation for subsequent feature fusion and recommendation prediction; Node feature updating, with the following calculation formula: ; In the above formula, , is a superedge set; is an attention weight of a node and a superedge; is a learnable parameter; is the node aggregation information of the linked superedge ; is the updated feature of node k; Multi-intent embedding generation, with the following calculation formula: ; In the above formula, For the first A vector representing an intention; For hypergraph network nodes; For the first Each intention mapping weight matrix; It is the bias vector; (·) is a non-linear activation function; Sparse intent selection, with the following calculation formula: ; ; In the above formula, is the computed intent score vector; is the output probability distribution; is the generalized Tsallis entropy regularizer to control the sparsity of the distribution; is to control sparsity; is the probability value for the jth intent; is the t-dimensional probability; Multi-intent fusion, with the following calculation formula: ; In the above formula, is the semantic feature corresponding to the i-th intent vector; is the semantic feature corresponding to the i-th intent vector; is the weight coefficient; used to adjust the fusion ratio between max-pooling and average-pooling; The fused multi-intent semantic vector is sent to the subsequent module.

6. The conversation recommendation method based on cross-category heterogeneous hypergraph multi-intention expression according to claim 3, characterized in that, In the S5, the fusion module (303) fuses the intra-session item transfer information learning layer and the inter-session item transfer information learning layer (302) to obtain a fused feature representation ; To obtain a comprehensive feature representation that combines local behavior dependence and global intent migration information, thereby providing a semantic complete user feature vector for final recommendation prediction; Computing the gating coefficients The formula is as follows: ; In the above formula, is a learnable parameter; is a bias term; represents a feature concatenation operation; intra-session item transfer information learning layer; is an inter-session item transfer information learning layer; is a sigmoid activation function used to control the fusion ratio of semantic information and star node information; Final fusion representation, with the following calculation formula: ; In the above formula, is the final fused representation of the star node and the multi-intent preferences.

7. The conversation recommendation method based on cross-category heterogeneous hypergraph multi-intention representation according to claim 1, characterized in that, The prediction module (401) is configured to perform recommendation result prediction on the feature representation output by the fusion module, by establishing a bidirectional dependency relationship of the sequence and calculating the recommendation probability of the candidate item, to realize accurate prediction of the next behavior of the user; The prediction module (401) includes an item set, a transposed position mapping, a softmax function, and a cross-entropy loss function, the item set includes all candidate item nodes , each represents a vectorized representation of the jth and candidate item, the transposed position mapping is used to capture bidirectional dependency between items in a sequence, and transposes position information of an input sequence, so that the model can not only learn forward context dependency, but also learn reverse dependency, the softmax function probabilistically processes a score vector to obtain a recommendation probability distribution of each candidate item, and the cross-entropy loss function is used to measure the difference between the model prediction probability distribution and the real user behavior distribution, to guide the optimization and update of the model parameters, so as to improve the accuracy and robustness of the prediction result. Transposed position mapping, with the following calculation formula: ; In the above formula, is the transpose of the position matrix; is the position mapped feature representation; Softmax function, with the following calculation formula: ; In the above formula, the softmax function is used to normalize the results into a probability distribution to determine the predicted scores of the candidate items; for the item and the included candidate items, for the final output recommendation result; Cross-entropy loss function, with the following calculation formula: ; In the above formula, is the predicted probability for the model; is the true label distribution; is the overall loss value for backpropagation to optimize the model parameters.

8. The conversation recommendation method based on cross-category heterogeneous hypergraph multi-intention representation according to claim 1, characterized in that, Also includes: The self-looping star graph module (201) is used for preprocessing the data. Perform graph structure construction to obtain the ring star graph structure; Cross-category heterogeneous hypergraph module (202) is used to classify items. Perform graph structure construction to obtain cross-category heterogeneous hypergraph structures; a session-in item transfer information learning layer (301) configured to learn representations for a self-loop star graph and obtain session-in representation learning; inter-session item transfer information learning layer (302) for learning representations across heterogeneous hypergraphs of different categories learning representations, obtaining inter-session representation learning; The fusion module (303) is used for learning in-session representations. Inter-session representation learning Perform fusion representation to obtain the fused feature representation; The prediction module (401) is configured to perform prediction on the fused feature representation obtained after the intra-session and inter-session representation learning fusion to obtain a final output result. perform prediction to obtain a final output result .

Images