A multi-level conversation recommendation method driven by hypergraph convolution

Abstract

The present invention provides a multi-level conversation recommendation method driven by hypergraph convolution, which relates to the technical field of conversation recommendation. Obtain target conversation data to generate a hypergraph; apply K-hop neighbor aggregation technology to generate a domain node graph; process the domain node graph through a graph convolution network to determine the timeliness of user interests and generate an initial feature matrix; input the initial feature matrix into a KAN network to extract potential features; introduce an adversarial learning module to generate perturbation features through a generative adversarial network and perform alternating training; input the output result of the KAN network and the embedding matrix after adversarial training into the prediction layer to generate the final recommendation result. The technical effect of accurately capturing the interaction dynamics between users and items in conversation recommendation is achieved.

Description

Technical Field

[0001] The present application relates to the technical field of conversation recommendation, and in particular to a multi-level conversation recommendation method driven by hypergraph convolution. Background Art

[0002] With the rapid development of information technology and the Internet, conversational recommendation systems have gradually become an important part of e-commerce and content recommendation platforms. With the increase in the number of online users and the complexity of user behavior, traditional recommendation systems face huge challenges. Traditional systems usually rely on long-term user profiles and historical data to predict user interests and needs. However, in today's fast-paced digital environment, user interests may change significantly in a short period of time, which makes recommendation systems that rely solely on long-term data seem a bit lagging and inflexible.

[0003] By analyzing the user's interactive behavior in a short period of time, the conversational recommendation system can provide real-time personalized recommendations without long-term user profiles. The design concept of this system is to quickly capture the user's current interests and needs based on the user's immediate behavior and contextual information during the session. For example, on an e-commerce website, a user may suddenly change his interest while browsing a certain category of goods. At this time, the conversational recommendation system can immediately adjust the recommended content and provide goods related to the user's current browsing behavior, thereby improving the user's shopping experience and satisfaction. This recommendation method is particularly important in scenarios where user behavior changes rapidly, because it can dynamically adjust the recommendation strategy based on the user's immediate needs and preferences. This not only improves the accuracy and relevance of the recommendation, but also enhances the interactivity between users and the platform. By analyzing the user's click, browse, search and other behaviors in real time, the conversational recommendation system can quickly respond to changes in user needs and provide content or product recommendations that are more in line with the user's current interests. This flexibility and immediacy make the conversational recommendation system a key tool for enterprises to improve user experience and increase user stickiness in a highly competitive market environment.

[0004] The session-based recommendation system is centered on session data and can capture users' short-term preferences and dynamic changes in preferences over time. However, existing session recommendation methods based on recurrent neural networks or graph neural networks ignore the complex item correlations in session-based data. In real scenarios, item conversions are usually triggered by the joint effect of previous item clicks, and there are many-to-many and high-order relationships between items. Simple graphs cannot fully describe this class set relationship. There are still many complex interactive dynamics between users and items that need to be captured. Traditional neural networks have many shortcomings in dealing with complex high-dimensional features, such as difficulty in parameter sharing, which limits the generalization ability and efficiency of the model; neural networks, due to their powerful learning ability, may overfit on training data, resulting in decreased generalization ability on unseen data; and poor interpretability.

[0005] Therefore, how to accurately capture the interaction dynamics between users and items in conversational recommendations has become a technical problem that needs to be solved urgently. Summary of the invention

[0006] In order to accurately capture the interaction dynamics between users and items in conversational recommendations, this application provides a multi-level conversational recommendation method driven by hypergraph convolution.

[0007] The present application provides a hypergraph convolution-driven multi-level conversation recommendation method that adopts the following technical solutions:

[0008] A hypergraph convolution-driven multi-level conversation recommendation method, comprising:

[0009] Acquire target session data, and model according to the target session data to generate a hypergraph;

[0010] Applying K-hop neighbor aggregation technology to process the hypergraph to generate a domain node graph;

[0011] The domain node graph is processed through a graph convolutional network, and the length and time interval of the user interaction sequence are analyzed in combination with the intent-aware attention mechanism to determine the timeliness of the user's interest and generate an initial feature matrix;

[0012] Inputting the initial feature matrix into the KAN network, integrating explicit features and extracting latent features through nonlinear feature mapping;

[0013] Introducing an adversarial learning module, generating perturbation features through a generative adversarial network and performing alternating training;

[0014] The output result of the KAN network and the embedding matrix after adversarial training are input into the prediction layer to generate the final recommendation result.

[0015] Optionally, the step of acquiring target session data and modeling according to the target session data to generate a hypergraph includes:

[0016] Collecting interactive session data between the user and the system, and performing data cleaning and standardization preprocessing on the interactive session data to generate target session data;

[0017] Splitting data by session in the target session data, and mapping the sessions into hyperedges;

[0018] A hypergraph structure is constructed according to the hyperedges, each hyperedge includes all project nodes in the session, and cross-session associations are formed through shared nodes.

[0019] Optionally, the K-hop neighbor aggregation technology includes:

[0020] Assign initial feature vectors to hypergraph nodes;

[0021] Traverse the K-hop range to aggregate neighbor node information and generate fine-grained context features;

[0022] A domain node graph is constructed based on the aggregation results, and the connectivity and rationality of the graph structure are optimized through verification.

[0023] Optionally, the step of processing the domain node graph through a graph convolutional network, analyzing the length and time interval of the user interaction sequence in combination with the intent-aware attention mechanism, determining the timeliness of the user's interest and generating an initial feature matrix includes:

[0024] Initializing a dual-channel convolutional network, setting initial weights and bias vectors, wherein the dual-channel convolutional network includes a first channel and a second channel;

[0025] Performing a convolution operation on the hypergraph using the first channel to generate node features;

[0026] Processing the domain node graph using the second channel to generate interaction features;

[0027] Aggregating the node features with the interaction features and applying an intent-aware attention mechanism to capture the timeliness of user interests;

[0028] The timeliness of the user's interests is judged and an initial feature matrix is ​​generated.

[0029] Optionally, the step of inputting the initial feature matrix into a KAN network, integrating explicit features through nonlinear feature mapping, and extracting latent features comprises:

[0030] Assume that the KAN network has L layers, and the weight matrix of each layer is and the bias vector , select the activation function as ReLU:

[0031] ;

[0032] The initial feature matrix Input into the KAN network for nonlinear feature mapping:

[0033] ;

[0034] The input features are combined and transformed layer by layer through multiple hidden layers, and the ability of piecewise linear mapping is used to extract potential features.

[0035] Optionally, the adversarial learning module is introduced to generate perturbation features through a generative adversarial network and perform alternating training;

[0036] Initialize the adversarial learning module, which consists of a generator G and a discriminator D. The parameters of the generator are and the parameters of the discriminator are ;

[0037] Use the generator to perturb the input feature z:

[0038] ;

[0039] Generating Adversarial Examples , the adversarial example adds a small perturbation in the feature space:

[0040] ;

[0041] in, is a small perturbation coefficient used to control the amplitude of the perturbation;

[0042] During the adversarial training process, the discriminator D aims to maximize the ability to distinguish between original features and adversarial features:

[0043] ;

[0044] The goal of the generator G is to minimize the ability of the discriminator, that is:

[0045] .

[0046] Optionally, the step of inputting the output result of the KAN network and the embedding matrix after adversarial training into a prediction layer to generate a final recommendation result includes:

[0047] The characteristics of the output results of the KAN network The embedding matrix E after adversarial training is input into the prediction layer so that the prediction layer can generate the final prediction result. :

[0048] ;

[0049] in, and are the weight matrix and bias vector of the fully connected neural network, Indicates vertical concatenation of the feature matrix and the embedding matrix;

[0050] The obtained prediction results are normalized using the softmax function to convert them into probability distribution:

[0051] ;

[0052] in, is the normalized output, is the exponential operation on each unnormalized prediction result.

[0053] In summary, the present application includes the following beneficial technical effects:

[0054] This application obtains the target session data to generate a hypergraph; applies K-hop neighbor aggregation technology to generate a domain node graph; processes the domain node graph through a graph convolutional network to determine the timeliness of user interests and generate an initial feature matrix; inputs the initial feature matrix into the KAN network to extract potential features; introduces an adversarial learning module to generate perturbation features through a generative adversarial network and conducts alternating training; inputs the output of the KAN network and the embedded matrix after adversarial training into the prediction layer to generate the final recommendation result. This achieves the technical effect of accurately capturing the dynamic interaction between users and items in conversational recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flowchart of the first embodiment of the multi-level conversation recommendation method driven by hypergraph convolution in the present application. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below through the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0057] The present application embodiment provides a hypergraph convolution driven multi-level session recommendation method, referring to Figure 1 , Figure 1 This is a flowchart of the first embodiment of the multi-level conversation recommendation method driven by hypergraph convolution in the present application.

[0058] In this embodiment, the hypergraph convolution driven multi-level conversation recommendation method includes the following steps:

[0059] Step S10: Obtain target session data, and model the target session data to generate a hypergraph.

[0060] It can be understood that the description of technical terms in this embodiment is as follows:

[0061] Hypergraph: A mathematical structure in graph theory, which is a generalization of traditional graphs. In a hypergraph, an edge (called a hyperedge) can connect any number of vertices (nodes or vertices). This is different from traditional graphs, where each edge can only connect two vertices. The form of a hypergraph is defined as H = (X, E), where X is a set of vertices and E is a set of non-empty subsets of X, which are called hyperedges.

[0062] KAN network: The full name is Kolmogorov-Arnold Networks, which is a new type of deep learning model that has shown great potential and advantages in many fields. The core feature of the KAN network is that it places activation functions on the edge of the network (on the weights), and these activation functions are learnable and are usually parameterized using B-splines. This design makes the KAN network more accurate and interpretable in handling complex function fitting, solving partial differential equations, etc.

[0063] K-hop neighbor aggregation technology: It is a message passing mechanism in Graph Neural Networks (GNNs) that allows nodes to aggregate not only the information of direct neighbors (1-hop neighbors), but also the information from more distant k-hop neighbors. This technology can enhance the expressive power of GNNs because it can capture a wider range of neighborhood information of nodes, which is very useful for understanding and processing complex structures and patterns in graphs.

[0064] Intent-aware attention mechanism: is a technique used in deep learning models, especially in the fields of natural language processing (NLP) and multimodal learning, to increase the model's attention to key information in the input data. This mechanism mimics the selective attention ability of humans when processing information, allowing the model to dynamically adjust its attention weights when processing input data, thereby highlighting important information and ignoring unimportant information.

[0065] Adversarial learning: is a machine learning method that improves the performance and robustness of models through competitive training between models. The core idea is to introduce two competing models, called the generator and the discriminator. The generator attempts to generate samples with authenticity, while the discriminator attempts to distinguish between samples generated by the generator and real samples. The two models improve their performance through repeated iterative training. The generator improves the quality of its generated samples by learning to compete with the discriminator, while the discriminator improves its accuracy by learning to distinguish between real samples and generated samples.

[0066] In a specific implementation, step S10 also includes: data collection and preprocessing, collecting interactive session data between users and the system. These data usually include detailed records of multiple users interacting with multiple projects in different time periods, expressed as ,in Represents the user, Indicates the timestamp, Represents the interaction between users and items. The raw data is cleaned, duplicates are removed, missing values ​​are processed, and items are standardized to ensure data consistency and completeness.

[0067] Session segmentation and mapping: The preprocessed interaction data is segmented according to the session, expressed as in Represents a session. Each session consists of several interactive items, which have a certain order in time. Next, each session is regarded as a hyperedge and mapped one by one into the hypergraph. In this process, the uniqueness of the items and the complete mapping of the item set in the session to a hyperedge are ensured.

[0068] Hyperedge construction and connection: In the hypergraph, create a hyperedge for each session: , where each Contains Session These hyperedges connect each other through the project nodes, which means that if two different sessions have shared project nodes, these hyperedges are connected through these shared nodes in the hypergraph.

[0069] It should be noted that the steps of obtaining target session data and modeling the target session data to generate a hypergraph include: collecting interactive session data between the user and the system, performing data cleaning and standardization preprocessing on the interactive session data to generate target session data; segmenting the data by session in the target session data, and mapping the session to a hyperedge;

[0070] A hypergraph structure is constructed based on hyperedges, each of which contains all project nodes in a session and forms cross-session associations through shared nodes.

[0071] Step S20: Apply K-hop neighbor aggregation technology to process the hypergraph to generate a domain node graph.

[0072] It can be understood that the K-hop neighbor aggregation technology includes: assigning initial feature vectors to hypergraph nodes; traversing the K-hop range to aggregate neighbor node information and generate fine-grained context features; building a domain node graph based on the aggregation results, and optimizing the connectivity and rationality of the graph structure through verification.

[0073] In the specific implementation, first, each node (i.e., project) in the hypergraph is initialized. This includes assigning an initial feature vector to each node. These features can be set based on the attributes of the project or historical interaction data. , assuming that a node has m attributes, its feature vector can be expressed as:

[0074] ;

[0075] in, Representative Node The j-th attribute value of .

[0076] K-hop neighbor aggregation: In the hypergraph, K-hop neighbor aggregation technology is applied. This technology collects the neighbor node information of each node by traversing K steps in the hypergraph. K-hop aggregation allows capturing the neighbor relationship of nodes within a multi-hop range, thereby obtaining richer contextual information. , its k-hop neighbor aggregation feature can be expressed as:

[0077] ;

[0078] in Representation Node The k-th layer neighbor set.

[0079] Domain node graph construction: Use the information obtained from K-hop neighbor aggregation to build a domain node graph. For each node , calculate its k-hop neighbor aggregation:

[0080] .

[0081] The edge set of the neighborhood node graph:

[0082] ;

[0083] For all nodes that are neighbors within the k-hop range and , there is an edge ( , ).

[0084] Graph structure optimization and verification: Once the domain node graph is built, it needs to be optimized and verified. This includes checking the connectivity of the graph, the integrity of node features, and the rationality of the overall structure. Through optimization, it is ensured that the domain node graph can accurately reflect the complex interactive relationship between users and items.

[0085] Step S30: Process the domain node graph through the graph convolutional network, analyze the length and time interval of the user interaction sequence in combination with the intent-aware attention mechanism, determine the timeliness of the user's interest and generate an initial feature matrix.

[0086] It should be noted that the steps of processing the domain node graph through the graph convolutional network, analyzing the length and time interval of the user interaction sequence in combination with the intent-aware attention mechanism, judging the timeliness of the user's interest and generating an initial feature matrix include: initializing the dual-channel convolutional network, setting the initial weights and bias vectors, the dual-channel convolutional network includes a first channel and a second channel; using the first channel to perform a convolution operation on the hypergraph to generate node features; using the second channel to process the domain node graph to generate interaction features; aggregating the node features with the interaction features and applying the intent-aware attention mechanism to capture the timeliness of the user's interest; judging the timeliness of the user's interest and generating an initial feature matrix.

[0087] In the specific implementation, the dual-channel convolutional network is initialized: two graph convolutional networks (GCNs) are initialized: one for processing the hypergraph (hypergraph convolution) and the other for processing the domain node graph (neighborhood graph convolution). Set the initial weights and the bias vector :

[0088] ;

[0089] ;

[0090] Hypergraph convolution processing: Use the first channel to perform convolution operations on the hypergraph. This process captures high-order correlation information between sessions by aggregating the features of the nodes in each hyperedge and generates more expressive node features. For each node in the hypergraph , the convolution operation at layer l can be expressed as:

[0091] ;

[0092] Among them, N(i) is the neighbor set of node i, is the normalization coefficient, is the activation function (ReLU function).

[0093] Neighborhood graph convolution processing: Using the second channel, the graph convolution network is applied to process the domain node graph. The neighborhood graph convolution reflects the interaction dynamics within the first-order and local neighborhood by aggregating the features of each node and its direct neighbors, ensuring that the generated node features can express the important information and relationships of the nodes at the local level. For each node in the domain node graph, , the convolution operation at layer l can be expressed as:

[0094] ;

[0095] in, , is the degree of nodes i and j.

[0096] Application of intent-aware attention mechanism: After aggregating the two convolutional features, the intent-aware attention mechanism is applied. This mechanism analyzes the length and time interval of the user interaction sequence, dynamically adjusts the feature weights, and highlights the part that is most closely related to the user's current intention, thereby effectively capturing the timeliness of the user's interest. and its neighbors , calculate the attention weight:

[0097] ;

[0098] Among them, a is the attention mechanism weight vector, and || represents the vector connection operation.

[0099] Timeliness judgment of user interests and strategy adjustment: Combine the results of dual-channel convolution and intent-aware attention mechanism to judge the timeliness of user interests:

[0100] ;

[0101] in, is the attention weight of node i, is a node feature.

[0102] By analyzing user sequence features, the model can identify items with high relevance to the user's current interests and adjust the recommendation strategy.

[0103] It can be understood that after analyzing the user's interests, the initial feature matrix is ​​constructed: the initial feature matrix is ​​constructed, which integrates the node features extracted from the two-channel convolutional network, assuming that the node features of the hypergraph convolution and the neighbor node graph are: and The initial feature matrix can be set as :

[0104] .

[0105] Step S40: Input the initial feature matrix into the KAN network, integrate the explicit features and extract the latent features through nonlinear feature mapping.

[0106] It can be understood that the steps of inputting the initial feature matrix into the KAN network, integrating the explicit features through nonlinear feature mapping and extracting the latent features include: assuming that the KAN network has L layers, and the weight matrix of each layer is and the bias vector , select the activation function as ReLU:

[0107] ;

[0108] The initial feature matrix Input into the KAN network for nonlinear feature mapping:

[0109] ;

[0110] The input features are combined and transformed layer by layer through multiple hidden layers, and the ability of piecewise linear mapping is used to extract potential features.

[0111] Step S50: introducing an adversarial learning module, generating perturbation features through a generative adversarial network and performing alternating training.

[0112] It should be noted that the adversarial learning module is introduced to generate perturbation features through the generative adversarial network and perform alternating training; the adversarial learning module is initialized. The adversarial learning module consists of a generator G and a discriminator D. The parameters of the generator are and the parameters of the discriminator are ; Use the generator to perturb the input feature z:

[0113] ;

[0114] Generating Adversarial Examples , adversarial examples add tiny perturbations in the feature space:

[0115] ;

[0116] in, is a small perturbation coefficient used to control the amplitude of the perturbation;

[0117] During the adversarial training process, the discriminator D aims to maximize the ability to distinguish between original features and adversarial features:

[0118] ;

[0119] The goal of the generator G is to minimize the ability of the discriminator, that is:

[0120] .

[0121] It should be noted that through adversarial learning, the model gradually generates more refined feature representations. These features can not only effectively cope with the interference of adversarial samples, but also maintain good performance under different data distributions, reflecting good generalization ability.

[0122] Step S60: Input the output result of the KAN network and the embedding matrix after adversarial training into the prediction layer to generate the final recommendation result. It can be understood that the step of inputting the output result of the KAN network and the embedding matrix after adversarial training into the prediction layer to generate the final recommendation result includes: inputting the feature of the output result of the KAN network The embedding matrix E after adversarial training is input into the prediction layer so that the prediction layer can generate the final prediction result. :

[0123] ;

[0124] in, and are the weight matrix and bias vector of the fully connected neural network, Indicates vertical concatenation of the feature matrix and the embedding matrix;

[0125] The obtained prediction results are normalized using the softmax function to convert them into probability distribution:

[0126] ;

[0127] in, is the normalized output, It is an exponential operation on each unnormalized prediction result. It ensures that the output is in the range (0,1) and the sum of all outputs is 1.

[0128] It should be noted that this embodiment proposes a hypergraph convolution-driven multi-level conversation recommendation method that integrates hypergraphs and intra-domain node networks. First, the hypergraph is processed by K-hop neighbor aggregation technology to construct an intra-domain node network, thereby fully capturing the complex interaction dynamics between users and items. Subsequently, the graph convolution network (GCN) technology is used, combined with an innovative intent-aware attention mechanism, to accurately judge the timeliness of user interests by analyzing the length and time interval of the sequence. The generated initial feature matrix is ​​input into the kan network (Kolmogorov-Arnold Net), which can efficiently integrate existing features and extract additional implicit features, significantly improving the expressiveness and interpretability of the model. To further enhance performance and robustness, an adversarial training module is introduced into the method to generate a more refined and generalizable feature representation. Through a multi-level and multi-angle feature extraction and learning process, combined with the efficient function approximation capability of the KAN network, this method can effectively process complex network structure data and greatly improve the accuracy, efficiency and reliability of the recommendation system. Finally, the hypergraph feature matrix obtained from the KAN network and the embedding matrix obtained through adversarial training are input into the prediction layer to generate the final prediction result. Compared with the existing technology, this method shows significant advantages in processing long-tail item recommendations, tracking dynamic changes in user interests, and realizing personalized recommendations. It is especially suitable for e-commerce and content recommendation systems in complex scenarios.

[0129] This embodiment obtains the target session data to generate a hypergraph; applies K-hop neighbor aggregation technology to generate a domain node graph; processes the domain node graph through a graph convolutional network to determine the timeliness of user interests and generate an initial feature matrix; inputs the initial feature matrix into the KAN network to extract potential features; introduces an adversarial learning module to generate perturbation features through a generative adversarial network and perform alternating training; inputs the output of the KAN network and the embedded matrix after adversarial training into the prediction layer to generate the final recommendation result. This achieves the technical effect of accurately capturing the dynamic interaction between users and items in conversational recommendations.

[0130] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present application. In practical applications, technicians in this field can select part or all of it according to actual needs to achieve the purpose of the present embodiment, and no limitation is made here.

[0131] In addition, for technical details not described in detail in this embodiment, please refer to the hypergraph convolution-driven multi-level conversation recommendation method provided in any embodiment of the present application, which will not be repeated here.

[0132] In addition, it should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0133] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0134] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory (ROM) / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods of each embodiment of the present application.

[0135] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A multi-level conversation recommendation method driven by hypergraph convolution, characterized in that: include: Acquire target session data, and model according to the target session data to generate a hypergraph; Applying K-hop neighbor aggregation technology to process the hypergraph to generate a domain node graph; The domain node graph is processed through a graph convolutional network, and the length and time interval of the user interaction sequence are analyzed in combination with the intent-aware attention mechanism to determine the timeliness of the user's interest and generate an initial feature matrix; Inputting the initial feature matrix into the KAN network, integrating explicit features and extracting latent features through nonlinear feature mapping; Introducing an adversarial learning module, generating perturbation features through a generative adversarial network and performing alternating training; The output result of the KAN network and the embedding matrix after adversarial training are input into the prediction layer to generate the final recommendation result; The step of obtaining target session data and modeling according to the target session data to generate a hypergraph includes: Collecting interactive session data between the user and the system, and performing data cleaning and standardization preprocessing on the interactive session data to generate target session data; Splitting data by session in the target session data, and mapping the sessions into hyperedges; Constructing a hypergraph structure according to the hyperedges, each hyperedge including all project nodes in a session, and forming cross-session associations through shared nodes; The step of processing the domain node graph through a graph convolutional network, analyzing the length and time interval of the user interaction sequence in combination with the intent-aware attention mechanism, determining the timeliness of the user's interest and generating an initial feature matrix includes: Initializing a dual-channel convolutional network, setting initial weights and bias vectors, wherein the dual-channel convolutional network includes a first channel and a second channel; Performing a convolution operation on the hypergraph using the first channel to generate node features; Processing the domain node graph using the second channel to generate interaction features; Aggregating the node features with the interaction features and applying an intent-aware attention mechanism to capture the timeliness of user interests; The timeliness of the user's interests is judged and an initial feature matrix is ​​generated.

2. The hypergraph convolution driven multi-level conversation recommendation method according to claim 1, characterized in that: The K-hop neighbor aggregation technology includes: Assign initial feature vectors to hypergraph nodes; Traverse the K-hop range to aggregate neighbor node information and generate fine-grained context features; A domain node graph is constructed based on the aggregation results, and the connectivity and rationality of the graph structure are optimized through verification.

3. The hypergraph convolution driven multi-level conversation recommendation method according to claim 1, characterized in that: The step of inputting the initial feature matrix into the KAN network, integrating explicit features through nonlinear feature mapping and extracting latent features comprises: Assume that the KAN network has L layers, and the weight matrix of each layer is and the bias vector , select the activation function as ReLU: The initial feature matrix Input into the KAN network for nonlinear feature mapping: The input features are combined and transformed layer by layer through multiple hidden layers, and the ability of piecewise linear mapping is used to extract potential features.

4. The hypergraph convolution driven multi-level conversation recommendation method according to claim 1, characterized in that: The adversarial learning module is introduced to generate perturbation features through a generative adversarial network and perform alternating training; Initialize the adversarial learning module, which consists of a generator G and a discriminator D. The parameters of the generator are and the parameters of the discriminator are ; Use the generator to perturb the input feature z: Generating Adversarial Examples , the adversarial sample adds a small perturbation in the feature space: in, is a small perturbation coefficient used to control the amplitude of the perturbation; During the adversarial training process, the discriminator D aims to maximize the ability to distinguish between original features and adversarial features: The goal of the generator G is to minimize the ability of the discriminator, that is: 。 5. The hypergraph convolution driven multi-level conversation recommendation method according to claim 1, characterized in that: The step of inputting the output result of the KAN network and the embedding matrix after adversarial training into the prediction layer to generate the final recommendation result includes: The characteristics of the output results of the KAN network The embedding matrix E after adversarial training is input into the prediction layer so that the prediction layer can generate the final prediction result. : in, and are the weight matrix and bias vector of the fully connected neural network, Indicates vertical concatenation of the feature matrix and the embedding matrix; The obtained prediction results are normalized using the softmax function to convert them into probability distribution: in, is the normalized output, is the exponential operation on each unnormalized prediction result.

Images