A conversation recommendation system based on decoupled representation learning

Abstract

The application belongs to the technical field of data mining and analysis, and particularly relates to a conversation recommendation system based on decoupled representation learning, comprising a data acquisition module, a decoupled representation learning module, a view construction module, a conversation embedding module and an aggregation module. The preference of a user may be based on certain factors, so the recommendation problem of SBR is refined to the feature level, which is more in line with the actual situation of the user when shopping, and the method is not limited to any scene and any data, has strong versatility, and also comprehensively considers the intra-conversation and inter-conversation information, so that the model has more available information, improves the recommendation performance, and the learning on the intra-conversation and inter-conversation information is limited to the feature level, in the learning process, the model autonomously and selectively learns effective information on certain features, which is more flexible, and is significantly helpful for noise reduction.

Description

Technical Field

[0001] This invention belongs to the field of data mining and analysis technology, specifically relating to a conversation recommendation system based on decoupled representation learning. Background Technology

[0002] Big data refers to massive amounts of data of unprecedented scale, variety, speed, and value. With the development of the internet and the Internet of Things (IoT), people's demand for data is increasing, while the speed and scale of data generation are also constantly growing. The emergence of big data technology enables the effective storage, management, and analysis of this data, thereby uncovering valuable information and knowledge. Against this backdrop of big data, recommendation systems have emerged. A recommendation system is an intelligent system that uses big data to analyze user behavior and interests. By analyzing vast amounts of data such as users' historical behavior, preferences, and social network information, it predicts products that users may be interested in and recommends them to them. Recommendation systems can help users quickly find products or content they like, improving user satisfaction and loyalty. Simultaneously, they can help businesses increase sales and user engagement.

[0003] Conversation-based recommender systems (SBR) are an emerging type of recommender system. They consider a short interaction by any anonymous user as a conversation and analyze the user's preferences solely based on that conversation to make recommendations. Therefore, SBR does not require any historical user data, resulting in strong versatility. Furthermore, SBR's focus on short-term user interests aligns well with current recommendation scenarios such as streaming media, thus attracting significant attention in recent years. Previous SBR models only focused on information within a single conversation, analyzing the connections between various products within a conversation to assess user interests from two perspectives: the user's overall interest performance during the conversation and the shifts in interest generated during the conversation.

[0004] However, a significant limitation of this type of model is that the information within the session is usually insufficient. This is because most sessions are very short, making it impossible for the model to analyze the two problems mentioned above. This makes it difficult to improve the performance of models that only analyze information within the session. Furthermore, user interests are a whole, but this ignores the fact that user interests are feature-driven, meaning that a user's preference for a certain product is often based on their preference for certain features of that product. Summary of the Invention

[0005] The purpose of this invention is to provide a conversational recommendation system based on decoupled representation learning, which can simulate the various latent features behind the products, and make these feature-level representations independent by weakening the distance correlation between these latent features, thereby eliminating redundant information, and introducing an attention mechanism to synthesize this information to finally obtain the user's preferences.

[0006] The specific technical solution adopted by this invention is as follows:

[0007] A conversation recommendation system based on decoupled representation learning includes a data acquisition module, a decoupled representation learning module, a view construction module, a conversation embedding module, and an aggregation module.

[0008] The data acquisition module is used to acquire a session recommendation dataset, wherein the session recommendation dataset includes a training set, a validation set, and a test set;

[0009] The decoupled representation learning module is used to acquire user interests related to products and refine user interests into multiple independent feature levels;

[0010] The view construction module is used to construct intra-session graphs and inter-session graphs based on the feature level of the product, and to determine the information propagation strategies of the intra-session graphs and inter-session graphs respectively;

[0011] The session embedding module is used to perform session embedding according to the information propagation strategy of the view;

[0012] The aggregation module is used to integrate information within and between sessions, score and sum all features of the product to obtain a comprehensive score for the product, and determine the user's satisfaction with the product based on the comprehensive score. The higher the comprehensive score, the higher the user's satisfaction.

[0013] In a preferred embodiment, the data acquisition module includes a data filtering unit, which is used to determine valid data and invalid data based on the number of times the product appears and the session length.

[0014] The data filtering unit is equipped with a primary filtering threshold and a secondary filtering threshold;

[0015] The first-level screening threshold is used to compare with the number of times a product appears, and product data with a number of appearances less than the first-level screening threshold is marked as invalid data, while other product data is marked as parameters to be screened.

[0016] The secondary filtering threshold is used to compare with the session length corresponding to the parameter to be filtered, and when the session length corresponding to the parameter to be filtered is less than or equal to the secondary filtering threshold, the corresponding product data is filtered out.

[0017] In a preferred embodiment, when the decoupled representation learning module is executed, the product is embedded into different spaces for independent representation, and the representation of the corresponding feature level is learned through different feature level embedding layers to obtain multiple mutually independent feature levels.

[0018] In a preferred embodiment, the view construction module includes an intra-session graph construction unit and an inter-session graph construction unit, wherein the intra-session graph construction unit is used to build a feature-level intra-session graph, and the inter-session graph construction unit is used to build a feature-level inter-session graph.

[0019] In a preferred embodiment, when the session inner graph construction unit is executed, an original session inner graph is established, where each vertex of the original session inner graph corresponds to each item appearing in the session, and the edges of the original session inner graph correspond to the sequential relationship in the session. Then, based on the original session inner graph, multiple feature-level sessions are evolved one by one to obtain multiple feature-level session inner graphs, wherein the edges of the original session inner graph and the feature-level session inner graphs are directional.

[0020] In a preferred embodiment, when the session graph construction unit is executed, an initial original session graph is established, wherein each vertex of the original session graph corresponds to a session of a product, and the vertex is the average pooling result of the representation of all products. Then, multiple feature-level session graphs are evolved based on the original session graph, wherein the edges of the original session graph and the feature-level session graphs are undirected.

[0021] In a preferred embodiment, the information propagation strategy includes information propagation of feature-level intra-session graphs and information propagation of feature-level inter-session graphs;

[0022] The information propagation of the feature-level intra-session graph adopts gated neural network convolution, and the number of rounds of the gated neural network convolution is set to three layers;

[0023] The information propagation of the feature-level inter-session graph adopts graph neural network convolution, and after each round of information propagation, the embedding of the convolution node is updated to the regularized embedding of the previous round plus the average of the embeddings of the surrounding nodes in the previous round.

[0024] In a preferred embodiment, the session embedding module includes intra-session session embedding and inter-session session embedding. The intra-session session embedding extracts intra-session graph information based on an attention mechanism and measures the importance of other items based on the last item the user interacted with. The inter-session session embedding is the vertex of the inter-session graph.

[0025] In a preferred embodiment, when the aggregation module is executed, it integrates intra-session and inter-session information to obtain a comprehensive score for each product interacting with the user, then sorts the products according to the comprehensive scores, and recommends products to the user one by one according to the sorting results.

[0026] A conversation recommendation terminal based on decoupled representation learning includes:

[0027] At least one processor;

[0028] and a memory communicatively connected to the at least one processor;

[0029] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to execute the aforementioned conversation recommendation system based on decoupled representation learning.

[0030] The technical effects achieved by this invention are as follows:

[0031] This invention takes into account that user preferences may be based on certain factors, and therefore refines the SBR recommendation problem to the feature level. This is more in line with the actual situation of users when shopping. Moreover, the method is not limited to any scenario or any data, and has strong versatility. It also comprehensively considers intra-session and inter-session information, giving the model more usable information and improving recommendation performance. Furthermore, the learning of intra-session and inter-session information is limited to the feature level. During the learning process, the model will autonomously and selectively learn effective information from certain features, making it more flexible and significantly helping with noise reduction. Attached Figure Description

[0032] Figure 1 This is a system module diagram provided by the present invention;

[0033] Figure 2 This is the system execution diagram provided by the present invention. Detailed Implementation

[0034] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0035] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0036] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in a preferred embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0037] Please see Figure 1 and Figure 2 As shown, the present invention provides a conversation recommendation system based on decoupled representation learning, including a data acquisition module, a decoupled representation learning module, a view construction module, a conversation embedding module, and an aggregation module;

[0038] The data acquisition module is used to acquire the session recommendation dataset, which includes a training set, a validation set, and a test set.

[0039] The decoupled representation learning module is used to acquire user interests related to products and refine user interests into multiple independent feature levels;

[0040] The view building module is used to construct intra-session and inter-session diagrams based on the feature level of the product, and to determine the information propagation strategies for the intra-session and inter-session diagrams respectively;

[0041] The session embedding module is used to perform session embedding based on the view's information propagation strategy;

[0042] The aggregation module is used to integrate information within and between sessions, score and sum all features of the product to obtain a comprehensive score for the product, and determine the user's satisfaction with the product based on the comprehensive score. The higher the comprehensive score, the higher the user's satisfaction.

[0043] In this invention, with the popularization and development of the Internet, people increasingly rely on online platforms for information acquisition, communication, and entertainment in their daily lives. In this process, conversation recommendation systems, as tools that can recommend relevant conversations based on users' interests and needs, have gradually become an important component of various online platforms. However, existing conversation recommendation systems still have certain limitations in processing large-scale data and meeting personalized needs. Decoupled representation learning is a method that maps high-dimensional data to a low-dimensional space. By reducing the dimensionality of the data, computational complexity can be effectively reduced, and model performance can be improved. In conversation recommendation systems, decoupled representation learning can help us better uncover users' interests and needs, thereby improving the accuracy and satisfaction of recommendations. In this embodiment, the conversation recommendation dataset is first collected through a data acquisition module. The dataset is then processed, and the input for each batch is set (for example, if 100 conversations are generated within the same range, then 100 conversations are considered as a batch). Depending on the dataset, all conversations from the last week or last month can be used to form the test set, and the rest as the training set. The proportion of training sets generated later is taken as 1%. 0% of the data is used as the validation set. If a product appears only in the test set but not in the training set, it is deleted. If the session length is less than 2 after deletion, the session is deleted. Then, the decoupled representation learning module is used to refine user interests, obtaining multiple independent feature levels related to the product. Then, the view construction module is used to construct intra-session and inter-session graphs, as well as information propagation strategies for intra-session and inter-session graphs. Then, the session embedding module executes session embedding according to the information propagation strategy of the view. Finally, the aggregation module integrates intra-session and inter-session information to obtain the comprehensive score of the product. Based on the comprehensive score, the user's satisfaction with the product is determined, thereby determining the products to recommend to the user. Unlike traditional recommendations, which usually directly compare products with user interests to score each product and select the highest-scoring products to recommend to the user, in this scheme, the product should correspond to multiple embeddings to match the user's interest in multiple features. Then, the embedding of the product on any feature is compared with the user's interest to obtain the user's liking for the product feature, or the product's score for the user. Finally, the scores on multiple features are summed to obtain the comprehensive score of the product.

[0044] In a preferred embodiment, the data acquisition module includes a data filtering unit, which is used to determine valid and invalid data based on the number of times the product appears and the session length.

[0045] The data filtering unit has a primary filtering threshold and a secondary filtering threshold.

[0046] The first-level filtering threshold is used to compare with the number of times a product appears, and product data with a number of appearances less than the first-level filtering threshold is marked as invalid data, while other product data is marked as parameters to be filtered;

[0047] The secondary filtering threshold is used to compare with the session length corresponding to the parameter to be filtered, and the corresponding product data is filtered out when the session length corresponding to the parameter to be filtered is less than or equal to the secondary filtering threshold.

[0048] In this implementation, the first-level filtering threshold is set to 3 to 5 times, and the second-level filtering threshold is set to 2 to 3 times. The second-level filtering threshold is compared with the session length only after the first-level filtering threshold and the number of times the product appears are completed, so as to avoid repeated comparison of invalid data. When the session length corresponding to the parameter to be filtered is less than or equal to the second-level filtering threshold, the corresponding product data is filtered out, and other data is marked as valid data and classified into the training set, test set and validation set respectively.

[0049] In a preferred implementation, when the decoupled representation learning module is executed, the product is embedded into different spaces for independent representation. The representation of the corresponding feature level is learned through different feature level embedding layers to obtain multiple mutually independent feature levels.

[0050] In this implementation, the purpose of decoupled representation learning is to learn independent feature-level representations by embedding items into different spaces. However, the generalized re-embedding actually adds a linear layer to the network, while the embedding in decoupled representation learning takes independence into account, striving to learn the corresponding feature-level representations through different feature-level embedding layers. Therefore, decoupled representation learning introduces distance correlation to measure the correlation of data in different spaces, and also takes weakening distance correlation as a learning objective to ensure that the embedding of different feature levels in each space is independent during the training process.

[0051] In a preferred embodiment, the view building module includes an intra-session graph building unit and an inter-session graph building unit. The intra-session graph building unit is used to build feature-level intra-session graphs, and the inter-session graph building unit is used to build feature-level inter-session graphs.

[0052] Secondly, when the session inner graph construction unit is executed, an original session inner graph is created. Each vertex of the original session inner graph corresponds to each product that appears in the session, and the edges of the original session inner graph correspond to the order relationship in the session. Then, based on the original session inner graph, multiple feature-level sessions are evolved one by one to obtain multiple feature-level session inner graphs. Among them, the edges of the original session inner graph and the feature-level session inner graphs are directional.

[0053] In this implementation, when the session in-graph construction unit is executed, it first establishes the most basic session in-graph. Each vertex of the session in-graph corresponds to each product that appears in the session, and the edges of the session in-graph correspond to the sequential relationship in the session. For example, if A points to B, it means that the user visited A first and then visited B in the session. In this view, the weight of all edges is 1, which only indicates the connection relationship. Then, based on this graph, multiple session in-graphs of various feature levels need to be evolved. These views have the same vertex and edge pointing relationship. The only difference is the weight, which is calculated by the cosine similarity of two related vertices (this is a technical means commonly used by those in the art and will not be elaborated on here). In this way, the same edge has different weights at different feature levels, which corresponds to the fact that two products may be similar in one feature but dissimilar in another feature.

[0054] Secondly, when the session graph construction unit is executed, an initial original session graph is established. Each vertex of the original session graph corresponds to a session of a product. This vertex is the average pooling result of the representation of all products. Then, multiple feature-level session graphs are evolved based on the original session graph. The edges of the original session graph and the feature-level session graphs are undirected.

[0055] In the above, within the same conversation Figure 1 Similarly, an initial inter-session graph needs to be built. The vertices of this graph are no longer products but sessions. Intuitively, all products contained in a session constitute a session vertex, and the embedding of this vertex is the average pooling result of the representations of all products. When two sessions both contain at least one common product, these two sessions are considered related, and an edge is created between them. The weight of this edge is the intersection of the products contained in the two sessions divided by the union of the products in the two sessions. Next, feature-level inter-session graphs need to be generated. Within the same feature level, sessions... Figure 1 Similarly, the weight is the only point that distinguishes these graphs, meaning that the connections between two graphs are not the same on different features, and the edges are undirected.

[0056] In a preferred embodiment, the information propagation strategy includes information propagation of feature-level intra-session graphs and information propagation of feature-level inter-session graphs.

[0057] Information propagation in feature-level intra-session graphs is achieved using gated neural network convolution, with the number of rounds of the gated neural network convolution set to three layers.

[0058] Information propagation in the feature-level inter-session graph uses graph neural network convolution. After each round of information propagation, the embedding of the convolution node is updated to the regularized embedding of the previous round plus the mean of the embeddings of the surrounding nodes in the previous round.

[0059] In this embodiment, the information propagation of the view can update the more accurate vertex embeddings by considering the neighboring nodes of each vertex. Since the intra-session graph is a directed graph and the inter-session graph is an undirected graph, different convolution strategies are adopted for the feature-level intra-session graph and inter-session graph. Like many SBR (Session Recommendation System) models, gated neural network convolution is used for the propagation of the intra-session graph. Compared with traditional graph neural networks, gated neural network convolution adds a gating mechanism, which can effectively filter the propagated information during information propagation. This brings advantages to the final learning of accurate representations. In addition, the gating filtering process can avoid the over-smoothing problem of graph neural networks to a certain extent. In addition, the number of rounds of gated neural network convolution is set to three layers. Since the inter-session graph is an undirected graph, the most basic graph neural network convolution is used as the information propagation method. After each round of information propagation, the embedding of the convolution node is updated to the regularized embedding of the previous round plus the average of the embeddings of the surrounding nodes of the node in the previous round.

[0060] In a preferred embodiment, the session embedding module includes intra-session session embedding and inter-session session embedding. Intra-session session embedding extracts intra-session graph information based on an attention mechanism and measures the importance of other items based on the last item the user interacted with. Inter-session session embedding is the vertex of the inter-session graph.

[0061] In this implementation, after the information propagation updates the vertex embeddings, it is necessary to extract information from the entire graph and finally extract the information needed to make recommendations, namely, the session embedding. The session embedding is also a simulation of the embedding of user interests. When embedding within a session, the attention mechanism commonly used in SBR is used to extract information from the graph within the session. Unlike other attention mechanisms that focus on the global picture, this attention mechanism pays more attention to the impact of the last item, because the last item that the user interacts with is most likely to be close to the user's final interest. Based on the last item, the importance of other items is measured, and finally, the information is aggregated into the final session embedding according to the attention mechanism. When embedding between sessions, since the vertices of the inter-session graph are the session embeddings, they can be directly used as the final learned session representation. In addition, since each feature-level graph is extracted as a session embedding corresponding to different features, one session corresponds to multiple embeddings, which correspond to the simulated user's interest in multiple features.

[0062] In a preferred implementation, when the aggregation module is executed, it integrates the information within and between sessions to obtain a comprehensive score for each product that interacts with the user, then sorts the products according to the comprehensive scores, and recommends products to the user one by one according to the sorting results.

[0063] In this implementation, after learning effective information from the intra-session graph and inter-session graph, it is necessary to integrate the intra-session and inter-session information. Simple weighted integration is difficult to balance the ratio of intra-session information to inter-session information, especially since inter-session information inevitably contains some unwanted out-of-session information, which may also hinder the recommendation. Therefore, an attention mechanism is used to integrate the information from the two sessions. The process mainly involves filtering out out-of-session information. The final integrated information is the session embedding output by the model, which is a simulation of user interests. Unlike traditional recommendations, which usually directly compare products with user interests to score each product and select the highest-scoring products to recommend to the user, products should correspond to multiple embeddings to match the user's interests in multiple features. The embedding of the product on any feature is compared with the user's interest to obtain the user's preference for that feature, or the product's score for the user. Finally, the scores on multiple features are summed to obtain the comprehensive score of the product.

[0064] The present invention also provides a session recommendation terminal based on decoupled representation learning, comprising:

[0065] At least one processor;

[0066] and memory that is communicatively connected to at least one processor;

[0067] The memory stores a computer program that can be executed by at least one processor, which enables the at least one processor to execute the aforementioned conversation recommendation system based on decoupled representation learning.

[0068] Those skilled in the art will understand that the session recommendation terminal described in this invention can be specifically designed and manufactured for the desired purpose, or it may include known devices in general-purpose computers. These devices have computer programs or applications stored therein, which can be selectively activated or reconfigured. Such computer programs can be stored in a device (e.g., computer)-readable medium or in any type of medium suitable for storing electronic instructions and coupled to a bus, including but not limited to any type of disk (including floppy disks, hard disks, optical disks, CD-ROMs, and magneto-optical disks), ROM (Read-Only Memory), RAM (Random Access Memory), EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, magnetic cards, or optical cards. That is, readable media includes any medium by which a device (e.g., a computer) stores or transmits information in a readable form.

[0069] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0070] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A conversation recommendation system based on decoupled representation learning, comprising a data acquisition module, a decoupled representation learning module, a view construction module, a conversation embedding module, and an aggregation module, characterized in that: The data acquisition module is used to acquire a session recommendation dataset, wherein the session recommendation dataset includes a training set, a validation set, and a test set; The decoupled representation learning module is used to acquire user interests related to products and refine user interests into multiple independent feature levels; The view construction module is used to construct intra-session graphs and inter-session graphs based on the feature level of the product, and to determine the information propagation strategies of the intra-session graphs and inter-session graphs respectively; The session embedding module is used to perform session embedding according to the information propagation strategy of the view; The aggregation module is used to integrate information within and between sessions, score and sum all features of the product to obtain a comprehensive score for the product, and determine the user's satisfaction with the product based on the comprehensive score. The higher the comprehensive score, the higher the user's satisfaction.

2. The conversation recommendation system based on decoupled representation learning according to claim 1, characterized in that: The data acquisition module includes a data filtering unit, which is used to determine valid and invalid data based on the number of times a product appears and the length of the session. The data filtering unit is equipped with a primary filtering threshold and a secondary filtering threshold; The first-level screening threshold is used to compare with the number of times a product appears, and product data with a number of appearances less than the first-level screening threshold is marked as invalid data, while other product data is marked as parameters to be screened. The secondary filtering threshold is used to compare with the session length corresponding to the parameter to be filtered, and when the session length corresponding to the parameter to be filtered is less than or equal to the secondary filtering threshold, the corresponding product data is filtered out.

3. The conversation recommendation system based on decoupled representation learning according to claim 1, characterized in that: When the decoupled representation learning module is executed, the product is embedded into different spaces for independent representation. The representation of the corresponding feature level is learned through different feature level embedding layers to obtain multiple mutually independent feature levels.

4. The conversation recommendation system based on decoupled representation learning according to claim 1, characterized in that: The view construction module includes an intra-session graph construction unit and an inter-session graph construction unit. The intra-session graph construction unit is used to build a feature-level intra-session graph, and the inter-session graph construction unit is used to build a feature-level inter-session graph.

5. A conversation recommendation system based on decoupled representation learning according to claim 4, characterized in that: When the session inner graph construction unit is executed, it establishes an original session inner graph. Each vertex of the original session inner graph corresponds to each product that appears in the session, and the edges of the original session inner graph correspond to the sequential relationship in the session. Then, based on the original session inner graph, multiple feature-level sessions are evolved one by one to obtain multiple feature-level session inner graphs. The edges of the original session inner graph and the feature-level session inner graphs are directional.

6. A conversation recommendation system based on decoupled representation learning according to claim 4, characterized in that: When the session graph construction unit is executed, it establishes an initial original session graph. Each vertex of the original session graph corresponds to a session of a product. The vertex is the average pooling result of the representation of all products. Then, it evolves multiple feature-level session graphs based on the original session graph. The edges of the original session graph and the feature-level session graphs are undirected.

7. The conversation recommendation system based on decoupled representation learning according to claim 1, characterized in that: The information propagation strategy includes information propagation within feature-level intra-session graphs and information propagation between feature-level inter-session graphs; The information propagation of the feature-level intra-session graph adopts gated neural network convolution, and the number of rounds of the gated neural network convolution is set to three layers; The information propagation of the feature-level inter-session graph adopts graph neural network convolution, and after each round of information propagation, the embedding of the convolution node is updated to the regularized embedding of the previous round plus the average of the embeddings of the surrounding nodes in the previous round.

8. A conversation recommendation system based on decoupled representation learning according to claim 1, characterized in that: The session embedding module includes intra-session session embedding and inter-session session embedding. Intra-session session embedding extracts intra-session graph information based on an attention mechanism and measures the importance of other products based on the last product the user interacted with. Inter-session session embedding is the vertex of the inter-session graph.

9. A conversation recommendation system based on decoupled representation learning according to claim 1, characterized in that: When the aggregation module is executed, it integrates the information within and between sessions to obtain a comprehensive score for each product that interacts with the user. Then, it sorts the products according to the comprehensive scores and recommends products to the user one by one based on the sorting results.

10. A conversation recommendation terminal based on decoupled representation learning, characterized in that: include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to execute the conversation recommendation system based on decoupled representation learning as described in any one of claims 1 to 8.

Images