Intelligent product recommendation method and system based on large model

Abstract

The invention relates to an intelligent product recommendation method and system based on a large model. The method comprises the steps of obtaining user information of a user to be recommended and product information of a product to be recommended; extracting user semantic features, user behavior features and natural language understanding features of the user information and product description features and product category features of the product information by using the large model, and performing feature fusion to obtain fused features; and obtaining a target recommendation score of each product relative to the to-be-recommended user, and recommending the product corresponding to the target recommendation score higher than a preset score to the to-be-recommended user as a target product. According to the method, the Transform model, the GNN model and the LLM model are combined, analysis and feature fusion of the user information and the product information are achieved, the extraction precision of the user semantic features and the product description features is improved, the user behavior features and the product category features can be effectively captured, and therefore the recommendation accuracy and the personalized level are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a product intelligent recommendation method and system based on a large model. Background Art

[0002] With the rapid development of e-commerce, product recommendation systems are becoming increasingly critical in improving user experience and increasing sales conversion. However, existing recommendation methods such as collaborative filtering and simple deep learning models have limitations when dealing with complex user behaviors and product descriptions. Existing methods rely on simple statistics or keyword matching when extracting user and product features, which makes it difficult to deeply understand semantic information and deep product features, resulting in inaccurate recommendation results that cannot meet personalized needs. At the same time, there is also the problem of incomplete consideration of features.

[0003] Therefore, it is urgent to provide a technical solution to solve the above problems. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a product intelligent recommendation method and system based on a large model.

[0005] In a first aspect, the present invention provides a product intelligent recommendation method based on a large model, and the technical solution of the method is as follows: Obtain user information of a user to be recommended and product information of multiple products to be recommended; The user semantic features, user behavior features and natural language understanding features of the user information, as well as the product description features and product category features of the product information are extracted using a large model, and feature fusion is performed to obtain fusion features; wherein the large model includes: a Transformer model, a GNN model and an LLM model; Based on the fusion features, an initial product recommendation model is constructed using a deep neural network, and the initial product recommendation model is trained to obtain a target product recommendation model; The fusion feature is input into the target product recommendation model to obtain the target recommendation score of each to-be-recommended product relative to the to-be-recommended user, and the to-be-recommended product corresponding to the target recommendation score higher than the preset score is recommended as the target product to the to-be-recommended user.

[0006] The beneficial effects of the product intelligent recommendation method based on a large model of the present invention are as follows: The method of the present invention realizes the analysis and feature fusion of user information and product information by combining the Transformer model, the GNN model and the LLM model, which not only improves the extraction accuracy of user semantic features and product description features, but also can effectively capture user behavior characteristics and product category characteristics, thereby improving the accuracy and personalization level of recommendations.

[0007] On the basis of the above scheme, the product intelligent recommendation method based on a large model of the present invention can also be improved as follows.

[0008] In an optional manner, the step of extracting user semantic features, user behavior features, and natural language understanding features of the user information, and product description features and product category features of the product information using a large model includes: Utilizing the Transformer model and the LLM model, parsing the user information and the product information, and extracting the user semantic features and the product description features; The GNN model is used to parse the interaction data between the user information and the product information, and an interaction graph is constructed through the interaction data to extract the user behavior characteristics of the user information and the product category characteristics of the product information.

[0009] In the above optional methods, the Transformer and LLM models are used to parse user and product information, which improves the accuracy of extracting user semantic features and product description features. At the same time, by using the GNN model to build an interaction graph, it can effectively capture user behavior characteristics and product category characteristics, and enhance the ability to express complex interaction relationships.

[0010] In an optional manner, it also includes: weighting the user behavior characteristics through a time decay mechanism to obtain weighted user behavior characteristics to reflect the impact of the recent behavior of the user to be recommended on the target product.

[0011] In the above optional method, applying the time decay mechanism to weight user behavior characteristics can more accurately reflect the impact of recent user behavior on recommendation results, and improve the sensitivity and responsiveness to dynamic changes in user interests during actual recommendations.

[0012] In an optional manner, the step of performing feature fusion to obtain fused features includes: The user semantic feature, the weighted user behavior feature, the natural language understanding feature, the product description feature, and the product category feature are fused by using a weighted summation method to obtain the fused feature.

[0013] Among the above optional methods, the weighted summation method is used for feature fusion, which can better integrate multiple features, thereby obtaining more comprehensive and accurate fusion features and improving the expressiveness of the subsequent target recommendation model.

[0014] In an optional manner, the step of constructing an initial product recommendation model using a deep neural network based on the fusion features includes: Determining the network structure of the deep neural network according to the fusion features; According to the network structure, the weights of the deep neural network are randomly initialized, and an initial recommendation score is obtained through a forward propagation process to complete the construction of the initial product recommendation model; Wherein, the network structure includes: an input layer, a hidden layer and an output layer; The input layer is used to receive the fused features and transmit the fused features to the hidden layer; The hidden layer is used to perform nonlinear transformation and multi-level feature extraction on the received fusion features to obtain a high-level feature representation that characterizes the complex relationship between the to-be-recommended user and each to-be-recommended product; The output layer is used to perform a linear transformation on the high-level feature representation to generate the initial recommendation score of each to-be-recommended product relative to the to-be-recommended user.

[0015] In the above optional method, by determining the network structure of the deep neural network according to the fusion features, randomly initializing the weights of the deep neural network, and then generating the initial recommendation score through forward propagation, an efficient initial product recommendation model can be constructed. The network structure design of the initial product recommendation model can not only make full use of the rich information in the fusion features, but also improve the personalization and accuracy of the recommendation through deep feature extraction. At the same time, a reasonable network structure and weight initialization method also accelerate the convergence speed of the model and improve the training efficiency.

[0016] In an optional manner, the step of training the initial product recommendation model to obtain a target product recommendation model includes: The initial product recommendation model is iteratively trained using the Adam optimization algorithm to obtain the target product recommendation model.

[0017] In the above optional method, the Adam optimization algorithm is used to iteratively train the initial recommendation model, which can optimize the model parameters more quickly and stably, improve the model's convergence speed and recommendation performance, and ultimately obtain a more accurate target product recommendation model.

[0018] In a second aspect, the present invention provides a product intelligent recommendation system based on a large model, and the technical solution of the system is as follows: The product intelligent recommendation system based on the big model includes: an acquisition module, a feature module, a construction module and a recommendation module; The acquisition module is used to: acquire user information of the user to be recommended and product information of multiple products to be recommended; The feature module is used to: use the big model to extract user semantic features, user behavior features and natural language understanding features of the user information, as well as product description features and product category features of the product information, and perform feature fusion to obtain fusion features; wherein the big model includes: Transformer model, GNN model and LLM model; The construction module is used to: construct an initial product recommendation model based on the fusion features using a deep neural network, and train the initial product recommendation model to obtain a target product recommendation model; The recommendation module is used to: input the fusion feature into the target product recommendation model, obtain the target recommendation score of each to-be-recommended product relative to the to-be-recommended user, and recommend the to-be-recommended product corresponding to the target recommendation score higher than the preset score as the target product to the to-be-recommended user.

[0019] The beneficial effects of the product intelligent recommendation system based on a large model of the present invention are as follows: The system of the present invention realizes the analysis and feature fusion of user information and product information by combining the Transformer model, the GNN model and the LLM model, which not only improves the extraction accuracy of user semantic features and product description features, but also can effectively capture user behavior characteristics and product category characteristics, thereby improving the accuracy and personalization level of recommendations.

[0020] Based on the above solution, the product intelligent recommendation system based on a large model of the present invention can also be improved as follows.

[0021] In an optional manner, the feature module is specifically used to: Utilizing the Transformer model and the LLM model, parsing the user information and the product information, and extracting the user semantic features and the product description features; The GNN model is used to parse the interaction data between the user information and the product information, and an interaction graph is constructed through the interaction data to extract the user behavior characteristics of the user information and the product category characteristics of the product information.

[0022] In the above optional methods, the Transformer and LLM models are used to parse user and product information, which improves the accuracy of extracting user semantic features and product description features. At the same time, by using the GNN model to build an interaction graph, it can effectively capture user behavior characteristics and product category characteristics, and enhance the ability to express complex interaction relationships.

[0023] In a third aspect, a technical solution of an electronic device of the present invention is as follows: It includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, the steps of the product intelligent recommendation method based on a large model of the present invention are implemented.

[0024] In a fourth aspect, the present invention provides a computer-readable storage medium having the following technical solution: Instructions are stored in the computer-readable storage medium. When the computer-readable storage medium reads the instructions, the computer-readable storage medium executes the steps of the large model-based product intelligent recommendation method of the present invention.

[0025] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings are only used to illustrate the embodiments and are not to be considered as limiting the present invention. In addition, the same reference symbols are used to represent the same components throughout the accompanying drawings. In the accompanying drawings: Figure 1 A schematic diagram of a flow chart of an embodiment of a method for intelligent product recommendation based on a large model of the present invention; Figure 2 A schematic diagram of the structure of an embodiment of a product intelligent recommendation system based on a large model of the present invention; Figure 3 The figure is a schematic structural diagram of an embodiment of an electronic device of the present invention. DETAILED DESCRIPTION

[0027] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0028] Figure 1 FIG. 1 is a flow chart showing an embodiment of a method for intelligent product recommendation based on a large model provided by the present invention, and the method is executed by a controller. Figure 1 As shown, the following steps are included: S1. Obtain user information of the user to be recommended and product information of multiple products to be recommended. In S1: 1) User information refers to various data related to the user to be recommended (i.e., the user who needs to receive the recommendation), which can help understand the user's characteristics, interests and behavior patterns. User information includes but is not limited to: basic attributes (such as user age, gender, occupation, geographic location and other demographic information), historical behavior data (such as user browsing history, click records, purchase records, evaluation and scoring records, etc.), interaction data (such as user clicks, favorites, shares, comments, etc. on products), semantic information (such as user search keywords, comment content, feedback text, etc. on the platform), and other preferences (such as user's clearly expressed interest tags, subscribed topics, preferred categories, etc.). By analyzing user information, we can better understand user preferences and recommend products that are more in line with user interests and needs. User information is mainly used to extract user semantic features, user behavior features and natural language understanding features in the recommendation process. The above multiple features can help build a target product recommendation model and generate personalized recommendation results.

[0029] 2) Product information refers to various attributes and descriptive data related to the recommended product (i.e., the product that needs to be recommended to the user), which can help understand the characteristics and categories of the product, so as to better match the user's needs. Product information includes but is not limited to: basic attributes (such as the product's name, brand, model, price, category and other basic information), descriptive information (such as detailed description of the product, function introduction, specification parameters, etc.), user evaluation (such as other users' evaluation, rating, comment text, etc. of the product), sales data (such as product sales, inventory status, promotional information, etc.), categories and tags (such as product classification information and associated tags). Product information is used to extract product description features and product category features in the recommendation process. The product description features and product category features will be combined with multiple features extracted from user information to build a target recommendation model and generate a recommendation score, and finally recommend the most relevant products to users.

[0030] Specifically, the controller obtains user information of a user to be recommended and product information of a plurality of products to be recommended.

[0031] S2. Using the big model to extract user semantic features, user behavior features, and natural language understanding features of the user information, as well as product description features and product category features of the product information, and perform feature fusion to obtain fusion features; wherein the big model includes: a Transformer model, a GNN model, and an LLM model. In S2: 1) User semantic features are features extracted by semantic analysis of user information (such as user search keywords, comments, feedback and other text data). User semantic features reflect the user's needs and interests at the natural language level. User semantic features can help understand the user's potential needs and interests, especially when the user does not have clear click or purchase behavior, semantic features can supplement the user portrait and improve the accuracy of recommendations. User semantic features can use Transformer models (such as BERT, GPT, etc.) and LLM models (large language models) to parse user text information, thereby extracting deep semantic information.

[0032] 2) User behavior features are features extracted based on the user's historical behavior data (such as browsing, clicking, purchasing, collecting, evaluating, etc.), reflecting the user's actual operations and preferences. User behavior features directly reflect the user's interests and preferences, and are one of the most direct and key features. By analyzing user behavior, products that are more in line with their historical interests can be recommended to users. User behavior features can use the GNN model (graph neural network) to parse the interaction data between users and products, and extract user behavior features by constructing an interaction graph. User behavior features can also be weighted through a time decay mechanism to reflect the impact of users' recent behavior on recommendation results.

[0033] 3) Natural language understanding features are features extracted through deep semantic analysis of natural language content generated by users (such as comments, feedback, questions, etc.). Natural language understanding features focus on understanding the user's language expression and the intention and emotion behind the language. Natural language understanding features can help better understand the user's emotions and needs, especially when the user generates a large amount of text data, natural language understanding features can provide richer user information. Natural language understanding features can use the LLM model (large language model) to parse and extract features from the user's natural language data.

[0034] 4) Product description features are features extracted by analyzing the product description information (such as the product's text description, function description, detailed parameters, etc.), reflecting the basic information and functional characteristics of the product. Product description features can help understand the characteristics and uses of the product, so as to better match user needs. Product description features are especially important for new products that do not have rich user behavior data. Product description features can use the Transformer model and LLM model to parse the product's text description information to extract deep semantic features.

[0035] 5) Product category features are features extracted by analyzing the category information of products (such as the classification, label, category, etc. to which the product belongs), reflecting the category attributes of the product and its position in the commodity system. Product category features can help understand the types and attributes of products, so that recommendations can be made based on user interest categories, which is especially important for category filtering and coarse-grained recommendations. Product category features can use the GNN model to parse the interaction data between users and products and extract product category features by building an interaction graph.

[0036] 6) Fusion features are feature representations obtained by combining and integrating multiple different types of features (such as user semantic features, user behavior features, natural language understanding features, product description features, and product category features). The purpose of fusion features is to integrate feature information of multiple different dimensions to provide a more comprehensive and rich feature representation. Fusion features can simultaneously consider the user's semantic information, behavioral preferences, natural language expressions, and product descriptions and category attributes, thereby improving the accuracy and personalization of recommendations. Feature fusion can be achieved through weighted fusion, splicing fusion, or deep neural network automatic learning of fusion weights, and can integrate different types of features into a unified feature representation for input into the subsequent target product recommendation model for further processing and prediction.

[0037] 7) Transformer is a neural architecture for processing sequence data, used for natural language processing (NLP) tasks such as machine translation. The Transformer model can be used to process and analyze sequence data, especially text data, such as user comments, product descriptions, etc. Furthermore, the Transformer model can parse user text information (such as comments, feedback) and product text description information, extract semantic features, and use pre-trained models such as BERT and GPT to extract high-quality semantic representations. The Transformer model can efficiently capture long-distance dependencies and complex semantic information, so it is very effective in extracting user semantic features and product description features. Through the Transformer model, a deep semantic representation can be obtained, thereby better understanding user needs and product features.

[0038] 8) The GNN model (graph neural network) is a neural network specifically designed to process graph-structured data. Users and products can be represented as nodes in the graph, and the interactions between users and products (such as clicks, purchases, and reviews) can be represented as edges in the graph. The GNN model can effectively capture the complex relationships and interaction patterns between users and products, and therefore plays a significant role in extracting user behavior features and product category features. Through the GNN model, a user-product interaction graph can be constructed, and deep interaction features and category features can be extracted from the interaction graph. For example, the GNN model can identify the potential interests of users and the similarities between products.

[0039] 9) The LLM model (Large Language Model) is a large-scale pre-trained language model based on the Transformer network structure, such as GPT-3, T5, etc. The LLM model is trained on a large-scale data set and can understand and generate natural language. The LLM model performs well in natural language processing tasks and can deeply understand the user's natural language expression, so it plays a significant role in extracting natural language understanding features. Through the LLM model, it is possible to parse natural language texts such as user comments, feedback, questions, etc., and extract semantic and emotional information from natural language texts. For example, the LLM model can understand the emotional tendencies and potential needs in user comments.

[0040] Specifically, the controller further uses the large model to extract user semantic features, user behavior features, and natural language understanding features of the user information, as well as product description features and product category features of the product information, and performs feature fusion to obtain fused features.

[0041] S3. Based on the fusion features, an initial product recommendation model is constructed using a deep neural network, and the initial product recommendation model is trained to obtain a target product recommendation model. In S3: 1) A deep neural network (DNN) is an artificial neural network with multiple hidden layers. It can perform nonlinear transformations on input data through multiple levels and extract high-level feature representations of the data. The nodes at each level process the input data through weighted summation and activation functions and pass it to the next layer. Furthermore, deep neural networks are used to build the initial product recommendation model. The main function of deep neural networks is to perform nonlinear transformations and multi-level feature extraction on fused features to obtain high-level feature representations that characterize the complex relationship between the user to be recommended and each product to be recommended. High-level feature representations can better capture the potential associations between users and products, thereby improving the accuracy of recommendations.

[0042] 2) The initial product recommendation model refers to the model obtained by randomly initializing the weights of the deep neural network according to the fusion features and the predefined network structure before the training process begins. At this time, the model has not been trained, and the output of the initial product recommendation model (initial recommendation score) is generated based on random weights. The role of the initial product recommendation model is to provide the infrastructure for the subsequent training process. Through the forward propagation process, the initial product recommendation model can generate the initial recommendation score for each product to be recommended relative to the user to be recommended. Although the initial recommendation score is not accurate enough, it provides a starting point for the training of the initial product recommendation model.

[0043] 3) The target product recommendation model refers to the initial product recommendation model optimized through the training process. During the training process, the Adam optimization algorithm is used to iteratively optimize the initial product recommendation model and adjust the network weights so that the recommendation score of the initial product recommendation model gradually approaches the true value. The final optimized model is the target product recommendation model. The role of the target product recommendation model is to provide accurate product recommendations. Through the training process, the target product recommendation model can better understand user preferences and product features, thereby generating a more accurate target recommendation score. By recommending the recommended products corresponding to the target recommendation scores that are higher than the preset scores to users, personalized and intelligent product recommendations are achieved.

[0044] Specifically, the controller further constructs an initial product recommendation model based on the fusion features using a deep neural network, and trains the initial product recommendation model to obtain a target product recommendation model.

[0045] S4, inputting the fusion feature into the target product recommendation model, obtaining the target recommendation score of each to-be-recommended product relative to the to-be-recommended user, and recommending the to-be-recommended product corresponding to the target recommendation score higher than the preset score as the target product to the to-be-recommended user, in S4: 1) The target recommendation score refers to the recommendation score generated by the target product recommendation model for each product to be recommended after the fusion features are input into the trained target product recommendation model. The target recommendation score reflects the recommendation priority or suitability of each product to be recommended relative to the user to be recommended. Specifically, the target recommendation score is a numerical value generated by the output layer after the fusion features are processed by a deep neural network. The main function of the target recommendation score is to quantify the recommendation value of each product to be recommended relative to the user to be recommended. By comparing the recommendation scores of different products, it can be determined which products are more in line with the user's interests and needs. The higher the target recommendation score, the more suitable the product is to be recommended to the user. Ultimately, products corresponding to target recommendation scores higher than the preset score will be recommended to the user.

[0046] 2) The preset score is a pre-set threshold used to determine which products to be recommended can be recommended to users. Only products with a target recommendation score higher than the preset score will be recommended. The preset score can be adjusted according to business needs, user preferences or other strategic factors. The preset score serves as a screening criterion to filter out products that are unsuitable or have a low recommendation priority. By setting the preset score, the length and quality of the recommendation list can be controlled to ensure that the products recommended to users have a high recommendation value. The preset score provides a flexible mechanism that allows the target product recommendation system to be adjusted according to different application scenarios and user needs.

[0047] Specifically, the controller further inputs the fusion feature into the target product recommendation model, obtains the target recommendation score of each to-be-recommended product relative to the to-be-recommended user, and recommends the to-be-recommended product corresponding to the target recommendation score higher than the preset score as the target product to the to-be-recommended user.

[0048] The technical solution of this embodiment realizes the analysis and feature fusion of user information and product information by combining the Transformer model, GNN model and LLM model, which not only improves the extraction accuracy of user semantic features and product description features, but also can effectively capture user behavior characteristics and product category characteristics, thereby improving the accuracy and personalization level of recommendations.

[0049] In an optional manner, the step of extracting user semantic features, user behavior features, and natural language understanding features of the user information, and product description features and product category features of the product information using a large model includes: Utilizing the Transformer model and the LLM model, parsing the user information and the product information, and extracting the user semantic features and the product description features; The GNN model is used to parse the interaction data between the user information and the product information, and an interaction graph is constructed through the interaction data to extract the user behavior characteristics of the user information and the product category characteristics of the product information.

[0050] It should be noted that the user information of the user to be recommended and the product information of the multiple products to be recommended can be obtained through the platform's log system, database and third-party data source.

[0051] In this embodiment, before using the Transformer model and the LLM model, the user information and product information need to be preprocessed. Further, the user information and text information are converted into text data, and then the text data is converted into a token sequence that can be accepted by the Transformer model and the LLM model, and the length of all input sequences is ensured to be consistent. By using a pre-trained word embedding model (such as Word2Vec, GloVe, or directly using the embedding layer of the Transformer model), the token sequence is converted into a vector representation.

[0052] For extracting multiple features using the Transformer model, the preprocessed user information text and product information text are first input into the embedding layer of the Transformer model, respectively, so as to obtain the embedding representation of user information and the embedding representation of product information under the Transformer model. Then, the embedding representation of user information and the embedding representation of product information under the Transformer model are respectively input into the encoder of the Transformer model, and the semantic features of user information and the semantic features of product information under the Transformer model are respectively obtained through the multi-head self-attention mechanism and the feedforward neural network. The semantic features of user information and the descriptive features of product information under the Transformer model are extracted from the output of the encoder. The semantic features of user information under the Transformer model can capture high-level semantic information such as user preferences and interests; the descriptive features of product information under the Transformer model can capture descriptive information such as product attributes and characteristics.

[0053] For extracting multiple features using the LLM model, the preprocessed user information text and product information text are first input into the embedding layer of the LLM model, respectively, so as to obtain the embedding representation of user information and the embedding representation of product information under the LLM model. Then, the embedding representation of user information and the embedding representation of product information under the LLM model are respectively input into the context encoder of the LLM model, and through the multi-layer Transformer structure, the semantic features of user information and the semantic features of product information under the LLM model are respectively obtained. The semantic features of user information and the descriptive features of product information under the LLM model are extracted from the output of the LLM model. The semantic features of user information under the LLM model can capture the user's interests, preferences and other information more comprehensively; the semantic features of product information under the LLM model can describe the various attributes and characteristics of the product in more detail.

[0054] The user semantic features and product description features under the Transformer model are fused with the user semantic features and product description features under the LLM model to obtain the final user semantic features and product description features.

[0055] Before using the GNN model, you first need to collect and organize the interaction data between user information and product information from log files or databases. The interaction data includes: user behavior records such as clicking, browsing, purchasing, and evaluating products. Then clean the interaction data to remove noise and outliers in the interaction data (such as invalid clicks, duplicate records, etc.). Next, format the interaction data into a form that builds a graph structure (such as a triple form of user ID, behavior type, and product ID). Use the collected and cleaned interaction data to build a user-product interaction graph. The nodes of the interaction graph represent users and products, and the edges of the interaction graph represent the interaction behaviors between users and products. Among them, the user in the user node is a node, and the node ID is the user ID; each product is a node, and the node ID is the product ID; the interaction behavior between the user and the product is an edge, and the weight of the edge can be determined according to the number of interactions, interaction type (such as click, purchase), etc.; the attributes of the edge can include information such as interaction type, interaction time, and number of interactions; the graph structure of the interaction graph can be undirected or directed. If the interaction is bidirectional (such as evaluation and feedback), an undirected graph is used. If the interaction is unidirectional (such as click and browse), a directed graph is used.

[0056] In order to extract user behavior features and product category features by parsing the interaction graph using the GNN model, the interaction graph is first embedded, including node embedding and edge embedding. Furthermore, node embedding is to embed user nodes and product nodes into a low-dimensional vector space, which can be achieved using graph embedding methods (such as Node2Vec, DeepWalk) or directly using the embedding layer of the GNN model. Edge embedding is to embed interaction edges into a low-dimensional vector space to capture information about interaction behavior. Then, a graph convolution operation is performed, using the graph convolution layer of the GNN model to perform convolution operations on the interaction graph, aggregate node and edge information, and gradually extract high-level features through multi-layer graph convolution. In the graph convolution operation, by propagating information on the interaction graph and updating the representation of nodes and edges, the representation of each node and edge can contain information about the neighboring nodes of the node currently being updated. Finally, through the initial embedding in the graph embedding stage and the information aggregation and updating in the graph convolution stage, the embedded representation of user nodes and the embedded representation of product nodes under the GNN model can be obtained. After multiple layers of graph convolution, the above two embedded representations contain rich node features and graph structure information, which can be directly used in the feature extraction stage to obtain user behavior features and product category features. Furthermore, user behavior features are extracted from the embedded representation of user nodes, and product category features are extracted from the embedded representation of product nodes. Among them, user behavior features can capture information such as user interaction patterns and preferences; product category features can capture product category information, related products and other information.

[0057] In the work of the graph convolution layer, when there are multiple users to be recommended, suppose there is an interaction graph , where V is the set of nodes and E is the set of edges. Each node v∈V has an initial representation (feature vector) .

[0058] In the In the convolutional layer, the representation of node v is is achieved by aggregating node v’s neighbor nodes in Layer Representation (where u is a neighbor node of v) and node v is in Layer Representation To update, the expression is: ; in, It is The weight matrix of the layer graph convolution; AGGREGATE is the aggregation function used to collect information from neighboring nodes; Represents the set of neighbor nodes of node v; is a non-linear activation function (such as ReLU).

[0059] Therefore, suppose there is a user-product interaction graph, where the user node includes u 1 and u 2 ; Product nodes include p 1 、p 2 ; The interaction edge is: (u 1 , p 1 ), (u 1 , p 2 ), (u 2 , p 1 ); In the graph convolution layer, the user node u 1 The neighbor is the product node p 1 and p 2 , product node p 1 The neighbor is user node u 1 and u 2 .

[0060] In the In the layer graph convolution operation, node u 1 Representation Can be aggregated through node u 1 Neighbor node p 1 Representation and p 2 Representation Thus updating. Node p 1 Representation Can be aggregated through the node p 1 Neighbor node u 1 Representation and u 2 Representation Thus updating.

[0061] If the average aggregation function is used (i.e., the representation of neighbor nodes is averaged), the update process is as follows: For updating user node u 1 , first, get the neighbor node p 1 Representation and p 2 Representation Then, for the neighbor node p 1 The representation and neighbor node p 2 The representation of taking the average is: ; in, is the user node u 1 In the The aggregated representation of the convolutional layer, u 1 represents the user node in the interaction graph, and m represents the user node u 1 In the Aggregation results when layering; is node u 1 The set of neighbor nodes of node u 1 All directly connected nodes, assuming that node u 1 There are two neighbor nodes p 1 and p 2 ,but ; is node u 1 The number of neighbor nodes, that is, the set The size of u 1 There are two neighbor nodes, then ; is the node u 1 The neighbor node representations of are averaged. Further, the representations of all neighbor nodes are added together and divided by the number of neighbor nodes. is the neighbor node p 1 In the The representation (embedded vector) of the graph convolution layer indicates that in the previous graph convolution layer, the neighbor node p 1 Embedded representation of ; is the neighbor node p 2 In the The representation (embedded vector) of the graph convolution layer indicates that in the previous graph convolution layer, the neighbor node p 2 Embedded table of is the node u 1 The representation of two neighbor nodes of and Take the average, if u 1 If there are only two neighbor nodes, the average is the sum of the representations of the two neighbor nodes divided by 2.

[0062] By adding the neighbor node p 1 Its own representation is combined with the aggregated neighbor information and a weight matrix is ​​applied and nonlinear activation function , thereby updating the representation of the node, which can be expressed as: ;in, It is The weight matrix of the graph convolution layer, The matrix is ​​used to aggregate neighbor information Perform a linear transformation, It is the aggregated neighbor information Apply a linear transformation by transforming Multiply by the weight matrix ,The purpose is to map the aggregated neighbor information into a new space.

[0063] Therefore, through information propagation, each node (such as u 1 and p 1 ) contains not only its own features, but also the information of neighboring nodes. Information propagation is carried out layer by layer. With the stacking of graph convolutional layers, the representation of nodes will contain more and more extensive neighbor information (i.e., a wider range of graph structure information), so that GNN can effectively utilize the topological relationship in the graph structure data, learn richer node and edge representations, and better extract user behavior characteristics and product category characteristics.

[0064] Specifically, the controller further utilizes the Transformer model and the LLM model to parse the user information and the product information, and extracts the user semantic features and the product description features; utilizes the GNN model to parse the interaction data between the user information and the product information, and constructs an interaction graph through the interaction data to extract the user behavior features of the user information and the product category features of the product information.

[0065] In an optional manner, it also includes: weighting the user behavior characteristics through a time decay mechanism to obtain weighted user behavior characteristics to reflect the impact of the recent behavior of the user to be recommended on the target product.

[0066] In this embodiment, first, the historical behavior data of the user to be recommended needs to be collected and sorted. Assume that the historical behavior data of the user includes: behavior type (such as click, purchase, browse, evaluation, etc.), behavior time (such as the timestamp of the user's behavior), and behavior object (such as the product ID or specific product information that the user has interacted with).

[0067] For more convenient expression, the following symbols are defined here: Represents the user's historical behavior sequence, each behavior b i (b 1 , b 2 , ..., b n ) contains the behavior type, timestamp, and product information of the interaction; t i Indicates behavior i Timestamp of occurrence; Indicates the current timestamp; Indicates behavior i The original user behavior feature vector.

[0068] In order to reflect the timeliness of user behavior, a time decay function (exponential decay function or linear decay function, etc.) is used to calculate the weight of each behavior. Taking the exponential decay function as an example, the definition is as follows: ; Among them, w i Yesi The time decay weight; α is the decay coefficient, which controls the decay speed. A larger α value can make the impact of recent behavior greater and the impact of long-term behavior weaken rapidly. It is behavior b i The time difference occurs.

[0069] For each behavior b i , through b i The corresponding time decay weight w i Perform weighted processing to obtain weighted user behavior characteristics , the specific calculation formula is as follows: ; in, is the weighted behavioral feature; w i is the weight calculated by the time decay function; is the original user behavior feature vector.

[0070] In order to obtain the weighted user behavior characteristics of the user, all weighted behavior characteristics can be aggregated (such as by weighted summation, averaging, etc.). If the weighted summation method is adopted, the weighted user behavior characteristics can be expressed as: ; in, is the weighted user behavior feature; is the weighted behavioral feature.

[0071] Furthermore, assuming that the current timestamp is , the specific time difference and weight calculation are as follows: t1 (timestamp of clicking product A): Assume ; t2 (timestamp of purchasing product B): Assume ; t3 (timestamp of browsing product C): Assume ; Assuming the attenuation coefficient α = 0.1, the weight is calculated as follows: ; Among them, w 1 is the time decay weight of product A; w 2 is the time decay weight of product B; w 3 is the time decay weight of product C.

[0072] Assume that the original user behavior feature vector of each behavior is as follows: ; in, is the original user behavior feature vector of product A; is the original user behavior feature vector of product B; is the original user behavior feature vector of product C.

[0073] Use weight w i Weight each behavioral feature vector: ; in, is the weighted user behavior feature vector of product A; is the weighted user behavior feature vector of product B; It is the weighted user behavior feature vector of product C.

[0074] All weighted user behavior feature vectors are aggregated to obtain weighted user behavior features: ; in, It is the weighted user behavior characteristics of the user to be recommended, reflecting the impact of the recent behavior of the user to be recommended on the target product.

[0075] Get weighted user behavior characteristics Afterwards, it can also be applied to models in the recommendation terminal (such as collaborative filtering, matrix decomposition, or deep learning models, etc.) to calculate the matching degree between the user to be recommended and the product to be recommended, thereby generating a personalized recommendation list.

[0076] Specifically, the controller further performs weighted processing on the user behavior characteristics through a time decay mechanism to obtain weighted user behavior characteristics to reflect the impact of the recent behavior of the user to be recommended on the target product.

[0077] In an optional manner, the step of performing feature fusion to obtain fused features includes: The user semantic feature, the weighted user behavior feature, the natural language understanding feature, the product description feature, and the product category feature are fused by using a weighted summation method to obtain the fused feature.

[0078] In this embodiment, multiple feature vectors are fused by weighted summation to obtain the final fused feature f∈R d : ; Among them, R d It means that f is a d-dimensional feature vector, where each component is a real number; is the user semantic feature weight; is the user semantic feature vector; is the weighted user behavior feature weight; is the weighted user behavior feature vector; is the natural language understanding feature weight; is the natural language understanding feature vector; is the product description feature weight; is the product description feature vector; is the product category feature weight; is the product category feature vector.

[0079] Furthermore, assuming that the dimension of each feature vector is d=128, the weights are: is 0.2; is 0.3; is 0.2; is 0.2; is 0.1, then the process of weighted summation is as follows: Assume the following feature vectors (take the first three dimensions as an example): =[0.8, 0.5, 0.3, ...]; =[0.6, 0.4, 0.5, ...]; =[0.7, 0.5, 0.4, ...]; =[0.9, 0.6, 0.4, ...]; =[0.5, 0.4, 0.3, ...].

[0080] Then the first three dimensions of the fused feature f can be calculated as follows: f[0]=0.2×0.8+0.3×0.6+0.2×0.7+0.2×0.9+0.1×0.5=0.73; f[1]=0.2×0.5+0.3×0.4+0.2×0.5+0.2×0.6+0.1×0.4=0.48; f[2]=0.2×0.3+0.3×0.5+0.2×0.4+0.2×0.4+0.1×0.3=0.37.

[0081] Therefore, the first 3 dimensions of the fused feature vector f are [0.73, 0.48, 0.37, ...].

[0082] Specifically, the controller further performs the feature fusion on the user semantic feature, the weighted user behavior feature, the natural language understanding feature, the product description feature, and the product category feature by means of weighted summation to obtain the fused feature.

[0083] In an optional manner, the step of constructing an initial product recommendation model using a deep neural network based on the fusion features includes: Determining the network structure of the deep neural network according to the fusion features; According to the network structure, the weights of the deep neural network are randomly initialized, and an initial recommendation score is obtained through a forward propagation process to complete the construction of the initial product recommendation model; Wherein, the network structure includes: an input layer, a hidden layer and an output layer; The input layer is used to receive the fused features and transmit the fused features to the hidden layer; The hidden layer is used to perform nonlinear transformation and multi-level feature extraction on the received fusion features to obtain a high-level feature representation that characterizes the complex relationship between the to-be-recommended user and each to-be-recommended product; The output layer is used to perform a linear transformation on the high-level feature representation to generate the initial recommendation score of each to-be-recommended product relative to the to-be-recommended user.

[0084] In this embodiment, the input layer is responsible for receiving the fused feature f, which has been fused together through the feature fusion step by the user semantic feature, the weighted user behavior feature, the natural language understanding feature, the product description feature, and the product category feature. Assuming that the dimension of the fused feature f is d, the input layer has d nodes, each of which corresponds to a dimension of the fused feature.

[0085] The hidden layer is used to perform nonlinear transformation and multi-level feature extraction on the received fusion features to obtain a high-level feature representation that characterizes the complex relationship between the recommended user and each recommended product. The hidden layer selects a multi-layer fully connected layer (i.e., multi-layer perceptron, MLP), and can select 2 to 3 hidden layers. The specific number of layers can be adjusted according to demand. Assuming that the input layer has d nodes, the first hidden layer can select 2d neurons, the second hidden layer can select d neurons, and so on. The common number of neurons can be 256, 512, 1024, etc., and the specific number is adjusted according to the data scale and task complexity. The activation function of the hidden layer can select ReLU (Rectified Linear Unit), which performs well in deep neural networks and can effectively alleviate the gradient disappearance problem.

[0086] The output layer performs a linear transformation on the high-level feature representation to generate the initial recommendation score of each product to be recommended relative to the user to be recommended. Since a recommendation score needs to be generated for each product to be recommended, the output layer has only one node, which outputs a scalar recommendation score. The activation function of the output layer can select a linear activation function, because the recommendation score is a continuous value and does not require nonlinear transformation.

[0087] In order to ensure the training stability of the network, the weights need to be initialized reasonably. Initialization methods include Xavier initialization (Glorot Initialization) and He initialization (He Initialization). The above two initialization methods can effectively avoid the gradient explosion or gradient disappearance problem. The bias is initialized to 0 or a constant close to 0.

[0088] The forward propagation process is as follows: First, the fusion feature f is input into the input layer of the network. For each hidden layer, the calculation formula is: ; in, It is The output of the layer, is a vector whose dimension depends on the The number of neurons in the layer; ReLU is the activation function, defined as ; ReLU replaces all negative values ​​in the input vector with 0, while positive values ​​remain unchanged; It is The weight matrix of the layer, is a two-dimensional matrix with dimensions ;in Indicates The number of neurons in the layer; It is The number of neurons in the layer; It is The output vector of the layer has the dimension ; It is The bias vector of the layer, dimension is , the bias vector is initialized to 0 or a value close to 0 at the beginning of training; is a matrix multiplication, the result is a vector with dimensions , It is The weight matrix of the layer is The product of the layer output vectors; is vector addition, which adds the bias vector Add to As a result, we get a dimension of Vector.

[0089] The calculation formula of the final output layer is: ; in, is the output of the model, which is a scalar (in regression problems) or a vector (in classification problems), representing the predicted recommendation score, that is, the initial recommendation score; is the weight matrix of the output layer, which is a two-dimensional matrix with dimension (in regression problems), where is the number of neurons in the output layer (i.e. the last hidden layer); is the output vector of the last hidden layer, with dimension ; is the bias of the output layer, which is a scalar (in regression problems); is a matrix multiplication, the result is a scalar (in regression problems), which is the product of the weight matrix of the output layer and the output vector of the last hidden layer; is vector addition, which will bias Add to Based on the result of .

[0090] Specifically, the controller further determines the network structure of the deep neural network based on the fusion features; according to the network structure, randomly initializes the weights of the deep neural network, and obtains the initial recommendation score through a forward propagation process to complete the construction of the initial product recommendation model.

[0091] In an optional manner, the step of training the initial product recommendation model to obtain a target product recommendation model includes: The initial product recommendation model is iteratively trained using the Adam optimization algorithm to obtain the target product recommendation model.

[0092] In this embodiment, before iterative training, the initial product recommendation model has generated an initial recommendation score through the forward propagation process. , that is, the recommendation score of each product to be recommended relative to the user to be recommended. The initial recommendation score can be expressed as: ; in, represents the initial recommendation score of the recommended product i by the user u to be recommended; is a model prediction function, which represents the prediction mechanism of the model and depends on the feature representation of users and products. And the parameters of the model , It can be a linear function, a deep neural network, or a factorization machine (FM); It is a high-level feature representation obtained by performing nonlinear transformation and multi-level feature extraction on the fusion features through the hidden layer of the deep neural network; Represents the parameters of the model.

[0093] To optimize the model, define a loss function to measure the initial recommendation score The real recommendation score The gap between them. The loss function is defined as: ; in, is the loss function of the model, which is used to measure the initial value of the model, that is, the initial recommendation score and the true value, i.e. the target recommendation score The difference between them, and a regularization term is added to prevent overfitting; is the number of users; is the quantity of the product; is a double summation symbol, which means summing all users u and all products i; is the weight of user u on product i; is the squared error between the predicted value and the true value; is the coefficient for normalizing the loss function to ensure that the value of the loss function does not depend on the number of users ; is the regularization parameter, which controls the strength of the regularization term and can be selected by cross-validation value; is the model parameter The square of the L2 norm is used for regularization. By adding regularization terms, the model is prevented from overfitting. is a regularization term used to control the size of model parameters to prevent overfitting. If it is too large, the model parameters Can be forced to be smaller to reduce the risk of overfitting.

[0094] The Adam optimization algorithm is an adaptive learning rate optimization algorithm. First, the parameters of the model are , and the parameters of the Adam optimization algorithm are initialized. The specific initialization content is as follows: Learning Rate Set to 0.001; Exponential decay rate of first-order moment estimate Set to 0.9 to control the weight of the previous gradient mean, which determines the influence of the previous gradient on the current estimate. The larger the value, the more attention is paid to the historical gradient information. Exponential decay rate of second-order moment estimate Set to 0.999, which is used to control the weight of the mean of the previous squared gradient, and determines the influence of the previous squared gradient on the current estimate. The larger the value, the more attention is paid to the historical squared gradient information. The small numerical stability parameter is set to ; Initialize the first-order moment estimate m and the second-order moment estimate v to 0.

[0095] Next, calculate the loss function For model parameters The gradient is: ; in, is the gradient, which represents the partial derivative of the loss function with respect to each parameter and is used to update the parameters; is the model parameter Gradient operator for the derivative.

[0096] By gradient It is possible to determine the direction of movement in parameter space that can most quickly reduce the value of the loss function. By calculating the gradient, it is possible to determine how to adjust the model parameters. To improve the predictive performance of the model.

[0097] Then, the biased first-order moment estimate (exponential moving average of the mean) and the biased second-order moment estimate (exponential moving average of the variance) are updated, where the first-order moment refers to the mean (i.e. expected value) and the second-order moment refers to the variance (exponential moving average of the square of the mean). In the optimization algorithm, the first-order moment and the second-order moment are used to represent the mean of the gradient and the mean of the square of the gradient, respectively. In addition, if the expected value of an estimator is not equal to the true value of the estimated parameter, the estimator is biased; if the expected value of an estimator is equal to the true value of the estimated parameter, the estimator is unbiased. The exponential moving average is a biased estimator that gradually approaches the true mean and variance over time. In the early iterations, the exponential moving average is biased. Especially at initialization, the first-order moment and the second moment will be initialized to 0, and due to the decay rate and Due to the influence of , the first-order moment and the second-order moment cannot accurately reflect the true gradient mean and variance at the beginning.

[0098] Furthermore, the update process is as follows: ; in, is the first-order moment estimate, that is, the exponential moving average of the mean of the gradient, which is used to estimate the mean of the gradient. Indicates that it is The estimated value of the iteration; is the first-order moment estimate of the previous moment, that is, The first-order moment estimate at the iteration; is the decay rate Complementary coefficients are used to balance the contribution of current gradients and historical gradients; is the second-order moment estimate, i.e., the exponential moving average of the mean of the square of the gradient, which is used to estimate the variance of the gradient. Indicates that it is The estimated value of the iteration; is the second-order moment estimate of the previous moment, that is, The second-order moment estimate at the iteration; is the square of the current gradient, i.e. The square of the gradient calculated at iteration; is the decay rate Complementary coefficients used to balance the contribution of the current squared gradient and the historical squared gradient.

[0099] In the initial stage, the exponential moving average and Since it is initialized to 0 and is affected by the exponential decay rate, it will be biased towards 0. Therefore, it is necessary to adjust it through deviation correction so that and Closer to the true mean and variance.

[0100] The specific deviation correction formula is as follows: Bias correction formula for first-order moment estimation: ; in, is the decay rate of The power reflects the impact of historical gradients on the current mean estimate; is a correction factor used to adjust Deviation; is the bias-corrected mean gradient estimate, In the initial stage, it increased value to compensate for the deviation.

[0101] Bias correction formula for second-order moment estimation: ; in, is the decay rate of Power, used to control the influence of the square of the historical gradient; is a bias correction factor used to adjust Deviation; is the estimate of the variance after bias correction, indicating that an unbiased estimate is obtained. An unbiased estimate means It is an unbiased estimate of the true gradient variance, eliminating the deviation caused by the exponential moving average in the initial stage. In the Adam optimization algorithm, Adjust the learning rate to make parameter updates more stable and efficient.

[0102] After the deviation correction, the parameters are adjusted according to the deviation-corrected gradient estimate in the Adam optimization algorithm to update the model parameters. , the parameter update formula is as follows: ; in, is the square root of the bias-corrected second moment estimate; is a small constant, usually set to , used to ensure the numerical stability of the denominator and prevent the denominator from being zero in the calculation; Represents the square root of the second-order moment after numerical stabilization. The purpose of dividing the mean of the gradient by the square root of the gradient variance is to adjust the effective step size of the gradient so that the parameter update can be adaptively adjusted according to both the direction of the gradient and the amplitude (variance) of the gradient. The parameter update amount is obtained by multiplying the ratio by the learning rate. By subtracting the update amount, the parameters are updated, allowing the model to gradually approach the optimal solution in each iteration.

[0103] Therefore, after multiple rounds of iterative training, the parameters of the model The optimization has been fully carried out to obtain the final target product recommendation model. At this point, the output of the target product recommendation model (i.e., the target recommendation score) can be expressed as: ; in, represents the target recommendation score between user u and item i by the target product recommendation model after multiple iterations; The feature vector (or latent vector) representing user u and item i can capture the implicit features of users and items; It is the parameter obtained by the model through multiple rounds of iterative optimization during the training process. It has reached a better state, can fit the training data well and has good generalization ability. It is the prediction function of the model, which calculates the recommendation score based on the input features and the trained model parameters.

[0104] Specifically, the controller further utilizes the Adam optimization algorithm to iteratively train the initial product recommendation model to obtain the target product recommendation model.

[0105] Figure 2 FIG. 2 is a schematic diagram showing a structural embodiment of a product intelligent recommendation system 200 based on a large model provided by the present invention. Figure 2 As shown, the system 200 includes: an acquisition module 210 , a feature module 220 , a construction module 230 and a recommendation module 240 .

[0106] The acquisition module 210 is used to: acquire user information of a user to be recommended and product information of a plurality of products to be recommended; The feature module 220 is used to: use the big model to extract the user semantic features, user behavior features and natural language understanding features of the user information, as well as the product description features and product category features of the product information, and perform feature fusion to obtain fusion features; wherein the big model includes: Transformer model, GNN model and LLM model; The construction module 230 is used to: construct an initial product recommendation model based on the fusion features using a deep neural network, and train the initial product recommendation model to obtain a target product recommendation model; The recommendation module 240 is used to: input the fusion features into the target product recommendation model, obtain the target recommendation score of each product to be recommended relative to the user to be recommended, and recommend the product to be recommended corresponding to the target recommendation score higher than the preset score as the target product to the user to be recommended.

[0107] In an optional manner, the feature module is specifically used to: Utilizing the Transformer model and the LLM model, parsing the user information and the product information, and extracting the user semantic features and the product description features; The GNN model is used to parse the interaction data between the user information and the product information, and an interaction graph is constructed through the interaction data to extract the user behavior characteristics of the user information and the product category characteristics of the product information.

[0108] The technical solution of this embodiment realizes the analysis and feature fusion of user information and product information by combining the Transformer model, GNN model and LLM model, which not only improves the extraction accuracy of user semantic features and product description features, but also can effectively capture user behavior characteristics and product category characteristics, thereby improving the accuracy and personalization level of recommendations.

[0109] The above-mentioned parameters and steps for each module to implement corresponding functions in the product intelligent recommendation system 200 based on the big model of this embodiment can refer to the parameters and steps in the embodiment of the product intelligent recommendation method based on the big model above, and will not be repeated here.

[0110] like Figure 3 As shown, an electronic device 300 of an embodiment of the present invention includes a processor 320, the processor 320 is coupled to a memory 310, and the memory 310 stores at least one computer program 330, and the at least one computer program 330 is loaded and executed by the processor 320, so that the electronic device 300 implements any of the above-mentioned large model-based product intelligent recommendation methods, specifically: The electronic device 300 may have relatively large differences due to different configurations or performances, and may include one or more processors 320 (Central Processing Units, CPU) and one or more memories 310, wherein the one or more memories 310 store at least one computer program 330, and the at least one computer program 330 is loaded and executed by the one or more processors 320, so that the electronic device 300 implements any one of the large model-based product intelligent recommendation methods provided in the above embodiments. Of course, the electronic device 300 may also have components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output, and the electronic device 300 may also include other components for implementing device functions, which will not be repeated here.

[0111] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A product intelligent recommendation method based on a large model, characterized in that: include: Obtain user information of a user to be recommended and product information of multiple products to be recommended; The user semantic features, user behavior features and natural language understanding features of the user information, as well as the product description features and product category features of the product information are extracted using a large model, and feature fusion is performed to obtain fusion features; wherein the large model includes: a Transformer model, a GNN model and an LLM model; Based on the fusion features, an initial product recommendation model is constructed using a deep neural network, and the initial product recommendation model is trained to obtain a target product recommendation model; The fusion feature is input into the target product recommendation model to obtain the target recommendation score of each to-be-recommended product relative to the to-be-recommended user, and the to-be-recommended product corresponding to the target recommendation score higher than the preset score is recommended as the target product to the to-be-recommended user.

2. The method for intelligent product recommendation based on a large model according to claim 1, characterized in that: The step of extracting user semantic features, user behavior features, and natural language understanding features of the user information, as well as product description features and product category features of the product information using a large model includes: Utilizing the Transformer model and the LLM model, parsing the user information and the product information, and extracting the user semantic features and the product description features; The GNN model is used to parse the interaction data between the user information and the product information, and an interaction graph is constructed through the interaction data to extract the user behavior characteristics of the user information and the product category characteristics of the product information.

3. The product intelligent recommendation method based on a large model according to claim 1 is characterized in that: Also includes: The user behavior characteristics are weighted by a time decay mechanism to obtain weighted user behavior characteristics to reflect the impact of the recent behavior of the user to be recommended on the target product.

4. The method for intelligent product recommendation based on a large model according to claim 3, characterized in that: The steps of performing feature fusion and obtaining fused features include: The user semantic feature, the weighted user behavior feature, the natural language understanding feature, the product description feature, and the product category feature are fused by using a weighted summation method to obtain the fused feature.

5. The method for intelligent product recommendation based on a large model according to claim 1, characterized in that: The step of constructing an initial product recommendation model using a deep neural network based on the fusion features includes: Determining the network structure of the deep neural network according to the fusion features; According to the network structure, the weights of the deep neural network are randomly initialized, and an initial recommendation score is obtained through a forward propagation process to complete the construction of the initial product recommendation model; Wherein, the network structure includes: an input layer, a hidden layer and an output layer; The input layer is used to receive the fused features and transmit the fused features to the hidden layer; The hidden layer is used to perform nonlinear transformation and multi-level feature extraction on the received fusion features to obtain a high-level feature representation that characterizes the complex relationship between the to-be-recommended user and each to-be-recommended product; The output layer is used to perform a linear transformation on the high-level feature representation to generate the initial recommendation score of each to-be-recommended product relative to the to-be-recommended user.

6. The method for intelligent product recommendation based on a large model according to claim 1, characterized in that: The step of training the initial product recommendation model to obtain a target product recommendation model includes: The initial product recommendation model is iteratively trained using the Adam optimization algorithm to obtain the target product recommendation model.

7. A product intelligent recommendation system based on a large model, characterized in that: include: Acquisition module, feature module, construction module and recommendation module; The acquisition module is used to: acquire user information of the user to be recommended and product information of multiple products to be recommended; The feature module is used to: use the big model to extract user semantic features, user behavior features and natural language understanding features of the user information, as well as product description features and product category features of the product information, and perform feature fusion to obtain fusion features; wherein the big model includes: Transformer model, GNN model and LLM model; The construction module is used to: construct an initial product recommendation model based on the fusion features using a deep neural network, and train the initial product recommendation model to obtain a target product recommendation model; The recommendation module is used to: input the fusion feature into the target product recommendation model, obtain the target recommendation score of each to-be-recommended product relative to the to-be-recommended user, and recommend the to-be-recommended product corresponding to the target recommendation score higher than the preset score as the target product to the to-be-recommended user.

8. The product intelligent recommendation system based on a large model according to claim 7 is characterized in that: The feature module is specifically used for: Utilizing the Transformer model and the LLM model, parsing the user information and the product information, and extracting the user semantic features and the product description features; The GNN model is used to parse the interaction data between the user information and the product information, and an interaction graph is constructed through the interaction data to extract the user behavior characteristics of the user information and the product category characteristics of the product information.

9. An electronic device, characterized in that: The electronic device includes a processor, which is coupled to a memory. The memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor so that the electronic device implements the large model-based product intelligent recommendation method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: At least one computer program is stored in the computer-readable storage medium, and the at least one computer program is loaded and executed by the processor so that the computer-readable storage medium implements the large model-based product intelligent recommendation method as described in any one of claims 1 to 6.

Images