Commodity search recommendation system based on AI

By processing product and user data from e-commerce platforms using large-scale AI-based language models and neural networks, standardized tags and embeddings are generated, solving the problems of data sparsity and cold start, achieving efficient personalized recommendations, and improving user experience and sales conversion rates.

CN118967259AInactive Publication Date: 2024-11-15NANJING XINTONG DIGITAL TECH CO LTD

Patent Information

Application Number
CN202411031726.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-30
Publication Date
2024-11-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Existing e-commerce platform product recommendation systems struggle to provide accurate and personalized recommendations when faced with data sparsity and cold start issues. Furthermore, the high complexity of feature engineering increases system development and maintenance costs, resulting in low user experience and sales conversion rates.

Method used

We employ an AI-based Large Language Model (LLM) to process product and user data, combining sentiment analysis and neural networks to generate standardized labels and embeddings, and use dynamic adjustment and adaptive learning mechanisms to optimize the recommendation model.

Benefits of technology

It improves the accuracy and personalization of recommendations, reduces the complexity and cost of feature engineering, enhances user experience and purchase conversion rates, and is suitable for global intelligent recommendations on cross-border e-commerce platforms.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118967259A_ABST
    Figure CN118967259A_ABST
Patent Text Reader

Abstract

The invention provides an AI-based commodity search recommendation system. The system comprises the steps of inputting commodity data into an LLM, generating commodity summary information, training a recommendation model and constructing a commodity category matching system. User historical evaluation is analyzed through LLM, user interest keywords are extracted, and user interest tags and commodity tags are generated; extracting features from commodity and user data, inputting the features into a neural network for model training, generating and storing; coding user and commodity features, using a neural network to carry out forward propagation and reverse propagation training, and generating and storing in a vector database; and inputting user or commodity information, and recommending a commodity most similar to a user or a commodity through model coding and calculation of cosine similarity between vectors. According to the method, artificial feature conversion, feature engineering and feature combination are omitted, LLM feature extraction and intelligent labeling are supplemented, a supervised learning method is adopted, model construction is very controllable and reliable, and recommendation accuracy is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of the Internet, and in particular to an AI-based commodity search and recommendation system. Background Art

[0002] With the rapid development of e-commerce, consumers often find it difficult to choose when faced with a large number of products. In order to improve user experience and promote product sales, e-commerce platforms widely adopt product recommendation systems. Traditional recommendation systems are mainly based on collaborative filtering and content recommendation methods. These methods can provide personalized recommendations to a certain extent, but there are also some significant problems. The collaborative filtering method is based on the user-item interaction matrix and recommends products based on the historical behavior of similar users. It is divided into user-based collaborative filtering and item-based collaborative filtering. Commonly used algorithms include K-nearest neighbor algorithm and matrix decomposition technology. The content recommendation method is based on the attributes of the product and the user's preferences. It makes recommendations by matching the characteristics of users and products, and uses text mining and information retrieval technology to extract key features from product descriptions.

[0003] The existing technology has the problem of data sparsity. The collaborative filtering method relies on the interaction data between users and products. When the data is sparse (that is, users have not rated most of the products), the recommendation effect will be greatly reduced. Data sparsity makes it difficult for the recommendation system to accurately capture the user's true preferences and needs. In addition, for the cold start problem, the system lacks sufficient interaction data for new users or new products, making it difficult to make accurate recommendations. Traditional methods perform poorly in cold start scenarios, and additional data and processing methods are needed to alleviate this problem. The complexity of feature engineering is also a problem. The content recommendation method relies on manually designed features and labeled data. The feature engineering process is cumbersome and time-consuming, requiring the participation of domain experts, which increases the cost of system development and maintenance. The semantic understanding ability of the existing technology is insufficient. Traditional methods have limited semantic understanding capabilities for user queries and product descriptions, and cannot fully capture the complex interests of users and the subtle differences in products, resulting in inaccurate recommendation results and reduced user satisfaction.

[0004] There will be some consequences if these problems are not solved. The first is the decline in user experience. Data sparsity and cold start problems make it difficult for the recommendation system to provide accurate and personalized recommendation services. Users have a poor experience when browsing and selecting products. Users cannot quickly find products that meet their needs and may turn to other e-commerce platforms. Secondly, the sales conversion rate is reduced. The accuracy and personalization of the recommendation system directly affect the exposure and click-through rate of the product, which in turn affects the sales conversion rate. If the system recommendation is not accurate enough, the user's desire to buy will decrease, and the platform's sales will be difficult to increase. Finally, the system development and maintenance costs are high. Traditional content recommendation methods rely on a large number of manually designed features and labeled data. The feature engineering process is complicated, which increases the cost of system development and maintenance. Feature engineering needs to be constantly updated and optimized, which consumes a lot of manpower and resources. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention proposes an AI-based commodity search and recommendation system, which uses LLM to identify users' evaluations of historically purchased commodities, extract user interest keywords and generate standardized labels, thereby improving the accuracy and diversity of recommendations and meeting the diverse needs of users. Moreover, by using a supervised learning method, the model construction process is controllable and reliable, the recommendation accuracy is high, and it can dynamically adapt to changes in user needs.

[0006] To achieve the above objectives, the present invention provides an AI-based commodity search and recommendation system, comprising:

[0007] Step S1: Input product data into a large language model (LLM), remove redundant information, generate product summary information, and integrate multimodal data (such as images, audio, and video) to provide a more comprehensive product description for training recommendation models and building a product category matching system;

[0008] Step S2: Analyze user historical evaluations through LLM, combine with sentiment analysis model, extract user interest keywords and sentiment information, and generate user interest tags and product tags;

[0009] Step S3: Extract features from product and user data and input them into the neural network for model training. Combine the collaborative filtering algorithm to generate and store user embedding and item embedding, and use dynamic adjustment and adaptive learning mechanisms to update the model in real time.

[0010] Step S4: Encode user and product features, use neural networks for forward propagation and back propagation training, generate user embedding and item embedding, store them in the vector database, and introduce privacy protection and security mechanisms;

[0011] Step S5: Input user or product information, encode through the model and calculate the cosine similarity between vectors, combine the multi-level recommendation strategy and intelligent dialogue system, and recommend the top-N products that are most similar to the user or product.

[0012] Further, step S1 includes:

[0013] Step S11: The product title, description, attribute data, and multimodal data such as product images, audio, and video are designed into prompt words (Prompt) and input into the large language model (LLM).

[0014] Step S12: Use LLM to process the product data, remove redundant information, generate standardized product summary information, and ensure the consistency and high quality of text, image, audio, video and other information.

[0015] Step S13: The product summary information generated in step S12 is used to train the recommendation model, provide input features, and optimize the generation capability of the LLM itself.

[0016] Step S14: Build a product category matching system, call LLM to infer the association and matching relationship between products, and map them into categories within the site to optimize the category interest expansion modeling during recommendation recall and sorting.

[0017] Further, step S2 includes:

[0018] Step S21: Identify the user's evaluation of historically purchased products through LLM, and extract implicit user interest keywords and sentiment information in combination with the sentiment analysis model.

[0019] Step S22: Use LLM to generate standardized user interest tags and product tags, analyze the user's evaluation content, extract interest keywords and sentiment information, and form a user sentiment profile.

[0020] Step S23: Based on the user behavior data and the text emotion recognition capability of LLM, identify the user's preference for the product, generate a user portrait and label it.

[0021] Further, step S3 includes:

[0022] Step S31: extract features from product and user data, input them into the neural network for model training, and combine with collaborative filtering algorithm to use the behavioral data of similar users to improve the recommendation ability of the model.

[0023] Step S32: Encode user and product information, map user features and product features into their respective high-dimensional spaces, and form independent sub-neural networks through embedding.

[0024] Step S33: During the training process, L2 regularization is used for dot product calculations, and no interaction is performed before the final cross layer.

[0025] Step S34: After the training is completed, a dynamic adjustment and adaptive learning mechanism is used to dynamically adjust the user embedding and item embedding according to the user's real-time feedback and store them in the vector database.

[0026] Further, step S4 includes:

[0027] Step S41: Classify and encode the features of users and products, use Label Encoding for category features, One-Hot encoding for sparse features, and Embedding encoding for dense features.

[0028] Step S42: Initialize parameters, use the uniform distribution method to randomly initialize the weight matrix, initialize the bias term to 0, set the learning rate to a minimum value, input the encoded feature matrix, calculate the result according to the method in step S3, and obtain the output Y.

[0029] Step S43: Calculate the gradient and use the stochastic gradient descent (SGD) method to update the network parameters along the negative direction of the loss function gradient.

[0030] Step S44: Use the binary cross entropy loss function to calculate the loss of the model output, back-propagate the gradient of the output layer to the hidden layer through the chain rule, and calculate the gradient of the model parameters.

[0031] Step S45: Repeat the forward propagation and back propagation processes for multiple iterations until the loss function converges or reaches a preset stop condition, and the training process ends.

[0032] Step S46: Introduce differential privacy technology to protect user data, ensure that user privacy is not leaked, and use security protocols and encryption technology to ensure the security of data transmission and storage.

[0033] Further, step S5 includes:

[0034] Step S51: Input user information, map it to the encoding space after model encoding, and generate the user's embedding vector through the neural network.

[0035] Step S52: Use the cosine similarity algorithm to calculate the distance between the user vector and the stored product vector, and combine the multi-level recommendation strategy (coarse-grained recommendation and fine-grained recommendation) to return the top-N products that are most similar to the user code.

[0036] Step S53: Input product information, map it to the encoding space after model encoding, and generate the embedding vector of the product.

[0037] Step S54: Use the cosine similarity algorithm to calculate the distance between the product vector and other product vectors, and return the top-N products that are most similar to the product code.

[0038] Step S55: Scenario judgment is performed based on the operation type of the front-end user, and the product categories constructed by LLM are combined for sorting and recommendation. The intelligent dialogue system interacts with the user to obtain the user's needs and feedback in real time, and dynamically adjust the recommendation strategy and content.

[0039] Furthermore, the formula in step S3 is as follows:

[0040] The model uses a binary cross entropy loss function to calculate the loss of the model output

[0041]

[0042] Where N is the number of samples, yi is the true value of the i-th sample, and pi is the model prediction probability of the i-th sample;

[0043] Finally, the cosine similarity of User Embedding and Item Embedding obtained by the neural network model is solved

[0044]

[0045] Where U and V represent vectors in the User Embedding and Item Embedding spaces, respectively.

[0046] Furthermore, the formula of step S4 is as follows:

[0047] Derived to find the gradient, using stochastic gradient descent (SGD) to update the network parameters along the negative direction of the loss function gradient;

[0048] The output of the model is ypred, the true value is ytrue, and the model uses the binary cross entropy loss function

[0049]

[0050] First, calculate the gradient of the output layer, that is, the gradient of the loss function to the output:

[0051]

[0052] Next, using the chain rule, the gradient of the output layer is back-propagated to the hidden layer, and the gradient of the loss function in the hidden layer to the model parameters is calculated.

[0053] Calculate the gradient of the output layer:

[0054]

[0055] Among them, z is the input of the output layer, that is, the value without activation function, and for the ReLU activation function

[0056] exist

[0057] Furthermore, it also includes calculating the gradients of all parameters in the network, including weights and bias terms, based on the gradients of the hidden layer. The processing method is as follows:

[0058] Compute the gradient of the parameters:

[0059]

[0060]

[0061] Update parameters: learning rate hyperparameters

[0062] α is the learning rate hyperparameter

[0063]

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1. The present invention provides an AI-based product search and recommendation system. Through the powerful semantic understanding and generation capabilities of LLM, the accuracy and personalization of product recommendations are improved. LLM can better understand the semantics of user queries and product descriptions, optimize the relevance of recommendation results, and identify users' evaluations of historically purchased products through LLM, extract user interest keywords and generate standardized labels, thereby improving the accuracy and diversity of recommendations and meeting the diverse needs of users.

[0066] 2. The present invention provides an AI-based product search and recommendation system, which eliminates the need for a large amount of labeled data and manually designed features, reduces the cost and complexity of feature engineering, and uses LLM for feature extraction to automatically generate standardized features for products and users.

[0067] 3. The present invention provides an AI-based product search and recommendation system, which uses a supervised method to optimize semantic embedding, improves the accuracy of the recommendation model, encodes and trains features through a deep neural network model, and achieves accurate recommendation results.

[0068] 4. The present invention provides an AI-based product search and recommendation system, which combines the interpretability of LLM and the automatic feature learning ability of deep neural networks, eliminating the need for manual feature engineering, and can achieve good results, especially when processing large-scale sparse data.

[0069] 5. The present invention provides an AI-based product search and recommendation system, which provides more accurate and personalized product recommendation services, improves user experience and purchase conversion rate. The system can accurately capture user interests and preferences, provide recommendations that meet user needs, and is applicable to various types of products and user groups of different sizes, with good versatility and scalability. It performs well on cross-border e-commerce platforms and in multilingual environments, and can provide global intelligent product recommendation services. It uses the cosine similarity algorithm to efficiently calculate the distance between vectors to achieve recommendation recall and sorting. It combines the product associations and matching relationships inferred by LLM to optimize the category interest expansion modeling during recommendation recall and sorting. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0071] Figure 1 This is a schematic diagram of the overall solution of the present invention.

[0072] Figure 2 This is the neural network structure diagram of the present invention DETAILED DESCRIPTION

[0073] The technical solution of the present invention will be more clearly and completely explained below through the description of the preferred embodiments of the present invention in combination with the accompanying drawings.

[0074] Terminology explanation:

[0075] 1. Large Language Model (LLM):

[0076] A deep learning-based natural language processing model that can understand and generate text. It learns the grammar, vocabulary, and semantics of a language by training on large amounts of text data.

[0077] 2. Deep Neural Network (DNN):

[0078] An artificial neural network that contains multiple hidden layers and processes data in a way that simulates the workings of the human brain. DNNs are often used for complex pattern recognition and prediction tasks.

[0079] 3. Feature Extraction:

[0080] Extract useful information or features from raw data so that machine learning algorithms can process and analyze the data more effectively.

[0081] 4. Semantic Understanding:

[0082] A computer or model's understanding of the meaning and context of text, able to identify and interpret meaning in language.

[0083] 5.Embedding:

[0084] A method for converting high-dimensional sparse data into low-dimensional dense vectors, often used in natural language processing and recommendation systems so that the model can process data more efficiently.

[0085] 6. User Embedding:

[0086] A low-dimensional representation of a user's feature vector, generated by a neural network, is used to represent the user's interests and preferences.

[0087] 7. Item Embedding:

[0088] The low-dimensional representation of the product feature vector is generated by a neural network and is used to represent the attributes and characteristics of the product.

[0089] 8. Cosine Similarity:

[0090] A metric used to measure the similarity between two vectors, expressed by calculating the cosine of the two vectors.

[0091] 9. Forward Propagation:

[0092] During the training of a neural network, the calculation process from the input layer to the output layer is performed through the weights and activation functions of each layer to generate the final output.

[0093] 10. Backward Propagation:

[0094] During neural network training, the gradient of the loss function is calculated and back-propagated from the output layer to the input layer through the chain rule to update the weights and bias terms to minimize the loss.

[0095] 11. Loss Function:

[0096] A function that measures the difference between the model's prediction results and the true value. Commonly used loss functions include mean square error (MSE) and binary cross-entropy.

[0097] 12. Activation Function: A function that introduces nonlinear characteristics. Commonly used activation functions include ReLU (Rectified Linear Unit), and the formula is:

[0098] 13. Gradient Descent:

[0099] An optimization algorithm that adjusts model parameters to minimize the loss by calculating the gradient of the loss function. Common variants include Stochastic Gradient Descent (SGD).

[0100] 14. Regularization:

[0101] In order to prevent the model from overfitting, a penalty term (such as L2 regularization) is added to the loss function to limit the size of the model parameters.

[0102] 15. Data Sparsity:

[0103] There are a lot of missing or zero values ​​in the dataset, which is common in the user-item interaction matrix in the recommendation system.

[0104] 16.Cold Start Problem:

[0105] The recommendation system has difficulty making accurate recommendations when there is not enough historical data on new users or new products.

[0106] 17. Feature Engineering:

[0107] The process of extracting and selecting features from raw data to improve model performance, either by hand-designed or automated methods.

[0108] 18. Training Set:

[0109] Data sets used to train machine learning models, through which the models adjust parameters and learn.

[0110] 19. Test Set:

[0111] The data set used to evaluate the performance of the model. The model will not be exposed to this data during the training process and is used to verify the generalization ability of the model.

[0112] 20.Supervised Learning:

[0113] A machine learning approach in which the model learns from labeled training data and predicts output labels based on the input data.

[0114] 21. Non-linear Transformation:

[0115] In a neural network, the activation function is used to perform a nonlinear transformation on the linear weighted summation result, enabling the model to represent complex relationships.

[0116] 22. Semantic Summary:

[0117] A simplified and standardized text description generated using LLM to represent the characteristics of a product or user.

[0118] like Figure 1 As shown, the present invention is specifically:

[0119] Use a large language model (LLM) to train text data to capture semantic information and contextual associations. Consider the language model as a feature extractor, design the original information of items and users (such as the title / text description / attributes / category of the product, the user's behavior sequence / context, etc.) into prompts, input them into the LLM, use the LLM to summarize the original corpus information of the product / user, generate a concise semantic expression, and output the semantic summary information, which will be used as a feature or through semantic mining of the user's potential interest preferences, and finally integrated into the decision-making process of the recommendation system for use in subsequent recommendation modules. The specific steps are as follows:

[0120] Step S1 is specifically as follows: the present invention constructs a prompt with data such as the product title, complete product text description, and attributes, thereby stimulating LLM to denoise and purify redundant text information of the product item based on its learned knowledge to generate a standard summary of the product such as the standard product name, product attributes, and other related information.

[0121] 1) Combine the product title with the text description

[0122] 2) Design prompt templates, remove redundant information in the text, remove irrelevant advertising terms or marketing language, merge repeated information, and explore features beyond the standard attributes of the product, as shown in Table 1

[0123] Table 1

[0124]

[0125] On the one hand, this generated information is used to train the recommendation model in the subsequent steps and provide input for the subsequent deep learning model; on the other hand, it is also used to iteratively optimize the generation capability of LLM itself.

[0126] In addition, LLM is also used in the present invention to build product catalog collocation. Based on the category system within the site, by calling the LLM model in batches and requesting LLM knowledge, the large model is used to infer the association and collocation between products for us, and then the LLM collocation categories are mapped to the categories within the site to build a category collocation system. The collocation I2I is introduced during the recommendation recall in "step 5 2)", and the category interest expansion modeling is performed during sorting. Here, we design some prompt templates to guide LLM to infer category collocation for us, as shown in Table 2:

[0127] Table 2

[0128]

[0129]

[0130] The more carefully designed the prompt template is, the more accurate the collocation catalogue that LLM can deduce for us will be. Then, substitute the target catalogue as a parameter into the prompt template, and batch request LLM to return the collocation catalogue for each target catalogue.

[0131] Step S2 is specifically as follows: the present invention also uses the LLM model to understand user intentions, extracts interests based on the user's behavior history, and generates intermediate semantic features for subsequent recommendation models. Make full use of the excellent text emotion recognition ability of the large model to extract the user's text evaluation of historically purchased goods, identify the user's preference for the product and generate a score; at the same time, the large model is used to identify and extract interest keywords in text evaluations, such as "exquisite packaging", "good quality", "high cost performance", etc., which are words with personal interest tendencies to generate standardized user interest tags and product tags, thereby establishing the correlation between users and products and improving the accuracy and diversity of product recommendations.

[0132] 1) We design a prompt template to analyze users’ historical product text reviews, as shown in Table 3;

[0133] Table 3

[0134]

[0135]

[0136] 2) Group the products, merge the product reviews into an array, substitute it into the prompt, and let the LLM analyze and understand it, and then return the processing results

[0137] 3) Label standardization: pre-set standard labels for product quality and price labels, and replace all labels with the most similar standard labels by LLM using text conversion capabilities.

[0138] 4) After LLM understands user behavior, it marks the products, mines products that users repurchase, and weights the basic score of LLM automatic scoring (weight value > 1) to get the weighted user score. The final score is obtained by taking the average of the user scores. The product is marked with the union of the product labels made by all users with a score greater than the average score. For conflicting labels in the same feature, the majority label is used to resolve the conflict.

[0139] 5) Label the user's profile. First, analyze the user's historical rating information extracted by LLM to determine whether the user has more positive ratings or more negative ratings. If there are more positive ratings, all the positive reviews of the user extracted by LLM are used; if the user has more negative ratings, all the negative reviews of the user extracted by LLM are used. The reviews used by the user using the above rules are merged, and the reviews extracted in the first round are sent back to LLM, which will perform a second summary of the user's preferences. For negative reviews, LLM is instructed to summarize the reasons for the negative reviews and reversely label the user. For example, if a user comments on a negative review: poor quality, it means that the user cares very much about quality.

[0140] Step S3 is specifically as follows: after extracting feature information from product and user data through statistical modeling and LLM model, input it into the constructed neural network for modeling. In the model, multi-layer neural networks are used to embed user-related features and product-related features respectively, forming two independent sub-neural networks respectively. There is no interaction between user features and product features before the last cross layer. In each layer of the hidden layer, the input is first linearly weighted and summed. The superposition of linear transformations is still linear. If there are only linear operations in the neural network, the entire network can only represent linear relationships and cannot learn and represent complex nonlinear relationships. Then, the Relu function is used to perform nonlinear transformation on the linear weighted summation result. As a nonlinear conversion function inside the neuron, the Relu function can introduce nonlinear characteristics to the network and map the linear combination of inputs to a more complex output space, so that the neural network can approximate any complex functional relationship, so that the neural network can learn complex patterns and features, and improve the network's expression and fitting capabilities. Then the results of this layer are input to the next layer for repeated iterations.

[0141] L 1 =W 1 ·X

[0142] L i =ReLu(W i L i-1 +B i ), i = 2, ..., N-1

[0143] Y=ReLu(W N L N-1 +B N )

[0144] Where Wi represents the weight matrix of the i-th layer, represents the bias term of the i-th layer, Li represents the output of the i-th layer of the network model, Y represents the output of the last layer, X represents the input matrix, that is, the matrix composed of vectors representing user or product features, and B i represents the bias term of the i-th layer, W N represents the weight matrix of the output layer, B N Represents the bias term of the output layer.

[0145] The model uses a binary cross entropy loss function to calculate the loss of the model output

[0146]

[0147] Where N is the number of samples, yi is the true value of the i-th sample, and pi is the model predicted probability of the ii-th sample.

[0148] Finally, the cosine similarity of the User Embedding and Item Embedding obtained by the neural network model is solved

[0149]

[0150] Where U and V represent vectors in the User Embedding and Item Embedding spaces, respectively.

[0151] Step S4 is specifically: model training, the neural network performs model training, and its operation steps are:

[0152] 1. Feature Encoding

[0153] First, classify and encode the characteristics of users and products

[0154] Use Label Encoding to process some category features of products or users, and automatically map each category value to an integer.

[0155] One-Hot encoding is used for some sparse features with small representation volume; Embedding encoding is used for dense features or no processing is performed and they are directly placed in the input matrix and sent to the neural network for processing.

[0156] 2. Forward Propagation

[0157] Initialize the parameters, use the uniform distribution method to randomly initialize the weight matrix W, initialize the bias B item to 0, and set the learning rate to a minimum value. After the encoded feature matrix is ​​input through calculation, the output Y is obtained by calculating the result according to the method in step 3.

[0158] 3. Back Propagation

[0159] Derivatives are obtained and the stochastic gradient descent (SGD) method is used to update the network parameters along the negative direction of the gradient of the loss function to reduce the value of the loss function.

[0160] The output of the model is ypred, the true value is ytrue, and the model uses the binary cross entropy loss function

[0161]

[0162] First, calculate the gradient of the output layer, that is, the gradient of the loss function to the output:

[0163]

[0164] Next, using the chain rule, the gradient of the output layer is back-propagated to the hidden layer, and the gradient of the loss function in the hidden layer to the model parameters is calculated.

[0165] Calculate the gradient of the output layer:

[0166]

[0167] in, represents the gradient of the loss function with respect to the input z of the output layer, represents the gradient of the loss function with respect to the predicted value, It represents the gradient of the predicted value to the input z of the output layer, where z is the input of the output layer, that is, the value without the activation function.

[0168] exist

[0169] Based on the gradient of the hidden layer, the gradient of all parameters in the network, including weights and bias terms, is calculated. The processing method is as follows:

[0170] Compute the gradient of the parameters:

[0171]

[0172]

[0173] in, is the gradient of the loss function with respect to the weight matrix W, is the gradient of the loss function with respect to the bias term b;

[0174] Update parameters: learning rate hyperparameters

[0175] α is the learning rate hyperparameter

[0176]

[0177] Where W is the updated weight matrix, W old is the weight matrix before updating, B is the bias term after updating, and B old Bias term before update

[0178] The above steps constitute one iteration. Repeat 2 and 3 for multiple iterations until the loss function converges or reaches the preset stop condition, and the training process ends.

[0179] The advantage of embedding based on the above steps is that user behavior information / item information is implicitly mined after multiple layers of linear weighting and nonlinear transformation.

[0180] Step S5 is specifically as follows:

[0181] 1) The input is user information. After the user information enters the model, the process of step 4 is repeated to re-encode the latest user Embedding vector. Then, the nearest neighbor algorithm of the vector is used to compare the result with the full item Embedding vector stored in the vector database to obtain the cosine similarity distance. The obtained distances are sorted and the top-N products most similar to the user code are returned;

[0182] 2) Input product information, obtain the product, and find the itemEmbedding of the product in the vector database index. Then use the nearest neighbor algorithm of the vector to compare the itemEmbedding vector with the full itemEmbedding vectors stored in the vector database except for the item to obtain the cosine similarity distance. Sorting is still required next. Unlike the u2i recommendation operation in 1), it is necessary to make scene judgments based on the operation type of the front-end user. Usually, for scenes where users have watched and watched again, it is necessary to find alternative products of the same type for the user. For scenes where users have bought, it is necessary to find related products for the user, such as recommending Bluetooth headsets to users after they buy a mobile phone. At this time, when sorting the distance obtained, it is necessary to expand and introduce the category directory matching built by LLM for us in step 1. After filtering through the category directory, the top-N products most similar to the itemEmbedding vector are finally returned.

[0183] As a specific implementation method,

[0184] Implementation 1: Multimodal product data processing

[0185] Step S11: Collect multimodal data:

[0186] Product title: For example, "Latest smartphones."

[0187] Product description: For example, "featuring a high-resolution camera and a long-lasting battery."

[0188] Product attribute data: such as "color: black, memory: 128GB".

[0189] Image: A high-resolution photo of the product.

[0190] Audio: Voice introduction of the product.

[0191] Video: A demonstration video of the product.

[0192] The data are integrated into a prompt: "Product name: the latest smartphone, description: with high-resolution camera and long-lasting battery, attributes: color: black, memory: 128GB, image: [image data], audio: [audio data], video: [video data]", and input into the LLM.

[0193] Step S12: Remove redundant information:

[0194] Text processing: Use natural language processing technology to remove redundant words and repeated information in product descriptions, such as descriptions where the word "new style" appears multiple times.

[0195] Image processing: Use image processing algorithms to remove noise and background clutter in images and enhance image clarity.

[0196] Audio Processing: Use audio processing algorithms to remove background noise and enhance speech clarity.

[0197] Video Processing: Use video processing techniques to remove distracting factors in the video, such as low light or shaking.

[0198] The final generated product summary information includes: "The latest smartphone with a high-resolution camera and long-life battery, color: black, memory: 128GB, high-definition pictures, clear voice introduction, demonstration video".

[0199] Step S13: Use the generated product summary information to train the recommendation model:

[0200] Input features: The generated standardized product summary information is used as input features to train the recommendation model.

[0201] Optimize LLM: Through a feedback loop, the generation ability of LLM is optimized when processing similar data, ensuring higher accuracy.

[0202] Step S14: Constructing a product category matching system:

[0203] Use LLM to infer the association and matching relationship between products, such as "smartphones and smart watches are the best match."

[0204] Map these associations to on-site categories to optimize category interest expansion modeling during recommendation recall and sorting, such as classifying smartphones and smart watches as "high-tech accessories."

[0205] Implementation 2: User Interest and Sentiment Analysis

[0206] Step S21: Analyze user reviews:

[0207] LLM is used to analyze the user's evaluation data on historically purchased products and extract keywords and sentiment. For example, from the user's comment "This phone has a very good camera and is very satisfied with its battery life", the keywords "camera" and "battery life" are extracted, and the sentiment is identified as "satisfied".

[0208] Step S22: Generate standardized labels:

[0209] The evaluation content is segmented and labeled to generate interest keywords, and the user's emotional attitude is determined through the sentiment analysis model. For example, the label "interest keywords: camera, battery life; sentiment: satisfaction" is generated.

[0210] Step S23: Generate user portrait:

[0211] Combine user behavior data (such as browsing, clicking, and purchasing records) with the sentiment analysis results of LLM to identify user preferences and degree of preference.

[0212] Generate detailed user portraits, such as "User portrait: likes high-quality camera phones, pays attention to battery life, and has high satisfaction."

[0213] Implementation 3: Feature extraction and dynamic model training

[0214] Step S31: Feature extraction and model training:

[0215] Extract features from product and user data and perform feature engineering, including normalization, standardization, and feature selection.

[0216] The data is input into a deep neural network for training, combined with a collaborative filtering algorithm, and the behavioral data of similar users is used to improve the performance of the recommendation model.

[0217] Step S32: Feature encoding and mapping:

[0218] Use the embedding layer to map user features and product features into high-dimensional vector space respectively to form independent sub-neural networks.

[0219] For example, the user feature "high-quality camera" is mapped to a point in the vector space, and the product feature "128GB memory" is mapped to another point in the vector space.

[0220] Step S33: Regularization during training:

[0221] During the training process, L2 regularization is used for dot product calculation to reduce the risk of overfitting of the model.

[0222] No feature interaction is performed before the cross layer, ensuring the simplicity and efficiency of the training process.

[0223] Step S34: Dynamic adjustment and adaptive learning:

[0224] Through online learning algorithms, user embedding and item embedding are updated in real time to ensure that the model always reflects the user's latest interests and behaviors.

[0225] The final generated embedding vector is stored in the vector database for subsequent recommendation.

[0226] Implementation 4: Feature Coding and Privacy Protection

[0227] Step S41: Feature encoding:

[0228] Use Label Encoding to encode categorical features and convert categorical values ​​into integer values.

[0229] One-Hot encoding is used for sparse features to convert each category into a binary vector.

[0230] Use Embedding encoding for dense features to map high-dimensional features into low-dimensional space.

[0231] Step S42: Parameter initialization:

[0232] The weight matrix is ​​randomly initialized using a uniform distribution method and the bias term is initialized to 0.

[0233] Set the learning rate to a minimum value, input the encoded feature matrix, and calculate the output result Y.

[0234] Step S43: Gradient calculation and update:

[0235] Calculate the gradient and use the stochastic gradient descent (SGD) method to update the network parameters along the negative direction of the gradient of the loss function.

[0236] The specific operations include calculating the gradient of the loss function and updating the weights and bias parameters of the network using the SGD algorithm.

[0237] Step S44: Loss calculation and back propagation:

[0238] Using the binary cross entropy loss function, the loss of the model output is calculated.

[0239] The gradient of the output layer is back-propagated to the hidden layer through the chain rule to calculate the gradient of the model parameters.

[0240] Step S45: Iterative training:

[0241] Repeat the forward propagation and back propagation process for multiple iterations until the loss function converges or the preset stopping condition is reached, and the training process ends.

[0242] Step S46: Privacy protection:

[0243] Differential privacy technology is introduced to add noise in data processing to protect the privacy of user data.

[0244] Use security protocols and encryption technology to ensure the security of data transmission and storage and prevent data leakage.

[0245] Implementation 5: Multi-level recommendation and intelligent interaction

[0246] Step S51: Generate user embedding vector:

[0247] Input user behavior data, encode it through the trained model, and generate the user's embedding vector to represent the user's characteristics in high-dimensional space.

[0248] Step S52: Calculate user-product similarity:

[0249] The cosine similarity algorithm is used to calculate the distance between the user vector and the stored product vector.

[0250] Combined with a multi-level recommendation strategy, coarse-grained recommendations are performed first, then fine-grained recommendations, and finally the top-N products most similar to the user code are returned.

[0251] Step S53: Generate product embedding vector:

[0252] Input product attribute data, encode it through the trained model, and generate the product's embedding vector to represent the product's characteristics in high-dimensional space.

[0253] Step S54: Calculate product-product similarity:

[0254] Use the cosine similarity algorithm to calculate the distance between the product vector and other product vectors, and return the top-N products that are most similar to the product code.

[0255] Step S55: Sorting and recommendation:

[0256] Make scene judgments based on the operation types of front-end users (such as browsing, clicking, and purchasing).

[0257] Combined with the product category collocation constructed by LLM, the product is sorted and recommended. For example, when a user browses for smartphones, related smart watches and headphones are recommended.

[0258] Interact with users through an intelligent dialogue system, obtain user needs and feedback in real time, dynamically adjust recommendation strategies and content, and provide more personalized services.

[0259] The system not only improves the accuracy and personalization of recommendations, but also adds innovation and technical advantages in multimodal data processing, sentiment analysis, dynamic adjustment, privacy protection, and intelligent interaction.

[0260] As a specific implementation case,

[0261] like Figure 1 As shown in the figure, the user information, product information and user product interaction behavior data of the platform are first collected through the data collection and transmission module. After a series of basic data sorting and cleaning operations such as sorting, deduplication, outlier processing, and default value supplementation, some static features and benchmark behavior sequences are obtained through statistical analysis.

[0262] Step L01. Use LLM to analyze product titles, product descriptions, and product tags, denoise, purify, and rewrite redundant text information of products to extract key information. Convert scattered information about products into semantically coherent text, and use the powerful language expression and logical reasoning capabilities of LLM to extract key information from product titles.

[0263] Step L02. Use LLM to regularize the product text features, and measure the semantic coherence of the regularized text through perplexity. Considering that most product titles are combinations of related words, we first segment the product titles to obtain a list of descriptive words, and then combine the product attribute list and the product text description to cleverly design the prompt template to limit LLM to select keywords to output a brief description of the product. In this way, the product title is regularized, and a semantically coherent text description is obtained, thereby forming rich product features.

[0264] Step L03. Based on the category system within the site, carefully design some prompt templates, such as: "Suppose you are a {computer installation expert}, please recommend {10} category catalogs for {computer} combinations for us", call the LLM model in batches, request LLM knowledge to obtain product catalog combinations, and then map the LLM combination categories into site categories, establish associations and mappings between site categories, and build a category combination system so that combination I2I can be introduced during recommendation recall and category interest expansion can be performed during sorting.

[0265] Step L04. Use LLM to analyze the user's query, feedback and evaluation text data, combine the user behavior sequence and context, summarize the user behavior, identify the user's preference for the product and generate a score; at the same time, use the large model to identify and extract interest keywords in the text evaluation, such as "exquisite packaging", "good quality", "high cost performance" and other words with personal interest tendencies to generate standardized user interest tags and product tags to profile the user, extract key information and context information to form the characteristics between the user and the product, so as to help the system better understand the relationship between the user and the product.

[0266] Step L05. Standardize and normalize the user interaction data extracted and standardized based on LLM, and standardize different preference weights and score the user's various behaviors such as browsing, collecting, adding to shopping carts, placing orders, repeat purchases, good reviews, and bad reviews. Given that the degree of user real-name registration in cross-border e-commerce is very low, the gender, age group, and geographic location information of user purchase behavior are supplemented by a statistical analysis method.

[0267] Step L06. Normalize the product features. Based on statistical analysis methods, use the average price of products in the same category as a benchmark to normalize the product price.

[0268] Step N01. Build a deep neural network model and construct a model structure. This structure includes: the input layer includes user features, content features and context features, which are encoded as the input of the model; the embedding layer: maps discrete product features and user features to low-dimensional dense vectors, obtains feature representations through learning, and improves the model's ability to represent features; the deep neural network layer, the deep neural network model usually includes multiple hidden layers, and learns the complex relationship between features through multi-layer neural networks; the output layer is usually a function used to output the predicted probability results. As shown in the attached Figure 2 As shown in the figure, the overall structure of the model is divided into input layer, representation layer and matching layer at the application logic level. First, the input layer processes data to obtain features; then the representation layer obtains user embedding and item embedding; and finally, the matching layer performs product recommendations.

[0269] Step N02. Associate and merge static statistics with data features extracted by LLM, and send the data to the attached Figure 2The input layer of the neural network shown in Figure 1. After constructing the Input layer, the Embedding layer of the corresponding features is constructed at the same time. User features in user feature training data include user dense features and user sparse features. The user's geographic location is used as the user's dense feature, and the dense features are embedded and encoded. The Embedding layer is used to map them to a high-dimensional space to capture more complex relationships. It can also be connected with other features without encoding and sent directly to the DNN of the representation layer. The user's sparse features are one-hot encoded and then sent to the Embedding layer. For user sparse features, first divide them into two lists according to discrete category features and sequence features. The gender and age group of the extracted user purchase behavior are used as sparse category features, and then the category feature list is traversed. Embedding layers are defined according to the vocabulary size of the category features, and then the Input layer is input into the corresponding Embedding layer; the user product interaction sequence (browsing, collection, adding to shopping cart, purchase record) and user tags and portrait information are used as sequence features, and then the sequence feature list is traversed to construct Embedding layers, and then the Input layer is input into the Embedding layer to obtain a group of vector sequences, which are pooled as needed. Similarly, the operation of product features on the product side is similar to that on the user side. When constructing the Embedding layer, the product name is used as a dense feature, the product brand and category directory are used as its sparse category features, and the product attributes and regularized features and marking labels extracted by LLM are processed as sparse sequence features. In addition, in order to increase the correlation between products, the LLM summary of product category collocation is also used as a sparse sequence feature to construct the Embedding layer.

[0270] Step N03. Divide the data set into a training set and a test set, select an appropriate data ratio, keep the data distribution of the training set and the test set consistent, and avoid bias in the evaluation results caused by uneven data distribution. According to the size of the training data set, this ratio is set at 8:2 in the present invention. For popular items, a suppression strategy is implemented, and a popularity-based random selection of negative samples algorithm is adopted. Randomly select from the global non-positive sample data as negative samples (but popular items are preferred. The more popular items are more likely to be used as negative samples. For example: users give bad reviews to products, the number of views or orders is large, but the current user has not browsed or interacted, etc.), to ensure the balance of the positive and negative sample ratios in the training set and increase the degree of model personalization.

[0271] Step N04. After the features in the input layer are spliced ​​together, they will be provided to the attached Figure 2The deep learning network models in the representation layer. User features and product features are converted into fixed-length vectors after passing through their two fully connected layers, and user embedding and item embedding with the same dimension are obtained here. The number of network layers and dimensions within each network can be different, but the dimensions of the final output must be the same so that operations can be performed in the matching layer. The user embedding and item embedding dimensions used in the present invention are both 32.

[0272] Step N05. After the model is trained, the user embedding and item embedding are obtained respectively and stored in the vector memory database. The user-side network model and the product-side network model are deployed online. The deployment timer service periodically requests the product-side network model to calculate the embedding vectors of all items and then stores and updates them. When each user request arrives, the user tower is requested to calculate the user's embedding vector.

[0273] Step N06. Deploy the network model and embedding vector online and recommend products online

[0274] 1) U2I recommendation: The vector nearest neighbor method can be used to calculate the top-N items that are most similar to each user (the vector has the highest cosine similarity). When making online recommendation recalls, the real-time user features are input into the model in combination with the user's latest behavior sequence. The model predicts the user vector and uses the vector retrieval tool to recall N similar items as candidate items as the result of the recall. After sorting, the recommended product list 4 is obtained.

[0275] Table 4

[0276] Recall Top-10 2.61% Top-50 10.20% Top-100 19.72% Top-500 64.5% Top-1000 83.19%

[0277] 2) i2i recommendation: The vector nearest neighbor method can be used to calculate the top-N items that are most similar to each item (with the highest cosine similarity), as shown in Table 5.

[0278] Table 5

[0279] Recall Top-10 3.81% Top-50 11.37% Top-100 19.91% Top-500 50.13% Top-1000 41.78%

[0280] Scenario a) of online recommendation recall: Based on the N items that the user has recently operated (such as product search, browsing or collection, adding to shopping cart, etc.), k similar items are recalled respectively, a total of N*k items as candidate items, and they are recommended to the user after being sorted.

[0281] Scenario b) Based on the product that the user is currently operating (such as browsing product details and placing an order), calculate the top-N items that are most similar to the user (with the highest cosine similarity), combine them with category interest expansion during sorting, and provide a product recommendation list.

[0282] When recommending online products, one or more recommendation recall strategies in the system can be used for specific scenarios. In the non-logged-in state, based on static statistical methods, the platform top-N products or the top-N products of the same category under the search keyword are provided to users. For the logged-in scenario, the u2i or i2i method is used for personalized recommendations to achieve the recommendation scenario that users look at again and again. Combined with the product catalog built by LLM, it can achieve the recommendation of related products in the scenario where users buy again and again.

[0283] The above specific implementations are only descriptions of the preferred implementations of the present invention, and do not limit the protection scope of the present invention. Without departing from the design concept and spirit of the present invention, various modifications, substitutions and improvements made by ordinary technicians in this field to the technical solution of the present invention based on the text description and drawings provided by the present invention should all fall within the protection scope of the present invention. The protection scope of the present invention is determined by the claims.

Claims

1. A product search and recommendation system based on AI, characterized in that: include: Step S1: Input product data into a large language model (LLM), remove redundant information, generate product summary information, and integrate multimodal data to provide product descriptions for training recommendation models and building a product category matching system; Step S2: Analyze user historical evaluations through LLM, combine with sentiment analysis model, extract user interest keywords and sentiment information, and generate user interest tags and product tags; Step S3: Extract features from product and user data and input them into the neural network for model training. Combine the collaborative filtering algorithm to generate and store user embedding and item embedding, and use dynamic adjustment and adaptive learning mechanisms to update the model in real time. Step S4: Encode user and product features, use neural networks for forward propagation and back propagation training, generate user embedding and item embedding, store them in the vector database, and introduce privacy protection and security mechanisms; Step S5: Input user or product information, encode through the model and calculate the cosine similarity between vectors, combine the multi-level recommendation strategy and intelligent dialogue system, and recommend the top-N products that are most similar to the user or product.

2. The AI-based product search and recommendation system according to claim 1, characterized in that: Step S1 includes: Step S11: designing product titles, descriptions, attribute data, and product image, audio, and video multimodal data into prompt words and inputting them into a large language model; Step S12: Use LLM to process the product data, remove redundant information, generate standardized product summary information, and ensure the consistency and high quality of text, image, audio and video information; Step S13: The product summary information generated in step S12 is used to train the recommendation model, provide input features, and optimize the generation capability of the LLM itself; Step S14: Build a product category matching system, call LLM to infer the association and matching relationship between products, and map them into categories within the site to optimize the category interest expansion modeling during recommendation recall and sorting.

3. The AI-based product search and recommendation system according to claim 1, characterized in that: Step S2 includes: Step S21: Identify the user's evaluation of historically purchased goods through LLM, and extract implicit user interest keywords and sentiment information in combination with the sentiment analysis model; Step S22: Use LLM to generate standardized user interest tags and product tags, analyze user evaluation content, extract interest keywords and sentiment information, and form a user sentiment profile; Step S23: Based on the user behavior data and the text emotion recognition capability of LLM, identify the user's preference for the product, generate a user portrait and label it.

4. The AI-based product search and recommendation system according to claim 1, characterized in that: Step S3 includes: Step S31: extract features from product and user data, input them into the neural network for model training, and combine with collaborative filtering algorithm to improve the recommendation ability of the model by using the behavior data of similar users; Step S32: Encode user and product information, map user features and product features into their respective high-dimensional spaces, and form independent sub-neural networks through embedding; Step S33: During the training process, L2 regularization is used for dot product calculation, and no interaction is performed before the last cross layer; Step S34: After the training is completed, a dynamic adjustment and adaptive learning mechanism is used to dynamically adjust the user embedding and item embedding according to the user's real-time feedback and store them in the vector database.

5. The AI-based product search and recommendation system according to claim 1, characterized in that: Step S4 includes: Step S41: Classify and encode the features of users and products, use Label Encoding for category features, One-Hot encoding for sparse features, and Embedding encoding for dense features; Step S42: Initialize parameters, use a uniform distribution method to randomly initialize the weight matrix, initialize the bias term to 0, set the learning rate to a minimum value, input the encoded feature matrix, calculate the result according to the method in step S3, and obtain the output Y; Step S43: Calculate the gradient and use the stochastic gradient descent method to update the network parameters along the negative direction of the loss function gradient; Step S44: using the binary cross entropy loss function, calculating the loss of the model output, back-propagating the gradient of the output layer to the hidden layer through the chain rule, and calculating the gradient of the model parameters; Step S45: Repeat the forward propagation and back propagation process for multiple iterations until the loss function converges or reaches a preset stop condition, and the training process ends; Step S46: Introduce differential privacy technology to protect user data, ensure that user privacy is not leaked, and use security protocols and encryption technology to ensure the security of data transmission and storage.

6. The AI-based product search and recommendation system according to claim 1, characterized in that: Step S5 includes: Step S51: input user information, map it to the encoding space after model encoding, and generate the user's embedding vector through the neural network; Step S52: Use the cosine similarity algorithm to calculate the distance between the user vector and the stored product vector, and combine the multi-level recommendation strategy to return the top-N products that are most similar to the user code; Step S53: Input product information, map it to the encoding space after model encoding, and generate the embedding vector of the product; Step S54: Use the cosine similarity algorithm to calculate the distance between the product vector and other product vectors, and return the top-N products that are most similar to the product code; Step S55: Perform scenario judgment based on the operation type of the front-end user, sort and recommend the product categories based on the LLM, and interact with the user through the intelligent dialogue system to obtain the user's needs and feedback in real time, and dynamically adjust the recommendation strategy and content.

7. The AI-based commodity search and recommendation system according to claim 4, characterized in that: The formula in step S3 is as follows: The model uses a binary cross entropy loss function to calculate the loss of the model output Where N is the number of samples, yi is the true value of the i-th sample, and pi is the model prediction probability of the i-th sample; Finally, the cosine similarity of User Embedding and Item Embedding obtained by the neural network model is solved Where U and V represent vectors in the User Embedding and Item Embedding spaces, respectively.

8. The AI-based commodity search and recommendation system according to claim 5, characterized in that: The formula of step S4 is as follows: Derived to find the gradient, using stochastic gradient descent (SGD) to update the network parameters along the negative direction of the loss function gradient; The output of the model is ypred, the true value is ytrue, and the model uses the binary cross entropy loss function L(and true ,and pred )=-y true log(y pred )-(1-and true )log(1-y pred ) (2) First, calculate the gradient of the output layer, that is, the gradient of the loss function to the output: Next, using the chain rule, the gradient of the output layer is back-propagated to the hidden layer, and the gradient of the loss function in the hidden layer to the model parameters is calculated. Calculate the gradient of the output layer: Among them, among them, represents the gradient of the loss function with respect to the input z of the output layer, represents the gradient of the loss function with respect to the predicted value, Represents the gradient of the predicted value to the input z of the output layer, where z is the input of the output layer, that is, the value without the activation function, and for the ReLU activation function exist 9. The AI-based commodity search and recommendation system according to claim 8, characterized in that: It also includes calculating the gradients of all parameters in the network, including weights and bias terms, based on the gradients of the hidden layer. The processing method is as follows: Compute the gradient of the parameters: is the gradient of the loss function with respect to the weight matrix W, is the gradient of the loss function with respect to the bias term b; Update parameters: learning rate hyperparameters α is the learning rate hyperparameter Where W is the updated weight matrix, W old is the weight matrix before updating, B is the bias term after updating, and B old The bias term before updating.

Citation Information

Patent Citations

  • Commodity recommendation system and method based on deep learning

    CN117078342A

  • Knowledge question and answer method, device and equipment and storage medium

    CN117171306A

  • Data processing method and related device

    CN117194766A

  • Commodity label generation method based on large language model and electronic equipment

    CN117788109A

  • Multi-modal large language model training method, correlation calculation method and label generation method

    CN118113901A

Cited By

  • Automobile product matching system and method based on large language model

    CN119624589A

  • An automotive product matching system and method based on large language models

    CN119624589B

  • Intelligent product recommendation method and system based on large model

    CN119624592A

  • Push strategy self-adaptive optimization method and system oriented to personalized requirements

    CN119963299A

  • Intelligent agent construction method and system for large model and intelligent recommendation

    CN120277272A