Fine arrangement and rough arrangement method based on deep neural network
Through the fine and rough arrangement method based on deep neural network, a multi-source data matrix is constructed, feature extraction and dimensionality reduction are performed, and local information refining and timing behavior analysis are combined with convolutional neural network and long-term memory network, which solves the existing technology The problem of insufficient accuracy and personalization of recommendation results has achieved more accurate and personalized user recommendations.
Patent Information
- Application Number
- CN202411494571.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-24
- Publication Date
- 2025-02-18
AI Technical Summary
The prior art is difficult to effectively capture the complex relationship between user behavior, product characteristics and environmental information, resulting in the accuracy and personalization of recommendation results being undesirable and difficult to adapt to changes in user interests and needs.
Using the fine and rough arrangement method based on deep neural network, a multi-source data matrix is constructed by obtaining user behavior records, product feature data, user portrait information and external context information, and performing feature extraction and dimensionality reduction, and dynamically hierarchically generate user preference vectors. , product feature vectors and environmental correlation vectors, combined with convolutional neural networks and long-term memory networks, local information refinement and timing behavior analysis are carried out to achieve refined dynamic sorting.
The ranking model has improved its ability to identify user interests and needs, improved the relevance and accuracy of recommendation results, and can adapt to changes in user interests and needs, and realize personalized recommendations.
Smart Images
Figure CN119477458A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of information recommendation, and more specifically, to a method for fine and coarse ranking based on a deep neural network. Background Art
[0002] Traditional ranking methods are usually based on simple linear models or shallow feature extraction, which cannot capture the complex relationships between user behavior, product characteristics, and environmental information, resulting in less than ideal accuracy and personalization of recommendation results. User behavior changes dynamically at different times. Traditional methods are often unable to effectively adapt to changes in user interests and needs, and are unable to adjust recommendation results in real time. As the amount of data continues to grow, the interactive relationship between users and products has become high-dimensional and diverse. Traditional ranking and recommendation methods find it difficult to effectively integrate and deeply mine multi-source data, resulting in ranking results that are difficult to meet the personalized needs of users. Existing deep neural networks still have the problem of ineffective fusion of coarse and fine ranking models in ranking tasks, making it difficult to capture user preferences in a refined manner. Summary of the Invention
[0003] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a fine sorting and coarse sorting method based on a deep neural network to solve the problems raised in the above-mentioned background technology.
[0004] To achieve the above object, the present invention provides the following technical solutions:
[0005] The fine and coarse sorting method based on deep neural network includes the following steps:
[0006] Obtain user behavior records, product feature data, user profile information, and external context information, build a multi-source data matrix, and map data of different dimensions into the feature space through embedding to form a user-product interaction matrix;
[0007] Performing feature extraction and dimensionality reduction on the user-product interaction matrix through a multi-layer neural network, dynamically stratifying the data by user behavior, product characteristics, and contextual conditions, and generating user preference vectors, product feature vectors, and environmental association vectors;
[0008] In the rough ranking stage, a convolutional neural network is applied to extract local information from the feature space to obtain the product's significance weights at the user behavior level and the feature level. Potential related products are preliminarily screened based on the product's significance weights at the user behavior level and the feature level to form a preliminary ranking list; the feature space includes the user preference vector, the product characteristic vector, and the environment association vector;
[0009] Performing temporal behavior analysis on the products in the preliminary ranking list using a long short-term memory network, dynamically modeling the interaction features between users and products at different time points, outputting a refined ranking list, and assigning a refined ranking score to each product;
[0010] Based on the refined ranking sequence, the inner product between the user preference vector and the product feature vector of each product is calculated to establish a relationship mapping between the features and quantify the degree of fit between each product and the user's preferences;
[0011] Compare the product scores in the refined ranking sequence with the actual user behavior, use the back propagation algorithm in the neural network to adjust the feature weights, complete the adaptive optimization of the model, and form a user behavior feature preference model;
[0012] Comprehensively rank the adaptively adjusted model and merge the results of fine ranking and coarse ranking to ensure that the product ranking results are consistent with the user's historical behavior habits and can adapt to the current context conditions, thus achieving refined dynamic ranking output.
[0013] In a preferred embodiment, user behavior records, product feature data, user profile information, and external context information are obtained to construct a multi-source data matrix. Data of different dimensions are mapped to the feature space through embedding to form a user-product interaction matrix, including:
[0014] User behavior data, product feature data, user profile information, and contextual environment information are processed through data preprocessing to form a multi-source data feature matrix;
[0015] Use the embedding model to perform low-dimensional mapping on multi-source data, convert the multi-source data into feature vectors, and form an embedded representation of the interaction between users and products;
[0016] All feature vectors are merged to generate a high-dimensional feature space and construct the user-product interaction matrix.
[0017] In a preferred embodiment, the user-product interaction matrix is subjected to feature extraction and dimensionality reduction through a multi-layer neural network, and the data is dynamically stratified according to user behavior, product characteristics, and contextual conditions to generate user preference vectors, product characteristic vectors, and environmental association vectors, including:
[0018] Use a multi-layer neural network to extract features from the user-product interaction matrix layer by layer to extract the interaction information between users and products;
[0019] Reduce the dimensionality of high-dimensional features to retain key features of user behavior, product characteristics, and environmental information;
[0020] The extracted features are classified according to user behavior, product attributes, and environmental relevance to generate user preference vectors, product feature vectors, and environmental relevance vectors.
[0021] In a preferred embodiment, in the rough ranking stage, a convolutional neural network is applied to extract local information from the feature space to obtain the product's significance weights at the user behavior level and the feature level. Potential related products are preliminarily screened based on the product's significance weights at the user behavior level and the feature level to form a preliminary ranking list; the feature space includes the user preference vector, the product characteristic vector, and the environment association vector, including:
[0022] Use convolutional neural networks to perform convolution operations on user preference vectors and product feature vectors, extract local features, and generate significant feature maps;
[0023] Based on the convolution results, the importance of the product in terms of user behavior and product features is weighted to generate a significance weight matrix;
[0024] Products are sorted according to the significance weight matrix to screen out potential related products that are highly matched with user interests and form a preliminary sorted list.
[0025] In a preferred embodiment, a long short-term memory network is used to perform temporal behavior analysis on the products in the preliminary ranking list, dynamically modeling the interaction characteristics between users and products at different time points, outputting a refined ranking list, and assigning a refined ranking score to each product, including:
[0026] Input the interaction data between users and products at different time points into the long short-term memory network to build a temporal behavior feature model between users and products;
[0027] Use long short-term memory networks to model time series data, capture the temporal feature changes of user behavior, and generate time-dependent feature representations;
[0028] According to the time series feature representation, a refined ranking score is calculated for each product in the preliminary ranking list, and a refined ranking list is output.
[0029] In a preferred embodiment, based on the refined ranking sequence, the inner product between the user preference vector and the product feature vector of each product is calculated, a relationship mapping between the features is established, and the degree of fit between each product and the user preference is quantified, including:
[0030] Calculate the inner product of the feature vector of each product in the refined ranking sequence and the user preference vector to generate a feature relationship score;
[0031] Using feature relationship scores, we can establish a relationship mapping between products and user preferences, and quantify the degree to which each product satisfies user needs.
[0032] The feature relationship mapping is normalized to generate a fit metric between the product and user preferences.
[0033] In a preferred embodiment, the product scores in the refined ranking sequence are compared with the actual user behavior, and the back propagation algorithm in the neural network is used to adjust the feature weights to complete the adaptive optimization of the model and form a user behavior feature preference model, including:
[0034] Compare the predicted scores of products in the refined ranking list with the actual user behavior data to generate prediction errors;
[0035] Use the back-propagation algorithm to adjust the feature weights of the neural network and optimize the model prediction performance;
[0036] According to the updated weight information, the model's preference expression for user behavior characteristics is dynamically adjusted to improve the user behavior characteristic preference model.
[0037] In a preferred embodiment, the adaptively adjusted model performs comprehensive ranking, integrating the results of fine ranking and coarse ranking to ensure that the product ranking results are consistent with the user's historical behavior habits and can adapt to the current context conditions, thereby achieving refined dynamic ranking output, including:
[0038] Fuse the data of fine sorting sequence and coarse sorting sequence to establish a fusion sorting model;
[0039] Adjust product ranking priorities based on the fusion ranking model to ensure that product ranking results are consistent with user historical preferences and real-time environment;
[0040] Output a comprehensive ranking list to ensure the accuracy of product ranking results and dynamic matching with user needs.
[0041] The technical effects and advantages of the method of the present invention are:
[0042] 1. By acquiring user behavior records, product feature data, user profile information, and external contextual information, a multi-source data matrix is constructed and mapped to the feature space using an embedding method, enabling the integration and modeling of user-product interaction data in a multidimensional feature space. By comprehensively considering the user's multidimensional feature information and utilizing deep neural networks to extract and reduce the dimensions of user preferences and product characteristics, the ranking model's ability to identify user interests and needs is enhanced. Through dynamic layering and local information extraction using convolutional neural networks, the significant correlation between user behavior characteristics and product characteristics can be effectively captured, enabling more accurate preliminary screening and improving the relevance and accuracy of recommendation results.
[0043] 2. Long-Short-Term Memory (LSTM) networks are used to analyze the temporal behavior of user-product interactions, capturing the temporal trends of user interests and adaptively optimizing the ranking model to ensure that ranking results align with users' dynamic needs. The fusion of fine and coarse ranking results, along with the adaptive adjustment of the backpropagation algorithm, continuously optimizes feature weights, enabling in-depth mining and modeling of user behavior characteristics. This improves the accuracy and personalization of information recommendations, effectively adapting to diverse user needs and dynamically adjusting based on real-time user behavior to ensure that product ranking results closely align with users' actual interests. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 This is a schematic diagram of a fine and coarse sorting method based on a deep neural network according to the present invention; DETAILED DESCRIPTION
[0045] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.
[0046] Reference Figure 1 .
[0047] The embodiment provides a method for fine and coarse sorting based on a deep neural network of the present invention, which includes the following steps:
[0048] Obtain user behavior records, product feature data, user profile information, and external context information, build a multi-source data matrix, and map data of different dimensions into the feature space through embedding to form a user-product interaction matrix;
[0049] Performing feature extraction and dimensionality reduction on the user-product interaction matrix through a multi-layer neural network, dynamically stratifying the data by user behavior, product characteristics, and contextual conditions, and generating user preference vectors, product feature vectors, and environmental association vectors;
[0050] In the rough ranking stage, a convolutional neural network is applied to extract local information from the feature space to obtain the product's significance weights at the user behavior level and the feature level. Potential related products are preliminarily screened based on the product's significance weights at the user behavior level and the feature level to form a preliminary ranking list; the feature space includes the user preference vector, the product characteristic vector, and the environment association vector;
[0051] Performing temporal behavior analysis on the products in the preliminary ranking list using a long short-term memory network, dynamically modeling the interaction features between users and products at different time points, outputting a refined ranking list, and assigning a refined ranking score to each product;
[0052] Based on the refined ranking sequence, the inner product between the user preference vector and the product feature vector of each product is calculated to establish a relationship mapping between the features and quantify the degree of fit between each product and the user preference;
[0053] Compare the product scores in the refined ranking sequence with the actual user behavior, use the back propagation algorithm in the neural network to adjust the feature weights, complete the adaptive optimization of the model, and form a user behavior feature preference model;
[0054] Comprehensively rank the adaptively adjusted model and merge the results of fine ranking and coarse ranking to ensure that the product ranking results are consistent with the user's historical behavior habits and can adapt to the current context conditions, thus achieving refined dynamic ranking output.
[0055] Specifically, we acquire user behavior records, product feature data, user profile information, and external context information, construct a multi-source data matrix, and map data of different dimensions into the feature space through embedding to form a user-product interaction matrix, including:
[0056] User behavior data, product feature data, user profile information, and contextual environment information are processed through data preprocessing to form a multi-source data feature matrix. For example, if a user searches for mobile phones on an e-commerce platform, user behavior data includes which mobile phones the user clicks on and how long they browse. Product feature data includes the brand, price, and specifications of the mobile phone. User profile information includes the user's age, gender, and interests. Contextual environment information may include the date of the day, promotional activities, etc.
[0057] The embedding model is used to perform low-dimensional mapping on multi-source data, converting the multi-source data into feature vectors to form an embedded representation of the interaction between users and products. For example, the number of views in user behavior and the price in product features are two data of different natures. Through the embedding model, they can be converted into feature vectors of the same dimension.
[0058] All feature vectors are merged to generate a high-dimensional feature space and construct the user-product interaction matrix.
[0059] Specifically, the user-product interaction matrix is subjected to feature extraction and dimensionality reduction through a multi-layer neural network, and the data is dynamically stratified according to user behavior, product characteristics, and contextual conditions to generate user preference vectors, product feature vectors, and environmental association vectors, including:
[0060] A multi-layer neural network is used to extract features layer by layer from the user-product interaction matrix, refining the interaction information between the user and the product. For example, if a user browses multiple types of mobile phones, the first layer of the network may extract the user's preference for mobile phone brand, the second layer extracts sensitivity to mobile phone price, and the third layer extracts preference for mobile phone performance.
[0061] Reduce the dimensionality of high-dimensional features to retain the key features of user behavior, product characteristics, and environmental information. For example, a 100-dimensional user feature vector is reduced to 20 dimensions representing the key features of user behavior.
[0062] The extracted features are classified according to user behavior, product attributes, and environmental relevance to generate user preference vectors, product characteristic vectors, and environmental relevance vectors. Among them, user browsing behavior on mobile phones may be more influenced by brand and price, and environmental information (such as whether there is a promotion on that day) will affect the user's purchasing decision. The user behavior and environmental information will be classified into different feature vectors.
[0063] Specifically, in the rough ranking stage, a convolutional neural network is applied to extract local information from the feature space to obtain the product's significance weights at the user behavior level and the feature level. Potential related products are preliminarily screened based on the product's significance weights at the user behavior level and the feature level to form a preliminary ranking list; the feature space includes the user preference vector, the product characteristic vector, and the environment association vector, including:
[0064] A convolutional neural network is used to perform a convolution operation on the user preference vector and the product feature vector, extracting local features and generating a saliency feature map. For example, if a user prefers a high-resolution screen and a certain phone excels in screen resolution, the convolutional neural network will extract both the user preference and the product features.
[0065] Based on the convolution results, the importance of the product in terms of user behavior and product features is weighted to generate a significance weight matrix;
[0066] Products are sorted according to the significance weight matrix to screen out potential related products that are highly matched with user interests and form a preliminary sorted list.
[0067] Specifically, the products in the preliminary ranking list are analyzed for temporal behavior using a long short-term memory network, the interaction characteristics between users and products at different time points are dynamically modeled, a refined ranking list is output, and a refined ranking score is assigned to each product, including:
[0068] The interaction data between users and products at different time points is fed into the LSTM network to build a temporal behavioral feature model between the user and the product. For example, if a user browsed a lot of mobile phones the day before and is primarily interested in a certain brand today, the LSTM network can capture this change in the user's focus on mobile phone brands and determine that the brand of mobile phone that the user is primarily interested in today is more important to the user.
[0069] Long Short-Term Memory (LSTM) networks are used to model time series data, capturing temporal changes in user behavior and generating time-dependent feature representations. For example, if a user has recently frequently browsed a high-resolution screen on their phone, the time-dependent feature representations generated by the LSTM network will highlight the user's preference for screen resolution.
[0070] According to the temporal feature representation, a refined ranking score is calculated for each product in the preliminary ranking list, and a refined ranking list is output; for example, if the number of times a user browses a certain mobile phone increases recently, the ranking score of the certain mobile phone will be higher and it will be ranked at the front of the refined ranking list.
[0071] Specifically, based on the refined ranking sequence, the inner product between the user preference vector and the product feature vector of each product is calculated, a relationship mapping between the features is established, and the degree of fit between each product and the user preference is quantified, including:
[0072] Calculate the inner product of the feature vector of each product in the refined ranking sequence and the user preference vector to generate a feature relationship score;
[0073] Using feature relationship scores, we establish a relationship mapping between products and user preferences, quantifying how well each product satisfies user needs. Using internal feature relationship scores, we create a mapping of how well all products align with user preferences. Products with high scores will be closer to the center of user preferences in the mapping.
[0074] The feature relationship mapping is normalized to generate a fit metric between the product and user preferences.
[0075] Specifically, the product scores in the refined ranking sequence are compared with the actual user behavior. The back propagation algorithm in the neural network is used to adjust the feature weights to complete the adaptive optimization of the model and form a user behavior feature preference model, including:
[0076] Compare the predicted scores of products in the refined ranking list with the actual user behavior data to generate prediction errors;
[0077] Use the back-propagation algorithm to adjust the feature weights of the neural network and optimize the model prediction performance;
[0078] According to the updated weight information, the model's preference expression for user behavior characteristics is dynamically adjusted to improve the user behavior characteristic preference model.
[0079] Specifically, the adaptively adjusted model performs comprehensive ranking, integrating the results of fine and coarse ranking to ensure that the product ranking results are consistent with the user's historical behavior habits and can adapt to the current context conditions, achieving refined dynamic ranking output, including:
[0080] The fine ranking sequence and the coarse ranking sequence are fused to establish a fusion ranking model. For example, the coarse ranking stage may screen out 50 potential mobile phones. The fine ranking stage conducts a more in-depth analysis of the 50 mobile phones and displays the top 10 to the user.
[0081] Based on the fusion ranking model, the product ranking priority is adjusted to ensure that the product ranking results are consistent with the user's historical preferences and the real-time environment. For example, if a user usually likes to browse smartwatches and searches for mobile phones during a promotion, the fusion ranking model will increase the ranking weight of mobile phones to ensure that mobile phones appear before smartwatches.
[0082] Output a comprehensive ranking list to ensure the accuracy of product ranking results and dynamic matching with user needs.
[0083] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via a wired (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.
[0084] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0085] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0086] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.
[0087] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.
[0088] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.
[0089] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.
[0090] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.
[0091] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for fine and rough sorting based on deep neural network, characterized in that: The following steps are involved: Obtain user behavior records, product feature data, user portrait information, and external context information, build a multi-source data matrix, map data of different dimensions to the feature space by embedding, and form a user-product interaction matrix; Extracting features and reducing the dimension of the user-product interaction matrix through a multi-layer neural network, dynamically stratifying the data according to user behavior, product features, and contextual conditions, and generating user preference vectors, product feature vectors, and environmental association vectors; In the rough sorting stage, a convolutional neural network is applied to extract local information from the feature space to obtain the significant weights of the products at the user behavior level and the feature level, and potential related products are preliminarily screened according to the significant weights of the products at the user behavior level and the feature level to form a preliminary sorting list; the feature space includes the user preference vector, the product characteristic vector and the environment association vector; Performing temporal behavior analysis on the products in the preliminary ranking list through a long short-term memory network, dynamically modeling the interaction features between users and products at different time points, outputting a refined ranking list, and assigning a refined ranking score to each product; Based on the refined ranking sequence, the inner product between the user preference vector and the product feature vector of each product is calculated, the relationship mapping between the features is established, and the degree of fit between each product and the user preference is quantified; Compare the product scores in the refined ranking sequence with the actual user behavior, use the back propagation algorithm in the neural network to adjust the feature weights, complete the adaptive optimization of the model, and form a user behavior feature preference model; Comprehensively sort the adaptively adjusted model and merge the results of fine and coarse sorting to ensure that the product sorting results are consistent with the user's historical behavior habits and can adapt to the current context conditions, thereby achieving refined dynamic sorting output.
2. The method for fine and rough sorting based on deep neural network according to claim 1, characterized in that: The method of obtaining user behavior records, product feature data, user portrait information and external context information, constructing a multi-source data matrix, and mapping data of different dimensions to the feature space by embedding to form a user-product interaction matrix includes: The user behavior data, the product feature data, the user portrait information and the context environment information are processed through data preprocessing to form a multi-source data feature matrix; Using an embedding model to perform low-dimensional mapping on the multi-source data, converting the multi-source data into feature vectors to form an embedding representation of the interaction between the user and the product; All feature vectors are merged to generate a high-dimensional feature space and construct the user-product interaction matrix.
3. The method for fine and rough sorting based on deep neural network according to claim 2, characterized in that: The method extracts features and reduces the dimension of the user-product interaction matrix through a multi-layer neural network, dynamically stratifies the data according to user behavior, product features and contextual conditions, and generates user preference vectors, product feature vectors and environment association vectors, including: Using a multi-layer neural network to extract features from the user-product interaction matrix layer by layer, and extracting the interaction information between the user and the product; Reduce the dimensionality of high-dimensional features and retain the key features of user behavior, product features and environmental information; The extracted features are classified according to user behavior, product attributes, and environmental relevance to generate the user preference vector, the product characteristic vector, and the environmental relevance vector.
4. The method for fine and rough sorting based on deep neural network according to claim 3, characterized in that: In the rough ranking stage, convolutional neural networks are used to extract local information from the feature space to obtain the significant weights of products at the user behavior level and feature level. Potential related products are preliminarily screened based on the significant weights of products at the user behavior level and feature level to form a preliminary ranking list, including: Use convolutional neural network to perform convolution operation on user preference vector and product feature vector, extract local features and generate significant feature map; Based on the convolution results, the importance of the product in terms of user behavior and product features is weighted to generate a significance weight matrix; The products are sorted according to the significance weight matrix, and potential related products that are highly matched with the user's interests are screened out to form the preliminary sorting list.
5. The method for fine and rough sorting based on deep neural network according to claim 4, characterized in that: The products in the preliminary ranking list are analyzed for temporal behavior through a long short-term memory network, the interaction features between users and products at different time points are dynamically modeled, a refined ranking list is output, and a refined ranking score is assigned to each product, including: Input the interaction data between users and products at different time points into the long short-term memory network to build a temporal behavior feature model between users and products; Use long short-term memory networks to model time series data, capture the temporal feature changes of user behavior, and generate time-dependent feature representations; According to the time series feature representation, a refined ranking score is calculated for each product in the preliminary ranking list, and the refined ranking list is output.
6. The method for fine and rough sorting based on deep neural network according to claim 5, characterized in that: The inner product between the user preference vector and the product feature vector of each product is calculated based on the refined sorting sequence, a relationship mapping between features is established, and the degree of fit between each product and the user preference is quantified, including: Performing inner product calculation on the feature vector of each product in the refined ranking sequence and the user preference vector to generate a feature relationship score; Using the feature relationship scores, a relationship mapping between products and user preferences is established to quantify the degree to which each product satisfies user needs; The feature relationship mapping is normalized to generate a fit metric between the product and user preferences.
7. The method for fine and rough sorting based on deep neural network according to claim 5, characterized in that: Compare the product scores in the refined ranking sequence with the actual user behavior, use the back propagation algorithm in the neural network to adjust the feature weights, complete the adaptive optimization of the model, and form a user behavior feature preference model, including: Compare the predicted scores of products in the refined ranking sequence with the actual behavior data of users to generate prediction errors; Use the back propagation algorithm to adjust the feature weights of the neural network and optimize the model prediction performance; According to the updated weight information, the model's preference expression for user behavior characteristics is dynamically adjusted to improve the user behavior characteristic preference model.
8. The method for fine and rough sorting based on deep neural network according to claim 5, characterized in that: The adaptively adjusted model is used for comprehensive sorting, integrating the results of fine sorting and rough sorting to ensure that the product sorting results are consistent with the user's historical behavior habits and can adapt to the current context conditions, achieving refined dynamic sorting output, including: The refined sorting sequence and the coarse sorting sequence are fused to establish a fusion sorting model; According to the fusion sorting model, adjust the sorting priority of products to ensure that the product sorting results are consistent with the user's historical preferences and real-time environment; Output a comprehensive ranking list to ensure the accuracy of product ranking results and dynamic matching with user needs.
9. A computer device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for fine and coarse sorting based on a deep neural network as described in any one of claims 1 to 8 is implemented.
10. A computer storage medium, characterized in that: It includes computer instructions. When the computer instructions are executed on an electronic device, the electronic device executes the fine sorting and coarse sorting method based on a deep neural network as described in any one of claims 1 to 8.