An intelligent recommendation system and method based on deep learning and big data fusion

Abstract

The present invention discloses an intelligent recommendation system and method based on the fusion of deep learning and big data, which relate to the fields of artificial intelligence and big data, and include a data acquisition module, a data preprocessing module, a deep learning model training module, a real-time recommendation module, a user feedback and iteration module and a recommendation result interpretation module. The system adopts a Transformer-based model structure, combined with an attention mechanism, to improve the model's ability to process complex data; utilizes distributed storage and parallel computing technology to achieve rapid processing and analysis of large-scale data; dynamically adjusts recommendation strategies based on real-time user behavior data to improve the timeliness and accuracy of recommendations; solves the problems of high training cost, large model complexity and poor interpretability of existing models when processing big data, and the present invention can efficiently process large-scale, high-dimensional data, improve recommendation efficiency, accurately capture users' complex and changeable interests and preferences, and improve recommendation accuracy.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and big data technology, and in particular to an intelligent recommendation system and method based on the fusion of deep learning and big data. Background Art

[0002] With the rapid development of information technology and the popularization of the Internet and mobile devices, the amount of data generated by users has grown exponentially. These data not only contain basic information about users, but also cover multi-dimensional information such as users' browsing history, purchase history, social media interactions, and geographic location. These data provide rich resources for intelligent recommendation systems, but also bring unprecedented challenges.

[0003] There are many intelligent recommendation systems on the market, most of which are based on traditional machine learning algorithms or simple statistical models. These systems have improved the accuracy of recommendations to a certain extent, but they still have the following problems: it is difficult to capture users' complex and changing interests and preferences; it is difficult to process large-scale, high-dimensional data, resulting in low recommendation efficiency; it is difficult to adapt to the dynamically changing market environment and user behavior. Some existing technologies use deep learning algorithms for recommendations, such as recommendation models based on recurrent neural networks (RNN) or convolutional neural networks (CNN). However, these models still have problems such as high training costs, high model complexity, and poor interpretability when processing big data.

[0004] Therefore, an intelligent recommendation system and method based on the fusion of deep learning and big data are needed to solve the above problems. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention discloses an intelligent recommendation system and method based on the fusion of deep learning and big data, which utilizes deep learning algorithms combined with big data processing technology to realize an efficient and accurate intelligent recommendation system and method. The system can be widely used in e-commerce, social media, content distribution and other fields to improve user experience and platform operation efficiency.

[0006] The present invention adopts the following technical solution: an intelligent recommendation system based on deep learning and big data fusion, comprising:

[0007] A data collection module, which is used to collect user behavior data, product information, and user portrait data from multiple data sources;

[0008] A data preprocessing module, which is used to clean, convert and normalize the collected data, and to construct user portraits and item portraits by integrating data from multiple data sources;

[0009] A deep learning model training module, wherein the deep learning model training module uses distributed storage and parallel computing to achieve rapid data processing and analysis, and the deep learning model training module includes a model selection unit, a model training unit, and a model performance evaluation unit. The model selection unit adaptively selects a deep learning model for training according to the application scenario and data type. The model selection unit uses an improved Transformer deep learning model to model and analyze image and text data, and uses an improved recurrent neural network sequence model to model and analyze sequence data to predict the user's future interest preferences. The improved Transformer deep learning model automatically learns and mines the user's behavior patterns, interest preferences, and potential needs in image and text data by combining an attention mechanism and an embedding layer. The improved recurrent neural network sequence model uses an attention mechanism to dynamically adjust the weights of different items in the user behavior sequence prediction results;

[0010] A real-time recommendation module, which generates a user-personalized product recommendation list based on the training model and dynamically adjusts the recommendation results through a real-time recommendation engine. The real-time recommendation engine triggers the recalculation of the training model and the update of the recommendation results according to the changes in the user's real-time behavior data monitored;

[0011] A user feedback and iteration module, which uses a user feedback mechanism to collect user feedback on recommendation results, and dynamically adjusts recommendation strategies and model parameters based on user feedback and real-time behavior data to achieve continuous optimization and iterative update of the training model;

[0012] The recommendation result explanation module displays the generation process and change trend of the recommendation results through a visualization platform, and uses an explanatory algorithm to perform explanatory analysis on the recommendation results.

[0013] Furthermore, the data collection module uses an application programming interface (API) to collect user browsing records, purchase behaviors, product details and user information in user databases, e-commerce platforms and log files in real time. The data preprocessing module cleans, converts and normalizes the original collected data through a data rule engine, and uses a self-supervised representation learning method to integrate user behavior characteristics, interest characteristics and item attribute characteristics to construct user portraits and item portraits.

[0014] Furthermore, the improved Transformer deep learning model includes an input embedding layer, a position encoding layer, a self-attention mechanism layer, a feedforward neural network layer, an adaptive parameter adjustment layer, an adaptive topology structure layer, a maximum pooling layer, a normalization layer, a residual connection layer and an output layer. The working method of the improved Transformer deep learning model includes:

[0015] Step 1: The input embedding layer embeds the preprocessed image or text data into a low-dimensional vector space using word vector processing or image convolution operation, and inputs the data into the improved Transformer deep learning model; the position encoding layer generates the position encoding of the input data using a linear combination of sine and cosine functions, and embeds the encoded position information into the existing low-dimensional vector space;

[0016] Step 2: Use multiple self-attention mechanism layers to construct a multi-layer self-attention structure, wherein the multi-layer self-attention structure divides the low-dimensional vector space data into multiple head sequences and inputs them into the self-attention mechanism layer, and processes the sequences input by the multiple self-attention mechanism layers in parallel, wherein the self-attention mechanism layer captures semantic information at different levels in the image or text data by calculating the attention weight of each element in the input sequence, and performs weighted average of the attention weights of the multiple head sequences to obtain the final attention representation;

[0017] Step 3, the feedforward neural network layer maps the attention representation to a high-dimensional feature representation through multiple fully connected layers, and uses an activation function to perform a combination of nonlinear transformations between the fully connected layers to abstract and extract feature representations layer by layer;

[0018] Step 4: Adopting the adaptive parameter adjustment layer to adaptively adjust and update the deep learning network parameters, and adopting the adaptive topology structure layer to dynamically adjust the deep learning network structure, the adaptive parameter adjustment layer adopts a dynamic parameter learning strategy to monitor the error change of the deep learning network, and adaptively adjusts the parameters according to the error change, and the adaptive topology structure layer adaptively adds or deletes the middle layer between the multi-layer self-attention structure and the feedforward neural network layer according to the characteristics of the input data;

[0019] Step 5: Use the maximum pooling layer to sample the extracted feature representation, select the maximum value in each pooling window as the sampling result, and input the extracted feature representation into the normalization layer for standardization processing;

[0020] Step 6, the residual connection layer adds the data directly input into the residual connection layer and the output data of the previous layer by constructing a cross-layer residual connection with the feedforward neural network layer to retain historical information;

[0021] Step 7: Input the processed data into the output layer and output the results according to different task requirements.

[0022] Furthermore, the improved recurrent neural network sequence model adopts recurrent neural network RNN ​​or long short-term memory network LSTM sequence modeling to capture the evolution and change of user behavior over time. The user behavior sequence is , Represents the input of user behavior data at the tth time step, and the modeling formula is: (1)

[0023] In formula (1), represents the user interest preference prediction result at the tth time step, , and Represents the input gate, forget gate, and output gate functions of the t-th time step, which are used to control the inflow and outflow of data information. represents the input of user behavior data at the t-1th time step, Represents the hidden state of user behavior data at time step t-1; dynamically adjusts the weights of different items during recommendation according to the input information at different time steps to reflect the user's personalized preferences. The calculation formula is: (2)

[0024] In formula (2), represents the attention weight at the t-th time step, , , and Represents the parameters of the self-attention mechanism, which is used for the linear transformation of user behavior data and hidden states. Represents the hidden state of the user behavior data at time step t.

[0025] Furthermore, the model training unit adopts a small batch gradient descent optimization algorithm to train the selection model, and adopts a regularization method to prevent model overfitting. The small batch gradient descent optimization algorithm updates the parameters of the selection model by randomly sampling batch data, and minimizes the loss function by iteratively updating the parameters of the selection model. The regularization method controls the number of selection model parameters by adding penalty terms. The model performance evaluation unit adopts a cross-validation method to evaluate the performance of the training model. The cross-validation method evaluates the performance of the training model by dividing the data set into subsets and performing repeated optimization training of the training model.

[0026] Furthermore, the real-time recommendation engine adopts a real-time stream processing framework to perform real-time updates and calculations on changes in user real-time behavior data. The real-time stream processing framework adopts a data caching mechanism to cache user behavior feature vectors in memory for processing, and adopts a multi-core processor for data parallel processing. The real-time stream processing framework adds new data points to the data set through a time window, updates the new data points, and obtains the membership vectors of the new data points in real time. The time window triggers recalculation and updating of recommendation results by inputting the new data points into the training model.

[0027] Furthermore, the visualization platform uses interactive charts and heat maps to visualize and interactively analyze the recommendation results. The visualization platform uses a user authentication mechanism to verify the identity of the accessing user, and records the number of times the cached data items are viewed through an access counter. The explanatory algorithm identifies key data features by calculating the degree of influence of each data feature on the recommendation results, and draws a curve chart to show the influence trend of changes in the characteristic value of a single variable on the target variable.

[0028] Furthermore, an intelligent recommendation method based on deep learning and big data fusion is provided, the intelligent recommendation method comprising the following steps:

[0029] Step 1: Collect user behavior data, product information, and user portrait data from multiple data sources;

[0030] Step 2: Clean, convert and normalize the collected data, and build user and item portraits by integrating data from multiple data sources;

[0031] Step 3: Select an improved deep learning algorithm, combine it with big data processing technology, train and optimize the model, and use distributed storage and parallel computing to achieve rapid data processing and analysis;

[0032] Step 4: Use the trained deep learning model to generate a personalized product recommendation list for users, and dynamically adjust the recommendation results through a real-time recommendation engine;

[0033] Step 5: Use the user feedback mechanism to collect user feedback on the recommendation results, and dynamically adjust the recommendation strategy and model parameters based on user feedback and real-time behavior data to achieve continuous optimization and iteration of the training model;

[0034] Step 6: Use a visualization platform to display the generation process and influencing factors of the recommendation results, and use an explanatory algorithm to perform an explanatory analysis of the recommendation results.

[0035] The beneficial effects of the present invention are:

[0036] 1. The present invention adopts a Transformer-based deep learning model, combined with attention mechanism and embedding layer technology, to model and analyze large-scale, high-dimensional user behavior data and item information, and uses distributed storage and parallel computing technology to realize rapid processing and analysis of big data, improve the real-time and efficiency of the recommendation system, deeply integrate deep learning algorithms with big data processing technology, and automatically learn complex features and patterns in big data through deep learning models to achieve more accurate and personalized recommendations. This deep integration not only improves the accuracy of the recommendation system, but also enhances the scalability and adaptability of the system.

[0037] 2. The present invention adopts a recurrent neural network (RNN) or a long short-term memory network (LSTM) sequence model to model and analyze user behavior sequences, predict users' future interest preferences, and combines the attention mechanism to dynamically adjust the weights of different items in the recommendation results, so that the recommendation results are more in line with users' current interests and needs. The personalized recommendation algorithm is optimized and innovated, and by introducing mechanisms such as user behavior sequence analysis and interest migration prediction, changes in user interests and potential needs can be more accurately captured.

[0038] 3. The present invention designs a real-time recommendation engine, which triggers the recalculation of the recommendation model and the update of the recommendation results by monitoring the real-time behavior data of users (, and establishes a user feedback mechanism to collect user feedback on the recommendation results for continuous optimization and iteration of the model, thus realizing the real-time recommendation function and introducing a dynamic adjustment mechanism, which can update the recommendation results in real time according to the real-time behavior and feedback of users, thereby improving the timeliness and accuracy of the recommendation.

[0039] 4. The present invention designs a recommendation result explanation module, which uses visualization technology to display the generation process and influencing factors of the recommendation results, and introduces an explanatory algorithm to perform an explanatory analysis of the recommendation results to help users understand the rationality and accuracy of the recommendation results. The present invention focuses on improving the interpretability and transparency of the recommendation model, and uses visualization technology and explanatory algorithms to help users understand the generation process and basis of the recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the overall system architecture of the present invention;

[0041] Figure 2 This is an architecture diagram of the deep learning model training module in the present invention;

[0042] Figure 3 Schematic diagram of the workflow of the improved Transformer deep learning model in the present invention;

[0043] Figure 4 Schematic diagram of the workflow of the overall method of the present invention;

[0044] Figure 5 Schematic diagram of the architecture of the improved Transformer deep learning model of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the attached embodiment of the present invention Figure 1 To Attachment Figure 5, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] The embodiment of the present invention discloses an intelligent recommendation system based on deep learning and big data fusion, including:

[0047] Data collection module, which is responsible for collecting data from multiple data sources (such as user behavior data, product information, and user portrait data). Collect user browsing, searching, purchasing and other behavior data through websites, mobile applications or IoT devices. Obtain detailed information about products from product databases, e-commerce platform APIs, etc., including descriptions, prices, attributes, etc. Integrate user registration information, social network data, historical transaction records, etc. to form a comprehensive profile of users.

[0048] Data preprocessing module: The data preprocessing module cleans, converts and normalizes the collected data to build user and item portraits. Through data cleaning operations, duplicate data is removed, missing values ​​are processed, and inconsistent data formats are resolved. Through feature conversion and normalization operations, the features of different data sources are uniformly encoded and normalized to facilitate subsequent model training. Through cluster analysis, feature extraction and other technical methods, a comprehensive description of users and items is established to build user and item portraits.

[0049] The deep learning model training module uses an improved deep learning algorithm (such as a Transformer-based model) combined with big data processing technology to train and optimize deep learning models, and uses distributed storage and parallel computing to process and analyze data. The deep learning model training module includes a model selection unit, a model training unit, and a model performance evaluation unit. The model selection unit selects a suitable deep learning model based on the application scenario and data type, such as an improved Transformer model for processing image and text data, and an improved recurrent neural network sequence model for processing user behavior sequence data. The model training unit uses a distributed computing platform to train the model and optimize the model parameters to predict user interest preferences. The model performance evaluation unit uses methods such as cross-validation or A / B testing to evaluate the prediction performance and accuracy of the model.

[0050] Real-time recommendation module, the real-time recommendation module generates a personalized product recommendation list based on the trained model, and dynamically adjusts the recommendation results through the real-time recommendation engine. The real-time recommendation engine triggers the recalculation of the model and the update of the recommendation results by monitoring the changes in the user's real-time behavior data, such as clicks, purchases, etc.

[0051] User feedback and iteration module: The user feedback and iteration module uses the user feedback mechanism to collect user feedback on the recommendation results for continuous optimization and iteration of the model, and dynamically adjusts the recommendation strategy and model parameters based on feedback and real-time behavior data. The feedback mechanism collects user satisfaction and preference information through questionnaires, user ratings, comments, etc. Dynamically adjust the recommendation strategy based on user feedback and real-time data, and update the recommendation algorithm to optimize the user experience and recommendation effect.

[0052] The recommendation result explanation module uses a visualization platform to display the generation process and change trend of the recommendation results, and uses explanatory algorithms to perform explanatory analysis on the recommendation results. The visualization platform uses an intuitive user interface to display the workflow, data flow, and changes in the recommendation results of the recommendation system. Explanatory algorithms (such as LIME, SHAP, etc.) explain the main driving factors and influencing factors behind each recommendation result.

[0053] This intelligent recommendation system integrates multiple key modules such as data collection, preprocessing, deep learning model training, real-time recommendation and user feedback. It improves the accuracy and personalization of the recommendation algorithm through big data and deep learning technology, while ensuring the real-time nature of the system and the effectiveness of the feedback mechanism.

[0054] The data collection module collects user browsing history, purchase behavior, product details, user information and other data from user databases, e-commerce platforms and log file systems in real time through integrated APIs. The API should support real-time data streaming or periodic data pulling to ensure the timeliness and integrity of the data. Or use HTTP client libraries (such as Python's requests or aiohttp) to write data crawling scripts and obtain data through API interfaces. Considering the amount of data and real-time requirements, it may be necessary to deploy a distributed crawling system or use a message queue (such as Kafka) to manage data flows. Store the captured data in an efficient data storage system, such as a NoSQL database (MongoDB, Cassandra) or a distributed file system (HDFS) for subsequent processing.

[0055] The data preprocessing module cleans, transforms and normalizes the raw data, and applies self-supervised representation learning methods to integrate user behavior, interests and item attribute characteristics. The implementation steps include the following:

[0056] Data cleaning: Remove duplicate, incomplete or abnormal data records. Correct or fill missing values, such as using the mean, median or machine learning-based methods to predict missing values.

[0057] Data conversion: Convert text data (such as user comments and product descriptions) into numerical features, which can be vectorized using models such as TF-IDF, Word2Vec, or BERT. Convert timestamp data into meaningful date and time formats, and extract features such as day of the week and time period.

[0058] Normalization: Normalize or standardize numerical features to eliminate the effects of different dimensions and distributions.

[0059] Self-supervised representation learning: Utilize user browsing, clicking, purchasing and other behavior sequences to learn user behavior patterns through models such as recurrent neural networks (RNN), long short-term memory networks (LSTM) or Transformer. Combine the user's historical purchase records and browsing preferences, and use clustering algorithms (such as K-means) or embedding models (such as matrix decomposition, Word2Vec variants) to identify the user's interest groups or topics. Based on the product details (such as price, brand, category, description, etc.), use similarity algorithms (such as cosine similarity) or deep learning methods (such as CNN, RNN processing text descriptions) to extract the key attribute features of the item. Fusion of user behavior features, interest features and item attribute features to construct user portraits and item portraits. This can be achieved through feature concatenation, weighted summation or more complex neural network models.

[0060] Portrait construction and application: Based on the results of preprocessing and feature integration, detailed user portraits and item portraits are constructed and applied to scenarios such as personalized recommendations, user analysis, and market strategy formulation.

[0061] The improved Transformer deep learning model includes an input embedding layer, a position encoding layer, a self-attention mechanism layer, a feedforward neural network layer, an adaptive parameter adjustment layer, an adaptive topology structure layer, a maximum pooling layer, a normalization layer, a residual connection layer and an output layer. The improved Transformer deep learning model specifically includes the following steps:

[0062] Input data and embed it into a low-dimensional vector space: For image data, use a pre-trained convolutional neural network (such as ResNet) to extract feature maps and flatten them into a series of vectors. For text data (mainly used for decoder input in generation tasks), use word embedding technology (such as Word2Vec or GloVe) to convert vocabulary into vectors. Merge the image vector and text vector (if applicable) and add position encoding to preserve position information.

[0063] Obtain feature representation: Use multiple self-attention mechanism layers, each layer divides the input data into multiple heads, and processes these heads in parallel to capture semantic information at different levels. In each self-attention layer, the attention weight of each element to other elements is calculated to capture important areas in the image or key words in the text. The attention weights of multiple head sequences are weighted averaged to obtain the final attention representation.

[0064] Obtain higher-level feature representations: Use multiple fully connected layers to further process the attention representation, and introduce nonlinear transformations through activation functions such as ReLU or GELU. Abstract and extract higher-level feature representations layer by layer to provide rich contextual information for the decoder to generate text.

[0065] Adaptive parameter and topology adjustment: During the training process, the adaptive parameter adjustment layer monitors the error changes of the network and adjusts the network parameters as needed. The adaptive topology layer dynamically adjusts the number of intermediate layers between the multi-layer self-attention structure and the feedforward neural network layer according to the characteristics of the input data and the training effect to optimize the model performance.

[0066] Pooling sampling and standardization: The maximum pooling layer samples the feature representation and selects the maximum value in each pooling window as the sampling result. The normalization layer standardizes the sampled feature representation to speed up the training process and improve the model stability.

[0067] Constructing residual connections: Constructing residual connections between feedforward neural network layers to add input data directly to the output of subsequent layers to alleviate the gradient vanishing problem in deep networks and retain more historical information.

[0068] The result output, the processed data is input into the output layer, usually a linear layer plus a softmax function, to generate an output vector of the vocabulary size.

[0069] The hardware working environment of the improved Transformer deep learning model includes

[0070] GPU: Since the Transformer model has high requirements for computing resources, especially the parallel computing capability of GPU, high-performance GPU is usually required for training and reasoning. For example, Nvidia's RTX series (such as RTX 2080Ti, RTX 3090, etc.) or Tesla series (such as Tesla V100, A100, etc.) GPUs are good choices. These GPUs have powerful parallel computing capabilities and high video memory, and can support the training of large-scale models.

[0071] Video memory: For large Transformer models, especially the improved version, a large video memory is usually required to store model parameters and intermediate calculation results. Therefore, it is crucial to choose a GPU with sufficient video memory.

[0072] CPU: Although GPU plays a major role in the training of deep learning models, CPU also plays an important role, especially in data preprocessing, model deployment and inference. Therefore, it is also necessary to choose a powerful CPU.

[0073] Hard disk: In order to store a large amount of training data and model parameters, a large-capacity hard disk is required. SSD (solid-state drive) is the first choice because of its fast read and write speed, but HDD (mechanical hard disk) can also be configured to expand storage capacity as needed.

[0074] Memory: Sufficient memory can ensure that data is not swapped to the hard disk due to insufficient memory during training, thus affecting the training speed.

[0075] Network equipment: The performance of network equipment (such as high-speed Ethernet switches, routers, etc.) also has an important impact on training speed. High-speed, low-latency network connections can ensure more efficient data synchronization and parameter updates between various computing nodes.

[0076] Cooling system: High-performance GPUs and CPUs generate a lot of heat during operation, so an effective cooling system is required to keep the hardware running stably.

[0077] Power supply: A stable power supply is essential to ensure the continuity and stability of the training process.

[0078] Cluster configuration (optional): For scenarios that require large-scale training, you can build a distributed cluster system with multiple machines and multiple cards. Each node is configured with a high-performance GPU and CPU and connected through a high-speed network. This can use the parallel computing power of multiple machines to accelerate the training process.

[0079] In summary, the hardware working environment of the improved Transformer deep learning model needs to include high-performance GPUs, CPUs, sufficient storage devices and network devices, as well as an effective cooling system and a stable power supply. In scenarios where large-scale training is required, you can also consider building a distributed cluster system to further improve training efficiency.

[0080] The laboratory configuration uses a Core i9 64+128G memory computer, and the simulation environment is established using simulation software. The on-site experimental environment is set up, the simulation data accuracy is 95%, and the algorithm operation error does not exceed 2.5%. The improved Transformer deep learning model (Group A) and the traditional Transformer deep learning model (Group B) are used for comparative experiments. The image information data of three users with different preferences are manually input. The users are entertainment enthusiasts, business people, and literary youths. Simulation work is performed in the experimental environment, and the training and reasoning time of the model, as well as the matching degree of the recommendation results are recorded. The experiment is repeated 5 times to calculate the average value, and the record is shown in Table 1.

[0081] Table 1 Results Statistics

[0082]

[0083] According to the experimental results, the improved Transformer deep learning model performs better in terms of training and reasoning time and the matching degree of recommendation results, indicating that this improved algorithm has better results in processing image information recommendation problems. It can be seen that the improved algorithm has higher practicality and reliability and can be applied to the actual scenarios of recommendation systems.

[0084] The laboratory configuration uses a Core i9 64+128G memory computer, and the simulation environment is established using simulation software. The on-site experimental environment is set up, the simulation data accuracy is 95%, and the algorithm operation error does not exceed 2.5%. The improved Transformer deep learning model (Group A) and the traditional Transformer deep learning model (Group B) are used for comparative experiments. The text information data of three users with different preferences are manually input. The users are entertainment enthusiasts, business people, and literary youth. Simulation work is performed in the experimental environment, and the training and reasoning time of the model, as well as the matching degree of the recommendation results are recorded. The experiment is repeated 5 times to calculate the average value, and the record is shown in Table 2.

[0085] Table 2 Results Statistics

[0086]

[0087] According to the experimental results, the improved Transformer deep learning model has better reasoning efficiency than the traditional Transformer model in the case of entertainment and business users, and the matching degree of recommendation results for all three types of users is also better than the traditional Transformer model. Therefore, the improved Transformer deep learning model also shows better results in processing text information recommendation problems and can be used in practical applications of recommendation systems.

[0088] The improved recurrent neural network sequence model uses recurrent neural network RNN ​​or long short-term memory network LSTM sequence modeling to capture the evolution and changes of user behavior over time. The user behavior sequence is , Represents the input of user behavior data at the tth time step, and the modeling formula is: (1)

[0089] In formula (1), represents the user interest preference prediction result at the tth time step, , and Represents the input gate, forget gate, and output gate functions of the t-th time step, which are used to control the inflow and outflow of data information. represents the input of user behavior data at the t-1th time step, Represents the hidden state of user behavior data at time step t-1; dynamically adjusts the weights of different items during recommendation according to the input information at different time steps to reflect the user's personalized preferences. The calculation formula is: (2)

[0090] In formula (2), represents the attention weight at the t-th time step, , , and Represents the parameters of the self-attention mechanism, which is used for the linear transformation of user behavior data and hidden states. Represents the hidden state of the user behavior data at time step t.

[0091] The hardware working environment of the improved recurrent neural network (RNN) sequence model mainly includes the following aspects:

[0092] GPU, improved RNN models, such as LSTM (Long Short-Term Memory Network) or GRU (Gated Recurrent Unit), have higher computational complexity when processing long sequence data. Therefore, using high-performance GPUs can significantly accelerate the training and reasoning process of the model. The parallel computing capability of GPUs is particularly important for processing a large number of matrix operations and loop operations in RNN.

[0093] CPU: Although GPUs dominate the training of RNN models, CPUs still play an important role in auxiliary tasks such as data preprocessing, model loading, and reasoning. Therefore, a powerful CPU is also necessary.

[0094] Hard disk, in order to store training data, model parameters and intermediate results, a large-capacity hard disk is required. SSD (solid-state drive) is popular because of its fast read and write speed, but HDD (mechanical hard disk) can also be used as an option for expanded storage.

[0095] Memory, sufficient memory is essential to ensure data fluency during training and reduce disk I / O operations. When training large RNN models, a large amount of memory may be required to store information such as the model's hidden state and gradients.

[0096] Network equipment, in scenarios where distributed training or remote access to data is required, the performance of network equipment (such as high-speed Ethernet switches, routers, etc.) has an important impact on the training speed. High-speed, low-latency network connections can ensure efficient transmission of data between multiple computing nodes.

[0097] Cooling system, high-performance GPU and CPU will generate a lot of heat during operation, so an effective cooling system is needed to keep the hardware running stably. Appropriate cooling measures can extend the service life of the hardware and improve the reliability of the system.

[0098] Power supply, a stable power supply is essential to ensure the continuity and stability of the training process. When training large RNN models, it is necessary to ensure that the power system can continue to provide sufficient power support.

[0099] Dedicated hardware (optional): With the continuous development of neural network hardware, some hardware accelerators designed specifically for neural networks have emerged, such as neural network processors (NNPs) and neural network accelerators (NNAs). These hardware accelerators can further improve the training and reasoning speed of RNN models, but usually require additional costs and integration work.

[0100] The laboratory configuration uses a Core i9 64+128G memory computer, and the simulation environment is established using simulation software. The on-site experimental environment is set up, the simulation data accuracy is 95%, and the algorithm operation error does not exceed 2.5%. The improved recurrent neural network sequence model (Group A) and the traditional recurrent neural network sequence model (Group B) are used for comparative experiments. The behavioral sequence data of three users with different preferences are manually input. The users are entertainment enthusiasts, business people, and literary youth. Simulation work is carried out in the experimental environment, and the training and reasoning time of the model, as well as the matching degree of the recommendation results are recorded. The experiment is repeated 5 times to calculate the average value, and the record is shown in Table 3.

[0101] Table 3 Results Statistics

[0102]

[0103] According to the experimental results, the improved recurrent neural network sequence model and the traditional recurrent neural network sequence model have little difference in training and reasoning time, but the improved model is better than the traditional model in terms of the matching degree of recommendation results. In the case of entertainment and business users, the matching degree of recommendation results reached 95.8% and 94.3% respectively, and it also showed better results in the case of literary and artistic users. Therefore, the improved recurrent neural network sequence model also shows better results in processing the problem of recommending behavioral sequence information, and can be used in the practical application of recommendation systems.

[0104] The model selection unit sends the model type and initialization parameters to the model training unit, and the model training unit performs model training based on this information. The model training unit sends the trained model to the model performance evaluation unit, and the evaluation unit uses the validation set or test set to evaluate the model performance. The evaluation results of the model performance evaluation unit are fed back to the model selection unit to provide a reference for the next model selection. The working methods of the model selection unit, model training unit, and model performance evaluation unit are as follows:

[0105] 1. Dataset division: The original dataset is divided into three parts: training set, validation set and test set. The training set is used to select model parameters, the validation set is used to adjust the model's hyperparameters and evaluate the model's performance, and the test set is used to evaluate the model's generalization ability.

[0106] 2. Model selection: Select an appropriate model structure according to task requirements and initialize model parameters.

[0107] 3. Select optimization algorithm: Use mini-batch gradient descent optimization algorithm to train the model. Each time, randomly extract a small batch of data from the training set, calculate the corresponding loss function and gradient, and update the model parameters. By iteratively updating the model parameters, the loss function is minimized.

[0108] 4. Regularization: To avoid overfitting of the model, regularization methods are used to penalize the model parameters. Common regularization methods include L1 regularization and L2 regularization. Regularization terms are added to the loss function to control the number of model parameters.

[0109] 5. Hyperparameter adjustment: Adjust the model's hyperparameters, such as learning rate, regularization strength, etc. Select the best hyperparameter combination by evaluating the model performance on the validation set.

[0110] 6. Cross-validation: Use cross-validation to evaluate the model. Divide the data set into K subsets, select one of the subsets as the validation set each time, and the remaining K-1 subsets as the training set. Perform K training and validation times to obtain the average performance of the model.

[0111] 7. Model evaluation: Finally, the test set is used to evaluate the generalization ability of the model, calculate indicators such as prediction accuracy, recall rate, F1 score, and select the best model as the final prediction model.

[0112] The real-time recommendation engine works as follows:

[0113] 1. Real-time stream processing framework selection: Select a suitable real-time stream processing framework based on business requirements and data scale. Common frameworks include Apache Storm, Apache Flink, Apache Kafka Streams, etc.

[0114] 2. Data caching: Cache the user behavior feature vector in memory and wait for processing. The caching mechanism improves data processing efficiency, reduces the number of database accesses, and avoids data loss. It can be optimized based on factors such as memory size and cache elimination algorithm.

[0115] 3. Parallel processing: Use multi-core processors to process data in parallel to improve computing efficiency. Use parallel computing frameworks, such as Apache Spark, to evenly distribute data to multiple computing nodes, and process data in parallel through different computing nodes to further improve data processing efficiency.

[0116] 4. Update and calculate the data set: Add new data points to the data set, and update and calculate the data set within a certain time window. The size of the time window can be set according to business needs and data size, usually a few seconds or minutes. When updating the data set, calculate its membership vector based on the user's behavior feature vector and store it in memory.

[0117] 5. Recalculation and update of recommendation results: When new data points are input into the training model, the model is recalculated and the recommendation results are updated. The model can be updated using batch processing or online learning to ensure the real-time and accuracy of the model. The recalculated recommendation results are finally provided to users for recommendation display.

[0118] 6. Monitoring and tuning: The real-time recommendation engine needs to monitor the operating status and performance indicators of each component in real time, such as latency, throughput, bottlenecks, etc., to promptly detect problems and adjust system parameters or expand the system architecture. At the same time, attention should also be paid to data security and privacy protection to avoid user data leakage and security risks.

[0119] The visualization platform displays interactive charts and heat maps based on the recommendation results, allowing users to clearly understand the characteristics, changing trends and interrelationships of the recommendation results, and provides interactive analysis functions such as zooming, scrolling, and mouse hovering, allowing users to flexibly adjust the chart range, highlight key information, and gain deeper business insights. The heat map shows the heat distribution of different data items through color changes. The darker the color, the higher the heat. In order to ensure the security and privacy of user data, the visualization platform uses a user authentication mechanism to verify the identity of the accessing user. Only authorized users can enter the system and view the recommendation results. By recording the access counter, you can also understand the number of times the cached data item has been viewed, so as to determine whether the recommendation result has reference value.

[0120] In order to help users better understand the basis and reasons for the recommendation results, the visualization platform also uses an explanatory algorithm to calculate the impact of each data feature on the recommendation results, identify key data features, and draw a curve chart to show the impact trend of the change in the feature value of a single variable on the target variable. In this way, users can better understand the logic behind the recommendation results and conduct deeper business analysis and prediction based on this.

[0121] An intelligent recommendation method based on deep learning and big data fusion, the intelligent recommendation method comprising the following steps:

[0122] Step 1: Collect user behavior data, product information, and user portrait data from multiple data sources;

[0123] Step 2: Clean, convert and normalize the collected data, and build user and item portraits by integrating data from multiple data sources;

[0124] Step 3: Select an improved deep learning algorithm, combine it with big data processing technology, train and optimize the model, and use distributed storage and parallel computing to achieve rapid data processing and analysis;

[0125] Step 4: Use the trained deep learning model to generate a personalized product recommendation list for users, and dynamically adjust the recommendation results through a real-time recommendation engine;

[0126] Step 5: Use the user feedback mechanism to collect user feedback on the recommendation results, and dynamically adjust the recommendation strategy and model parameters based on user feedback and real-time behavior data to achieve continuous optimization and iteration of the training model;

[0127] Step 6: Use a visualization platform to display the generation process and influencing factors of the recommendation results, and use an explanatory algorithm to perform an explanatory analysis of the recommendation results.

[0128] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. An intelligent recommendation system based on deep learning and big data fusion, characterized in that: include: A data collection module, which is used to collect user behavior data, product information, and user portrait data from multiple data sources; A data preprocessing module, which is used to clean, convert and normalize the collected data, and to construct user portraits and item portraits by integrating data from multiple data sources; A deep learning model training module, wherein the deep learning model training module uses distributed storage and parallel computing to achieve rapid data processing and analysis. The deep learning model training module includes a model selection unit, a model training unit, and a model performance evaluation unit. The model selection unit adaptively selects a deep learning model for training according to the application scenario and data type. The model selection unit uses an improved Transformer deep learning model to model and analyze image and text data, and uses an improved recurrent neural network sequence model to model and analyze sequence data to predict the user's future interest preferences. The improved Transformer deep learning model automatically learns and mines users’ behavior patterns, interest preferences, and potential needs in image and text data by combining attention mechanisms and embedding layers. The improved Transformer deep learning model includes an input embedding layer, a position encoding layer, a self-attention mechanism layer, a feedforward neural network layer, an adaptive parameter adjustment layer, an adaptive topology structure layer, a maximum pooling layer, a normalization layer, a residual connection layer and an output layer. The working method of the improved Transformer deep learning model includes: Step 1: The input embedding layer embeds the preprocessed image or text data into a low-dimensional vector space using word vector processing or image convolution operation, and inputs the data into the improved Transformer deep learning model; the position encoding layer generates the position encoding of the input data using a linear combination of sine and cosine functions, and embeds the encoded position information into the existing low-dimensional vector space; Step 2: Use multiple self-attention mechanism layers to construct a multi-layer self-attention structure, wherein the multi-layer self-attention structure divides the low-dimensional vector space data into multiple head sequences and inputs them into the self-attention mechanism layer, and processes the sequences input by the multiple self-attention mechanism layers in parallel, wherein the self-attention mechanism layer captures semantic information at different levels in the image or text data by calculating the attention weight of each element in the input sequence, and performs weighted average of the attention weights of the multiple head sequences to obtain the final attention representation; Step 3, the feedforward neural network layer maps the attention representation to a high-dimensional feature representation through multiple fully connected layers, and uses an activation function to perform a combination of nonlinear transformations between the fully connected layers to abstract and extract feature representations layer by layer; Step 4: Adopting the adaptive parameter adjustment layer to adaptively adjust and update the deep learning network parameters, and adopting the adaptive topology structure layer to dynamically adjust the deep learning network structure, the adaptive parameter adjustment layer adopts a dynamic parameter learning strategy to monitor the error change of the deep learning network, and adaptively adjusts the parameters according to the error change, and the adaptive topology structure layer adaptively adds or deletes the middle layer between the multi-layer self-attention structure and the feedforward neural network layer according to the characteristics of the input data; Step 5: Use the maximum pooling layer to sample the extracted feature representation, select the maximum value in each pooling window as the sampling result, and input the extracted feature representation into the normalization layer for standardization processing; Step 6, the residual connection layer adds the data directly input into the residual connection layer and the output data of the previous layer by constructing a cross-layer residual connection with the feedforward neural network layer to retain historical information; Step 7: Input the processed data into the output layer and output the results according to different task requirements; The improved recurrent neural network sequence model adopts an attention mechanism to dynamically adjust the weights of different items in the user behavior sequence prediction results; The improved recurrent neural network sequence model adopts recurrent neural network RNN ​​or long short-term memory network LSTM sequence modeling to capture the evolution and change of user behavior over time. The user behavior sequence is x = {x1, x2, ..., x t-1 ,x t }, x t Represents the input of user behavior data at the tth time step, and the modeling formula is: In formula (1), Y(t) represents the user interest preference prediction result at the tth time step, i t 、f t and t represents the input gate, forget gate, and output gate functions of the t-th time step, which are used to control the inflow and outflow of data information. t-1 represents the input of user behavior data at the t-1th time step, h t-1 Represents the hidden state of user behavior data at time step t-1; dynamically adjusts the weights of different items during recommendation according to the input information at different time steps to reflect the user's personalized preferences. The calculation formula is: In formula (2), W t represents the attention weight at the t-th time step, v a ,q a 、u a and b a Represents the parameters of the self-attention mechanism, which is used for the linear transformation of user behavior data and hidden states, h t Represents the hidden state of user behavior data at time step t; A real-time recommendation module, which generates a user-personalized product recommendation list based on the training model and dynamically adjusts the recommendation results through a real-time recommendation engine. The real-time recommendation engine triggers the recalculation of the training model and the update of the recommendation results according to the changes in the user's real-time behavior data monitored; A user feedback and iteration module, which uses a user feedback mechanism to collect user feedback on recommendation results, and dynamically adjusts recommendation strategies and model parameters based on user feedback and real-time behavior data to achieve continuous optimization and iterative update of the training model; The recommendation result explanation module displays the generation process and change trend of the recommendation results through a visualization platform, and uses an explanatory algorithm to perform explanatory analysis on the recommendation results. The explanatory algorithm identifies key data features by calculating the degree of influence of each data feature on the recommendation results, and displays the influence trend of the change of the characteristic value of a single variable on the target variable by drawing a curve chart.

2. The intelligent recommendation system based on deep learning and big data fusion according to claim 1, characterized in that: The data collection module uses an application programming interface (API) to collect user browsing records, purchase behaviors, product details and user information in user databases, e-commerce platforms and log files in real time. The data preprocessing module cleans, converts and normalizes the original collected data through a data rule engine, and uses a self-supervised representation learning method to extract and integrate user behavior characteristics, interest characteristics and item attribute characteristics to construct user portraits and item portraits.

3. The intelligent recommendation system based on deep learning and big data fusion according to claim 1, characterized in that: The model training unit adopts a small batch gradient descent optimization algorithm to train the selection model, and adopts a regularization method to prevent model overfitting. The small batch gradient descent optimization algorithm updates the parameters of the selection model by randomly extracting batch data, and minimizes the loss function by iteratively updating the parameters of the selection model. The regularization method controls the number of selection model parameters by adding penalty terms. The model performance evaluation unit adopts a cross-validation method to evaluate the performance of the training model. The cross-validation method evaluates the performance of the training model by dividing the data set into subsets and performing repeated optimization training of the training model.

4. The intelligent recommendation system based on deep learning and big data fusion according to claim 1, characterized in that: The real-time recommendation engine adopts a real-time stream processing framework to perform real-time updates and calculations on changes in user real-time behavior data. The real-time stream processing framework adopts a data caching mechanism to cache user behavior feature vectors in memory for processing, and adopts a multi-core processor for data parallel processing. The real-time stream processing framework adds new data points to the data set through a time window, updates the new data points, and obtains the membership vectors of the new data points in real time. The time window triggers recalculation and updating of recommendation results by inputting the new data points into the training model.

5. The intelligent recommendation system based on deep learning and big data fusion according to claim 1, characterized in that: The visualization platform uses interactive charts and heat maps to visualize and interactively analyze the recommendation results. The visualization platform uses a user identity authentication mechanism to verify the identity of the accessing user and records the number of times the cached data items are viewed through an access counter.

6. An intelligent recommendation method based on deep learning and big data fusion, applied to an intelligent recommendation system based on deep learning and big data fusion as described in any one of claims 1 to 5, characterized in that: The intelligent recommendation method comprises the following steps: Step 1: Collect user behavior data, product information, and user portrait data from multiple data sources; Step 2: Clean, convert and normalize the collected data, and build user and item portraits by integrating data from multiple data sources; Step 3: Select an improved deep learning algorithm, combine it with big data processing technology, train and optimize the model, and use distributed storage and parallel computing to achieve rapid data processing and analysis; Step 4: Use the trained deep learning model to generate a personalized product recommendation list for users, and dynamically adjust the recommendation results through a real-time recommendation engine; Step 5: Use the user feedback mechanism to collect user feedback on the recommendation results, and dynamically adjust the recommendation strategy and model parameters based on user feedback and real-time behavior data to achieve continuous optimization and iteration of the training model; Step 6: Use a visualization platform to display the generation process and influencing factors of the recommendation results, and use an explanatory algorithm to perform an explanatory analysis of the recommendation results.

Images