User behavior analysis and personalized recommendation method and system

Abstract

The invention relates to the technical field of user behavior data processing, and discloses a user behavior analysis and personalized recommendation method and system, and the method comprises the steps: collecting multi-dimensional behavior data of a user, and carrying out the feature extraction of the multi-dimensional behavior data; key features are defined based on behavior data, redundant features are reduced through a feature selection algorithm, the dimension of numerical features is unified through normalization processing, and an input feature vector used for a deep learning model is generated; processing the input feature vector by adopting a hybrid deep learning architecture, and generating a user-commodity interaction model in combination with the user behavior sequence and the commodity features; performing real-time prediction on the new user behavior data based on a deep learning model, and dynamically adjusting a recommendation strategy according to a prediction result; and displaying a user behavior analysis result through a visual interface, and continuously optimizing a recommendation strategy in combination with an A / B test framework. According to the method, the defects of a traditional recommendation system in the aspects of multi-source data processing, recommendation individuation and the like are effectively overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of user behavior data processing, in particular to a user behavior analysis and personalized recommendation method and system. BACKGROUND

[0002] With the rapid development of e-commerce and digital retail industry, the scale and complexity of user behavior data are growing exponentially. Traditional user behavior analysis methods mainly rely on statistical models or simple rule systems, which are difficult to effectively process massive, multi-modal, unstructured user behavior data.

[0003] In recent years, deep learning technology has made significant progress in image recognition, natural language processing and time series modeling, providing the possibility of building more intelligent and accurate user behavior analysis systems. However, how to organically integrate multiple deep learning models and combine multi-source heterogeneous data for efficient modeling is still one of the key challenges in the current technical field. Especially in the e-commerce recommendation scenario, how to use deep learning technology to achieve comprehensive understanding, prediction and personalized recommendation of user behavior still has many technical bottlenecks.

[0004] Therefore, a user behavior analysis and personalized recommendation method and system are proposed. SUMMARY

[0005] Therefore, the present application proposes a user behavior analysis and personalized recommendation method and system, which solves the shortcomings of traditional recommendation systems in multi-source data processing, user behavior modeling, recommendation personalization and system interpretability, and has good practicality and promotion prospect.

[0006] To achieve the above purpose, in a first aspect, the present application provides a user behavior analysis and personalized recommendation method, comprising:

[0007] Collecting multi-dimensional behavior data of users and performing data cleaning, standardization processing and feature extraction of unstructured data;

[0008] Defining key features based on the behavior data, reducing redundant features through feature selection algorithm, and using normalization processing to unify the dimension of numerical features, generating input feature vectors for deep learning model;

[0009] Using a hybrid deep learning architecture to process the input feature vectors, and combining user behavior sequences and product features to generate a user-product interaction model;

[0010] Based on the trained deep learning model, real-time prediction is performed on new user behavior data to generate user consumption tendency and interest labels, and the recommendation strategy is dynamically adjusted according to the prediction results to generate a personalized recommendation list;

[0011] The user behavior analysis result is displayed through a visual interface, and a recommendation strategy is continuously optimized in combination with an A / B test framework.

[0012] Preferably, the multi-dimensional behavior data of the user is collected, and data cleaning, standardization processing and feature extraction of unstructured data are performed, specifically:

[0013] The multi-dimensional behavior data of the user is collected through website logs, mobile application interfaces or third-party services, and the behavior data includes browsing records, click streams, shopping cart operations, purchase histories, search queries and evaluation feedback;

[0014] Through the data cleaning, outliers, duplicate records and noise data are removed, and the timestamp field is uniformly formatted;

[0015] Through unstructured data processing, the user evaluation text is segmented, stop word filtered and sentiment analyzed to extract emotion labels;

[0016] Through multi-modal data encoding, the user basic attributes, product attributes and behavior sequences are mapped to low-dimensional vectors through One-Hot encoding or Embedding layer, and are spliced with image features to form multi-modal input.

[0017] Preferably, the feature extraction includes static feature extraction and dynamic feature generation, specifically:

[0018] The static feature extraction models semantic relevance of product categories, price intervals and brand attributes through hierarchical encoding;

[0019] The dynamic feature generation calculates behavior frequency, conversion rate and short-term preference indicators based on user behavior sequences, and uses a sliding window to count recent interest trends of the user.

[0020] Preferably, the input feature vector is processed using a hybrid deep learning architecture, and a user-item interaction model is generated combining user behavior sequences and product features, including:

[0021] The hybrid deep learning architecture includes a recurrent neural network or a long short-term memory network for processing time series data, a Transformer model for extracting text semantics, and a convolutional neural network for product image feature extraction;

[0022] A double-tower structure is used to encode user features and product features respectively, and user-item similarity is calculated through dot product;

[0023] In sequence modeling, self-attention dynamically weights key nodes of user behavior;

[0024] We employ a combination of supervised and unsupervised learning, generating unpurchased items as negative samples through negative sampling, and introducing adversarial training to improve the model's robustness.

[0025] Preferably, the dynamically adjusted recommendation strategy includes:

[0026] The online inference engine performs forward propagation on real-time user behavior data, outputs probability distributions, and generates comprehensive prediction results.

[0027] The final recommendation list is generated by combining collaborative filtering, content-based recommendation, and hybrid recommendation strategies, and the impact weight of outdated behaviors is reduced by a time decay function.

[0028] The recommendation model parameters are dynamically adjusted based on user interaction with the recommendation results, and the recommendation basis is analyzed through SHAP values ​​to enhance the interpretability of the model.

[0029] Preferably, the step of displaying user behavior analysis results through a visual interface and continuously optimizing the recommendation strategy using an A / B testing framework includes:

[0030] The visualization interface displays the user group's interest distribution, consumption cycle trends, and behavioral path heatmaps, and provides a user profile tag filtering function;

[0031] The A / B testing framework randomly groups users to test the conversion rate differences of different recommendation strategies and outputs statistical significance conclusions to guide strategy iteration.

[0032] Based on user segmentation results, targeted discount rules are generated, and combined with sentiment analysis results, customer service intervention or product optimization suggestions are triggered.

[0033] Preferably, it also includes a privacy protection mechanism, specifically:

[0034] Hash encryption or anonymize sensitive user information;

[0035] The model is trained on a local device using federated learning techniques, avoiding the need to upload raw user data.

[0036] Ensure that data processing procedures comply with privacy regulations and provide a user feedback channel to dynamically optimize model logic.

[0037] Secondly, the present invention provides a user behavior analysis and personalized recommendation system, comprising:

[0038] The data acquisition module is used to collect multi-dimensional behavioral data from users and perform data cleaning, standardization, and feature extraction on unstructured data.

[0039] a feature module, configured to define key features based on the behavior data, reduce redundant features through a feature selection algorithm, and normalize the dimensions of numerical features to generate an input feature vector for a deep learning model;

[0040] a training module, configured to process the input feature vector using a hybrid deep learning architecture and generate a user-item interaction model in combination with user behavior sequences and product features;

[0041] a prediction module, configured to perform real-time prediction on new user behavior data based on the trained deep learning model, generate user consumption tendencies and interest labels, and dynamically adjust a recommendation strategy based on the prediction results to generate a personalized recommendation list;

[0042] a display module, configured to display user behavior analysis results through a visual interface and continuously optimize the recommendation strategy in combination with an A / B testing framework.

[0043] Preferably, the collection module specifically includes:

[0044] multidimensional behavior data of users are collected through website logs, mobile application interfaces, or third-party services, and the behavior data includes browsing records, click streams, shopping cart operations, purchase histories, search queries, and evaluation feedback;

[0045] abnormal values, duplicate records, and noise data are removed through the data cleaning, and timestamp fields are uniformly formatted;

[0046] user evaluation text is segmented, stop word filtered, and sentiment analyzed through unstructured data processing to extract emotional labels;

[0047] user basic attributes, product attributes, and behavior sequences are mapped to low-dimensional vectors through One-Hot encoding or Embedding layers through multi-modal data encoding, and are spliced with image features to form multi-modal inputs.

[0048] Preferably, the feature extraction includes static feature extraction and dynamic feature generation, specifically:

[0049] the static feature extraction models semantic relevance of product categories, price intervals, and brand attributes through hierarchical encoding;

[0050] the dynamic feature generation calculates behavior frequency, conversion rate, and short-term preference indicators based on user behavior sequences, and uses a sliding window to statistically analyze recent interest trends of users.

[0051] The embodiment of the application discloses a user behavior analysis and personalized recommendation method and system. The method enhances data quality and availability by unified data cleaning, standardization and feature extraction of unstructured data of multi-dimensional user behavior data, and provides high-quality input for subsequent model training; the expression ability of the feature vector and the model training efficiency are improved by defining key features, removing redundant information by using a feature selection algorithm, and normalizing numerical value type features, which helps to improve the overall system performance; a hybrid deep learning architecture including a recurrent neural network, a long short-term memory network, a Transformer and a convolutional neural network is adopted, which can effectively process text, image and time series behavior data, thereby generating a more accurate and comprehensive user-item interaction model; based on the trained deep learning model, real-time prediction is performed on new user behavior data, user consumption tendency and interest labels are generated, and the recommendation strategy is dynamically adjusted combined with the prediction result, which effectively improves the personalized level and conversion rate of the recommendation; the user behavior analysis result is displayed through a visual interface, and the recommendation strategy is continuously optimized combined with an A / B test framework, which provides intuitive data insight and scientific decision basis for enterprises, and helps to improve user experience and business value. BRIEF DESCRIPTION OF DRAWINGS

[0052] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not intended to be limiting in

[0053] Figure 1 A flowchart of a user behavior analysis and personalized recommendation method provided by an embodiment of the application;

[0054] Figure 2 A structure schematic diagram of a user behavior analysis and personalized recommendation system provided by an embodiment of the application. DETAILED DESCRIPTION

[0055] Exemplary embodiments of the present disclosure will be described in detail with reference to the drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in conjunction with the embodiments.

[0056] As Figure 1As shown, in some embodiments of the present application, the present embodiment provides a user behavior analysis and personalized recommendation method, specifically, the method comprises the following steps:

[0057] Step S101, collect multi-dimensional behavior data of the user, and perform data cleaning, standardization processing and feature extraction of unstructured data.

[0058] As described above, this step involves systematic collection and preprocessing of user behavior data on digital platforms. Multi-dimensional behavior data refers to various types of behavior records generated by users during the use of websites, mobile applications or other interactive platforms, including but not limited to browsing path, click operation, search query, product collection, shopping cart addition, transaction completion, evaluation feedback, etc.

[0059] In order to ensure the data quality and accuracy of subsequent modeling and recommendation process, a series of cleaning and standardization processing of the original data is needed after the collection is completed. Data cleaning mainly corrects or removes abnormal values, missing fields, duplicate records, etc.; standardization processing is used to unify the data format and measurement unit of different fields, so that various data can be compared and modeled on the same scale.

[0060] In addition, for unstructured data (such as user comments, search keywords, etc.), feature extraction needs to be performed through natural language processing technology to convert it into a numerical vector representation that can be used for model training, thereby improving the model's understanding of semantic information.

[0061] For example, a complete behavior process of a user on an e-commerce platform may include the following typical behaviors: browsing a certain type of product page; clicking on a specific product link multiple times to view details; adding the product to the shopping cart but not immediately purchasing; searching for related keywords and viewing search results; finally completing the order and submitting an evaluation. The above behavior data can be collected through website logs, front-end point-in-time, API interface calls, third-party analysis tools, etc.

[0062] In the data cleaning stage, the following problems can be identified and processed: delete invalid click events (such as robot simulated clicks); fill in missing timestamp fields; unify time format (such as "2025-07-16" and "16 / 07 / 2025" to standard format); filter abnormal browsing times (such as single-page browsing times exceeding a reasonable threshold). In terms of unstructured data processing, for example, the user's comment "This product is very good and cost-effective", the positive emotion label and keyword features can be extracted through word segmentation, stop word filtering, sentiment analysis, etc. for subsequent model use.

[0063] It should be noted that in specific implementation scenarios, on the basis of the above scheme, a multi-channel data fusion scheme can also be adopted, that is, not only limited to a single platform (such as a web page or a mobile terminal), but also external channels such as social media, third-party payment platforms, and customer service systems can be integrated to realize more comprehensive user portrait construction. Privacy desensitization processing scheme, that is, in the collection and processing stage, the user identity (such as mobile phone number, device ID) is hashed and encrypted or anonymized to meet data compliance requirements, while not affecting the subsequent modeling effect. Streaming processing scheme, that is, not only supports batch data collection, but also integrates real-time data stream processing mechanism to realize instant response and rapid modeling of user behavior, and enhances the dynamic adaptability of the system. The above optional schemes are within the protection scope of the present application.

[0064] Step S102, defining key features based on the behavior data, reducing redundant features through a feature selection algorithm, and using normalization processing to unify the dimension of numerical features, to generate an input feature vector for a deep learning model.

[0065] As described above, this step involves feature engineering processing of the collected user behavior data to construct high-quality input feature vectors suitable for deep learning models. Defining key features means extracting feature dimensions that can reflect user interests, preferences, or potential behavior tendencies from the original behavior data according to business objectives and data analysis requirements.

[0066] In actual application, the original data usually contains a large number of redundant or low-correlation fields, so a feature selection algorithm (such as information gain, LASSO regression, random forest importance score, etc.) is needed to filter out core features that have a significant impact on recommendation results, thereby reducing model complexity and improving training efficiency.

[0067] In addition, different types of numerical features often have different dimensions (for example, the number of clicks is an integer, and the price of a product may be a floating-point number). In order to eliminate the modeling bias caused by unit differences, these features need to be normalized to fall within a unified numerical interval (such as [0, 1] or [-1, 1]), ensuring that the model gives fair weights to each feature during training.

[0068] Finally, the features after feature selection and normalization processing are organized into a feature vector according to a certain structure as input data for subsequent deep learning models.

[0069] For example, in the recommendation system of an e-commerce platform, the following key features can be extracted from the user's browsing, clicking, collecting, adding to cart, and purchasing behaviors: the number of page views by the user in the past 7 days; the time interval of the user's last click on a certain type of product; the average price of a certain type of product in the user's historical orders; the frequency of a specific category word in the user's search keywords; the proportion of positive / negative sentiment polarity in the user's reviews.

[0070] In the feature selection stage, invalid behavior records such as "page dwell time less than 1 second" can be removed by evaluating the correlation between each feature and the target variable (such as whether to purchase), and behavior indicators with more predictive value are retained. In the normalization process, assuming that a feature "number of clicks on products by the user in the past 30 days" has a value range of [0, 500], and another feature "product price" has a value range of [10, 10000], the maximum and minimum normalization method can be applied to map both features to the [0, 1] interval, avoiding the price feature from dominating the model due to its large numerical value.

[0071] Finally, the above filtered and normalized multiple features are combined into a structured feature vector, such as [0.68, 0.42, 0.35, 0.71, 0.92], which can be used as input data for the deep learning model for subsequent behavior prediction and recommendation calculation.

[0072] It should be noted that in specific implementation scenarios, dynamic feature updating schemes can be adopted based on the above-mentioned schemes, i.e., new feature dimensions can be automatically identified and introduced according to the changing trend of user behavior, such as adding "user cross-device access frequency" and "behavior activity change before holidays", to enhance the model's ability to adapt to time-varying behavior; context information schemes can be combined, i.e., not only considering the user's own historical behavior, but also introducing context features (such as current geographic location, device type, access time period, etc.), to build a more rich feature space and improve the scene perception ability of the recommendation; graph structure feature schemes can be introduced, i.e., in the feature construction process, the interaction relationship between users and products can be modeled as a graph structure, and graph features such as node degree, path length, and community affiliation can be extracted to further enrich the input expression of the model. The above optional schemes are within the scope of the present application.

[0073] Step S103, using a hybrid deep learning architecture to process the input feature vector, and combining user behavior sequences and product features to generate a user-product interaction model.

[0074] As mentioned above, this step involves building and applying a hybrid deep learning architecture for modeling user behavior data after preprocessing and feature engineering. The hybrid deep learning architecture refers to the organic combination of multiple neural network structures to fully leverage their respective advantages in different data types or modeling tasks, thereby improving the overall modeling effect.

[0075] The architecture typically includes but is not limited to: Recurrent Neural Network (RNN) or Long Short-Term Memory Network (LSTM) for capturing the time series characteristics of user behavior, modeling the user's long-term interest and short-term preference; Convolutional Neural Network (CNN) for extracting visual features from product images, enhancing the understanding of product appearance attributes; Transformer model for processing unstructured text information (such as search words, reviews, etc.), mining the semantic expression therein; Multi-Layer Perceptron (MLP) or Multi-Tower structure for integrating feature vectors from different sources, achieving efficient user-item matching modeling.

[0076] During model training, the system takes the user's historical behavior sequence (such as the time sequence of browsing, clicking, collecting, purchasing, etc.) and product features (such as category, price, image, description text, etc.) as input, jointly models them through the above hybrid architecture, and finally generates a user-item interaction model that accurately reflects the potential interaction between users and products. The core goal of this model is to predict the user's interest in a certain product, providing a basis for subsequent recommendation decisions.

[0077] For example, in the application scenario of an e-commerce platform, assume that a user has completed the following behaviors in the past period of time: browsed multiple sports shoe product pages; clicked to view the details of a pair of A brand running shoes; added a pair of B brand running shoes to the shopping cart but did not place an order; searched for the keyword "light and breathable running shoes"; and finally purchased a pair of C brand sports shoes. At the same time, the system obtains the feature information of the product, including its category "sports shoes", brand "C", price range "mid-range", picture showing black and breathable design, and product description containing keywords such as "cushioning", "light", and "suitable for long-distance running".

[0078] Based on this information, the system uses: LSTM network to model the user's behavior sequence, identify the behavior path from "browsing" to "purchase", and recognize the evolution of interest; CNN network to extract visual features such as color and style from product images; Transformer model to analyze product description text and identify product advantages represented by keywords such as "cushioning" and "light"; Multi-Tower structure to encode user features and product features respectively, and calculate the similarity score between the two through dot product. Finally, the system outputs a numerical value representing the user's interest in the product, which is used for recommendation ranking or click-through rate prediction.

[0079] It should be noted that in specific implementation scenarios, on the basis of the above scheme, the following optional schemes can be adopted: an attention mechanism scheme, i.e., introducing a self-attention (Self-Attention) mechanism in time series modeling, dynamically weighting the importance of key nodes in user historical behavior, and improving the sensitivity of the model to changes in user interest; a graph neural network scheme, i.e., constructing the interaction relationship between users and commodities into a graph structure, modeling high-order relationship propagation paths using a graph neural network (GNN), and further improving the diversity and generalization ability of recommendations; a multi-task joint scheme, i.e., predicting multiple targets (such as click rate, conversion rate, rating, and repeat purchase probability) simultaneously when constructing a user-commodity interaction model, and improving the overall performance of the model by sharing the underlying feature representation; and a contrastive learning mechanism scheme, i.e., guiding the model to learn more discriminative user and commodity embedding representations by contrasting positive and negative sample pairs, and improving the accuracy of the recommendation system. The above optional schemes are all within the protection scope of the present application.

[0080] Step S104, based on the trained deep learning model, real-time prediction is performed on new user behavior data, user consumption tendency and interest labels are generated, and the recommendation strategy is dynamically adjusted according to the prediction results to generate a personalized recommendation list.

[0081] As described above, this step involves applying the trained deep learning model to the online inference phase in actual business scenarios. New user behavior data refers to the latest interaction behavior of users continuously collected after the model is deployed, such as browsing, clicking, searching, collecting, adding to cart, ordering, etc. These behavior data usually have strong timeliness and can reflect the current interest status and potential consumption tendency of users.

[0082] The system inputs the above real-time behavior data into the trained deep learning model, and the model outputs corresponding prediction results, including but not limited to: the preference degree of a user for a certain commodity category; the probability of a user's purchase behavior in the future period of time; and label classification of whether the user is a high-value user, a loss risk user, or an ordinary user, etc. Based on the above prediction results, the system further generates user interest labels (such as "quality priority customer", "high-frequency buyer in a certain field", "price-sensitive user", etc.), and dynamically adjusts the recommendation strategy in combination with the current context information (such as time, location, device type, etc.).

[0083] Dynamically adjusting the recommendation strategy means that the system can automatically select different recommendation logic or weight combinations according to the prediction results, such as preferentially displaying high-profit goods when a user shows high purchase willingness, or pushing coupons or recall strategies when a user is on the verge of loss. Finally, the system generates a personalized recommendation list according to the adjusted recommendation strategy and returns it to the front-end application interface for the user to view.

[0084] For example, a user completes the following behaviors in a visit to an e-commerce platform: browses multiple smart watch product pages; clicks on a product detail page of a product with heart rate monitoring function; searches for the keyword "health monitoring smart watch"; adds a product to the shopping cart but does not immediately place an order. The system transmits the above behavior data into the trained deep learning model in real time, and the model predicts that the probability of the user purchasing such products within the next 24 hours is 78%, and labels it as an interest tag of "paying attention to health class electronic products".

[0085] Subsequently, the recommendation system makes the following dynamic adjustments based on this prediction result: prioritizing the display of smart wearable devices related to health monitoring in the recommendation list; at the same time, reducing the exposure weight of irrelevant categories (such as clothing category) goods; combining user historical preferences (such as preference for high-end brands), filtering out low-priced inferior goods; if it is detected that the user has browsed multiple times but not ordered, a coupon pop-up window or a time-limited discount prompt is triggered. Finally, the system generates a personalized recommendation list that highly matches the user's current interest and consumption ability, improving the user's click and conversion probability.

[0086] It should be noted that in specific implementation scenarios, on the basis of the above scheme, a multi-strategy fusion scheme can also be adopted, that is, when generating the recommendation list, the system can simultaneously call multiple algorithms such as collaborative filtering, content-based recommendation, and popularity-based recommendation, and dynamically allocate the weights of each strategy according to the user portrait, so as to balance accuracy and diversity; adopt a time decay weight adjustment scheme, that is, when calculating the user's interest, give higher weight to more recent behaviors and reduce the influence of historical behaviors, so as to more accurately capture the user's current intention; adopt a user feedback closed loop scheme, that is, the system can dynamically adjust the model parameters based on the user's subsequent behaviors (such as clicking, purchasing, sliding and skipping, etc.) to the recommendation results, so that the recommendation system has a certain self-optimization ability. The above optional schemes are within the protection scope of the present application.

[0087] Step S105, displaying the user behavior analysis result through a visual interface, and continuously optimizing the recommendation strategy based on an A / B test framework.

[0088] As described above, this step involves presenting the key data and model output results generated during the operation of the user behavior analysis and the recommendation system to the operation personnel or business decision makers in an intuitive and easy-to-understand manner. The visual interface refers to presenting complex user behavior characteristics, interest distribution, and product preferences in a graphical form through charts, heat maps, trend curves, and tag clouds, etc., to help users quickly grasp the overall portrait and behavior rules of platform users.

[0089] In addition, to ensure that the recommendation strategy can continuously optimize and adapt to changing user needs, the system also introduces an A / B testing framework. This framework supports dividing user groups into multiple control groups within the same time period, each group using different recommendation algorithms or parameter configurations. By comparing the performance of each group in click-through rate, conversion rate, dwell time, and other core indicators, the effectiveness of different strategies is evaluated, and scientific evidence is provided for subsequent strategy iteration.

[0090] Through the above mechanisms, not only the transparency and operability of data analysis are improved, but also a closed-loop feedback path is provided for continuous improvement of the recommendation system, thus achieving dynamic improvement of recommendation effectiveness and significant enhancement of operational efficiency.

[0091] For example, in an e-commerce platform, the system displays the following user behavior analysis results through a visual interface: user interest distribution chart, showing the most popular product categories in the current user group, such as "sports clothing", "digital products", "baby products", etc., presented in pie chart or bar chart form; user behavior path heat map, showing the main browsing path and loss nodes of users from the homepage to the ordering process, helping to identify problems in page design or process setup; user clustering tag cloud, showing user tags based on consumption inclination, such as "high-value users", "price-sensitive users", "new customer conversion potential users", etc., facilitating fine-grained operations; recommendation effectiveness overview table, showing key performance indicators such as average click-through rate, add-to-cart rate, conversion rate under the current recommendation strategy, for reference by operational personnel.

[0092] At the same time, the system starts an A / B testing experiment, randomly dividing users into three groups: Group A uses the original recommendation strategy (control group); Group B uses the new recommendation sorting logic with a time decay factor; Group C uses the recommendation weight adjustment mechanism that integrates user sentiment analysis. After a week of operation, the system automatically calculates the differences in key indicators for each group and finds that the click-through rate of Group C has increased by 8%, and the conversion rate has increased by 6%, with statistical significance. Based on this, the system recommends that the recommendation strategy of Group C be used as the new default strategy online, and records the data from this experiment for subsequent strategy optimization analysis.

[0093] It should be noted that in a specific implementation scenario, on the basis of the above scheme, a multi-dimensional user group and behavior insight linkage analysis scheme can also be adopted, that is, on the visual interface, the user can filter the user group according to multiple dimensions such as region, device type, access time period, and synchronously view the behavior mode and recommendation response situation, to improve the flexibility and depth of operation analysis; an automatic experiment management mechanism scheme is adopted, that is, the A / B test framework can support automatic creation, running, monitoring and termination of experiment tasks, and automatically generate strategy update suggestions according to the preset threshold, to reduce manual intervention and improve test efficiency; a knowledge base and backtracking mechanism scheme is adopted, that is, the system can automatically record the configuration, result and final decision of each A / B test, form the knowledge accumulation of strategy evolution, and support quick rollback or reuse of historical strategy versions. The above optional schemes all belong to the protection scope of the present application.

[0094] In some embodiments of the present application, in order to realize efficient and accurate user behavior analysis and personalized recommendation, it is necessary to first systematically collect and deeply preprocess multi-dimensional behavior data of users. The multi-dimensional behavior data of users is collected, and data cleaning, standardization processing and feature extraction of unstructured data are performed, specifically:

[0095] The multi-dimensional behavior data of users is collected through website logs, mobile application interfaces or third-party services, and the behavior data includes browsing records, click streams, shopping cart operations, purchase history, search queries and evaluation feedback;

[0096] The data cleaning removes outliers, duplicate records and noise data, and uniformly formats the timestamp field;

[0097] The unstructured data processing performs word segmentation, stop word filtering and sentiment analysis on user evaluation text to extract emotion labels;

[0098] Through multi-modal data encoding, user basic attributes, product attributes and behavior sequences are mapped to low-dimensional vectors through One-Hot encoding or Embedding layer, and are spliced with image features to form multi-modal input.

[0099] As described above, first, the system collects user behavior data through multiple channels, including but not limited to website logs, mobile application interfaces (APIs) and third-party service interfaces. The collected data covers various interactive behaviors of users on the platform, such as records of browsing product pages, click operation paths, shopping cart addition and deletion behaviors, order purchase history, search keyword input, and evaluation feedback on goods or services. The above behavior data reflects the user's interest preferences, consumption ability and potential intentions from different angles, and is the basis for subsequent modeling and recommendation.

[0100] Due to the quality issues of the original collected data, such as outliers, duplicate records or invalid fields, it needs to be cleaned. The cleaning process includes identifying and removing data records that do not conform to logic, such as inconsistent timestamp formats, illegal operation sequences in behavior sequences, etc. In addition, the system also uniformly formats all timestamp fields to ensure that time information from different sources can be accurately parsed and compared.

[0101] When facing unstructured data, especially user-submitted evaluation text, the system uses natural language processing technology for in-depth analysis. Specifically, the text content is first segmented, that is, continuous strings are divided into semantic word units; then common but meaningless stop words (such as "of", "was", "is", etc.) are filtered out to reduce redundant information; finally, through sentiment analysis algorithm to judge the sentiment tendency of the text, extract emotion labels (such as positive, negative, neutral), so as to enhance the model's understanding ability of user's subjective attitude.

[0102] In order to make the collected various data can be effectively used by deep learning model, the system further encodes and fuses the data. For user basic attributes (such as gender, age), commodity attributes (such as category, brand) and behavior sequence (such as click order, browsing path) and other discrete variables, the system uses One-Hot encoding or Embedding embedding layer to map them into low-dimensional vector form. At the same time, the system also integrates visual features from commodity images, and splices them with the above structured features to form a unified multi-modal input vector. This vector will be used as input data for subsequent feature engineering and deep learning model, supporting more comprehensive and accurate user behavior modeling and personalized recommendation generation.

[0103] In some embodiments of the present application, in order to improve the expression ability and prediction accuracy of user behavior analysis and personalized recommendation model, the system further extracts features after data preprocessing. The feature extraction includes static feature extraction and dynamic feature generation, specifically:

[0104] The static feature extraction models the semantic association of commodity categories, price intervals and brand attributes through hierarchical coding;

[0105] The dynamic feature generation calculates the behavior frequency, conversion rate and short-term preference index based on the user behavior sequence, and uses a sliding window to calculate the recent interest trend of the user.

[0106] As described above, in the static feature extraction part, the system structurally models the inherent attributes of the commodity. Specifically, for discrete attributes such as commodity category, price range, brand, etc., the system uses a hierarchical coding method to model the semantic correlation. For example, commodity categories can be coded according to a hierarchical structure such as "electronic products > mobile phones > smartphones", price ranges can be divided into "low price, medium price, high price" and other levels, and brands can be grouped or vectorized according to market positioning and user perception. Through this hierarchical processing method, not only the semantic relationship between attributes can be preserved, but also the ability to identify the substitutability or correlation between similar commodities can be enhanced, thereby improving the accuracy and interpretability of the recommendation.

[0107] In the dynamic feature generation part, the system extracts time series features that can reflect the interest changes of the user based on the user's historical behavior sequence. Specifically, the system first calculates the behavior frequency of the user within a certain time window, such as the total number of behaviors such as browsing, clicking, adding to cart, and ordering; secondly, it calculates the conversion rate between key behaviors, such as the conversion proportion from "clicking on a commodity" to "adding to cart" and then to "completing an order", which is used to evaluate the strength of the user's purchase intention; in addition, the system also constructs short-term preference indicators, such as the commodity category that the user recently focuses on and the preferred price range, which are used to capture the current interest tendency of the user. In order to better reflect the dynamic change trend of the user's interest, the system introduces a sliding window mechanism, i.e. continuously updating the behavior statistics within a certain time range (such as the last 7 days), so that the model can perceive the evolution of the user's interest in real time, thereby improving the timeliness and personalization of the recommendation.

[0108] Through the combination of the above static features and dynamic features, the system can more comprehensively depict the interaction between the user and the commodity, providing high-quality input features with semantic expression ability for the training of subsequent deep learning models, significantly improving the modeling efficiency and prediction performance of the recommendation system.

[0109] In some embodiments of the present application, in order to more comprehensively model the complex interaction between the user and the commodity, the system uses a hybrid deep learning architecture that combines the advantages of multiple neural network structures, which can simultaneously process multi-modal input data and effectively capture the temporal evolution characteristics of user behavior and the multi-dimensional attribute information of commodities. The hybrid deep learning architecture is used to process the input feature vector, and a user-commodity interaction model is generated by combining the user behavior sequence and the commodity features, including:

[0110] The hybrid deep learning architecture includes a recurrent neural network or a long short-term memory network for processing time series data, a Transformer model for extracting text semantics, and a convolutional neural network for commodity image feature extraction;

[0111] The user features and the commodity features are encoded respectively by using a double-tower structure, and the user-commodity similarity is calculated by dot product;

[0112] The key nodes of the user behavior are dynamically weighted by introducing self-attention mechanism in sequence modeling.

[0113] The combination of supervised learning and unsupervised learning is adopted, the non-purchased commodities are generated as negative samples by negative sampling, and the model robustness is improved by introducing adversarial training.

[0114] As described above, first, the system uses recurrent neural network (RNN) or long short-term memory network (LSTM) to model the user's behavior sequence. These behaviors include but are not limited to browsing, clicking, searching, adding to cart, and purchasing, etc., which have obvious time sequence characteristics. Through the RNN / LSTM structure, the system can extract the long-term dependence relationship in the user behavior, so as to more accurately depict the user's interest evolution path. Second, for commodity description text information (such as commodity title, detail page introduction, etc.), the system introduces the Transformer model for semantic modeling. Transformer can effectively identify the key semantic units in the text through self-attention mechanism, and generate high-quality text embedding representation, so as to enhance the model's understanding ability of commodity attributes. In addition, the system also uses convolutional neural network (CNN) to extract features of commodity images. Through multi-level convolution operation on commodity pictures, CNN can automatically extract visual features in the image, such as color distribution, texture features, commodity style, etc., to provide more rich perceptual information for subsequent recommendation decision.

[0115] After completing the above feature encoding, the system further uses a double-tower structure to model the user features and the commodity features respectively. Among them, the user tower is responsible for integrating the user's behavior sequence, static attributes and dynamic preference features; the commodity tower is responsible for fusing the image features, text description and basic attribute information of the commodity. The two towers output corresponding high-dimensional vector representations respectively, and then calculate the similarity score between the user and the commodity through dot product operation as the basis for recommendation ranking.

[0116] In order to further improve the model's ability to identify key behavior nodes in the user behavior sequence, the system introduces self-attention mechanism in the sequence modeling process. This mechanism can dynamically allocate weights according to the importance of user historical behavior, so that the model pays more attention to the key behavior nodes that truly reflect the user's intention, thereby improving the prediction accuracy and personalization degree.

[0117] In terms of model training, the system adopts a combination of supervised learning and unsupervised learning. In the supervised learning phase, the system uses the user's real historical purchase behavior as positive samples, and through negative sampling technology, randomly selects several samples from the non-purchased goods as negative samples to optimize the discrimination ability of the model. In the unsupervised learning phase, the system introduces an adversarial training mechanism, simulates the distribution of real user behavior through the game process between the generator and the discriminator, and further enhances the generalization ability and robustness of the model.

[0118] In some embodiments of the present application, in order to improve the real-time response ability and recommendation quality of personalized recommendation, the system introduces a dynamic adjustment recommendation strategy mechanism, which can continuously optimize the recommendation results according to the latest behavior data and interaction feedback of the user, thereby enhancing the accuracy of the recommendation and user satisfaction. The dynamic adjustment recommendation strategy includes:

[0119] Forward propagation of real-time behavior data of the user through the online inference engine, output of probability distribution and generation of comprehensive prediction results;

[0120] Combining collaborative filtering, content-based recommendation and hybrid recommendation strategy to generate the final recommendation list, and reducing the influence weight of outdated behavior through a time decay function;

[0121] Dynamic adjustment of recommendation model parameters according to the user's interaction behavior on the recommendation results, and analysis of the basis for recommendation through SHAP value to enhance the explainability of the model.

[0122] As described above, first, the system receives the real-time behavior data of the user through the online inference engine and inputs it into the trained deep learning model for forward propagation calculation. This process can be completed in milliseconds, outputting the user's interest probability distribution for various goods. Based on this distribution, the system further generates a comprehensive prediction result to guide the subsequent recommendation sorting and display logic.

[0123] In generating the final recommendation list, the system adopts a multi-strategy fusion recommendation method, combining the advantages of multiple recommendation algorithms. Specifically, the system combines collaborative filtering recommendation (based on the behavior preferences of similar users), content-based recommendation (based on the matching degree of the attributes of the goods and the user's historical preferences), and hybrid recommendation strategy (integrating multiple recommendation logics and weighted fusion) to improve the diversity and accuracy of the recommendation results. At the same time, the system introduces a time decay function to assign a time-decreasing weight to the user's historical behavior, ensuring that the model pays more attention to recent behavior and reduces the influence of outdated behavior on the recommendation results, thereby better reflecting the user's current interest state.

[0124] In addition, to achieve continuous optimization of the recommendation system, the system also supports dynamic adjustment of recommendation model parameters based on user interaction behavior with the recommendation results. For example, when a user clicks on a recommended product or adds it to the shopping cart, the system can use these behaviors as feedback signals to update the relevant parameters in the recommendation model, allowing the model to adapt more quickly to changing trends in user interest. To further enhance the transparency and credibility of the recommendation system, the system also introduces a SHAP value analysis mechanism. SHAP (Shapley Additive Explanations) is a feature importance explanation method based on game theory, which can quantify the influence of each input feature on the recommendation result. Through this mechanism, the system can analyze the decision-making basis behind why a certain type of product is recommended to the user, such as because the user has recently frequently browsed similar products or because there are highly relevant product categories in their historical purchase records. This explanation mechanism not only helps to improve user experience, but also provides operations personnel with clearer insights into model behavior, facilitating subsequent strategy optimization.

[0125] In some embodiments of the present application, to improve the readability of user behavior analysis results and the scientificity of recommendation system strategy adjustment, the system introduces a continuous optimization mechanism combining a visualization interface display mechanism and an A / B testing framework. This mechanism not only helps operations personnel to intuitively understand user behavior characteristics and interest distribution, but also iteratively optimizes recommendation strategies through experimental verification, thereby improving the intelligent level and business value of the overall system. The continuous optimization of recommendation strategies through the visualization interface display of user behavior analysis results and the A / B testing framework includes:

[0126] The visualization interface displays the interest distribution, consumption cycle trend, and behavior path heat map of the user group, and provides a filtering function for user portrait labels;

[0127] The A / B testing framework randomly groups users to test the conversion rate difference of different recommendation strategies and outputs statistically significant conclusions to guide strategy iteration;

[0128] According to the user grouping results, generate targeted preferential rules, and trigger customer service intervention or product optimization suggestions combined with the emotional analysis results.

[0129] As mentioned above, first, in terms of visual interface display, the system presents the results of user behavior analysis in a graphical manner, facilitating the operation personnel to quickly grasp the overall portrait and behavior regularity of platform users. Specifically: the system displays the popularity of different commodity categories in the current platform through user group interest distribution graph, for example, in the form of pie chart or bar chart to show the attention heat of "digital products", "maternal and child products", "home appliances" and other categories; through the consumption cycle trend chart, the time distribution regularity of user purchase behavior is presented, such as the change of weekly activity, the consumption peak before and after holidays, etc., to help formulate more accurate marketing rhythm; the user behavior path heat map is used to show the main browsing path and loss node of users from access to order, to identify unreasonable page jump and low conversion rate key links; at the same time, the system also provides user portrait label filtering function, allowing operation personnel to filter specific user groups according to "high value users", "price sensitive users", "new customer conversion potential users" and other labels, and view their corresponding behavior characteristics and recommendation response, to provide data support for fine operation.

[0130] Secondly, in the application of A / B test framework, the system supports the comparison and evaluation of different recommendation strategies and the verification of their effects. Specifically: the system randomly divides users into multiple experimental groups, each group uses different recommendation algorithm configuration or parameter combination, for example, one group uses collaborative filtering based recommendation logic, and another group uses time decay factor optimized sorting logic; within the set time period, the system automatically collects the click rate, add-to-cart rate, conversion rate and other key indicators of each group of users, and performs statistical analysis; finally, the system outputs statistically significant conclusions to determine which recommendation strategy performs better in actual application, thereby providing a scientific basis for subsequent strategy online or model update.

[0131] In addition, in order to further exert the value of data analysis, the system also generates specific business optimization suggestions combined with user segmentation and emotion analysis results: based on user segmentation results (such as high-frequency purchase users, potential loss users, price-sensitive users, etc.), the system automatically generates targeted discount rules, for example, pushing recall coupons to loss risk users, providing exclusive discounts to high value users, to improve user stickiness and repurchase rate; at the same time, the system combines the emotion analysis results of user evaluation text, when detecting that a certain product has more negative feedback, it can trigger the active intervention mechanism of customer service, or send optimization suggestions to the product team to promote product improvement or service upgrade.

[0132] In some embodiments of the present application, in order to protect user data security and meet the requirements of relevant privacy regulations, this embodiment introduces a privacy protection mechanism, which runs through the whole process of data collection, model training and strategy optimization, to ensure that the recommendation effect is improved while effectively reducing the risk of user privacy leakage. The privacy protection mechanism includes:

[0133] Hash encryption or anonymization of user sensitive information;

[0134] Training models on local devices using federated learning techniques, avoiding uploading raw user data;

[0135] Ensuring data processing complies with privacy regulations and providing user feedback channels to dynamically optimize model logic.

[0136] As mentioned above, first, in the data processing phase, the system performs hash encryption or anonymization on information related to user identity recognition or sensitive behavior. For example, user's mobile phone number, device ID, real name, and other identifiable information will be converted into unique identifiers through irreversible hash algorithms; user's behavior records will also be removed from direct association with their identity, making them unable to be restored to the original user information. This processing method not only retains the value of data for modeling analysis, but also effectively prevents the leakage of user privacy.

[0137] Second, to further reduce the security risks brought by the transmission and centralized storage of raw user data, the system introduces federated learning technology. This technology allows the model to be trained on the user's local device (such as mobile phone, tablet computer), and only uploads the model parameter update results to the server side for aggregation, without uploading the user's real behavior data. This distributed training method not only improves data security, but also enhances user trust in the platform's data usage.

[0138] In addition, the system strictly follows the requirements of existing privacy protection laws and regulations in the overall data processing process. The system monitors key links such as data collection range, use purpose, and storage period through the built-in compliance checking module, ensuring that each operation has a legal basis, and supporting user's right to access, correct, delete, and other rights. At the same time, in order to enhance the transparency and user participation of the system, the system also provides user feedback channels. Users can view their interest tags, recommendation preference settings, and mark or feedback inappropriate recommended content through the platform interface. These feedback information will be used as an important reference for model optimization, to dynamically adjust the recommendation logic, so that the recommended results are more in line with the actual needs and acceptance level of users.

[0139] Compared with the prior art, the embodiment of the application discloses a user behavior analysis and personalized recommendation method, which enhances data quality and availability by unified data cleaning, standardization and feature extraction of unstructured data of multi-dimensional user behavior data, and provides high-quality input for subsequent model training; the expression ability of the feature vector and the model training efficiency are improved by defining key features, removing redundant information by using a feature selection algorithm, and normalizing numerical features, which helps to improve the overall system performance; a hybrid deep learning architecture including a recurrent neural network (RNN), a long short-term memory network (LSTM), a Transformer and a convolutional neural network (CNN) is adopted, which can effectively process text, image and time series behavior data, thereby generating a more accurate and comprehensive user-item interaction model; based on the trained deep learning model, new user behavior data is predicted in real time, user consumption tendency and interest labels are generated, and the recommendation strategy is dynamically adjusted according to the prediction result, thereby effectively improving the personalized level and conversion rate of the recommendation; the user behavior analysis result is displayed through a visual interface, and the recommendation strategy is continuously optimized in combination with an A / B test framework, thereby providing intuitive data insight and scientific decision basis for enterprises, and helping to improve user experience and business value.

[0140] Based on the same inventive concept as the above method, the embodiment of the application also proposes a user behavior analysis and personalized recommendation system, as shown in Figure 2 , which is a structural schematic diagram of a user behavior analysis and personalized recommendation system, the system comprising:

[0141] a collection module for collecting multi-dimensional behavior data of users and performing data cleaning, standardization and feature extraction of unstructured data on the multi-dimensional behavior data;

[0142] a feature module for defining key features based on the behavior data, reducing redundant features by using a feature selection algorithm, and using normalization processing to unify the dimension of numerical features, thereby generating an input feature vector for a deep learning model;

[0143] a training module for processing the input feature vector by using a hybrid deep learning architecture, and generating a user-item interaction model in combination with user behavior sequences and product features;

[0144] a prediction module for predicting new user behavior data in real time based on the trained deep learning model, generating user consumption tendency and interest labels, and dynamically adjusting a recommendation strategy according to the prediction result, thereby generating a personalized recommendation list;

[0145] a display module for displaying user behavior analysis results through a visual interface, and continuously optimizing the recommendation strategy in combination with an A / B test framework.

[0146] Preferably, the collection module, in particular:

[0147] Collecting multi-dimensional behavior data of users through website logs, mobile application interfaces or third-party services, the behavior data including browsing records, click streams, shopping cart operations, purchase histories, search queries and evaluation feedbacks;

[0148] Removing outliers, duplicate records and noise data through the data cleaning, and uniformly formatting the timestamp field;

[0149] Performing word segmentation, stopword filtering and sentiment analysis on user evaluation texts through unstructured data processing to extract emotional labels;

[0150] Mapping user basic attributes, commodity attributes and behavior sequences into low-dimensional vectors through One-Hot encoding or Embedding layers through multi-modal data encoding, and splicing image features to form multi-modal inputs.

[0151] Preferably, the feature extraction includes static feature extraction and dynamic feature generation, in particular:

[0152] The static feature extraction models semantic relevance of commodity categories, price intervals and brand attributes through hierarchical encoding;

[0153] The dynamic feature generation calculates behavior frequency, conversion rate and short-term preference indicators based on user behavior sequences, and utilizes a sliding window to count recent interest change trends of users.

[0154] Those skilled in the art will understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer-usable program code.

[0155] The present application is described with reference to flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the flowcharts and / or block diagrams. Figure 1 one flow or multiple flows and / or blocksFigure 1 means for performing the function specified in the block or blocks.

[0156] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 flow or flows and / or blocks Figure 1 means for performing the function specified in the block or blocks.

[0157] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 flow or flows and / or blocks Figure 1 steps for performing the function specified in the block or blocks.

[0158] Finally, it should be noted that the above-mentioned embodiments are merely intended to illustrate the technical solutions of the present application, rather than limit the same. Although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and any modification or replacement without departing from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A method for user behavior analysis and personalized recommendation, characterized in that, include: Collect multi-dimensional behavioral data from users, and perform data cleaning, standardization, and feature extraction on unstructured data. Based on the behavioral data, key features are defined, redundant features are reduced through feature selection algorithms, and the units of numerical features are unified by normalization processing to generate input feature vectors for deep learning models. The input feature vector is processed using a hybrid deep learning architecture, and a user-product interaction model is generated by combining user behavior sequences with product features. Based on the trained deep learning model, new user behavior data is predicted in real time to generate user consumption tendencies and interest tags, and the recommendation strategy is dynamically adjusted according to the prediction results to generate a personalized recommendation list. The results of user behavior analysis are displayed through a visual interface, and the recommendation strategy is continuously optimized by combining an A / B testing framework.

2. The method as described in claim 1, characterized in that, The process of collecting multi-dimensional user behavior data, and then performing data cleaning, standardization, and feature extraction on unstructured data, specifically includes: The system collects multi-dimensional user behavior data through website logs, mobile application interfaces, or third-party services. The behavior data includes browsing history, click stream, shopping cart operations, purchase history, search queries, and evaluation feedback. The data cleaning process removes outliers, duplicate records, and noisy data, and the timestamp field is uniformly formatted. Unstructured data processing is used to segment user review texts, filter stop words, and perform sentiment analysis to extract sentiment tags. Multimodal data encoding maps user basic attributes, product attributes, and behavioral sequences into low-dimensional vectors through One-Hot encoding or Embedding layers, and then concatenates them with image features to form a multimodal input.

3. The method as described in claim 2, characterized in that, The feature extraction includes static feature extraction and dynamic feature generation, specifically: The static feature extraction uses hierarchical coding to model the semantic association between product categories, price ranges, and brand attributes; The dynamic feature generation is based on calculating behavior frequency, conversion rate and short-term preference indicators based on user behavior sequence, and uses a sliding window to statistically analyze the recent interest change trend of users.

4. The method as described in claim 1, characterized in that, The process of using a hybrid deep learning architecture to process the input feature vector and combining user behavior sequences with product features to generate a user-product interaction model includes: The hybrid deep learning architecture includes a recurrent neural network or long short-term memory network for processing temporal data, a Transformer model for extracting text semantics, and a convolutional neural network for extracting product image features. A dual-tower structure is used to encode user features and product features separately, and user-product similarity is calculated by dot product. Introduce self-attention dynamic weighted key nodes of user behavior into sequence modeling; We employ a combination of supervised and unsupervised learning, generating unpurchased items as negative samples through negative sampling, and introducing adversarial training to improve the model's robustness.

5. The method as described in claim 1, characterized in that, The dynamically adjusted recommendation strategy includes: The online inference engine performs forward propagation on real-time user behavior data, outputs probability distributions, and generates comprehensive prediction results. The final recommendation list is generated by combining collaborative filtering, content-based recommendation, and hybrid recommendation strategies, and the impact weight of outdated behaviors is reduced by a time decay function. The recommendation model parameters are dynamically adjusted based on user interaction with the recommendation results, and the recommendation basis is analyzed through SHAP values ​​to enhance the interpretability of the model.

6. The method as described in claim 1, characterized in that, The process of displaying user behavior analysis results through a visual interface and continuously optimizing the recommendation strategy using an A / B testing framework includes: The visualization interface displays the user group's interest distribution, consumption cycle trends, and behavioral path heatmaps, and provides a user profile tag filtering function; The A / B testing framework randomly groups users to test the conversion rate differences of different recommendation strategies and outputs statistical significance conclusions to guide strategy iteration. Based on user segmentation results, targeted discount rules are generated, and combined with sentiment analysis results, customer service intervention or product optimization suggestions are triggered.

7. The method as described in claim 1, characterized in that, It also includes privacy protection mechanisms, specifically: Hash encryption or anonymize sensitive user information; The model is trained on a local device using federated learning techniques, avoiding the need to upload raw user data. Ensure that data processing procedures comply with privacy regulations and provide a user feedback channel to dynamically optimize model logic.

8. A user behavior analysis and personalized recommendation system, characterized in that, include: The data acquisition module is used to collect multi-dimensional behavioral data from users and perform data cleaning, standardization, and feature extraction on unstructured data. The feature module is used to define key features based on the behavioral data, reduce redundant features through feature selection algorithms, and use normalization processing to unify the dimensions of numerical features to generate input feature vectors for deep learning models. The training module uses a hybrid deep learning architecture to process the input feature vector and combines user behavior sequences with product features to generate a user-product interaction model. The prediction module is used to make real-time predictions on new user behavior data based on the trained deep learning model, generate user consumption tendencies and interest tags, and dynamically adjust the recommendation strategy according to the prediction results to generate a personalized recommendation list. The presentation module is used to display user behavior analysis results through a visual interface and to continuously optimize the recommendation strategy in conjunction with an A / B testing framework.

9. The system as described in claim 8, characterized in that, The acquisition module is specifically: The system collects multi-dimensional user behavior data through website logs, mobile application interfaces, or third-party services. The behavior data includes browsing history, click stream, shopping cart operations, purchase history, search queries, and evaluation feedback. The data cleaning process removes outliers, duplicate records, and noisy data, and the timestamp field is uniformly formatted. Unstructured data processing is used to segment user review texts, filter stop words, and perform sentiment analysis to extract sentiment tags. Multimodal data encoding maps user basic attributes, product attributes, and behavioral sequences into low-dimensional vectors through One-Hot encoding or Embedding layers, and then concatenates them with image features to form a multimodal input.

10. The system as described in claim 9, characterized in that, The feature extraction includes static feature extraction and dynamic feature generation, specifically: The static feature extraction uses hierarchical coding to model the semantic association between product categories, price ranges, and brand attributes; The dynamic feature generation is based on calculating behavior frequency, conversion rate and short-term preference indicators based on user behavior sequence, and uses a sliding window to statistically analyze the recent interest change trend of users.

Images