E-commerce marketing propaganda system based on behavior analysis

Abstract

The invention relates to the technical field of big data, in particular to an e-commerce marketing propaganda system based on behavior analysis, which comprises a global user portrait module, an intelligent recommendation engine, a marketing fatigue management module, a supply chain collaboration module, a lightweight terminal module and a budget and value management module. In the prior art, user portraits are constructed only depending on single channel data such as online clicking or purchase records, so that user interest modeling is incomplete and lagged; according to the method, all-channel behavior data such as APP, Web, offline POS and social media are integrated, a user-commodity-scene heterogeneous graph is constructed by using a graph neural network, and an interest attenuation period (for example, the weight is reduced by 50% after the interest of mother and infant users lasts for 18 months) is dynamically captured through an LSTM model; for example, after a user tries on a certain style of clothes offline, the system associates online behaviors in real time and recommends commodities of the same style, the cross-scene conversion rate is improved by 32%, and the user portrait coverage degree is improved by 60%.

Description

Technical Field

[0001] This invention relates to the field of big data technology, and in particular to an e-commerce marketing and promotion system based on behavioral analysis. Background Technology

[0002] Existing e-commerce systems still face significant challenges in data integration, real-time response, user experience optimization, and supply chain collaboration. Traditional user profiling relies heavily on single-channel behavioral data such as online clicks or purchase records, lacking deep integration with cross-channel scenarios like offline consumption and social media interactions. This results in incomplete and lagging interest modeling. For example, after a user tries on products in a physical store, the online recommendation system often fails to capture this behavior in time, missing cross-scenario conversion opportunities. Furthermore, the dynamic nature of user interests over time is not fully explored, and static tagging systems struggle to reflect short-term behavioral fluctuations and long-term interest decay. For instance, the phased demand for maternal and infant products causes recommendation results to deviate from the user's actual needs.

[0003] In terms of recommendation strategies, most systems are still limited to single methods such as collaborative filtering or content matching, failing to effectively integrate multimodal data such as product review text, visual features, real-time weather, and geographic location. Traditional collaborative filtering cannot analyze users' deep needs for product functions, such as waterproof features, while rule-based recommendations lack flexible adaptation to sudden scenarios, such as automatically increasing the weight of rain gear during rainfall. Furthermore, insufficient control over the diversity of recommendation results leads to user fatigue due to repeated exposure to similar products, resulting in a high close rate. Marketing fatigue management often relies on fixed frequency limits or simple demotion, lacking scientific quantitative models. This can lead to the loss of high-value users due to excessive push notifications, while also overlooking opportunities to activate dormant users.

[0004] The disconnect between the supply chain and the recommendation system further impacts user experience. Popular items are heavily recommended even when inventory is low, leading to stock shortages after users place orders, while slow-moving items accumulate in warehouses due to insufficient exposure. Logistics scheduling relies on human experience, making it difficult to respond promptly to traffic congestion or inclement weather, resulting in frequent delivery delays. For example, during heavy rain, delivery routes were not dynamically adjusted, causing a surge in order delays and a decline in user satisfaction.

[0005] On the terminal side, mobile recommendation models are typically bulky, with high loading latency and energy consumption, and frequent service interruptions in weak network environments. Traditional centralized data processing poses privacy risks, and uploading user behavior data to the cloud raises compliance controversies. Purely localized processing, on the other hand, struggles to achieve continuous model optimization, creating a zero-sum game between efficiency and privacy. Furthermore, existing systems generally neglect dynamic analysis of user spending power and cycles, recommending high-priced items even when budgets are tight at the end of the month, leading to a sharp drop in conversion rates. They also lack fine-grained value segmentation, such as identifying high-net-worth users, making it difficult to provide differentiated services.

[0006] Despite the gradual application of technologies such as graph neural networks and federated learning in recent years, existing solutions still suffer from problems such as low accuracy in cross-channel ID mapping, poor efficiency in multimodal feature fusion, and high error rates in sales prediction. For example, collaborative filtering significantly reduces recommendation effectiveness in data-sparse scenarios, while the communication overhead and model accuracy of federated learning are difficult to balance. These limitations restrict the refined operational capabilities of e-commerce platforms, necessitating a technical solution that integrates full-domain data, multimodal recommendation, dynamic fatigue management, and real-time supply chain collaboration to achieve a dual improvement in user experience and business efficiency while ensuring privacy. Summary of the Invention

[0007] To overcome the problems mentioned in the background art, the present invention proposes an e-commerce marketing and promotion system based on behavior analysis.

[0008] The technical solution of this invention is: an e-commerce marketing and promotion system based on behavior analysis, comprising:

[0009] The comprehensive user profiling module integrates user behavior data from multiple channels and constructs a heterogeneous graph of users, products, and scenarios using a graph neural network model, while building user interest profiles using an LSTM model. By integrating user behavior data from multiple channels and using a graph neural network model to construct a heterogeneous graph of users, products, and scenarios, as well as an LSTM model to build user interest profiles, this module can gain a more comprehensive and in-depth understanding of user needs and preferences, providing solid data support for subsequent precision marketing and effectively improving the targeting and effectiveness of marketing activities.

[0010] The intelligent recommendation engine uses a BERT model to parse product review tags and a ResNet-50 model to extract visual style features. Combined with manufacturer data, it constructs a multi-strategy fusion recommendation system. This technical solution can more accurately capture product characteristics and user preferences, enabling personalized recommendations, improving user satisfaction and purchase conversion rates, thereby driving e-commerce business growth.

[0011] The marketing fatigue management module uses an exponential decay model to predict the decline in user click-through rates and dynamically adjusts push strategies. This effectively prevents user fatigue caused by excessive marketing. This technical solution helps maintain user interest in marketing campaigns, improves user experience, and also helps improve the efficiency of marketing resource utilization.

[0012] The supply chain collaboration module uses the Prophet algorithm to predict sales for the next 7 days and combines it with a dynamic safety stock formula to achieve intelligent replenishment. This technical solution optimizes inventory management, reduces inventory backlog and stockout risks, and improves the responsiveness and flexibility of the supply chain, thereby ensuring the smooth operation of e-commerce business.

[0013] The lightweight terminal module uses 1-bit quantization technology to compress the DNN model, enables local processing of user data through federated learning, and dynamically adapts to device performance. This solution significantly reduces the model's performance requirements, improves the terminal device's operating efficiency, protects user privacy, and expands the system's applicability, allowing more users to enjoy seamless e-commerce services.

[0014] The budget and value management module analyzes user price sensitivity based on historical orders, predicts monthly spending budgets, and adjusts recommendation strategies in conjunction with payday cycles. This technological solution helps merchants more accurately grasp user spending power and habits, formulate reasonable pricing and marketing strategies, improve marketing effectiveness and sales, and also contribute to enhancing user loyalty and satisfaction.

[0015] As a preferred embodiment, the full-domain user profiling module, when integrating user behavior data from multiple channels, constructing a heterogeneous graph of user-product-scenario through a graph neural network model, and building user interest profiles through LSTM, specifically includes:

[0016] S11: Data Acquisition and Synchronization. This feature uses a data tracking SDK to capture user behavior events in real-time on the app and web, and synchronizes offline POS transaction data to the data center. This comprehensive integration of user behavior data from multiple channels enables the system to acquire user behavior data in different scenarios, providing a rich and comprehensive data foundation for subsequent user profiling and facilitating a more accurate understanding of user needs and behavioral patterns.

[0017] S12: Identity Fusion. A unique identifier is generated based on the device fingerprint and associated with the user account upon login, while historical behavioral data is merged under the user account. This solves the problem of user identification across different devices or channels, enabling cross-device and cross-channel tracking of user behavior. Through identity fusion, the system can more accurately associate user behavioral data, improving the accuracy and completeness of user profiles.

[0018] S13: Short-term interest capture. Using a Flink sliding window to analyze high-frequency behaviors over the past hour, the system updates user interest tags in real time. This allows for the real-time capture of short-term changes in user interests. By analyzing recent user behavior data, the system can promptly identify user interests and update interest tags in real time, making recommended content more aligned with current user preferences and improving the timeliness and accuracy of recommendations.

[0019] S14: Long-term interest modeling. This involves analyzing historical behavior using XGBoost, calculating preference weights, and predicting interest decay cycles using an LSTM model to obtain the user's long-term interest model. This allows for in-depth analysis of users' long-term interests and preferences. By analyzing users' historical behavioral data, the system can calculate the preference weights for different goods or services and predict interest decay cycles, thus constructing a long-term interest model. This helps the system to more comprehensively understand user needs and provide more personalized recommendation services.

[0020] S15: Feature fusion and compression concatenates the obtained user interest tags and long-term interest models into a 512-dimensional original vector, and then uses an encoder to reduce the dimensionality of the 512-dimensional original vector into a 64-dimensional dense feature vector. This optimizes the storage and computational efficiency of user profiles. Through feature fusion and compression, the system can reduce the dimensionality of data and storage space while retaining key information, improving the running efficiency of subsequent recommendation algorithms. At the same time, the dimensionality-reduced dense feature vector is easier to process and analyze, contributing to improved overall system performance.

[0021] As a preferred approach, when constructing a multi-strategy fusion recommendation system based on a BERT model to parse product review tags, extracting visual style features using a ResNet-50 model, and combining manufacturer data, the intelligent recommendation engine specifically includes:

[0022] S21: Data Input. Input a 64-dimensional dense feature vector, product database data, geographic location data, weather data, and time data. This step provides the recommendation system with multi-dimensional information input by providing these data. This comprehensive data input method helps the system more accurately understand user needs and contextual environment, laying a solid foundation for the subsequent recommendation process, thereby improving the accuracy and personalization of recommendations.

[0023] S22: In the recall phase, 10a candidate products are selected from the product database through collaborative filtering, content recall, and best-selling product supplementation. During this phase, the system uses various strategies, including collaborative filtering, content recall, and best-selling product supplementation, to select a large number of candidate products from the product database. This efficiently filters products related to user interests from a massive amount of data, providing a rich candidate set for subsequent fine-grained ranking, while ensuring the diversity and coverage of recommendations.

[0024] S23: In the ranking stage, based on the Wide&Deep model, user characteristics, product attributes, geographic location data, weather data, and time data are integrated to predict product click-through rates and dynamically weight them to quantify the intensity of user interest in each product. Combined with business rules, the ranking priority is optimized to select a selected products from 10a candidate products. This process comprehensively considers multiple factors such as user characteristics, product attributes, and contextual environment to achieve more accurate product ranking and improve the click-through rate and conversion rate of recommendation results.

[0025] S24: In the re-ranking phase, products are sorted using fatigue filtering and diversity control, while manual rules are injected to generate the final product list. Fatigue filtering removes products that users have seen but not clicked within a certain period, while diversity control manages the variety of product categories, price distribution, and visual differences displayed on each screen. The re-ranking phase optimizes product sorting through fatigue filtering and diversity control, and injects manual rules to generate the final product list. Fatigue filtering helps remove products that users have seen but not clicked within a certain period, preventing user fatigue; diversity control ensures the diversity of products displayed on each screen in terms of category, price distribution, and visual differences, improving user experience. This further optimizes the quality of the recommendation list, increasing user satisfaction and platform activity.

[0026] As a preferred approach, when selecting 10n candidate products from the product database through collaborative filtering, content recall, and best-selling product supplementation, the principle and steps of collaborative filtering include:

[0027] S31: Similarity calculation, which calculates the cosine of the angle between the 64-dimensional dense feature vectors of users. The principle formula is as follows:

[0028]

[0029] Among them, u i Indicates the intensity of user u's behavior towards the product, v i This represents the intensity of user v's behavior towards a product; the similarity between users is measured by calculating the cosine of the angle between their 64-dimensional dense feature vectors. This calculation method quantifies the similarity between user behavior patterns, providing an accurate metric for finding similar users and helping to improve the accuracy of collaborative filtering recommendations.

[0030] S32: Similar user filtering uses an approximate nearest neighbor algorithm to accelerate calculations, quickly finding the Top 100 similar users from tens of millions of users. The similar user list is updated fully at fixed intervals, with active user data updated incrementally every hour. This combination of full periodic updates and hourly incremental updates ensures both filtering efficiency and the timeliness of the similar user list. This helps the system promptly capture changes in user interests, improving the real-time performance and accuracy of recommendations.

[0031] S33: Candidate product aggregation: This function aggregates products purchased by similar users, filters products with low ratings and low exposure / conversion rates, and selects the top 3a products for the candidate pool based on a weighted ranking of purchase frequency and rating. This process, based on the purchasing behavior of similar users, filters out products more likely to align with the target user's interests while eliminating low-quality or unpopular products, thus improving the overall quality of the candidate products.

[0032] As a preferred approach, when selecting 10n candidate products from the product database through collaborative filtering, content recall, and best-selling product supplementation, the principle and steps of content recall include:

[0033] S41: User Interest Tag Extraction. This step involves obtaining explicit tags from the 64-dimensional vector generated by the user profile module and assigning weights to these tags based on behavior intensity. This process accurately extracts user interest tags and quantifies the degree of user preference for different interest points. This helps the system gain a deeper understanding of user needs and provides precise targets for subsequent content retrieval.

[0034] S42: Product Content Analysis. A pre-trained BERT model is used to encode product reviews, extract keywords, and calculate keyword importance. Common words are filtered out, and differentiated feature words are retained. Specifically, during keyword extraction, a graph neural network is used to construct a product-keyword relationship graph to capture implicit associations. This step allows for in-depth mining of valuable information in product reviews, extracting keywords that represent product characteristics, and capturing the implicit associations between products and keywords through the graph, providing high-quality product features for subsequent matching and filtering.

[0035] S43: Matching and Filtering. This step calculates the similarity between user interest tag vectors and product keyword vectors using cosine similarity, and filters products based on a set threshold. This process accurately matches products that match user interests and filters out irrelevant products using thresholds, improving recommendation accuracy and user satisfaction.

[0036] As a preferred embodiment, the marketing fatigue management module, when predicting the downward trend of user click-through rate through an exponential decay model and dynamically adjusting the push strategy, specifically includes:

[0037] S51: Decay model modeling, using a decay exponential model to describe the decay law of click-through rate with the number of impressions:

[0038] CTR(t) = CTR0 × e -βt ;

[0039] Wherein, CTR(t) is the click-through rate after t exposures, CTR0 is the baseline click-through rate of the first exposure, β is the decay coefficient, and t is the cumulative number of exposures. The decay coefficient β is determined by fitting user behavior data. The decay exponential model is used to describe the decay law of click-through rate with the number of exposures. It can accurately predict the trend of user click-through rate decreasing with the increase of exposures, providing a scientific basis for subsequent fatigue judgment and push strategy adjustment, and helping to avoid user fatigue caused by excessive push.

[0040] S52: Fatigue assessment. The system calculates the user's fatigue score in real time and determines whether the user has entered a fatigued state based on a set threshold. The formula for calculating the user fatigue score is as follows:

[0041] It calculates user fatigue scores in real time and determines whether a user has entered a fatigue state based on a set threshold. This allows for timely detection of user fatigue levels, providing accurate timing for adjustments to subsequent push notification strategies. This ensures timely adjustments before user fatigue sets in, improving user experience.

[0042] S53: Frequency control strategy adjustment. Based on user fatigue scores, the exposure interval is dynamically extended and the recommendation strategy is switched. Pop-up frequency is reduced for highly fatigued users, while cross-category products are introduced. This allows for flexible adjustment of the push strategy according to user fatigue levels, reducing user aversion to repetitive content, and increasing the novelty of recommendations by introducing cross-category products, thereby improving user engagement.

[0043] S54: Creative content updates utilize material carousel and intelligent cropping technologies to increase visual differentiation on the first screen. This maintains the diversity and freshness of recommended content, attracting user attention and improving click-through rates and conversion rates through enhanced visual differentiation.

[0044] S55: Strategy Implementation Output. This output combines fatigue scores, frequency control rules, and content update results to generate the final push strategy. This ensures the scientific validity and effectiveness of the push strategy by comprehensively considering multiple factors to generate the most suitable push strategy for the current user state, thereby improving marketing results.

[0045] As a preferred option, the supply chain collaboration module, when predicting sales for the next 7 days using the Prophet algorithm and combining it with a dynamic safety stock formula to achieve intelligent replenishment, specifically includes:

[0046] S61: Demand Forecasting. Using the Prophet model, historical sales, seasonality, and promotional plans are analyzed to predict product demand for the next 7 days. A safety stock is calculated using a dynamic safety stock formula. An alert is triggered when inventory falls below 120% of the predicted sales volume. The dynamic safety stock formula is as follows:

[0047] S = μ + Zσ;

[0048] Where S is the safety stock threshold, μ is the average daily sales volume, Z is the coefficient, and σ is the standard deviation of sales volume, used to measure the volatility of sales. The Prophet model is used to analyze historical sales, seasonality, and promotional plans to predict the demand for goods in the next 7 days, and the safety stock is calculated using a dynamic safety stock formula. This accurately predicts future sales, providing a scientific basis for inventory management. Furthermore, the application of the dynamic safety stock formula ensures that inventory levels meet demand without causing excessive overstocking.

[0049] S62: Allocation Decisions. Real-time monitoring of warehouse inventory. When inventory levels fall below the safety stock threshold, a replenishment task is generated. Based on real-time inventory levels and geographical distance, replenishment is prioritized from the nearest regional warehouse, and the replenishment quantity is optimized using a cost model. This system enables timely response to inventory shortages, reduces replenishment costs, and improves supply chain efficiency through optimized allocation decisions.

[0050] S63: Logistics optimization, integrating real-time traffic API for dynamic route planning and using genetic algorithms to solve for optimal routes for multiple vehicles; This allows for dynamic adjustment of delivery routes based on real-time traffic conditions, improving delivery efficiency, reducing transportation costs and time, and enhancing user satisfaction.

[0051] As a preferred embodiment, the lightweight terminal module, when using 1-bit quantization technology to compress the DNN model, achieving local processing of user data through federated learning, and dynamically adapting to device performance, specifically includes:

[0052] S71: Model compression utilizes a pass-through estimator to binarize 32-bit floating-point weights during training and employs binary masks to accelerate matrix operations during inference. This significantly reduces model size and computational complexity, enabling efficient operation on resource-constrained devices and improving user experience.

[0053] S72: Local inference. A 4MB lightweight model is loaded onto the terminal device for real-time inference, switching to local caching in weak network environments. This ensures stable recommendation services even in weak network conditions, reducing network dependence and improving system availability and response speed through local caching technology.

[0054] S73: Federated learning updates model gradients locally on the terminal using user behavior data, adds Gaussian noise to protect privacy, and encrypts and uploads gradients to the cloud for aggregation every 24 hours. After a global model update, the model is then distributed back to the terminal. This allows for continuous model iteration and optimization while protecting user privacy. By fully utilizing the computing resources of the terminal device through federated learning technology, it improves the efficiency and accuracy of model training.

[0055] S74: Output results. The terminal returns a recommendation list in real time. In weak network conditions, it relies on caching to provide offline service, while simultaneously using federated learning to iterate the model. This ensures that the terminal device can provide stable recommendation services in any network environment, improves user experience through offline caching technology, and continuously optimizes and updates the model through federated learning.

[0056] As a preferred option, the budget and value management module, when analyzing user price sensitivity based on historical orders, predicting monthly spending budgets, and adjusting recommendation strategies in conjunction with payday cycles, specifically includes:

[0057] S81: Price Range Analysis. By statistically analyzing the price distribution of users' historical orders and calculating a price sensitivity index, this feature identifies price-sensitive and stable users, dynamically filtering recommended product price ranges. It accurately pinpoints users' price sensitivity, providing product recommendations that better match their consumption preferences, thereby improving user satisfaction and purchase conversion rates.

[0058] S82: Budget Forecasting. Based on a linear regression model, this predicts the user's monthly spending limit and dynamically adjusts the recommendation strategy according to the pay cycle. It can anticipate the user's spending power in advance and recommend higher-value products when the user has ample funds, thereby increasing sales; and recommend more cost-effective products when the user is short on funds, preventing users from abandoning purchases due to high prices and improving user experience.

[0059] S83: Value Operations. This uses the RFM model to calculate user value scores, categorizing users into high-value and medium-to-low-value levels. High-value users receive dedicated customer service and priority shipping, while low-value users trigger coupon recall strategies. The formula for calculating user value scores is as follows:

[0060] V = 0.4 × R norm +0.3×F norm +0.3×M norm ;

[0061] Where V represents the user value score, and R... norm F is the normalized score for the most recent consumption time. norm M is the normalized consumption frequency score. nprm The system generates a normalized consumption amount score; it uses the RFM model to calculate user value scores, categorizing users into high-value and low-to-medium-value levels. High-value users receive dedicated customer service and priority shipping, while low-value users are targeted with a coupon recall strategy. This approach allows for differentiated services based on user value segmentation, increasing loyalty and repurchase rates among high-value users, while activating low-value users through coupon recall strategies, thereby boosting overall user activity.

[0062] S84: Output and Execution. It integrates price ranges, budget forecasts, and value stratification results to generate dynamic price filtering conditions and operational actions. This allows for the integration of multiple information sources to generate more accurate and personalized recommendation strategies, improving recommendation effectiveness and operational efficiency.

[0063] As a preferred approach, the full-domain user profiling module, intelligent recommendation engine, marketing fatigue management module, supply chain collaboration module, lightweight terminal module, and budget and value management module achieve synergy through the following steps:

[0064] S91: Data Acquisition and Integration. This integrates user-end event tracking data, business system orders, inventory, and external data. It uses a real-time Kafka pipeline to clean up invalid exposures and resolve device ID conflicts. This comprehensive and accurate data acquisition provides a solid foundation for subsequent user profiling and intelligent recommendation generation, while ensuring data accuracy and consistency.

[0065] S92: User profile construction, capturing short-term interests in real time, modeling long-term features offline, and compressing 512-dimensional features to 64 dimensions to dynamically generate lightweight user vectors. This efficient user profile construction retains key user features while reducing computational complexity, providing strong support for subsequent intelligent recommendations.

[0066] S93: Intelligent recommendation generation. The recall layer integrates collaborative filtering, content matching, and a backup of popular recommendations. The ranking layer uses a Wide & Deep model to predict click-through rate, and the re-ranking layer filters out overexposure and controls category diversity. This enables the generation of more accurate and diverse recommendation lists, improving the accuracy of the recommendation system and user satisfaction.

[0067] S94: Marketing fatigue control, dynamically extending exposure intervals based on an exponential decay model, and employing content carousel and intelligent cropping; This avoids user fatigue due to overexposure while maintaining the freshness and appeal of recommended content, thus increasing user engagement;

[0068] S95: Supply chain coordination, using the Prophet model to predict sales, dynamically allocate resources, and optimize logistics. This allows for advance sales forecasting and inventory adjustments, ensuring timely and accurate product supply. Simultaneously, logistics optimization reduces transportation costs and time, improving supply chain efficiency.

[0069] S96: Terminal display and feedback. It adapts to multi-device performance through a 1-bit quantization model, updates gradients locally using federated learning, and loads cached recommendations in weak network conditions. This ensures the efficient operation of the recommendation system on different devices while protecting user privacy and enabling continuous model iteration and optimization, thus improving user experience and recommendation performance.

[0070] The beneficial effects of this invention are:

[0071] 1. Compared to existing technologies that rely solely on single-channel data such as online clicks or purchase records to build user profiles, resulting in incomplete and delayed user interest modeling, this invention integrates behavioral data from APP, Web, offline POS, social media, and other channels. It utilizes graph neural networks to construct a heterogeneous user-product-scenario graph and dynamically captures interest decay cycles using an LSTM model (e.g., the weight of maternal and infant users' interests decreases by 50% after 18 months). For example, after a user tries on a piece of clothing offline, the system correlates it with online behavior in real time and recommends similar products, achieving a 32% increase in cross-scenario conversion rate, a 60% increase in user profile coverage, and a reduction in long-term interest prediction error to 8%.

[0072] 2. Compared to existing technologies that use single recommendation strategies such as collaborative filtering or keyword matching, which cannot analyze product functional requirements (such as "waterproof" and "breathable") and visual style differences, this invention extracts semantic tags from product reviews using the BERT model (e.g., converting "good breathability" into the "breathability" tag), combines it with ResNet-50 to extract visual features from product images (e.g., identifying "minimalist style" and "retro style"), and dynamically integrates real-time scene data such as weather and geographical location (e.g., increasing the weight of rain gear by 1.5 times on rainy days). This multimodal recommendation increases click-through rate by 58%, improves the conversion rate of scene-adapted products by 26%, and achieves a visual style matching accuracy of over 85%.

[0073] 3. Compared to existing technologies that use fixed push frequencies or simple demotion for fatigue management, leading to the loss of high-value users or insufficient activation of dormant users, this invention quantifies the downward trend of user click-through rate based on an exponential decay model. When the exposure effect is lower than 60% of the benchmark value, the push interval is dynamically extended (e.g., from 30 minutes to 2 hours), and A / B / C material carousel is started (the template is changed every 2 hours). For example, after a user receives pushes of the same type of products for 3 consecutive days, the system automatically inserts 30% new cross-category products, the user close rate decreases by 40%, the exposure ratio of high-margin products remains stable at 10%, and the repurchase rate of fatigued users increases by 25%.

[0074] 4. Compared to existing technologies where recommendation systems are disconnected from the supply chain, leading to high stockout rates (30%) for best-selling products or stockpiling of slow-moving goods, this invention uses the Prophet algorithm to predict sales for the next 7 days (error rate <9%), combined with a dynamic safety stock formula (S = μ + 1.96σ) to monitor inventory levels in real time. When the inventory of a product is 120% lower than the predicted value, cross-warehouse transfer is automatically triggered (response time < 4 hours), and logistics routes are optimized based on genetic algorithms (such as avoiding congested sections due to heavy rain). For example, when a popular product is running low on inventory, the system prioritizes transferring goods from the nearest warehouse and marks them as "in stock for immediate delivery," reducing the stockout rate by 62%, increasing next-day delivery coverage to 92%, and decreasing the logistics delay rate by 43%.

[0075] 5. Compared to existing technologies that use cloud-based recommendation models of hundreds of MB, resulting in high mobile loading latency (200ms), frequent service interruptions in weak network conditions, and significant risks associated with centralized processing of user privacy data, this invention compresses the model to 4MB using 1-bit quantization technology (accuracy loss <3%), improves inference speed by 4 times to 58ms, and employs federated learning to achieve local gradient updates (adding Gaussian noise to protect privacy) and encrypts the upload of aggregation parameters. For example, users can still obtain recommendations based on local caching even without a network, achieving 99% availability in weak network conditions and reducing the risk of privacy leakage by 92%. At the same time, the model is continuously optimized through millions of gradient iterations per day, improving recommendation accuracy by 15%. Attached Figure Description

[0076] Figure 1 The diagram shown is a schematic representation of the structure of the e-commerce marketing and promotion system based on behavior analysis according to the present invention.

[0077] Figure 2 The diagram illustrates the workflow of the e-commerce marketing and promotion system based on behavior analysis according to the present invention. Detailed Implementation

[0078] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0079] Please see Figure 1-Figure 2 This invention provides an embodiment: an e-commerce marketing and promotion system based on behavior analysis, comprising:

[0080] Full-domain user profile module:

[0081] This is used to integrate user behavior data from multiple channels, construct a heterogeneous graph of users, products, and scenarios using a graph neural network model, and build user interest profiles using LSTM. Specifically, it includes:

[0082] Data collection and synchronization: Real-time capture of user behavior events on the APP and Web terminals via the event tracking SDK, and synchronization of offline POS transaction data to the data center; Identity fusion: Generate unique identifiers based on device fingerprints, associate the unique identifiers with user accounts when users log in, and merge historical behavior data under the user account; Short-term interest capture: Use Flink sliding window to count high-frequency behaviors in the past hour and update user interest tags in real time; Long-term interest modeling: Analyze historical behavior through XGBoost, calculate preference weights, and predict the interest decay cycle through an LSTM model to obtain the user's long-term interest model; Feature fusion and compression: Concatenate the obtained user interest tags and long-term interest model into a 512-dimensional original vector, and reduce the dimensionality of the 512-dimensional original vector into a 64-dimensional dense feature vector through an encoder;

[0083] Intelligent recommendation engine:

[0084] This system is used to parse product review tags based on the BERT model, extract visual style features using the ResNet-50 model, and combine manufacturer data to build a multi-strategy fusion recommendation system, specifically including:

[0085] The data input includes a 64-dimensional dense feature vector, product database data, geographic location data, weather data, and time data. In the recall phase, collaborative filtering, content recall, and best-selling product supplementation are used to select 10a candidate products from the product database. In the ranking phase, a Wide&Deep model is used to fuse user features, product attributes, geographic location data, weather data, and time data to predict product click-through rates. Dynamic weighting is applied to quantify the intensity of user interest in each product, and business rules are combined to optimize ranking priorities, selecting a selected products from the 10a candidate products. In the re-ranking phase, products are ranked using fatigue filtering and diversity control, while manual rules are injected to generate the final product list.

[0086] Among these, fatigue filtering removes products that users have seen but not clicked within a certain period, while diversity control controls the diversity of product categories, price distributions, and visual differences displayed on each screen. When selecting 10n candidate products from the product database through collaborative filtering, content recall, and best-selling product supplementation, the principle and steps of collaborative filtering include: similarity calculation, which calculates the cosine of the angle between the 64-dimensional dense feature vectors of users. The principle formula is: Among them, u i Indicates the intensity of user u's behavior towards the product, v iThis represents the intensity of user v's behavior towards a product; similar user filtering uses an approximate nearest neighbor algorithm to accelerate calculations, quickly finding the Top 100 similar users from tens of millions of users, and updating the similar user list at fixed intervals with incremental updates to active user data every hour; candidate product aggregation involves statistically analyzing products purchased by similar users, filtering out products with low ratings and low exposure conversion rates, sorting by purchase frequency and rating, and selecting the Top 3a products to enter the candidate pool.

[0087] In the process of selecting 10n candidate products from the product database through collaborative filtering, content recall, and best-selling product supplementation, the principle and steps of content recall include: user interest tag extraction, obtaining explicit tags from the 64-dimensional vector of the user profile module, and assigning weights to explicit tags according to behavioral intensity; product content analysis, using a pre-trained BERT model to encode product reviews, extract keywords, calculate keyword importance, filter common words, and retain differentiated feature words. During keyword extraction, a graph neural network is used to construct a product-keyword relationship graph to capture implicit associations; matching and filtering, calculating the similarity between the user interest tag vector and the product keyword vector using cosine similarity, and filtering products according to a set threshold.

[0088] Marketing fatigue management module:

[0089] This tool is used to predict the decline in user click-through rates using an exponential decay model and to dynamically adjust push strategies, specifically including:

[0090] The decay model is used to describe the decay of click-through rate (CTR) with the number of exposures: CTR(t) = CTR0 × e -βt Where CTR(t) is the click-through rate after t exposures, CTR0 is the baseline click-through rate of the first exposure, β is the decay coefficient, and t is the cumulative number of exposures. The decay coefficient β is determined by fitting user behavior data. Fatigue assessment involves real-time calculation of user fatigue scores and determination of user fatigue based on a set threshold. The formula for calculating the user fatigue score is as follows: Frequency control strategy adjustment: dynamically extend the exposure interval and switch recommendation strategies based on user fatigue scores, reduce pop-up frequency for highly fatigued users, and introduce cross-category products; creative content update: use material carousel and intelligent cropping technology to update creative content, increase the visual difference of the first screen; strategy effective output: combine fatigue scores, frequency control rules and content update results to generate the final push strategy;

[0091] Supply chain collaboration module;

[0092] This tool is used to predict sales over the next 7 days using the Prophet algorithm, and combines this with a dynamic safety stock formula to achieve intelligent replenishment. Specifically, it includes:

[0093] Demand forecasting uses the Prophet model to analyze historical sales, seasonality, and promotional plans to predict product demand for the next 7 days. A dynamic safety stock formula is used to calculate safety stock, triggering an alert when inventory falls below 120% of predicted sales. The dynamic safety stock formula is: S = μ + Zσ, where S is the safety stock threshold, μ is the average daily sales, Z is a coefficient, and σ is the standard deviation of sales, used to measure sales volatility. Allocation decisions involve real-time monitoring of warehouse inventory. When inventory falls below the safety stock threshold, a replenishment task is generated. Based on real-time inventory levels and geographical distance, replenishment is prioritized from the nearest regional warehouse, and the replenishment quantity is optimized using a cost model. Logistics optimization integrates real-time traffic APIs to dynamically plan delivery routes, using a genetic algorithm to solve for optimal routes for multiple vehicles.

[0094] Lightweight terminal module:

[0095] This technology is used to compress DNN models using 1-bit quantization, achieve local processing of user data through federated learning, and dynamically adapt to device performance. Specifically, it includes:

[0096] Model compression involves binarizing 32-bit floating-point weights during training using a pass-through estimator and accelerating matrix operations during inference using binary masks. Local inference loads a lightweight 4MB model onto the terminal device for real-time inference, switching to local caching in weak network conditions. Federated learning updates the model gradient locally on the terminal using user behavior data, adds Gaussian noise to protect privacy, and encrypts and uploads gradients to the cloud for aggregation every 24 hours. After a global model update, the results are distributed back to the terminal. Output results are returned to the terminal in real-time, with offline service provided via caching in weak network conditions, while model iteration is achieved through federated learning.

[0097] Budget and Value Management Module:

[0098] This tool is used to analyze user price sensitivity based on historical orders, predict monthly spending budgets, and adjust recommendation strategies in conjunction with payday cycles. Specifically, it includes:

[0099] Price range analysis involves statistically analyzing the price distribution of users' historical orders and calculating a price sensitivity index to identify price-sensitive and stable users, dynamically filtering recommended product price ranges. Budget prediction uses a linear regression model to predict users' monthly spending ceiling and dynamically adjusts recommendation strategies based on payday cycles. Value management employs an RFM model to calculate user value scores, categorizing users into high-value, medium-low-value, and high-value levels. High-value users receive dedicated customer service and priority shipping, while low-value users trigger coupon recall strategies. The formula for calculating user value scores is: V = 0.4 × R. norm +0.3×F norm +0.3×M normWhere V represents the user value score, and R... norm F is the normalized score for the most recent consumption time. norm M is the normalized consumption frequency score. norm The normalized consumption amount score is used for output and execution. Based on the combined results of price range, budget forecast, and value stratification, dynamic price filtering conditions and operational actions are generated.

[0100] The specific workflow of this system is as follows:

[0101] Step 1: Data collection and integration. Integrate user-end tracking data, business system orders, inventory, and external data. Clean up invalid exposures and fix device ID conflicts through Kafka real-time pipeline.

[0102] Step 2: User profile construction, real-time capture of short-term interests, offline modeling of long-term features, and compression of 512-dimensional features to 64-dimensional features to dynamically generate lightweight user vectors;

[0103] Step 3: Intelligent recommendation generation, recall layer integrates collaborative filtering, content matching and best-selling product backup, ranking layer uses Wide & Deep model to predict click rate, and re-ranking layer filters fatigue exposure and controls category diversity;

[0104] Step 4: Marketing fatigue control, dynamically extending the exposure interval based on the exponential decay model, and using material carousel and intelligent cropping;

[0105] Step 5: Supply chain coordination, using the Prophet model to predict sales, dynamically allocate resources, and optimize logistics;

[0106] Step Six: Terminal Display and Feedback. Adapt the model to multi-terminal performance through 1-bit quantization, update gradients locally through federated learning, and load cached recommendations in weak network conditions.

[0107] Example 1: Optimizing the Cross-Channel Shopping Experience

[0108] Scenario Description: After trying on a pair of sneakers in a physical store, users can browse similar products through a mobile app. The system recommends suitable products in real time and ensures inventory supply, achieving a seamless shopping experience.

[0109] Detailed implementation process:

[0110] 1. Building a comprehensive user profile

[0111] Data collection: Offline POS systems capture try-on records (SKU, time, store location), while the tracking SDK collects in-app browsing behavior in real time (clicks on the sneaker details page, scrolling speed). A temporary ID is generated using the device fingerprint (IP + device ID), which is linked to the user's mobile phone number after login, merging offline try-on data with online browsing data.

[0112] Interest Modeling: Short-term interest: Flink sliding window (30 minutes) counts users' consecutive clicks on 3 athletic shoes, labeling them with the real-time tag "Athletic Shoes - Breathability". Long-term interest: XGBoost analyzes users' historical purchase cycles (e.g., buying athletic shoes every 3 months), and LSTM predicts the interest decay point to be 3 weeks later. Feature compression: The 512-dimensional behavior vector (trying on + browsing) is compressed to 64 dimensions using AutoEncoder and stored in Redis.

[0113] 2. Intelligent recommendation engine triggered

[0114] Recall Phase: Collaborative Filtering: Calculate the top 50 similar users (cosine similarity > 0.85) and aggregate their purchased 10 sneakers. Content Recall: BERT analyzes keywords from user reviews of products they tried on (e.g., "good breathability") and matches them with tags for similar products. Bestseller Supplement: Insert the top 10 best-selling sneakers of the day.

[0115] Ranking Phase: The Wide&Deep model takes a 64-dimensional user vector, product price (200-500 yuan), and inventory status (50 items) as input to predict CTR. Dynamic weighting: Due to the user's location in the city where the fitting store is located, a local inventory sufficiency coefficient is multiplied by 1.5.

[0116] Reordering phase: Remove the 5 sports shoes that users have seen in the past 24 hours to ensure that each screen displays 3 categories (such as sports shoes, sports socks, and protective gear).

[0117] 3. Marketing fatigue management

[0118] Attenuation model monitoring: If a user is exposed for 5 consecutive times without clicking, fatigue control is triggered (CTR drops to 55% of the baseline value).

[0119] Strategy Adjustment: Extend the recommendation interval to 2 hours, insert 30% fitness accessories (such as fitness trackers), and display images and text of "breathable" and "cushioned" models in a rotating format.

[0120] 4. Supply chain collaborative response

[0121] Sales forecast: Prophet predicts that the sneakers will sell 200 pairs in the next 7 days (95% confidence interval), with a safety stock of 240 pairs.

[0122] Allocation execution: The local warehouse has only 180 pairs left, so 60 pairs will be transferred from a nearby warehouse (response time 3 hours), and the logistics route will avoid peak sections.

[0123] 5. Lightweight terminal display

[0124] The 1-bit quantization model generates a recommendation list within 58ms on the app, and loads the locally cached "This Week's Hot Sneakers" when the network is weak. Federated learning updates user click data and encrypts and uploads gradient parameters (daily average traffic <100KB).

[0125] 6. Budget and Value Management

[0126] Price filtering: For users with a PSI of 0.52 (price sensitive), products priced above 500 yuan are filtered out, and products in the 200-400 yuan range are recommended. Salary cycle adaptation: For users whose salary is paid on the 5th of each month, new products are recommended from the 5th to the 12th, and discounted products are recommended after the 25th.

[0127] Results: Cross-channel conversion rate increased by 45%, out-of-stock rate decreased by 50%, and user repurchase rate increased by 30% the next day.

[0128] Example 2: Adaptation to Sudden Weather Scenarios

[0129] Scenario Description: During heavy rain, the system dynamically recommends rain gear and ensures regional inventory and delivery timeliness, thereby improving the ability to respond to emergency needs.

[0130] Detailed implementation process:

[0131] 1. Update of user profiles across the entire domain

[0132] External data integration: Connect to the meteorological API, mark the probability of rainfall in the user's city as >90%, and activate the user's historical rain gear purchase records (such as the user's purchase of an umbrella 3 months ago).

[0133] Real-time interest capture: Flink window counts users searching for "umbrella" 3 times within 30 minutes and generates short-term tag "heavy rain - emergency needs".

[0134] 2. Optimized intelligent recommendation engine

[0135] Recall Phase: Content Recall: BERT extracts comment tags such as "waterproof" and "windproof" and matches them with umbrellas, raincoats, and rain boots. Context Recall: Based on LBS (Location-Based Services) geofencing, it recommends instant delivery products (such as 1-hour delivery rain gear) in the user's city.

[0136] Ranking phase: The Wide&Deep model input rainfall coefficient ×1.8 and sufficient inventory product weight ×2.0, improving the predicted umbrella CTR to 12%.

[0137] Rearrangement stage: Each screen displays three categories: rain gear, waterproof shoe covers, and dehumidifiers, with a visual difference of >50% (different colors / scene main images).

[0138] 3. Marketing fatigue control

[0139] Attenuation monitoring: If a user receives 5 rain gear notifications within 30 minutes but does not click on them, a fatigue flag is triggered.

[0140] Strategy shift: Reduce the exposure of rain gear, insert 30% related categories (such as waterproof backpacks), and rotate the content to "commuter style" and "children's style" raincoats.

[0141] 4. Supply Chain Emergency Response

[0142] Sales Forecast: Prophet predicts a 300% increase in rain gear sales in areas affected by heavy rain, and has revised its dynamic safety stock forecast upwards to 150%. Logistics Scheduling: Genetic algorithms plan delivery routes, avoiding flooded areas, and expedited orders are assigned to dedicated electric vehicles (delivery within 2 hours).

[0143] 5. Real-time terminal adaptation

[0144] A 1-bit model loads locally cached high-selling rain gear (such as folding umbrellas), and federated learning encrypts and uploads user click behavior. A "Must-Have Rainstorm Emergency" special page is displayed in weak network environments (cache update cycle is 1 hour).

[0145] 6. Value-based tiered operation

[0146] VIP users: Users with a Value > 90 will be offered a premium automatic umbrella (priced at 299 yuan), with dedicated customer service providing next-day delivery guarantee. Price-sensitive users: Users with a PSI > 0.5 will be recommended a disposable raincoat for 9.9 yuan, with a discount coupon available.

[0147] Results: Rain gear conversion rate increased by 60%, regional stockout rate decreased by 70%, and delivery delay rate decreased by 45%.

[0148] Example 3: In-depth operation of high-value users

[0149] Scenario Description: Identify high-net-worth users and push luxury goods recommendations, providing exclusive services and priority fulfillment to enhance customer lifetime value.

[0150] Detailed implementation process:

[0151] 1. Segmentation of user profiles across the entire domain

[0152] RFM calculation: The user purchased luxury goods 3 times in the past month (M = 15,000 yuan), R = 7 days (most recent purchase), F = 3 times / month, Value = 0.4×0.9 + 0.3×0.8 + 0.3×0.95 = 89 points (high value).

[0153] Interest modeling: LSTM predicts that users' interest cycle for luxury goods is 2 months, and labels preferences as "handmade leather goods" and "limited editions".

[0154] 2. Customized Intelligent Recommendation Engine

[0155] Recall Phase: Collaborative Filtering: Aggregating the top 10 luxury goods (such as Hermès handbags) purchased by similar VIP users. Content Recall: BERT extracts tags such as "handmade" and "rare leather" and matches them with product descriptions.

[0156] Sorting phase: Wide&Deep model weights high-margin products (profit margin > 40%) by 1.5, prioritizing the display of limited edition items.

[0157] Rearrangement Phase: Fatigue Filtering: VIP users are exempt from regular frequency control, but the daily exposure limit is 20 times. Diversity Control: Each screen mixes luxury goods, customized services, and high-end experience activities.

[0158] 3. Precise supply chain assurance

[0159] Inventory forecast: Prophet forecasts monthly sales of 50 units for a limited-edition handbag, sets a safety stock of 60 units, and locks in inventory during the pre-sale phase.

[0160] Priority fulfillment: After a user places an order, a dedicated vehicle will be automatically assigned for delivery (delivery within 6 hours), and the logistics tracker will be synchronized to the APP in real time.

[0161] 4. Premium Terminal Experience

[0162] 1-bit model loading for 3D product preview (high-end device exclusive feature); displaying locally cached brand story videos in weak network conditions. Federated learning protects browsing history, uploading only gradient parameters (differential privacy noise σ = 0.1).

[0163] 5. Budget and Value Strategy

[0164] Consumption Forecast: Linear regression predicts a user's monthly spending cap of 20,000 yuan, and new products (such as a watch priced at 12,000 yuan) are recommended in the first week after payday. Churn Warning: If the user's purchase interval is extended from 15 days to 30 days, an exclusive gift package (free maintenance service) is triggered.

[0165] Results: High-value user retention rate increased by 40%, luxury goods GMV increased by 55%, and fulfillment timeliness of exclusive orders improved by 60%.

[0166] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. An e-commerce marketing and promotion system based on behavioral analysis; characterized in that: include: The full-domain user profile module is used to integrate user behavior data from multiple channels, construct a heterogeneous graph of users, products, and scenarios through a graph neural network model, and construct user interest profiles through LSTM. The intelligent recommendation engine is used to parse product review tags based on the BERT model, extract visual style features through the ResNet-50 model, and combine manufacturer data to build a multi-strategy fusion recommendation system. The marketing fatigue management module is used to predict the downward trend of user click-through rate through an exponential decay model and dynamically adjust the push strategy. The supply chain collaboration module is used to predict sales for the next 7 days using the Prophet algorithm and combine it with a dynamic safety stock formula to achieve intelligent replenishment. The lightweight terminal module is used to compress DNN models using 1-bit quantization technology, enable local processing of user data through federated learning, and dynamically adapt to device performance. The budget and value management module is used to analyze users' price sensitivity based on historical orders, predict monthly spending budgets, and adjust recommendation strategies in conjunction with payday cycles.

2. The e-commerce marketing and promotion system based on behavior analysis according to claim 1, characterized in that: The comprehensive user profiling module integrates user behavior data from multiple channels, constructs a heterogeneous graph of users, products, and scenarios using a graph neural network model, and builds user interest profiles using LSTM. Specifically, this includes: S11: Data collection and synchronization. The system uses a data tracking SDK to capture user behavior events on the APP and Web in real time and synchronizes offline POS transaction data to the data center. S12: Identity fusion, generating a unique identifier based on the device fingerprint, associating the unique identifier with the user account when the user logs in, and merging historical behavior data under the user account; S13: Short-term interest capture, using Flink sliding window to count high-frequency behaviors in the past hour and update user interest tags in real time; S14: Long-term interest modeling. Historical behavior is analyzed using XGBoost, preference weights are calculated, and the interest decay cycle is predicted using an LSTM model to obtain the user's long-term interest model. S15: Feature fusion and compression, which concatenates the obtained user interest tags and long-term interest model into a 512-dimensional original vector, and then uses an encoder to reduce the dimensionality of the 512-dimensional original vector into a 64-dimensional dense feature vector.

3. The e-commerce marketing and promotion system based on behavior analysis according to claim 2, characterized in that: When an intelligent recommendation engine parses product review tags based on the BERT model, extracts visual style features using the ResNet-50 model, and combines manufacturer data to build a multi-strategy fusion recommendation system, it specifically includes: S21: Data input, input 64-dimensional dense feature vector, product database data, geographic location data, weather data and time data; S22: During the recall phase, 10a candidate products are selected from the product database data through collaborative filtering, content recall, and best-selling product supplementation. S23: In the sorting stage, based on the Wide&Deep model, user characteristics, product attributes, geographic location data, weather data, and time data are integrated to predict the product click-through rate and dynamically weight it to quantify the intensity of user interest in each product. Combined with business rules, the sorting priority is optimized to select a selected products from 10a candidate products. S24: In the rearrangement stage, products are sorted through fatigue filtering and diversity control, while manual rules are injected to generate the final product list. Fatigue filtering removes products that have been exposed to but not clicked by users in the past period, and diversity control controls the diversity of product categories, price distribution and visual differences displayed on each screen.

4. The e-commerce marketing and promotion system based on behavior analysis according to claim 3, characterized in that: When selecting 10n candidate products from the product database through collaborative filtering, content recall, and best-selling product supplementation, the principles and steps of collaborative filtering include: S31: Similarity calculation, calculates the cosine of the angle between the 64-dimensional dense feature vectors of users; S32: Similar user filtering uses an approximate nearest neighbor algorithm to accelerate calculations and quickly find the Top 100 similar users from tens of millions of users. It also uses a fixed period to fully update the list of similar users and incrementally updates active user data every hour. S33: Candidate product aggregation, statistics on products purchased by similar users, filtering products with low ratings and low exposure conversion rates, sorting by purchase frequency and rating, and selecting the Top 3a products to enter the candidate pool.

5. The e-commerce marketing and promotion system based on behavior analysis according to claim 4, characterized in that: When selecting 10n candidate products from the product database through collaborative filtering, content recall, and best-selling product supplementation, the principle and steps of content recall include: S41: User interest tag extraction, obtain explicit tags from the 64-dimensional vector of the user profile module, and assign weights to the explicit tags according to the intensity of the behavior; S42: Product content analysis. Use a pre-trained BERT model to encode product reviews, extract keywords, calculate keyword importance, filter common words, and retain differentiated feature words. In the process of extracting keywords, use a graph neural network to construct a product-keyword relationship graph to capture implicit associations. S43: Matching and filtering. Calculate the similarity between user interest tag vectors and product keyword vectors using cosine similarity, and filter products based on a set threshold.

6. The e-commerce marketing and promotion system based on behavior analysis according to claim 5, characterized in that: The marketing fatigue management module, when predicting the decline in user click-through rates using an exponential decay model and dynamically adjusting push strategies, specifically includes: S51: Decay model modeling, using the decay exponential model to describe the decay law of click rate with the number of exposures; S52: Fatigue assessment, calculates user fatigue score in real time, and determines whether the user has entered a fatigue state based on the set threshold; S53: Frequency control strategy adjustment, dynamically extend the exposure interval and switch the recommendation strategy based on the user fatigue score, reduce the frequency of pop-up windows for highly fatigued users, and introduce cross-category products; S54: Creative content updates, using material carousel and intelligent cropping technology to update creative content, increasing the visual difference of the first screen; S55: Strategy activation output, taking into account fatigue score, frequency control rules and content update results, to generate the final push strategy.

7. The e-commerce marketing and promotion system based on behavior analysis according to claim 6, characterized in that: When the supply chain collaboration module predicts sales for the next 7 days using the Prophet algorithm and combines this with a dynamic safety stock formula to achieve intelligent replenishment, it specifically includes: S61: Demand Forecasting. By analyzing historical sales, seasonality, and promotional plans using the Prophet model, the demand for goods in the next 7 days is predicted. Safety stock is calculated using a dynamic safety stock formula, and an alert is triggered when the inventory is 120% lower than the predicted sales. S62: Allocation decision, real-time monitoring of inventory in each warehouse, when the inventory level is lower than the safety stock threshold, a replenishment task is generated, based on real-time inventory level and geographical distance, priority is given to transferring goods from the nearest regional warehouse, and the replenishment quantity is optimized by combining the cost model; S63: Logistics optimization, integrating real-time traffic API to dynamically plan delivery routes, and using genetic algorithms to solve for the optimal routes for multiple vehicles.

8. The e-commerce marketing and promotion system based on behavior analysis according to claim 7, characterized in that: The lightweight terminal module compresses the DNN model using 1-bit quantization technology, achieves local processing of user data through federated learning, and dynamically adapts to device performance. Specifically, this includes: S71: Model compression, which binarizes 32-bit floating-point weights during training using a pass-through estimator and accelerates matrix operations during inference using a binary mask; S72: Local inference, loads a 4MB lightweight model to the terminal device for real-time inference, and switches to local cache in weak network environments; S73: Federated learning updates the model gradient locally on the terminal using user behavior data, adds Gaussian noise to protect privacy, encrypts and uploads the gradient to the cloud for aggregation every 24 hours, and distributes it to the terminal after the global model is updated. S74: Output results. The terminal returns a recommendation list in real time. In the event of a weak network, offline service is provided by relying on caching. At the same time, model iteration is achieved through federated learning.

9. The e-commerce marketing and promotion system based on behavior analysis according to claim 8, characterized in that: The budget and value management module, when analyzing user price sensitivity based on historical orders, predicting monthly spending budgets, and adjusting recommendation strategies in conjunction with payday cycles, specifically includes: S81: Price range analysis. By statistically analyzing the price distribution of users' historical orders and calculating the price sensitivity index, it identifies price-sensitive users and stable users, and dynamically filters the price range of recommended products. S82: Budget forecasting, based on a linear regression model to predict users' monthly spending limit, and dynamically adjusts the recommendation strategy in conjunction with the pay cycle; S83: Value Operations. The RFM model is used to calculate user value scores, and users are divided into high-value and medium-low-value levels. Dedicated customer service and priority delivery services are launched for high-value users, and coupon recall strategies are triggered for low-value users. S84: Output and Execution. It integrates price bands, budget forecasts, and value stratification results to generate dynamic price filtering conditions and operational actions.

10. The e-commerce marketing and promotion system based on behavior analysis according to claim 9, characterized in that: The full-domain user profiling module, intelligent recommendation engine, marketing fatigue management module, supply chain collaboration module, lightweight terminal module, and budget and value management module achieve synergy through the following steps: S91: Data collection and integration, integrating user-end embedded data, business system orders, inventory and external data, and cleaning up invalid exposures and fixing device ID conflicts through Kafka real-time pipeline; S92: User profile construction, real-time capture of short-term interests, offline modeling of long-term features, and compression of 512-dimensional features to 64-dimensional features to dynamically generate lightweight user vectors; S93: Intelligent recommendation generation, recall layer integrates collaborative filtering, content matching and best-selling product backup, ranking layer uses Wide&Deep model to predict click rate, and re-ranking layer filters fatigued exposure and controls category diversity. S94: Marketing fatigue control, dynamically extending the exposure interval based on the exponential decay model, and using material carousel and intelligent cropping; S95: Supply chain linkage, using the Prophet model to predict sales, dynamically allocate resources and optimize logistics; S96: Terminal display and feedback, adapting multi-terminal performance through 1-bit quantization model, locally updating gradients through federated learning, and loading cached recommendations in weak network conditions.

Images