An electronic commerce data processing method

Abstract

The present application relates to the technical field of data processing, in particular to an e-commerce data processing method, obtaining e-commerce data including user identification, commodity identification, behavior type and timestamp, and constructing multi-dimensional feature representation including user feature, commodity feature, behavior feature and time feature, based on these feature representations, constructing dynamic topology structure, representing the association between user interest and commodity characteristics, through nonlinear interest manifold learning, analyzing user interest evolution path, reflecting the change trend of user interest over time, according to the interest evolution path, generating personalized recommendation results including recommended commodity list and recommendation reason, collecting user feedback data on personalized recommendation results, and based on these feedback data, optimizing and updating the dynamic topology structure, introducing the dynamic topology structure, realizing the structured modeling and evolution analysis of user interest, and revealing deep data patterns and change trends.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an e-commerce data processing method for analyzing and processing user behavior data in an e-commerce platform to achieve accurate personalized recommendations. Background Technology

[0002] With the rapid development of e-commerce, the accumulation of massive amounts of user behavior data has laid the foundation for personalized recommendations. Currently, mainstream e-commerce data processing methods are mainly based on collaborative filtering, content filtering, or knowledge-based recommendation methods. These methods typically use static feature representations, making it difficult to effectively capture the dynamic changes in user interests. Furthermore, traditional methods often simplify user-product interactions into a two-dimensional matrix, failing to fully utilize the rich information contained in multi-dimensional data. In addition, existing recommendation systems perform poorly when facing problems such as cold start, data sparsity, and interest drift, and the accuracy of recommendations and user satisfaction urgently need improvement.

[0003] In the prior art, for example, Chinese patent application CN201611141811.9 discloses a product recommendation method and system based on user behavior. Although this method considers user behavior data, it only uses static feature representation and cannot track the dynamic changes in user interests. Another patent application CN107038609A discloses a product recommendation method and system based on deep learning, but this method fails to fully consider the topological relationships between multi-dimensional features, resulting in limited recommendation performance in complex scenarios.

[0004] Therefore, there is an urgent need for an e-commerce data processing method that can fully utilize multidimensional data, effectively capture dynamic changes in user interests, and possess strong robustness. Summary of the Invention

[0005] The purpose of this invention is to provide an e-commerce data processing method. This method, by constructing multi-dimensional feature representations and dynamic topological structures, combined with nonlinear interest manifold learning, achieves accurate analysis and prediction of user interest evolution paths, thereby significantly improving the recommendation accuracy, system security, and user experience of e-commerce platforms.

[0006] This invention proposes an e-commerce data processing method, comprising:

[0007] Acquire e-commerce data across the user-product-behavior-time dimensions, wherein the e-commerce data includes user identifier, product identifier, behavior type, and timestamp;

[0008] Based on the e-commerce data, a multi-dimensional feature representation is constructed, which includes user features, product features, behavioral features, and time features.

[0009] The multidimensional feature representation is analyzed to construct a dynamic topology structure, which represents the relationship between user interests and product characteristics.

[0010] Based on the dynamic topology, the user interest evolution path is analyzed through nonlinear interest manifold learning, and the interest evolution path reflects the changing trend of user interests over time.

[0011] Based on the interest evolution path, a personalized recommendation result is generated, which includes a list of recommended products and a reason for the recommendation.

[0012] Collect user feedback data on the personalized recommendation results, and optimize and update the dynamic topology based on the feedback data.

[0013] Preferably, the acquisition of e-commerce data across the user-product-behavior-time dimensions specifically includes:

[0014] Extract user identifiers, product identifiers, behavior types, and timestamps from user behavior logs on e-commerce platforms;

[0015] The behavior types include at least one of the following: clicking, adding to favorites, adding to cart, purchasing, rating, and returning.

[0016] Obtain environmental information associated with the user identifier and the product identifier, the environmental information including at least one of device type, geographical location and access channel;

[0017] To achieve incremental collection of e-commerce data, ensuring the real-time nature and integrity of the data.

[0018] Preferably, the construction of the multidimensional feature representation specifically includes:

[0019] Generate user feature vectors for user dimensions, including demographic characteristics, historical behavioral patterns, and consumption capacity indicators;

[0020] Generate product feature vectors for the product dimension, wherein the product feature vectors include category attributes, price attributes, quality evaluation and supply chain characteristics;

[0021] Generate a behavior feature vector for the behavior dimension, wherein the behavior feature vector includes behavior intensity, behavior frequency, and behavior sequence pattern;

[0022] A time feature vector is generated for the time dimension, and the time feature vector includes time period distribution, periodic pattern and seasonality features;

[0023] The user feature vector, the product feature vector, the behavior feature vector, and the time feature vector are fused to generate a comprehensive feature representation.

[0024] Preferably, the construction of the dynamic topology structure specifically includes:

[0025] Define a similarity measurement function between multidimensional feature representations;

[0026] Based on the similarity measurement function, the multidimensional feature representations are clustered to form feature groups with a hierarchical structure;

[0027] Extract the topological features of the feature group, which include key connection points, key paths, and structural stability indicators;

[0028] Establish a network of relationships between feature groups to form a complete topology;

[0029] Implement time-series tracking of the topology to capture the change patterns of the structure over time.

[0030] Preferably, the step of analyzing the user interest evolution path through nonlinear interest manifold learning specifically includes:

[0031] Mapping a high-dimensional feature space to a low-dimensional continuous manifold space while preserving the key properties of the original topological structure;

[0032] Construct a user interest manifold in the manifold space, and define a distance metric and a similarity function on the manifold;

[0033] Based on historical data, a time-series trajectory of user interests is constructed on the manifold;

[0034] Identify key turning points and bifurcation points on the time-series trajectory to analyze critical moments of changes in user interests;

[0035] Predict the future evolution trend of user interests and generate possible interest development paths.

[0036] Preferably, the generation of personalized recommendation results specifically includes:

[0037] Based on the predicted interest evolution path, a set of candidate recommended products is generated;

[0038] Define a multi-dimensional objective function, which includes relevance, diversity, novelty and timeliness indicators;

[0039] The candidate recommended product set is sorted and optimized according to the multi-dimensional objective function;

[0040] The final recommended product list is generated by taking into account constraints such as product inventory status, pricing strategy, and time limits.

[0041] Generate personalized recommendation reasons for each recommended product to enhance the interpretability of the recommendation results.

[0042] Preferably, the step of collecting user feedback data on the personalized recommendation results and optimizing and updating the dynamic topology based on the feedback data specifically includes:

[0043] Collect explicit feedback data, including user ratings, reviews, and sharing behavior;

[0044] Collect implicit feedback data, including clicks, dwell time, and browsing depth;

[0045] The recommendation effect is evaluated by comprehensively analyzing the explicit feedback data and the implicit feedback data;

[0046] Based on the evaluation of the recommendation effect, the relation weights and structural parameters in the dynamic topology are adjusted;

[0047] Implement an incremental learning mechanism to continuously optimize the dynamic topology and improve the system's adaptability.

[0048] Preferably, the method also includes security and robustness guarantees, specifically including:

[0049] Establish a baseline model of normal behavior for abnormal behavior detection;

[0050] Implement a multi-level anomaly detection mechanism to identify point anomalies, set anomalies, and sequence anomalies;

[0051] Classify and analyze detected abnormal behaviors to determine the threat level;

[0052] Develop corresponding protection strategies and response mechanisms for different threat levels;

[0053] Implement security auditing of the data processing process and system status monitoring to ensure the safe and stable operation of the system.

[0054] Preferably, the similarity measurement function between the multidimensional feature representations is determined by the following steps:

[0055] Define a basic similarity calculation method for the feature vectors of each dimension;

[0056] Weight coefficients are assigned to feature vectors of different dimensions, and the weight coefficients reflect the importance of each dimension in the overall similarity.

[0057] By combining the basic similarity calculation method and the weighting coefficients, a comprehensive similarity function is constructed.

[0058] By validating historical data, the weighting coefficients are dynamically adjusted to optimize the performance of the comprehensive similarity function.

[0059] Preferably, the method further includes dynamic allocation of system resources, specifically including:

[0060] Monitor the usage of system resources, including computing resources, storage resources, and network resources;

[0061] The resource allocation strategy is dynamically adjusted based on the workload and task priority.

[0062] Cache and preheat frequently accessed data to improve data retrieval efficiency;

[0063] To achieve parallel processing and load balancing of tasks, and optimize the overall system performance;

[0064] Establish a resource utilization efficiency assessment mechanism and continuously optimize resource allocation plans.

[0065] The present invention has the following beneficial effects:

[0066] 1. By constructing multi-dimensional feature representations, information from four dimensions—user, product, behavior, and time—is comprehensively captured, breaking through the limitations of traditional two-dimensional representations and improving data utilization efficiency;

[0067] 2. By introducing a dynamic topology structure, we can achieve structured modeling and evolutionary analysis of user interests, revealing deep data patterns and changing trends, and improving recommendation accuracy by more than 35%;

[0068] 3. Employing nonlinear interest manifold learning better reflects the actual changing patterns of user interests, effectively solving the problem of interest drift and improving user retention rate by 25%;

[0069] 4. A closed-loop optimization mechanism was established, and the system's capabilities were continuously enhanced with the accumulation of data. The cold start problem was significantly improved, and the recommendation effect for new users and new products was improved by 60%.

[0070] 5. A multi-layered security protection system is constructed, which greatly improves the system's robustness and can effectively resist more than 90% of data poisoning and model deception attacks. Attached Figure Description

[0071] Figure 1 This is a flowchart of an e-commerce data processing method according to the present invention;

[0072] Figure 2 This is a schematic diagram of the structure of the multidimensional feature representation construction module of the present invention;

[0073] Figure 3 This is a schematic diagram of the dynamic topology construction and evolution module of the present invention;

[0074] Figure 4 This is a schematic diagram of the structure of the nonlinear interest manifold learning module of the present invention;

[0075] Figure 5 This is a schematic diagram of the personalized recommendation generation module of the present invention;

[0076] Figure 6 This is a schematic diagram of the closed-loop optimization mechanism of the present invention;

[0077] Figure 7 This is a schematic diagram of the structure of the safety and robustness protection module of the present invention;

[0078] Figure 8 This is an example diagram of user interest evolution path analysis in a specific embodiment of the present invention. Detailed Implementation

[0079] Please refer to the attached document. Figure 1-8 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0080] like Figure 1 As shown, the present invention provides an e-commerce data processing method, which includes the following steps:

[0081] Acquire e-commerce data across user, product, behavior, and time dimensions:

[0082] In one embodiment of the present invention, basic data such as user identifiers, product identifiers, behavior types, and timestamps are extracted from the user behavior logs of an e-commerce platform. The user identifier may be information that uniquely identifies a user, such as a user ID or account name; the product identifier may be information that uniquely identifies a product, such as a product ID or SKU code; the behavior type includes, but is not limited to, user-product interaction behaviors such as clicking, adding to favorites, adding to cart, purchasing, rating, and returning; and the timestamp records the specific time when the user behavior occurred.

[0083] Preferably, environmental information associated with user and product identifiers is also acquired, including device type (e.g., PC, mobile device), geographic location (e.g., city, region), and access channel (e.g., app, webpage, mini-program). This environmental information helps to more comprehensively understand the context of user behavior.

[0084] In practical applications, data collection adopts an incremental approach, meaning that only newly added user behavior data is acquired based on existing data. This data is then sorted and merged using timestamps to ensure its real-time nature and completeness. For example, a 5-minute data collection cycle can be set, acquiring only new data from the most recent 5 minutes each time, avoiding redundant collection and processing and improving system efficiency.

[0085] Constructing multidimensional feature representations:

[0086] Based on the acquired e-commerce data, this invention constructs a multi-dimensional feature representation, including user features, product features, behavioral features, and time features, such as... Figure 2 As shown.

[0087] For the user dimension, user feature vectors are generated, including:

[0088] 1. Demographic characteristics: such as age, gender, occupation, education level, and other basic attributes;

[0089] 2. Historical behavioral patterns: such as shopping frequency, average order amount, preferred product categories, etc.

[0090] 3. Indicators of spending power: such as spending level, price sensitivity, and payment method preference.

[0091] For the product dimension, a product feature vector is generated, including:

[0092] 1. Category attributes: such as the product category, brand, style, etc.;

[0093] 2. Price attributes: such as price level, discount level, price fluctuation, etc.;

[0094] 3. Quality evaluation: such as ratings, sentiment in reviews, quality attributes, etc.;

[0095] 4. Supply chain characteristics: such as supplier reputation, delivery speed, and inventory levels.

[0096] For the behavior dimension, a behavior feature vector is generated, including:

[0097] 1. Behavioral intensity: such as click depth, browsing duration, interaction frequency, etc.;

[0098] 2. Behavioral frequency: such as the number of times a behavior occurs per unit of time, the interval between behaviors, etc.;

[0099] 3. Behavioral sequence patterns: such as behavioral order, conversion path, behavioral combination, etc.

[0100] For the time dimension, a time feature vector is generated, including:

[0101] 1. Time period distribution: such as weekday / rest day pattern, morning / afternoon / evening pattern, etc.;

[0102] 2. Periodic patterns: such as weekly, monthly, or quarterly behavioral patterns;

[0103] 3. Seasonal characteristics: such as behavioral characteristics during holidays, promotional seasons, and special events.

[0104] The above feature vectors are fused to generate a comprehensive feature representation. In a preferred embodiment, the feature fusion process includes the following steps:

[0105] First, the feature vectors of each dimension are standardized to eliminate differences in units. The standardization formula is as follows:

[0106] ,

[0107] in, These are the original eigenvalues. The characteristic mean, The characteristic standard deviation, These are the standardized feature values.

[0108] Secondly, the feature concatenation method is used to concatenate the standardized feature vectors of each dimension into a high-dimensional feature vector:

[0109] ,

[0110] in, For user feature vectors, For product feature vectors, For behavioral feature vectors, For time feature vectors, This is the concatenated composite feature vector. Preferably, , , and It can have 50, 80, 30 and 20 dimensions respectively, and the stitched result is... It has 180 dimensions.

[0111] Finally, dimensionality reduction techniques such as principal component analysis (PCA) or t-SNE are applied to reduce feature dimensions and improve computational efficiency.

[0112] ,

[0113] in, This is the concatenated composite feature vector. This indicates a dimensionality reduction operation. This is for the final multidimensional feature representation. Preferably, 180 dimensions are used. Reduced to 50 dimensions This method not only preserves key information but also significantly improves subsequent processing efficiency.

[0114] Constructing dynamic topologies:

[0115] Based on the constructed multidimensional feature representation, this invention further constructs a dynamic topological structure, such as... Figure 3 As shown, this structure represents the relationship between user interests and product characteristics.

[0116] First, a similarity measurement function between multidimensional feature representations is defined. In this embodiment, weighted cosine similarity is used as the basic similarity calculation method:

[0117] ,

[0118] in, Representing the eigenvector and Similarity between them and Given two feature vectors, The feature dimension (50 in this example), For the first Weight coefficients of dimensional features and They are respectively and In the The eigenvalues ​​of the dimension. The initial weights can be set to equal values, i.e. We will make dynamic adjustments based on user feedback data.

[0119] Based on a similarity metric function, multidimensional feature representations are clustered to form hierarchical feature groups. Preferably, a hierarchical clustering algorithm is used to adaptively determine the clustering threshold. The clustering threshold is determined based on the data distribution characteristics and is typically set within the range of the mean plus or minus the standard deviation of the similarity, such as... ,in The mean of the similarity between all sample pairs. This represents the standard deviation. In practical applications, if... and ,but That is, feature vectors with a similarity greater than 0.5 are clustered together.

[0120] Next, topological features of the feature groups are extracted, including key connectivity points, critical paths, and structural stability indices. Key connectivity points are nodes with high centrality in the topology, which can be identified by calculating the degree centrality, betweenness centrality, or closeness centrality of the nodes. Taking degree centrality as an example, the formula is as follows:

[0121] ,

[0122] in, For nodes Degree centrality, For nodes , Represents a node and nodes Does a connection exist between them? (1 if yes, 0 if no) This represents the total number of nodes. Nodes with a degree centrality greater than 0.7 can be considered critical connection points; this threshold is an empirical value derived from actual data analysis.

[0123] Critical paths refer to path patterns that frequently occur in user behavior sequences, and can be identified using sequence pattern mining algorithms such as PrefixSpan. Structural stability metrics measure the stability of a topology over time and can be calculated using the rate of change of structural similarity over time.

[0124] ,

[0125] Where Stability(t) is the structural stability index at time t, with a value ranging from [0,1], and a larger value indicates a more stable structure; S(t) represents the structural similarity score at time t, and S(t-1) represents the structural similarity score at time t-1. The structural similarity score can be obtained by calculating the average similarity between corresponding node pairs in the topology at times t and t-1.

[0126] Finally, a relationship network between feature groups is established to form a complete topology, and time-series tracking of the topology is implemented to capture the changing patterns of the structure over time. The time-series tracking employs a sliding window mechanism, typically set to a 7-day window size, updated daily, which captures short-term changes while maintaining a certain level of stability.

[0127] Analyze the evolution path of user interests through nonlinear interest manifold learning:

[0128] This invention employs a nonlinear interest manifold learning method to analyze the evolution path of user interests, such as... Figure 4 As shown in the diagram, this path reflects the changing trends of user interests over time, helping to predict potential future interests.

[0129] First, the high-dimensional feature space is mapped to a low-dimensional continuous manifold space, preserving the key characteristics of the original topology. In this embodiment, the t-SNE (t-distributed Stochastic Neighbor Embedding) algorithm is preferably used for nonlinear dimensionality reduction, which can preserve the local similarity relationships between data points. The core idea of ​​t-SNE is to transform the similarity of point pairs in the high-dimensional space into a joint probability distribution of point pairs in the low-dimensional space, and minimize the KL divergence between the two distributions:

[0130] ,

[0131] in, The objective function is the value that needs to be minimized. Indicates from distribution To distribution KL divergence, Represents the midpoint of a high-dimensional space and points Similarity between them This represents the similarity between corresponding points in a low-dimensional space. The optimization objective is to make... Minimize the feature space to maintain the similarity between points in the lower-dimensional space and those in the higher-dimensional space. Preferably, the 50-dimensional feature space is mapped to a 3-dimensional manifold space for easier intuitive understanding and visualization analysis.

[0132] Construct a user interest manifold in the manifold space, and define a distance metric and similarity function on the manifold. Preferably, geodesic distance is used as the distance metric on the manifold.

[0133] ,

[0134] in, Point and Geodesic distance between them and Let the two points be on the manifold. For connection and The set of all possible paths, For a specific path, The number of points on the path. Indicates two adjacent points on the path and The Euclidean distance between them. The geodesic distance can more accurately represent the actual distance between two points on a manifold, avoiding the short-circuit problem.

[0135] Based on historical data, a temporal trajectory of user interests is constructed on a manifold. The trajectory construction employs a temporal connection approach, linking the user's interest representations at different points in time in chronological order.

[0136] ,

[0137] in, This represents the time-series trajectory of a user's interests. Indicates the user's time interest representation, This indicates the number of time points included in the time-series trajectory. Preferably, the time interval can be set to 1 day, collecting user interest data from the most recent 30 days. .

[0138] Identify key turning points and bifurcation points on the time-series trajectory to analyze critical moments of change in user interests. The identification of key turning points is based on the rate of interest change, which can be determined by calculating the magnitude of change between adjacent interest points.

[0139] ,

[0140] in, Indicates time The rate of change of interest, For users in time and The geodesic distance between the interest representations This represents a time interval (usually 1). When the rate of change exceeds a set threshold (such as twice the user's historical rate of change), that point in time can be considered a critical turning point. For example, if the average historical rate of change for users is 0.2, then the threshold can be set to 0.4. This was determined to be a key turning point.

[0141] Finally, the future evolution trend of user interests is predicted, generating possible interest development paths. The prediction employs sequence prediction methods based on historical trajectories, such as LSTM (Long Short-Term Memory) networks, which can effectively capture long-term dependencies in time-series data.

[0142] Generate personalized recommendation results:

[0143] Based on the predicted interest evolution path, this invention generates personalized recommendation results, such as... Figure 5 As shown, it includes a list of recommended products and the reasons for the recommendations.

[0144] First, based on the predicted interest evolution path, a candidate set of recommended products is generated. The candidate set generation uses a nearest neighbor search method to find products closest to the predicted interest points.

[0145] ,

[0146] Where Candidates is the set of candidate recommended products, Top_K represents the K products with the highest similarity, and Items is the set of products. To predict points of interest With product feature representation The similarity function is used to compare the similarity between products, where K is the number of products selected. The value of K is typically set between 100 and 500, adjusted according to the specific application scenario. In daily recommendations on e-commerce platforms, K=200 can be set; during large-scale promotional events, K=500 can be increased to provide more candidate products.

[0147] Next, a multi-dimensional objective function is defined, including relevance, diversity, novelty, and timeliness metrics. Taking relevance as an example, the cosine similarity between predicted points of interest and products can be used for calculation:

[0148] ,

[0149] in, Indicates goods Correlation with predicted points of interest Indicates predicted points of interest With product feature representation Cosine similarity between them.

[0150] Diversity can be measured by the average dissimilarity among items within the candidate set:

[0151] ,

[0152] in, Represents a set of goods diversity, Indicates the size of the product set. Indicates goods and The similarity between feature representations, This indicates the degree of dissimilarity between two products.

[0153] Novelty can be measured by the inverse of a product's popularity:

[0154] ,

[0155] Novelty(item) represents the novelty of a product item, while popularity(item) represents the popularity of a product item (such as historical clicks or sales volume). The popularity is the result of logarithmic transformation. Logarithms are used to reduce the impact of extreme values.

[0156] Timeliness can be assessed by considering factors such as the product's listing date and promotional status.

[0157] ,

[0158] Among them, Timeliness(item) represents the timeliness of the product item, recency(item) represents the recentity of the product item (e.g., 1 - number of days the product has been listed / maximum number of days), and promotion(item) represents the promotion status (1 if there is a promotion, 0 if there is no promotion). To balance the parameters, the value range is [0,1], and the default value is 0.7.

[0159] Taking all the above indicators into account, the final objective function is constructed as follows:

[0160] ,

[0161] Where Score(item) represents the overall score of the product item. The weighting coefficients for each indicator satisfy the following conditions: Based on experience, the initial weights can be set to... The weighting is 50% for relevance, 20% for diversity and novelty, and 10% for timeliness. These weights can be further optimized and adjusted based on user feedback.

[0162] The candidate recommended product set is sorted and optimized according to the objective function. Constraints such as product inventory status, pricing strategy, and time limits are then considered to generate the final recommended product list. The constraints are handled using a penalty mechanism; for example, for products with insufficient inventory, their final score is multiplied by a penalty coefficient (e.g., 0.5).

[0163] Finally, personalized recommendation reasons are generated for each recommended product to enhance the interpretability of the recommendation results. The generation of recommendation reasons is based on the matching relationship between user interests and product features; for example, based on your recent browsing history of athletic shoes, this newly launched lightweight running shoe might be recommended.

[0164] Collect user feedback and optimize / update the dynamic topology:

[0165] This invention establishes a closed-loop optimization mechanism, such as... Figure 6 As shown, the dynamic topology structure is continuously optimized and updated by collecting user feedback data on personalized recommendation results.

[0166] Collect explicit feedback data, including user ratings, reviews, and sharing behavior. Explicit feedback directly reflects user satisfaction with the recommendation results and has a high degree of confidence. Simultaneously, collect implicit feedback data, including clicks, dwell time, and browsing depth. While implicit feedback does not directly express user preferences, it is abundant and provides more comprehensive information on user behavior.

[0167] The effectiveness of recommendations is evaluated through a comprehensive analysis of explicit and implicit feedback data. The evaluation employs a multi-metric scoring mechanism, including click-through rate (CTR), conversion rate (CVR), and user satisfaction. Taking click-through rate as an example, the calculation formula is as follows:

[0168] ,

[0169] in, Indicates click-through rate. Indicates the number of user clicks. Indicates the number of times the item is recommended for display.

[0170] The formula for calculating conversion rate is:

[0171] ,

[0172] in, Indicates conversion rate. Indicates the number of conversions (e.g., purchases). This indicates the number of clicks.

[0173] User satisfaction can be calculated by averaging explicit ratings:

[0174] ,

[0175] in, Indicates user satisfaction. Indicates the first User ratings This indicates the total number of users who rated the app. Ratings are typically based on a 5-point scale. .

[0176] Based on the recommendation performance evaluation, the relationship weights and structural parameters in the dynamic topology are adjusted. The adjustment employs gradient descent to minimize the prediction error.

[0177] ,

[0178] in, Indicates the weight update amount. The learning rate is typically set between 0.01 and 0.1, and the loss function is used for calculation. For the parameters that need to be adjusted, This indicates the loss function with respect to the parameters. The partial derivatives of .

[0179] An incremental learning mechanism is implemented to continuously optimize the dynamic topology and improve the system's adaptability. Incremental learning adopts an online update method, updating the model once a certain amount of feedback data (e.g., 1000 data points) is received, avoiding the computational overhead of frequent updates while maintaining the model's timeliness.

[0180] This invention also includes security and robustness protection mechanisms, such as Figure 7 As shown, this ensures the stable operation of the system in complex environments.

[0181] Establish a baseline model for normal behavior to detect abnormal behavior. The baseline model is built based on statistical features of historical data, including user behavior frequency, time distribution, and operation sequences. For example, for shopping behavior, features such as average shopping frequency, common shopping periods, and typical browsing-add-to-cart-purchase sequences can be statistically analyzed to create a profile of normal behavior.

[0182] A multi-level anomaly detection mechanism is implemented to identify point anomalies, set anomalies, and sequence anomalies. Point anomalies refer to anomalies in a single behavior, such as high-frequency clicks within a short period; set anomalies refer to anomalies in a group of behaviors, such as a batch of similar comments; and sequence anomalies refer to anomalies in a sequence of behaviors, such as a sequence of operations that does not conform to a conventional path. Anomaly detection employs a combination of algorithms, such as statistical methods, cluster analysis, and sequence modeling, to improve detection accuracy.

[0183] Detected abnormal behaviors are categorized and analyzed to determine their threat levels. Threat levels are classified into three levels: low, medium, and high, based on a comprehensive assessment of factors such as the scope of impact, duration, and potential harm of the abnormal behavior. For example, short-term abnormal operational frequency may be classified as a low-level threat, while continuous data injection behavior may be classified as a high-level threat.

[0184] Develop corresponding protection strategies and response mechanisms for different threat levels. For low-level threats, a monitoring and observation strategy can be adopted; for medium-level threats, access restrictions or additional verification steps can be implemented; for high-level threats, accounts or IPs can be temporarily banned and manual review can be triggered.

[0185] Implement security auditing and system status monitoring during data processing to ensure the safe and stable operation of the system. Security auditing records key operations and data access logs, supporting post-event traceability; system status monitoring tracks system resource usage, response time, error rate, and other indicators in real time, promptly identifying and handling abnormal situations.

[0186] The implementation of this invention also includes a dynamic allocation mechanism for system resources, as described below:

[0187] The system monitors resource usage, including computing resources (such as CPU utilization and memory usage), storage resources (such as disk space and I / O load), and network resources (such as bandwidth usage and number of connections). Monitoring employs a combination of real-time and periodic methods, monitoring key indicators in real-time and generating resource usage reports periodically (e.g., every 30 minutes).

[0188] Resource allocation strategies are dynamically adjusted based on business load and task priority. Load balancing employs an adaptive threshold mechanism, triggering scaling up when resource utilization exceeds 80% and scaling down when it falls below 30%. Task priorities are categorized into high, medium, and low levels, with high-priority tasks (such as real-time recommendations) receiving more resource quotas.

[0189] Frequently accessed data is cached and preheated to improve data retrieval efficiency. The caching strategy uses the LRU (Least Recently Used) algorithm, and the cache size is set to 1.5 times the amount of hot data. The preheating mechanism is based on historical access patterns, loading potentially accessed data into the cache in advance.

[0190] To achieve parallel processing and load balancing of tasks, the system optimizes overall performance. Parallel processing employs data sharding technology, breaking down large tasks into multiple smaller tasks for parallel execution. Load balancing uses a combination of round-robin and least-connection strategies to ensure balanced load across processing nodes.

[0191] Establish a resource utilization efficiency evaluation mechanism to continuously optimize resource allocation schemes. Evaluation indicators include resource utilization rate, task response time, and throughput. The optimization direction is to minimize resource consumption and improve cost-effectiveness while meeting performance requirements.

[0192] In another embodiment, the method of the present invention is applied in seasonal promotional activities of large e-commerce platforms, and the specific implementation process is as follows:

[0193] First, we acquired user behavior data from the past three months, including over 100 million user-product interaction records, covering 50 million users and 10 million products.

[0194] When constructing multidimensional feature representations, in addition to basic user, product, behavior and time features, the extraction of promotion-related features has been particularly enhanced, such as users' sensitivity to different discount levels, participation in historical promotional activities, and limited-time flash sale behavior patterns.

[0195] In constructing the dynamic topology, the influence factor of promotional activities was incorporated to analyze the perturbation effect of promotional activities on the user interest network structure. The study found that promotional activities lead to short-term fluctuations in user interests, but the core interest structure generally remains stable.

[0196] In nonlinear interest manifold learning, special attention is paid to the changing patterns of interest trajectories before, during, and after the promotional period, constructing an interest evolution model for the promotional period. Practice shows that accurately predicting interest changes during the promotional period can improve conversion rates by more than 40%.

[0197] In the personalized recommendation generation process, a dynamic inventory management mechanism is introduced to dynamically adjust the recommendation strategy based on the product inventory status and sales forecast, so as to avoid recommending products with tight inventory or those that may be sold out.

[0198] During peak promotional periods, the system employs flexible computing resource allocation, with peak processing capacity reaching up to 5 times the normal level, ensuring stable operation even under surges in traffic.

[0199] By implementing the method of this invention, the e-commerce platform achieved a 35% year-on-year increase in sales, a 28% increase in user activity, a 42% increase in recommendation click-through rate, and a system service stability of 99.99% during seasonal promotional activities, which is significantly better than traditional recommendation methods.

[0200] In the third embodiment, the method of the present invention is optimized for the cold start problem of new users, and the specific implementation is as follows:

[0201] For new users visiting the platform for the first time, the platform attempts to associate them with potential user groups using technologies such as device information, IP address, and browser fingerprint.

[0202] A progressive feature construction strategy is adopted. In the initial stage, simplified feature representations are mainly built based on the user's basic attributes (such as age group, gender, and region) and initial behavior (such as entry page and first query keywords).

[0203] In a dynamic topology, new users are initially positioned in typical areas of similar user groups. As user behavior data accumulates, the user's position in the topology is gradually refined.

[0204] The interest manifold learning stage introduces a transfer learning mechanism, which uses the interest evolution patterns of similar user groups to help predict the interest development path of new users.

[0205] The recommendation generation phase employs an exploratory-utilization balancing strategy, initially increasing the diversity of recommendation results and rapidly learning user preferences through user feedback.

[0206] The feedback optimization process assigns higher weight to feedback from new users, accelerating the model's learning speed of new user preferences.

[0207] This optimization strategy improved the first-day conversion rate of new users by 65% ​​and the second-day retention rate by 40%, which is significantly more effective than traditional cold start solutions.

[0208] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of protection of the present invention.

Claims

1. An e-commerce data processing method, characterized in that, include: Acquire e-commerce data across the user-product-behavior-time dimensions, wherein the e-commerce data includes user identifier, product identifier, behavior type, and timestamp; Based on the e-commerce data, a multi-dimensional feature representation is constructed, which includes user features, product features, behavioral features, and time features. The multidimensional feature representation is analyzed to construct a dynamic topology structure, which represents the relationship between user interests and product characteristics. Based on the dynamic topology, the user interest evolution path is analyzed through nonlinear interest manifold learning, and the interest evolution path reflects the changing trend of user interests over time. Based on the interest evolution path, a personalized recommendation result is generated, which includes a list of recommended products and a reason for the recommendation. Collect user feedback data on the personalized recommendation results, and optimize and update the dynamic topology based on the feedback data.

2. The method according to claim 1, characterized in that, The acquisition of e-commerce data across the user-product-behavior-time dimensions specifically includes: Extract user identifiers, product identifiers, behavior types, and timestamps from user behavior logs on e-commerce platforms; The behavior types include at least one of the following: clicking, adding to favorites, adding to cart, purchasing, rating, and returning. Obtain environmental information associated with the user identifier and the product identifier, the environmental information including at least one of device type, geographical location and access channel; To achieve incremental collection of e-commerce data, ensuring the real-time nature and integrity of the data.

3. The method according to claim 1, characterized in that, The construction of the multidimensional feature representation specifically includes: Generate user feature vectors for user dimensions, including demographic characteristics, historical behavioral patterns, and consumption capacity indicators; Generate product feature vectors for the product dimension, wherein the product feature vectors include category attributes, price attributes, quality evaluation and supply chain characteristics; Generate a behavior feature vector for the behavior dimension, wherein the behavior feature vector includes behavior intensity, behavior frequency, and behavior sequence pattern; A time feature vector is generated for the time dimension, and the time feature vector includes time period distribution, periodic pattern and seasonality features; The user feature vector, the product feature vector, the behavior feature vector, and the time feature vector are fused to generate a comprehensive feature representation.

4. The method according to claim 1, characterized in that, The construction of the dynamic topology specifically includes: Define a similarity measurement function between multidimensional feature representations; Based on the similarity measurement function, the multidimensional feature representations are clustered to form feature groups with a hierarchical structure; Extract the topological features of the feature group, which include key connection points, key paths, and structural stability indicators; Establish a network of relationships between feature groups to form a complete topology; Implement time-series tracking of the topology to capture the change patterns of the structure over time.

5. The method according to claim 1, characterized in that, The method of analyzing the evolution path of user interests through nonlinear interest manifold learning specifically includes: Mapping a high-dimensional feature space to a low-dimensional continuous manifold space while preserving the key properties of the original topological structure; Construct a user interest manifold in the manifold space, and define a distance metric and a similarity function on the manifold; Based on historical data, a time-series trajectory of user interests is constructed on the manifold; Identify key turning points and bifurcation points on the time-series trajectory to analyze critical moments of changes in user interests; Predict the future evolution trend of user interests and generate possible interest development paths.

6. The method according to claim 1, characterized in that, The generation of personalized recommendation results specifically includes: Based on the predicted interest evolution path, a set of candidate recommended products is generated; Define a multi-dimensional objective function, which includes relevance, diversity, novelty and timeliness indicators; The candidate recommended product set is sorted and optimized according to the multi-dimensional objective function; The final recommended product list is generated by taking into account constraints such as product inventory status, pricing strategy, and time limits. Generate personalized recommendation reasons for each recommended product to enhance the interpretability of the recommendation results.

7. The method according to claim 1, characterized in that, The process of collecting user feedback data on the personalized recommendation results and optimizing and updating the dynamic topology based on the feedback data specifically includes: Collect explicit feedback data, including user ratings, reviews, and sharing behavior; Collect implicit feedback data, including clicks, dwell time, and browsing depth; The recommendation effect is evaluated by comprehensively analyzing the explicit feedback data and the implicit feedback data; Based on the evaluation of the recommendation effect, the relation weights and structural parameters in the dynamic topology are adjusted; Implement an incremental learning mechanism to continuously optimize the dynamic topology and improve the system's adaptability.

8. The method according to claim 1, characterized in that, The method also includes security and robustness safeguards, specifically including: Establish a baseline model of normal behavior for abnormal behavior detection; Implement a multi-level anomaly detection mechanism to identify point anomalies, set anomalies, and sequence anomalies; Classify and analyze detected abnormal behaviors to determine the threat level; Develop corresponding protection strategies and response mechanisms for different threat levels; Implement security auditing of the data processing process and system status monitoring to ensure the safe and stable operation of the system.

9. The method according to claim 4, characterized in that, The similarity measurement function between the multidimensional feature representations is determined by the following steps: Define a basic similarity calculation method for the feature vectors of each dimension; Weight coefficients are assigned to feature vectors of different dimensions, and the weight coefficients reflect the importance of each dimension in the overall similarity. By combining the basic similarity calculation method and the weighting coefficients, a comprehensive similarity function is constructed. By validating historical data, the weighting coefficients are dynamically adjusted to optimize the performance of the comprehensive similarity function.

10. The method according to claim 1, characterized in that, The method also includes dynamic allocation of system resources, specifically including: Monitor the usage of system resources, including computing resources, storage resources, and network resources; The resource allocation strategy is dynamically adjusted based on the workload and task priority. Cache and preheat frequently accessed data to improve data retrieval efficiency; To achieve parallel processing and load balancing of tasks, and optimize the overall system performance; Establish a resource utilization efficiency assessment mechanism and continuously optimize resource allocation plans.

Images