Data classification method and device for e-commerce platform, equipment and storage medium

Abstract

This application relates to the field of data processing technology, and discloses a data classification method, apparatus, device, and storage medium for e-commerce platforms. The method includes: processing user behavior feature matrices and product information data using a graph neural network to obtain a product knowledge graph; performing interaction pattern analysis on the user behavior feature matrix and product knowledge graph using a temporal attention mechanism to obtain a temporal interaction feature vector; performing feature fusion processing through a multimodal feature fusion network to obtain a fused feature vector; and performing classification processing on the fused feature vector through a hierarchical attention decision network to obtain an initial classification result and then performing data classification optimization processing to obtain the target classification data. This application, by introducing a temporal attention mechanism and a multimodal feature fusion network, combined with a hierarchical attention decision network and an interpretability rule processor, achieves deep fusion of user behavior and product features, improving the accuracy and interpretability of classification.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a data classification method, apparatus, device, and storage medium for e-commerce platforms. Background Technology

[0002] Currently, data classification methods on e-commerce platforms primarily rely on manual annotation and rule matching. Traditional classification methods typically employ keyword-based text matching, statistical collaborative filtering, and classification algorithms based on simple machine learning. These methods mainly focus on the static attribute features of products, such as product titles, descriptions, and categories. Meanwhile, some improved methods are beginning to incorporate user behavior data, analyzing user clicks, favorites, and purchases to assist in product classification, and are starting to explore the use of deep learning models to enhance classification performance.

[0003] However, existing e-commerce platform data classification methods have the following shortcomings: First, the use of user behavior data remains at a simple statistical level, failing to fully explore the temporal evolution characteristics contained in the user behavior sequence; second, existing methods often process product features and user behavior features separately, lacking an effective feature fusion mechanism; third, the classification results lack interpretability, making it difficult to understand the basis for classification decisions; and finally, existing methods fail to handle the hierarchical characteristics of the product category system well, resulting in low classification accuracy. Summary of the Invention

[0004] This application provides a data classification method, apparatus, device, and storage medium for e-commerce platforms. By introducing a temporal attention mechanism and a multimodal feature fusion network, combined with a hierarchical attention decision network and an interpretability rule processor, it achieves deep integration of user behavior and product features, thereby improving the accuracy and interpretability of classification.

[0005] Firstly, this application provides a data classification method for e-commerce platforms. The method includes: performing multi-dimensional feature extraction processing on user interaction behavior data to obtain a user behavior feature matrix; processing the user behavior feature matrix and product information data using a graph neural network to obtain a product knowledge graph; performing interaction pattern analysis processing on the user behavior feature matrix and the product knowledge graph using a temporal attention mechanism to obtain a temporal interaction feature vector; performing feature fusion processing on the user behavior feature matrix, the product knowledge graph, and the temporal interaction feature vector using a multimodal feature fusion network to obtain a fused feature vector; performing classification processing on the fused feature vector using a hierarchical attention decision network to obtain an initial classification result; and performing data classification optimization processing on the initial classification result using an interpretable rule processor to obtain target classification data.

[0006] Secondly, this application provides a data classification device for an e-commerce platform, the data classification device for an e-commerce platform comprising:

[0007] The data acquisition module is used to perform multi-dimensional feature extraction on user interaction behavior data to obtain a user behavior feature matrix.

[0008] The processing module is used to perform graph neural network processing on the user behavior feature matrix and product information data to obtain a product knowledge graph;

[0009] The analysis module is used to perform interaction pattern analysis on the user behavior feature matrix and the product knowledge graph through a temporal attention mechanism to obtain a temporal interaction feature vector.

[0010] The fusion module is used to perform feature fusion processing on the user behavior feature matrix, the product knowledge graph, and the temporal interaction feature vector through a multimodal feature fusion network to obtain a fused feature vector;

[0011] The classification module is used to classify the fused feature vector through a hierarchical attention decision network to obtain an initial classification result.

[0012] The optimization module is used to perform data classification optimization processing on the initial classification results through an interpretable rule processor to obtain target classification data.

[0013] A third aspect of this application provides a computer device in which the memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via a bus, and when the machine-readable instructions are executed by the processor, the steps of the data classification method for an e-commerce platform described above are performed.

[0014] A fourth aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the aforementioned data classification method for an e-commerce platform.

[0015] The technical solution provided in this application comprehensively captures user behavior features such as search, clicks, and purchases by performing multi-dimensional feature extraction on user interaction behavior data, effectively improving the accuracy of user behavior profiling. By processing the user behavior feature matrix and product information data using graph neural networks, a product knowledge graph containing rich semantic information and structural relationships is constructed, providing comprehensive knowledge support for product classification. A temporal attention mechanism is employed to analyze the interaction patterns of the user behavior feature matrix and product knowledge graph, accurately capturing the dynamic changes in user interests and improving the classification system's understanding of user preferences. A multimodal feature fusion network is introduced to fuse the user behavior feature matrix, product knowledge graph, and temporal interaction feature vectors, achieving effective integration of features from different sources and enhancing the completeness and complementarity of feature representation. A hierarchical attention decision network is used for classification processing, establishing a flexible multi-level classification system and improving the accuracy and hierarchy of classification results. Finally, an interpretable rule processor is used for data classification optimization, ensuring not only the accuracy of the classification results but also providing clear explanations of the classification criteria, enhancing the interpretability and credibility of the classification results. The overall solution significantly improves the accuracy and efficiency of product categorization on e-commerce platforms through the organic combination of multiple innovative technologies. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of an embodiment of the data classification method for an e-commerce platform in this application.

[0018] Figure 2 This is a schematic diagram of one embodiment of a data classification device for an e-commerce platform in this application.

[0019] Figure 3 This is a schematic diagram of the structure of the computer device in the embodiments of this application. Detailed Implementation

[0020] This application provides a data classification method, apparatus, device, and storage medium for e-commerce platforms. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0021] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the data classification method for e-commerce platforms in this application includes:

[0022] Step S101: Perform multi-dimensional feature extraction processing on the user interaction behavior data to obtain the user behavior feature matrix;

[0023] Step S102: Perform graph neural network processing on the user behavior feature matrix and product information data to obtain a product knowledge graph;

[0024] Step S103: Perform interaction pattern analysis on the user behavior feature matrix and product knowledge graph using a temporal attention mechanism to obtain a temporal interaction feature vector;

[0025] Step S104: The user behavior feature matrix, product knowledge graph, and temporal interaction feature vector are fused using a multimodal feature fusion network to obtain a fused feature vector;

[0026] Step S105: Classify the fused feature vectors using a hierarchical attention decision network to obtain the initial classification result;

[0027] Step S106: The initial classification results are processed by the interpretability rule processor to optimize the data classification and obtain the target classification data.

[0028] It is understood that the executing entity of this application can be a data classification device used on an e-commerce platform, or it can be a terminal or a server; no specific limitation is made here. This application's embodiment uses a server as an example for illustration.

[0029] Specifically, user interaction data is processed by reading user search term sequences, click sequences, and purchase sequences on e-commerce platforms to form basic behavioral sequence data. For each user's behavioral sequence, time-dimensional features such as access time distribution, shopping cycle, and dwell time are recorded to construct a time feature vector. By analyzing the relationship between product categories visited before and after the user, the correlation between search terms and purchased products, and the degree of response to promotional activities, a behavioral association graph is generated. The importance of nodes in the behavioral association graph is calculated using the PageRank algorithm to obtain a user interest weight vector. The behavioral sequence data is then time-sliced, and dynamic interest features are extracted using a Long Short-Term Memory network to finally obtain a user behavior feature matrix. Based on the obtained user behavior feature matrix, product information is further processed. First, product titles, descriptions, and specifications are segmented into words, keyword sequences are extracted, and word vectors are encoded. The co-occurrence and conversion relationships of products in the user behavior feature matrix are analyzed to obtain product association data. Product text features and association data are fused to construct product node feature vectors. Hierarchical relationships are established through the product category system, and semantic relationships between products are derived using graph reasoning methods. By using graph neural networks to learn the embeddings of nodes, a knowledge graph containing complete information about the products is ultimately formed.

[0030] After obtaining the user behavior feature matrix and the product knowledge graph, the interaction patterns between them are analyzed using a temporal attention mechanism. First, user behavior is divided into time windows, and decay weights are assigned to behavior data in different time periods. User behavior is matched with the product knowledge graph to extract user-product interaction sequences. Attention weights at different time points are calculated using a temporal attention network, and cross-features are extracted and sequence encoded. Temporal pattern recognition is performed on the encoded sequences to obtain interaction pattern data, which is ultimately converted into a temporal interaction feature vector. The above three features are then processed through a multimodal feature fusion network. First, feature space mapping is performed on the user behavior features and the product knowledge graph, and their mutual attention features are calculated. The mutual attention features are aligned with the temporal interaction features, and feature information is supplemented through a residual connection network. Cross-modal interaction and reconstruction are performed on the aligned features, and compression mapping is performed using an autoencoder. A multi-head attention mechanism is used to capture the relationships between different features, and finally, feature normalization and nonlinear transformation are performed to obtain a fused feature vector.

[0031] For the fused feature vectors, a hierarchical attention decision network is used for classification. First, the system is divided into multiple levels according to the product category system, and weights are assigned to nodes at different levels. The feature attention weights for each level are calculated, and an adaptive discrimination threshold is set. Features are extracted hierarchically through a hierarchical feature extraction network, and the correlation between nodes is calculated. A classification path probability graph is constructed and pruned to obtain a set of candidate paths. Candidate paths are scored and multi-path integration is performed. Finally, an initial classification result is obtained through confidence evaluation and a soft voting mechanism. Finally, an interpretable rule processor is used to optimize the initial classification result. Rules are extracted from the classification result and ranked by importance to generate rule weights. Consistency checks are performed on the rules, and conflicting rules are identified and reconciled. The classification decision path is tracked, and the contribution of features to the classification result is analyzed. Based on the analysis results, classification criteria are extracted, classification labels are corrected, and finally, target classification data containing class labels, classification criteria, and confidence scores is generated.

[0032] For example, an e-commerce platform categorizes mobile phone products. Analyzing user A's behavioral data over a month reveals that this user browses mobile phones extensively during weekday evenings (8:00 PM - 10:00 PM), with an average session duration of 3 minutes, primarily focusing on search terms related to camera features and battery life. Using the PageRank algorithm, the user's interest weight for camera features is calculated to be 0.8, and for battery life, 0.7. Product information processing reveals that a certain mobile phone's description contains keywords such as "100 million pixels" and "5000mAh large battery," leading to its classification as a high-end camera phone using a graph neural network. Analyzing the interaction sequence between user behavior and product features shows that the user's click-through rate for this type of phone increased over time, from 5% initially to 15% later. After multimodal feature fusion and hierarchical classification, this phone is categorized as "Mobile Phone / High-End Mobile Phone / Photography Phone," with a classification confidence score of 0.85. The primary classification criteria are the product's camera configuration features and the user's interest matching.

[0033] In this embodiment, multi-dimensional feature extraction processing of user interaction behavior data comprehensively captures user search, click, and purchase behavior features, effectively improving the accuracy of user behavior profiling. By processing the user behavior feature matrix and product information data using graph neural networks, a product knowledge graph containing rich semantic information and structural relationships is constructed, providing comprehensive knowledge support for product classification. A temporal attention mechanism is employed to analyze the interaction patterns of the user behavior feature matrix and product knowledge graph, accurately capturing the dynamic changes in user interests and improving the classification system's understanding of user preferences. A multi-modal feature fusion network is introduced to fuse the user behavior feature matrix, product knowledge graph, and temporal interaction feature vectors, achieving effective integration of features from different sources and enhancing the completeness and complementarity of feature representation. A hierarchical attention decision network is used for classification processing, establishing a flexible multi-level classification system and improving the accuracy and hierarchy of classification results. Finally, an interpretable rule processor is used for data classification optimization, ensuring not only the accuracy of the classification results but also providing clear explanations of the classification criteria, enhancing the interpretability and credibility of the classification results. The overall solution significantly improves the accuracy and efficiency of product categorization on e-commerce platforms through the organic combination of multiple innovative technologies.

[0034] In one specific embodiment, the process of performing step S101 may specifically include the following steps:

[0035] (1) Query record processing of user interaction behavior data, extract user search term sequence, click sequence and purchase sequence to obtain basic behavior sequence data, and perform time feature analysis processing on basic behavior sequence data to obtain user time feature vector;

[0036] (2) Perform sequence association analysis on the basic behavior sequence data and user time feature vector to obtain the user behavior association graph, and use the PageRank algorithm to calculate the node importance of the user behavior association graph to obtain the user interest weight vector.

[0037] (3) Perform time-series slicing on the user time feature vector and the user interest weight vector to obtain time-series interest segments, and perform sequence encoding on the time-series interest segments to obtain the user interest evolution sequence.

[0038] (4) The user interest evolution sequence is processed by a long short-term memory network to extract sequence features, resulting in a dynamic interest feature vector. The dynamic interest feature vector and the user interest weight vector are then concatenated to obtain a combined feature vector.

[0039] (5) Perform dimension normalization on the combined feature vectors to obtain standardized feature vectors, and perform matrix transformation on the standardized feature vectors to obtain the user behavior feature matrix.

[0040] Specifically, the system queries and processes user interaction data generated on e-commerce platforms, retrieving each user's historical search records, click records, and purchase records from the database. Search term sequences record the user's entered search keywords and their timestamps; click sequences record the IDs of products viewed and the duration of each visit; and purchase sequences record the order information of the final purchased products, forming basic behavioral sequence data. Temporal feature analysis is then performed on this basic behavioral sequence data to extract features such as the distribution of user activity periods, shopping cycles, and time intervals between various behaviors, generating a user time feature vector to characterize the user's time-dimensional behavioral patterns. For example, a user's peak daily activity period is from 8:00 PM to 10:00 PM, with an average of one search every three days and an average dwell time of two minutes per product view. These temporal characteristics are encoded into the feature vector.

[0041] Further sequence association analysis was performed on the basic behavioral sequence data and user time feature vectors to uncover the relationships between user behaviors. This included analyzing the semantic relevance between adjacent search terms, the conversion relationship from search to click, and the conversion relationship from click to purchase, thus constructing a user behavior association graph. The PageRank algorithm was then used to calculate the importance score of each behavioral node in the constructed association graph, generating a user interest weight vector. During the algorithm process, nodes directly related to purchase behavior were assigned higher initial weights, and the final importance score of each node was calculated through iterative propagation. For example, in the mobile phone category, the user's search term "camera phone" node scored 0.8, indicating a strong interest in the mobile phone's camera function.

[0042] Subsequently, the user time feature vector and user interest weight vector are processed by time-series slicing, dividing the complete behavior sequence into multiple time-series interest segments according to time windows (such as days, weeks, and months). For each time-series interest segment, the user's behavioral features and interest distribution within that time period are extracted and encoded into a fixed-dimensional vector representation using a sequence encoder, forming a user interest evolution sequence reflecting the changing trends of user interests. Taking a user as an example, their interest gradually shifts from initially focusing on basic mobile phone functions to focusing on camera performance; this evolution trend is encoded in the sequence.

[0043] Next, a Long Short-Term Memory (LSTM) network is used to process the user interest evolution sequence, which can capture long-term dependencies in the sequence data. Through the network's gating mechanism, interest features from different periods are selectively retained and forgotten, generating a feature vector that expresses the dynamic changes in user interests. This dynamic interest feature vector is concatenated with the previously obtained user interest weight vector to obtain a combined feature vector that includes both the current interest distribution and interest evolution information. For example, if a user's current interest weight for the photography function is 0.8, and historical trends show a stable upward trend, this information is integrated into the combined feature vector. Finally, the combined feature vector undergoes dimensionality normalization, mapping feature values ​​from different dimensions to the [0,1] interval to eliminate the influence of dimensions. The normalized feature vector is then processed through linear or nonlinear transformations to adjust the feature distribution and generate the final user behavior feature matrix. Each row of this matrix corresponds to a user, and each column corresponds to a feature dimension, comprehensively describing the user's behavioral characteristics and interest preferences. Based on this feature matrix, more accurate product recommendations and personalized services can be implemented subsequently. For the aforementioned user, based on their feature matrix showing a high level of concern for camera performance, the system will prioritize recommending mobile phone products with outstanding camera capabilities.

[0044] In one specific embodiment, the process of performing step S102 may specifically include the following steps:

[0045] (1) Perform text segmentation on the product information data, extract the keyword sequence from the product title, description and specification parameters to obtain the product feature word sequence, and perform word vector encoding on the product feature word sequence to obtain the product text feature vector;

[0046] (2) Perform product association analysis on the user behavior feature matrix, extract the co-occurrence relationship and conversion relationship between products, obtain product association data, and perform feature fusion processing on the product association data and product text feature vector to obtain product node feature vector;

[0047] (3) Perform category-level matching on the feature vectors of product nodes, construct the category relationship of products, obtain product category structure data, and deduce product relationship from the product category structure data through graph reasoning method to obtain product semantic relationship data;

[0048] (4) Perform graph structure construction on the semantic relationship data of the products, establish the connection relationship between product nodes, obtain the initial knowledge graph, and perform node embedding learning on the initial knowledge graph through graph neural network to obtain the product node embedding vector;

[0049] (5) Perform knowledge fusion processing on the product node embedding vector and product semantic relationship data to obtain a product knowledge graph.

[0050] Specifically, the product information data from e-commerce platforms is processed by segmenting text content such as product titles, detailed descriptions, and specifications into word units using word segmentation technology. When extracting keyword sequences, key information such as product model, brand, and core parameters is retained to form a product feature word sequence. For each feature word, a word vector encoding model is used to calculate its vector representation in the semantic space, resulting in a product text feature vector. For example, the description text of a mobile phone, "6.7-inch OLED curved screen, 100-megapixel main camera, 5000mAh large battery," is segmented to obtain a feature word sequence, with each word mapped to a fixed-dimensional vector. The relationships between products are analyzed based on a user behavior feature matrix, counting the frequency of different products being viewed or purchased by the same user and calculating the co-occurrence probability between product pairs. Simultaneously, the conversion path from browsing one product to ultimately purchasing another is analyzed to obtain the conversion relationships between products. These statistically obtained product association data are then fused with the product text feature vectors to generate product node feature vectors containing multi-dimensional information. For example, if it is found that after a user browses high-end camera phones, there is an 80% probability that they will also pay attention to other camera phones in the same price range, this correlation is integrated into the node features.

[0051] The product node feature vectors are matched against the category system to map products to the pre-defined category hierarchy of the e-commerce platform, determining the category path to which each product belongs. A formalized product category structure data is constructed based on the product's category affiliation. Graph reasoning methods are used to expand the category structure, deriving implicit semantic relationships between products through transitivity and similarity rules, generating product semantic relationship data. For example, the category path for "mobile phone" is categorized as "mobile phone / smartphone / camera phone," and similarity relationships with other mobile phones with prominent camera functions are derived.

[0052] A graph structure is constructed based on product semantic relationship data, with products as nodes and various relationships between products as edges, forming an initial knowledge graph. Graph neural network technology is used to learn representations for the nodes in the graph. During message passing, the network aggregates information from neighboring nodes, learning product node embedding vectors that contain structural information. For example, for a camera phone node, its embedding vector will incorporate feature information from neighboring nodes such as related accessories and similar products. Finally, the product node embedding vectors are fused with the product semantic relationship data to construct a complete product knowledge graph. This graph not only contains basic product attribute information but also rich knowledge such as hierarchical relationships, similarity relationships, and association relationships between products. For example, in the final knowledge graph, each mobile phone product is connected to entities such as its accessories, competitors, and applicable scenarios, forming a complete knowledge network. Based on this knowledge graph, the system can better understand the relationships between products, providing knowledge support for subsequent classification tasks. For example, when classifying newly listed mobile phones, the category information and attribute features of similar products in the knowledge graph can be referenced.

[0053] In one specific embodiment, the process of executing step S103 may specifically include the following steps:

[0054] (1) Divide the user behavior feature matrix into time windows to obtain user behavior time series segments, and perform time decay weighting on the user behavior time series segments to obtain time series weight data.

[0055] (2) Perform interactive matching processing on the temporal weight data and the product knowledge graph, extract the user-product interaction sequence, obtain the interaction sequence data, and calculate the temporal features of the interaction sequence data through the temporal attention network to obtain the temporal attention weight;

[0056] (3) Cross-feature extraction is performed on the interaction sequence data and temporal attention weights to obtain user-product interaction features, and the user-product interaction features are sequence encoded to obtain the interaction encoding vector;

[0057] (4) Perform temporal pattern recognition processing on the interaction coding vector to obtain interaction pattern data, and perform feature combination processing on the interaction pattern data and temporal weight data to obtain combined interaction features.

[0058] (5) Perform dimensional transformation on the combined interaction features to obtain the temporal interaction feature vector.

[0059] Specifically, the user behavior feature matrix is ​​divided into time windows, and the continuous time series is segmented into multiple time-series segments of fixed length (e.g., hours, days, weeks). A time decay factor is introduced into the user behavior data in each time-series segment, so that the weight of behavior data further away from the current time is lower, generating time-series weighted data. For example, if a user's browsing history over the past 30 days is segmented by day, the weight of the behavior on day i is set to 0.95. i This has led to greater attention being paid to recent actions.

[0060] User behavior data with temporal weights is interactively matched with a product knowledge graph to record user interactions with different products at different times, extracting user-product interaction sequences. These interaction sequences are then processed using a temporal attention network to calculate the importance weights of interactions at different times. For example, when a user is browsing on their phone, comparing a quick browse of 0-2 minutes with a detailed examination of 2-5 minutes, the latter receives a higher attention weight, indicating that the user is more interested in the product.

[0061] Based on interaction sequence data and calculated temporal attention weights, cross-feature extraction is performed to analyze the correlation patterns between user behavior and product attributes. User action sequences (such as clicks, favorites, adding to cart, and purchases) are cross-combined with product attribute features (such as price, category, and performance parameters) to generate user-product interaction features. These interaction features are encoded using a sequence encoder, transforming the variable-length interaction sequences into fixed-dimensional interaction encoding vectors. For example, recording that a user browsed multiple mobile phones at different price points and then focused on mid-to-high-end camera phones, this behavioral pattern is encoded into a vector. Temporal pattern recognition is performed on the interaction encoding vectors to analyze the patterns of user behavior changes over time. Temporal patterns such as user interest shift trajectories and shopping decision paths are identified to generate interaction pattern data. The identified interaction pattern data is combined with the previous temporal weight data to obtain combined interaction features containing temporal information. For example, if the system discovers that a user initially focused on the basic functions of a mobile phone, then gradually shifted their attention to camera performance, and recently showed a strong purchase intention, this evolutionary pattern is preserved in the combined features.

[0062] Finally, the combined interaction features undergo dimensionality transformation. Through operations such as dimensionality reduction or feature transformation, the features are mapped to a suitable dimensional space to generate the final temporal interaction feature vector. This feature vector reflects both the user's historical behavioral trajectory and current interests, providing important information for subsequent classification tasks. For example, based on a user's temporal interaction feature vector, it may show that the user's attention to camera phones is continuously increasing, and their recent browsing time and interaction frequency are both high. This information can be used to predict the user's purchase intention and product category preference.

[0063] In one specific embodiment, the process of executing step S104 may specifically include the following steps:

[0064] (1) Perform feature space mapping on the user behavior feature matrix to obtain the user behavior projection vector, and perform graph structure embedding on the product knowledge graph to obtain the product graph vector;

[0065] (2) Perform mutual attention calculation on the user behavior projection vector and the product map vector to obtain interactive attention features, and perform feature alignment processing on the interactive attention features and the temporal interactive feature vector to obtain the alignment feature matrix;

[0066] (3) The aligned feature matrix is ​​supplemented by a residual connection network to obtain supplemented feature data, and cross-modal feature extraction is performed on the supplemented feature data to obtain modal interaction features;

[0067] (4) Perform feature reconstruction processing on the modal interaction features to obtain the reconstructed feature vector, and perform compression mapping on the reconstructed feature vector through an autoencoder to obtain compressed feature data;

[0068] (5) Perform multi-head attention calculation on the compressed feature data to obtain multi-head feature vectors, and perform feature normalization on the multi-head feature vectors to obtain normalized feature data.

[0069] (6) Perform feature concatenation processing on the normalized feature data and modal interaction features to obtain the concatenated feature vector, and perform nonlinear transformation on the concatenated feature vector through a multilayer perceptron to obtain transformed feature data.

[0070] (7) Perform feature selection processing on the transformed feature data to obtain the fused feature vector.

[0071] Specifically, the user behavior feature matrix is ​​spatially mapped by projecting user behavior data into a new feature space through linear or nonlinear transformations, generating user behavior projection vectors. Simultaneously, the product knowledge graph is structurally embedded, encoding the nodes and edge relationships in the graph into vector representations, resulting in product graph vectors. For example, user browsing, favorites, and purchase behaviors are mapped to a 128-dimensional vector space, and product attributes and relationship information are also mapped to a vector space of the same dimension. Mutual attention scores are calculated between the user behavior projection vectors and the product graph vectors to analyze the correlation strength between the two types of features, generating interactive attention features. These interactive attention features are aligned with temporal interactive feature vectors along the feature dimensions, ensuring that features from different sources can be compared and fused in the same feature space, forming an aligned feature matrix. For example, if a user's behavior vector for the camera phone category and the graph vector of products in that category yield a high attention score, it indicates a high correlation between user behavior and product features.

[0072] The aligned feature matrix is ​​supplemented using a residual connection network, adding new feature representations while retaining the original feature information to obtain supplementary feature data. Cross-modal feature extraction is then performed on this supplementary feature data to analyze the complementary relationships between features from different modalities (user behavior, product attributes, temporal relationships), generating modal interaction features. For example, by cross-referencing user purchase behavior features with product price range features, user preference patterns for mid-to-high-end priced products can be discovered.

[0073] Modal interaction features are reconstructed by generating reconstructed feature vectors through feature transformation and combination. An autoencoder is then used to compress the reconstructed feature vectors, reducing feature dimensionality while retaining key information, resulting in compressed feature data. For example, 512-dimensional interaction features can be compressed to 256 dimensions using an autoencoder, retaining the most representative feature information. A multi-head attention mechanism is then employed to process the compressed feature data, calculating attention weights from multiple feature subspaces to generate multi-head feature vectors. These multi-head feature vectors are then normalized, scaling the feature values ​​to a uniform numerical range, resulting in normalized feature data. For instance, eight attention heads can be used simultaneously to focus on different feature dimensions, with each head generating a 64-dimensional feature representation.

[0074] Normalized feature data is concatenated with previous modal interaction features to form a more comprehensive feature representation, resulting in a concatenated feature vector. A multilayer perceptron is then used to perform a nonlinear transformation on the concatenated feature vector to enhance its expressive power, generating transformed feature data. For example, a three-layer perceptron can be used with 512, 256, and 128 neurons in the hidden layers, employing ReLU as the activation function. Finally, feature selection is performed on the transformed feature data, filtering out the most representative feature dimensions based on their importance scores to generate the final fused feature vector. This feature vector integrates user behavior, product knowledge, and temporal information, providing comprehensive feature support for subsequent classification tasks. For instance, a user's fused feature vector might show a high level of interest in features such as camera functionality, screen quality, and battery life within the mobile phone category; this information can be used to predict specific product categories the user might be interested in.

[0075] In one specific embodiment, the process of executing step S105 may specifically include the following steps:

[0076] (1) Perform multi-level category division on the fused feature vector to obtain category hierarchical structure data, and perform node weight allocation on the category hierarchical structure data to obtain the hierarchical node weights.

[0077] (2) Perform hierarchical attention calculation on the fused feature vector and hierarchical node weights to obtain hierarchical feature weights, and perform adaptive threshold adjustment on the hierarchical feature weights to obtain hierarchical discrimination thresholds;

[0078] (3) The fused feature vector is subjected to hierarchical feature extraction through a hierarchical feature extraction network to obtain a hierarchical feature matrix, and the node correlation calculation is performed on the hierarchical feature matrix to obtain node association data.

[0079] (4) Perform path probability calculation on the node association data to obtain a classification path probability graph, and perform path pruning on the classification path probability graph to obtain a candidate path set.

[0080] (5) Perform path scoring processing on the candidate path set and the hierarchical discrimination threshold to obtain path score data, and perform multi-path integration processing on the path score data to obtain integrated classification data;

[0081] (6) Perform confidence assessment on the integrated classification data to obtain classification confidence, and assign labels to the classification confidence through a soft voting mechanism to obtain the initial classification results.

[0082] Specifically, the fused feature vectors are divided into multi-level categories, constructing a hierarchical classification structure based on the e-commerce platform's pre-defined product category system. From top-level broad categories to bottom-level subcategories, a complete hierarchical category structure is formed. Each node in the category hierarchy is weighted according to its importance and distinguishability, generating hierarchical node weights. For example, the hierarchical structure of the mobile phone category can be divided into levels such as "mobile phone / smartphone / camera phone," with each level node receiving different weight values ​​based on the number of products it contains and its feature distinguishability. Hierarchical attention is calculated by combining the fused feature vectors with the hierarchical node weights to analyze the importance of features at different levels, obtaining hierarchical feature weights. Adaptive discrimination thresholds are set for the feature weights at different levels, dynamically adjusting the threshold values ​​to control classification accuracy, generating hierarchical discrimination thresholds. For example, in the mobile phone category, the discrimination threshold for the first-level category "mobile phone" is set to 0.6, the threshold for the second-level category "smartphone" is set to 0.7, and the threshold for the third-level category "camera phone" is set to 0.8, ensuring that classification accuracy improves layer by layer. A hierarchical feature extraction network is used to process the fused feature vector, extracting feature representations at different levels to form a hierarchical feature matrix. The correlation between nodes at different levels is calculated, and the similarity and dependency of node features are analyzed to generate node association data. For example, the feature correlation between the "camera phone" node and other mobile phone category nodes is analyzed to identify a set of nodes with significant associations.

[0083] Path probability calculations are performed on node-related data to construct a classification path probability graph from the root node to the leaf node. Path pruning removes low-probability classification paths, retaining the most likely ones to form a candidate path set. For example, in classifying a product, possible paths include "mobile phone / smartphone / camera phone" and "mobile phone / smartphone / gaming phone." Paths with higher confidence are retained through probability calculation and pruning. Each classification path is scored based on the candidate path set and a hierarchical discrimination threshold, generating path score data. Multiple possible classification paths are integrated, and the classification results are combined using weighted averaging or voting to obtain integrated classification data. For example, a mobile phone product scores 0.85 on the "camera phone" path and 0.65 on the "gaming phone" path; the final classification direction is determined through integration. The integrated classification data undergoes confidence evaluation, calculating the reliability score of the classification results to generate classification confidence. A soft voting mechanism is used to assign labels to classification results with higher confidence, comprehensively considering the voting results of multiple classification paths to obtain the initial classification result. For example, for a certain mobile phone product, after confidence assessment and soft voting, it was finally classified as a "camera phone" with a confidence level of 0.85, which meets the preset discrimination threshold requirements.

[0084] In one specific embodiment, the process of executing step S106 may specifically include the following steps:

[0085] (1) Extract rules from the initial classification results to obtain a set of classification rules, and sort the classification rule set by rule importance to obtain rule weight data;

[0086] (2) Perform rule matching processing on the initial classification results and rule weight data to obtain the rule coverage matrix, and perform consistency verification processing on the rule coverage matrix to obtain rule consistency data;

[0087] (3) Perform conflict rule identification processing on the rule consistency data to obtain rule conflict data, and perform rule reconciliation processing on the rule conflict data to obtain a set of reconciliation rules;

[0088] (4) Perform decision path tracing on the harmonization rule set to obtain the classification decision path, and perform feature contribution analysis on the classification decision path to obtain feature interpretation data;

[0089] (5) Extract classification criteria from the feature interpretation data to obtain a classification criteria vector, and then perform label correction on the classification criteria vector and the initial classification results to obtain corrected classification labels;

[0090] (6) Perform category structure mapping processing on the modified category labels to obtain target category data, which includes product category labels, classification basis descriptions and confidence scores.

[0091] Specifically, rules are extracted from the initial classification results, and key judgment conditions and feature thresholds in the classification decision process are analyzed to form a set of classification rules. Each rule in the rule set is evaluated for importance based on its coverage, accuracy, and discriminative power, generating rule weight data. For example, for mobile phone classification, a rule like "phones with over 50 million pixels and a photo score greater than 85 are classified as camera phones" is extracted, and weights are assigned based on the rule's discriminative effect. The initial classification results are matched with the rule weight data to check whether each classification result meets the corresponding rule conditions, generating a rule coverage matrix. A consistency check is performed on the rule coverage matrix to analyze the coordination and complementarity between different rules, obtaining rule consistency data. If a mobile phone satisfies multiple classification rules simultaneously, it is necessary to evaluate whether there are any logical conflicts between these rules.

[0092] Analyzing rule consistency data identifies potentially conflicting rule combinations, generating rule conflict data. Conflicts are resolved through rule reconciliation, which may involve merging rules, adjusting priorities, or adding supplementary conditions, ultimately resulting in a harmonized set of rules. For example, when a phone simultaneously meets the classification rules for both a camera phone and a gaming phone, conflict reconciliation is achieved by analyzing the weights of core parameters to determine the most suitable classification. Decision path tracing is then performed on the harmonized rule set to reconstruct key decision points and judgment criteria during the classification process, forming a complete classification decision path. The influence of each feature in the decision path on the final classification result is analyzed, feature contribution is calculated, and feature explanation data is generated. For instance, analysis might reveal that the primary basis for classifying a phone as a camera phone is its camera configuration and image processing capabilities.

[0093] Based on feature interpretation data, key classification criteria are extracted, and qualitative and quantitative judgment conditions are integrated into a classification criterion vector. The initial classification results are reviewed and corrected using this vector to ensure accuracy and interpretability, resulting in revised classification labels. For example, analysis might reveal that while a certain phone has excellent camera capabilities, its primary marketing focus and target user group lean towards gaming performance; therefore, the classification label is adjusted accordingly. Finally, the revised classification labels are mapped onto the e-commerce platform's category structure, generating target classification data containing multi-dimensional information. This classification data not only includes specific product category labels but also detailed explanations of the classification criteria and a confidence score for the classification results, providing a basis for subsequent product management and user services. For example, the final classification data for a certain phone shows the category label as "phone / smartphone / gaming phone," with classification criteria including core features such as "Snapdragon 8 series processor, 16GB large memory, 144Hz high refresh rate screen," and a confidence score of 0.92, indicating high reliability of the classification results.

[0094] The data classification method for e-commerce platforms in the embodiments of this application has been described above. The data classification apparatus for e-commerce platforms in the embodiments of this application are described below. Please refer to [link / reference]. Figure 2 One embodiment of the data classification device for e-commerce platforms in this application includes:

[0095] The acquisition module 201 is used to perform multi-dimensional feature extraction processing on user interaction behavior data to obtain a user behavior feature matrix.

[0096] Processing module 202 is used to perform graph neural network processing on the user behavior feature matrix and product information data to obtain a product knowledge graph;

[0097] Analysis module 203 is used to perform interaction pattern analysis on the user behavior feature matrix and the product knowledge graph through a temporal attention mechanism to obtain a temporal interaction feature vector;

[0098] The fusion module 204 is used to perform feature fusion processing on the user behavior feature matrix, the product knowledge graph and the temporal interaction feature vector through a multimodal feature fusion network to obtain a fused feature vector;

[0099] The classification module 205 is used to classify the fused feature vector through a hierarchical attention decision network to obtain an initial classification result;

[0100] The optimization module 206 is used to perform data classification optimization processing on the initial classification result through an interpretable rule processor to obtain target classification data.

[0101] Through the collaborative efforts of the aforementioned components, multi-dimensional feature extraction of user interaction data comprehensively captures user search, click, and purchase behaviors, effectively improving the accuracy of user behavior profiling. By processing the user behavior feature matrix and product information data using graph neural networks, a product knowledge graph rich in semantic information and structural relationships is constructed, providing comprehensive knowledge support for product classification. A temporal attention mechanism is employed to analyze interaction patterns in the user behavior feature matrix and product knowledge graph, accurately capturing the dynamic changes in user interests and improving the classification system's understanding of user preferences. A multi-modal feature fusion network is introduced to fuse the user behavior feature matrix, product knowledge graph, and temporal interaction feature vectors, achieving effective integration of features from different sources and enhancing the completeness and complementarity of feature representations. A hierarchical attention decision network is used for classification processing, establishing a flexible multi-level classification system and improving the accuracy and hierarchy of classification results. Finally, an interpretable rule processor is used for data classification optimization, ensuring not only the accuracy of the classification results but also providing clear explanations of the classification criteria, enhancing the interpretability and credibility of the classification results. The overall solution significantly improves the accuracy and efficiency of product categorization on e-commerce platforms through the organic combination of multiple innovative technologies.

[0102] Based on the same technical concept, embodiments of this application also provide a computer device. (Refer to...) Figure 3 The diagram shown is a structural schematic of a computer device 300 provided in an embodiment of this application, including a processor 301, a memory 302, and a bus 303. The memory 302 stores execution instructions and includes a main memory 3021 and an external memory 3022. The main memory 3021, also called internal memory, is used to temporarily store computational data in the processor 301 and data exchanged with external memory 3022 such as a hard disk. The processor 301 exchanges data with the external memory 3022 through the main memory 3021. When the computer device 300 is running, the processor 301 and the memory 302 communicate via the bus 303.

[0103] This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the data classification method for an e-commerce platform.

[0104] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0105] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A data classification method for e-commerce platforms, characterized in that, The data classification method for e-commerce platforms includes: performing multi-dimensional feature extraction processing on user interaction behavior data to obtain a user behavior feature matrix; The user behavior feature matrix and product information data are processed using a graph neural network to obtain a product knowledge graph. This step specifically includes: performing text segmentation on the product information data, extracting keyword sequences from product titles, descriptions, and specifications to obtain product feature word sequences, and performing word vector encoding on these product feature word sequences to obtain product text feature vectors; performing product association analysis on the user behavior feature matrix, extracting co-occurrence and conversion relationships between products to obtain product association data, and performing feature fusion processing on the product association data and the product text feature vectors to obtain product node feature vectors; performing category-level matching on the product node feature vectors to construct product category relationships to obtain product category structure data, and performing product relationship deduction on the product category structure data using graph reasoning methods to obtain product semantic relationship data; performing graph structure construction processing on the product semantic relationship data to establish connections between product nodes to obtain an initial knowledge graph, and performing node embedding learning on the initial knowledge graph using a graph neural network to obtain product node embedding vectors; and performing knowledge fusion processing on the product node embedding vectors and the product semantic relationship data to obtain the product knowledge graph. The user behavior feature matrix and the product knowledge graph are processed by a temporal attention mechanism to analyze interaction patterns and obtain a temporal interaction feature vector. The user behavior feature matrix, the product knowledge graph, and the temporal interaction feature vector are fused using a multimodal feature fusion network to obtain a fused feature vector. The fused feature vectors are classified using a hierarchical attention decision network to obtain an initial classification result; The initial classification results are then processed by an interpretable rule processor to optimize the data classification and obtain the target classification data.

2. The data classification method for e-commerce platforms according to claim 1, characterized in that, The process of extracting multidimensional features from user interaction behavior data to obtain a user behavior feature matrix includes: processing user interaction behavior data for query records, extracting user search term sequences, click sequences, and purchase sequences to obtain basic behavior sequence data, and performing time feature analysis on the basic behavior sequence data to obtain a user time feature vector. The basic behavioral sequence data and the user time feature vector are subjected to sequence association analysis to obtain a user behavior association graph. The user behavior association graph is then processed by the PageRank algorithm to calculate the node importance and obtain a user interest weight vector. The user time feature vector and the user interest weight vector are subjected to time-series slicing to obtain time-series interest segments, and the time-series interest segments are subjected to sequence encoding to obtain user interest evolution sequences. The user interest evolution sequence is processed by a long short-term memory network to extract sequence features, resulting in a dynamic interest feature vector. The dynamic interest feature vector and the user interest weight vector are then concatenated to obtain a combined feature vector. The combined feature vectors are normalized to obtain standardized feature vectors, and the standardized feature vectors are then subjected to matrix transformation to obtain the user behavior feature matrix.

3. The data classification method for e-commerce platforms according to claim 1, characterized in that, The step of performing interaction pattern analysis and processing on the user behavior feature matrix and the product knowledge graph through a temporal attention mechanism to obtain a temporal interaction feature vector includes: dividing the user behavior feature matrix into time windows to obtain user behavior time segments, and performing time decay weighting processing on the user behavior time segments to obtain temporal weight data. The temporal weight data and the product knowledge graph are subjected to interactive matching processing to extract user-product interaction sequences, obtain interaction sequence data, and the interaction sequence data is processed by a temporal attention network to calculate temporal features and obtain temporal attention weights. Cross-feature extraction is performed on the interaction sequence data and the temporal attention weights to obtain user-product interaction features, and the user-product interaction features are then subjected to sequence encoding to obtain an interaction encoding vector. The interaction encoding vector is subjected to temporal pattern recognition processing to obtain interaction pattern data, and the interaction pattern data and the temporal weight data are subjected to feature combination processing to obtain combined interaction features. The combined interaction features are subjected to dimensionality transformation to obtain the temporal interaction feature vector.

4. The data classification method for e-commerce platforms according to claim 1, characterized in that, The step of performing feature fusion processing on the user behavior feature matrix, the product knowledge graph, and the temporal interaction feature vector through a multimodal feature fusion network to obtain a fused feature vector includes: performing feature space mapping processing on the user behavior feature matrix to obtain a user behavior projection vector, and performing graph structure embedding processing on the product knowledge graph to obtain a product graph vector. The user behavior projection vector and the product map vector are subjected to mutual attention calculation to obtain interactive attention features, and the interactive attention features and the temporal interactive feature vector are subjected to feature alignment processing to obtain an aligned feature matrix. The aligned feature matrix is ​​supplemented by a residual connection network to obtain supplemented feature data, and cross-modal feature extraction is performed on the supplemented feature data to obtain modal interaction features. The modal interaction features are subjected to feature reconstruction processing to obtain reconstructed feature vectors, and the reconstructed feature vectors are compressed and mapped through an autoencoder to obtain compressed feature data. Multi-head attention computation is performed on the compressed feature data to obtain multi-head feature vectors, and feature normalization is performed on the multi-head feature vectors to obtain normalized feature data. The normalized feature data and the modal interaction features are concatenated to obtain a concatenated feature vector, and the concatenated feature vector is nonlinearly transformed through a multilayer perceptron to obtain transformed feature data. The transformed feature data is subjected to feature selection processing to obtain the fused feature vector.

5. The data classification method for e-commerce platforms according to claim 1, characterized in that, The step of classifying the fused feature vector through a hierarchical attention decision network to obtain an initial classification result includes: performing multi-level category division on the fused feature vector to obtain category hierarchical structure data, and performing node weight allocation on the category hierarchical structure data to obtain hierarchical node weights. The fused feature vector and the hierarchical node weights are subjected to hierarchical attention calculation to obtain hierarchical feature weights, and the hierarchical feature weights are subjected to adaptive threshold adjustment to obtain hierarchical discrimination thresholds. The fused feature vector is subjected to hierarchical feature extraction through a hierarchical feature extraction network to obtain a hierarchical feature matrix. The hierarchical feature matrix is ​​then processed to calculate node correlation to obtain node association data. The path probability calculation is performed on the node-related data to obtain a classification path probability map, and the classification path probability map is then pruned to obtain a candidate path set. The candidate path set and the hierarchical discrimination threshold are subjected to path scoring processing to obtain path score data, and the path score data is subjected to multi-path integration processing to obtain integrated classification data. The integrated classification data is subjected to confidence evaluation to obtain classification confidence, and the classification confidence is then used to assign labels to the classification confidence through a soft voting mechanism to obtain the initial classification result.

6. The data classification method for e-commerce platforms according to claim 1, characterized in that, The step of performing data classification optimization processing on the initial classification result through an interpretable rule processor to obtain target classification data includes: performing rule extraction processing on the initial classification result to obtain a set of classification rules, and sorting the set of classification rules by rule importance to obtain rule weight data; The initial classification results and the rule weight data are subjected to rule matching processing to obtain a rule coverage matrix, and the rule coverage matrix is ​​subjected to consistency verification processing to obtain rule consistency data. The rule consistency data is subjected to conflict rule identification processing to obtain rule conflict data, and the rule conflict data is subjected to rule reconciliation processing to obtain a reconciliation rule set; The set of harmonization rules is subjected to decision path tracing processing to obtain classification decision paths, and the classification decision paths are subjected to feature contribution analysis processing to obtain feature explanation data. The feature interpretation data is processed to extract classification criteria to obtain a classification criteria vector, and the classification criteria vector and the initial classification result are processed to correct the labels to obtain corrected classification labels. The modified category labels are subjected to category structure mapping processing to obtain the target category data, wherein the target category data includes the product category label, classification basis description and confidence score.

7. A data classification device for an e-commerce platform, used to implement the data classification method for an e-commerce platform as described in any one of claims 1 to 6, characterized in that, The data classification device for e-commerce platforms includes: a data acquisition module, used to perform multi-dimensional feature extraction processing on user interaction behavior data to obtain a user behavior feature matrix; The processing module is used to perform graph neural network processing on the user behavior feature matrix and product information data to obtain a product knowledge graph; The analysis module is used to perform interaction pattern analysis on the user behavior feature matrix and the product knowledge graph through a temporal attention mechanism to obtain a temporal interaction feature vector. The fusion module is used to perform feature fusion processing on the user behavior feature matrix, the product knowledge graph, and the temporal interaction feature vector through a multimodal feature fusion network to obtain a fused feature vector; The classification module is used to classify the fused feature vector through a hierarchical attention decision network to obtain an initial classification result. The optimization module is used to perform data classification optimization processing on the initial classification results through an interpretable rule processor to obtain target classification data.

8. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the data classification method for an e-commerce platform as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the data classification method for e-commerce platforms as described in any one of claims 1 to 6.

Images