User tagging management and demand analysis system based on big data e-commerce

Abstract

The invention discloses a user tagging management and demand analysis system based on big data e-commerce, and the system comprises the following steps: a collection module which is used for collecting multi-modal behavior data, and generating a context-aware behavior representation vector; the label extraction module is used for constructing a content feature channel, a behavior feature channel and an emotion feature channel and extracting a multi-granularity label candidate set; the atlas construction module is used for constructing a label atlas according to the multi-granularity label candidate set; the intention reasoning module is used for inputting the label atlas and the current behavior representation of the user into an intention reasoning network and outputting a potential demand representation vector of the user; the clustering module is used for inputting the user potential demand representation vector into an improved grey wolf demand clustering algorithm for clustering analysis to generate a user demand group tag; and the structure output module is used for generating a user-label-demand ternary structure. According to the method, multi-channel modeling and the improved grey wolf algorithm are fused, and accurate identification of e-commerce user tags and demands is realized.

Description

Technical Field

[0001] The present invention relates to the field of big data analysis and e-commerce intelligent operation technology, and in particular to a user tagging management and demand analysis system based on big data e-commerce. Background Art

[0002] With the continuous growth in the number of e-commerce platform users and the increasing variety of products, platforms are facing unprecedented information overload and personalized operational challenges. Effectively identifying user behavioral characteristics, building accurate user profiles, and predicting potential user needs have become key technical directions for improving user stickiness, optimizing recommendation effectiveness, and enhancing platform operational efficiency. Traditional user profiling methods primarily rely on static attributes and shallow behavioral characteristics, such as age, gender, purchase history, and number of views, generating labels through rule configuration or simple clustering. While such methods have some applicability in the early stages, they have significant limitations when dealing with multi-source heterogeneous data, dynamic interest migration, and deep behavioral intentions.

[0003] First, existing user tagging systems are mostly static and preset, lacking the ability to fully model the diversity of user behavior and contextual information. User behavior on e-commerce platforms is time-sensitive, multimodal, and emotion-driven. Traditional systems struggle to capture the true intentions and contextual states behind user behavior. Tag update cycles are long, granular, and interpretable, making it difficult to support high-frequency iterations of personalized recommendations and refined marketing needs. Second, current relationship modeling between tags mostly relies on keyword co-occurrence or shared category divisions, lacking the ability to build structured graphs. This makes it difficult to establish semantic propagation paths and behavioral similarity connections between tags, resulting in a fragmented and severely redundant tagging system, further limiting the precise expression of user interest structures.

[0004] When it comes to user demand analysis, traditional methods typically rely on statistical methods or shallow clustering, relying on explicit behavioral data within a fixed window. This makes it difficult to reveal users' deeper behavioral patterns or implicit demand intentions. Most systems ignore the dynamic interactive relationship between user tags and behaviors, making it impossible to predict demand trends from user status changes, resulting in delayed recommendations and slow policy responses. In addition, the performance of existing clustering algorithms in high-dimensional feature spaces is often limited. Clustering results are often greatly affected by initial values, are prone to falling into local optimality, and have weak boundary determination capabilities. Especially when faced with massive high-dimensional user demand representation vectors, clustering quality and stability are difficult to guarantee, making it impossible to provide a reliable basis for group division for subsequent recommendation systems.

[0005] In the clustering optimization process, some studies have attempted to introduce swarm intelligence algorithms, such as the standard grey wolf optimization algorithm and particle swarm optimization algorithm, to improve the initial center selection problem of the traditional K-means method. However, these methods generally lack adaptive mechanisms that match the data density structure and are difficult to handle in real e-commerce data scenarios with uneven user distribution and drastic density fluctuations. At the same time, there is a lack of flexible search strategies between individuals in the clustering process, and conventional algorithms have a single update path, making it difficult to effectively escape the local optimal area when faced with complex high-dimensional feature representations. In addition, the system lacks the fusion and utilization of label maps and demand clustering results, failing to achieve a closed-loop structure from user labeling to demand forecasting to product management, resulting in fragmented data value and limited intelligence.

[0006] Therefore, how to provide a user tagging management and demand analysis system based on big data e-commerce is an urgent problem that technical personnel in this field need to solve. Summary of the Invention

[0007] One purpose of the present invention is to propose a user labeling management and demand analysis system based on big data e-commerce. The present invention integrates multi-channel feature modeling, label graph construction and improved gray wolf clustering algorithm. The system realizes multi-granularity label extraction and potential demand group identification of e-commerce platform users, and constructs a user-label-demand ternary structure for precise recommendation and strategy management. It has the advantages of rich label expression, accurate intent recognition, high clustering accuracy, and strong system response intelligent linkage. It can effectively improve the accuracy of platform user portraits and the level of operational intelligence, and significantly enhance the e-commerce platform's user understanding and product strategy decision-making capabilities in a big data environment.

[0008] The user tagging management and demand analysis system based on big data e-commerce according to an embodiment of the present invention includes the following modules:

[0009] The acquisition module is used to collect multimodal behavior data of users on the e-commerce platform, perform time alignment and vectorization processing, and generate context-aware behavior representation vectors;

[0010] A label extraction module is used to construct a content feature channel, a behavior feature channel, and an emotion feature channel based on the context-aware behavior representation vector, and extract a multi-granularity label candidate set;

[0011] A graph construction module is used to construct a tag graph based on the multi-granularity tag candidate set, where the graph nodes are tags and the graph edges are generated based on behavior co-occurrence relationships, semantic similarity, and emotion correlation;

[0012] An intention reasoning module is used to input the label map and the user's current behavior representation into the intention reasoning network and output a user potential demand representation vector;

[0013] A clustering module, configured to input the user potential demand representation vector into an improved Grey Wolf demand clustering algorithm for cluster analysis to generate user demand group labels;

[0014] The structure output module is used to generate a user-label-demand ternary structure based on user demand group labels, and is used for product strategy management on e-commerce platforms.

[0015] The user tagging management and demand analysis method based on big data e-commerce according to an embodiment of the present invention includes the following steps:

[0016] S1. Collect multimodal behavior data of users on e-commerce platforms, perform time alignment and vectorization processing, and generate context-aware behavior representation vectors;

[0017] S2. Based on the context-aware behavior representation vector, construct a content feature channel, a behavior feature channel, and an emotion feature channel, and extract a multi-granularity label candidate set;

[0018] S3. Construct a tag graph based on the multi-granular tag candidate set, where the graph nodes are tags and the graph edges are constructed based on behavior co-occurrence relationships, semantic similarity, and emotion correlation;

[0019] S4. Based on the label graph and the user's current behavior representation, input the intention reasoning network and output the user's potential demand representation vector;

[0020] S5. Taking the user potential demand representation vector as input, performing cluster analysis using an improved Grey Wolf demand clustering algorithm to generate user demand group labels, wherein the improved Grey Wolf demand clustering algorithm includes a density-sensitive scheduling function and a local jump search mechanism;

[0021] S6. Based on the user demand group label, generate a user-label-demand ternary structure for use in product strategy management of the e-commerce platform.

[0022] Optionally, the multimodal behavior data includes the user's click behavior, search history, dwell time, purchase behavior, comment content, collection and add-to-cart behavior, operating device information and login geographic location.

[0023] Optionally, the S2 specifically includes:

[0024] S21, splitting the context-aware behavior representation vector into three categories: content, behavior, and emotion;

[0025] S22. Build a content feature channel to generate a content semantic embedding vector based on the product titles, descriptions, and category texts involved in user browsing, searching, and collecting behaviors;

[0026] S23. Build a behavior feature channel, extract user click frequency, stay duration, operation path sequence, and behavior transition probability, and generate a user behavior pattern vector;

[0027] S24. Construct an emotional feature channel, extract emotional polarity labels based on user comment text and search statements, and quantify them into emotional embedding vectors;

[0028] S25, fusing the output vectors of the content feature channel, the behavior feature channel, and the emotion feature channel, and using an attention mechanism to weightedly integrate multi-channel features to form a user multimodal label generation representation;

[0029] S26 , performing multi-granularity tag mapping on the user multimodal tag generation representation, selecting corresponding tags in a predefined tag space based on content themes, behavioral preferences, and emotional tendencies, to form a multi-granularity tag candidate set.

[0030] Optionally, the S3 specifically includes:

[0031] S31, taking the multi-granularity label candidate set as a node set of the label graph, where each node corresponds to a label;

[0032] S32. Traverse the user's historical behavior data and count the co-occurrence relationships of tags in the user's behavior sequence. If two tags appear in the same user's continuous behavior window with a frequency greater than a set frequency threshold, establish a behavior co-occurrence connection edge between the tags and set the initial weight of the edge based on the co-occurrence frequency and the behavior time interval;

[0033] S33, encoding the text semantics corresponding to each label, calculating the semantic similarity between the labels, and establishing a semantic connection edge when the distance between two labels in the semantic space is less than a set threshold;

[0034] S34. Extract the emotional polarity contained in user comments, evaluation texts, and search statements, and determine the emotional consistency in tag association behavior. If two tags are accompanied by the same emotional response in more than 50% of users, establish an emotional empathy connection edge.

[0035] S35. Unify and fuse the behavioral co-occurrence connection edges, semantic connection edges, and emotional empathy connection edges, assign different weight coefficients to each of them using a weighting mechanism, merge them to generate a complete multi-dimensional label graph edge weight structure, and finally output the label graph.

[0036] Optionally, the S4 specifically includes:

[0037] S41. Construct a representation vector of the user's current behavior, aggregate the user's recent behavior sequence through the attention mechanism, and obtain the contextual feature representation of the user's behavior state;

[0038] S42. Construct an intention inference network, which includes a state perception layer, a behavior prediction layer, and a decision layer. The state perception layer receives the context feature representation of the user's behavior state and the tag graph node features, and fuses them to form a comprehensive state feature.

[0039] S43, the behavior prediction layer adopts a multi-layer perceptron structure to receive the comprehensive state features and generate candidate demand representations;

[0040] S44, the decision layer is based on the Actor-Critic network, iteratively trained through the policy gradient method, and uses the user's real feedback as a reward signal to make an execution intention selection decision on the candidate demand representation;

[0041] S45. Output the optimal intention representation determined by the Actor-Critic network as the user's potential demand representation vector.

[0042] Optionally, the S5 specifically includes:

[0043] S51, the user's potential demand representation vector constitutes a feature set X={x1,x2,…,x N}, where x N represents the potential demand representation vector of the Nth user, N represents the total number of users, and the number of cluster categories is set to K;

[0044] S52, initialize the search population of the improved gray wolf clustering algorithm, each individual in the search population corresponds to a set of cluster centers to be optimized C = {c1, c2, ..., c K}, c K represents the Kth cluster center, and the roles are divided according to the individual fitness ranking, including main wolf, guide wolf, follower wolf and ordinary search wolf;

[0045] S53. Build density-sensitive scheduling function:

[0046]

[0047] Among them, ρ(x i ) represents the density of the potential demand vector of the i-th user, exp represents the exponential function, x i represents the potential demand representation vector of the i-th user, x j represents the potential demand representation vector of the jth user, σ represents the kernel width parameter, and ||·|| represents the Euclidean distance;

[0048] S54. The common search wolf updates the center according to the cluster centers of the main wolf, the guide wolf, and the follower wolf:

[0049]

[0050] in, represents the position of the kth cluster center in the t+1 generation, represents the kth cluster center vector solution corresponding to the main wolf in the tth generation, represents the kth cluster center vector solution corresponding to the guiding wolf in the tth generation, represents the kth cluster center vector solution corresponding to the wolf in the tth generation, ρ max Represents the maximum density of all users’ potential demand representation vectors, ∈ k represents the offset vector;

[0051] S55. In the final stage of each generation, with a preset probability P jump Select some individuals from the ordinary search wolves to trigger the local jump search mechanism. The update strategy is:

[0052]

[0053] in, represents the position of the kth cluster center in the tth generation, η represents the jump amplitude adjustment coefficient, sign(rand-0.5) represents the sign function of the jump direction, rand represents a random number ranging from 0 to 1, and τ represents the disturbance amplitude;

[0054] S56. The fitness of each individual in the population after each round of iteration is:

[0055]

[0056] Among them, F i represents the fitness of the potential demand representation vector of the i-th user, C k represents the potential demand representation vector of all users in the kth cluster center, c k represents the kth cluster center, λ represents the density balance adjustment parameter, Var(ρ(C k )) represents the density variance of all user potential demand representation vectors in the k-th cluster center;

[0057] S57 , repeating steps S54 - S56 until any convergence condition is met or the maximum number of iterations is reached, and finally outputting the cluster label to which each user's demand belongs as the user demand group label.

[0058] Optionally, the S52 specifically includes:

[0059] S521. In each iteration, the individuals in the population are sorted based on the fitness corresponding to all current gray wolf individuals;

[0060] S522: Select the individual with the highest fitness value as the master wolf, and the corresponding cluster center as the current optimal clustering solution;

[0061] S523: Select the individual with the second highest fitness as the guide wolf, select the individual with the third highest fitness as the follower wolf, and mark the remaining individuals as ordinary search wolves. The cluster centers of the guide wolf and the follower wolf are used to guide the update direction of the ordinary search wolves.

[0062] S524. After each generation of iteration, the fitness of all individuals is re-evaluated and the roles are updated to achieve dynamic changes in role division.

[0063] Optionally, the convergence condition specifically includes: the maximum norm of the change in the cluster center position is less than a set change threshold and the change in the optimal value of the individual fitness is less than a set accuracy threshold.

[0064] Optionally, the S6 specifically includes:

[0065] S61. Based on the obtained user demand group label, bind each user to the corresponding user demand group label to form an initial user-demand mapping relationship;

[0066] S62: Determine the most relevant core tag node of the user in the tag graph to form a user-tag association structure;

[0067] S63: Associating the user, label, and requirement information based on the initial mapping relationship between user and requirement and the association structure between user and label to generate a complete user-label-requirement ternary structure;

[0068] S64, analyzing the user-label-demand ternary structure, and performing differential analysis on the demand preferences of different groups;

[0069] S65. Based on the user-label-demand ternary structure and the difference analysis results, provide customized labels and demand information for the product recommendation engine of the e-commerce platform for product strategy management of the e-commerce platform.

[0070] The beneficial effects of the present invention are:

[0071] First, this invention introduces a multimodal behavior collection and feature channel division mechanism, uniformly aligning and vectorizing the multidimensional user behaviors on e-commerce platforms, including clicks, searches, comments, add-to-cart, device, and geographic information. This effectively overcomes the limitations of traditional systems that model user behavior in a single, time-sensitive manner. By constructing content feature channels, behavioral feature channels, and emotional feature channels, the system fully extracts and integrates fine-grained user features across the three dimensions of product text, behavioral path, and emotional expression, significantly improving the expressiveness and diversity of user tags.

[0072] Secondly, by constructing a multidimensional tag graph that integrates behavioral co-occurrence, semantic similarity, and sentiment, we achieve structural modeling and semantic propagation of tag relationships, effectively addressing the issues of isolated tags and missing graph structure in existing technologies. The generated tag graph not only provides static tag mapping capabilities but also reflects changes in user interests through the evolution of the graph structure, facilitating the dynamic evolution and contextual adaptation of the tag system.

[0073] Furthermore, this invention introduces an intent-based reasoning network to fuse label graphs with users' current behaviors to generate a vector representing their potential needs, thereby enhancing the system's ability to identify users' underlying preferences and potential intentions. The intent-based reasoning network employs a three-layer architecture consisting of state perception, predictive reasoning, and strategic decision-making. Combined with reinforcement learning, it enables continuous self-optimization based on historical behavior feedback, providing higher-quality input representations for subsequent demand clustering.

[0074] Furthermore, the improved Grey Wolf demand clustering algorithm proposed in this paper builds on the traditional Grey Wolf optimization by introducing a density-sensitive scheduling function and a local jump search mechanism. This addresses the problems of existing clustering methods, such as unclear boundary identification and susceptibility to local optima when dealing with high-dimensional user demand representations. The algorithm adaptively adjusts the guidance weights of samples in different density areas and dynamically jumps within local regions to search for cluster centers, significantly improving the stability and discriminability of clustering results, thereby achieving more accurate user group segmentation.

[0075] Finally, by outputting the user-tag-demand ternary structure, the system establishes a logical closed loop between user profiling, interest expression, and product strategy, truly achieving intelligent linkage from perception to analysis, from understanding to application. This structure not only supports the strategic deployment of personalized recommendations and refined marketing, but also provides data-driven support for product management and supply decision-making on the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0077] Figure 1 This is a schematic diagram of the structure of the user tagging management and demand analysis system based on big data e-commerce proposed by the present invention;

[0078] Figure 2 This is a flow chart of the user tagging management and demand analysis method based on big data e-commerce proposed by the present invention. DETAILED DESCRIPTION

[0079] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0080] refer to Figure 1 , a user tagging management and demand analysis system based on big data e-commerce, including the following modules:

[0081] The acquisition module is used to collect multimodal behavior data of users on the e-commerce platform, perform time alignment and vectorization processing, and generate context-aware behavior representation vectors;

[0082] A label extraction module is used to construct a content feature channel, a behavior feature channel, and an emotion feature channel based on the context-aware behavior representation vector, and extract a multi-granularity label candidate set;

[0083] A graph construction module is used to construct a tag graph based on the multi-granularity tag candidate set, where the graph nodes are tags and the graph edges are generated based on behavior co-occurrence relationships, semantic similarity, and emotion correlation;

[0084] An intention reasoning module is used to input the label map and the user's current behavior representation into the intention reasoning network and output a user potential demand representation vector;

[0085] A clustering module, configured to input the user potential demand representation vector into an improved Grey Wolf demand clustering algorithm for cluster analysis to generate user demand group labels;

[0086] The structure output module is used to generate a user-label-demand ternary structure based on user demand group labels, and is used for product strategy management on e-commerce platforms.

[0087] Through a modular system architecture, key functions such as multimodal behavior collection, label extraction, graph construction, intent reasoning, intelligent clustering, and strategy output are integrated into the same system, forming a complete user labeling and demand analysis closed loop for e-commerce platforms. The system not only has a clear structure of the data processing flow, but also realizes the linkage of the entire process from data perception to decision feedback at the system design level. Especially after the introduction of the improved gray wolf clustering algorithm and intent reasoning network, the system's identification of user needs is more accurate, dynamic, and predictive, breaking through the limitations of traditional recommendation models that rely on explicit behavior or static labels. The overall system has the significant advantages of strong real-time performance, high intelligence, and strong label semantic expression, providing strong technical support for improving the recommendation accuracy, user experience, and product strategy efficiency of e-commerce platforms.

[0088] refer to Figure 2 The user tagging management and demand analysis method based on big data e-commerce includes the following steps:

[0089] S1. Collect multimodal behavior data of users on e-commerce platforms, perform time alignment and vectorization processing, and generate context-aware behavior representation vectors;

[0090] S2. Based on the context-aware behavior representation vector, construct a content feature channel, a behavior feature channel, and an emotion feature channel, and extract a multi-granularity label candidate set;

[0091] S3. Construct a tag graph based on the multi-granular tag candidate set, where the graph nodes are tags and the graph edges are constructed based on behavior co-occurrence relationships, semantic similarity, and emotion correlation;

[0092] S4. Based on the label graph and the user's current behavior representation, input the intention reasoning network and output the user's potential demand representation vector;

[0093] S5. Taking the user potential demand representation vector as input, performing cluster analysis using an improved Grey Wolf demand clustering algorithm to generate user demand group labels, wherein the improved Grey Wolf demand clustering algorithm includes a density-sensitive scheduling function and a local jump search mechanism;

[0094] S6. Based on the user demand group label, generate a user-label-demand ternary structure for use in product strategy management of the e-commerce platform.

[0095] The present invention clarifies the execution process between each functional module in the system, standardizes the full-process operation steps of the system from data collection to label structure output, and makes the method highly operable and engineering implementable. Through this process division, the structured transformation of behavioral data, the semantic fusion of label maps, the deep modeling of user intentions and efficient clustering division can be realized. Each step is progressively developed based on the results of the previous stage, which enhances the logical consistency and processing efficiency of data flow. At the same time, this claim also provides a standard process framework for the technical refinement of subsequent dependent claims, ensuring the high stability of system operation and the ability to coordinate and optimize each link, reflecting good system integration and expansion potential.

[0096] In this embodiment, the multimodal behavior data includes the user's click behavior, search history, dwell time, purchase behavior, comment content, collection and add-to-cart behavior, operating device information and login geographic location.

[0097] In this embodiment, S2 specifically includes:

[0098] S21, splitting the context-aware behavior representation vector into three categories: content, behavior, and emotion;

[0099] S22. Build a content feature channel to generate a content semantic embedding vector based on the product titles, descriptions, and category texts involved in user browsing, searching, and collecting behaviors;

[0100] S23. Build a behavior feature channel, extract user click frequency, stay duration, operation path sequence, and behavior transition probability, and generate a user behavior pattern vector;

[0101] S24. Construct an emotional feature channel, extract emotional polarity labels based on user comment text and search statements, and quantify them into emotional embedding vectors;

[0102] S25, fusing the output vectors of the content feature channel, the behavior feature channel, and the emotion feature channel, and using an attention mechanism to weightedly integrate multi-channel features to form a user multimodal label generation representation;

[0103] S26 , performing multi-granularity tag mapping on the user multimodal tag generation representation, selecting corresponding tags in a predefined tag space based on content themes, behavioral preferences, and emotional tendencies, to form a multi-granularity tag candidate set.

[0104] This step solves the problems of insufficient semantic coverage and context failure caused by the single channel of traditional tagging systems by introducing a three-channel tag extraction mechanism. The system analyzes user behavior from three perspectives: content, behavior, and emotion, and integrates them into a multi-granularity tag representation, significantly enhancing the semantic hierarchy and behavior-driven adaptability of the tags. In particular, the introduction of the emotional channel enables the subjective preferences and emotional reactions expressed by users in natural language to be effectively extracted, thereby enriching the emotional dimension and personality characteristics of the tagging system. This fusion tag generation strategy enhances the multi-dimensional interpretation ability of tags, making subsequent graph modeling and recommendation systems more targeted and in-depth in user understanding.

[0105] In this embodiment, S3 specifically includes:

[0106] S31, taking the multi-granularity label candidate set as a node set of the label graph, where each node corresponds to a label;

[0107] S32. Traverse the user's historical behavior data and count the co-occurrence relationships of tags in the user's behavior sequence. If two tags appear in the same user's continuous behavior window with a frequency greater than a set frequency threshold, establish a behavior co-occurrence connection edge between the tags and set the initial weight of the edge based on the co-occurrence frequency and the behavior time interval;

[0108] S33, encoding the text semantics corresponding to each label, calculating the semantic similarity between the labels, and establishing a semantic connection edge when the distance between two labels in the semantic space is less than a set threshold;

[0109] S34. Extract the emotional polarity contained in user comments, evaluation texts, and search statements, and determine the emotional consistency in tag association behavior. If two tags are accompanied by the same emotional response in more than 50% of users, establish an emotional empathy connection edge.

[0110] S35. Unify and fuse the behavioral co-occurrence connection edges, semantic connection edges, and emotional empathy connection edges, assign different weight coefficients to each of them using a weighting mechanism, merge them to generate a complete multi-dimensional label graph edge weight structure, and finally output the label graph.

[0111] The tag graph structure constructed in this step effectively solves the problem of lack of structured expression and semantic connection between existing tag systems. By introducing three types of edge connections: behavioral co-occurrence, semantic similarity, and emotional consistency, a multi-dimensional coupling relationship is established between tags, forming a graph structure with communication capabilities and semantic retention. This graph can not only be used for tag similarity calculation, but also provides a structured basis for the evolution of user interests, thereby supporting dynamic recommendations and preference migration analysis. The edge weight fusion mechanism in the graph realizes the unified scheduling of associations of different dimensions, enhances the expression stability and reasoning logic of the tag graph, and provides a solid foundation for the subsequent embedding of graph neural networks and user modeling.

[0112] In this embodiment, the S4 specifically includes:

[0113] S41. Construct a representation vector of the user's current behavior, aggregate the user's recent behavior sequence through the attention mechanism, and obtain the contextual feature representation of the user's behavior state;

[0114] S42. Construct an intention inference network, which includes a state perception layer, a behavior prediction layer, and a decision layer. The state perception layer receives the context feature representation of the user's behavior state and the tag graph node features, and fuses them to form a comprehensive state feature.

[0115] S43, the behavior prediction layer adopts a multi-layer perceptron structure to receive the comprehensive state features and generate candidate demand representations;

[0116] S44, the decision layer is based on the Actor-Critic network, iteratively trained through the policy gradient method, and uses the user's real feedback as a reward signal to make an execution intention selection decision on the candidate demand representation;

[0117] S45. Output the optimal intention representation determined by the Actor-Critic network as the user's potential demand representation vector.

[0118] The intent reasoning module significantly improves the system's ability to model users' deep intentions by constructing a three-layer network structure consisting of state perception, prediction generation, and policy decision-making. This module not only integrates the current behavioral context and label graph embedding, but also uses the Actor-Critic mechanism to implement a self-learning strategy based on behavioral feedback, making the user's intent reasoning results more forward-looking and adjustable. Compared to traditional recommendation models that rely solely on historical behavior to predict explicit interests, this structure can actively learn user interest transfer trends and predict potential needs that are not explicitly expressed, enhancing the system's intelligent and adaptive recommendation capabilities and improving the platform's core competitiveness in precise recommendations and intelligent marketing.

[0119] In this embodiment, the S5 specifically includes:

[0120] S51, the user's potential demand representation vector constitutes a feature set X={x1,x2,…,x N}, where x N represents the potential demand representation vector of the Nth user, N represents the total number of users, and the number of cluster categories is set to K;

[0121] S52, initialize the search population of the improved gray wolf clustering algorithm, each individual in the search population corresponds to a set of cluster centers to be optimized C = {c1, c2, ..., c K}, c K represents the Kth cluster center, and the roles are divided according to the individual fitness ranking, including main wolf, guide wolf, follower wolf and ordinary search wolf;

[0122] S53. Build density-sensitive scheduling function:

[0123]

[0124] Among them, ρ(x i ) represents the density of the potential demand vector of the i-th user, exp represents the exponential function, x i represents the potential demand representation vector of the i-th user, x j represents the potential demand representation vector of the jth user, σ represents the kernel width parameter, and ||·|| represents the Euclidean distance;

[0125] S54. The common search wolf updates the center according to the cluster centers of the main wolf, the guide wolf, and the follower wolf:

[0126]

[0127] in, represents the position of the kth cluster center in the t+1 generation, represents the kth cluster center vector solution corresponding to the main wolf in the tth generation, represents the kth cluster center vector solution corresponding to the guiding wolf in the tth generation, represents the kth cluster center vector solution corresponding to the wolf in the tth generation, ρ max Represents the maximum density of all users’ potential demand representation vectors, ∈ k represents the offset vector;

[0128] S55. In the final stage of each generation, with a preset probability P jump Select some individuals from the ordinary search wolves to trigger the local jump search mechanism. The update strategy is:

[0129]

[0130] in, represents the position of the kth cluster center in the tth generation, η represents the jump amplitude adjustment coefficient, sign(rand-0.5) represents the sign function of the jump direction, rand represents a random number ranging from 0 to 1, and τ represents the disturbance amplitude;

[0131] S56. The fitness of each individual in the population after each round of iteration is:

[0132]

[0133] Among them, F i represents the fitness of the potential demand representation vector of the i-th user, C k represents the potential demand representation vector of all users in the kth cluster center, c k represents the kth cluster center, λ represents the density balance adjustment parameter, Var(ρ(C k )) represents the density variance of all user potential demand representation vectors in the k-th cluster center;

[0134] S57 , repeating steps S54 - S56 until any convergence condition is met or the maximum number of iterations is reached, and finally outputting the cluster label to which each user's demand belongs as the user demand group label.

[0135] By introducing a density-sensitive scheduling function and a local jump search mechanism, the improved Grey Wolf clustering algorithm proposed in this step significantly improves the accuracy and stability of user demand group segmentation. The density-aware scheduling function effectively addresses the problem of mis-clustering that traditional clustering algorithms often encounter when dealing with uneven sample distributions, while the jump search mechanism enhances the search process's ability to escape local optima, improving the algorithm's global search performance in high-dimensional feature spaces. This method is particularly well-suited for real-world e-commerce environments where the potential demand space of users exhibits uneven density and fuzzy boundaries, providing a more robust foundation for precise recommendation and clustering strategies.

[0136] In this embodiment, the S52 specifically includes:

[0137] S521. In each iteration, the individuals in the population are sorted based on the fitness corresponding to all current gray wolf individuals;

[0138] S522: Select the individual with the highest fitness value as the master wolf, and the corresponding cluster center as the current optimal clustering solution;

[0139] S523: Select the individual with the second highest fitness as the guide wolf, select the individual with the third highest fitness as the follower wolf, and mark the remaining individuals as ordinary search wolves. The cluster centers of the guide wolf and the follower wolf are used to guide the update direction of the ordinary search wolves.

[0140] S524. After each generation of iteration, the fitness of all individuals is re-evaluated and the roles are updated to achieve dynamic changes in role division.

[0141] By clearly defining the roles of leader, guide, and follower wolves within the clustering algorithm and implementing a dynamic update mechanism, the hierarchical control capabilities of the gray wolf optimization structure were enhanced. This role division provides the optimization strategy with a clear guiding direction and diverse update paths, improving the algorithm's search efficiency and the flexibility of cluster center updates. Furthermore, the dynamic evaluation and role reallocation mechanism avoids the rigidity of the population structure, helping to adapt to search needs at different stages and overall improving the clustering algorithm's adaptability and optimization performance.

[0142] In this embodiment, the convergence conditions specifically include: the maximum norm of the change of the cluster center position is less than a set change threshold and the change in the optimal value of the individual fitness is less than a set accuracy threshold.

[0143] By implementing a dual convergence judgment mechanism, the convergence efficiency and stability of the clustering process are improved at the algorithm control level. By setting both cluster center position change thresholds and fitness change thresholds, the system effectively terminates unnecessary iterations while ensuring clustering quality, saving computing resources and improving real-time performance. This mechanism balances optimization quality and system efficiency, making it suitable for e-commerce platforms where both clustering stability and responsiveness are crucial, demonstrating excellent engineering controllability.

[0144] In this embodiment, S6 specifically includes:

[0145] S61. Based on the obtained user demand group label, bind each user to the corresponding user demand group label to form an initial user-demand mapping relationship;

[0146] S62: Determine the most relevant core tag node of the user in the tag graph to form a user-tag association structure;

[0147] S63: Associating the user, label, and requirement information based on the initial mapping relationship between user and requirement and the association structure between user and label to generate a complete user-label-requirement ternary structure;

[0148] S64, analyzing the user-label-demand ternary structure, and performing differential analysis on the demand preferences of different groups;

[0149] S65. Based on the user-label-demand ternary structure and the difference analysis results, provide customized labels and demand information for the product recommendation engine of the e-commerce platform for product strategy management of the e-commerce platform.

[0150] By constructing a user-tag-demand ternary structure, a closed-loop mechanism is established, linking user interest expressions with product strategy. This structure not only integrates user behavioral preferences, tag semantic expression, and group demand clustering results, but also serves multiple business modules such as product recommendations, marketing strategy development, and user tier management, achieving full-chain utilization of the platform's data assets. It has significant practical application value in precise recommendations, personalized operations, and supply-demand matching, providing a structural foundation for the intelligent upgrade of e-commerce platforms.

[0151] Example 1:

[0152] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the data analysis and intelligent recommendation system of a large-scale national comprehensive e-commerce platform, and active user behavior data over a three-month period were selected as test samples. The total amount of data exceeded 48 million, involving approximately 570,000 daily active users, approximately 2.1 million products, and approximately 130 million cumulative comment texts, constituting a typical e-commerce behavior data scenario with multimodality, large volume, and strong temporal sequence.

[0153] The platform previously employed a recommendation system based on static tags, with user profiles primarily based on registration information and shopping history, and product recommendations executed through a collaborative filtering algorithm. Due to the single tag and delayed updates, the recommendation results deviated significantly from the user's current interests, with the click-through rate (CTR) remaining at around 6.8% for a long time and the conversion rate (CVR) less than 1.5%. Especially with the ever-expanding product pool and complex and ever-changing user behavior, traditional methods struggled to quickly respond to changes in interests, and the system lacked the ability to deeply understand them. This resulted in frequent problems such as cold starts and misplaced recommendations, severely restricting the platform's recommendation efficiency and user experience.

[0154] In the implementation scheme of the present invention, the platform first deploys a multimodal behavior acquisition module to time-align and vectorize the user's click stream, search terms, purchase behavior, comment text, geographic location, device ID and other data to generate a context-aware behavior vector. On this basis, a three-channel label generation module for content, behavior and emotion is constructed. For example, a user searches for smart bracelets and heart rate monitoring within a week, browses and purchases the Huawei Band 9, and expresses emotional tendencies such as accuracy and long battery life when commenting on other health devices. The system will identify the user's potential interests in multiple label dimensions such as smart health wearable devices, function-oriented users, and positive emotion-driven users.

[0155] Subsequently, by constructing a tag graph, establishing behavioral co-occurrence relationships, semantic relevance, and sentiment consistency between tags, and using an intent inference network to combine the graph with current behavior, the system dynamically generates a vector of the user's potential needs. For this user, the system further identifies that their interest in health monitoring products has evolved from initial browsing to a high level of purchase intent, and further determines that their next product interests will be high-end smartwatches and multi-functional health devices.

[0156] The potential demand vectors of all users were subjected to high-dimensional clustering analysis using an improved Grey Wolf clustering algorithm. This algorithm, by incorporating a density-sensitive scheduling function and a localized jump search mechanism into the traditional Grey Wolf optimization framework, effectively overcomes the issues of uneven density in user feature spaces and blurred interest boundaries. The resulting clustering outputs categorized user demand groups into 32 groups, with clear themes and behavioral orientations evident, such as those focused on sports and health, those focused on improving work-from-home living, and those focused on short-term gifting.

[0157] The platform synchronized the user-tag-demand triad into its product recommendation engine and operational strategy system. During the Spring Festival, it promoted holiday gift boxes, brand-co-branded packages, and other related products to a group focused on short-term gift-giving. This boosted the recommendation click-through rate to 12.6% and the conversion rate to 3.2%. Furthermore, by optimizing product display placement based on the tag characteristics of users who were working from home, sales of high-priced office equipment increased by 22.1%.

[0158] Data results show that after the implementation of this invention, the average user stay time increased by 31%, the product exposure matching degree increased by about 37%, the system response time was controlled within 2.1 seconds, and the overall order completion rate of the platform increased by 14.7% year-on-year, which is significantly better than the traditional tag recommendation system. Substantial breakthroughs have been achieved in the depth of user behavior understanding, demand prediction accuracy, and recommendation strategy response efficiency.

[0159] This embodiment verifies that the present invention has extremely high practicality and promotion value in a real e-commerce platform environment, especially in key technical dimensions such as label system expression ability, user intent modeling and demand clustering accuracy, achieving a cross-generational leap, and providing a practical technical path for intelligent upgrades in big data e-commerce scenarios.

[0160] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A user tagging management and demand analysis system based on big data e-commerce, characterized by: Includes the following modules: The acquisition module is used to collect multimodal behavior data of users on the e-commerce platform, perform time alignment and vectorization processing, and generate context-aware behavior representation vectors; A label extraction module is used to construct a content feature channel, a behavior feature channel, and an emotion feature channel based on the context-aware behavior representation vector, and extract a multi-granularity label candidate set; A graph construction module is used to construct a tag graph based on the multi-granularity tag candidate set, where the graph nodes are tags and the graph edges are generated based on behavior co-occurrence relationships, semantic similarity, and emotion correlation; An intention reasoning module is used to input the label map and the user's current behavior representation into the intention reasoning network and output a user potential demand representation vector; A clustering module, configured to input the user potential demand representation vector into an improved Grey Wolf demand clustering algorithm for cluster analysis to generate user demand group labels; The structure output module is used to generate a user-label-demand ternary structure based on user demand group labels, and is used for product strategy management on e-commerce platforms.

2. The user tagging management and demand analysis system based on big data e-commerce according to claim 1 is characterized in that: The modules are implemented as follows: S1. Collect multimodal behavior data of users on e-commerce platforms, perform time alignment and vectorization processing, and generate context-aware behavior representation vectors; S2. Based on the context-aware behavior representation vector, construct a content feature channel, a behavior feature channel, and an emotion feature channel, and extract a multi-granularity label candidate set; S3. Construct a tag graph based on the multi-granular tag candidate set, where the graph nodes are tags and the graph edges are constructed based on behavior co-occurrence relationships, semantic similarity, and emotion correlation; S4. Based on the label graph and the user's current behavior representation, input the intention reasoning network and output the user's potential demand representation vector; S5. Taking the user potential demand representation vector as input, performing cluster analysis using an improved Grey Wolf demand clustering algorithm to generate user demand group labels, wherein the improved Grey Wolf demand clustering algorithm includes a density-sensitive scheduling function and a local jump search mechanism; S6. Based on the user demand group label, generate a user-label-demand ternary structure for use in product strategy management of the e-commerce platform.

3. The user tagging management and demand analysis system based on big data e-commerce according to claim 2 is characterized in that: The multimodal behavior data includes the user's click behavior, search history, dwell time, purchase behavior, comment content, collection and add-to-cart behavior, operating device information and login geographic location.

4. The user tagging management and demand analysis system based on big data e-commerce according to claim 2 is characterized in that: The S2 specifically includes: S21, splitting the context-aware behavior representation vector into three categories: content, behavior, and emotion; S22. Build a content feature channel to generate a content semantic embedding vector based on the product titles, descriptions, and category texts involved in user browsing, searching, and collecting behaviors; S23. Build a behavior feature channel, extract user click frequency, stay duration, operation path sequence, and behavior transition probability, and generate a user behavior pattern vector; S24. Construct an emotional feature channel, extract emotional polarity labels based on user comment text and search statements, and quantify them into emotional embedding vectors; S25, fusing the output vectors of the content feature channel, the behavior feature channel, and the emotion feature channel, and using an attention mechanism to weightedly integrate multi-channel features to form a user multimodal label generation representation; S26 , performing multi-granularity tag mapping on the user multimodal tag generation representation, selecting corresponding tags in a predefined tag space based on content themes, behavioral preferences, and emotional tendencies, to form a multi-granularity tag candidate set.

5. The user tagging management and demand analysis system based on big data e-commerce according to claim 2 is characterized in that: The S3 specifically includes: S31, taking the multi-granularity label candidate set as a node set of the label graph, where each node corresponds to a label; S32. Traverse the user's historical behavior data and count the co-occurrence relationships of tags in the user's behavior sequence. If two tags appear in the same user's continuous behavior window with a frequency greater than a set frequency threshold, establish a behavior co-occurrence connection edge between the tags and set the initial weight of the edge based on the co-occurrence frequency and the behavior time interval; S33, encoding the text semantics corresponding to each label, calculating the semantic similarity between the labels, and establishing a semantic connection edge when the distance between two labels in the semantic space is less than a set threshold; S34. Extract the emotional polarity contained in user comments, evaluation texts, and search statements, and determine the emotional consistency in tag association behavior. If two tags are accompanied by the same emotional response in more than 50% of users, establish an emotional empathy connection edge. S35. Unify and fuse the behavioral co-occurrence connection edges, semantic connection edges, and emotional empathy connection edges, assign different weight coefficients to each of them using a weighting mechanism, merge them to generate a complete multi-dimensional label graph edge weight structure, and finally output the label graph.

6. The user tagging management and demand analysis system based on big data e-commerce according to claim 2 is characterized in that: The S4 specifically includes: S41. Construct a representation vector of the user's current behavior, aggregate the user's recent behavior sequence through the attention mechanism, and obtain the contextual feature representation of the user's behavior state; S42. Construct an intention inference network, which includes a state perception layer, a behavior prediction layer, and a decision layer. The state perception layer receives the context feature representation of the user's behavior state and the tag graph node features, and fuses them to form a comprehensive state feature. S43, the behavior prediction layer adopts a multi-layer perceptron structure to receive the comprehensive state features and generate candidate demand representations; S44, the decision layer is based on the Actor-Critic network, iteratively trained through the policy gradient method, and uses the user's real feedback as a reward signal to make an execution intention selection decision on the candidate demand representation; S45. Output the optimal intention representation determined by the Actor-Critic network as the user's potential demand representation vector.

7. The user tagging management and demand analysis system based on big data e-commerce according to claim 2 is characterized in that: The S5 specifically includes: S51, the user's potential demand representation vector constitutes a feature set X={x1,x2,…,x N }, where x N represents the potential demand representation vector of the Nth user, N represents the total number of users, and the number of cluster categories is set to K; S52, initialize the search population of the improved gray wolf clustering algorithm, each individual in the search population corresponds to a set of cluster centers to be optimized C = {c1, c2, ..., c K }, c K represents the Kth cluster center, and the roles are divided according to the individual fitness ranking, including main wolf, guide wolf, follower wolf and ordinary search wolf; S53. Build density-sensitive scheduling function: Among them, ρ(x i ) represents the density of the potential demand vector of the i-th user, exp represents the exponential function, x i represents the potential demand representation vector of the i-th user, x j represents the potential demand representation vector of the jth user, σ represents the kernel width parameter, and ||·|| represents the Euclidean distance; S54. The common search wolf updates the center according to the cluster centers of the main wolf, the guide wolf, and the follower wolf: in, represents the position of the kth cluster center in the t+1 generation, represents the kth cluster center vector solution corresponding to the main wolf in the tth generation, represents the kth cluster center vector solution corresponding to the guiding wolf in the tth generation, represents the kth cluster center vector solution corresponding to the wolf in the tth generation, ρ max Represents the maximum density of all users’ potential demand representation vectors, ∈ k represents the offset vector; S55. In the final stage of each generation, with a preset probability P jump Select some individuals from the ordinary search wolves to trigger the local jump search mechanism. The update strategy is: in, represents the position of the kth cluster center in the tth generation, η represents the jump amplitude adjustment coefficient, sign(rand-0.5) represents the sign function of the jump direction, rand represents a random number ranging from 0 to 1, and τ represents the disturbance amplitude; S56. The fitness of each individual in the population after each round of iteration is: Among them, F i represents the fitness of the potential demand representation vector of the i-th user, C k represents the potential demand representation vector of all users in the kth cluster center, c k represents the kth cluster center, λ represents the density balance adjustment parameter, Var(ρ(C k )) represents the density variance of all user potential demand representation vectors in the k-th cluster center; S57 , repeating steps S54 - S56 until any convergence condition is met or the maximum number of iterations is reached, and finally outputting the cluster label to which each user's demand belongs as the user demand group label.

8. The user tagging management and demand analysis system based on big data e-commerce according to claim 7 is characterized in that: The S52 specifically includes: S521. In each iteration, the individuals in the population are sorted based on the fitness corresponding to all current gray wolf individuals; S522: Select the individual with the highest fitness value as the master wolf, and the corresponding cluster center as the current optimal clustering solution; S523: Select the individual with the second highest fitness as the guide wolf, select the individual with the third highest fitness as the follower wolf, and mark the remaining individuals as ordinary search wolves. The cluster centers of the guide wolf and the follower wolf are used to guide the update direction of the ordinary search wolves. S524. After each generation of iteration, the fitness of all individuals is re-evaluated and the roles are updated to achieve dynamic changes in role division.

9. The user tagging management and demand analysis system based on big data e-commerce according to claim 7 is characterized in that: The convergence conditions specifically include: the maximum norm of the cluster center position change is less than a set change threshold and the change in the optimal value of the individual fitness is less than a set accuracy threshold.

10. The user tagging management and demand analysis system based on big data e-commerce according to claim 2 is characterized in that: The S6 specifically includes: S61. Based on the obtained user demand group label, bind each user to the corresponding user demand group label to form an initial user-demand mapping relationship; S62: Determine the most relevant core tag node of the user in the tag graph to form a user-tag association structure; S63: Associating the user, label, and requirement information based on the initial mapping relationship between user and requirement and the association structure between user and label to generate a complete user-label-requirement ternary structure; S64, analyzing the user-label-demand ternary structure, and performing differential analysis on the demand preferences of different groups; S65. Based on the user-label-demand ternary structure and the difference analysis results, provide customized labels and demand information for the product recommendation engine of the e-commerce platform for product strategy management of the e-commerce platform.

Images