User clustering method and system based on behavior path analysis and computer equipment

Abstract

The invention is suitable for the technical field of data processing, and provides a user clustering method and system based on behavior path analysis and computer equipment, and the method comprises the steps: collecting interaction behavior data of users, and constructing a behavior path for each user; analyzing a preset feature of the behavior path, and adaptively selecting a matched first or second clustering process for the behavior path according to an analysis result of the preset feature; when the clustering process is a first clustering process, dividing the users into different user clusters by iteratively pruning the features and repeatedly executing the clustering process on the basis of the subsequence features which are formed by continuous atomic behaviors and extracted from a behavior path; and when the process is a second clustering process, dividing the users into different user clusters based on a path semantic vector generated by performing aggregation operation on the low-dimensional embedded vector corresponding to each atomic behavior in the behavior path. The problems that an existing clustering model is single, feature expression is shallow, and a clustering result is difficult to explain and apply are solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a user clustering method, system, and computer device based on behavioral path analysis. Background Technology

[0002] With the rapid development of mobile internet and information technology, information flow products (such as news apps, short video applications, and social platforms) have become the main channels for users to obtain information. In this process, platforms have accumulated massive amounts of high-dimensional user behavior data. How to effectively analyze and utilize this data to accurately profile and segment users, and then provide personalized recommendations, optimize operational strategies, and improve user experience, has become a core technological challenge in the information flow field.

[0003] User clustering, as an important data mining technique, is widely used in information flow user analysis. Its core idea is to group users with similar behavioral patterns into the same group (cluster), thereby revealing the potential needs and preferences of different user groups. Currently, there are various algorithms for information flow user clustering, such as K-Means, DBSCAN (Density-Based Noise-Based Applied Spatial Clustering), and agglomerative hierarchical clustering. These algorithms have proven effective in specific scenarios: K-Means is widely used for processing large-scale datasets due to its simplicity and high computational efficiency; DBSCAN can discover clusters of arbitrary shapes and has good robustness to noisy data. In recent years, some methods have also attempted to combine deep learning techniques with traditional clustering algorithms to improve the accuracy and efficiency of clustering.

[0004] However, existing clustering methods based on information flow user behavior still face many challenges and shortcomings in practical applications:

[0005] Insufficient feature representation: Traditional clustering methods rely heavily on static statistical features of users (such as click-through rate and usage time), neglecting the dynamic and temporal nature of user behavior. User behavior is often a continuous, context-dependent sequence (i.e., a "user behavior path"), such as "browsing news -> clicking related recommendations -> watching videos -> posting comments." Using only statistical features will lose the deep logic and intent contained in the behavioral sequence, resulting in coarse-grained clustering results. At the same time, standard clustering algorithms are easily dominated by the common behavioral patterns of most users (such as watching short videos), causing some small but highly valuable niche user groups (such as deep consumers of content in specific fields) to be submerged and unidentifiable.

[0006] Data sparsity and high dimensionality issues: User behavior data in news feeds is inherently sparse, meaning that the interactions of a single user (such as clicks, likes, and comments) are very limited compared to the massive content library. This makes it difficult to accurately calculate the similarity between users. Simultaneously, as content categories and user behavior types become increasingly diverse, the dimensionality of user features increases dramatically. This not only increases computational complexity but also easily leads to the "curse of dimensionality," where the distance between any two points tends to be equal in high-dimensional space, rendering distance metrics meaningless and severely impacting clustering results.

[0007] The challenges of algorithm selection and parameter setting: Different clustering algorithms make different assumptions about data distribution and clustering objectives, making the selection of the most suitable algorithm a challenge in practical applications. More importantly, many mainstream algorithms (such as K-Means and DBSCAN) rely on preset key parameters, such as the number of clusters K in K-Means, and the neighborhood radius eps and minimum neighbor count minPts for core objects in DBSCAN. The setting of these parameters has a decisive impact on the quality of clustering results, but in practice, effective guidance methods are often lacking, requiring repeated trials based on human experience. This is not only inefficient, but unreasonable parameter settings can also lead to poor clustering results or distorted outcomes.

[0008] Poor interpretability and limited applicability of the results: Even if traditional algorithms can generate clustering results, these results often lack intuitive interpretability. Even if operations personnel or analysts see user groups, it's difficult to understand the specific behavioral patterns and underlying logic of each group—that is, "why these users are grouped together." This makes it difficult to translate clustering results into concrete, actionable operational strategies. For example, clustering results may be too coarse and unbalanced. For instance, it might find that a group "likes videos and text / images" accounts for more than 50% of users, but it's unclear whether a particular group "likes watching entertainment news after dinner" or "likes browsing sports videos late at night." Or, it might not be sensitive enough to dynamic changes in user behavior. For example, it might find that the activity of a large category of "video and text / image preference users" is increasing, but it's impossible to further segment their consumption scenarios and content preferences. This makes it difficult to promptly identify and locate subgroups with increased activity on a particular interest point, ultimately making clustering analysis unable to effectively empower business decisions and hindering refined operations.

[0009] In summary, existing information flow user clustering methods have significant shortcomings in feature representation capabilities, handling sparse high-dimensional data, addressing parameter sensitivity, and providing interpretable and practical analytical results. Therefore, a new technical solution is urgently needed to more effectively process user behavior sequence data, overcome the limitations of traditional methods, and achieve more accurate and insightful user segmentation. Summary of the Invention

[0010] Therefore, the purpose of this invention is to provide a user clustering method, system, and computer device based on behavioral path analysis, so as to fundamentally solve the problems of existing clustering models being too simple, having shallow feature representations, and having clustering results that are difficult to interpret and apply.

[0011] According to an embodiment of the present invention, a user clustering method based on behavioral path analysis is provided, the method comprising:

[0012] Collect user interaction behavior data and construct a behavior path for each user consisting of atomic behaviors arranged in chronological order;

[0013] Analyze at least one preset feature of the behavior path, and adaptively select a matching first clustering process or a second clustering process for the behavior path based on the analysis results of the preset features;

[0014] When the first clustering process is selected, based on the subsequence features composed of continuous atomic behaviors extracted from the behavior path, users are divided into different user clusters through an iterative feature pruning process and repeated clustering to achieve multi-level user segmentation.

[0015] When the second clustering process is selected, users are divided into different user clusters based on the path semantic vector generated by the aggregation operation of the low-dimensional embedding vectors corresponding to each atomic behavior in the behavior path.

[0016] In addition, the user clustering method based on behavior path analysis according to the above embodiments of the present invention may also have the following additional technical features:

[0017] Furthermore, the step of iteratively pruning features and repeatedly performing clustering based on the subsequence features extracted from the behavioral path, which consist of continuous atomic behaviors, includes:

[0018] The behavior path is traversed using a sliding window of a preset length to extract all subsequence features of continuous atomic behaviors that conform to the preset length, thus forming an initial feature subspace.

[0019] Within the feature subspace used in the current round, user behavior paths are clustered to obtain the user clusters for this round;

[0020] Identify and remove statistically significant subsequence features that define the user cluster in this round, in order to form a new, pruned feature subspace for the next iteration;

[0021] Determine whether the preset iteration termination condition is met. If not, use the new feature subspace as the feature subspace used in the current round and return to execute the step of clustering user behavior paths.

[0022] Furthermore, the step of identifying and removing statistically significant subsequence features that define the user cluster in this round includes:

[0023] For each user cluster generated in this round of clustering, a contingency table is constructed to describe the distribution of features of each subsequence inside and outside the user cluster;

[0024] Calculate the chi-square statistic of each subsequence feature relative to the user cluster based on the contingency table;

[0025] Identify and select the subsequence features with the highest global chi-square statistic of a preset number as salient features that need to be removed from the current feature space.

[0026] Furthermore, the step of dividing users into different user clusters based on the path semantic vector generated by aggregating the low-dimensional embedding vectors corresponding to each atomic behavior in the behavioral path includes:

[0027] Obtain the low-dimensional embedding vector corresponding to each atomic behavior in the behavior path;

[0028] Calculate the weight of each atomic behavior in the behavior path, whereby the weight is determined by the importance of the atomic behavior in the current behavior path and its rarity in all user behavior paths.

[0029] Multiply the low-dimensional embedding vector of each atomic behavior by its corresponding weight to obtain a weighted vector;

[0030] An aggregation operation is performed to sum and normalize all weighted vectors in the behavior path to generate the path semantic vector, and users are divided into different user clusters based on the path semantic vector.

[0031] Furthermore, the step of calculating the weight of each atomic behavior in the behavioral path includes:

[0032] The frequency of the atomic behavior occurring in the current behavior path is obtained as a measure of its importance;

[0033] Obtain the total number of user paths containing the atomic behavior, and calculate the inverse frequency value based on the total number of user paths, as a measure of the rarity;

[0034] The final weight of the atomic behavior is obtained by multiplying the measure of importance by the measure of rarity.

[0035] Further, the step of adaptively selecting a matching first clustering process or a second clustering process for the behavior path includes:

[0036] Obtain the user's preset characteristics, which are at least one of the following: user registration duration, length of the behavior path, or number of unique atomic behaviors contained in the behavior path;

[0037] Determine whether the analysis result of the preset feature meets the first preset condition or the second preset condition;

[0038] When the analysis result meets the first preset condition, a first clustering process is selected to match the behavior path;

[0039] When the analysis results meet the second preset condition, a second clustering process is selected to match the behavior path.

[0040] Furthermore, the method also includes:

[0041] After users are divided into different user clusters through the first clustering process, the statistically significant subsequence features that are removed in each iteration of clustering are combined as user cluster profile labels of the corresponding level, thereby generating hierarchical multi-granular profiles for users.

[0042] After users are divided into different user clusters through the second clustering process, the center vector of each user cluster in the low-dimensional embedding space is calculated, and the atomic behavior that is semantically closest to the center vector is found as the interest label of the user cluster.

[0043] Furthermore, the step of collecting user interaction behavior data and constructing a behavior path for each user consisting of atomic behaviors arranged in chronological order includes:

[0044] Each user interaction is encoded into a structured atomic behavior based on the context in which the interaction occurs, the type of content being interacted with, and the user's consumption characteristics.

[0045] The continuous consumption behavior indicators are discretized using a data-driven non-equidistant binning method to generate discrete codes for the consumption characteristics, including user dwell time and content completion rate.

[0046] Another embodiment of the present invention aims to provide a user clustering system based on behavioral path analysis, the system comprising:

[0047] The behavior path construction module is used to collect user interaction behavior data and construct a behavior path for each user consisting of atomic behaviors arranged in chronological order.

[0048] The clustering process selection module is used to analyze at least one preset feature of the behavior path, and adaptively select a matching first clustering process or a second clustering process for the behavior path based on the analysis results of the preset features.

[0049] The first clustering module is used to divide users into different user clusters based on the subsequence features composed of continuous atomic behaviors extracted from the behavior path when the clustering process selection module selects the first clustering process, so as to achieve multi-level division of users.

[0050] The second clustering module is used to divide users into different user clusters based on the path semantic vector generated by the aggregation operation of the low-dimensional embedding vectors corresponding to each atomic behavior in the behavior path when the clustering process selection module selects the second clustering process.

[0051] Another embodiment of the present invention aims to provide a computer device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the user clustering method based on behavioral path analysis as described above.

[0052] The user clustering method based on behavior path analysis provided in this invention analyzes the preset features of user behavior paths and adaptively selects either a first or second clustering process for each user. This achieves a disruptive shift from a "one-size-fits-all" approach to a "personalized" clustering model. It matches efficient and intuitive explicit pattern clustering for new users with simple behavior patterns and semantic clustering that deeply uncovers internal connections for older users with complex behavior paths, thereby significantly improving the clustering accuracy for different types of users and the overall system efficiency. In the first clustering process, it extracts subsequence features composed of continuous atomic behaviors and iteratively prunes these features while repeatedly performing clustering, achieving a multi-level, multi-granular "onion-like" analysis of user groups. This ensures that while identifying the mainstream user group, it can further discover more subtle and commercially valuable niche user patterns within subgroups, fundamentally solving the problem of mainstream patterns masking subtle patterns. In the second clustering process, it obtains low-dimensional embedding vectors of atomic behaviors and then uses importance and rarity... Weighted aggregation based on jointly determined weights generates path semantic vectors, enabling precise capture of deep semantics in user behavior. This effectively overcomes the high-dimensional sparsity and "curse of dimensionality" problems associated with older user behavior data, ensuring accurate grouping of users with similar potential interests even when behavioral appearances vary widely. After user segmentation, salient features removed in each iteration are combined into hierarchical profiles, or atomic behaviors closest to cluster center semantics are used as interest tags. This achieves automated and reliable interpretation of clustering results, providing business stakeholders with intuitive and understandable user group profiles and clear decision-making basis for subsequent refined operations. By encoding user interaction behaviors into structured atomic behaviors and discretizing continuous consumption indicators using a data-driven non-equidistant binning method, refined and business-oriented feature engineering of raw behavioral data is achieved. This ensures that every feature input into the clustering model contains rich contextual information and genuine business insights, fundamentally improving the quality and effectiveness of clustering. This solves the problems of existing clustering models being simplistic, having shallow feature representations, and being difficult to interpret and apply. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the user clustering method based on behavior path analysis in the first embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram of the user clustering system based on behavior path analysis in the second embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of the structure of the computer device in the third embodiment of the present invention;

[0056] The following detailed description of the embodiments will further illustrate the present invention in conjunction with the above-described accompanying drawings. Detailed Implementation

[0057] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0058] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0060] Example 1

[0061] Please see Figure 1 The image shows a user clustering method based on behavioral path analysis in the first embodiment of the present invention. For ease of explanation, only the parts related to the embodiment of the present invention are shown. The user clustering method based on behavioral path analysis provided by the embodiment of the present invention includes:

[0062] Step S10: Collect user interaction behavior data and construct a behavior path for each user consisting of atomic behaviors arranged in chronological order;

[0063] In one embodiment of the present invention, the step of collecting user interaction behavior data and constructing a behavior path for each user consisting of atomic behaviors arranged in a time sequence specifically includes:

[0064] Each user interaction is encoded into a structured atomic behavior based on the context in which the interaction occurs, the type of content being interacted with, and the user's consumption characteristics.

[0065] The continuous consumer behavior indicators are discretized using a data-driven non-equidistant binning method to generate discrete codes for consumer characteristics. These indicators include user dwell time and content completion rate.

[0066] Specifically, the aforementioned collection of user interaction data and the construction of a behavior path for each user, consisting of sequentially arranged atomic behaviors, primarily serves to transform the user's raw, fragmented interaction logs into a structured, information-rich, and precisely reflective temporal sequence of user intent and behavioral patterns. The implementation steps specifically include:

[0067] First, the event tracking SDK (Software Development Kit) deployed on the user's client (such as an App or Mini Program) is responsible for capturing all user interactions in real time. These interactions include, but are not limited to, page views, element clicks, swipes, input, and audio / video playback duration. Whenever an event occurs, the event tracking SDK immediately generates a raw log containing core information and reports it to the backend data receiving server via an asynchronous network request. This raw log directly reflects the user's underlying operations, and its data structure typically includes a unique user identifier (UserID), an event timestamp (Timestamp), an event name (EventName), and event properties (EventProperties). The unique user identifier includes, for example, a device ID or login account, used to associate all actions of the same user. The event timestamp is usually accurate to milliseconds and is the basis for constructing the time-series path. The event name defines the basic type of behavior. The event properties are JSON objects containing contextual information about the event, such as the ID of the clicked news item, the source page, and its position on the page.

[0068] Furthermore, after receiving the raw logs, the backend data receiving server will perform preliminary processing and verification. To eliminate errors introduced by differences in client environments or network issues, the raw logs will undergo data cleaning, specifically including: removing forged data generated by test accounts, correcting abnormal timestamps caused by inaccurate device clocks, and filling in missing fields caused by reporting failures or version compatibility issues.

[0069] Furthermore, the pre-processed log stream is transformed into atomic behaviors. First, scenario encoding is performed. At this stage, a pre-maintained scenario mapping dictionary is queried based on the page and module identifiers in the logs. This dictionary maps various specific identifiers, which may have version differences, to a standardized, unique scenario ID. This eliminates data inconsistencies caused by client heterogeneity and ensures a unified benchmark for behavioral analysis within the same functional scenario. Next, content type and category parsing and encoding are performed. The content metadata service interface is called based on the unique content identifier in the logs. This service returns detailed attributes of the content, including its content type and category. Content types include text, images, videos, or advertisements, while content categories include sports and entertainment. These content types and categories are then mapped to preset content type IDs and content category IDs, respectively, enabling the differentiation of user preferences for different content types and categories. Finally, consumption features are generated and encoded. These features are generated by discretizing continuous consumption behavior indicators using a data-driven non-equidistant binning method. Specifically, for continuous metrics such as dwell time or content completion rate, traditional equidistant binning (e.g., intervals of 10 seconds) cannot accurately reflect the natural clustering of user behavior patterns. To address this issue, this embodiment of the invention employs a data-driven process. Specifically, firstly, a large-scale sample of data, such as tens of millions of dwell time records, is randomly extracted from a historical database. Then, this one-dimensional data is used as input, and an unsupervised learning algorithm based on one-dimensional K-Means clustering is applied. The goal of this algorithm is to divide data points into a predetermined number of clusters (e.g., K=5) and identify the natural boundaries between clusters. After the K-Means clustering algorithm is completed, the boundary values ​​of each cluster (e.g., for dwell time, these might be 2.5 seconds, 13.8 seconds, 45.2 seconds, or 120.0 seconds) are adopted as thresholds for non-equidistant binning. These thresholds are determined by the data distribution itself and can effectively distinguish statistically significant different behavioral patterns, such as "scanning," "brief browsing," "normal reading," and "deep immersion." Based on the aforementioned thresholds, a binning rule model is generated. When processing new interactions in real time, the duration of the interaction (e.g., 15 seconds) is recorded, and the model is applied to determine the range the interaction falls into (e.g., between 13.8 seconds and 45.2 seconds), thus assigning it a unique, discrete consumption feature ID. This process directly reflects the user's actual consumption habits, greatly enhancing the information content of atomic behaviors. Through encoding across these three dimensions, each user interaction is transformed into an information-rich and structurally consistent atomic behavior.

[0070] After all interactions are encoded into atomic behaviors, all atomic behaviors belonging to the same user are aggregated together based on the user's unique identifier. Then, they are strictly sorted in ascending order according to the original timestamp of each atomic behavior record. This sorted sequence of atomic behaviors ultimately constitutes a behavior path that fully reflects the sequence and pattern of the user's actions.

[0071] Furthermore, to address the boundary issues of user sessions, the method in this embodiment introduces a session segmentation preprocessing step. Specifically, if the time interval between two consecutive atomic actions of a user exceeds a preset threshold (e.g., 30 minutes), the previous session is considered to have ended, and the subsequent action initiates a new session. Therefore, when constructing behavior paths, a special session-ending atomic action can be inserted at the session boundary, or paths from different sessions can be processed separately. In this embodiment, by constructing a user's long-term spanning behaviors into one or more behavior paths, long-term, cross-session behavior patterns are captured. The generated behavior path is a high-quality data structure with highly condensed information, clear temporal relationships, and can be directly used for subsequent complex model analysis.

[0072] Step S20: Analyze at least one preset feature of the behavior path, and adaptively select a matching first clustering process or a second clustering process for the behavior path based on the analysis results of the preset features.

[0073] In one embodiment of the present invention, the step of adaptively selecting the matching first clustering process or second clustering process for the behavior path specifically includes:

[0074] Obtain user preset characteristics, which are at least one of the following: user registration duration, length of behavior path, or number of unique atomic behaviors contained in behavior path;

[0075] Determine whether the analysis results of the preset features meet the first preset condition or the second preset condition;

[0076] When the analysis results meet the first preset condition, select the first clustering process to match the behavior path;

[0077] When the analysis results meet the second preset conditions, the second clustering process is selected to match the behavior path.

[0078] Specifically, for each input user behavior path, three preset features are calculated: user registration duration, behavior path length, and the number of unique atomic behaviors. For the user registration duration, the user center database is queried based on the user ID to which the behavior path belongs to obtain the user's registration time. The registration duration is then calculated by subtracting the registration time from the current time. This feature is a direct basis for determining whether a user is new, as new users typically exhibit more singular and exploratory behavior patterns. For the behavior path length, the total number of atomic behaviors constituting the behavior path is directly counted. The length of the behavior path is a basic indicator of user activity and the total amount of behavioral information. Shorter behavior paths usually indicate sparse behavioral data, suitable for explicit pattern matching; while longer paths contain richer behavioral patterns, suitable for deep semantic mining. For the number of unique atomic behaviors, all atomic behaviors in the behavior path are traversed and deduplicated to obtain the number of unique atomic behaviors contained in the behavior path. This feature directly reflects the diversity of user behavior. A smaller number indicates a narrower range of user behavior and higher repetitive patterns; a larger number indicates a wider range of functions explored and content consumed, with more complex behavioral patterns.

[0079] Through the above calculations, each behavioral path is assigned a feature vector composed of these three quantitative indicators. Based on the quantitative feature analysis results obtained above, a predefined set of decision rules is applied to automatically select and match the most suitable clustering process for the behavioral path. The core logic of this decision rule set lies in distinguishing between users with "simple behavioral patterns suitable for explicit pattern mining" and users with "complex behavioral patterns suitable for implicit semantic understanding." Specifically, the following judgment logic is executed:

[0080] The first pre-defined condition is determined by whether the feature analysis results of the behavioral path meet any of the following conditions: the user's registration duration is less than a pre-defined "new user" duration threshold (e.g., 7 days); the length of the behavioral path is less than a pre-defined "sparse behavior" length threshold (e.g., 100 atomic behaviors); or the number of unique atomic behaviors is less than a pre-defined "low diversity" number threshold (e.g., 20 types). When the analysis results meet the first pre-defined condition (i.e., any one of the above conditions is true), the behavioral path is considered to have low complexity. Such behavioral paths typically belong to new users or inactive users, whose behavioral patterns are relatively fixed and superficial. For these behavioral paths, capturing their explicit, specific behavioral sequences (e.g., whether a certain onboarding process has been completed) is more important and effective than mining their deep semantics. Therefore, a matching first clustering process is selected for this behavioral path, namely the iterative pruning clustering method based on sub-sequence features in step S30 below.

[0081] The second pre-set condition is determined as follows: If the feature analysis results of the path do not meet the first pre-set condition, it is further determined whether it meets the second pre-set condition, namely: the user's registration duration is greater than or equal to the "new user" duration threshold; the length of the behavioral path is greater than or equal to the "sparse behavior" length threshold; and the number of unique atomic behaviors is greater than or equal to the "low diversity" number threshold. When the above analysis results meet the second pre-set condition, that is, all the above conditions are true at the same time, the behavioral path is determined to have high complexity. Such behavioral paths usually belong to mature and active users whose behavior is driven by deep and diverse interests, and simple behavioral sequences are difficult to summarize their true intentions. For such behavioral paths, a method that can understand the context and semantic relationship between behaviors is required. Therefore, a matching second clustering process is selected for this behavioral path, namely the semantic clustering method based on embedding vector aggregation in step S40 below.

[0082] Step S30: When the first clustering process is selected, based on the subsequence features composed of continuous atomic behaviors extracted from the behavior path, users are divided into different user clusters through an iterative feature pruning process and repeated clustering to achieve multi-level segmentation of users.

[0083] In one embodiment of the present invention, the step of classifying users into different user clusters based on the subsequence features composed of continuous atomic behaviors extracted from the behavior path, through an iterative feature pruning process and repeated clustering, specifically includes:

[0084] A sliding window of a preset length is used to traverse the behavior path and extract all subsequence features of continuous atomic behaviors that conform to the preset length, forming an initial feature subspace.

[0085] Within the feature subspace used in the current round, user behavior paths are clustered to obtain the user clusters for this round;

[0086] Identify and remove statistically significant subsequence features that define the user cluster in this round, in order to form a new, pruned feature subspace for the next iteration;

[0087] Determine whether the preset iteration termination condition is met. If not, use the new feature subspace as the feature subspace used in the current round and return to execute the step of clustering user behavior paths.

[0088] Furthermore, the steps described above for identifying and removing statistically significant subsequence features that define the user cluster in this round specifically include:

[0089] For each user cluster generated in this round of clustering, a contingency table is constructed to describe the distribution of features of each subsequence inside and outside the user cluster;

[0090] The chi-square statistic of each subsequence feature relative to the user cluster is calculated based on the contingency table;

[0091] Identify and select the subsequence features with the highest global chi-square statistic of a preset number as salient features that need to be removed from the current feature space.

[0092] Specifically, when selecting the first clustering process to match user behavior paths, the core objective is to mine and utilize explicit, continuous behavioral patterns within the behavior paths. Through the iterative pruning clustering method described in this embodiment, a multi-level, coarse-to-fine segmentation of the user group is achieved. This process abandons the traditional one-time clustering method, replacing it with a dynamic, onion-peeling analysis process. This process mainly includes four mutually iterative key steps: extraction and space construction of sub-sequence features, user clustering in the current feature space, identification and pruning of core features, and control and termination of iteration. The goal of this stage is to transform each user behavior path into a vector composed of sub-sequence features that can be used for clustering calculations.

[0093] First, a preset subsequence feature length is set, typically referred to as N in N-gram (in this embodiment, N = 3). Then, a sliding window of length N is used to traverse each user behavior path from beginning to end. At each position of the sliding window, N consecutive atomic behaviors constitute a "subsequence feature". For example, for the path A->B->C->D->E, the sliding window will sequentially extract three 3-gram subsequence features: {A->B->C}, {B->C->D}, and {C->D->E}. Further, this process is performed on all user behavior paths, and all extracted unique subsequence features are aggregated to construct a global, high-dimensional initial feature subspace. Subsequently, each user behavior path can be represented as a vector in this feature subspace, where each dimension of the vector corresponds to a unique subsequence feature, the value of which can be the frequency of that feature in the path.

[0094] Secondly, after constructing the feature vectors, clustering is performed on all users within the feature subspace used in the current round. In this embodiment, a top-down split-level hierarchical clustering algorithm is used. In the first iteration, the clustering algorithm uses the complete initial feature subspace. It treats all users as a large root node and then finds an optimal split point to divide them into two initial user clusters with the greatest differences, such as cluster A and cluster B. The split is based on the distance between user path vectors; users with closer distances are considered to have more similar behavioral patterns. The goal is to identify and mask the key features that define the clustering results of this round, forcing the next round of clustering to focus on more subtle differences. Specifically, after obtaining the user clusters for this round (such as cluster A and cluster B), the following operations are performed to identify and remove statistically significant subsequence features: For each user in cluster A, their path vector is analyzed. Then, for each subsequence feature in the feature subspace, a 2x2 contingency table is constructed. This contingency table describes the distribution of the feature's presence or absence "within cluster A" and "outside cluster A (i.e., cluster B)". Based on this contingency table, the chi-square statistic is calculated for each subsequence feature. A higher chi-square value indicates a stronger ability of the feature to distinguish users inside and outside cluster A; that is, the feature is a core identifier defining cluster A. At this point, a preset number (e.g., Top 10) of subsequence features with the highest global chi-square statistics are identified and selected. After identifying these core features, they are temporarily "removed" from the feature subspace used in the current round. This operation creates a smaller, pruned new feature subspace for the next iteration.

[0095] Finally, after completing one pruning iteration, it is determined whether the preset iteration termination conditions are met. These iteration termination conditions include, but are not limited to, reaching the preset target number of clusters, having too few users in the current cluster, or the pruned feature subspace no longer providing sufficient discriminative power.

[0096] If the termination condition is not met, the next iteration will begin. For example, cluster A generated in the previous round is selected as the new analysis object. Then, the new, pruned feature subspace is used as the feature subspace for the current round, and the user clustering steps under the current feature space described above are executed again. At this point, since the most salient feature defining cluster A has been removed, the clustering algorithm will be forced to further divide cluster A into more refined subclusters, such as cluster A1 and cluster A2, based on those minor, more subtle feature differences.

[0097] The process of clustering, identification, pruning, and re-clustering is continuously repeated. Each iteration further divides a more focused user subgroup into more refined features, ultimately forming a tree-like, multi-layered user group structure.

[0098] Therefore, when the behavioral path is determined to be simple in structure and sparse in information (usually corresponding to new users or low-activity users), the first clustering process is initiated. This allows for the discovery of explicit, literal, continuous patterns in user behavior, and multi-granularity partitioning is achieved through iterative pruning clustering. At this point, by extracting explicit sub-sequence features, the key conversion path of new users is accurately captured. New user behavior is typically limited and goal-oriented, such as completing onboarding or making an initial purchase. Therefore, this process can directly identify these fixed patterns, thus solving the problem of ambiguous operational strategies caused by the inability to understand key new user behaviors, and providing direct and clear data support for product optimization and user onboarding. Simultaneously, by extracting sub-sequence features (i.e., N-grams) of continuous atomic behaviors, the process focuses on capturing direct, continuous, and fixed-length local patterns in user behavior. For example, it accurately identifies a typical, superficial behavioral combination such as "opening the app -> browsing the recommended feed -> clicking on images and articles." New user behavioral paths are usually short, and their behavioral patterns are not yet fixed; their behavioral intentions are mainly reflected through simple, repetitive exploratory operations. This process enables the precise and efficient capture of explicit patterns during these exploration phases, thereby accurately categorizing new users into groups such as "video explorers," "text and image viewers," or "feature trial users." For users with sparse data and insufficient contextual information, forcibly learning their "deep semantics" is unrealistic and ineffective; instead, directly extracting explicit patterns from their limited behavior is more robust and effective.

[0099] By iteratively pruning features and repeatedly performing clustering, a deep, multi-layered analysis of seemingly simple user groups is achieved. Even new users exhibit subtle differences in behavior. Each iteration identifies and separates the most significant group-specific features. For example, the first round identifies "deep interaction," and the second round identifies "preference for sports content" within the "deeply interactive" group. At this point, the final clustering result and its causes (i.e., the features pruned in each round) are completely transparent. Therefore, by masking the most significant features in each iteration, the algorithm is forced to focus on secondary behavioral patterns. This solves the problem of traditional clustering's overly coarse granularity in segmenting simple behavioral patterns, resulting in truly multi-granular, progressively interpretable user profiles.

[0100] Step S40: When the second clustering process is selected, users are divided into different user clusters based on the path semantic vector generated by the aggregation operation of the low-dimensional embedding vectors corresponding to each atomic behavior in the behavior path.

[0101] In one embodiment of the present invention, the step of dividing users into different user clusters based on the path semantic vector generated by the aggregation operation of the low-dimensional embedding vectors corresponding to each atomic behavior in the behavior path specifically includes:

[0102] Obtain the low-dimensional embedding vector corresponding to each atomic behavior in the behavior path;

[0103] Calculate the weight of each atomic behavior in the behavior path, where the weight is determined by the importance of the atomic behavior in the current behavior path and its rarity in all user behavior paths;

[0104] Multiply the low-dimensional embedding vector of each atomic behavior by its corresponding weight to obtain a weighted vector;

[0105] An aggregation operation is performed to sum and normalize all weighted vectors in the behavior path to generate a path semantic vector, and users are divided into different user clusters based on the path semantic vector.

[0106] Furthermore, the steps for calculating the weight of each atomic behavior in the above-mentioned behavioral path specifically include:

[0107] Obtain the frequency of atomic behaviors occurring in the current behavior path as a measure of importance;

[0108] Obtain the total number of user paths containing atomic behaviors, and calculate the inverse frequency value based on the total number of user paths, which serves as a measure of rarity.

[0109] The final weight of the atomic behavior is obtained by multiplying the measure of importance by the measure of rarity.

[0110] Furthermore, the steps for obtaining the low-dimensional embedding vector corresponding to each atomic behavior in the behavior path specifically include:

[0111] Use all user behavior paths as training data;

[0112] The training corpus was trained using a neural network model based on the Skip-gram architecture and optimized with negative sampling.

[0113] Extract the low-dimensional embedding vector corresponding to each unique atomic behavior from the trained neural network model.

[0114] Specifically, when selecting a matching second clustering process for user behavior paths, the core objective is to go beyond the literal sequence of behaviors and delve deeper into the semantics and intentions hidden behind them, thereby accurately segmenting users with complex and diverse behavior patterns. This second clustering process primarily transforms the symbolic behavior path into a low-dimensional, dense mathematical vector—the path semantic vector—and then performs clustering within this vector space, enabling each unique atomic behavior to learn a low-dimensional vector representation that can characterize its meaning. These vectors are not manually set but are automatically learned from massive amounts of user behavior data through unsupervised learning. This second clustering process mainly includes three key steps: obtaining the embedding vectors of atomic behaviors, generating a weighted aggregation of path semantic vectors, and segmenting users based on the path vectors.

[0115] Specifically, the entire set of user behavior paths is first treated as a massive text corpus, where each behavior path corresponds to a sentence and each atomic behavior corresponds to a word. Then, a technique similar to word2vec in natural language processing (in this embodiment, the Skip-gram model) is used to train this corpus. The neural network model learns the correlation and substitutability between behaviors by analyzing the context of each atomic behavior in the massive number of paths (i.e., other atomic behaviors preceding and following it). After training, the neural network model generates a low-dimensional dense vector of a fixed dimension (e.g., 128 dimensions) for each unique atomic behavior in the vocabulary. These vectors possess the semantic characteristic that "atomic behaviors with similar meanings or frequently co-occurring in the vector space have corresponding vectors that are spatially close to each other." For example, the vectors for the atomic behaviors "watching food videos" and "reading recipe articles" would have a very high similarity. These pre-trained embedding vectors are stored for subsequent real-time retrieval. When processing a specific user behavior path, each atomic behavior in that path is traversed, and its corresponding low-dimensional embedding vector is directly retrieved from storage. This allows a sequence consisting of multiple embedded vectors of atomic behaviors to be aggregated into a single path semantic vector that can represent the core semantics of the entire path.

[0116] Furthermore, to achieve this objective, this embodiment of the invention employs a weighted aggregation strategy to ensure that key behaviors have a higher influence during the aggregation process. The specific implementation process is as follows: For each atomic behavior in the path, its weight is calculated. This weight is jointly determined by two factors: Importance metric (TF) and Rarity metric (IDF). The Importance metric calculates the frequency of the atomic behavior's occurrence in the current behavior path. If an atomic behavior appears repeatedly in a single behavior path, it usually means that it is particularly important to that user. The Rarity metric calculates the total number of user paths containing the atomic behavior and obtains an inverse frequency value based on the total number of user paths. If an atomic behavior appears only in the behavior paths of a few users, it indicates that it is a highly distinctive and rare behavior. Finally, the Importance metric and the Rarity metric are multiplied to obtain the final weight of the atomic behavior. After obtaining the final weight of each atomic behavior, the low-dimensional embedding vector of each atomic behavior is multiplied by its corresponding weight to obtain a weighted vector. Then, all weighted vectors in the behavior path are summed. Finally, the summed vector is normalized (e.g., L2 norm normalization) to eliminate the influence of path length. After this series of operations, a single, fixed-dimensional path semantic vector is generated. This vector mathematically condenses the core information of the entire path, and its direction points to the semantic center of the user's primary behavioral intent within that path.

[0117] After generating path semantic vectors for all users to be analyzed, user segmentation is performed in this low-dimensional dense space composed of path vectors. At this point, all users' path semantic vectors are used as input, and a mature clustering algorithm (K-Means algorithm in this embodiment) is applied. The K-Means algorithm groups path vectors that are spatially close into the same cluster based on the Euclidean distance between the vectors. Since the path semantic vectors themselves contain deep semantic information, even if two users' original behavioral paths have nothing in common literally, as long as their core intentions are similar, their path vectors will be close to each other in space, thus being correctly classified into the same user cluster. Finally, this process outputs a set of user clusters, where users within each cluster have a high degree of homogeneity in behavioral intentions and interest preferences.

[0118] Therefore, when a behavioral path is determined to be structurally complex and information-rich (typically corresponding to mature, active users), a second clustering process is initiated. This process is not limited to local, continuous combinations of behaviors, but treats each atomic behavior as a "word." By learning its contextual relationships across all user behavioral paths (the global corpus), a low-dimensional embedding vector representing its semantics is generated. Thus, it focuses on the deep, logically related, but not direct, discontinuous connections between behaviors, thereby accurately segmenting users with diverse behavioral patterns. For example, it understands that "reading financial articles" and "watching stock analysis videos" are highly semantically related, even if these two behaviors are far apart in the path. In this case, by transforming atomic behaviors into low-dimensional embedding vectors, a deep understanding of the user's true intent is achieved. Among these, older users have long behavioral paths and complex, diverse behavioral patterns; their core interests are often reflected through cross-scenario, discontinuous combinations of behaviors. This process typically penetrates behavioral appearances to accurately capture implicit associations based on deep interests, thereby identifying user groups that simple sequence patterns cannot detect, such as "sports enthusiasts" (who watch games, read reports, and participate in discussions) and "high-value content contributors." This is crucial for achieving in-depth user engagement, personalized recommendations, and monetization. While the first clustering process cannot group them into one category, this second clustering process learns from the context that the two behaviors are semantically equivalent. This solves the problem of explicit pattern matching failing due to the diversity of user behaviors, greatly improving the accuracy of clustering.

[0119] By aggregating high-dimensional, sparse behavioral paths into low-dimensional, dense path semantic vectors, efficient processing and generalization of complex data are achieved. Since long paths of experienced users generate an astronomical number of subsequence combinations, using the first clustering process would result in an explosive feature space (curse of dimensionality) and extremely sparse data, leading to clustering failure. This process, however, effectively solves the curse of dimensionality and data sparsity problems by mapping all behaviors to a unified low-dimensional, dense vector space. This not only greatly improves computational efficiency but also enables the model to generalize and understand unseen new behavioral patterns. Furthermore, weighted aggregation, determined by weights jointly defined by "Importance" (TF) and "Scarcity" (IDF), amplifies the influence of those unique behaviors that best represent user personality (high IDF values). Therefore, the weighted aggregation mechanism ensures that the clustering results better reflect users' core, stable interests, rather than being influenced by a large number of routine, indistinguishable behaviors (such as "opening the app"). This makes the segmentation of complex users both accurate and robust, greatly improving the quality of clustering.

[0120] Furthermore, in one embodiment of the present invention, the above method further includes:

[0121] After users are divided into different user clusters through the first clustering process, the statistically significant subsequence features that are removed in each iteration of clustering are combined as user cluster profile labels of the corresponding level, thereby generating hierarchical multi-granular profiles for users.

[0122] After users are divided into different user clusters through the second clustering process, the center vector of each user cluster in the low-dimensional embedding space is calculated, and the atomic behavior that is semantically closest to the center vector is found as the interest label of the user cluster.

[0123] Specifically, after obtaining multi-level user segmentation through the first clustering process, the intermediate information generated during its iterative pruning process is used to generate a hierarchical, coarse-to-fine multi-granular profile for each user. The specific implementation process is as follows: the output of the first clustering process is a tree-like hierarchical structure. For example, all users are first divided into cluster A and cluster B, and then cluster A is further subdivided into cluster A1 and cluster A2. During each split, a set of subsequence features defining the current segmentation and possessing the highest statistical significance is identified and removed. At this point, the entire clustering tree construction process is traced back, and the following operations are performed: when initially dividing all users into clusters A and B, the core feature combination defining cluster A is recorded (e.g., "high-frequency search -> clicked news subsequence"). This feature combination is directly used as the profile label for cluster A at the top level (first layer). When cluster A is further divided into clusters A1 and A2, the new core features defining cluster A1 after removing the core features of the first layer are also recorded (e.g., "high-frequency sports news -> comment subsequence"). Ultimately, the complete profile label for cluster A1 is generated by combining the profile label of its parent node (cluster A) with its own profile label. Therefore, the user profile of cluster A1 is a hierarchical description: "News search user" -> "Sports news deep interaction preference user". In this way, the salient features removed in each iteration become part of the corresponding level of user cluster profile. Finally, for each bottom-level user cluster, a clear, multi-granular profile path from the root node to the leaf node is generated, fully revealing the evolution of the user group's behavioral characteristics from general to specific.

[0124] After user segmentation is achieved through the second clustering process, since this clustering process is based on an abstract low-dimensional vector space, a method is needed to translate spatial locations back into business language to generate interest tags for user clusters. The specific implementation process is as follows: The output of the second clustering process is a set of flat user clusters. For each user cluster, the path semantic vectors of all users within that cluster are obtained. Then, the average of these vectors is calculated element-wise to obtain a center vector representing the average semantics or centroid of the user cluster in the low-dimensional embedding space. Mathematically, this center vector is the convergence point of the overall behavioral intent of the user group. The calculated center vector is then compared with the embedding vectors of all unique atomic behaviors stored in the pre-trained embedding model. The comparison method is to calculate the cosine similarity between the center vector and each atomic behavior vector. The higher the cosine similarity, the more consistent the directions of the two in the semantic space, i.e., the closer their meanings. Finally, a predetermined number (e.g., Top 5) of atomic behaviors with the highest cosine similarity to the center vector of the user cluster are identified. These atomic behaviors are the most accurate descriptions of the core interests of the user cluster. For example, the calculated preset number of atomic behaviors might be: "Evening - Novel reading - Completed", "Late at night - Community post - Comment", "Before bed - Text and image information - Long stay". These specific, high-frequency semantic behaviors combined together constitute the final interest tag for this user cluster: "Late-night immersive consumption of text and image information and novels". In this way, mathematical concepts in low-dimensional embedded space can be mapped back to specific, understandable user behaviors, giving each user cluster a clear and focused interest tag.

[0125] Furthermore, in one embodiment of the present invention, the above method further includes:

[0126] The identified user clusters are associated with a predefined library of operational strategies;

[0127] When a user's cluster affiliation changes, or a new user is assigned to a specific user cluster, an operational action associated with the user cluster is automatically triggered from the operational strategy library. These actions may include adding the user to a specific A / B test group, pushing a customized content list, or sending personalized marketing messages.

[0128] Specifically, this step involves deeply associating dynamic user cluster affiliations with a predefined operational strategy library and establishing an event-driven triggering mechanism to automate and refine interventions in user lifecycle changes. First, a configuration backend or interface is provided for operations personnel to define and input various operational strategies. Each strategy in the operational strategy library is an atomic operational action, and its data structure includes at least: a strategy identifier, a target user cluster profile tag, an action type, and an action carrier. The strategy identifier is a unique strategy identifier. The target user cluster profile tag is the user group that the strategy aims to reach, such as "video explorers among new users" or "late-night novel enthusiasts among mature users." The action type is the specific form of intervention, such as "assigning a comparative test group," "pushing a customized content list," or "sending a personalized marketing message." The action carrier is the specific material information required to execute the action. For example, for "assigning a comparative test group," the carrier is the name of the experimental group (e.g., "New player interface group B"); for "pushing a customized content list," the carrier is the identifier of the customized content pool or the recommendation algorithm; for "sending a personalized marketing message," the carrier is the message template identifier and the associated coupon code.

[0129] After the clustering process (whether it's the first or second clustering process) generates new or updated user cluster profile tags, operations personnel can bind one or more preset operational strategies to these profile tags in the strategy library. This association forms the basis for subsequent automated execution. The goal of this stage is to establish a monitoring system capable of sensing changes in user cluster affiliation in real time and generating a trigger event for downstream consumption when a change occurs. The user's final cluster affiliation (i.e., their user cluster identifier or profile tag) is stored as a core field in the user profile system. Any changes to this field are captured, and after the clustering process runs periodically (e.g., updated daily) and completes its calculations, it writes the latest clustering results for each user to the user profile database. At this point, two key change events occur: new user affiliation determination and existing user affiliation change. New user affiliation determination occurs when a newly registered user is clustered for the first time, and their cluster affiliation field changes from "undefined" to a specific user cluster tag. Existing user affiliation change occurs when an existing user's behavior pattern changes significantly, causing them to migrate from an old user cluster to a new user cluster.

[0130] Furthermore, the user profile database is deeply integrated with a message middleware (such as a message queue service). By leveraging database change data capture technology or proactively publishing at the business logic level, any write or update operation to the user clustering affiliation field is automatically captured and encapsulated into a standardized "user clustering change event." This event message contains at least the user identifier, the user cluster label before the change, and the new user cluster label after the change. This event is published to a designated message queue for subsequent execution engine subscription and processing. Furthermore, one or more dedicated "operational action execution engine" services are deployed. These services continuously listen for and consume the "user clustering change events" published to the message queue in the previous stage. When an event is received, the execution engine performs the following automated process: First, the execution engine parses the user identifier and the new user cluster profile label from the event. Second, using this new profile label as the query key, it retrieves all associated operational strategies from the operational strategy library built in the first stage. Further, for each retrieved strategy, the execution engine parses its "action type" and "action carrier" and calls the corresponding downstream business system interface to complete the operation. For example: If the action type is "Assign a comparison test group," the engine will call the comparison test platform's interface to add the user's identifier to the specified experimental group in the carrier. If the action type is "Push a customized content list," the engine will call the message push center's interface to trigger a push notification for the user containing the customized list specified in the carrier. If the action type is "Send a personalized marketing message," the engine will call the marketing automation platform's interface to generate and send a personalized in-app message or SMS for the user.

[0131] Furthermore, in one embodiment of the present invention, the above method further includes:

[0132] On a user interface, a hierarchical, multi-granular profile is visualized in the form of an interactive tree diagram or galaxy diagram.

[0133] Receive user selection operations for any user cluster node on the user interface;

[0134] In response to the selection operation, the core user interest tags, demographic characteristics distribution, and significant improvements or declines in various indicators of the user cluster compared to the parent node user cluster are dynamically displayed.

[0135] Specifically, the calculated user cluster hierarchical or semantic relationships are first transformed into graphical elements on the user interface. In this embodiment of the invention, two optional visualization modes are provided: a hierarchical tree diagram and a semantic galaxy diagram, to accommodate the results produced by different clustering processes.

[0136] For constructing hierarchical tree diagrams, this mode is specifically designed to display the multi-level user segmentation results with parent-child relationships generated by the first clustering process. First, the entire clustering result is organized into a tree-like data structure, where the root node represents all users, each child node represents a user cluster, and the connections between nodes indicate their derivation relationships. This data structure is sent to the front-end application of the user interface. Upon receiving it, the front-end application invokes the graphics rendering engine to draw an interactive tree diagram on the canvas. Each node in the diagram is clearly labeled with the name of the user cluster it represents and the number of users it contains, intuitively demonstrating the hierarchical segmentation of the user group from macro to micro levels.

[0137] For constructing semantic galaxy maps, this model is primarily used to display the flattened user segmentation results generated by the second clustering process, emphasizing the semantic similarity between clusters. First, the center vector of each user cluster in the low-dimensional embedding space is obtained. Then, a dimensionality reduction technique is used to project these high-dimensional center vectors onto a two-dimensional plane, generating the coordinates of each user cluster on the plane. During this process, semantically similar clusters have coordinates close to each other on the two-dimensional plane. These coordinates, along with information such as the number of users in each cluster, are sent to the front-end. Upon receiving this information, the front-end application renders each user cluster as a "star," its position in the "galaxy map" determined by its two-dimensional coordinates, and its size proportional to the number of users. In this way, analysts can clearly see which user groups share similar core interests.

[0138] Furthermore, on the user interface, whether it's a rectangular node in a tree diagram or a circular star in a galaxy diagram, each graphical element representing a user cluster is bound to a selection event listener. When an analyst clicks on any user cluster node using a mouse or touch device, the front-end application immediately captures this selection action. At this point, it accurately identifies which user cluster was selected and obtains its unique cluster identifier.

[0139] Upon receiving a user's selection of a cluster node, the system uses the cluster's unique identifier as an index to instantly query, calculate, and dynamically display in-depth analysis data about that user's cluster on a specific information panel within the interface. The displayed content primarily includes the following three aspects:

[0140] Display of core user interest tags: Based on the source flow of the user cluster, the corresponding tag generation logic is invoked. If the user cluster was generated by the first clustering process, its path in the clustering tree is traced back, and the significant sub-sequence features that were removed at each level are combined to form a hierarchical profile tag and displayed. If the user cluster was generated by the second clustering process, the set of atomic behaviors that are semantically closest to its center vector is displayed as its core interest tag.

[0141] The system displays demographic characteristics: based on all user identifiers within the selected user cluster, it queries the user profile database in real time and aggregates the demographic information of these users. For example, it calculates and displays the gender ratio, age distribution, and geographic city tier distribution of the user cluster in chart form. This allows analysts to quickly understand the natural attributes of this behavioral group.

[0142] Key performance indicator (KPI) comparative analysis: To highlight the uniqueness of this user group, several key business metrics are compared with corresponding metrics of its parent user group (in the tree diagram) or all users (in the galaxy diagram). These metrics may include 7-day retention rate, average daily usage time, and content consumption completion rate. The relative percentage increase or decrease for each metric is calculated, and the most significant differences are highlighted. For example, the panel might display "7-day retention rate: 25% higher than parent node," and "Video content consumption completion rate: 10% lower than parent node." This allows analysts to immediately understand the core value and potential risks of this user group.

[0143] Furthermore, in one embodiment of the present invention, the above method further includes:

[0144] Continuously monitor the cohesion or dispersion of behavioral patterns within each user cluster;

[0145] When the cohesion of a target user cluster is lower than a preset threshold, or the dispersion is higher than a preset threshold, the target user cluster is marked as a failed cluster.

[0146] The clustering process dynamic selection step and selective clustering execution step in steps S20-S40 above are automatically re-executed for users in the failed cluster or all users to generate updated user clusters.

[0147] Specifically, in this embodiment of the invention, a periodic monitoring task is also deployed (e.g., executed daily). This task performs the following calculations for all currently valid user clusters:

[0148] Cohesion calculation: For each user cluster, the monitoring task obtains the latest behavioral paths of all users within it. Then, based on which clustering process initially generated the user cluster, the corresponding distance metric is used to calculate the pairwise behavioral path similarity between all users within the cluster, ultimately obtaining an average similarity value. This average similarity value is the cohesion score of the user cluster. A higher score indicates a more consistent user behavior pattern within the cluster.

[0149] Dispersion calculation: Simultaneously, the monitoring task also calculates the centroid of the user cluster (for the first clustering process, this could be the center of the feature distribution; for the second clustering process, it could be the average center of the path semantic vector). Then, it calculates the behavioral path distance from each user within the cluster to this centroid, ultimately obtaining an average distance or the variance of the distance. This value is the dispersion score of the user cluster. A higher score indicates that the user behavior patterns within the cluster are more dispersed and inconsistent.

[0150] At this point, the calculated cohesion and dispersion scores, along with the timestamps, are recorded in the user cluster's health history archive, forming a time series curve. In this embodiment of the invention, two configurable health thresholds are set for both cohesion and dispersion. These thresholds are determined by data analysts based on business experience and historical data distribution. For example, the cohesion threshold is set to 0.7, and the dispersion threshold is set to 0.4.

[0151] In this process, after completing the health calculation, the periodic monitoring task immediately compares the latest score with a preset threshold and executes the following judgment logic: if the cohesion score of a target user cluster is consistently lower than the preset cohesion threshold (e.g., below 0.7 for three consecutive days), or if the dispersion score of a target user cluster is consistently higher than the preset dispersion threshold (e.g., above 0.4 for three consecutive days), it is determined that the behavioral pattern of that user cluster has significantly diverged, and its original definition can no longer accurately summarize all users within it. At this time, the target user cluster is automatically marked as a failed cluster. This marking triggers an internal alarm, notifying operations and analysis personnel that the definition of this user group needs attention and preparing for subsequent automatic re-clustering.

[0152] When one or more user clusters are marked as failed clusters, a re-clustering task is automatically triggered according to a preset strategy. Re-clustering can have two execution scopes:

[0153] Local re-clustering: If the number of failed clusters is small, only all users marked as "failed clusters" are extracted to form a user pool to be analyzed.

[0154] Global re-clustering: If the number of failed clusters exceeds a certain proportion, or if a fixed global update cycle is set (e.g., once a week), then all users will be used as the user pool to be analyzed.

[0155] At this point, whether it's local or global re-clustering, the user pool to be analyzed will be re-entered into the forefront of the method of this invention. That is, for each user, the clustering process in step S20 above is dynamically selected (i.e., analyzing their latest behavioral path characteristics), and the most suitable first or second clustering process is matched. Subsequently, the matched clustering process (i.e., step S30 or step S40) is selectively executed to perform a completely new segmentation of these users. This process ultimately generates updated user clusters. The new cluster division not only corrects the old, invalid clusters but may also discover new group patterns that have emerged due to the evolution of user behavior. The updated user clusters and their profile labels will overwrite the old segmentation results and automatically establish new associations with the operational strategy library, thereby ensuring that the entire user segmentation system always remains highly synchronized with the actual user behavior.

[0156] In summary, the user clustering method based on behavior path analysis in the above embodiments of the present invention, by analyzing the preset features of user behavior paths and adaptively selecting the first or second clustering process for each user, achieves a disruptive shift from a "one-size-fits-all" to a "personalized" clustering model. It matches efficient and intuitive explicit pattern clustering for new users with simple behavior patterns, and semantic clustering that can deeply explore internal connections for older users with complex behavior paths, thereby greatly improving the clustering accuracy for different types of users and the overall system efficiency. In the first clustering process, by extracting subsequence features composed of continuous atomic behaviors and iteratively pruning features and repeatedly executing clustering, a multi-level, multi-granular "onion-like" analysis of user groups is achieved. This ensures that while identifying the mainstream user group, more subtle and commercially valuable niche user patterns can be discovered within subgroups, fundamentally solving the problem of mainstream patterns masking subtle patterns. In the second clustering process, by obtaining low-dimensional embedding vectors of atomic behaviors and using importance and... Weighted aggregation based on sparseness-determined weights generates path semantic vectors, enabling precise capture of deep semantics in user behavior. This effectively overcomes the high-dimensional sparsity and "curse of dimensionality" problems associated with older user behavior data, ensuring accurate grouping of users with similar potential interests even when behavioral appearances vary widely. After user segmentation, salient features removed in each iteration are combined into hierarchical profiles, or atomic behaviors closest to cluster center semantics are used as interest tags. This achieves automated and reliable interpretation of clustering results, providing business stakeholders with intuitive and understandable user group profiles and clear decision-making basis for subsequent refined operations. By encoding user interactions into structured atomic behaviors and discretizing continuous consumption indicators using a data-driven non-equidistant binning method, refined and business-oriented feature engineering of raw behavioral data is achieved. This ensures that every feature input into the clustering model contains rich contextual information and genuine business insights, fundamentally improving the quality and effectiveness of clustering. This solves the problems of existing clustering models being simplistic, having shallow feature representations, and being difficult to interpret and apply.

[0157] Example 2

[0158] Please see Figure 2 This is a schematic diagram of a user clustering system based on behavioral path analysis provided in the second embodiment of the present invention. For ease of explanation, only the parts related to the embodiment of the present invention are shown. The system includes:

[0159] The behavior path construction module 11 is used to collect user interaction behavior data and construct a behavior path for each user consisting of atomic behaviors arranged in chronological order.

[0160] Clustering process selection module 12 is used to analyze at least one preset feature of the behavior path, and adaptively select a matching first clustering process or a second clustering process for the behavior path based on the analysis results of the preset features.

[0161] The first clustering module 13 is used to divide users into different user clusters based on the subsequence features composed of continuous atomic behaviors extracted from the behavior path when the clustering process selection module 12 selects the first clustering process, so as to achieve multi-level division of users.

[0162] The second clustering module 14 is used to divide users into different user clusters based on the path semantic vector generated by the aggregation operation of the low-dimensional embedding vectors corresponding to each atomic behavior in the behavior path when the clustering process selection module 12 selects the second clustering process.

[0163] Furthermore, in one embodiment of the present invention, the first clustering module 13 includes:

[0164] The subsequence feature extraction unit is used to traverse the behavior path using a sliding window of a preset length, extract all subsequence features that conform to the preset length of continuous atomic behavior, and form an initial feature subspace.

[0165] The behavior path clustering unit is used to cluster user behavior paths within the feature subspace used in the current round to obtain the user clusters for this round.

[0166] The pruning unit is used to identify and remove statistically significant subsequence features that define the user cluster in this round, in order to form a new, pruned feature subspace for the next iteration.

[0167] The loop unit is used to determine whether the preset iteration termination condition is met. If not, the new feature subspace is used as the feature subspace for the current round, and the process returns to the step of clustering user behavior paths.

[0168] Furthermore, in one embodiment of the present invention, the pruning unit includes:

[0169] The contingency table construction subunit is used to construct a contingency table describing the distribution of features of each subsequence inside and outside the user cluster for each user cluster generated in this round of clustering;

[0170] The chi-square statistic calculation unit is used to calculate the chi-square statistic of each subsequence feature relative to the user cluster based on the contingency table;

[0171] The pruning sub-unit is used to identify and select a preset number of subsequence features with the highest global chi-square statistic as salient features that need to be removed from the current feature space.

[0172] Furthermore, in one embodiment of the present invention, the second clustering module 14 includes:

[0173] A low-dimensional embedding vector acquisition unit is used to acquire the low-dimensional embedding vector corresponding to each atomic behavior in the behavior path;

[0174] The weight calculation unit is used to calculate the weight of each atomic behavior in the behavior path. The weight is determined by the importance of the atomic behavior in the current behavior path and its rarity in all user behavior paths.

[0175] The weighted unit is used to multiply the low-dimensional embedding vector of each atomic behavior with its corresponding weight to obtain a weighted vector;

[0176] The path semantic vector generation unit is used to perform an aggregation operation that sums and normalizes all weighted vectors in the behavior path to generate the path semantic vector, and to divide users into different user clusters based on the path semantic vector.

[0177] Furthermore, in one embodiment of the present invention, the weight calculation unit includes:

[0178] An importance metric acquisition subunit is used to acquire the frequency of the atomic behavior in the current behavior path, as a metric for the importance.

[0179] The rarity metric acquisition subunit is used to acquire the total number of user paths containing the atomic behavior, and calculate the inverse frequency value based on the total number of user paths, as the rarity metric.

[0180] The weight calculation subunit is used to multiply the measure of importance by the measure of rarity to obtain the final weight of the atomic behavior.

[0181] Furthermore, in one embodiment of the present invention, the clustering process selection module 12 includes:

[0182] A preset feature acquisition unit is used to acquire preset features of the user, wherein the preset feature is at least one of the following: user registration duration, length of the behavior path, or number of unique atomic behaviors contained in the behavior path;

[0183] The result analysis unit is used to determine whether the analysis result of the preset feature meets the first preset condition or the second preset condition.

[0184] The first selection unit is used to select a matching first clustering process for the behavior path when the result analysis unit determines that the analysis result meets the first preset condition.

[0185] The second selection unit is used to select a matching second clustering process for the behavior path when the result analysis unit determines that the analysis result meets the second preset condition.

[0186] Furthermore, in one embodiment of the present invention, the system further includes:

[0187] The multi-granularity profile generation module is used to combine the statistically significant subsequence features that are removed in each iteration of clustering after the user is divided into different user clusters through the first clustering process, and use them as user cluster profile labels of the corresponding level, thereby generating hierarchical multi-granular profiles for users.

[0188] The interest tag acquisition module is used to calculate the center vector of each user cluster in the low-dimensional embedding space after the users are divided into different user clusters through the second clustering process, and find the atomic behavior that is semantically closest to the center vector as the interest tag of the user cluster.

[0189] Furthermore, in one embodiment of the present invention, the behavior path construction module 11 includes:

[0190] Atomic behavior encoding units are used to encode each user interaction into a structured atomic behavior, based on the scenario in which the interaction occurs, the type of content being interacted with, and the user's consumption characteristics.

[0191] The consumption feature encoding unit is used to discretize continuous consumption behavior indicators using a data-driven non-equidistant binning method to generate discrete codes for the consumption features, including user dwell time and content completion rate.

[0192] Furthermore, in one embodiment of the present invention, the system further includes:

[0193] The association module is used to associate the segmented user clusters with a predefined operational strategy library;

[0194] The operational action triggering module is used to automatically trigger an operational action associated with a user cluster from the operational strategy library when a user's cluster affiliation changes or a new user is assigned to a specific user cluster. The operational actions include adding the user to a specific A / B test group, pushing a customized content list, or sending personalized marketing messages.

[0195] Furthermore, in one embodiment of the present invention, the system further includes:

[0196] The visualization module is used to visualize hierarchical, multi-granular images in the form of interactive tree diagrams or galaxy diagrams on a single user interface.

[0197] The selection operation receiving module is used to receive the user's selection operation on any user cluster node on the user interface;

[0198] The selection operation response module is used to respond to selection operations and dynamically display the core user interest tags, demographic characteristics distribution, and significant improvements or declines in various indicators compared to the parent node user cluster.

[0199] Furthermore, in one embodiment of the present invention, the system further includes:

[0200] The monitoring module is used to continuously monitor the cohesion or dispersion of behavioral patterns within each user cluster.

[0201] The marking module is used to mark a target user cluster as a failed cluster when the cohesion of a target user cluster is lower than a preset threshold or the dispersion is higher than a preset threshold.

[0202] The update module is used to automatically re-execute the steps of the clustering process selection module, the first clustering module, and the second clustering module for users within the failed cluster or for all users, in order to generate an updated user cluster.

[0203] The user clustering system based on behavioral path analysis provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0204] Example 3

[0205] In another aspect, the present invention also proposes a computer device, see [link to relevant documentation]. Figure 3 The computer device shown is a third embodiment of the present invention, including a memory 20, a processor 10, and a program 30 stored in the memory 20 and executable on the processor. When the processor 10 executes the program 30, it implements the user clustering method based on behavior path analysis as described in the above embodiment.

[0206] In some embodiments, the processor 10 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip, used to run program code stored in memory 20 or process data, such as executing access restriction programs.

[0207] The memory 20 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (such as SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 20 can be an internal storage unit of a computer device, such as the hard disk of the computer device. In other embodiments, the memory 20 can also be an external storage device of a computer device, such as a Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device. Furthermore, the memory 200 can include both internal and external storage units of the computer device. The memory 200 can be used not only to store application software and various types of data installed on the computer device, but also to temporarily store data that has been output or will be output.

[0208] It should be pointed out that, Figure 3 The structure shown does not constitute a limitation on the computer device. In other embodiments, the computer device may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0209] This invention also provides a readable storage medium storing a program that, when executed by a processor, implements the user clustering method based on behavioral path analysis as described in the above embodiments.

[0210] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units or modules as needed, that is, the internal structure of the storage device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application.

[0211] Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "storage medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0212] More specific examples (a non-exhaustive list) of readable storage media include: electrical connections (electronic devices) with one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, the storage medium can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in memory.

[0213] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0214] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0215] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A user clustering method based on behavioral path analysis, characterized in that, The method comprises: collecting interaction behavior data of users and constructing a behavior path composed of atomic behaviors arranged in time sequence for each user; analyzing at least one preset feature of the behavior path, and adaptively selecting a matching first clustering process or a second clustering process for the behavior path according to the analysis result of the preset feature; when the first clustering process is selected, dividing users into different user clusters by a process of iteratively pruning features and repeatedly performing clustering based on subsequence features composed of continuous atomic behaviors extracted from the behavior path, to realize multi-level division of users; when the second clustering process is selected, dividing users into different user clusters based on a path semantic vector generated by performing aggregation operation on low-dimensional embedding vectors corresponding to each atomic behavior in the behavior path. 2.The user clustering method based on behavior path analysis of claim 1, wherein, The step of iteratively pruning features and repeatedly performing clustering based on subsequence features composed of continuous atomic behaviors extracted from the behavior path comprises: traversing the behavior path using a preset length sliding window to extract all subsequence features of continuous atomic behaviors conforming to the preset length, to form an initial feature subspace; performing clustering on user behavior paths in the feature subspace used in the current round to obtain user clusters in the current round; identifying and removing subsequence features with statistical significance that define user clusters in the current round to form a pruned new feature subspace for the next round of iteration; determining whether a preset iteration termination condition is met, and if not, using the new feature subspace as the feature subspace used in the current round and returning to perform the step of clustering user behavior paths. 3.The user clustering method based on behavior path analysis of claim 2, wherein, The step of identifying and removing subsequence features with statistical significance that define user clusters in the current round comprises: for each user cluster generated by clustering in the current round, constructing a contingency table describing the distribution of each subsequence feature in and out of the user cluster; calculating the chi-square statistic of each subsequence feature with respect to the user cluster based on the contingency table; identifying and selecting a preset number of subsequence features with the highest global chi-square statistic as significant features that need to be removed from the current feature space. 4.The user clustering method based on behavior path analysis of claim 1, wherein, The step of dividing users into different user clusters based on a path semantic vector generated by performing aggregation operation on low-dimensional embedding vectors corresponding to each atomic behavior in the behavior path comprises: obtaining low-dimensional embedding vectors corresponding to each atomic behavior in the behavior path; calculating the weight of each atomic behavior in the behavior path, which is determined by the importance of the atomic behavior in the current behavior path and the rarity of the atomic behavior in all user behavior paths; multiplying the low-dimensional embedding vector of each atomic behavior by its corresponding weight to obtain a weighted vector; performing summation and normalization aggregation operation on all weighted vectors in the behavior path to generate the path semantic vector, and dividing users into different user clusters based on the path semantic vector.

5. The user clustering method based on behavioral path analysis according to claim 4, wherein, The step of calculating the weight of each atomic behavior in the behavior path comprises: obtaining a frequency of occurrence of the atomic behavior in a current behavior path as a measure of the importance; obtaining a total number of user paths containing the atomic behavior, and calculating an inverse frequency value as a measure of the rarity according to the total number of user paths; multiplying the measure of the importance and the measure of the rarity to obtain a final weight of the atomic behavior. 6.The user clustering method based on behavior path analysis of claim 1, wherein, The step of adaptively selecting a matching first clustering process or a second clustering process for the behavior path comprises: obtaining a preset feature of the user, the preset feature being at least one of a user registration duration, a length of the behavior path, or a number of unique atomic behaviors contained in the behavior path; determining whether an analysis result of the preset feature meets a first preset condition or a second preset condition; when the analysis result meets the first preset condition, selecting a matching first clustering process for the behavior path; when the analysis result meets the second preset condition, selecting a matching second clustering process for the behavior path. 7.The user clustering method based on behavior path analysis of claim 1, wherein, The method further comprises: when the users are divided into different user clusters through the first clustering process, combining sub-sequence features that are excluded in each iteration clustering and have statistical significance as user cluster portrait labels of a corresponding level, thereby generating hierarchical multi-granularity portraits for the users; when the users are divided into different user clusters through the second clustering process, calculating a center vector of each user cluster in a low-dimensional embedding space, and finding an atomic behavior that is semantically closest to the center vector as an interest label of the user cluster. 8.The user clustering method based on behavior path analysis of claim 1, wherein, The step of collecting interaction behavior data of the users and constructing a behavior path composed of atomic behaviors arranged in time sequence for each user comprises: encoding each interaction behavior of the user into a structured atomic behavior according to a scenario in which the interaction occurs, a content type of the interaction, and a consumption feature of the user; discretizing continuous consumption behavior indicators including a user stay duration and a content completion degree through a data-driven non-uniform binning method to generate discrete encoding of the consumption feature.

9. A user clustering system based on behavioral path analysis, characterized by, The system comprises: a behavior path construction module configured to collect interaction behavior data of the users and construct a behavior path composed of atomic behaviors arranged in time sequence for each user; a clustering process selection module configured to analyze at least one preset feature of the behavior path, and adaptively select a matching first clustering process or a second clustering process for the behavior path according to an analysis result of the preset feature; a first clustering module configured to, when the clustering process selection module selects the first clustering process, divide the users into different user clusters based on sub-sequence features composed of continuous atomic behaviors extracted from the behavior path through a process of iteratively pruning features and repeatedly performing clustering, to realize multi-level division of the users; The second clustering module is configured to divide users into different user clusters based on a path semantic vector generated by performing an aggregation operation on the low-dimensional embedding vectors corresponding to the atomic behaviors in the behavior path when the second clustering process is selected by the clustering process selection module.

10. A computer device, comprising: A computer readable storage medium storing a program, the program being executable by a processor to implement the user clustering method based on behavior path analysis according to any one of claims 1-8.

Images