Activity matching method and system based on user behaviors

Abstract

The invention relates to the technical field of user behavior analysis, and discloses an activity matching method and system based on user behaviors, and the method comprises the steps: carrying out the full-dimension collection of behavior data of a user in a multi-channel marketing environment, and obtaining a structured user behavior data set; constructing dynamic user portrait data based on the structured user behavior data set; performing three-layer tagging processing of basic attributes, content features and experience values on the original activity data to obtain structured activity tag data; performing asynchronous iterative calculation and intelligent weighted matching according to the dynamic user portrait data and the structured activity tag data to obtain a personalized activity recommendation list; and executing a multi-brand cooperation strategy and global constraint processing on the personalized activity recommendation list to obtain a target matching scheme meeting cross-scene requirements. According to the method, coordinated activity recommendation in a multi-brand environment is realized, and the technical problem of mutual conflict of different brand activity recommendation is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of user behavior analysis, and particularly to an activity matching method and system based on user behavior. Background Art

[0002] The traditional marketing model has been difficult to meet the diversified needs of enterprises and consumers. Existing digital marketing systems generally have the problem of data silos, and user behavior data is scattered in different channel platforms, resulting in enterprises being unable to comprehensively grasp the user demand characteristics. Activity recommendations often rely on simple demographic characteristics or historical purchase records, lacking an understanding of users' deep interests and behavior patterns. In addition, the traditional keyword matching mode is too rigid to flexibly associate user preferences with multi-scenario activities, resulting in a significant gap between the recommended results and the actual needs of users.

[0003] When existing technical solutions deal with the problem of user behavior and activity matching, they often adopt synchronous calculation methods, which cannot cope with the real-time processing requirements of massive user data and are difficult to consider users' historical behaviors, content preferences, and multi-scenario linkage relationships at the same time. Especially in a complex marketing environment involving multi-brand collaboration, there is a lack of an effective brand relationship processing mechanism, resulting in problems such as brand conflicts and user experience fragmentation in the recommended results. At the same time, existing solutions are difficult to achieve precise recommendations while protecting user privacy, making the contradiction between data value and user privacy protection increasingly prominent. Summary of the Invention

[0004] The present invention provides an activity matching method and system based on user behavior. This method realizes coordinated activity recommendations in a multi-brand environment and solves the technical problem of conflicting activity recommendations between different brands.

[0005] In a first aspect, the present invention provides an activity matching method based on user behavior. The activity matching method based on user behavior includes: Collecting the behavior data of users in a multi-channel marketing environment in all dimensions to obtain a structured user behavior data set; Constructing dynamic user portrait data based on the structured user behavior data set; Performing three-layer tagging processing on the original activity data for basic attributes, content features, and experience value to obtain structured activity tag data; Performing asynchronous iterative calculation and intelligent weighted matching according to the dynamic user portrait data and the structured activity tag data to obtain a personalized activity recommendation list; Executing a multi-brand collaboration strategy and global constraint processing on the personalized activity recommendation list to obtain a target matching solution that meets cross-scenario requirements.

[0006] Second aspect, the present invention provides an activity matching system based on user behavior, and the activity matching system based on user behavior includes: A collection module, configured to perform all-dimensional collection on the behavior data of a user in a multi-channel marketing environment to obtain a structured user behavior data set; A construction module, configured to construct dynamic user portrait data based on the structured user behavior data set; A processing module, configured to perform three-layer tagging processing on the original activity data for basic attributes, content features, and experience value to obtain structured activity tag data; A calculation module, configured to perform asynchronous iterative calculation and intelligent weighted matching according to the dynamic user portrait data and the structured activity tag data to obtain a personalized activity recommendation list; A collaboration module, configured to execute a multi-brand collaboration strategy and global constraint processing on the personalized activity recommendation list to obtain a target matching solution that meets cross-scenario requirements.

[0007] In the technical solution provided by the present invention, through the all-channel data collection and five-dimensional data structure organization of web pages, mobile applications, offline interactions, and social media, the data island problem of traditional methods is solved. At the same time, by separating the explicit behavior and implicit preference processing and the time series analysis model, a dynamic user portrait containing an eight-dimensional feature vector is constructed to accurately perceive the change of user preference; the three-layer tagging processing of the basic attributes, content features, and experience value of the activity, combined with the activity association analysis, realizes the multi-dimensional three-dimensional description of the activity content and significantly improves the matching accuracy; based on the asynchronous iterative calculation of the user-activity bipartite graph model, combined with the three-channel matching and Q-learning reinforcement learning, the performance bottleneck of traditional synchronous calculation is broken through, effectively avoiding the "information cocoon" problem; through brand network analysis, collaborative optimization processing, and integer programming model, the coordinated recommendation of activities in a multi-brand environment is realized, and the technical problem of mutual conflict in the recommendation of different brand activities is solved; by using the explicit and implicit feedback channels, multi-level learning engines, and abnormal pattern recognition mechanisms, a complete feedback-driven closed-loop system is constructed, enabling the matching algorithm to continuously self-optimize and realizing the long-term stable improvement of the system performance.

[0008] Other features and advantages of the present invention will be described in the following specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention are achieved and obtained by the structures specifically pointed out in the specification, claims, and drawings.

[0009] To make the above objectives, features, and advantages of the present invention more obvious and understandable, the following specific preferred embodiments are given, and in conjunction with the accompanying drawings, the detailed description is as follows. Description of the Drawings

[0010] Figure 1 Schematic diagram of an embodiment of the activity matching method based on user behavior in an embodiment of the present invention; Figure 2 Schematic diagram of an embodiment of the activity matching system based on user behavior in an embodiment of the present invention. Detailed implementation manners

[0011] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0012] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but optionally further includes other unlisted steps or units, or optionally further includes other steps or units inherent to these processes, methods, products or devices.

[0013] For ease of understanding of this embodiment, a method for activity matching based on user behavior disclosed in the embodiments of the present invention will be introduced in detail first. As Figure 1 shown, the method includes the following steps: 101. Collect all-dimensional behavior data of users in a multi-channel marketing environment to obtain a structured user behavior data set; It can be understood that the execution subject of the present invention may be an activity matching system based on user behavior, or a terminal or a server. Specifically, it is not limited here. An example will be given with the server as the execution subject in the embodiments of the present invention.

[0014] Specifically, through the web behavior tracking module, mobile application monitoring module, offline interaction capture module, and social media analysis module, the behavioral data of users in multiple channels is collected to obtain the original behavioral data. In the web behavior tracking module, by embedding behavior tracking codes, such as JavaScript scripts or pixel tags, in the web page, real-time monitoring of user behaviors such as clicks, scrolls, dwell times, and page jump paths is achieved. The mobile application monitoring module records the operation behaviors of users in mobile applications through the SDK (Software Development Kit) integrated in the APP, including detailed data such as startup, function usage, page browsing, and button clicks. The offline interaction capture module obtains the interaction data of users in offline scenarios through intelligent devices (such as cameras, sensors, RFID devices, etc.), such as behaviors like product browsing, trying on, and purchasing in physical stores. At the same time, the social media analysis module obtains the interaction data of users in social media in real time by connecting to the social platform API interface, including likes, comments, shares, follows, and topic participation, which reflect the interests and social behavior preferences of users. Perform timestamp unification, user ID anonymization, and behavior type standardization processing on the original behavioral data. Standardize the time formats of the behavioral data recorded on different channels and different devices and convert them into a unified timestamp format. On this basis, perform anonymization processing of the user ID. Through hashing algorithms (such as SHA-256, MD5, etc.), the original user identifiers (such as mobile phone numbers, email addresses, device IDs, etc.) are converted into irreversible anonymous IDs, so as to achieve the tracking and analysis of user behaviors without exposing the personal information of users. Standardize the behavior types, uniformly encode the behavior event names, operation types, etc. existing in different data sources, and eliminate the problem of inconsistent behavior semantics caused by different data sources, obtaining the initially processed behavioral data. Use a distributed architecture to desensitize and compress the initially processed behavioral data at the data source. During the desensitization process, perform fuzzification processing on sensitive information (such as the accuracy of geographical location, personal preference data in specific user operation records, etc.) to ensure that the privacy of users will not be leaked during data transmission. At the same time, data compression technologies (such as Snappy, Gzip, etc.) significantly reduce the volume of data packets, improve data transmission efficiency, and reduce network bandwidth occupancy. After desensitization and compression are completed, the data is packaged into a secure transmission data packet and transmitted to the central data processing center through an encrypted channel. The encrypted channel can effectively prevent eavesdropping, tampering, and forgery of data during transmission, ensuring the confidentiality and integrity of the data. Transmit the secure transmission data packet through the encrypted channel to the central data processing center for parsing to obtain centralized behavioral data. The parsing process includes decoding and decompressing the data format, and mapping and converting various fields in the data to ensure that the data from different sources maintains consistent structure in the same data model.After data parsing is completed, start the data quality monitoring and detection process, and review the data according to preset rules (such as the integrity of data fields, the rationality of numerical ranges, the consistency of time series, etc.). Once abnormal data (such as null values, duplicate data, data beyond reasonable ranges) is found, automatically screen out this abnormal data to avoid deviations in subsequent analysis results caused by data quality problems, and obtain valid behavior data. Organize and integrate the valid behavior data according to a five-dimensional data structure of user, time, behavior, scenario, and object to obtain a structured user behavior dataset. The user dimension records the anonymized user ID and its basic attributes (such as age group, gender preference, city where located, etc.); the time dimension provides the specific time point when the behavior occurs and time series characteristics (such as behavior frequency, activity level in a specific time period, etc.); the behavior dimension describes the operation types of users (such as click, browse, purchase, interaction, etc.); the scenario dimension is used to distinguish the channels where the behavior occurs (such as online, mobile, offline, social platforms, etc.) and specific scenarios (such as promotion activity pages, offline stores, social media posts, etc.); the object dimension represents the specific objects associated with user behavior (such as products, services, brands, activity content, etc.).

[0015] 102. Construct dynamic user portrait data based on the structured user behavior dataset; Specifically, the structured user behavior dataset is separated into explicit behavior data and implicit preference data. Explicit behavior data includes the explicit operations of users in a multi-channel environment, such as direct behaviors like clicking, browsing, purchasing, signing up for activities, etc. This type of data is suitable for calculating weights through the direct assignment method due to its strong certainty, quantifiability, and easy acquisition. At the same time, implicit preference data is the user preference inferred from the user's behavior patterns, behavior frequencies, and the correlations between behaviors. For example, the potential interest of users in a certain type of activity, the long-term preference for a specific brand, etc. This type of data is weighted by the time decay method to reflect the timeliness of user preferences. The time decay weighting method assigns higher weights to behaviors closer to the current time, making the model more sensitive to capturing changes in users' recent interests. Through this step, the weighted behavior feature data reflects the preference intensity of users in different time dimensions. Based on the weighted behavior feature data, a preliminary user feature vector is constructed. The preliminary user feature vector covers multi-dimensional feature information of users, including demographic characteristics (such as age, gender, region), interest preferences (the degree of interest in different types of activities and products), consumption ability (obtained through the analysis of historical consumption behaviors), brand interaction (the frequency and depth of interaction with various brands), time patterns (the time periods when users are active in a day or a week), social influence (analyzing the influence of users in the group based on social media interaction data), scene preferences (the behavior patterns of users in specific scenarios), and activity responses (the enthusiasm and feedback for participating in activities). Similar user group feature migration is performed on the preliminary user feature vector. Feature migration uses clustering algorithms (such as K-Means, GMM, etc.) to find user groups similar to the current user's characteristics, analyzes the collective characteristics of these groups, and migrates some representative or predictive characteristics to the feature vector of the current user. Historical change trend analysis is performed on the complete feature vector to extract the temporal characteristics of user preferences over time. Through time series analysis methods, the change trajectories of users in specific characteristics are analyzed. The complete feature vector and the temporal characteristics are integrated to generate a standardized data structure containing user ID, feature vector, confidence score, and timestamp, and finally, dynamic user portrait data is obtained. The user ID is used to uniquely identify the user, the feature vector records the feature distribution of the user in each dimension, the confidence score represents the reliability and data coverage of the current portrait, and the timestamp marks the latest update time of the portrait, supporting the dynamic update mechanism of the portrait.

[0016] 103. Perform three-layer tagging processing on the original activity data for basic attributes, content features, and experience value to obtain structured activity tag data; Specifically, receive the basic activity information including the activity name, time, location, content description, target audience, and participation method, which constitute the original data of the activity. Process the original activity data through a rule engine to extract the basic attributes of the activity, including activity type, time attribute, location attribute, participation threshold, and scale, etc. These basic attributes constitute the first-level basic attribute tags of the activity. Perform semantic analysis on the description text in the original activity data through natural language processing technology to extract keywords and map them to a preset content classification system to obtain the second-level content feature tags of the activity. Through text analysis and semantic understanding technologies in natural language processing. By performing operations such as word segmentation, named entity recognition, and keyword extraction on the activity description, identify the keywords related to the activity, and map these keywords to a preset content classification system, such as technology, education, art, entertainment, sports, etc. Combine the activity objectives, forms, and feedback information from historical participants in the original activity data to conduct an experience value analysis to obtain the third-level experience value tags of the activity. Evaluate the activity from the perspective of user experience and analyze what kind of experience value the activity can bring to participants. The activity objectives and forms can reflect the purpose and organization method of the activity itself, while the feedback from historical participants can reflect the actual feelings and experiences of the participants in the activity. Conduct an activity relevance analysis on the first-level basic attribute tags, second-level content feature tags, and third-level experience value tags to obtain the relationships of marked series activities, complementary activities, and competitive activities. By calculating the similarity between activity tags, identify those activities with similar content features, that is, "series activities"; by analyzing the complementarity of activities to users in different scenarios, identify "complementary activities"; by comparing the activity objectives and forms, find those activities that compete with the current activity, that is, "competitive activities". Integrate the first-level basic attribute tags, second-level content feature tags, third-level experience value tags, and the association tags between activities, and assign weight values to each tag to obtain structured activity tag data including activity ID, tag set, weight value, timeliness, and reference to related activities.

[0017] 104. Perform asynchronous iterative calculation and intelligent weighted matching based on the dynamic user portrait data and the structured activity tag data to obtain a personalized activity recommendation list; Specifically, based on the dynamic user profile data and structured activity label data, a user-activity bipartite graph model is created. In this model, the nodes are divided into two categories, representing the user feature vector and the activity feature vector respectively, and the edge weights between the nodes represent the initial matching degree between the user and the activity. By mapping the user features and activity features into the same graph, the bipartite graph model reveals the potential connections between the user and different activities. The calculation of the initial matching degree is based on vector similarity calculation. For example, the cosine similarity or Euclidean distance is used to measure the matching degree between the features in the user profile and the features in the activity label. On the user-activity bipartite graph model, multi-channel matching calculations of content similarity calculation, collaborative filtering analysis, and context factor evaluation are sequentially performed to obtain the content matching degree score result, the collaborative matching degree score result, and the context matching degree score result. The content similarity calculation is achieved by comparing the explicit interests in the user profile and the content feature labels of the activity. For example, the preference of the user for a specific theme activity is matched with the keywords in the activity content, and natural language processing technologies such as TF-IDF and Word2Vec are used. Collaborative filtering analysis utilizes the similarity of the behavior of the user group. By analyzing the activity participation records of other users with similar behaviors to the target user, the activities that the target user is interested in are predicted. The context factor evaluation combines external information such as the time, geographical location, and device type where the user is located to adjust the matching degree score, ensuring that the recommended activities not only match the user's interests in terms of content but also are feasible to participate in the actual scenario. For example, mechanisms such as geographical distance calculation and time window filtering are introduced to improve the accuracy of the matching. To improve the intelligence and dynamics of the matching, according to the richness of the user data and the characteristics of the activity, an adaptive weight fusion calculation is performed on the content matching degree score result, the collaborative matching degree score result, and the context matching degree score result to obtain the comprehensive matching degree score. The allocation of the adaptive weight is achieved by analyzing the comprehensiveness of the user profile data and the significance of the activity characteristics. For example, for users with rich data, more reliance is placed on the content matching and collaborative filtering scores, while for new users or users with sparse data, the weight of the context factor is increased to avoid the "cold start" problem. The fusion calculation process adopts the weighted average or the method based on Bayesian optimization to dynamically adjust the weights of each matching channel, enabling the model to adapt to the differences of different users and activity scenarios. The comprehensive matching degree score is input into an asynchronous Q-learning reinforcement learning model for dynamic optimization. In the Q-learning model, the interaction between the user and the system is regarded as a reinforcement learning process, where the user's behavior feedback (such as clicks, participation, evaluations, etc.) is used as the reward signal. Through the iterative update of the Q value, the system gradually learns the optimal activity recommendation strategy. Asynchronous Q-learning can achieve efficient parallel computing in large-scale user and activity data scenarios, avoiding the slow convergence problem of the traditional Q-learning model in the big data environment.By introducing an exploration-exploitation balance mechanism, while recommending familiar activities, the model introduces new types of activities to users to avoid excessive convergence of recommendation results and improve the diversity and user satisfaction of recommendations. The exploration-exploitation balance algorithm is executed based on the target matching strategy, and the matching results are dynamically adjusted to achieve a balance between the security and novelty of the recommended content. For example, when calculating the recommendation list, through the ε-greedy strategy or the UCB-based algorithm, a certain probability is reserved for exploring new activities among the activities with high matching degrees. By identifying the preference associations of users in different scenarios, cross-scenario combination analysis is performed on the candidate recommendation set to generate a personalized activity recommendation list containing activity IDs, matching scores, matching reasons, and expected participation probabilities. The cross-scenario combination analysis dynamically adjusts the activity sorting and display methods in the recommendation list by analyzing the behavior patterns of users in different situations.

[0018] 105. Execute the multi-brand collaboration strategy and global constraint processing on the personalized activity recommendation list to obtain a target matching solution that meets the cross-scenario requirements.

[0019] Specifically, by applying activity capacity constraints, time conflict constraints, and geographical distance constraints, the recommendation list is globally screened to obtain a preliminary matching solution that meets the hard limit conditions. The activity capacity constraint is based on the upper limit of the number of participants in each activity and the current number of registered participants, ensuring that the recommended activities will not prevent users from actually participating due to exceeding the capacity. The time conflict constraint filters out the recommendations that conflict with the user's existing schedule in terms of time by analyzing the user's existing schedule and the time of the recommended activities, avoiding the situation where users cannot participate due to overlapping time in actual operation. The geographical distance constraint calculates the actual distance between the recommended activities and the user based on the user's current location or permanent residence. For offline activities, a distance threshold is set. For example, based on the acceptable driving or walking time range of the user, only the activities that the user is likely to actually attend are retained. Brand network analysis is performed on the solution based on brand relationship data. By identifying the relationships between competing brands, complementary brands, and cooperative brands, brand association marking results are generated. This analysis relies on a brand relationship database that contains the competition history between brands, joint activity records, cross-brand user behavior data, etc. By constructing a brand network model and using graph analysis algorithms to calculate the relationship strength and type between brand nodes, it is identified which brands have direct competition relationships in a specific activity field, which brands are complementary, or which brands have long-term cooperative relationships. For example, if brand A and brand B have jointly held joint promotion activities many times in history, these two brands are marked as cooperative brands, while if brand C and brand D often compete for user groups in similar activities, they will be marked as competing brands. Based on the brand association marking results, brand collaborative optimization processing is performed on the preliminary matching solution. In the brand collaborative optimization process, different recommendation rules are set according to different brand relationships. For example, for competing brands, to avoid users receiving recommendations for multiple similar brand activities in a short period of time, the number of such activities in the recommendation list is restricted, thus avoiding the user's selection difficulties or reduced brand preference caused by brand conflicts. For complementary brands, through enhanced recommendations, the activities of related brands are combined and recommended. For the activities of known cooperative brands, joint activities are given priority recommendations to enhance the attractiveness of the activities by leveraging the synergy between brands. To prevent the recommendation strategy from overly relying on specific brands or activity types, combined with the user's historical activity participation frequency and feedback, fatigue assessment and adjustment are performed on the target combination to obtain an activity combination with a reasonable recommendation rhythm. Fatigue assessment determines whether users have developed aesthetic fatigue for a certain type of activity by analyzing the frequency of the user's participation in activities of a certain brand and the participation feedback (such as post-participation evaluations, participation conversion rates, etc.) over a certain period of time. For example, if the user has participated in activities of the same brand multiple times in a short period and the feedback is mediocre, the system will appropriately reduce the recommendation frequency of this brand's activities, introduce new brands and new activities through algorithms to ensure the reasonableness of the recommendation rhythm, and keep users fresh about the recommended content.Input the activity combination with a reasonable recommendation rhythm into an integer programming model for overall value maximization operation to optimize the recommendation plan. The integer programming model quantifies the value of activity recommendations (such as participation probability, brand benefits, user satisfaction) into a mathematical form by establishing an objective function. Under the premise of considering multiple constraints such as activity capacity, time arrangement, brand relationship, and user preferences, it solves the optimal solution of the activity combination. This process uses optimization algorithms such as linear programming and mixed integer programming to ensure the maximization of user value and brand value within the limited recommendation space by calculating the optimal strategy for resource allocation. Perform personalized processing on the display strategy for the optimal activity plan according to the user's information reception preference to obtain a target matching plan that meets cross-scenario requirements. In the display strategy processing, select the most appropriate display form and display timing according to the user's behavior patterns under different devices, different times, and different interaction habits.

[0020] Multi - dimensionally collect user response data of the target matching solution that meets cross - scenario requirements through explicit feedback channels and implicit feedback channels. The explicit feedback channels include users' explicit interaction behaviors on the platform, such as clicks, sign - ups, purchases, evaluations, and scores and comments after participating in activities. This type of data has high certainty and directness. The implicit feedback channels, on the other hand, obtain implicit feedback data by capturing users' indirect behavior characteristics, such as the duration of stay on the activity page, browsing paths, interaction frequencies, cursor movement trajectories, and the open rate of activity push messages. These data reflect changes in users' potential interests and preferences. By combining explicit and implicit feedback data, a user feedback dataset is constructed. Apply a differential weight strategy to calculate scores for different types of feedback in the user feedback dataset to obtain weighted user feedback scores. The differential weight strategy dynamically assigns different weights according to the signal - to - noise ratio of the feedback type, the certainty of the feedback behavior, and the relevance to the activity match. For example, a higher weight is given to users' explicit sign - up or participation behaviors, while a lower weight is given to cases where users only browse but do not generate participation behaviors. For the potential interests reflected in implicit feedback (such as long - term stays and frequent clicks on detailed content), the weight value is dynamically adjusted through time decay or behavior frequency analysis. Input the weighted user feedback scores into a multi - level learning engine for gradient processing. The multi - level learning engine includes three major modules: short - term immediate update, medium - term batch update, and long - term deep reconstruction, which are hierarchically processed for short - term fluctuations of user interests, periodic updates of stable features, and long - term optimization of core algorithms respectively. In the short - term immediate update, according to the latest user feedback, the temporary interest tags in the user profile are adjusted in real - time. For example, when a user frequently participates in a certain type of activity in a short period, the system immediately increases the priority of relevant activities in the recommendation list. The medium - term batch update periodically (such as weekly or monthly) collects and accumulates feedback data to batch - update the stable features in the user profile, such as users' long - term interest preferences, consumption habits, and behavior patterns. The long - term deep reconstruction module is based on deep learning models (such as reinforcement learning, convolutional neural networks) to conduct an overall reconstruction of the recommendation algorithm, ensuring that the system can perform deep adaptive adjustments in the face of long - term changes in user behavior, thus avoiding the problem of recommendation failure caused by outdated models. After the multi - level learning engine completes the output of optimized parameters, conduct A / B test comparison and analysis on different matching strategies based on multi - dimensional optimized parameters to quantify the actual effects of different recommendation strategies. The A / B test randomly divides users into a control group and an experimental group, applies different recommendation strategies respectively, and compares the performance differences of these two groups in key indicators (such as click - through rate, participation rate, conversion rate, user satisfaction, etc.) to evaluate the advantages and disadvantages of different strategies. For example, by comparing the "content - based recommendation" strategy and the "collaborative filtering recommendation" strategy, observe the response differences of users in activity participation to select a more effective recommendation method.During the A / B testing process, not only the overall effect is analyzed, but also the strategy performance of different user groups and different scenarios is segmented to obtain more refined effect evaluation results. Based on the A / B testing, an anomaly pattern recognition mechanism is used to identify and extract features from the abnormal phenomena in the evaluation results of strategy effect differences, obtaining user interest change points and reaction anomaly points. The anomaly pattern recognition mechanism can quickly identify situations in user behavior that are significantly different from previous patterns through statistical analysis and machine learning (such as clustering analysis, anomaly detection model Isolation Forest, etc.). For example, a certain type of activity suddenly receives a large amount of attention, or the behavior characteristics of a certain user group change significantly. This mechanism can not only help the system capture the changing trends of user interests and timely discover potential problems in the recommendation system, such as recommended content deviating from the actual needs of users and some recommendation strategies having adverse effects. To convert the identified user interest change points and reaction anomaly points into specific recommendation optimization plans, a full-link effect attribution model is used for analysis. The full-link effect attribution model integrates the entire link data of users from receiving recommendations, browsing activities, participating in activities to final feedback, and adopts the method of "synergistic analysis of three libraries in one", that is, data linkage between the user portrait library, activity label library, and brand relationship library. Through attribution analysis models (such as Shapley value decomposition, multi-touch attribution model), the contribution of each link to the final recommendation effect is quantified, thereby accurately identifying the key factors affecting the recommendation effect. Through full-link analysis, intelligent recommendation results supporting full-link marketing are obtained. This result includes preset performance indicators (such as expected participation rate, brand exposure, activity conversion rate), and automatically generates a problem diagnosis report, clearly pointing out the weak links in the current recommendation strategy, and providing targeted improvement suggestions in combination with historical data and model predictions.

[0021] In the embodiments of the present invention, through the full-channel data collection and five-dimensional data structure organization of web pages, mobile applications, offline interactions, and social media, the data island problem of traditional methods is solved. At the same time, by separating and processing explicit behaviors and implicit preferences and using a time-series analysis model, a dynamic user profile containing eight-dimensional feature vectors is constructed to accurately perceive changes in user preferences; the activities are processed with three-layer tagging of basic attributes, content features, and experience values, and combined with activity association analysis, a multi-dimensional three-dimensional description of the activity content is realized, significantly improving the matching accuracy; based on the asynchronous iterative calculation of the user-activity bipartite graph model, combined with three-channel matching and Q-learning reinforcement learning, the performance bottleneck of traditional synchronous calculation is broken through, effectively avoiding the "information cocoon" problem; through brand network analysis, collaborative optimization processing, and integer programming models, coordinated activity recommendations in a multi-brand environment are realized, solving the technical problem of mutual conflicts in the recommendation of different brand activities; by using explicit and implicit feedback channels, a multi-level learning engine, and an abnormal pattern recognition mechanism, a complete feedback-driven closed-loop system is constructed, enabling the matching algorithm to continuously self-optimize and realizing long-term stable improvement of the system performance.

[0022] In a specific embodiment, the process of executing step 101 may specifically include the following steps: Collect user full-channel behavior data through a web behavior tracking module, a mobile application monitoring module, an offline interaction capture module, and a social media analysis module to obtain raw behavior data; Perform timestamp unification, user ID anonymization, and behavior type standardization processing on the raw behavior data to obtain preliminarily processed behavior data; Use a distributed architecture to desensitize and compress the preliminarily processed behavior data at the data source to obtain a secure transmission data packet; Transmit the secure transmission data packet through an encrypted channel to a central data processing center for parsing to obtain centralized behavior data, and perform data quality monitoring and detection on the centralized behavior data to screen out abnormal data and obtain valid behavior data; Organize and integrate the valid behavior data according to a five-dimensional data structure of user, time, behavior, scenario, and object to obtain a structured user behavior data set.

[0023] Specifically, user behavior data is comprehensively collected from different channels through a web behavior tracking module, a mobile application monitoring module, an offline interaction capture module, and a social media analysis module. In the web behavior tracking module, by implanting tracking scripts (such as JavaScript code, pixel tracking tools, etc.) in web pages, various behaviors of users on the website are recorded, including page views, clicks, scrolls, form fills, and shopping cart operations, etc. For example, when a user browses a certain product page on an e-commerce platform, the page URL, the product link clicked by the user, the user's stay time on the page, and the interaction frequency with the page content are recorded. The mobile application monitoring module realizes real-time monitoring of user operation behaviors by integrating an SDK (Software Development Kit) into the mobile APP, including operations such as APP startup, page switching, button clicks, and shopping processes. The offline interaction capture module obtains users' offline behavior data through intelligent devices (such as NFC devices, cameras, sensors, etc.). For example, in a shopping mall, the user's movement trajectory is recorded through Wi-Fi probes, or the user's staying time in front of a certain exhibition booth is counted through cameras and face recognition technology. The social media analysis module obtains the interaction data of users on social media platforms through API interfaces or web scraping technologies. Perform normalized preprocessing on the original behavior data. Perform unified timestamp processing, converting all time data into a unified timestamp format (such as UNIX timestamp), so that in subsequent time series analysis, the time dimensions of all data can be accurately aligned. User ID anonymization processing is to protect user privacy and security during the data analysis process. By using a hash algorithm (such as SHA-256) or an encryption algorithm, the user's real identity information (such as mobile phone number, email, device ID, etc.) is converted into an irreversible anonymous ID, thus effectively avoiding the risk of user privacy leakage. Behavior type standardization processing is to unify different expressions that describe the same behavior in data from different sources, which helps to eliminate data semantic ambiguities when fusing multi-channel behaviors. After completing the preprocessing, use a distributed architecture (such as big data processing platforms based on Apache Kafka, Flink, Spark, etc.) to perform desensitization and compression processing on the preliminarily processed behavior data at the data source to generate secure transmission data packets. Data desensitization targets fields involving personal privacy or sensitive information, and through masking, encryption, or obfuscation processing, keeps the data secure during transmission. Data compression significantly reduces the data volume by using compression algorithms such as Snappy and Gzip, to improve transmission efficiency and reduce storage costs. The secure transmission data packets are transmitted to the central data processing center through an encrypted channel (such as the SSL / TLS protocol). In the central data processing center, these data packets undergo decryption and decompression operations and are restored to centralized behavior data.During the data parsing process, format and map data fields from different sources, standardize the field names, data types, and data formats of each data source, and eliminate potential problems caused by inconsistent data formats. After completing the data parsing, perform data quality monitoring and detection on the centralized behavioral data. Through predefined rules and models (such as field integrity detection, numerical rationality verification, time series continuity analysis, etc.), screen out abnormal data to obtain valid behavioral data. Organize and integrate the valid behavioral data according to the five-dimensional data structure of user, time, behavior, scenario, and object, and finally construct a structured user behavior dataset. In this five-dimensional data model, the user dimension includes the user ID and basic information in the user profile. The time dimension includes the specific time when the behavior occurs and the time series characteristics. The behavior dimension records the specific operation types of the user (such as click, purchase, share, etc.). The scenario dimension describes the channels and devices where the behavior occurs (such as web pages, mobile devices, offline scenarios, social platforms, etc.). The object dimension defines the specific targets involved in the user behavior (such as products, activities, brands, content, etc.).

[0024] In a specific embodiment, the process of executing step 102 may specifically include the following steps: Separate the structured user behavior dataset into explicit behavior data and implicit preference data; Apply the direct assignment method to calculate weights for the explicit behavior data, and apply the time decay weighted method to calculate weights for the implicit preference data to obtain weighted behavior feature data; Construct a preliminary user feature vector based on the weighted behavior feature data, including demographics, interest preferences, consumption ability, brand interaction, time pattern, social influence, scenario preference, and activity response; Perform similar user group feature migration on the preliminary user feature vector to obtain a complete feature vector, and perform historical change trend analysis on the complete feature vector to obtain time series features reflecting the time change characteristics of user preferences; Integrate the complete feature vector and the time series features into a standardized data structure including user ID, feature vector, confidence score, and timestamp to obtain dynamic user portrait data.

[0025] Specifically, the structured user behavior dataset is separated into explicit behavior data and implicit preference data. Explicit behavior data includes users' explicit operations, such as clicks, purchases, sign-ups for activities, ratings, etc. These data are sourced from users' specific interactions on web pages, mobile applications, and offline scenarios. Since such behavior data is direct and clear, the weights are calculated using the direct assignment method. The direct assignment method assigns fixed weights to each behavior based on the actual value of the behavior. Implicit preference data infers users' potential interests through their indirect behaviors, such as the duration of stay on a specific page, the speed of swiping the page, the frequency of repeated visits to specific types of content, etc. The weights for implicit preference data are calculated using the time decay weighting method. The time decay weighting method can dynamically adjust the weights of behaviors, making the most recent behaviors have a higher influence, while the influence of earlier behaviors gradually weakens. This method is implemented through an exponential decay model, and the weight calculation formula is:

[0026] Among them, represents the weighted value of the th implicit behavior, is the basic weight value, preset according to the influence of the behavior type; is the current time, is the time when the behavior occurred, is the time decay coefficient, which controls the weight decay speed. The time decay mechanism makes the system more sensitive to capturing the changes in user interests. Based on the weighted feature data of explicit and implicit behaviors, a preliminary feature vector of the user is constructed. The feature vector synthesizes the multi-dimensional features of the user, including demographics (age, gender, occupation), interest preferences (preferences for different activity types), consumption ability (evaluating purchasing power through historical consumption records), brand interaction (the frequency and depth of user interaction with the brand), time pattern (the time period when the user is active, such as being more active in the evening or on weekends), social influence (the influence of the user on the social platform, such as the number of comments, likes, and shares), scenario preference (the behavior pattern of the user in a specific scenario, such as being more inclined to browse activities on the mobile device or participate in activities through the PC), and activity response (the participation rate and feedback score of the user for the recommended activities). Similar user group feature migration is performed on the preliminary user feature vector. By analyzing the user groups similar to the current user's characteristics, potential characteristics not shown by the current user are inferred. This method of feature migration can effectively solve the cold start problem, especially when the user data is scarce or the behavior pattern changes rapidly. By learning from the behavior patterns of similar users, the recommendation system can better predict the potential needs of the user. Historical change trend analysis is performed on the complete feature vector to obtain time series features. This is achieved through time series analysis models (such as moving average models, LSTM long short-term memory networks, etc.). For example, by analyzing the change trends of different features of the user in the past six months, it is identified which features show an upward trend, which features tend to be stable or decline. Trend analysis is applicable to the adjustment of interest preferences and is used to predict the behavior changes of the user at a specific time point. The complete feature vector and the time series features are integrated into a standardized data structure containing user ID, feature vector, confidence score, and timestamp to obtain dynamic user portrait data. The user ID is used to uniquely identify the user, the feature vector represents the user's interests, preferences, and behavior characteristics in the form of multi-dimensional numerical values, the confidence score represents the credibility of the current user portrait. For example, when the user behavior data is relatively rich, the confidence is high (such as 0.9), while when the user data is scarce or updated slowly, the confidence is low (such as 0.6). The timestamp is used to mark the last update time of the user portrait, supporting the system to automatically refresh the user portrait according to the dynamic changes of user behavior.

[0027] In a specific embodiment, the process of executing step 103 may specifically include the following steps: Receive the basic activity information including activity name, time, location, content description, target audience, and participation method to obtain the original activity data; Extract the basic attributes of activity type, time attribute, location attribute, participation threshold, and scale from the original activity data through a rule engine to obtain the first-level basic attribute labels; Using natural language processing technology to perform semantic analysis on the description text in the original activity data to extract keywords, and obtaining second-level content feature tags mapped to a preset content classification system; Combining the activity objectives, forms, and historical participant feedback in the original activity data to conduct an experience value analysis, and obtaining third-level experience value tags representing knowledge acquisition, emotional satisfaction, and social expansion characteristics; Conducting an activity relevance analysis on the first-level basic attribute tags, second-level content feature tags, and third-level experience value tags, and obtaining association tags marking the relationships of series activities, complementary activities, and competitive activities; Integrating the first-level basic attribute tags, second-level content feature tags, third-level experience value tags, and association tags and assigning weight values to obtain structured activity tag data including activity ID, tag set, weight value, timeliness, and references to related activities.

[0028] Specifically, the basic information of the activity including the name, time, location, content description, target audience and participation method of the activity is received to form the original data of the activity. This information is automatically collected through the management background, activity publishing platform or third-party data interface. The original data of the activity is parsed by the rule engine, and basic attributes such as activity type, time attribute, location attribute, participation threshold and scale are extracted from it to generate the first-level basic attribute label. Based on the preset rule library, the rule engine automatically extracts these key information from the activity data through condition matching and logical judgment. In this process, the rule engine can not only process simple explicit information, but also expand the expression dimension of the activity label through rule reasoning. The content description in the original data of the activity is semantically analyzed by natural language processing technology. Representative keywords are extracted from the activity description text through word segmentation, part-of-speech tagging, keyword extraction, topic analysis and other technologies, and these keywords are mapped to the preset content classification system to generate the second-level content feature label. Combined with the activity goals, forms and historical participant feedback in the original data of the activity, the experience value analysis is carried out to obtain the third-level experience value label that represents the characteristics of knowledge acquisition, emotional satisfaction and social expansion. In this process, the goal of the activity is analyzed, and the experience value that the activity can provide is derived by combining the goal characteristics with the participant feedback in the historical data (such as satisfaction, participation frequency, and interaction depth). At the same time, the actual experience effect of the activity is evaluated by quantitatively analyzing the behavior data of participants in different activities (such as interaction frequency and feedback evaluation scores), and this experience is reflected in the label weight in a quantitative form. The activity correlation analysis is performed on the first-level basic attribute tags, the second-level content feature tags, and the third-level experience value tags to identify the relationship between activities, including marking series activities, complementary activities, and competitive activities, and then generate related tags. Activity correlation analysis realizes intelligent association between activities by constructing a multi-dimensional label network of activities. For example, based on graph analysis algorithms (such as PageRank and community detection algorithms), it automatically discovers which activities have a high degree of association in the user behavior network, so that related activities can be used as part of the linkage recommendation when recommending. The first-level basic attribute tags, the second-level content feature tags, the third-level experience value tags, and the related tags are integrated, and a weight value is assigned to each tag to obtain structured activity tag data containing activity ID, tag set, weight value, timeliness, and related activity references. In this process, the calculation of the weight value is dynamically adjusted by comprehensively considering the actual performance of the activity, the importance of the tag, and the user's behavior preferences, for example, through a weighted calculation formula:

[0029] in, represents the comprehensive label score of the activity, is the weight value of the label, indicating the importance of the label in a specific activity. is the feature value of the label, calculated based on the actual matching degree between the label and the activity. The timeliness is calculated by analyzing the start time and duration of the activity to calculate the effective period of the label.

[0030] In a specific embodiment, the process of executing step 104 may specifically include the following steps: Create a user-activity bipartite graph model with user feature vectors and activity feature vectors as nodes and the initial matching degree as the edge weight based on dynamic user profile data and structured activity label data; Perform multi-channel matching calculations including content similarity calculation, collaborative filtering analysis, and context factor evaluation on the user-activity bipartite graph model to obtain content matching degree scoring results, collaborative matching degree scoring results, and context matching degree scoring results; Perform adaptive weight fusion calculation on the content matching degree scoring results, collaborative matching degree scoring results, and context matching degree scoring results according to the richness of user data and activity characteristics to obtain a comprehensive matching degree score; Input the comprehensive matching degree score into an asynchronous Q-learning reinforcement learning model for dynamic optimization to obtain a target matching strategy considering user feedback; Execute an exploration-exploitation balance algorithm based on the target matching strategy and adjust the matching results to obtain a candidate recommendation set including user-familiar types and new types of activities; Perform cross-scenario combination analysis on the candidate recommendation set by identifying the preference associations of the user in different scenarios to obtain a personalized activity recommendation list including activity ID, matching score, matching reason, and expected participation probability.

[0031] Specifically, a user-activity bipartite graph model is created based on dynamic user portrait data and structured activity label data, with user feature vectors and activity feature vectors as nodes and the initial matching degree as the edge weight. The user feature vectors are derived from the user's dynamic portrait data, which includes information such as the user's interest preferences, behavior characteristics, and historical participation. The activity feature vectors are generated from the activity's label data, including features such as the activity type, content, audience group, time, and location. By converting these feature data into vector form, the relationship between users and activities is clearly expressed in the graph model, and the edge weights are initialized based on the similarity between the user feature vectors and the activity feature vectors. Multichannel matching calculations of content similarity calculation, collaborative filtering analysis, and context factor evaluation are sequentially performed on the user-activity bipartite graph model to gradually optimize the matching degree between users and activities. The content similarity calculation measures the similarity between user features and activity labels. The cosine similarity is used for the content similarity calculation to calculate the similarity between the user feature vector and the activity label vector. Collaborative filtering analysis is carried out, considering the activity participation of the user's neighbors or other users with similar behaviors to the user, so as to infer the activities that the user is interested in. The context factor evaluation dynamically adjusts the activity matching degree according to information such as the user's current behavior, environment, and time. By integrating these context factors, the matching degree score between the user and the activity is adjusted to better meet the user's current needs. After the above three matching calculations, the content matching degree score result, collaborative matching degree score result, and context matching degree score result are obtained. Adaptive weight fusion calculation is performed on the content matching degree score result, collaborative matching degree score result, and context matching degree score result according to the richness of user data and activity characteristics. By weighted average or other weighting strategies, the matching results of different dimensions are integrated to obtain the comprehensive matching degree score. For example, for a user with rich historical behavior data, the influence of collaborative filtering and context factors is greater, so higher weights are assigned. For a new user, the content similarity calculation will occupy a higher weight because the user has less behavior data and more relies on content features to infer their interests. The formula for the comprehensive matching degree is expressed in the following form:

[0032] Among them, represents the final comprehensive matching degree, , , are the content matching degree, collaborative matching degree, and context matching degree respectively, , , are the corresponding weight coefficients. The weight coefficients are dynamically adjusted according to the characteristics of different users and activities. For example, is automatically adjusted according to the richness of the user's behavior data, and It is adjusted according to the social interactivity and timeliness of the activities. The comprehensive matching degree score is input into the asynchronous Q-learning reinforcement learning model for dynamic optimization. The goal of the reinforcement learning model is to continuously adjust the matching strategy according to the user's feedback (such as the number of times of participating in activities, activity evaluations, etc.) to improve the accuracy of the recommendation results. The Q-learning algorithm continuously optimizes the matching strategy through the balance between exploration and exploitation in order to obtain the optimal activity recommendation strategy. Through the balance between exploration and exploitation, the system moderately explores activities that the user has not explicitly shown interest in and makes more frequent recommendations for activities with known user interests, thereby continuously improving the user's participation and satisfaction. After executing the exploration-exploitation balance algorithm based on the target matching strategy, a candidate recommendation set including activities of familiar types and new types for the user is generated. The candidate recommendation set includes the user's known interested activities and also includes some speculative activities based on the user preference model to enhance the diversity and freshness of the activities. In order to improve the accuracy of the recommendation, cross-scenario combination analysis is performed on the candidate recommendation set. The user's preferences vary in different scenarios. By identifying preference associations and dynamically adjusting the recommended content according to the user's scenario status, the recommended activities not only conform to the user's interests but also meet their actual needs at that time. After multi-dimensional matching and optimization, a personalized activity recommendation list including activity ID, matching score, matching reason, and expected participation probability is generated. In this list, the activity ID identifies the specific activity recommended, the matching score reflects the matching degree between the user and the activity, the matching reason explains why this activity is recommended (such as "similar to the user's historical interests" or "based on the participation behavior of similar users"), and the expected participation probability predicts the possibility of the user participating in this activity based on the comprehensive matching degree and historical data.

[0033] In a specific embodiment, the process of executing step 105 may specifically include the following steps: Apply activity capacity constraints, time conflict constraints, and geographical distance constraints to globally screen the personalized activity recommendation list to obtain a preliminary matching plan that meets the hard limit conditions; Perform brand network analysis on the preliminary matching plan based on brand relationship data to identify competitive brand, complementary brand, and cooperative brand relationships therein, and obtain the brand association marking result; Perform brand collaborative optimization processing on the preliminary matching plan according to the brand association marking result to obtain a target combination that avoids competition conflicts, strengthens complementary experiences, and prioritizes joint activities; Combine the user's historical activity participation frequency and feedback situation to evaluate and adjust the fatigue degree of the target combination to obtain an activity combination with a reasonable recommendation rhythm; Input the activity combination with a reasonable recommendation rhythm into an integer programming model for overall value maximization operation to obtain the optimal activity plan; Perform personalized processing on the display strategy of the optimal activity plan according to the user information reception preference to obtain a target matching plan that meets the cross-scenario requirements. The target matching plan that meets the cross-scenario requirements includes an activity combination list, a recommended order, a display strategy, and an expected effect indicator.

[0034] Specifically, consider the impacts of activity capacity constraints, time conflict constraints, and geographical distance constraints. Activity capacity constraints mean that the number of participants in some activities is limited, and more users cannot be accommodated beyond this number. Time conflict constraints involve the time arrangements of users. The time of some activities conflicts with the existing arrangements of users, resulting in their inability to participate. Geographical distance constraints are based on the locations of users. If the distance to an activity is too far, it will affect the willingness of users to participate. When making personalized activity recommendations, globally screen the activities in the recommendation list to obtain a preliminary matching plan that meets the hard limit conditions. Conduct brand network analysis on the preliminary matching plan based on brand relationship data, and identify the brand relationships between activities based on the brand network. Brand relationship analysis not only considers the types of activities but also analyzes based on the competitive, complementary, or cooperative relationships between brands to obtain the brand association marking results. The brand association marking results help to optimize the quality of subsequent activity recommendations, avoid recommending users to participate in two competing activities, and at the same time emphasize complementary activity combinations that can provide more value to users. Based on the brand association marking results, perform brand collaborative optimization processing to optimize the activity combinations in the preliminary matching plan, avoid competition conflicts, strengthen complementary experiences, and give priority to recommending joint activities. Through this optimization, effectively improve the user's activity participation experience and at the same time reduce the user's boredom with repeated or conflicting activities. Combine the user's historical activity participation frequency and feedback to conduct fatigue assessment and adjustment of the target combination. The purpose of fatigue assessment is to analyze the types and frequencies of activities the user has participated in and judge whether the user has developed interest fatigue in a certain type of activity. The user's feedback (such as ratings and evaluations after activities) is used to adjust the recommendation strategy. If an activity receives a high evaluation from the user, similar activities will be recommended first. If the user shows fatigue with a certain type of activity, the system will appropriately reduce its appearance frequency in the recommendation list to ensure the diversity and freshness of activity recommendations. Through fatigue assessment, obtain an activity combination with a reasonable recommendation rhythm. Input the activity combination with a reasonable recommendation rhythm into an integer programming model for overall value maximization operation. The integer programming model is optimized according to the characteristics of activities, the needs of users, and the constraint conditions between activities to obtain an optimal activity plan. The goal of the model is to maximize the overall value of the recommended activity combination, which is a comprehensive evaluation based on various factors such as user participation, satisfaction, and brand influence. The optimal activity plan is processed for display strategy personalization based on the user's information reception preferences. The purpose of display strategy personalization is to determine how to display the recommended activities in the most appropriate way according to the user's behavior habits and preferences. For example, some users prefer to receive recommended activities via email, while other users are more inclined to receive information through the notification function of the mobile application. According to these preferences, select the appropriate display channels to ensure that the recommended information can attract users to participate to the greatest extent. At the same time, the display strategy includes the optimization of the recommendation order.For example, based on the matching degree of the user's past participation history and activities, it is decided which activities to be preferentially presented to the user, thereby improving the user's willingness to participate. The recommended activity combinations include an activity list, accompanied by a recommended order, a display strategy, and expected effect indicators. The expected effect indicators are used to measure the effect of the recommendation strategy, such as the user click-through rate, participation rate, etc. These indicators are used to evaluate and improve future recommendation strategies.

[0035] In a specific embodiment, the method for performing activity matching based on user behavior further includes the following steps: Multi-dimensionally collect the user response data of the target matching solution that meets the cross-scenario requirements through explicit feedback channels and implicit feedback channels to obtain a user feedback data set containing direct behavior data and indirect reaction indicators; Apply a differential weight strategy to calculate the scores of different types of feedback in the user feedback data set to obtain weighted user feedback scores; Input the weighted user feedback scores into a multi-level learning engine for short-term immediate update, medium-term batch update, and long-term deep reconstruction for gradient processing to obtain multi-dimensional optimization parameters for temporary interest adjustment, stable feature update, and core algorithm optimization; Based on the multi-dimensional optimization parameters, conduct A / B test comparison and analysis on different matching strategies to obtain the evaluation result of the strategy effect difference; Use an abnormal pattern recognition mechanism to identify and extract the abnormal phenomena in the evaluation result of the strategy effect difference to obtain user interest change points and reaction abnormal points; Through a full-link effect attribution model, conduct a three-library integrated collaborative analysis on the user interest change points and the obtained reaction abnormal points to obtain an intelligent recommendation result that supports full-link marketing. The intelligent recommendation result that supports full-link marketing includes preset performance indicators, problem diagnosis, and improvement suggestions.

[0036] Specifically, the feedback information of users is collected through the explicit feedback channel and the implicit feedback channel. The explicit feedback channel includes the direct behavioral feedback of users, such as clicks, comments, ratings, etc., while the implicit feedback channel includes the interests and behaviors indirectly shown by users, such as indicators like browsing duration, stay time, bounce rate, etc. By combining these two types of data, the response and preference of users to the recommendation activity are reflected. The explicit feedback data includes the ratings or likes of users for the activity, while the implicit feedback is obtained by analyzing indicators such as the page browsing path and activity participation frequency of users. These data are combined to form a user feedback dataset containing direct behavioral data and indirect response indicators. A differential weight strategy is applied to calculate the scores of different types of feedback in the user feedback dataset. For example, the direct ratings and comments of users are more important than the browsing duration or the number of clicks, because these feedbacks can more directly reflect the real interests and activity preferences of users. For the explicit feedback data, a higher weight is assigned; while for the implicit feedback data, the weight is dynamically adjusted according to its relevance to the actual interests of users to obtain the weighted user feedback score. Assume that the weight of the explicit feedback is defined as and the weight of the implicit feedback is , then the weighted user feedback score is expressed as:

[0037] where represents the explicit feedback data, represents the implicit feedback data, and and These are the weights of explicit and implicit feedback. Through the weighting method, the impacts of different feedback sources on user interests are comprehensively considered, so as to more accurately reflect user preferences. The weighted user feedback scores are input into a multi-level learning engine for processing. The multi-level learning engine includes three levels: short-term immediate update, medium-term batch update, and long-term deep reconstruction. The short-term immediate update is used to process the changes in user interests that require quick responses, and perform immediate optimization based on the recent changes in user activities; the medium-term batch update focuses on analyzing a large amount of data within a certain time window to update the user profile and interest model; while the long-term deep reconstruction comprehensively analyzes and optimizes the core interests of users through deep learning and long-term data accumulation. The goal of this process is to obtain multi-dimensional optimization parameters for temporary interest adjustment, stable feature update, and core algorithm optimization through continuous gradient processing. The short-term immediate update can quickly respond to users' strong interests in a certain activity, while the long-term deep reconstruction can identify the long-term change trends of user preferences. Based on these optimization parameters, A / B test comparison analysis is carried out on different matching strategies to evaluate the effect differences of different strategies. The A / B test is to divide users into two or more groups, apply different matching strategies respectively, and then compare their effects, such as indicators like user engagement, click-through rate, conversion rate, etc. Through comparative analysis, it is identified which matching strategy is more suitable for a specific user group, and then the recommendation strategy is optimized. For example, Group A applies a content-based recommendation strategy, while Group B applies a collaborative filtering-based strategy. If the participation rate of Group A is significantly higher than that of Group B, the system further adjusts the matching strategy, takes the strategy of Group A as the mainstream method, and optimizes the strategy of Group B. The anomaly pattern recognition mechanism is used to identify and extract the abnormal phenomena in the evaluation results of the strategy effect differences. Through pattern recognition, it is analyzed which factors lead to the abnormal fluctuations in the strategy effects, such as sudden changes in user interests or over-recommendation of certain activities. By extracting these abnormal points, the user interest change points and response abnormal points are analyzed to understand which factors are interfering with the recommendation effect. Through the full-link effect attribution model, three-library integrated collaborative analysis is carried out on the user interest change points and response abnormal points to improve the accuracy of recommendations. The three-library integration refers to the collaborative analysis of the user behavior library, activity library, and feedback library. These three libraries work together to help the system more comprehensively understand the correlation between user behavior, activity characteristics, and feedback information. Through multi-dimensional collaborative analysis, intelligent recommendation results supporting full-link marketing are generated. These recommendation results include preset performance indicators such as click-through rate, conversion rate, etc., as well as problem diagnosis and improvement suggestions. The problem diagnosis part can analyze the deficiencies of the current recommendation strategy, such as the recommended activity type not meeting user interests, or the user fatigue caused by too high a recommendation frequency; while the improvement suggestions will provide specific optimization plans according to the analysis results, such as adjusting the activity type, optimizing the recommendation frequency, or improving the display strategy, etc.

[0038] The above describes the activity matching method based on user behavior in the embodiments of the present invention. Next, the activity matching system based on user behavior in the embodiments of the present invention will be described. Please refer to Figure 2 , an embodiment of the activity matching system based on user behavior in the embodiments of the present invention includes: A collection module 201, configured to perform all-dimensional collection on the behavior data of users in a multi-channel marketing environment to obtain a structured user behavior data set; A construction module 202, configured to construct dynamic user portrait data based on the structured user behavior data set; A processing module 203, configured to perform three-layer tagging processing on the original activity data for basic attributes, content features, and experience value to obtain structured activity tag data; A calculation module 204, configured to perform asynchronous iterative calculation and intelligent weighted matching according to the dynamic user portrait data and the structured activity tag data to obtain a personalized activity recommendation list; A collaboration module 205, configured to execute a multi-brand collaboration strategy and global constraint processing on the personalized activity recommendation list to obtain a target matching solution that meets cross-scenario requirements.

[0039] Through the collaborative cooperation of the above-mentioned various components, through the full-channel data collection and five-dimensional data structure organization of web pages, mobile applications, offline interactions, and social media, the data island problem of traditional methods is solved. At the same time, by separating the explicit behavior and implicit preference processing and the time series analysis model, a dynamic user portrait containing eight-dimensional feature vectors is constructed to accurately perceive the change of user preferences; the activity is subjected to three-layer tagging processing for basic attributes, content features, and experience value, combined with activity association analysis, to realize the multi-dimensional three-dimensional description of the activity content and significantly improve the matching accuracy; based on the asynchronous iterative calculation of the user-activity bipartite graph model, combined with three-channel matching and Q-learning reinforcement learning, the performance bottleneck of traditional synchronous calculation is broken through, effectively avoiding the "information cocoon" problem; through brand network analysis, collaborative optimization processing, and integer programming model, coordinated activity recommendations in a multi-brand environment are realized, solving the technical problem of mutual conflict in activity recommendations of different brands; by using explicit and implicit feedback channels, multi-level learning engines, and abnormal pattern recognition mechanisms, a complete feedback-driven closed-loop system is constructed, enabling the matching algorithm to continuously self-optimize and realizing the long-term stable improvement of system performance.

[0040] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the above-described system, system, and unit can refer to the corresponding processes in the foregoing method embodiments and will not be repeated here.

[0041] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0042] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An activity matching method based on user behavior, characterized in that: The method comprises: Collect all dimensions of user behavior data in a multi-channel marketing environment to obtain a structured user behavior data set; Building dynamic user portrait data based on the structured user behavior data set; The original activity data is processed into three-layer labels: basic attributes, content features, and experience value, to obtain structured activity label data; Performing asynchronous iterative calculation and intelligent weighted matching according to the dynamic user portrait data and the structured activity tag data to obtain a personalized activity recommendation list; A multi-brand collaboration strategy and global constraint processing are executed on the personalized activity recommendation list to obtain a target matching solution that meets cross-scenario requirements.

2. The activity matching method based on user behavior according to claim 1, characterized in that: The user behavior data in the multi-channel marketing environment is collected in all dimensions to obtain a structured user behavior data set, including: Collect user omni-channel behavior data through web behavior tracking module, mobile application monitoring module, offline interaction capture module and social media analysis module to obtain original behavior data; Performing timestamp unification, user ID anonymization and behavior type standardization processing on the original behavior data to obtain preliminarily processed behavior data; Desensitizing and compressing the preliminarily processed behavior data at the data source by using a distributed architecture to obtain a secure transmission data packet; Transmitting the secure transmission data packet to a central data processing center through an encrypted channel for parsing to obtain centralized behavior data, and performing data quality monitoring and detection on the centralized behavior data to screen out abnormal data and obtain valid behavior data; The effective behavior data is organized and integrated according to the five-dimensional data structure of user, time, behavior, scene and object to obtain a structured user behavior data set.

3. The activity matching method based on user behavior according to claim 1, characterized in that: The constructing dynamic user portrait data based on the structured user behavior data set includes: Separating the structured user behavior data set into explicit behavior data and implicit preference data; Applying a direct assignment method to the explicit behavior data to calculate weights, and applying a time decay weighting method to the implicit preference data to calculate weights, to obtain weighted behavior feature data; Constructing a preliminary user feature vector including demographics, interest preferences, consumption capacity, brand interaction, time pattern, social influence, scenario preference and activity response based on the weighted behavioral feature data; Performing similar user group feature migration on the preliminary user feature vector to obtain a complete feature vector, and performing historical change trend analysis on the complete feature vector to obtain a time series feature reflecting the time change characteristics of user preferences; The complete feature vector and the time series feature are integrated into a standardized data structure including a user ID, a feature vector, a confidence score and a timestamp to obtain dynamic user portrait data.

4. The activity matching method based on user behavior according to claim 1, characterized in that: The three-layer labeling of basic attributes, content features and experience value is performed on the original activity data to obtain structured activity label data, including: Receive basic information about the activity including the name, time, location, content description, target audience and participation method, and obtain the original data of the activity; The rule engine extracts the basic attributes of activity type, time attribute, location attribute, participation threshold and scale from the original activity data to obtain the first-level basic attribute label; Using natural language processing technology to perform semantic analysis on the description text in the original data of the activity to extract keywords, and obtain second-level content feature labels mapped to a preset content classification system; Combine the activity goals, forms and historical participant feedback in the original data of the activity to perform experience value analysis, and obtain a third-level experience value label representing the characteristics of knowledge acquisition, emotional satisfaction and social expansion; Performing activity correlation analysis on the first-level basic attribute tags, the second-level content feature tags, and the third-level experience value tags to obtain association tags marking relationships among series activities, complementary activities, and competitive activities; The first-level basic attribute tags, the second-level content feature tags, the third-level experience value tags and the associated tags are integrated and assigned weight values ​​to obtain structured activity tag data including activity ID, tag set, weight value, timeliness and related activity references.

5. The activity matching method based on user behavior according to claim 1, characterized in that: The step of performing asynchronous iterative calculation and intelligent weighted matching according to the dynamic user portrait data and the structured activity tag data to obtain a personalized activity recommendation list includes: Based on the dynamic user portrait data and the structured activity label data, a user-activity bipartite graph model is created, in which nodes are user feature vectors and activity feature vectors, and edge weights are initial matching degrees; Performing content similarity calculation, collaborative filtering analysis, and multi-channel matching calculation of context factor evaluation on the user-activity bipartite graph model in sequence to obtain content matching score results, collaborative matching score results, and context matching score results; According to the richness of user data and activity characteristics, adaptive weight fusion calculation is performed on the content matching score result, the collaborative matching score result and the context matching score result to obtain a comprehensive matching score; Inputting the comprehensive matching score into an asynchronous Q-learning reinforcement learning model for dynamic optimization to obtain a target matching strategy that takes user feedback into account; Execute an exploration-exploitation balance algorithm based on the target matching strategy, and adjust the matching results to obtain a candidate recommendation set including activities of types familiar to the user and new types; By identifying the preference associations of users in different scenarios, a cross-scenario combination analysis is performed on the candidate recommendation set to obtain a personalized activity recommendation list including activity IDs, matching scores, matching reasons, and expected participation probabilities.

6. The activity matching method based on user behavior according to claim 1, characterized in that: The performing of multi-brand collaboration strategy and global constraint processing on the personalized activity recommendation list to obtain a target matching solution that meets cross-scenario requirements includes: Applying activity capacity constraints, time conflict constraints, and geographic distance constraints, the personalized activity recommendation list is globally screened to obtain a preliminary matching solution that meets the hard constraint conditions; Performing brand network analysis on the preliminary matching scheme based on brand relationship data, identifying the competitive brands, complementary brands and cooperative brand relationships therein, and obtaining brand association marking results; Performing brand collaborative optimization processing on the preliminary matching scheme according to the brand association tag result to obtain a target combination of avoiding competitive conflicts, strengthening complementary experiences and prioritizing joint activities; The fatigue of the target combination is evaluated and adjusted based on the user's historical activity participation frequency and feedback, so as to obtain an activity combination with a reasonable recommendation rhythm; Input the activity combination with a reasonable recommended rhythm into the integer programming model to perform overall value maximization calculation to obtain the optimal activity plan; The display strategy of the optimal activity plan is personalized according to the user information receiving preference to obtain a target matching plan that meets cross-scenario needs. The target matching plan that meets cross-scenario needs includes an activity combination list, a recommended order, a display strategy, and expected effect indicators.

7. The activity matching method based on user behavior according to claim 1, characterized in that: The activity matching method based on user behavior also includes: The user response data of the target matching solution that meets the cross-scenario requirements is collected in multiple dimensions through an explicit feedback channel and an implicit feedback channel to obtain a user feedback data set including direct behavior data and indirect response indicators; Applying a differentiated weight strategy to score different types of feedback in the user feedback data set to obtain a weighted user feedback score; Input the weighted user feedback scores into a multi-level learning engine for short-term instant update, mid-term batch update, and long-term deep reconstruction for gradient processing to obtain multi-dimensional optimization parameters for temporary interest adjustment, stable feature update, and core algorithm optimization; Based on the multi-dimensional optimization parameters, A / B test comparison analysis is performed on different matching strategies to obtain evaluation results of strategy effect differences; Using the abnormal pattern recognition mechanism to identify and extract features from the abnormal phenomena in the strategy effect difference evaluation results, and obtain the user interest change points and reaction abnormal points; Through the full-link effect attribution model, the three-database integrated collaborative analysis is performed on the user interest change points and the response abnormal points to obtain intelligent recommendation results that support full-link marketing. The intelligent recommendation results that support full-link marketing include preset performance indicators, problem diagnosis and improvement suggestions.

8. An activity matching system based on user behavior, characterized in that: Used to execute the activity matching method based on user behavior according to any one of claims 1 to 7, the activity matching system based on user behavior comprises: The collection module is used to collect all-dimensional user behavior data in a multi-channel marketing environment to obtain a structured user behavior data set; A construction module, used to construct dynamic user portrait data based on the structured user behavior data set; The processing module is used to perform three-layer labeling processing on the original activity data, including basic attributes, content characteristics and experience value, to obtain structured activity label data; A calculation module, used for performing asynchronous iterative calculation and intelligent weighted matching according to the dynamic user portrait data and the structured activity tag data to obtain a personalized activity recommendation list; The collaboration module is used to execute multi-brand collaboration strategy and global constraint processing on the personalized activity recommendation list to obtain a target matching solution that meets cross-scenario requirements.

Images