Intelligent workbench search optimization method and system based on multi-scenario business

Abstract

The present invention relates to the field of intelligent workbench search optimization technology, and in particular to an intelligent workbench search optimization method and system based on multi-scenario business. The method comprises the following steps: collecting user behavior data of platform users to obtain user behavior data; performing behavior time series analysis on user behavior data to obtain user behavior sequence; constructing a user tag library according to the user behavior sequence to obtain a user tag library; extracting content information from the workbench and constructing a content feature vector library to obtain a content feature vector library; constructing a scene context library according to the user tag library and the content feature vector library to obtain a scene context library. The present invention solves the problem of lack of intelligent recommendation and prediction in traditional methods by taking users as the center, actively perceiving user needs, and providing accurate and personalized information push and search assistance, thereby improving the intelligence level, user experience and work efficiency of the intelligent workbench.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent workbench search optimization, and in particular to an intelligent workbench search optimization method and system based on multi-scenario business. Background Art

[0002] With the advancement of enterprise digital transformation, intelligent workbenches based on multi-scenario business have emerged, aiming to integrate different systems and applications to provide users with one-stop services. As one of the core functions of the intelligent workbench, search optimization has become particularly important. However, traditional intelligent workbench search optimization methods based on multi-scenario business lack intelligent recommendations and predictions, and are mostly passive. Users need to actively enter keywords to obtain information, and need to constantly try different keywords to find the required information. The search process is cumbersome, affecting the user experience, missing a lot of valuable but unsearched information, resulting in reduced information utilization. It takes a lot of time to screen and judge the relevance of search results. The search efficiency is low and is often limited to known keywords. It is difficult to actively discover new and potentially valuable information. There is a lack of the ability to actively recommend information based on user characteristics and needs, and it is impossible to provide targeted services.

[0003] The system lacks the ability to proactively perceive user needs and make recommendations and predictions. For example, when a user is preparing for a meeting, the system cannot proactively recommend relevant materials. Personalized recommendation algorithms are mostly based on historical user behavior data and lack an understanding of the user's current scenario and intentions, resulting in inaccurate recommendation results. Summary of the invention

[0004] Based on this, it is necessary to provide an intelligent workbench search optimization method and system based on multi-scenario business to solve at least one of the above technical problems.

[0005] To achieve the above purpose, a smart workbench search optimization method based on multi-scenario business includes the following steps:

[0006] Step S1: Under the authorization of the platform users, user behavior data of the platform users are collected to obtain user behavior data; behavior time series analysis is performed on the user behavior data to obtain a user behavior sequence; a user tag library is constructed according to the user behavior sequence to obtain a user tag library; content information is extracted from the workbench and a content feature vector library is constructed to obtain a content feature vector library; a scene context library is constructed according to the user tag library and the content feature vector library to obtain a scene context library;

[0007] Step S2: Optimize similarity propagation based on the user tag library to obtain a user similarity matrix; extract user-content interaction based on the content feature matrix to obtain a user-content interaction matrix; assign interaction matrix weights to the user-content interaction matrix based on the scene context library and the content feature vector library to obtain a user-content interaction weighted matrix; construct an association rule library based on the user-content interaction weighted matrix to obtain an association rule library;

[0008] Step S3: Real-time scene capture of platform users to obtain current scene tags; rule priority sorting according to current scene tags and the associated rule library to obtain a recommendation rule sorting set; user interest vector calculation according to the user tag library to obtain the user interest vector; personalized recommendation list generation according to the recommendation rule sorting set and the user interest vector to obtain a personalized recommendation list;

[0009] Step S4: extracting user behavior trajectories according to user behavior sequences to obtain user behavior trajectory data; dividing the user behavior trajectory data into time windows to obtain a user behavior time window data set; constructing a frequent behavior pattern library for the user behavior time window data set to obtain a frequent behavior pattern library; receiving search terms from platform users in real time to obtain current search term data; predicting search intent based on the current search term data and the frequent behavior pattern library to obtain search intent prediction data;

[0010] Step S5: Collect multi-channel feedback through the workbench to obtain user feedback data; iteratively optimize the search data for the personalized recommendation list and search intent prediction data based on the user feedback data to realize the intelligent workbench search work.

[0011] The present invention builds a multi-level and multi-dimensional information foundation for the entire system by collecting user behavior data, building a user tag library, a content feature vector library and a scene context library, realizing the digitization of user portraits, content portraits and scene information, and laying a solid data foundation for subsequent personalized recommendations, search intent prediction and iterative optimization. Through user similarity calculation, user-content interaction extraction, scene weight fusion and association rule mining, a network of association relationships between users, content and scenes is constructed, which can more accurately portray the user's interest preferences, the semantic information of the content and the influencing factors of the scene, and provide a more accurate and richer basis for personalized recommendations and search intent prediction. Through real-time scene capture, rule matching, personalized sorting, diversity optimization and visual display, scene-aware personalized recommendations are realized, which can provide the freshest and highest-quality content based on the user's real-time scene, interest preferences and information needs, effectively improving user experience and platform value. Through user behavior trajectory analysis, frequent behavior pattern mining, semantic matching and target behavior prediction, search intention prediction based on historical behavior is realized. It can predict the user's search target based on the user's historical search behavior and current search terms, and recommend related search terms and content to help users find the required information more quickly and accurately, and improve the user's search experience. Through multi-channel collection of user feedback, semantic processing, data annotation, feature extraction and model optimization, a data-driven iterative optimization closed loop is constructed, which can continuously improve the recommendation effect and search quality of the intelligent workbench search, make the system more intelligent and personalized, and provide users with a more convenient, efficient and satisfactory search experience. Therefore, the present invention provides a smart workbench search optimization method based on multi-scenario business. By focusing on the user, integrating multi-scenario perception, artificial intelligence and data mining calculation, actively perceiving user needs, providing accurate and personalized information push and search assistance, and establishing a data-driven iterative optimization closed loop, the problem of lack of intelligent recommendation and prediction in the traditional smart workbench search optimization method based on multi-scenario business is solved, which greatly improves the intelligence level, user experience and work efficiency of the smart workbench.

[0012] Preferably, step S1 comprises the following steps:

[0013] Step S11: Collecting user behavior data of platform users with authorization from platform users to obtain user behavior data; performing behavior time series analysis on the user behavior data to obtain a user behavior sequence; performing scene tagging on the user behavior sequence to obtain a scene tag set; constructing a user tag library based on the user behavior sequence to obtain a user tag library;

[0014] Step S12: extracting content information from the workbench to obtain content metadata; performing content semantic analysis on the content metadata to obtain a content semantic vector;

[0015] Step S13: performing content feature fusion on the content semantic vector and constructing a content feature vector library to obtain a content feature vector library;

[0016] Step S14: Associating scene information according to the scene tag set and performing vector encoding to obtain a scene feature vector;

[0017] Step S15: Perform scene context association on the user tag library and the content feature vector library according to the scene feature vector and construct a scene context library to obtain a scene context library.

[0018] The present invention can restore the user's behavior on the platform and understand the user's behavior habits and preferences by collecting user behavior data and performing time series analysis. At the same time, by performing scene tagging on the user behavior sequence, the user's behavior can be linked to the specific scene, providing a basis for subsequent scene-based recommendation and search. By constructing a user tag library based on the user behavior sequence, each user can be labeled with a personalized tag, which is convenient for the system to better understand the user and provide more accurate services. By extracting and semantically analyzing the workbench content information, the unstructured content data can be converted into a structured data representation, which is convenient for the system to store, retrieve and analyze. At the same time, by converting the content into a semantic vector, the semantic information of the content can be better captured, and the accuracy of search and recommendation can be improved. By fusing the content semantic vector with other content features, a more comprehensive and richer content feature vector library can be constructed, providing a more accurate basis for subsequent personalized recommendation and search. By associating scene information and performing vector encoding according to the scene tag set, the abstract scene information can be converted into a specific vector representation, which is convenient for the system to calculate and analyze, and provides support for subsequent scene perception and personalized recommendation. By performing scene context association on the user tag library and the content feature vector library based on the scene feature vector and constructing a scene context library, the association relationship between users, content and scenes can be established, so that the system can recommend more relevant content according to the user's interests and needs in different scenarios.

[0019] Preferably, step S11 includes the following steps:

[0020] Step S111: registering basic information of platform users and collecting user behavior data with the authorization of the platform users, and obtaining basic information and user behavior data;

[0021] Step S112: performing user behavior analysis on the user behavior data and performing behavior time series analysis to obtain a user behavior sequence;

[0022] Step S113: identifying scene elements of the user behavior sequence and marking the scene to obtain a scene label set;

[0023] Step S114: performing user behavior cluster analysis according to the user behavior sequence to obtain user behavior clusters;

[0024] Step S115: Generate user tags and construct a user tag library based on user basic information, user behavior clusters, and scene tag sets to obtain a user tag library.

[0025] The present invention can not only ensure user privacy and security by collecting user basic information and behavior data with user authorization, but also provide a data basis for building a more comprehensive user portrait. By integrating user basic information and behavior data, the user's interests and needs can be identified more accurately, thereby providing more accurate personalized services. Parsing and time series analysis of user behavior data can convert the original, messy user behavior data into structured, orderly information, which is convenient for the system to store, retrieve and analyze, and provides a basis for subsequent user behavior pattern mining and scene recognition. Through scene element recognition and scene labeling, the user's behavior can be linked to specific scenes, such as work, study, entertainment, etc., so as to better understand the user's behavioral motivations and information needs in different scenes, and provide support for subsequent scene-based recommendations and searches. Cluster analysis based on user behavior sequences can be performed to divide users with similar behavior patterns into the same user behavior cluster, so as to discover the behavioral differences between different user groups, and provide a basis for more accurate user portrait characterization and personalized services. By combining basic user information, user behavior clusters and scenario labels, more comprehensive and accurate labels can be generated for each user, such as the user's occupation, interests and hobbies, current scenario, etc., thereby building a more refined user label library and providing a more accurate user portrait for subsequent personalized recommendations and searches.

[0026] Preferably, step S2 comprises the following steps:

[0027] Step S21: construct a user tag matrix according to the user tag library to obtain a user tag matrix; perform tag co-occurrence statistics according to the user tag matrix to obtain a tag co-occurrence matrix;

[0028] Step S22: Calculate user similarity based on the tag co-occurrence matrix to obtain a user initial similarity matrix; perform similarity propagation optimization on the user initial similarity matrix to obtain a user similarity matrix;

[0029] Step S23: performing user interaction behavior analysis on the user behavior sequence to obtain user interaction behavior data; constructing a content feature matrix according to the content feature vector library to obtain a content feature matrix;

[0030] Step S24: extracting user-content interaction from the user interaction behavior data according to the content feature matrix to obtain a user-content interaction matrix;

[0031] Step S25: Calculate scene attention weights according to the scene context library, the content feature vector library, and the user interaction behavior data to obtain scene attention weight data; assign interaction matrix weights to the user-content interaction matrix according to the scene attention weight data to obtain a user-content interaction weighted matrix;

[0032] Step S26: performing association rule mining according to the user-content interaction weighted matrix to obtain an initial association rule set;

[0033] Step S27: Filter the initial association rule set by rule confidence using a preset rule confidence threshold to obtain a high-confidence association rule set; merge and simplify the high-confidence association rule set to obtain an association rule base.

[0034] The present invention can quantify the association between user tags by constructing a user tag matrix and a tag co-occurrence matrix. For example, users with similar tags are more likely to have potential associations. This provides a data basis for subsequent calculation of user similarity and mining association rules. First, the initial similarity of users is calculated according to the tag co-occurrence matrix, and then optimized using the similarity propagation algorithm. The association between users can be more accurately portrayed. Even if two users do not have direct common tags, as long as they have indirect tag associations with other users, they can be captured by the algorithm, thereby improving the diversity and coverage of recommendation results. By analyzing the user behavior sequence, extracting the interaction between users and content, and constructing a content feature matrix according to the content feature vector library, user behavior and content information can be associated, providing data support for subsequent mining of user-content interaction patterns and construction of association rules. According to the content feature matrix, user-content interaction extraction is performed on user interaction behavior data, and a user-content interaction matrix is ​​constructed. The interaction intensity between users and different contents can be quantified. For example, the number of times and duration of a user browsing a certain content, etc., provide more accurate user preference information for personalized recommendation. By calculating the scene attention weight, the user's attention to different content can be dynamically adjusted according to the scene the user is currently in. For example, users pay more attention to work-related documents and tools in work scenes, and pay more attention to entertainment content in leisure scenes, thereby improving the scene relevance and user satisfaction of recommendation results. By using the association rule mining algorithm, the association relationship between content features that users frequently interact with in different scenes can be mined from the user-content interaction weighted matrix. For example, after searching for a certain keyword, what types of results will users often click on? These association rules can be used to predict the user's search intention and recommend related content. Through rule confidence filtering, rule merging and rule simplification, redundant and low-confidence association rules can be removed, and the most representative and valuable rules can be retained, thereby building a streamlined and efficient association rule library to improve the efficiency and accuracy of search and recommendation.

[0035] Preferably, step S25 comprises the following steps:

[0036] Step S251: Generate a scene vector according to the scene context library to obtain a scene vector set; contextualize the user interaction behavior data according to the scene vector set to obtain contextualized user behavior data;

[0037] Step S252: performing interactive behavior semantic vectorization on the user interactive behavior data to obtain an interactive behavior semantic vector; performing content semantic vector extraction based on a content feature vector library to obtain a content semantic vector;

[0038] Step S253: Calculate the scene attention weight according to the scene vector set, the interaction behavior semantic vector and the content semantic vector to obtain scene attention weight data;

[0039] Step S254: performing scene weight fusion according to the scene attention weight data and the scene-based user behavior data to obtain fused scene weight data;

[0040] Step S255: assigning interaction matrix weights to the user-content interaction matrix according to the fused scene weight data to obtain a user-content interaction weighted matrix.

[0041] The present invention can link user behavior with specific scenarios by converting scenario information into vector representation and associating it with user interaction behavior data. For example, it can distinguish the behavioral differences between users browsing technical documents in work scenarios and browsing news information in leisure scenarios, providing a more accurate basis for subsequent scenario-based recommendations. Converting user interaction behavior and content information into semantic vectors can better capture the semantic relevance between user behavior and content. For example, the behavior of users searching for "machine learning" and browsing "Python tutorials" can be calculated with a higher similarity through semantic vectors, thereby improving the accuracy and relevance of recommendation results. Using the scenario attention mechanism, the user's attention to different content can be dynamically adjusted according to the user's current scenario. For example, in the work scenario, the user pays more attention to work-related documents and tools, while in the leisure scenario, the user pays more attention to entertainment content, thereby improving the scenario adaptability and user satisfaction of the recommendation results. Fusion of scenario attention weights in different scenarios can more comprehensively consider the user's behavioral preferences in different scenarios. For example, users often browse technical documents in work scenarios and often watch movies in leisure scenarios. The fused scenario weights can reflect the user's comprehensive interest preferences in different scenarios, thereby improving the personalization of recommendation results. By weighting the user-content interaction matrix according to the fused scenario weights, the intensity of user-content interaction in different scenarios can be adjusted. For example, the interaction weight of users with work-related content in work scenarios can be increased, and the interaction weight of users with work-related content in leisure scenarios can be reduced, so that the recommendation results are more in line with the actual needs of users in different scenarios.

[0042] Preferably, step S3 comprises the following steps:

[0043] Step S31: Real-time scene capture of platform users to obtain current scene labels; scene rule matching of current scene labels and associated rule bases to obtain candidate recommendation rule sets;

[0044] Step S32: sorting the candidate recommendation rule sets by rule priority according to the association rule library to obtain a recommendation rule sorting set; calculating the user interest vector according to the user tag library and the user similarity matrix to obtain the user interest vector;

[0045] Step S33: performing content retrieval on the content feature vector library according to the recommendation rule sorting set to obtain a content candidate set; performing content interest score calculation according to the user interest vector and the content candidate set to obtain a content interest list;

[0046] Step S34: performing diversity sorting optimization on the content interest list to obtain a diversity content recommendation list; performing time decay adjustment on the diversity content recommendation list to obtain a content recommendation time decay list;

[0047] Step S35: Visualize the recommendation results according to the content recommendation time decay list to obtain a personalized recommendation list.

[0048] The present invention can quickly filter out recommendation rules related to the user's current scenario by capturing user scenarios in real time and matching scenario tags with an association rule library. For example, when the user is in a work scenario, the system will give priority to recommending documents and tools related to work, thereby improving the timeliness and pertinence of the recommendation results. Prioritizing candidate recommendation rules and calculating user interest vectors in combination with user tag libraries and user similarity matrices can ensure that the recommendation results meet both current scenario requirements and user personalized preferences, thereby avoiding recommendation results that are too single or inconsistent with user interests. According to recommendation rules and user interest vectors, content is retrieved from a content feature vector library, and content interest scores are calculated, thereby narrowing a huge content library to a candidate set that is highly relevant to user interests and the current scenario, thereby improving recommendation efficiency and accuracy. Diversity sorting and optimization of the content interest list can avoid recommendation results that are too concentrated on a certain type or theme. For example, while recommending technical documents, some related industry information or learning resources can also be recommended to improve user experience and broaden user horizons. Time decay adjustment can be performed to recommend the latest published or updated content to users, thereby ensuring the freshness of recommendation results. Visually displaying the recommendation results, such as in the form of cards or lists, and highlighting the key information of the content, such as title, summary, pictures, etc., can help users quickly browse and understand the recommended content and improve the user experience.

[0049] Preferably, step S4 comprises the following steps:

[0050] Step S41: extracting user behavior trajectories from the user tag library, user behavior sequence, and scene context library to obtain user behavior trajectory data;

[0051] Step S42: Divide the user behavior trajectory data into time windows to obtain a user behavior time window data set; construct a frequent behavior pattern library for the user behavior time window data set to obtain a frequent behavior pattern library;

[0052] Step S43: receiving search terms from platform users in real time to obtain current search term data; performing search term semantic vectorization on the current search term data to obtain a search term semantic vector;

[0053] Step S44: performing historical behavior semantic matching on the search word semantic vector and the frequent behavior pattern library to obtain a matching behavior pattern; performing a predicted target behavior inference based on the matching behavior pattern and the current search word data to obtain predicted target behavior data;

[0054] Step S45: Generate search word recommendations based on the predicted target behavior data to obtain a search word prediction list; calculate search content relevance based on the search word semantic vector and the content feature vector library to obtain content relevance score data;

[0055] Step S46: perform sorting and recommendation of related content based on the content relevance score data to obtain a related content prediction list; perform search intent prediction based on the search term prediction list and the related content prediction list to obtain search intent prediction data.

[0056] The present invention extracts user behavior trajectory data by integrating user tags, behavior sequences and scene context information, which can more comprehensively and three-dimensionally characterize the user's historical behavior characteristics, and provide a richer information basis for subsequent behavior pattern mining and intention prediction. Dividing the user behavior trajectory data into time windows and constructing a frequent behavior pattern library can identify the user's behavior patterns and preference changes in different time periods. For example, there are differences in the user's search behavior on weekdays and weekends, thereby providing a basis for more accurate search intention prediction. Converting the search terms entered by the user into semantic vectors can break through the limitations of traditional keyword matching and improve the coverage and accuracy of search results. Semantically matching the search term semantic vector with the frequent behavior pattern library and predicting the user's target behavior can predict the user's current search target based on the user's historical behavior, thereby providing a reference for search term recommendations and related content recommendations. Recommending search terms based on the predicted target behavior and calculating the relevance score between the search terms and the content can help users find the required information more quickly and accurately, and give priority to displaying learning resources that are highly relevant to these keywords. By combining the search term prediction list and the related content prediction list, we can more comprehensively judge the user's search intention. The user's search intention is to learn knowledge and skills related to machine learning, thereby providing a more accurate basis for search result optimization and personalized services.

[0057] Preferably, step S42 includes the following steps:

[0058] Step S421: Calculate the time interval of the user behavior trajectory data to obtain the user behavior timestamp data;

[0059] Step S422: dividing the user behavior trajectory data into time windows according to the user behavior timestamp data to obtain a user behavior time window data set;

[0060] Step S423: mining the behavior sequence pattern in the window on the user behavior time window data set to obtain an initial frequent behavior pattern set;

[0061] Step S424: performing time interval statistics on the initial frequent behavior pattern set to obtain time interval frequent behavior patterns; performing pattern support calculation on the time interval frequent behavior patterns to obtain support frequent behavior patterns;

[0062] Step S425: Filter the support frequent behavior patterns using a preset support threshold and construct a frequent behavior pattern library to obtain a frequent behavior pattern library.

[0063] The present invention can capture the temporal characteristics of user behavior by calculating the time intervals between user behaviors. For example, it can be used to understand how long it usually takes for a user to search or click on other links after browsing a certain page, thereby providing a more refined time scale for subsequent behavior pattern mining. By dividing the user behavior trajectory into multiple time windows according to the timestamp information, the user behavior can be segmented into smaller granularities, which facilitates the system to analyze the user's behavior patterns in different time periods. For example, there are differences in the user's search behavior in the morning and afternoon. By performing frequent pattern mining on the user behavior sequence in each time window, it is possible to find the behavior combinations that often occur in users in a specific time period. For example, users often search for technical keywords first and then browse related technical documents on weekday mornings. By performing time interval statistics and support calculations on the initial frequent behavior pattern, the temporal characteristics of the behavior pattern can be further refined. For example, it is possible to understand how long it usually takes for a user to click on the search result after searching for a certain keyword, and the prevalence of this behavior pattern among all users. By filtering frequent behavior patterns using a preset support threshold, we can remove patterns that appear occasionally or have low support, and retain those common and representative behavior patterns, thereby building a more streamlined and efficient frequent behavior pattern library, providing a more reliable basis for search intent prediction.

[0064] Preferably, step S5 comprises the following steps:

[0065] Step S51: Collect multi-channel feedback through the workbench to obtain user feedback data;

[0066] Step S52: semantically categorize the user feedback data into feedback behavior to obtain a semantically categorized feedback behavior label;

[0067] Step S53: annotating the user feedback data according to the semantic feedback behavior label to obtain a feedback annotation data set;

[0068] Step S54: extracting feedback features from the feedback annotation data set to obtain a feedback feature vector;

[0069] Step S55: Perform search data iterative optimization on the personalized recommendation list and search intent prediction data according to the feedback annotation data set and the feedback feature vector to implement the intelligent workbench search work.

[0070] The present invention collects user feedback data through multiple channels, which can more comprehensively understand the user's satisfaction with search results and recommendation results, and provide richer reference basis for subsequent algorithm optimization. By semantically processing user feedback data, the user's implicit feedback can be converted into more explicit, machine-understandable labels, which is convenient for subsequent algorithm analysis and model training. By labeling user feedback data according to semantic feedback labels, a high-quality training data set can be constructed for training and optimizing recommendation algorithms and search models. By extracting feedback feature vectors from feedback labeling data sets, the user's feedback behavior can be linked to specific features, helping the algorithm to better understand the reasons for user feedback. By iteratively optimizing recommendation algorithms and search models using feedback labeling data sets and feedback feature vectors, the accuracy, relevance and user satisfaction of recommendation results and search results can be continuously improved. For example, the parameters of the recommendation algorithm and the weight of the model can be adjusted according to user feedback, thereby making the system more intelligent and personalized.

[0071] Preferably, the present invention further provides an intelligent workbench search optimization system based on multi-scenario services, which is used to execute the intelligent workbench search optimization method based on multi-scenario services as described above, and the intelligent workbench search optimization system based on multi-scenario services comprises:

[0072] The multi-dimensional portrait construction module is used to collect user behavior data with the authorization of platform users to obtain user behavior data; perform behavior time series analysis on user behavior data to obtain user behavior sequences; construct a user tag library based on the user behavior sequences to obtain a user tag library; extract content information from the workbench and construct a content feature vector library to obtain a content feature vector library; construct a scene context library based on the user tag library and the content feature vector library to obtain a scene context library;

[0073] The association rule mining module is used to optimize similarity propagation based on the user tag library to obtain a user similarity matrix; extract user-content interaction based on the content feature matrix to obtain a user-content interaction matrix; assign interaction matrix weights to the user-content interaction matrix based on the scene context library and the content feature vector library to obtain a user-content interaction weighted matrix; construct an association rule library based on the user-content interaction weighted matrix to obtain an association rule library;

[0074] The scene intelligent recommendation module is used to capture the scene of the platform users in real time and obtain the current scene label; sort the rule priorities according to the current scene label and the associated rule library to obtain the recommended rule sorting set; calculate the user interest vector according to the user tag library to obtain the user interest vector; generate a personalized recommendation list according to the recommended rule sorting set and the user interest vector to obtain a personalized recommendation list;

[0075] The search dynamic prediction module is used to extract user behavior trajectories according to user behavior sequences to obtain user behavior trajectory data; divide the user behavior trajectory data into time windows to obtain a user behavior time window data set; construct a frequent behavior pattern library for the user behavior time window data set to obtain a frequent behavior pattern library; receive search terms from platform users in real time to obtain current search term data; predict search intent based on current search term data and the frequent behavior pattern library to obtain search intent prediction data;

[0076] The search feedback optimization module is used to collect multi-channel feedback through the workbench to obtain user feedback data; based on the user feedback data, the personalized recommendation list and search intent prediction data are iteratively optimized to realize the intelligent workbench search work.

[0077] The present invention builds a multi-level and multi-dimensional information foundation for the entire system by collecting user behavior data, building a user tag library, a content feature vector library and a scene context library, realizing the digitization of user portraits, content portraits and scene information, and laying a solid data foundation for subsequent personalized recommendations, search intent prediction and iterative optimization. Through user similarity calculation, user-content interaction extraction, scene weight fusion and association rule mining, a network of association relationships between users, content and scenes is constructed, which can more accurately portray the user's interest preferences, the semantic information of the content and the influencing factors of the scene, and provide a more accurate and richer basis for personalized recommendations and search intent prediction. Through real-time scene capture, rule matching, personalized sorting, diversity optimization and visual display, scene-aware personalized recommendations are realized, which can provide the freshest and highest-quality content based on the user's real-time scene, interest preferences and information needs, effectively improving user experience and platform value. Through user behavior trajectory analysis, frequent behavior pattern mining, semantic matching and target behavior prediction, search intention prediction based on historical behavior is realized. It can predict the user's search target based on the user's historical search behavior and current search terms, and recommend related search terms and content to help users find the required information more quickly and accurately, and improve the user's search experience. Through multi-channel collection of user feedback, semantic processing, data annotation, feature extraction and model optimization, a data-driven iterative optimization closed loop is constructed, which can continuously improve the recommendation effect and search quality of the intelligent workbench search, make the system more intelligent and personalized, and provide users with a more convenient, efficient and satisfactory search experience. Therefore, the present invention provides a smart workbench search optimization method based on multi-scenario business. By focusing on the user, integrating multi-scenario perception, artificial intelligence and data mining calculation, actively perceiving user needs, providing accurate and personalized information push and search assistance, and establishing a data-driven iterative optimization closed loop, the problem of lack of intelligent recommendation and prediction in the traditional smart workbench search optimization method based on multi-scenario business is solved, which greatly improves the intelligence level, user experience and work efficiency of the smart workbench. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 A schematic diagram of the steps of a smart workbench search optimization method based on multi-scenario business;

[0079] Figure 2 for Figure 1 Detailed implementation steps of step S3 in FIG.

[0080] Figure 3 for Figure 1 Detailed implementation steps of step S4 in FIG.

[0081] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0082] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0083] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0084] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0085] To achieve this, please refer to Figures 1 to 3 , an intelligent workbench search optimization method based on multi-scenario business, comprising the following steps:

[0086] Step S1: Under the authorization of the platform users, user behavior data of the platform users are collected to obtain user behavior data; behavior time series analysis is performed on the user behavior data to obtain a user behavior sequence; a user tag library is constructed according to the user behavior sequence to obtain a user tag library; content information is extracted from the workbench and a content feature vector library is constructed to obtain a content feature vector library; a scene context library is constructed according to the user tag library and the content feature vector library to obtain a scene context library;

[0087] Step S2: Optimize similarity propagation based on the user tag library to obtain a user similarity matrix; extract user-content interaction based on the content feature matrix to obtain a user-content interaction matrix; assign interaction matrix weights to the user-content interaction matrix based on the scene context library and the content feature vector library to obtain a user-content interaction weighted matrix; construct an association rule library based on the user-content interaction weighted matrix to obtain an association rule library;

[0088] Step S3: Real-time scene capture of platform users to obtain current scene tags; rule priority sorting according to current scene tags and the associated rule library to obtain a recommendation rule sorting set; user interest vector calculation according to the user tag library to obtain the user interest vector; personalized recommendation list generation according to the recommendation rule sorting set and the user interest vector to obtain a personalized recommendation list;

[0089] Step S4: extracting user behavior trajectories according to user behavior sequences to obtain user behavior trajectory data; dividing the user behavior trajectory data into time windows to obtain a user behavior time window data set; constructing a frequent behavior pattern library for the user behavior time window data set to obtain a frequent behavior pattern library; receiving search terms from platform users in real time to obtain current search term data; predicting search intent based on the current search term data and the frequent behavior pattern library to obtain search intent prediction data;

[0090] Step S5: Collect multi-channel feedback through the workbench to obtain user feedback data; iteratively optimize the search data for the personalized recommendation list and search intent prediction data based on the user feedback data to realize the intelligent workbench search work.

[0091] In the embodiment of the present invention, reference Figure 1 The above is a schematic diagram of the steps of the intelligent workbench search optimization method based on multi-scenario services of the present invention. In this example, the intelligent workbench search optimization method based on multi-scenario services includes the following steps:

[0092] Step S1: Under the authorization of the platform users, user behavior data of the platform users are collected to obtain user behavior data; behavior time series analysis is performed on the user behavior data to obtain a user behavior sequence; a user tag library is constructed according to the user behavior sequence to obtain a user tag library; content information is extracted from the workbench and a content feature vector library is constructed to obtain a content feature vector library; a scene context library is constructed according to the user tag library and the content feature vector library to obtain a scene context library;

[0093] In an embodiment of the present invention, first, with the user's authorization, the user's operation log on the platform is recorded, including browsing, searching, downloading and other behaviors, to form user behavior data. Then, the user behavior data is sorted in chronological order to form a user behavior sequence. And by analyzing the user behavior sequence, such as counting high-frequency behaviors, mining behavior patterns, etc., a user tag library is constructed to give each user a personalized tag. At the same time, information of various contents, such as document titles, abstracts, keywords, etc., is extracted from the workbench, and the content is converted into a vector representation using natural language processing technology to construct a content feature vector library. Finally, the user tag library and the content feature vector library are combined to construct a scenario context library to describe the association between users and content in different scenarios.

[0094] Step S2: Optimize similarity propagation based on the user tag library to obtain a user similarity matrix; extract user-content interaction based on the content feature matrix to obtain a user-content interaction matrix; assign interaction matrix weights to the user-content interaction matrix based on the scene context library and the content feature vector library to obtain a user-content interaction weighted matrix; construct an association rule library based on the user-content interaction weighted matrix to obtain an association rule library;

[0095] In the embodiment of the present invention, the similarity between users is calculated based on the user tag library, and the user similarity matrix is ​​optimized using the similarity propagation algorithm to characterize the association relationship between users. At the same time, a user-content interaction matrix is ​​constructed based on the user behavior sequence and the content feature vector library to represent the interaction intensity between users and different contents. Then, the user-content interaction matrix is ​​weighted according to the scene context library and the content feature vector library to obtain a user-content interaction weighted matrix to highlight the different preferences of users for content in different scenarios. Finally, the association rule mining algorithm is used to mine the association relationship between the content features that users frequently interact with in different scenarios from the user-content interaction weighted matrix to construct an association rule library.

[0096] Step S3: Real-time scene capture of platform users to obtain current scene tags; rule priority sorting according to current scene tags and the associated rule library to obtain a recommendation rule sorting set; user interest vector calculation according to the user tag library to obtain the user interest vector; personalized recommendation list generation according to the recommendation rule sorting set and the user interest vector to obtain a personalized recommendation list;

[0097] In the embodiment of the present invention, the user's operation behavior, time, location, device and other information are collected in real time to identify the user's current scene, and the relevant recommendation rules are retrieved from the association rule library according to the scene label, and the rules are sorted according to indicators such as confidence and support. At the same time, the user's interest vector for different content is calculated based on the user's label and the user similarity matrix. Then, according to the recommendation rules and the user's interest vector, the content is retrieved, sorted and filtered from the content feature vector library to form a personalized recommendation list, which is displayed to the user in a visual manner.

[0098] Step S4: extracting user behavior trajectories according to user behavior sequences to obtain user behavior trajectory data; dividing the user behavior trajectory data into time windows to obtain a user behavior time window data set; constructing a frequent behavior pattern library for the user behavior time window data set to obtain a frequent behavior pattern library; receiving search terms from platform users in real time to obtain current search term data; predicting search intent based on the current search term data and the frequent behavior pattern library to obtain search intent prediction data;

[0099] In an embodiment of the present invention, the user tag library, user behavior sequence and scenario context library are integrated to extract the user's historical behavior trajectory and divide it according to the time window. Then, a frequent pattern mining algorithm is used to mine frequently occurring user behavior sequences from the user behavior time window data set, and a frequent behavior pattern library is constructed based on the time interval information. At the same time, the search terms entered by the user are received in real time and converted into semantic vectors. Next, the semantic similarity between the search term semantic vector and each behavior pattern in the frequent behavior pattern library is calculated, the pattern with higher similarity is selected, and the user's next behavior is predicted based on the matching pattern. Finally, relevant search terms and content are recommended based on the predicted behavior, and the user's search intention is comprehensively judged in combination with the search terms and recommended content.

[0100] Step S5: Collect multi-channel feedback through the workbench to obtain user feedback data; perform search data iterative optimization on the personalized recommendation list and search intent prediction data according to the user feedback data to realize the intelligent workbench search work;

[0101] In the embodiment of the present invention, user feedback on search results and recommendation results, such as likes, dislikes, and favorites, is collected through multiple channels, and natural language processing technology is used to convert user feedback into semantic tags. Then, the user feedback data is annotated according to the semantic feedback tags, and features related to user feedback are extracted to form a feedback feature vector. Finally, the feedback annotation data set and feedback feature vector are used to train and optimize the personalized recommendation algorithm and search intent prediction model to improve the accuracy of recommendation results and search results and user satisfaction.

[0102] Preferably, step S1 comprises the following steps:

[0103] Step S11: Collecting user behavior data of platform users with authorization from platform users to obtain user behavior data; performing behavior time series analysis on the user behavior data to obtain a user behavior sequence; performing scene tagging on the user behavior sequence to obtain a scene tag set; constructing a user tag library based on the user behavior sequence to obtain a user tag library;

[0104] Step S12: extracting content information from the workbench to obtain content metadata; performing content semantic analysis on the content metadata to obtain a content semantic vector;

[0105] Step S13: performing content feature fusion on the content semantic vector and constructing a content feature vector library to obtain a content feature vector library;

[0106] Step S14: Associating scene information according to the scene tag set and performing vector encoding to obtain a scene feature vector;

[0107] Step S15: Perform scene context association on the user tag library and the content feature vector library according to the scene feature vector and construct a scene context library to obtain a scene context library.

[0108] In an embodiment of the present invention, after the platform user agrees to the authorization agreement, the user's operation log is recorded to form a user behavior database. The user behavior data is parsed using natural language processing technology to extract key behaviors and operation objects, and arranged in chronological order to form a user behavior sequence. According to the pre-defined scenario rules or models, each behavior in the user behavior sequence is scene-tagged. The clustering algorithm is used to perform cluster analysis on the behavior sequences of all users, and users with similar behavior patterns are divided into the same user behavior cluster. Combined with the user's basic information, the behavior cluster to which the user belongs, and the scenario label, a corresponding label is generated for each user, and a user label library is constructed. Information on various contents, such as document titles, abstracts, keywords, authors, etc., is extracted from the workbench to form a content metadata database. Then, the natural language processing technology is used to perform semantic analysis on the content metadata, and the text information is converted into a vector representation to form a content semantic vector. The content semantic vector is processed by dimensionality reduction, feature selection, etc., and other content features, such as document type, release time, popularity, etc., are integrated to construct a content feature vector library. According to each scene tag in the scene tag set, the relevant scene information is associated, and the scene information is encoded into a vector using Word2Vec and other technologies to form a scene feature vector. Using the scene feature vector, the semantic similarity between each user tag in the user tag library and each content feature vector in the content feature vector library is calculated, and the tag-feature pairs with similarity higher than the set threshold are stored in the scene context library to describe the association relationship between users and content in different scenarios.

[0109] Preferably, step S11 includes the following steps:

[0110] Step S111: registering basic information of platform users and collecting user behavior data with the authorization of the platform users, and obtaining basic information and user behavior data;

[0111] Step S112: performing user behavior analysis on the user behavior data and performing behavior time series analysis to obtain a user behavior sequence;

[0112] Step S113: identifying scene elements of the user behavior sequence and marking the scene to obtain a scene label set;

[0113] Step S114: performing user behavior cluster analysis according to the user behavior sequence to obtain user behavior clusters;

[0114] Step S115: Generate user tags and construct a user tag library based on user basic information, user behavior clusters, and scene tag sets to obtain a user tag library.

[0115] In the embodiment of the present invention, after the platform user agrees to the authorization agreement, the user is guided to fill in or import personal basic information, such as name, department, position, professional skills, etc. At the same time, the user's platform operation log is recorded, including but not limited to the user's browsing history, search terms, click content, stay time, operation time, location, device and other information, to form a user behavior database. The user behavior data is cleaned and preprocessed to remove invalid data and noise data. Then, the user behavior data is parsed using event extraction, named entity recognition and other technologies, and key behaviors and operation objects, such as "browsing documents", "searching keywords", "downloading files", etc., are extracted and arranged in chronological order to form a user behavior sequence. According to the pre-defined scene elements and rules, each behavior in the user behavior sequence is identified and marked. For example, whether it is working time is determined according to the operation time, whether it is in the company is determined according to the location information, and whether a mobile device is used is determined according to the device information. Combining these elements, the user's behavior can be marked as "work scene", "meeting scene", "commuting scene", etc. Clustering analysis is performed on the behavior sequences of all users using a clustering algorithm, such as the K-means algorithm or the DBSCAN algorithm. By calculating the similarity between user behavior sequences, users with similar behavior patterns are grouped into the same user behavior cluster. For example, users who frequently browse technical documents and search for code examples can be classified into the "technical R&D personnel" cluster. Considering the basic information of the user, the behavior cluster to which the user belongs, and the scenario label, corresponding labels are generated for each user, and a user label library is constructed. For example, a user whose basic information is a software engineer and who frequently browses technical documents and searches for code examples on weekday mornings can be labeled "technical R&D personnel", "Java programming", "software development", etc.

[0116] Preferably, step S2 comprises the following steps:

[0117] Step S21: construct a user tag matrix according to the user tag library to obtain a user tag matrix; perform tag co-occurrence statistics according to the user tag matrix to obtain a tag co-occurrence matrix;

[0118] Step S22: Calculate user similarity based on the tag co-occurrence matrix to obtain a user initial similarity matrix; perform similarity propagation optimization on the user initial similarity matrix to obtain a user similarity matrix;

[0119] Step S23: performing user interaction behavior analysis on the user behavior sequence to obtain user interaction behavior data; constructing a content feature matrix according to the content feature vector library to obtain a content feature matrix;

[0120] Step S24: extracting user-content interaction from the user interaction behavior data according to the content feature matrix to obtain a user-content interaction matrix;

[0121] Step S25: Calculate scene attention weights according to the scene context library, the content feature vector library, and the user interaction behavior data to obtain scene attention weight data; assign interaction matrix weights to the user-content interaction matrix according to the scene attention weight data to obtain a user-content interaction weighted matrix;

[0122] Step S26: performing association rule mining according to the user-content interaction weighted matrix to obtain an initial association rule set;

[0123] Step S27: Filter the initial association rule set by rule confidence using a preset rule confidence threshold to obtain a high-confidence association rule set; merge and simplify the high-confidence association rule set to obtain an association rule base.

[0124] In the embodiment of the present invention, a user-label matrix is ​​constructed according to a user label library, wherein the rows of the matrix represent users, the columns represent labels, and the matrix elements represent the weight of the user having the label, which can be calculated using methods such as TF-IDF. Then, according to the user-label matrix, the frequency of co-occurrence between labels is counted, and a label co-occurrence matrix is ​​constructed, wherein the matrix elements represent how many users two labels co-occur. According to the label co-occurrence matrix, the initial similarity between users is calculated. Then, similarity propagation algorithms such as SimRank are used to iteratively calculate the similarity between users until convergence, and the final user similarity matrix is ​​obtained. The user behavior sequence is analyzed, and the interaction behavior between users and content is extracted to form user interaction behavior data. At the same time, according to the content feature vector library, a content feature matrix is ​​constructed, wherein the rows of the matrix represent content, the columns represent features, and the matrix elements represent the degree to which the content has the features. According to the content feature matrix, the content ID in the user interaction behavior data is converted into the corresponding feature vector, and a user-content interaction matrix is ​​constructed, wherein the rows of the matrix represent users, the columns represent content features, and the matrix elements represent the interaction intensity between users and the content features, which can be calculated using indicators such as interaction frequency and interaction time. According to the scene context library, the association between user tags and content features in the current scene is extracted, and the scene attention weight is calculated to indicate the degree of attention of users to different content features in different scenes. Then, according to the scene attention weight, the user-content interaction matrix is ​​weighted to obtain the user-content interaction weighted matrix to highlight the different preferences of users for content in different scenes. Using the Apriori algorithm or the FP-Growth algorithm, the user-content interaction weighted matrix is ​​mined for association rules to find the association between content features that users frequently interact with in different scenes, for example, users often browse "technical documents" and "code examples" in the "work scene". The rule confidence threshold is set to filter out association rules with low confidence, retain association rules with high confidence, and form a high-confidence association rule set. Then, the high-confidence association rule set is merged and streamlined, redundant rules are removed, and the most concise and representative association rules are retained to finally form an association rule library.

[0125] Preferably, step S25 comprises the following steps:

[0126] Step S251: Generate a scene vector according to the scene context library to obtain a scene vector set; contextualize the user interaction behavior data according to the scene vector set to obtain contextualized user behavior data;

[0127] Step S252: performing interactive behavior semantic vectorization on the user interactive behavior data to obtain an interactive behavior semantic vector; performing content semantic vector extraction based on a content feature vector library to obtain a content semantic vector;

[0128] Step S253: Calculate the scene attention weight according to the scene vector set, the interaction behavior semantic vector and the content semantic vector to obtain scene attention weight data;

[0129] Step S254: performing scene weight fusion according to the scene attention weight data and the scene-based user behavior data to obtain fused scene weight data;

[0130] Step S255: assigning interaction matrix weights to the user-content interaction matrix according to the fused scene weight data to obtain a user-content interaction weighted matrix.

[0131] In an embodiment of the present invention, first, according to the user tags and content features corresponding to each scene in the scene context library, Word2Vec and other methods are used to represent the scene as a low-dimensional dense scene vector to form a scene vector set. Then, according to the scene label of each interactive behavior in the user interactive behavior data, the user interactive behavior data is associated with the scene vector to form scenario-based user behavior data, for example, user A browsed "technical document B" under the "work scene". Using word vector models such as Word2Vec or GloVe, the interactive behaviors (for example, "browse", "search") in the user interactive behavior data are converted into corresponding interactive behavior semantic vectors. At the same time, the semantic vectors of the content involved in the user interactive behavior data are extracted from the content feature vector library. Using the attention mechanism, the correlation between the scene vector, the interactive behavior semantic vector and the content semantic vector is calculated to obtain the scene attention weight data, which indicates the degree of attention paid to different aspects of the content when the user performs a specific interactive behavior in a specific scene. According to the scene attention weight data and the scenario user behavior data, the scene attention weights in different scenes are fused. For example, the weighted average method can be used to obtain the fused scene weight data, which represents the comprehensive attention of users to content in different scenes. According to the fused scene weight data, the user-content interaction matrix is ​​weighted to obtain the user-content interaction weight matrix. Specifically, the value of each element in the user-content interaction matrix is ​​multiplied by the corresponding fused scene weight to adjust the intensity of user-content interaction in different scenes.

[0132] Preferably, step S3 comprises the following steps:

[0133] Step S31: Real-time scene capture of platform users to obtain current scene labels; scene rule matching of current scene labels and associated rule bases to obtain candidate recommendation rule sets;

[0134] Step S32: sorting the candidate recommendation rule sets by rule priority according to the association rule library to obtain a recommendation rule sorting set; calculating the user interest vector according to the user tag library and the user similarity matrix to obtain the user interest vector;

[0135] Step S33: performing content retrieval on the content feature vector library according to the recommendation rule sorting set to obtain a content candidate set; performing content interest score calculation according to the user interest vector and the content candidate set to obtain a content interest list;

[0136] Step S34: performing diversity sorting optimization on the content interest list to obtain a diversity content recommendation list; performing time decay adjustment on the diversity content recommendation list to obtain a content recommendation time decay list;

[0137] Step S35: Visualize the recommendation results according to the content recommendation time decay list to obtain a personalized recommendation list.

[0138] As an example of the present invention, refer to Figure 2 As shown, in this example, step S3 includes:

[0139] Step S31: Real-time scene capture of platform users to obtain current scene labels; scene rule matching of current scene labels and associated rule bases to obtain candidate recommendation rule sets;

[0140] In the embodiment of the present invention, the user's operation behavior, time, location, equipment and other information are collected in real time, and the scene the user is currently in is determined according to the preset scene recognition rules, such as "work scene" or "meeting scene", and the scene is marked as the current scene label. Then, according to the current scene label, the association rules related to the scene are retrieved from the association rule library to form a candidate recommendation rule set.

[0141] Step S32: sorting the candidate recommendation rule sets by rule priority according to the association rule library to obtain a recommendation rule sorting set; calculating the user interest vector according to the user tag library and the user similarity matrix to obtain the user interest vector;

[0142] In the embodiment of the present invention, the candidate recommendation rule set is prioritized according to the confidence, support, time decay and other indicators of the association rule to obtain the recommendation rule ranking set, and the rules with higher rankings are more in line with the user's needs in the current scenario. At the same time, the user's interest vector is calculated according to the user's tag and user similarity matrix to characterize the user's preference for different content.

[0143] Step S33: performing content retrieval on the content feature vector library according to the recommendation rule sorting set to obtain a content candidate set; performing content interest score calculation according to the user interest vector and the content candidate set to obtain a content interest list;

[0144] In the embodiment of the present invention, according to the content feature conditions in the recommendation rule sorting set, the content that meets the conditions is retrieved from the content feature vector library to form a content candidate set. Then, the similarity between the user interest vector and the feature vector of each content in the content candidate set is calculated as the content interest score, and a content interest list is formed according to the score.

[0145] Step S34: performing diversity sorting optimization on the content interest list to obtain a diversity content recommendation list; performing time decay adjustment on the diversity content recommendation list to obtain a content recommendation time decay list;

[0146] In the embodiment of the present invention, in order to avoid the recommendation results being too single, the content interest list is optimized for diversity sorting. For example, factors such as the subject, type, and source of the content can be considered to ensure that the recommendation list contains a variety of content. In addition, the diversity content recommendation list can also be adjusted for time decay according to the release time of the content, so that content with a more recent release time is ranked at a higher position.

[0147] Step S35: Visualize the recommendation results according to the content recommendation time decay list to obtain a personalized recommendation list.

[0148] In an embodiment of the present invention, the top-ranked content is selected based on the content recommendation time decay list and visualized in a user-friendly manner. For example, the content title, summary, picture and other information can be organized into a card format to form a final personalized recommendation list for display to the user.

[0149] Preferably, step S4 comprises the following steps:

[0150] Step S41: extracting user behavior trajectories from the user tag library, user behavior sequence, and scene context library to obtain user behavior trajectory data;

[0151] Step S42: Divide the user behavior trajectory data into time windows to obtain a user behavior time window data set; construct a frequent behavior pattern library for the user behavior time window data set to obtain a frequent behavior pattern library;

[0152] Step S43: receiving search terms from platform users in real time to obtain current search term data; performing search term semantic vectorization on the current search term data to obtain a search term semantic vector;

[0153] Step S44: performing historical behavior semantic matching on the search word semantic vector and the frequent behavior pattern library to obtain a matching behavior pattern; performing a predicted target behavior inference based on the matching behavior pattern and the current search word data to obtain predicted target behavior data;

[0154] Step S45: Generate search word recommendations based on the predicted target behavior data to obtain a search word prediction list; calculate search content relevance based on the search word semantic vector and the content feature vector library to obtain content relevance score data;

[0155] Step S46: perform sorting and recommendation of related content based on the content relevance score data to obtain a related content prediction list; perform search intent prediction based on the search term prediction list and the related content prediction list to obtain search intent prediction data.

[0156] As an example of the present invention, refer to Figure 3 As shown, in this example, step S4 includes:

[0157] Preferably, step S4 comprises the following steps:

[0158] Step S41: extracting user behavior trajectories from the user tag library, user behavior sequence, and scene context library to obtain user behavior trajectory data;

[0159] In an embodiment of the present invention, a user tag library, a user behavior sequence, and a scenario context library are integrated to extract the user's historical behavior trajectory to form user behavior trajectory data. For example, user A browsed technical document B, searched for keyword C, downloaded file D, and other information in a work scenario at time T1.

[0160] Step S42: Divide the user behavior trajectory data into time windows to obtain a user behavior time window data set; construct a frequent behavior pattern library for the user behavior time window data set to obtain a frequent behavior pattern library;

[0161] In the embodiment of the present invention, the user behavior trajectory data is divided into multiple time windows according to the preset time window size (for example, 30 minutes), and the user behavior in each time window constitutes a user behavior time window data. Then, a frequent pattern mining algorithm, such as the Apriori algorithm or the FP-Growth algorithm, is used to mine frequently occurring user behavior sequences from the user behavior time window data set to build a frequent behavior pattern library. For example, after searching for the keyword "machine learning", the user is likely to browse "Python tutorial" and "deep learning framework".

[0162] Step S43: receiving search terms from platform users in real time to obtain current search term data; performing search term semantic vectorization on the current search term data to obtain a search term semantic vector;

[0163] In the embodiment of the present invention, the search words input by the user on the platform are received in real time, and a word vector model such as Word2Vec or GloVe is used to convert the search words into corresponding semantic vectors to facilitate subsequent semantic matching.

[0164] Step S44: performing historical behavior semantic matching on the search word semantic vector and the frequent behavior pattern library to obtain a matching behavior pattern; performing a predicted target behavior inference based on the matching behavior pattern and the current search word data to obtain predicted target behavior data;

[0165] In the embodiment of the present invention, the semantic similarity between the search word semantic vector and each behavior pattern in the frequent behavior pattern library is calculated, for example, using cosine similarity. The frequent behavior patterns with similarity higher than a preset threshold are selected as matching behavior patterns. Then, based on the matching behavior patterns and the current search word, the user's next possible behavior is predicted. For example, after searching for "machine learning", the user may search for "deep learning".

[0166] Step S45: Generate search word recommendations based on the predicted target behavior data to obtain a search word prediction list; calculate search content relevance based on the search word semantic vector and the content feature vector library to obtain content relevance score data;

[0167] In the embodiment of the present invention, based on the predicted target behavior data, search terms related to the predicted behavior are recommended. For example, if it is predicted that the user may search for "deep learning" next, search terms such as "deep learning framework" and "deep learning algorithm" are recommended to form a search term prediction list. At the same time, based on the search term semantic vector and the content feature vector library, the relevance score between the search term and each content is calculated, for example, using cosine similarity.

[0168] Step S46: performing ranking and recommendation of related content according to the content relevance score data to obtain a related content prediction list; performing search intent prediction according to the search word prediction list and the related content prediction list to obtain search intent prediction data;

[0169] In the embodiment of the present invention, the content is sorted according to the content relevance score data, and the content with high relevance is recommended to form a related content prediction list. Finally, the search term prediction list and the related content prediction list are combined to comprehensively judge the user's search intention. For example, after the user searches for "machine learning", "deep learning framework" and "Python tutorial" are recommended, and the user's search intention is to learn knowledge and skills related to machine learning.

[0170] Preferably, step S42 includes the following steps:

[0171] Step S421: Calculate the time interval of the user behavior trajectory data to obtain the user behavior timestamp data;

[0172] In the embodiment of the present invention, the timestamp of each behavior in the user behavior trajectory data is analyzed, and the time interval between adjacent behaviors is calculated. For example, user A browsed document B at 10:00 and searched for keyword C at 10:05, then the time interval between these two behaviors is 5 minutes.

[0173] Step S422: dividing the user behavior trajectory data into time windows according to the user behavior timestamp data to obtain a user behavior time window data set;

[0174] In the embodiment of the present invention, the user behavior trajectory data is divided into multiple time windows according to the preset time window size (for example, 30 minutes) and time interval information to ensure that the behavior in each time window occurs continuously. For example, all the behaviors of user A from 9:00 to 9:30 are divided into one time window, and the behaviors from 9:30 to 10:00 are divided into another time window, and so on.

[0175] Step S423: mining the behavior sequence pattern in the window on the user behavior time window data set to obtain an initial frequent behavior pattern set;

[0176] In an embodiment of the present invention, a frequent pattern mining algorithm, such as an Apriori algorithm or an FP-Growth algorithm, is used for each user behavior time window data set to mine user behavior sequences that frequently appear in the time window to form an initial frequent behavior pattern set. For example, in many 30-minute time windows, users often search for the keyword "Python" first and then browse "Python tutorial".

[0177] Step S424: performing time interval statistics on the initial frequent behavior pattern set to obtain time interval frequent behavior patterns; performing pattern support calculation on the time interval frequent behavior patterns to obtain support frequent behavior patterns;

[0178] In the embodiment of the present invention, the time interval distribution of each behavior pattern in the initial frequent behavior pattern set is analyzed. For example, the interval after the user searches for the keyword "Python" is counted, and the frequency of each time interval is calculated to form a time interval frequent behavior pattern. Then, the support of each time interval frequent behavior pattern is calculated, that is, the proportion of the pattern appearing in all time windows. For example, in 80% of the time windows, the user will browse the "Python tutorial" within 5 minutes after searching for "Python".

[0179] Step S425: Filter the support frequent behavior patterns using a preset support threshold and construct a frequent behavior pattern library to obtain a frequent behavior pattern library;

[0180] In an embodiment of the present invention, a support threshold is set, for example, 0.6, to filter out frequent behavior patterns of time intervals whose support is lower than the threshold, retain patterns with high support, and store these patterns in a frequent behavior pattern library for subsequent search intent prediction.

[0181] Preferably, step S5 comprises the following steps:

[0182] Step S51: Collect multi-channel feedback through the workbench to obtain user feedback data;

[0183] Step S52: semantically categorize the user feedback data into feedback behavior to obtain a semantically categorized feedback behavior label;

[0184] Step S53: annotating the user feedback data according to the semantic feedback behavior label to obtain a feedback annotation data set;

[0185] Step S54: extracting feedback features from the feedback annotation data set to obtain a feedback feature vector;

[0186] Step S55: Perform search data iterative optimization on the personalized recommendation list and search intent prediction data according to the feedback annotation data set and the feedback feature vector to implement the intelligent workbench search work.

[0187] In an embodiment of the present invention, user feedback on search results and recommended results is collected through multiple channels, for example, buttons such as like, dislike, and favorite are added to the search results page, allowing users to rate or comment on recommended results, collecting feedback information actively submitted by users, etc., to form a user feedback database. Natural language processing technology is used to perform semantic analysis on user feedback data, extract user feedback behavior and feedback objects, and convert them into semantic feedback behavior labels, for example, convert "like" into "positive", convert "dislike" into "negative", convert "favorite" into "interested", etc. According to the semantic feedback behavior labels, the user feedback data is annotated, for example, the search results or recommended results liked by the user are annotated as "positive", and the results disliked by the user are annotated as "negative", forming a feedback annotation data set for training and optimizing recommendation algorithms and search models. Features related to user feedback are extracted from the feedback annotation data set, such as user features (e.g., user tags, user behavior history), content features (e.g., content theme, content type), scene features (e.g., time, location, device), etc., and these features are converted into numerical vectors to form feedback feature vectors. Feedback annotation datasets and feedback feature vectors are used to train and optimize personalized recommendation algorithms and search intent prediction models. For example, machine learning algorithms can be used to adjust the parameters of the recommendation algorithm and the weights of the model based on the user's feedback behavior and feedback characteristics to improve the accuracy of recommendation results and search results and user satisfaction.

[0188] Preferably, the present invention further provides an intelligent workbench search optimization system based on multi-scenario services, which is used to execute the intelligent workbench search optimization method based on multi-scenario services as described above, and the intelligent workbench search optimization system based on multi-scenario services comprises:

[0189] The multi-dimensional portrait construction module is used to collect user behavior data with the authorization of platform users to obtain user behavior data; perform behavior time series analysis on user behavior data to obtain user behavior sequences; construct a user tag library based on the user behavior sequences to obtain a user tag library; extract content information from the workbench and construct a content feature vector library to obtain a content feature vector library; construct a scene context library based on the user tag library and the content feature vector library to obtain a scene context library;

[0190] The association rule mining module is used to optimize similarity propagation based on the user tag library to obtain a user similarity matrix; extract user-content interaction based on the content feature matrix to obtain a user-content interaction matrix; assign interaction matrix weights to the user-content interaction matrix based on the scene context library and the content feature vector library to obtain a user-content interaction weighted matrix; construct an association rule library based on the user-content interaction weighted matrix to obtain an association rule library;

[0191] The scene intelligent recommendation module is used to capture the scene of the platform users in real time and obtain the current scene label; sort the rule priorities according to the current scene label and the associated rule library to obtain the recommended rule sorting set; calculate the user interest vector according to the user tag library to obtain the user interest vector; generate a personalized recommendation list according to the recommended rule sorting set and the user interest vector to obtain a personalized recommendation list;

[0192] The search dynamic prediction module is used to extract user behavior trajectories according to user behavior sequences to obtain user behavior trajectory data; divide the user behavior trajectory data into time windows to obtain a user behavior time window data set; construct a frequent behavior pattern library for the user behavior time window data set to obtain a frequent behavior pattern library; receive search terms from platform users in real time to obtain current search term data; predict search intent based on current search term data and the frequent behavior pattern library to obtain search intent prediction data;

[0193] The search feedback optimization module is used to collect multi-channel feedback through the workbench to obtain user feedback data; based on the user feedback data, the personalized recommendation list and search intent prediction data are iteratively optimized to realize the intelligent workbench search work.

[0194] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0195] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A smart workbench search optimization method based on multi-scenario business, characterized in that: The following steps are involved: Step S1: Collecting user behavior data of platform users with the authorization of platform users to obtain user behavior data; Conduct behavioral time series analysis on user behavior data to obtain user behavior sequence; construct a user tag library based on the user behavior sequence to obtain a user tag library; extract content information from the workbench and construct a content feature vector library to obtain a content feature vector library; construct a scene context library based on the user tag library and the content feature vector library to obtain a scene context library; Step S2: Optimize similarity propagation based on the user tag library to obtain a user similarity matrix; extract user-content interaction based on the content feature matrix to obtain a user-content interaction matrix; Assign interaction matrix weights to the user-content interaction matrix according to the scene context library and the content feature vector library to obtain a user-content interaction weighted matrix; construct an association rule library according to the user-content interaction weighted matrix to obtain an association rule library; Step S3: Real-time scene capture of platform users to obtain current scene tags; rule priority sorting according to current scene tags and associated rule library to obtain recommendation rule sorting set; user interest vector calculation according to user tag library to obtain user interest vector; personalized recommendation list generation according to recommendation rule sorting set and user interest vector to obtain personalized recommendation list, wherein step S3 is specifically as follows: Step S31: Real-time scene capture of platform users to obtain current scene labels; scene rule matching of current scene labels and associated rule bases to obtain candidate recommendation rule sets; Step S32: sorting the candidate recommendation rule sets by rule priority according to the association rule library to obtain a recommendation rule sorting set; calculating the user interest vector according to the user tag library and the user similarity matrix to obtain the user interest vector; Step S33: performing content retrieval on the content feature vector library according to the recommendation rule sorting set to obtain a content candidate set; performing content interest score calculation according to the user interest vector and the content candidate set to obtain a content interest list; Step S34: performing diversity sorting optimization on the content interest list to obtain a diversity content recommendation list; Performing time decay adjustment on the diversity content recommendation list to obtain a content recommendation time decay list; Step S35: Visualize the recommendation results according to the content recommendation time decay list to obtain a personalized recommendation list; Step S4: extracting user behavior trajectories according to the user behavior sequence to obtain user behavior trajectory data; Divide the user behavior trajectory data into time windows to obtain a user behavior time window data set; construct a frequent behavior pattern library for the user behavior time window data set to obtain a frequent behavior pattern library; receive search terms from platform users in real time to obtain current search term data; predict search intent based on current search term data and the frequent behavior pattern library to obtain search intent prediction data; Step S5: Collect multi-channel feedback through the workbench to obtain user feedback data; Based on user feedback data, the personalized recommendation list and search intent prediction data are iteratively optimized to realize intelligent workbench search work.

2. The intelligent workbench search optimization method based on multi-scenario business according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: Collecting user behavior data of platform users with authorization from platform users to obtain user behavior data; performing behavior time series analysis on the user behavior data to obtain a user behavior sequence; performing scene tagging on the user behavior sequence to obtain a scene tag set; constructing a user tag library based on the user behavior sequence to obtain a user tag library; Step S12: extracting content information from the workbench to obtain content metadata; performing content semantic analysis on the content metadata to obtain a content semantic vector; Step S13: performing content feature fusion on the content semantic vector and constructing a content feature vector library to obtain a content feature vector library; Step S14: Associating scene information according to the scene tag set and performing vector encoding to obtain a scene feature vector; Step S15: Perform scene context association on the user tag library and the content feature vector library according to the scene feature vector and construct a scene context library to obtain a scene context library.

3. The intelligent workbench search optimization method based on multi-scenario business according to claim 2 is characterized in that: Step S11 includes the following steps: Step S111: registering the basic information of the platform user and collecting the user behavior data with the authorization of the platform user to obtain the basic information and behavior data of the user; Step S112: performing user behavior analysis on the user behavior data and performing behavior time series analysis to obtain a user behavior sequence; Step S113: identifying scene elements of the user behavior sequence and marking the scene to obtain a scene label set; Step S114: performing user behavior cluster analysis according to the user behavior sequence to obtain user behavior clusters; Step S115: Generate user tags and construct a user tag library based on user basic information, user behavior clusters, and scene tag sets to obtain a user tag library.

4. The intelligent workbench search optimization method based on multi-scenario business according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: construct a user tag matrix according to the user tag library to obtain a user tag matrix; perform tag co-occurrence statistics according to the user tag matrix to obtain a tag co-occurrence matrix; Step S22: Calculate user similarity based on the tag co-occurrence matrix to obtain a user initial similarity matrix; perform similarity propagation optimization on the user initial similarity matrix to obtain a user similarity matrix; Step S23: performing user interaction behavior analysis on the user behavior sequence to obtain user interaction behavior data; constructing a content feature matrix according to the content feature vector library to obtain a content feature matrix; Step S24: extracting user-content interaction from the user interaction behavior data according to the content feature matrix to obtain a user-content interaction matrix; Step S25: Calculate scene attention weights according to the scene context library, the content feature vector library, and the user interaction behavior data to obtain scene attention weight data; assign interaction matrix weights to the user-content interaction matrix according to the scene attention weight data to obtain a user-content interaction weighted matrix; Step S26: performing association rule mining according to the user-content interaction weighted matrix to obtain an initial association rule set; Step S27: Filter the initial association rule set by rule confidence using a preset rule confidence threshold to obtain a high-confidence association rule set; merge and simplify the high-confidence association rule set to obtain an association rule base.

5. The intelligent workbench search optimization method based on multi-scenario business according to claim 4 is characterized in that: Step S25 includes the following steps: Step S251: Generate a scene vector according to the scene context library to obtain a scene vector set; contextualize the user interaction behavior data according to the scene vector set to obtain contextualized user behavior data; Step S252: performing interactive behavior semantic vectorization on the user interactive behavior data to obtain an interactive behavior semantic vector; performing content semantic vector extraction based on a content feature vector library to obtain a content semantic vector; Step S253: Calculate the scene attention weight according to the scene vector set, the interaction behavior semantic vector and the content semantic vector to obtain scene attention weight data; Step S254: performing scene weight fusion according to the scene attention weight data and the scene-based user behavior data to obtain fused scene weight data; Step S255: assigning interaction matrix weights to the user-content interaction matrix according to the fused scene weight data to obtain a user-content interaction weighted matrix.

6. The intelligent workbench search optimization method based on multi-scenario business according to claim 1 is characterized in that: Step S4 includes the following steps: Step S41: extracting user behavior trajectories from the user tag library, user behavior sequence, and scene context library to obtain user behavior trajectory data; Step S42: Divide the user behavior trajectory data into time windows to obtain a user behavior time window data set; construct a frequent behavior pattern library for the user behavior time window data set to obtain a frequent behavior pattern library; Step S43: receiving search terms from platform users in real time to obtain current search term data; performing search term semantic vectorization on the current search term data to obtain a search term semantic vector; Step S44: performing historical behavior semantic matching on the search word semantic vector and the frequent behavior pattern library to obtain a matching behavior pattern; performing a predicted target behavior inference based on the matching behavior pattern and the current search word data to obtain predicted target behavior data; Step S45: Generate search word recommendations based on the predicted target behavior data to obtain a search word prediction list; calculate search content relevance based on the search word semantic vector and the content feature vector library to obtain content relevance score data; Step S46: perform sorting and recommendation of related content based on the content relevance score data to obtain a related content prediction list; perform search intent prediction based on the search term prediction list and the related content prediction list to obtain search intent prediction data.

7. The intelligent workbench search optimization method based on multi-scenario business according to claim 6 is characterized in that: Step S42 includes the following steps: Step S421: Calculate the time interval of the user behavior trajectory data to obtain the user behavior timestamp data; Step S422: dividing the user behavior trajectory data into time windows according to the user behavior timestamp data to obtain a user behavior time window data set; Step S423: mining the behavior sequence pattern in the window on the user behavior time window data set to obtain an initial frequent behavior pattern set; Step S424: performing time interval statistics on the initial frequent behavior pattern set to obtain time interval frequent behavior patterns; performing pattern support calculation on the time interval frequent behavior patterns to obtain support frequent behavior patterns; Step S425: Filter the support frequent behavior patterns using a preset support threshold and construct a frequent behavior pattern library to obtain a frequent behavior pattern library.

8. The intelligent workbench search optimization method based on multi-scenario business according to claim 1 is characterized in that: Step S5 includes the following steps: Step S51: Collect multi-channel feedback through the workbench to obtain user feedback data; Step S52: semantically categorize the user feedback data into feedback behavior to obtain a semantically categorized feedback behavior label; Step S53: annotating the user feedback data according to the semantic feedback behavior label to obtain a feedback annotation data set; Step S54: extracting feedback features from the feedback annotation data set to obtain a feedback feature vector; Step S55: Perform search data iterative optimization on the personalized recommendation list and search intent prediction data according to the feedback annotation data set and the feedback feature vector to implement the intelligent workbench search work.

9. The intelligent workbench search optimization system based on multi-scenario business is characterized by: Used to execute the intelligent workbench search optimization method based on multi-scenario services as claimed in claim 1, the intelligent workbench search optimization system based on multi-scenario services comprises: The multi-dimensional portrait construction module is used to collect user behavior data with the authorization of platform users to obtain user behavior data; perform behavior time series analysis on user behavior data to obtain user behavior sequences; construct a user tag library based on the user behavior sequences to obtain a user tag library; extract content information from the workbench and construct a content feature vector library to obtain a content feature vector library; construct a scene context library based on the user tag library and the content feature vector library to obtain a scene context library; The association rule mining module is used to optimize similarity propagation based on the user tag library to obtain a user similarity matrix; extract user-content interaction based on the content feature matrix to obtain a user-content interaction matrix; assign interaction matrix weights to the user-content interaction matrix based on the scene context library and the content feature vector library to obtain a user-content interaction weighted matrix; construct an association rule library based on the user-content interaction weighted matrix to obtain an association rule library; The scene intelligent recommendation module is used to capture the scene of the platform users in real time and obtain the current scene label; sort the rule priorities according to the current scene label and the associated rule library to obtain the recommended rule sorting set; calculate the user interest vector according to the user tag library to obtain the user interest vector; generate a personalized recommendation list according to the recommended rule sorting set and the user interest vector to obtain a personalized recommendation list; the scene intelligent recommendation module is used to: Step S31: Real-time scene capture of platform users to obtain current scene labels; scene rule matching of current scene labels and associated rule bases to obtain candidate recommendation rule sets; Step S32: sorting the candidate recommendation rule sets by rule priority according to the association rule library to obtain a recommendation rule sorting set; calculating the user interest vector according to the user tag library and the user similarity matrix to obtain the user interest vector; Step S33: performing content retrieval on the content feature vector library according to the recommendation rule sorting set to obtain a content candidate set; performing content interest score calculation according to the user interest vector and the content candidate set to obtain a content interest list; Step S34: performing diversity sorting optimization on the content interest list to obtain a diversity content recommendation list; performing time decay adjustment on the diversity content recommendation list to obtain a content recommendation time decay list; Step S35: Visualize the recommendation results according to the content recommendation time decay list to obtain a personalized recommendation list; The search dynamic prediction module is used to extract user behavior trajectories according to user behavior sequences to obtain user behavior trajectory data; divide the user behavior trajectory data into time windows to obtain a user behavior time window data set; construct a frequent behavior pattern library for the user behavior time window data set to obtain a frequent behavior pattern library; receive search terms from platform users in real time to obtain current search term data; predict search intent based on current search term data and the frequent behavior pattern library to obtain search intent prediction data; The search feedback optimization module is used to collect multi-channel feedback through the workbench to obtain user feedback data; based on the user feedback data, the personalized recommendation list and search intent prediction data are iteratively optimized to realize the intelligent workbench search work.

Images