Content matching method and system based on dynamic combination characteristics

Abstract

The invention relates to the technical field of feature matching, and discloses a dynamic combination feature-based content matching method, which comprises the following steps of: acquiring user behavior data, content attribute data and a service scene label in real time; generating a user behavior feature vector according to the user behavior data, generating a scene adaptation feature vector according to the service scene label, and fusing the user behavior feature vector and the scene adaptation feature vector to generate a dynamic combination feature vector; calculating the matching degree of the user and the content based on the dynamic combination feature vector; sorting or recommending the content according to the matching degree, and dynamically optimizing a matching strategy through a real-time feedback mechanism; in the live broadcast recommendation scene, the recommendation priority is dynamically adjusted according to the recent reservation behavior of the user and the service target; and in the course matching scene, calling a user social relation chain and an occupational portrait corresponding to the user, and generating a course recommendation list in combination with the user social relation chain and the occupational portrait. The scene adaptability of the recommendation system can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of feature matching, in particular to a content matching method and system based on dynamic combination features. BACKGROUND

[0002] Current content recommendation systems generally rely on static tags (such as user basic attributes, historical behaviors) for matching, and are difficult to cope with dynamic changes in user interest and complex needs of business scenarios. For example, when user behavior fluctuates over time and business goals (such as lead generation or precision marketing) require differentiated strategies, existing methods lack the ability to dynamically respond to real-time behavior and scene requirements, resulting in recommended results deviating from the actual preferences of users, affecting platform conversion efficiency and user experience.

[0003] In addition, the existing technology has limitations in multi-dimensional data fusion, which limits the scene adaptability and personalization ability of the recommendation system; as known from the above, how to improve the scene adaptability of the recommendation system still needs to be solved. SUMMARY

[0004] In order to improve the scene adaptability of the recommendation system, the present application provides a content matching method and system based on dynamic combination features.

[0005] In the first aspect, the present application provides a content matching method based on dynamic combination features, which adopts the following technical scheme:

[0006] A content matching method based on dynamic combination features, comprising:

[0007] Real-time collection of user behavior data, content attribute data and business scenario tags, wherein the user behavior data includes user operation behavior and timestamp, the content attribute data includes live broadcast / course metadata, and the business scenario tags include preset business goals;

[0008] According to the user behavior data, the user behavior features are weighted and calculated to generate a user behavior feature vector, the weight of the content attribute data is dynamically adjusted according to the business scenario tags to generate a scene adaptation feature vector, and the user behavior feature vector and the scene adaptation feature vector are fused to generate a dynamic combination feature vector, wherein the weight of the scene adaptation feature vector is adjusted in real time according to the business goal;

[0009] Based on the dynamic combination feature vector, the matching degree of the user and the content is calculated; the content is sorted or recommended according to the matching degree, and the matching strategy is dynamically optimized through a real-time feedback mechanism;

[0010] In a live broadcast recommendation scenario, the recommendation priority is dynamically adjusted according to recent appointment behaviors of a user and a business target; in a course matching scenario, a user social relationship chain and a professional portrait corresponding to the user are called, and a course recommendation list is generated by combining the user social relationship chain and the professional portrait.

[0011] By adopting the technical solution, user behaviors, content attributes and business scenario data are collected in real time, a user behavior feature vector and a scenario adaptation feature vector are dynamically weighted and fused into a dynamic combination feature, the weight is adjusted in real time in combination with a business target (such as flow diversion and conversion), and finally a matching degree is calculated and a recommendation strategy is optimized, so that the scenario adaptability of the recommendation system is improved - in a live broadcast scenario, high-heat content is preferentially pushed, and in a course scenario, personalized recommendation is generated in combination with a social relationship chain and a professional portrait, and user interest and business demand are considered.

[0012] Optionally, in the process of collecting user behavior data, the method further includes:

[0013] User behavior data is synchronously acquired through a multi-source data collection module, and abnormal data is filtered through a data cleaning rule;

[0014] User behaviors are classified according to operation types, and time series data is generated in combination with time stamps, wherein the operation types include clicking, browsing, sharing and collecting;

[0015] The time series data is subjected to sliding window processing, and a user interest preference change trend is dynamically extracted.

[0016] By adopting the technical solution, the accuracy of behavior data is ensured through multi-source data collection and cleaning, structured time series data is constructed through operation type classification and time stamp, and a user interest change trend is dynamically captured in combination with sliding window processing, so that the precision and response speed of the recommendation system can be improved.

[0017] Optionally, in the process of generating the user behavior feature vector, the method further includes:

[0018] User behavior data is mapped into a multi-dimensional feature space, wherein the multi-dimensional feature space includes behavior frequency, behavior duration and behavior intensity;

[0019] An initial weight is assigned to each behavior type, and the weight is dynamically adjusted through a time decay function;

[0020] A user behavior pattern recognition module is introduced, periodic characteristics of user behaviors are recognized through a rule engine, and the periodic characteristics are determined as a basis for dynamic weight adjustment.

[0021] By adopting the technical scheme, the user behavior is mapped into multi-dimensional features (such as frequency and time length), the weight is dynamically adjusted, and the periodic behavior mode is identified, so that a more accurate user behavior feature vector is generated, and the personalized matching effect of the recommendation system is improved.

[0022] Optionally, in the dynamic adjustment process of the scene adaptation feature vector, the method further comprises:

[0023] According to the business scene label, the content attribute weight rule is defined, wherein the high-heat live broadcast weight is increased by 30%, and the professional matching course weight is increased by 20%;

[0024] The weight rule is dynamically adjusted in combination with the real-time business target, wherein the live broadcast with top 3 reservation numbers is preferentially pushed;

[0025] The weight parameter is updated in real time through an artificial audit interface or an automatic rule engine, and a weight adjustment log is recorded.

[0026] By adopting the technical scheme, the content attribute weight is dynamically adjusted according to the business scene, the recommendation priority is optimized in combination with the real-time target (such as flow diversion and conversion), and the weight parameter is flexibly updated through an artificial or automatic mode, so that the recommendation strategy is flexibly adapted to different scene requirements, and the matching accuracy and business adaptability are improved.

[0027] Optionally, in the fusion process of the dynamically combined feature vector, the method further comprises:

[0028] The user behavior feature vector and the scene adaptation feature vector are linearly combined according to a preset ratio, wherein the user behavior accounts for 60%, and the scene adaptation accounts for 40%; and the combined feature vector is normalized.

[0029] A multi-dimensional evaluation module is introduced, and the final dynamically combined feature vector is generated in combination with the user historical behavior preference and the scene target.

[0030] By adopting the technical scheme, the user behavior feature and the scene adaptation feature are linearly combined, and normalization processing and multi-dimensional evaluation are performed, so that the dynamically combined feature vector generated can reflect the real interest of the user and adapt to the business target, thereby improving the individualization degree of the recommendation result and the scene matching accuracy.

[0031] Optionally, in the real-time feedback mechanism, the method further comprises:

[0032] Interaction data of the user on the recommended content is collected, and the interaction data includes a click rate, a viewing time length, and a completion rate;

[0033] An A / B test module is used to compare the performance of different matching strategies, and corresponding test results are obtained;

[0034] The strategy update is automatically triggered according to the test result, and a strategy change log is recorded.

[0035] By adopting the technical solution, the user interaction data is collected, the effects of different matching strategies are evaluated in combination with A / B testing, the recommendation strategy is automatically optimized and updated, the system is ensured to continuously adapt to user behavior changes and business requirements, and the stability and adaptive ability of the recommendation effect are improved.

[0036] Optionally, in the recommendation process of the course matching scenario, the method further comprises:

[0037] The user's professional portrait information and social relationship chain are called, the professional portrait information includes a professional field and investment preferences, and the social relationship chain includes a customer manager sharing path;

[0038] The high conversion path associated with the user is identified by a social relationship chain analysis module;

[0039] The courses are matched based on the professional portrait, the recommendation priority is adjusted based on the high conversion path, and the courses are recommended.

[0040] By adopting the technical solution, the high conversion path is identified and the recommendation priority is adjusted by combining the user's professional portrait and the social relationship chain, so that the course recommendation not only meets the user's professional needs, but also improves the recommendation credibility and conversion effect by using the social relationship.

[0041] In a second aspect, the application provides a content matching method based on dynamic combination features, which adopts the following technical solution:

[0042] A content matching system based on dynamic combination features comprises:

[0043] A data acquisition module is configured to acquire user behavior data, content attribute data and business scenario labels in real time, wherein the user behavior data includes user operation behavior and a timestamp, the content attribute data includes metadata of a live broadcast / course, and the business scenario labels include preset business targets;

[0044] A dynamic combination feature vector generation module is configured to perform weighted calculation on user behavior features based on the user behavior data, generate a user behavior feature vector, dynamically adjust the weight of the content attribute data based on the business scenario labels, generate a scenario adaptation feature vector, and fuse the user behavior feature vector and the scenario adaptation feature vector to generate a dynamic combination feature vector, wherein the weight of the scenario adaptation feature vector is adjusted in real time according to the business target;

[0045] A matching strategy matching module is configured to calculate the matching degree between a user and content based on the dynamic combination feature vector, sort or recommend the content according to the matching degree, and dynamically optimize the matching strategy through a real-time feedback mechanism;

[0046] The recommendation list generation module dynamically adjusts the recommendation priority according to the recent appointment behavior of the user and the business target in a live broadcast recommendation scene; in a course matching scene, the user social relationship chain and the professional portrait corresponding to the user are called, and the user social relationship chain and the professional portrait are combined to generate a course recommendation list.

[0047] In a third aspect, the application provides a content matching system based on dynamically combined features, which adopts the following technical solution:

[0048] A content matching system based on dynamically combined features, comprising a processor, wherein the processor runs a program of the content matching method based on dynamically combined features according to any one of the preceding embodiments.

[0049] In a fourth aspect, the application provides a storage medium, which adopts the following technical solution:

[0050] A storage medium storing a program of the content matching method based on dynamically combined features according to any one of the preceding embodiments.

[0051] In summary, the application has at least one of the following beneficial technical effects:

[0052] By collecting user behavior, content attributes and business scene data in real time, user behavior features and scene adaptation features are dynamically generated and fused, and the matching degree of the user and the content is accurately calculated. In combination with live broadcast and course recommendation scenes, user appointment behavior, social relationship chain and professional portrait are used to realize personalized recommendation and improve the business adaptation and user satisfaction of the recommendation result.

[0053] Through time series modeling, dynamic weighting, multi-dimensional feature fusion and A / B test feedback mechanism, it is ensured that the recommendation system can quickly respond to changes in user interest and different business targets. In specific scenes, the recommendation priority is further optimized in combination with social communication path and professional matching logic, thereby effectively improving the accuracy, flexibility and scene adaptation ability of the recommendation system. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 is a flowchart of a content matching method based on dynamically combined features according to an exemplary embodiment.

[0055] Figure 2 is a structural block diagram of a content matching system based on dynamically combined features according to an exemplary embodiment. DETAILED DESCRIPTION

[0056] The embodiments of the application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0057] In the description of the specification, the description of the terms "certain embodiments", "one embodiment", "some embodiments", "illustrative embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the described embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in one or more embodiments or examples.

[0058] The embodiment of the present application discloses a content matching method based on dynamic combination features, referring to Figure 1 , comprising:

[0059] S100, real-time collection of user behavior data, content attribute data and business scene label, wherein the user behavior data includes user operation behavior and timestamp, the content attribute data includes live / course metadata, and the business scene label includes a preset business target;

[0060] In the embodiment of the present application, in the process of collecting user behavior data, the method further comprises:

[0061] The user behavior data, content attribute data and business scene label are synchronously acquired through a multi-source data acquisition module. The user behavior data includes user operation behavior (such as clicking, browsing, sharing, and collecting) and its occurrence timestamp; the content attribute data covers live or course metadata (such as title, classification, and heat index); and the business scene label contains a preset business target (such as "draining" and "customer conversion"). This process is realized through APP end embedding, server log, third-party API and other modules, ensuring the comprehensiveness and real-time nature of the data source.

[0062] After acquiring the original data, S100 filters abnormal data through data cleaning rules. For example, invalid clicks (such as accidental touch operations), repeated records (such as repeated collection of the same course) or abnormal timestamps (such as future timestamps) are removed, ensuring the data quality for subsequent analysis.

[0063] The user behavior is classified according to operation types (clicking, browsing, sharing, and collecting), and structured time series data is generated in combination with the timestamp. For example, user A clicks live X 3 times, browses course Y for 5 minutes, and shares course Z once within 1 hour. These behaviors are mapped as time series, facilitating subsequent analysis

[0064] The time series data is analyzed by using a sliding window (such as the behavior in the last 7 days) to identify the change trend of the user interest. For example, if user B frequently books a live broadcast of “macroeconomic analysis” recently, but reduces the attention to the “investment foundation” course, the sliding window processing can capture the interest deviation.

[0065] By synchronously collecting user behavior, content attribute and business scenario data from multiple sources, filtering abnormal information by combining data cleaning rules, classifying and converting the user behavior into structured time series, and dynamically capturing the interest change trend by using a sliding window, high-quality and real-time input is provided for the subsequent steps, thereby guaranteeing the data reliability.

[0066] In S200, the user behavior feature is weighted and calculated according to the user behavior data, to generate a user behavior feature vector. The weight of the content attribute data is dynamically adjusted according to the business scenario label, to generate a scenario adaptation feature vector. The user behavior feature vector and the scenario adaptation feature vector are fused to generate a dynamically combined feature vector. The weight of the scenario adaptation feature vector is adjusted in real time according to the business target.

[0067] In the embodiments of the present application, S200 includes the following sub-steps:

[0068] In step 1, based on the user behavior data (such as time series and interest trend), S200 first performs weighted calculation on the user behavior feature:

[0069] Behavior type weighting: According to the operation type (click, browse, share, collect), an initial weight is assigned (such as a click weight of 0.8 and a browse weight of 0.5), to reflect the influence degree of different behaviors on the user interest.

[0070] Time decay processing: In combination with the time stamp, a time decay function (such as an exponential decay function) is applied to the historical behavior, so that the weight of the recent behavior is higher. For example, the weight of the booking behavior of the user last week is 50% of the current behavior.

[0071] Multi-dimensional feature mapping: The weighted behavior data is mapped to a multi-dimensional feature space (such as behavior frequency, behavior duration and behavior intensity), to form a user behavior feature vector.

[0072] In step 2, according to the business scenario label (such as “draining” and “customer conversion”), S200 dynamically adjusts the weight of the content attribute data:

[0073] For different business targets, the weight rules are defined. For example, in the draining scenario, the weight of the high-heat live broadcast is increased by 30%, and the top 3 live broadcasts are preferentially pushed. In the customer conversion scenario, the weight of the professional matching course is increased by 20%, and the related courses are recommended in combination with the user professional portrait.

[0074] Real-time parameter adjustment: dynamically update weight parameters according to real-time business needs (such as 1 hour before the start of the show) through manual review interface or automatic rule engine, and record adjustment log.

[0075] Real-time parameter adjustment: dynamically update weight parameters according to real-time business needs (such as 1 hour before the start of the show) through manual review interface or automatic rule engine, and record adjustment log.

[0076] 3. Fusion of user behavior feature vector and scene adaptation feature vector:

[0077] Linear combination: weighted sum according to preset proportion (such as user behavior accounting for 60% and scene adaptation accounting for 40%), balance user interest and business target;

[0078] Normalization processing: normalize the fused feature vector (such as Min-Max standardization), eliminate the dimensional difference of different dimensions, and ensure the stability of calculation;

[0079] Multi-dimensional evaluation: combine user historical behavior preference (such as "macroeconomic" live broadcast preference degree) and scene target (such as "customer conversion" scene high single price course), generate final dynamic combination feature vector.

[0080] Through weighted user behavior (such as click, browse), dynamic adjustment of scene weight (such as priority push of high heat live broadcast in flow scene) and fusion of multi-dimensional features (user interest and business target), generate dynamic combination feature vector, provide accurate input for core matching logic.

[0081] In the embodiment of the application, in the generation process of the user behavior feature vector, the method further comprises:

[0082] 1. Multi-dimensional feature space mapping: through mapping user behavior data (such as click, browse, share) into multi-dimensional feature space containing behavior frequency (operation times per unit time), behavior duration (single operation duration) and behavior intensity (such as sharing, collecting and other high value behaviors), realize comprehensive quantification of user behavior. For example, user A clicks live broadcast X 5 times in 3 days (frequency = 5 / 3), watches live broadcast X for 20 minutes (duration = 20), and shares live broadcast X 2 times (intensity = 2). Through the introduction of behavior intensity, the value difference of different behaviors to user interest (such as sharing > click > browse) can be distinguished, so as to improve the ability of feature vector to describe user's real interest, and avoid one-sidedness of single dimension.

[0083] 2. Initial weight and time decay dynamic adjustment: When assigning initial weights, set different weights for different behavior types according to business experience or historical data (such as click = 0.8, browse = 0.5, share = 1.2) to reflect their importance to user interest. At the same time, apply time decay rules to historical behavior combined with timestamps, so that the weight of recent behavior is higher (such as the weight of the user's recent appointment live is higher than the historical appointment), so as to dynamically respond to the short-term fluctuations of user interest (such as from "investment foundation" to "high-level strategy").

[0084] 3. User behavior pattern recognition module: Analyze the periodic characteristics of user behavior (such as appointment live on Wednesday every week) through a rule engine (such as time series clustering algorithm), and use the identified rules (such as "Wednesday active") as the basis for dynamic weight adjustment. For example, during the recommendation period on Wednesday, the user's weight for "macroeconomic" live is increased by 20%, thereby optimizing the recommendation priority. This module improves the prediction ability of the feature vector by mining potential behavior patterns (such as active during weekdays), and dynamically adjusts the weight in combination with periodic characteristics, so that the recommendation is more in line with user behavior habits (such as pushing high-heat live on Wednesday).

[0085] In the embodiment of the present application, during the dynamic adjustment process of the scene adaptation feature vector, the method further comprises:

[0086] 1. Define content attribute weight rules according to business scenario labels: In the generation process of the scene adaptation feature vector, first define the weight rules of the content attributes according to the pre-set business scenario labels (such as "driving" and "customer conversion"). For example, in the driving scenario, the weight of high-heat live will be increased by 30% to preferentially push live with high appointment number or viewing volume; in the customer conversion scenario, the weight of professional matching courses will be increased by 20% to match the user's professional field (such as "corporate legal person" recommending "hedge" courses). These rules are determined through pre-set business logic or historical data statistics, for example, based on the conversion effect of high-heat live in the past driving activities, or the matching degree analysis of professional portrait and course classification, so as to ensure that the recommended content is directly related to the scene goal.

[0087] 2. Dynamically adjust weight rules based on real-time business targets: Based on preset rules, the system dynamically adjusts weight rules according to real-time business targets (such as "1 hour before the start of the show to attract traffic" and "quarter-end course conversion"). For example, in the live streaming traffic attraction scenario, if the number of reservations for a certain live streaming suddenly increases, the system will temporarily increase its weight from 30% to 50%, so that it will be displayed preferentially in the recommendation; in the course conversion scenario, if the real-time reservation data of a certain type of course (such as "finance") is significantly higher than that of other types, the system will dynamically increase its weight to enhance the recommendation effect. The adjustment rule is triggered by real-time business indicators (such as reservation number and conversion rate), for example, when the reservation number of a certain live streaming exceeds the threshold, the priority push logic is automatically activated, thereby ensuring that the recommendation strategy can quickly respond to business changes.

[0088] 3. Update weight parameters in two ways: manual review interface and automatic rule engine; In the manual review interface, the operator can manually adjust the weight according to the business needs (such as temporarily increasing the weight of a certain course), and submit changes through the approval process; In the automatic rule engine, the system will automatically trigger weight adjustment according to the preset conditions (such as reservation number exceeding threshold) without human intervention. Each adjustment will record the weight parameter change log, including time, operator, pre / post-adjustment value, etc. for subsequent backtracking analysis. For example, if a manual adjustment leads to a decline in recommendation effect, the log can be used to quickly locate the problem and restore the historical parameters, thereby ensuring the controllability and traceability of strategy adjustment.

[0089] In the fusion process of dynamically combining feature vectors, the method further comprises:

[0090] 1. In the fusion process of dynamically combining feature vectors, the system first weights and sums the user behavior feature vector and the scene adaptation feature vector according to a preset proportion. For example, the weight proportion of the user behavior feature vector is 60%, and the weight proportion of the scene adaptation feature vector is 40%. This proportion allocation is determined based on business experience or historical data optimization to ensure a balance between user interest (such as click, browse, etc.) and business target (such as traffic attraction, conversion, etc.). For example, in the traffic attraction scenario, if high-heat live streaming needs to be preferentially pushed, the system can dynamically adjust the proportion and increase the scene adaptation weight to 50% to enhance business adaptability.

[0091] 2. The combined feature vector needs to be normalized to eliminate the range difference of different dimensions. For example, some feature values may be distributed in the interval of 0-100, while other feature values may be distributed in the interval of 0-1. Through normalization (such as mapping all feature values to the uniform interval of 0-1), it is avoided that some features occupy an unreasonable dominant position in the matching calculation due to the range difference of the values. For example, if a feature value is originally 100, it may become 1 after normalization, and other feature values are scaled proportionally, thereby ensuring the fairness and stability of the matching algorithm.

[0092] 3. Introduce a multi-dimensional evaluation module that combines user historical behavior preferences (such as long-term interest in "macroeconomic" live broadcasts) with current scenario goals (such as "customer conversion" scenarios that require priority for professional matching courses), and perform secondary calibration on the dynamically combined feature vector. For example:

[0093] Historical preference reinforcement: If a user frequently books "financial analysis" live broadcasts in the past, the module will dynamically increase the weight of this category to ensure that the recommended content matches their long-term interests;

[0094] Scenario goal adaptation: In the course conversion scenario, if the user's profession is "corporate officer", the module will enhance the weight of the "hedge" course, even if its matching degree is slightly lower than other courses, it will still be displayed to meet business goals.

[0095] S300, calculate the matching degree between the user and the content based on the dynamically combined feature vector; sort or recommend the content according to the matching degree, and dynamically optimize the matching strategy through real-time feedback mechanism.

[0096] In the embodiments of the present application, S300 includes the following sub-steps:

[0097] 1. Calculate the matching degree based on the dynamically combined feature vector: Based on the dynamically combined feature vector (fusing user behavior features and scenario adaptation features), the matching degree between the user and the content is calculated through algorithm. For example, the system compares the user's interest features (such as recent booking behavior, professional preference) and the attribute features of the live broadcast (such as popularity, category, time arrangement) to determine whether they are compatible. The higher the matching degree, the stronger the relevance between the user and the content. In addition, the system will further optimize the matching result by combining user historical preferences (such as past clicked course types) and current business goals (such as priority for high-heat live broadcasts in the lead-in scenario), to ensure that the recommendations are both in line with user interests and adapt to business needs.

[0098] 2. After calculating the matching degree, the system sorts or recommends the content according to the matching degree:

[0099] Sorting logic: content with high matching degree is displayed first. For example, if user A's matching degree for "macroeconomic analysis" live broadcast is 90%, and for "investment foundation" live broadcast is 60%, the former will be displayed first in the recommendation list. At the same time, the system will adjust the final sorting according to business rules (such as higher priority for high-heat live broadcasts in the lead-in scenario).

[0100] Recommendation logic: In the course matching scenario, if user B's professional field is "corporate officer", the system will preferentially recommend courses that match their profession (such as "hedge"), even if their matching degree is slightly lower than other courses, but because of higher professional adaptability, they will still be displayed first.

[0101] 3. The system continuously optimizes the matching strategy through a real-time feedback mechanism:

[0102] Data collection: Collect user interaction data on recommended content in real time, such as click rate (whether the user clicks on the recommendation), viewing time (whether the user continues to watch the live broadcast), and completion rate (whether the user completes the course learning).

[0103] A / B testing: Compare the effects of different matching strategies. For example, strategy A may be based on user interest ranking, and strategy B may combine professional image and social relationship chain. Through test data (such as click rate difference), determine the better solution.

[0104] Strategy update: Automatically adjust the matching strategy according to the test results. For example, if the recommendation strategy of a certain live broadcast leads to a decrease in user click rate, the system will switch to other strategies (such as increasing the weight of high-heat live broadcasts), and record the change log for subsequent analysis.

[0105] In the real-time feedback mechanism in the embodiments of the present application, the method further comprises:

[0106] 1. The real-time feedback mechanism first collects user interaction data on recommended content in real time through the burying point technology, including click rate (whether the user clicks on the recommendation), viewing time (whether the user continues to watch the live broadcast), and completion rate (whether the user completes the course learning) and other key indicators. For example, the click rate counts the proportion of the number of times the user clicks on a specific live broadcast or course in the recommendation list; the viewing time records the actual viewing time of the user (such as an average of 30 minutes); and the completion rate calculates the proportion of users who complete the course learning (such as 70 out of 100 users).

[0107] 2. Compare the effects of different matching strategies through the A / B testing module. For example, strategy A is based on user interest ranking (such as preferentially pushing the "macroeconomic" live broadcast that the user has long been interested in), and strategy B combines scene goals and social relationship chain ranking (such as preferentially pushing high-heat live broadcasts in the lead-in scene). A / B testing randomly divides users into experimental and control groups, applies different strategies to each group, and calculates the click rate, viewing time, and completion rate of each group. For example, if the click rate of strategy A is 15% and the click rate of strategy B is 10%, then strategy A is better. Through data comparison, the system can scientifically verify the effect of the strategy and avoid subjective decision bias.

[0108] 3. According to the results of A / B testing, the system automatically triggers the update of the matching strategy. For example, if the recommendation strategy of a certain live broadcast leads to a decrease in user click rate, the system will switch to other strategies (such as increasing the weight of high-heat live broadcasts); if the completion rate of a certain type of course is significantly higher than that of other types, the system will dynamically adjust the weight of course classification. Each strategy update records the change log (including time, pre / post-adjustment parameters, operator, etc.) for subsequent analysis and backtracking. For example, if a certain strategy adjustment leads to a decrease in user engagement, the log can be used to quickly locate the problem and restore historical parameters.

[0109] S400, in the live broadcast recommendation scenario, the recommendation priority is dynamically adjusted according to the user's recent reservation behavior and business target; in the course matching scenario, the user's social relationship chain and professional portrait are called, and a course recommendation list is generated by combining the user's social relationship chain and professional portrait.

[0110] In the embodiments of the present application, S400 includes the following sub-steps:

[0111] 1. In the live broadcast recommendation scenario, the system dynamically adjusts the recommendation priority according to the user's recent reservation behavior and the current business target. For example:

[0112] User's recent reservation behavior: the system analyzes the user's reservation records in the past week (such as reserving live broadcasts of “macroeconomic analysis” and “investment basics”), and judges the user's interest preferences. If the user frequently reserves a certain type of topic (such as “financial policy interpretation”), the system will preferentially recommend related content live broadcasts.

[0113] Business target adaptation: adjust the recommendation strategy according to the current business needs (such as “draining” and “conversion”). For example, in the drainage stage, preferentially push high-heat live broadcasts (such as the top 3 live broadcasts in terms of reservation number), and in the conversion stage, focus on recommending live broadcasts that the user has reserved but not watched, in order to improve the completion rate.

[0114] Dynamic adjustment mechanism: the system monitors user behavior changes in real time (such as shifting from “investment basics” to “advanced strategies”), and adjusts the recommendation priority in combination with the business target. For example, if the user has recently reserved multiple live broadcasts on the topic of “hedge”, and the current business target is “customer conversion”, the system will increase the weight of live broadcasts in this field.

[0115] 2. In the course matching scenario, the system calls the user's social relationship chain and professional portrait, and generates a personalized course recommendation list by combining the two. For example:

[0116] Social relationship chain analysis: The system analyzes the learning behaviors of the user's friends, colleagues, or peers (e.g., friend A booked a "Hedging" course, and friend B completed a "Financial Derivatives" course) and generates recommendations based on the popular courses in the social circle. For example, if most people in the user's social circle have booked a "Risk Management" course, the system will prioritize related content for recommendation.

[0117] Professional portrait adaptation: Courses are matched according to the user's professional field (e.g., "Corporate Officer" or "Financial Analyst"). For example, a corporate officer may be recommended "Hedging" and "Tax Planning" courses, while a financial analyst may be recommended "Quantitative Investment" and "Market Trend Analysis" courses.

[0118] Dynamic generation of recommendation list: The system combines the popular courses in the social relationship chain and the courses matched by the professional portrait to generate the final recommendation list. For example, if the user's profession is "Corporate Officer" and many people in their social circle have booked a "Hedging" course, the system will place this course at the top of the recommendation list.

[0119] By dynamically adjusting the priority of live broadcast recommendations (based on the user's recent behavior and business goals) and matching courses with the social relationship chain and professional portrait, the recommendation strategy is made more accurate and scenario-based.

[0120] In the embodiments of the present application, in the recommendation process of the course matching scenario, the method further comprises:

[0121] 1. In the course matching scenario, the system first retrieves the user's professional portrait information and social relationship chain data. The professional portrait information includes the user's professional field (e.g., "Corporate Officer" or "Financial Analyst") and investment preference (e.g., "Conservative" or "Aggressive"), which is obtained through user registration information, historical behavior records, and interaction data. The social relationship chain includes the user's interaction path with the account manager (e.g., Account Manager A recommended a "Hedging" course, and Account Manager B recommended a "Tax Planning" course), as well as the learning behavior data of friends and colleagues (e.g., Friend A booked and completed a certain course).

[0122] 2. Through the social relationship chain analysis module, the user's associated high-conversion path (i.e., the course recommendation path that the user is more likely to accept and complete) is identified. For example, if the "Hedging" course recommended by a certain account manager has a completion rate of 80% in the user's social circle, this path is marked as a high-conversion path; if Friend A booked and completed a "Financial Derivatives" course, while Friend B booked but did not complete a "Risk Management" course, the recommendation priority of the "Financial Derivatives" course is higher. By analyzing these paths, the system can determine which courses are more suitable for the user's social circle learning habits and the account manager's recommendation effectiveness.

[0123] 3. The system integrates professional image and high conversion path information to dynamically adjust the priority of course recommendations. For example, according to the user's professional field, match the course classification (such as "corporate officer" matches "hedge" course, "financial analyst" matches "quantitative investment" course), and give higher weight to high conversion path courses (such as "hedge" course recommended by customer manager) in the matching results. If the user's occupation is "corporate officer", and the "hedge" course recommended by the customer manager has a high completion rate in the social circle, this course will be displayed in the recommendation list in priority. For example, the system will place the "hedge" course at the top of the recommendation list, while reducing the weight of other low conversion path courses, ensuring that the recommended content not only meets the user's professional needs, but also meets the learning habits of the social circle.

[0124] Parent user C recently booked multiple live broadcasts on the education platform on the topic of "children's learning methods" and collected the "mathematical thinking training" course. The system collects its behavior data (such as clicks, reservations), content attributes (such as course classification, popularity), and business scenario tags (such as "knowledge expansion" and "exam preparation") through S100, and generates structured time series after cleaning abnormal data. For example, user C booked 3 "learning method" live broadcasts in 7 days, but did not complete the "mathematical thinking training" course. At this time, S200 dynamically adjusts the weight according to the business scenario: in the "knowledge expansion" stage, the system increases the weight of high-heat live broadcasts (such as the top 3 "learning method" live broadcasts) by 30%, generating a scenario-adapted feature vector; at the same time, the user behavior feature vector is dominant (60% weight) due to recent high-frequency reservations for "learning methods", and the dynamically combined feature vector is preferentially pushed after fusion.

[0125] When user C enters the "exam preparation" stage, the system calculates the matching degree based on the dynamically combined feature vector through S300. For example, user C's professional image shows that he is a "primary school parent" who needs to match "math score improvement" related courses, and the "mathematical thinking training" course recommended by other parents in the social relationship chain has a 80% completion rate in the user's friend circle (high conversion path). At this time, S400 dynamically adjusts the recommendation priority by combining the professional image (primary school parents need math tutoring courses) and the social path: increase the weight of the "mathematical thinking training" course by 20%, and preferentially display it in the recommendation list. At the same time, the real-time feedback mechanism verifies the strategy effect through A / B testing, and if the user click rate increases by 15% compared with historical data, the system automatically solidifies the strategy and records logs for subsequent optimization.

[0126] In the above case, the system improves the scene adaptability through multi-source data collection (S100) and dynamic weight adjustment (S200): in the "knowledge expansion" stage, high-heat live broadcasts are preferentially pushed to attract the attention of parents; in the "exam tutoring" stage, the user demand is accurately matched in combination with the professional portrait (primary school parents need mathematics tutoring) and the high-conversion path of social interaction (the completion rate of the "mathematical thinking training" course recommended by other parents is high). In addition, the real-time feedback mechanism (S300) continuously optimizes the strategy through indicators such as click rate and course completion rate, ensuring that the recommended content not only meets the interests of parents but also serves the educational goals. Finally, the system flexibly switches strategies in different scenes, increasing the conversion rate of user C from "focusing on learning methods" to "completing the mathematics score improvement course" by 25%, verifying the direct driving effect of scene adaptability on user retention and satisfaction of education platforms.

[0127] The embodiment of the application discloses a content matching system based on dynamic combined features, referring to Figure 2 , comprising:

[0128] The data collection module 001 is used for collecting user behavior data, content attribute data and business scene labels in real time, wherein the user behavior data includes user operation behavior and time stamp, the content attribute data includes live broadcast / course metadata, and the business scene label includes a preset business target;

[0129] The dynamic combined feature vector generation module 002 is used for performing weighted calculation on user behavior features according to the user behavior data, generating a user behavior feature vector, dynamically adjusting the weight of the content attribute data according to the business scene label, generating a scene adaptation feature vector, and fusing the user behavior feature vector and the scene adaptation feature vector to generate a dynamic combined feature vector, wherein the weight of the scene adaptation feature vector is adjusted in real time according to the business target;

[0130] The matching strategy matching module 003 is used for calculating the matching degree of the user and the content based on the dynamic combined feature vector, sorting or recommending the content according to the matching degree, and dynamically optimizing the matching strategy through a real-time feedback mechanism;

[0131] The recommendation list generation module 004 is used for dynamically adjusting the recommendation priority according to the recent appointment behavior of the user and the business target in a live broadcast recommendation scene, and for calling the user social relationship chain and the professional portrait corresponding to the user, and generating a course recommendation list in combination with the user social relationship chain and the professional portrait in a course matching scene.

[0132] The embodiment of the application also discloses a content matching system based on dynamic combined features, comprising a processor, and a program of the content matching method based on dynamic combined features in any one of the above embodiments is run in the processor.

[0133] The embodiment of the present application further discloses a storage medium, which stores the program of the content matching method based on the dynamic combination features.

[0134] Although the embodiments of the present application have been shown and described above, it should be understood by those skilled in the art that the above embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A content matching method based on dynamic combination features, characterized in that, include: Real-time collection of user behavior data, content attribute data, and business scenario tags. User behavior data includes user actions and timestamps, content attribute data includes metadata for live streams / courses, and business scenario tags include preset business objectives. Based on user behavior data, user behavior features are weighted and calculated to generate user behavior feature vectors. The weights of content attribute data are dynamically adjusted according to business scenario tags to generate scenario-adaptive feature vectors. The user behavior feature vectors and scenario-adaptive feature vectors are then fused to generate dynamically combined feature vectors. The weights of the scenario-adaptive feature vectors are adjusted in real time according to business objectives. The matching degree between users and content is calculated based on the dynamic combination of feature vectors; the content is sorted or recommended according to the matching degree, and the matching strategy is dynamically optimized through a real-time feedback mechanism. In live streaming recommendation scenarios, recommendation priorities are dynamically adjusted based on users' recent booking behavior and business goals; in course matching scenarios, users' social relationship chains and professional profiles are retrieved, and a course recommendation list is generated by combining the user's social relationship chains and professional profiles.

2. The content matching method based on dynamic combination features according to claim 1, characterized in that, The methods used in collecting user behavior data also include: User behavior data is acquired synchronously through a multi-source data acquisition module, and abnormal data is filtered out through data cleaning rules. User behaviors are categorized by operation type and combined with timestamps to generate time series data. The operation types include click, browse, share, and favorite. A sliding window process is used to dynamically extract trends in user interest preferences by processing time-series data.

3. The content matching method based on dynamic combination features according to claim 1, characterized in that, The method further includes the following steps in generating the user behavior feature vector: User behavior data is mapped into a multidimensional feature space, which includes behavior frequency, behavior duration, and behavior intensity. An initial weight is assigned to each behavior type, and the weight is dynamically adjusted using a time decay function. A user behavior pattern recognition module is introduced to identify the periodic characteristics of user behavior through a rule engine, and the periodic characteristics are used as the basis for dynamic weight adjustment.

4. The content matching method based on dynamic combination features according to claim 1, characterized in that, During the dynamic adjustment of the scene adaptation feature vector, the method further includes: Content attribute weighting rules are defined based on business scenario tags, with high-popularity live streams receiving a 30% weight increase and career matching courses receiving a 20% weight increase. The weighting rules are dynamically adjusted based on real-time business objectives, with priority given to pushing the top 3 live streams in terms of reservation numbers. The weight parameters are updated in real time through a manual review interface or an automated rule engine, and weight adjustment logs are recorded.

5. The content matching method based on dynamic combination features according to claim 1, characterized in that, The method further includes the following steps in the fusion process of the dynamically combined feature vectors: The user behavior feature vector and the scene adaptation feature vector are linearly combined according to a preset ratio, where user behavior accounts for 60% and scene adaptation accounts for 40%; the combined feature vector is then normalized. A multi-dimensional evaluation module is introduced to generate a final dynamic combination feature vector by combining users' historical behavioral preferences with scenario objectives.

6. The content matching method based on dynamic combination features according to claim 1, characterized in that, The real-time feedback mechanism further includes the following methods: Collect user interaction data on recommended content, including click-through rate, viewing time, and completion rate; The performance of different matching strategies is compared using the A / B testing module to obtain the corresponding test results. The policy is automatically updated based on the test results, and the policy change log is recorded.

7. The content matching method based on dynamic combination features according to claim 1, characterized in that, In the course matching scenario recommendation process, the method further includes: Access user professional profile information and social relationship chain. Professional profile information includes professional field and investment preferences, and social relationship chain includes the account manager sharing path. Identify high-conversion paths associated with users through the social relationship chain analysis module; The course categories are matched with professional profiles, and the recommendation priority is adjusted based on the high conversion paths.

8. A content matching system based on dynamic combination features, characterized in that, include: The data acquisition module is used to collect user behavior data, content attribute data, and business scenario tags in real time. The user behavior data includes user operation behavior and timestamps, the content attribute data includes metadata of live broadcasts / courses, and the business scenario tags include preset business objectives. The dynamic feature vector generation module calculates user behavior features by weighting them based on user behavior data to generate user behavior feature vectors. It dynamically adjusts the weights of content attribute data based on business scenario tags to generate scenario-adaptive feature vectors. The user behavior feature vectors and scenario-adaptive feature vectors are then fused to generate dynamic feature vectors. The weights of the scenario-adaptive feature vectors are adjusted in real time according to business objectives. The matching strategy matching module calculates the matching degree between users and content based on dynamically combined feature vectors; sorts or recommends content according to the matching degree, and dynamically optimizes the matching strategy through a real-time feedback mechanism. The recommendation list generation module dynamically adjusts recommendation priorities based on users' recent booking behavior and business goals in live streaming recommendation scenarios; in course matching scenarios, it retrieves users' social relationship chains and professional profiles, and combines these to generate a course recommendation list.

9. A content matching system based on dynamic combination features, characterized in that, Includes a processor, wherein the processor runs a program of the content matching method based on dynamic combination features as described in any one of claims 1-7.

10. A storage medium, characterized in that, The program stores the content matching method based on dynamic combination features as described in any one of claims 1-7.

Images