Intelligent media content publishing system

Abstract

The invention relates to the technical field of electronic digital data processing, in particular to an intelligent media content publishing system, which is characterized in that a user tag analysis module classifies new users and distributes user tags based on initial user information; the short video set generation module generates a short video set according to the user tag; the video pushing module sends videos in the short video set to a new user and plays the videos in sequence; the user feedback acquisition module acquires feedback information of the user after the user watches the first N videos; the integral calculation module generates an integral combination of each short video according to feedback information of all new users of the same label; the preference degree calculation module calculates the preference degree of the first N videos based on the integral combination and predicts the prediction preference degree of the videos which are not played yet; and the video sorting optimization module is used for adjusting the unplayed videos based on the prediction preference degree so as to carry out collective pushing on the short videos after analysis is carried out in combination with the common behavior data of all the new users, thereby improving the content publishing efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic digital data processing, and particularly relates to an intelligent media content publishing system. BACKGROUND

[0002] Media content publishing refers to the process of disseminating various media materials, such as articles, pictures, audio, video, etc., to the public through media platforms.

[0003] Short video collection type pushing is an intelligent content distribution mode based on user interest and behavior data. The platform analyzes the user's viewing history, interaction behavior and preferences through algorithms, integrates multiple related short videos into a theme or a scene-based collection, and makes a centralized recommendation. This way not only improves the coherence and immersion of the content, but also improves the user's viewing efficiency and experience.

[0004] When pushing media content to new users, due to insufficient data in the early stage, it takes a long time to collect data of target users before a relatively stable push can be formed, thereby reducing the efficiency of content publishing. SUMMARY

[0005] The purpose of the present application is to provide an intelligent media content publishing system that can analyze the behavior data of all new users in a certain period of time and then push short videos in a collection type, thereby improving the efficiency of content publishing.

[0006] To achieve the above purpose, the present application provides an intelligent media content publishing system, which comprises a user tag analysis module, a short video set generation module, a video pushing module, a user feedback collection module, an integral calculation module, a preference degree calculation module and a video sorting optimization module.

[0007] The user tag analysis module is used for classifying new users and assigning user tags based on initial user information.

[0008] The short video set generation module is used for generating short video sets according to the user tags.

[0009] The video pushing module is used for sending videos in the short video set to new users and playing them in order.

[0010] The user feedback collection module is used for collecting feedback information of users after they watch the first N videos, which includes the completion rate, likes, collections, sharing operations and dwell time.

[0011] The integral calculation module is used for generating integral combinations of each short video according to the feedback information of all new users with the same tag, which includes completion integral, behavior activity integral and social interaction integral.

[0012] The preference degree calculation module is configured to calculate the preference degrees of the first N videos based on the integral combination and predict the predicted preference degrees of the videos that have not been played.

[0013] The video ranking optimization module is configured to adjust the videos that have not been played based on the predicted preference degrees, and the adjustment includes deleting short videos that are lower than a preference degree threshold and increasing short videos of the same type as the short videos that are higher than the preference degree threshold.

[0014] The user label analysis module includes a user information acquisition unit, a data processing unit, and a matching unit.

[0015] The user information acquisition unit is configured to receive initial registration information of a new user, and the initial registration information includes identity data and selection preference data.

[0016] The data processing unit is configured to perform cleaning and standardization processing on the initial registration information.

[0017] The matching unit is configured to perform multi-dimensional matching on the cleaned user data according to a preset label system and a rule engine to generate a preliminary user label set.

[0018] The short video set generation module includes a label acquisition unit, a short video screening unit, a sorting unit, and an integration unit.

[0019] The label acquisition unit is configured to acquire a user label set of a user.

[0020] The short video screening unit is configured to screen short video resources that match the user label from a short video content library according to a preset label matching strategy.

[0021] The sorting unit is configured to perform sorting processing on the screened short video resources based on matching degree weights.

[0022] The integration unit is configured to generate a target short video set based on the sorting result.

[0023] The sorting unit includes a weighted matching degree calculation subunit, an arrangement subunit, a sorting correction subunit, and a selection subunit.

[0024] The weighted matching degree calculation subunit is configured to calculate a weighted matching degree between each short video resource in a candidate short video set and a user label.

[0025] The arrangement subunit is configured to arrange the candidate short video resources in descending order according to the weighted matching degrees.

[0026] The sorting correction subunit is configured to correct the descending arrangement based on video heat.

[0027] The selection unit is configured to intercept the first N items of the ranked short video resources according to a preset number.

[0028] The user feedback collection module includes a data collection unit, a data structuring unit, and a storage unit.

[0029] The data collection unit is configured to collect interactive behavior data of a user during the process of watching the first N videos when the number of videos watched by the user accumulatively reaches a preset value N, the interactive behavior data including a complete play status and a complete play rate of each video, whether the user performs a like operation on the video, whether the user adds the video to a collection list, whether the user performs a sharing operation on the video, and a stay duration of the user on each video.

[0030] The data structuring unit is configured to clean and structure the collected interactive behavior data to generate standardized user feedback records.

[0031] The storage unit is configured to store the user feedback records in a user behavior database and mark the records as association data between corresponding video content and a user ID.

[0032] The integral calculation module includes a feedback information summarizing unit and an integral calculation unit.

[0033] The feedback information summarizing unit is configured to obtain user feedback records of all new users under the same user label category.

[0034] The integral calculation unit is configured to map the user feedback record data to a corresponding integral model and quantitatively process each integral after mapping; the integral model includes a complete play integral triggered if a user completes video playback, an interactive integral triggered if the user likes, collects, or shares a video, and an active degree integral triggered if the user stays on a video page for more than a set threshold.

[0035] The preference degree calculation module includes an interest preference degree calculation unit, a matching model generating unit, and a prediction unit.

[0036] The interest preference degree calculation unit is configured to calculate the interest preference degree of a user for the first N short video resources currently displayed based on the content preference integral generated based on the feedback information of the user.

[0037] The matching model generating unit is configured to construct a matching model between a played video content feature vector and an unplayed video content feature vector.

[0038] The prediction unit is configured to match candidate short video resources that have not been played using the matching model and combine the interest preference degree to obtain a corresponding predicted preference degree.

[0039] The specific steps of constructing the matching model between the played video content feature vector and the unplayed video content feature vector include: extracting the content feature vector of each played video, the content feature vector including but not limited to video tags, categories, creator attributes, publishing time, heat indicators, and video duration; extracting the video content feature vector of each unplayed candidate video; and constructing a matching model based on content similarity.

[0040] The video sorting optimization module includes a preference degree acquisition unit, a video screening unit, a video matching unit, and a video adding unit.

[0041] The preference degree acquisition unit is configured to acquire a predicted preference degree of a user for a candidate unplayed short video, the predicted preference degree being calculated by the matching model according to the historical behavior of the user and the video content features.

[0042] The video screening unit is configured to perform a deletion operation on a short video whose predicted preference degree is lower than a preference degree threshold, so as to eliminate low-relevance content in the recommendation result.

[0043] The video matching unit is configured to extract the content features of a short video whose predicted preference degree is higher than the preference degree threshold, and find similar short videos in the candidate library.

[0044] The video adding unit is configured to add the found similar short videos to the short video set, the number of the similar short videos being the same as the number of the deleted short videos.

[0045] The specific steps of deleting the short videos lower than the preference degree threshold include: traversing each video in the candidate short video set; comparing the predicted preference degree of the video with a preset minimum preference degree threshold; and removing the video from the candidate set if the preference degree of the video is lower than the threshold.

[0046] The intelligent media content publishing system of the application, the user tag analysis module is used for preliminary classification of new users, and based on the initial user information (such as age, gender, geographical location, interest, etc.) provided when registering, the corresponding user tags are assigned. These tags constitute the initial image of the user, providing a basis for subsequent content recommendation. The short video set generation module selects short videos matching the potential interests of users from a large number of video resources based on the user tag analysis results, and generates a preliminary short video set. This set contains video content of various types or themes suitable for this group of users to watch, ensuring content diversity while meeting individual needs. The video push module pushes the videos in the above short video set to new users and plays them in a set order. This module supports functions such as automatic playback and loop playback, ensuring that users can smoothly watch multiple videos and collect enough behavior data for subsequent analysis. The user feedback collection module starts after the user watches the first N videos and collects real-time user viewing feedback information. The feedback information specifically includes but is not limited to: video completion rate, like operation, collection behavior, sharing action, and average duration of each video. These data reflect the user's true interest in different video content. The score calculation module calculates the score of each short video based on the feedback information of all new users with the same tag in multiple dimensions, generating a corresponding integral combination. This integral combination includes: completion integral (reflecting video appeal), behavior activity integral (reflecting user interaction frequency), and social interaction integral (measuring user sharing and dissemination ability). Through comprehensive scoring, the system can quantitatively evaluate the quality and popularity of each video content. The preference degree calculation module further calculates the overall preference degree of the user for the first N videos played using the above integral combination, and predicts the user's predicted preference degree for the videos not yet played based on machine learning algorithms. This module captures and predicts user interests by modeling user behavior patterns. The video sorting optimization module intelligently sorts and adjusts the videos not yet played according to the predicted preference degree. Specifically, the system will delete short videos below the preference degree threshold to avoid low-quality or irrelevant content affecting user experience; at the same time, it will also increase short videos of the same content above the preference degree threshold to strengthen the user's interest in the theme, thereby achieving dynamic optimization and personalized customization of the content library. The short videos can be pushed in a collective manner after analyzing the behavior data of all new users in a certain time period, thereby improving content publishing efficiency. It effectively improves the first-time use satisfaction and platform retention rate of new users, and has wide application prospects and promotional value. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0048] Figure 1 is a structural diagram of an intelligent media content publishing system of the present application.

[0049] Figure 2 is a structural diagram of a user tag analysis module of the present application.

[0050] Figure 3 is a structural diagram of a short video set generation module of the present application.

[0051] Figure 4 is a structural diagram of a sorting unit of the present application.

[0052] Figure 5 is a structural diagram of a user feedback collection module of the present application.

[0053] Figure 6 is a structural diagram of an integral calculation module of the present application.

[0054] Figure 7 is a structural diagram of a preference degree calculation module of the present application.

[0055] Figure 8 is a structural diagram of a video sorting optimization module of the present application.

[0056] The user tag analysis module 101, the short video set generation module 102, the video pushing module 103, the user feedback collection module 104, the integral calculation module 105, the preference degree calculation module 106, the video sorting optimization module 107, the user information acquisition unit 108, the data processing unit 109, the matching unit 110, the label acquisition unit 111, the short video screening unit 112, the sorting unit 113, the integration unit 114, the weighted matching degree calculation subunit 115, the arrangement subunit 116, the sorting correction subunit 117, the selection subunit 118, the data collection unit 119, the data structuring unit 120, the storage unit 121, the feedback information summary unit 122, the integral calculation unit 123, the interest preference degree calculation unit 124, the matching model generation unit 125, the prediction unit 126, the preference degree acquisition unit 127, the video screening unit 128, the video matching unit 129, and the video adding unit 130. DETAILED DESCRIPTION

[0057] Embodiments of the present application are described below in detail with reference to the drawings, examples of which are shown in the accompanying drawings, wherein the same or similar numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0058] The present application provides an intelligent media content publishing system, comprising a user tag analysis module 101, a short video set generation module 102, a video pushing module 103, a user feedback collection module 104, an integral calculation module 105, a preference degree calculation module 106 and a video sorting optimization module 107; the user tag analysis module 101 is used for classifying new users and assigning user tags based on initial user information; the short video set generation module 102 is used for generating a short video set according to the user tags; the video pushing module 103 is used for sending videos in the short video set to new users and playing them in order; the user feedback collection module 104 is used for collecting feedback information of users after watching the first N videos, the feedback information including complete playback rate, likes, collections, sharing operations and stay time; the integral calculation module 105 is used for generating integral combinations of each short video according to the feedback information of all new users with the same tag, the integral combinations including complete playback integral, behavior activity integral and social interaction integral; the preference degree calculation module 106 is used for calculating the preference degree of the first N videos based on the integral combinations and predicting the predicted preference degree of the videos not yet played; the video sorting optimization module 107 adjusts the videos not yet played based on the predicted preference degree, the adjustment including deleting short videos below the preference degree threshold and increasing short videos of the same kind of short videos above the preference degree threshold.

[0059] In the present embodiment, the user tag analysis module 101 is used to preliminarily classify new users and assign corresponding user tags based on the initial user information provided when they register, such as age, gender, geographical location, interests, etc. These tags constitute the initial profile of the user and provide a basis for subsequent content recommendation. The short video set generation module 102 selects short videos that match the potential interests of the user from a large number of video resources based on the user tag analysis results, and generates a preliminary short video set. This set contains videos of various types or themes suitable for this group of users to watch, ensuring content diversity while meeting individual needs. The video pushing module 103 pushes the videos in the above short video set to new users and plays them in a set order. This module supports functions such as automatic play and loop play to ensure that users can smoothly watch multiple videos and collect enough behavioral data for subsequent analysis. The user feedback collection module 104 starts after the user watches the first N videos and collects real-time feedback information. The feedback information specifically includes but is not limited to: video completion rate, like operation, collection behavior, sharing action, and average dwell time of each video. These data reflect the user's true interest in different video content. The score calculation module 105 calculates the multi-dimensional score of each short video based on the feedback information of all new users with the same tag, and generates the corresponding integral combination. This integral combination includes: completion integral (reflecting video appeal), behavior activity integral (reflecting user interaction frequency), and social interaction integral (measuring user sharing and spreading ability). Through comprehensive scoring, the system can quantitatively evaluate the quality and popularity of each video content. The preference degree calculation module 106 further calculates the overall preference degree of the user for the first N videos played using the above integral combination, and predicts the predicted preference degree of the user for the videos not yet played based on machine learning algorithms. This module captures and predicts the user's interest by modeling user behavior patterns. The video sorting optimization module 107 intelligently sorts and adjusts the videos not yet played according to the predicted preference degree. Specifically, the system will delete short videos below the preference degree threshold to avoid low-quality or irrelevant content affecting user experience; at the same time, it will also increase the same content of short videos above the preference degree threshold to strengthen the user's interest in the theme, thereby achieving dynamic optimization and personalized customization of the content library.

[0060] In summary, the present application effectively improves the first-time use satisfaction and platform retention rate of new users by constructing a complete user profile, content generation, feedback collection, data analysis and recommendation optimization closed-loop system, and has wide application prospects and promotional value.

[0061] The user tag analysis module 101 includes a user information acquisition unit 108, a data processing unit 109, and a matching unit.

[0062] The user information acquisition unit 108 is configured to receive initial registration information of a new user, the initial registration information including identity data and selection preference data; the data processing unit 109 is configured to perform cleaning and standardization processing on the initial registration information; and the matching unit is configured to perform multi-dimensional matching on the cleaned user data according to a preset label system and a rule engine, to generate a preliminary user label set.

[0063] The user information acquisition unit 108 is configured to receive and store initial registration information submitted by a new user during a registration process. The information mainly includes two aspects: one is identity data, such as the basic attributes of the user's gender, age, region, occupation background, device type, etc.; the other is selection preference data, i.e., subjective preference information such as interest labels, favorite content types, and topics of interest that the user actively checks during registration. By collecting multi-dimensional initial information, the unit provides comprehensive data support for subsequent label matching. The data processing unit 109 performs cleaning and standardization processing on the obtained initial registration information. Specifically, the unit first performs denoising, error correction, and missing value filling on the original data to eliminate data anomalies caused by user misfilling or system input errors; second, according to a unified data format standard, it converts unstructured or semi-structured user information into structured data, facilitating subsequent analysis and use; in addition, the unit also supports desensitization processing of sensitive information (such as mobile phone numbers, ID numbers, etc.), to ensure user privacy security and comply with platform compliance requirements. The matching unit performs multi-dimensional matching analysis on the user data that has been cleaned and standardized based on a preset label system and a rule engine, to finally generate a preliminary user label set for the user. The label system is a complete set of classification labels predefined by the platform, covering population attribute labels (such as age interval, gender), behavior characteristic labels (such as active time period, click habit), interest preference labels (such as "technology enthusiasts", "fashion experts", "food lovers"), etc. The rule engine automatically determines which labels are suitable for the current user according to the set logical relationship and weight strategy, and assigns a corresponding confidence score. For example, if a user fills in "sports and fitness" as their interest, and their city is a first-tier city, and their age is between 25-35 years old, the system may assign them related labels such as "urban youth", "healthy lifestyle", "short video high-frequency viewer", etc.

[0064] The short video set generation module 102 includes a label acquisition unit, a short video screening unit 112, a sorting unit 113, and an integration unit 114; the label acquisition unit is configured to acquire a user label set of a user; the short video screening unit 112 is configured to screen short video resources matched with the user label from a short video content library according to a preset label matching strategy; the sorting unit 113 is configured to sort the screened short video resources based on a matching degree weight; and the integration unit 114 is configured to generate a target short video set based on a sorting result.

[0065] The label acquisition unit is configured to acquire a user label set of a new user from the user label analysis module 101. The label set includes demographic attribute labels (such as age interval, gender, and region) of the user, behavior characteristic labels (such as active time period and device use habit), and interest preference labels (such as “tourist”, “pet lover”, and “food making”), which constitute a multi-dimensional description of the user's interest preference. The label acquisition unit supports real-time calling and caching mechanism, so as to ensure that the latest user portrait data can be quickly and accurately acquired in the short video generation process. The short video screening unit 112 screens short video resources matched with the user label from a large short video content library of the platform according to a preset label matching strategy. The strategy can be based on keyword matching, semantic analysis, image recognition, and other technical means to compare the metadata (such as title, label, classification, cover picture, and audio feature) of each short video with the user label, calculate the matching degree, and screen a candidate video set. For example, if the label of a user includes “outdoor exploration”, the system will preferentially screen short video resources with related labels such as “mountain climbing”, “camping”, and “hiking”. The sorting unit 113 further processes the screened short video resources, and sorts them according to the matching degree weight between each video and the user label. The matching degree weight is determined by multiple factors, including but not limited to label coincidence degree, historical click rate, complete play rate, and interactive heat. Meanwhile, in order to avoid aesthetic fatigue caused by single content, the system can also introduce a diversity control factor to appropriately retain some non-fully matched but potentially attractive videos under the premise of ensuring overall relevance, so as to improve the richness and explorability of the user's viewing experience. The integration unit 114 generates a target short video set for the user based on the sorting result and in combination with the playback strategy (such as playback quantity limit and content type distribution ratio) set by the platform. The set usually includes a certain number (such as 20-50) of short videos, covers high-matching-degree interest content of the user, and is supplemented with a small amount of exploratory content, so as to form an initial playback list with reasonable structure, high-quality content, and good user experience. The integrated short video set is pushed to a video playback queue, and waits for subsequent playback and feedback collection.

[0066] The sorting unit 113 includes a weighted matching degree calculation subunit 115, an arrangement subunit 116, a sorting correction subunit 117, and a selection subunit 118. The weighted matching degree calculation subunit 115 is configured to calculate the weighted matching degree between each short video resource in the candidate short video set and the user label. The arrangement subunit 116 is configured to arrange the candidate short video resources in descending order according to the weighted matching degree. The sorting correction subunit 117 is configured to correct the descending arrangement based on the video heat. The selection subunit 118 is configured to cut the top N items of the sorted short video resources according to a preset number.

[0067] The weighted matching degree calculation subunit 115 is configured to calculate the weighted matching degree between each video resource in the candidate short video set output by the short video screening unit 112 and the user label. Specifically, the subunit first extracts the metadata information of each short video, including title keywords, label system, classification attributes, content theme, and the like. Then, by combining the preliminary user label set of the user and by setting different weight coefficients (for example, the proportion of “interest preference” is high, and the proportion of “demographic attribute” is low), a comprehensive matching score is calculated. For example, if a short video contains two labels “travel” and “photography”, and the user label also contains similar labels such as “tourism” and “landscape photography”, the matching degree score of the video will be higher. In addition, the system can also introduce a machine learning model to continuously optimize the weight distribution strategy according to historical user behavior feedback and improve the matching accuracy.

[0068] The arrangement subunit 116 sorts the videos in descending order after obtaining the weighted matching degrees of all candidate short videos, forming a preliminary recommendation sequence. This subunit supports multiple sorting algorithms, such as direct descending order sorting based on scores, or introducing a bucketing mechanism to divide videos into different priority intervals, providing a flexible basis for subsequent correction. The sorting correction subunit 117 further introduces a video popularity factor to dynamically adjust the preliminary sorting results. The popularity factor usually includes multiple dimensions, such as the video's historical play count, like rate, collection count, share count, and real-time exposure frequency within the platform. This subunit adjusts the original sorting by setting a popularity weight factor. For example, for two videos with similar matching degrees, if one has significantly higher popularity indicators, the system can appropriately improve its sorting position to balance user interest and content popularity, improving overall recommendation effectiveness and user click willingness. The selection subunit 118 extracts the top N short video resources from the sorted list according to the platform's preset rules (such as initial recommendation video quantity limit) after sorting and correction, as part of the final short video set for pushing. The value of N can be adaptively adjusted according to user type, device screen size, network status, etc., such as setting 20 videos for the first recommendation of new users, and dynamically expanding or compressing the recommendation quantity based on feedback.

[0069] The user feedback collection module 104 includes a data collection unit 119, a data structuring unit 120, and a storage unit 121.

[0070] The data collection unit 119 collects user interaction behavior data during the viewing of the first N videos when the user's cumulative completed video viewing count reaches a preset value N. The interaction behavior data includes the completion status and completion rate of each video, whether the user likes the video, whether the user adds the video to the collection list, whether the user shares the video, the user's dwell time on each video, and the data structuring unit 120 cleans and structures the collected interaction behavior data to generate standardized user feedback records. The storage unit 121 stores the user feedback records in the user behavior database and marks them as association data between corresponding video content and user ID.

[0071] The data collection unit 119 starts comprehensive collection of user interaction behavior when the user's cumulative completed video viewing count reaches a preset value N. Specifically, this unit monitors and records multiple types of interaction behavior data generated during the viewing of the first N short videos, including but not limited to:

[0072] Completion status and completion rate: whether the user watched a certain video completely, and the proportion of actual viewing time to total video time.

[0073] Like operation: whether the user has liked the video;

[0074] Collection behavior: whether the user has added the video to a personal collection or a specific playlist;

[0075] Sharing operation: whether the user has forwarded the video to other users within the platform or to external social networks;

[0076] Stay duration: the average viewing time of the user on each video, used to judge the user's interest level and content appeal.

[0077] The above behavior data is reported in real time by the client SDK or front-end point mechanism, and is transmitted to the back-end processing module through the message queue or API interface, ensuring the timeliness and accuracy of data collection.

[0078] The data structuring unit 120 cleans, denoises, and standardizes the collected raw interaction behavior data, converting it into structured data in a unified format. This unit first performs validity check on the raw log data, eliminating abnormal data (such as excessive stay time, repeated clicks, etc.), and then maps various behavior events to standard fields according to the platform-defined data model, such as "video_id", "user_id", "is_liked", "watch_duration", "share_count", etc., thus generating user feedback records with clear semantics. In addition, this unit also supports multi-dimensional behavior feature aggregation analysis, extracting basic feature vectors that can be used for subsequent modeling. The storage unit 121 writes the structured user feedback records into the system's user behavior database, and establishes the association between video content and user ID. These records are usually stored in high-performance databases (such as HBase, Elasticsearch, ClickHouse, etc.) in the form of key-value pairs or documents, facilitating subsequent fast retrieval and batch analysis. At the same time, each feedback record is timestamped and labeled, to support statistical analysis by time window, user group, or video category dimensions. In addition, the system also synchronizes key behavior data to the data warehouse for offline training of recommendation models.

[0079] In summary, the user feedback collection module 104 acquires real-time interaction behavior of users through the data collection unit 119, improves data quality and usability with the help of the data structuring unit 120, and completes efficient persistent storage by the storage unit 121, building a complete chain from user behavior perception to data sedimentation. This module not only provides the system with rich user interest signals, but also lays a solid data foundation for subsequent optimization of personalized recommendation algorithms and adjustment of content strategies.

[0080] The integral calculation module 105 includes a feedback information aggregation unit 122 and an integral calculation unit 123; the feedback information aggregation unit 122 is configured to obtain user feedback records of all new users under the same user label category; the integral calculation unit 123 is configured to map the user feedback record data to a corresponding integral model, and quantize each integral after mapping; the integral model includes a complete playback integral triggered if a user completes video playback, an interaction integral triggered if a user likes, collects or shares a video, and a behavior activity integral triggered if a user stays on a video page for more than a set threshold.

[0081] The feedback information aggregation unit 122 obtains user feedback records generated by a group of new users with the same user label category from a user behavior database. These records are derived from the interaction data collected and structured in the user feedback collection module 104, including but not limited to the complete playback of the first N videos watched by each user, like behavior, collection behavior, sharing behavior, and dwell time, and other key indicators. By aggregating the behavior data of multiple users under the same label, this unit can form a behavior statistics set for a specific user group, providing a unified data basis for subsequent integral calculation. In addition, this unit also supports extracting data by time window (such as the last 24 hours, 7 days) or behavior type classification to adapt to the needs of different recommendation strategies.

[0082] The integral calculation unit 123 maps the above feedback information to the corresponding integral dimension based on the preset integral model, and quantizes each behavior indicator, finally generating a multi-dimensional integral combination for each short video. The integral model mainly includes the following three integral types:

[0083] Complete playback integral: triggered when a user completes the complete playback of a certain short video. This integral reflects the attractiveness and integrity of the video content, and is usually calculated by weighting according to the complete playback rate (such as 100% complete playback for full score, partial complete playback for proportional decrease).

[0084] Interaction integral: triggered when a user performs like, collect or share operations on the video. These three behaviors represent the user's recognition, retention intention and dissemination intention for the video content, and the system can score comprehensively according to the behavior weight (such as sharing > collection > like), reflecting the social influence and user stickiness of the video.

[0085] Behavior activity integral: triggered when a user stays on a video page for more than a set threshold (such as 60% of the total video duration or an absolute time such as 8 seconds). This integral measures the user's attention and immersion to the video content, the longer the dwell time, the more attractive the content, and the higher the corresponding integral.

[0086] The integral calculation unit 123 converts the original behavior data into standardized integral values by means of weighted summation, normalization processing, etc., and dynamically adjusts in combination with historical data. For example, the basic integral coefficient can be appropriately reduced for popular videos, and an incentive scoring mechanism can be set for emerging high-quality content, so as to realize benign guidance of the platform content ecology.

[0087] The preference degree calculation module 106 includes an interest preference degree calculation unit 124, a matching model generation unit 125, and a prediction unit 126. The interest preference degree calculation unit 124 is configured to calculate the interest preference degree of a user for the first N short video resources currently displayed based on the content preference integral generated based on the feedback information of the user. The matching model generation unit 125 is configured to construct a matching model between the content feature vectors of the played videos and the content feature vectors of the unplayed videos. The prediction unit 126 is configured to match the candidate short video resources that have not been played by using the matching model and in combination with the interest preference degree, to obtain the corresponding predicted preference degree.

[0088] The interest preference degree calculation unit 124 is configured to calculate the interest preference degree of a user for the first N short video resources currently displayed based on the feedback information of the user and in combination with the content preference integral output by the integral calculation module 105. Specifically, this unit first receives the structured feedback records from the user feedback collection module 104 and the multi-dimensional integral (such as the complete playback integral, the interaction integral, and the behavior activity integral) generated by the integral calculation module 105, and then fuses each index into a value representing the overall interest intensity of the user according to a preset weighting formula or a neural network model, i.e., the interest preference degree. For example, if a user shows high complete playback rate, frequent like and collection behaviors on multiple videos, the interest preference degree of the user for this type of content will be assigned a higher value, reflecting the strong viewing intention and content matching degree of the user.

[0089] The matching model generation unit 125 constructs a matching model between the content feature vectors of the played videos and the content feature vectors of the unplayed candidate videos, to support subsequent interest migration and preference prediction. This unit encodes the metadata features (such as title keywords, label system, classification attributes, image features, audio semantics, etc.) of each video into a vector representation of a unified dimension by extracting the metadata features; at the same time, a deep matching model (such as a Dual-Tower Model, a Collaborative Filtering Model CF, or a Graph Neural Network GNN) is trained by using the interest preference degree of the user as a target variable, to measure the similarity and correlation between any two videos. This model not only considers the matching relationship of the content itself, but also fuses the feedback signals of the user historical behaviors, thereby improving the accuracy and generalization ability of the prediction.

[0090] The prediction unit 126 then performs matching analysis on the candidate short video resources that have not been played based on the matching model described above, and combines the user's current interest preference to predict the prediction preference of these videos. The specific process is as follows: first, the similarity score between each candidate video and the user's watched video is calculated using the matching model; second, the historical interest preference weight of the user is combined to obtain the overall prediction preference score of the candidate video; finally, the candidate video is sorted according to the score result, and the input basis for the subsequent video sorting optimization module 107 is provided. In addition, this unit also supports the introduction of time decay factor, context environment perception (such as time period, device type) and other enhancement factors, to further improve the dynamic adaptation ability and personalized expression ability of the prediction model.

[0091] In summary, the preference calculation module 106 accurately quantifies the user's interest in the played video through the interest preference calculation unit 124, establishes the matching logic between the contents by means of the matching model generation unit 125, and completes the prediction preference evaluation of the unplayed video by the prediction unit 126, thereby forming a complete user interest modeling and content recommendation prediction mechanism. This module not only improves the intelligent level of the platform recommendation system, but also provides a solid technical support for realizing personalized content pushing and improving user satisfaction.

[0092] The specific steps of constructing the matching model between the played video content feature vector and the unplayed video content feature vector include: extracting the content feature vector of each played video, which includes but is not limited to video tags, categories, creator attributes, publishing time, hotness indicators, and video duration; extracting the video content feature vector of each unplayed candidate video; constructing a matching model based on content similarity.

[0093] Extract the content feature vector of each played video. The so-called "content feature vector" refers to expressing the video content in a structured and numerical way for subsequent modeling and calculation. The vector is usually composed of multiple dimensions, mainly including but not limited to the following types of information:

[0094] Video tags: automatically or manually annotated keyword tags by the platform, used to describe the theme content of the video (such as "food making", "fitness training", "travel guide", etc.);

[0095] Category attribute: the major category classification to which the video belongs, such as "entertainment", "education", "life", "technology", etc.;

[0096] Creator attribute: information related to the creator who published the video, such as the number of fans, content style, historical video quality score, etc.;

[0097] Upload time: the time when the video is uploaded, which can be used to determine the freshness and timeliness of the content;

[0098] Heat index: including the historical play count, like count, collection count, and sharing frequency of the video, which reflect the user's attention;

[0099] Video duration: the duration of the video, which affects the user's viewing experience and content density.

[0100] After encoding, these features can form a high-dimensional numerical vector, which represents the video in the feature space.

[0101] Secondly, extract the video content feature vector of each unplayed candidate video. For candidate short video resources that have not been watched by users in the system, the content features also need to be extracted according to the above standards, and standardized feature vectors are generated. This process usually uses natural language processing (NLP), image recognition, audio semantic analysis and other multi-modal technologies to ensure that different types of video content can be uniformly modeled and compared.

[0102] Finally, build a matching model based on content similarity. The core goal of this model is to measure the content relevance between played and unplayed videos, so as to infer the user's interest preference for unwatched videos. The specific methods include:

[0103] Use the cosine similarity (Cosine Similarity) mathematical method to calculate the similarity score between two video feature vectors;

[0104] Introduce deep learning models (such as Siamese network, Dual-Tower Model, Graph Neural Network GNN, etc.) to learn more complex feature mapping relationships using neural network structures;

[0105] Combine user historical behavior data (such as completion rate, interaction points, dwell time, etc.) to weight and integrate the similarity score to form the final matching score;

[0106] Sort the matching results to provide the association between video content and user interest for the prediction unit 126.

[0107] The video ranking optimization module 107 includes a preference acquisition unit 127, a video screening unit 128, a video matching unit 129, and a video adding unit 130. The preference acquisition unit 127 is configured to acquire a predicted preference of a user for a candidate unplayed short video, where the predicted preference is calculated by a matching model based on historical behaviors of the user and content features of the video. The video screening unit 128 is configured to perform a deletion operation on a short video whose predicted preference is lower than a preference threshold, so as to eliminate low-relevance content in a recommendation result. The video matching unit 129 is configured to, for a short video whose predicted preference is higher than the preference threshold, extract content features of the short video, and search for similar short videos in a candidate library. The video adding unit 130 is configured to add the found similar short videos to a short video set, where the number of the similar short videos is the same as the number of deleted short videos.

[0108] The preference acquisition unit 127 is configured to acquire, from the preference calculation module 106, a predicted preference of a user for all candidate unplayed short videos. The predicted preference is calculated by a matching model based on similarity between historical viewing behaviors (such as a complete playback rate, interactive behaviors, and a stay duration) of the user and a content feature vector of a video, and reflects a possibility of the user producing positive feedback in a future viewing process of the video. The unit supports a real-time calling and caching mechanism, so as to ensure that the latest predicted score result can be quickly acquired in a ranking optimization process.

[0109] The video screening unit 128 evaluates all candidate short videos according to their predicted preferences, and performs a screening operation. Specifically, the unit sets a preference threshold (Threshold), and determines, as content with a low matching degree of interest of a user, a short video whose predicted preference is lower than the threshold. The system automatically deletes the short video from a recommendation list, so as to eliminate low-relevance and low-attractive video resources, and improve quality and accuracy of an overall recommendation result. The threshold can be dynamically adjusted according to platform strategies, user types, or recommendation stages, so as to adapt to recommendation requirements in different scenarios.

[0110] The video matching unit 129 further extracts a content feature vector of a short video whose predicted preference is higher than the preference threshold, including but not limited to a video tag, a category attribute, a creator style, a heat index, and the like. The unit searches for similar short videos in a candidate short video library based on the content features, so as to ensure that the found similar short videos are consistent with the high-preference video in terms of theme, style, and quality, and the like.

[0111] The video adding unit 130 adds the same short video found by the video matching unit 129 to the current short video set, while ensuring that the number of newly added short videos is equal to the number of low-preference short videos filtered out, so that the overall length of the recommendation queue is maintained. This "replacement" update mechanism not only preserves the continuity and stability of the short video playlist, but also improves the coherence and immersion of the user's viewing experience by introducing more high-quality, high-matching same content. In addition, the system can continuously optimize and adaptively adjust this replacement strategy according to the user's subsequent feedback behavior.

[0112] In summary, the video sorting optimization module 107 obtains the user interest prediction result through the preference obtaining unit 127, removes low-relevance content using the video filtering unit 128, mines high-quality same videos using the video matching unit 129, and completes dynamic supplement and optimization of recommended content through the video adding unit 130. This module effectively improves the intelligent level and personalization ability of the short video recommendation system, not only enhances the user experience satisfaction, but also provides strong support for the healthy development of the platform content ecosystem.

[0113] The specific steps of deleting short videos below the preference threshold include: traversing each video in the candidate short video set; comparing its predicted preference with the preset minimum preference threshold; if the preference of the video is less than the threshold, removing it from the candidate set.

[0114] Traverse each video resource in the candidate short video set. After completing the user predicted preference calculation, the system takes the candidate unplayed video list as the processing object, and analyzes each video in it one by one in a sequential or parallel manner. This candidate set is usually generated by the platform according to content library, user tags, hotness strategy and other conditions, and has obtained the corresponding predicted preference score through the preference calculation module 106.

[0115] Secondly, compare the predicted preference of the video with the preset minimum preference threshold. This threshold is a benchmark value set by the platform based on historical data modeling, A / B test results and business objectives, and is used to measure whether the video has enough recommendation potential. For example, if the predicted preference of a video is 0.65, and the system sets the minimum threshold to 0.7, the video will be considered as low-relevance content. This threshold can be dynamically adjusted according to different user groups (such as new users, active users), different time periods (such as peak hours, off-peak hours) or different content categories (such as news, entertainment), to improve the flexibility and adaptability of the recommendation system.

[0116] Finally, if the prediction preference degree of the video is less than the set threshold, it is removed from the candidate video set. This operation is automatically performed by the system by updating the data structure of the recommendation queue (such as array, linked list, queue, etc.), and eliminating the video entries that do not meet the recommendation standard. The removed video no longer enters the subsequent push process, thereby effectively avoiding low-quality, low-matching content from interfering with the user's viewing experience, and improving the accuracy and effectiveness of the overall recommendation.

[0117] In addition, in practical applications, the system can also introduce a gradient elimination mechanism or a multi-level screening strategy, that is, set multiple preference degree intervals (such as high, medium, and low), which correspond to different processing strategies (such as priority display, weight reduction, and direct elimination), to further refine the recommendation control logic. At the same time, all deleted video information can be recorded to the log system for subsequent behavior modeling, content quality evaluation, or algorithm optimization.

[0118] The above only discloses a preferred embodiment of the present application, of course, cannot limit the scope of the present application, those skilled in the art can understand that all or part of the above-mentioned embodiments can be implemented, and equivalent changes made according to the claims of the present application still belong to the scope covered by the present application.

Claims

1. An intelligent media content publishing system, characterized in that, comprising a user tag analysis module, a short video set generation module, a video pushing module, a user feedback collection module, an integral calculation module, a preference degree calculation module and a video sorting optimization module; the user tag analysis module is used for classifying new users and assigning user tags based on initial user information; the short video set generation module is used for generating a short video set according to the user tags; the video pushing module is used for sending videos in the short video set to new users and playing them in order; the user feedback collection module is used for collecting feedback information of users after they watch the first N videos, the feedback information including complete playback rate, likes, collections, sharing operations and dwell time; the integral calculation module is used for generating integral combinations of each short video according to the feedback information of all new users with the same tag, the integral combination including complete playback integral, behavior activity integral and social interaction integral; the preference degree calculation module is used for calculating the preference degrees of the first N videos based on the integral combinations and predicting the predicted preference degrees of the videos not yet played; the video sorting optimization module adjusts the videos not yet played based on the predicted preference degrees, the adjustment including deleting short videos below the preference degree threshold and increasing short videos of the same kind of short videos above the preference degree threshold.

2. The intelligent media content publishing system of claim 1, characterized in that, the user tag analysis module includes a user information acquisition unit, a data processing unit and a matching unit; the user information acquisition unit is used for receiving initial registration information of new users, the initial registration information including identity data and selection preference data; the data processing unit is used for cleaning and standardizing the initial registration information; the matching unit is used for performing multi-dimensional matching on the cleaned user data according to a preset tag system and a rule engine to generate a preliminary user tag set.

3. The intelligent media content publishing system of claim 2, characterized in that, the short video set generation module includes a tag acquisition unit, a short video screening unit, a sorting unit and an integration unit; the tag acquisition unit is used for acquiring the user tag set of the user; the short video screening unit is used for screening short video resources matching the user tags from a short video content library according to a preset tag matching strategy; the sorting unit is used for sorting the screened short video resources based on matching degree weight; the integration unit is used for generating a target short video set based on the sorting result.

4. The intelligent media content publishing system of claim 3, characterized in that, the sorting unit includes a weighted matching degree calculation subunit, an arrangement subunit, a sorting correction subunit and a selection subunit; the weighted matching degree calculation subunit is used for calculating the weighted matching degree between each short video resource in the candidate short video set and the user tags; the arrangement subunit is used for arranging the candidate short video resources in descending order according to the weighted matching degree; The sorting modifier unit is configured to modify the descending order based on video popularity. The selection subunit is configured to cut the first N items of the sorted short video resources according to a preset number. 5.The intelligent media content publishing system of claim 4, wherein, The user feedback collection module comprises a data collection unit, a data structuring unit, and a storage unit. The data collection unit is configured to collect interactive behavior data of a user during the process of watching the first N videos when the number of videos watched by the user accumulatively reaches a preset value N, the interactive behavior data comprising a complete play status and a complete play rate of each video. whether the user performs a like operation on the video; whether the user adds the video to a collection list; whether the user performs a sharing operation on the video; and the length of time the user stays on each video. The data structuring unit is configured to clean and structure the collected interactive behavior data to generate standardized user feedback records. The storage unit is configured to store the user feedback records in a user behavior database and mark the records as association data between corresponding video content and user IDs. 6.The intelligent media content publishing system of claim 5, wherein, The feedback information summary unit is configured to obtain user feedback records of all new users under the same user label category. The integral calculation unit is configured to map the user feedback record data to a corresponding integral model and quantitatively process each integral after mapping; the integral model comprises a complete play integral triggered if a user completes video playback, an interactive integral triggered if a user likes, collects, or shares a video, and an active degree integral triggered if a user stays on a video page for more than a set threshold. 7.The intelligent media content publishing system of claim 6, wherein, The preference degree calculation module comprises an interest preference degree calculation unit, a matching model generation unit, and a prediction unit. The interest preference degree calculation unit is configured to calculate the interest preference degree of a user for the first N short video resources currently displayed based on the content preference integral generated from the feedback information of the user. The matching model generation unit is configured to construct a matching model between the content feature vectors of played videos and unplayed videos. The prediction unit is configured to match candidate short video resources that have not been played using the matching model and combine the interest preference degree to obtain a corresponding predicted preference degree. 8.The intelligent media content publishing system of claim 7, wherein, The specific steps of constructing a matching model between the content feature vectors of played videos and unplayed videos comprise: extracting the content feature vectors of each played video, the content feature vectors comprising but not limited to video labels, categories, creator attributes, publishing times, popularity indicators, and video lengths; extracting the video content feature vectors of each unplayed candidate video; and constructing a matching model based on content similarity. 9.The intelligent media content publishing system of claim 8, wherein, the video ranking optimization module comprises a preference acquisition unit, a video screening unit, a video matching unit and a video adding unit; the preference acquisition unit is configured to acquire a predicted preference of a user for a candidate unplayed short video, the predicted preference being calculated by a matching model according to historical behaviors of the user and content features of the video; the video screening unit is configured to perform a deletion operation on a short video whose predicted preference is lower than a preference threshold, so as to eliminate low-relevance content in the recommendation result; the video matching unit is configured to, for a short video whose predicted preference is higher than the preference threshold, extract content features thereof, and search for similar short videos in a candidate library; the video adding unit is configured to add the found similar short videos to the short video set, the number of the similar short videos being the same as the number of the deleted short videos. 10.The intelligent media content publishing system of claim 9, wherein, the specific steps of deleting the short videos lower than the preference threshold comprise: traversing each video in the candidate short video set; comparing the predicted preference of the video with a preset minimum preference threshold; and if the preference of the video is lower than the threshold, removing the video from the candidate set.

Images