A data information acquisition system and a data information acquisition method

Abstract

This invention relates to a data information acquisition system, comprising a terminal-side acquisition unit and a server-side processing unit. The terminal-side acquisition unit is deployed on a user's video playback terminal, and the server-side processing unit is deployed on a cloud server. In this invention, a quantitative model integrating static attributes, dynamic behavior, and time decay characteristics is used to transform fuzzy requirements into calculable preference degrees. This solves the problem of traditional algorithms struggling to quantify behavioral weights and time effects, improving the accuracy of capturing users' true video needs. Furthermore, the model's dynamic iterative update mechanism can optimize parameters in real time, adapting to users' short-term interest switching and long-term stable needs, avoiding the demand bias caused by the rigidity of traditional algorithm models. Moreover, the model's matching dimensions, behavioral types, and weight coefficients can all be adjusted according to different scenarios such as short videos, long videos, and professional domain videos, making it widely applicable.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a data information acquisition system and a data information acquisition method. Background Technology

[0002] With the development of big data and artificial intelligence technologies, intelligent data collection and information push have become core means for video platforms to improve user experience. Existing technologies primarily focus on user information verification, interest point mining, and bandwidth adaptation to optimize data collection and push performance.

[0003] For example, Chinese invention patent CN117556151A discloses a method and system for intelligent data collection and information push based on big data. Through steps such as enterprise user account registration and verification, industry keyword matching, and interest category association, it establishes a connection between users and an industry category database to achieve targeted information push. Chinese invention patent CN108282541A discloses a method and system for information push based on big data. Through steps such as establishing browsing points of interest, bandwidth-adapted bitrate, and dynamic updating of points of interest, it optimizes the adaptability of information push.

[0004] However, the aforementioned existing technologies still have significant shortcomings in video data acquisition and demand capture scenarios: users' interactive behaviors with videos are complex and varied, including positive behaviors such as clicking and liking, as well as negative behaviors such as fast-forwarding and swiping. Moreover, behavioral preferences change dynamically over time. Traditional algorithms rely only on single behaviors or static labels for modeling, making it difficult to quantify the demand weight of different behaviors and the impact of time decay. This results in the inability to accurately capture users' real needs for video data, thereby affecting the targeting of data acquisition and the accuracy of push notifications. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] This invention provides a data information acquisition system and method, which solves the problems of complex and ever-changing user video interaction behaviors and the difficulty of traditional algorithms in accurately capturing users' real needs in existing intelligent data acquisition and information push methods.

[0007] The problem.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, the present invention provides the following technical solution: a data information acquisition system, comprising a terminal-side acquisition unit and a server-side processing unit. The terminal-side acquisition unit is deployed on a user's video playback terminal and includes: a user behavior perception module for real-time acquisition of dynamic interaction behavior data between the user and the video; a terminal environment perception module for acquiring terminal hardware characteristics, network environment characteristics, and playback environment characteristics; a lightweight video feature acquisition module for acquiring basic extractable features from the video terminal and performing standardized encoding; and an edge preprocessing module for deduplicating, removing outliers, and calibrating timestamps on the acquired multi-dimensional data, and pushing the preprocessed data to the server-side processing unit at a fixed frequency, while simultaneously caching local short-term behavior data. The server-side processing unit is deployed on a cloud server and includes: a data receiving and fusion module for receiving preprocessed data from each terminal-side acquisition unit; a user static attribute management module for storing and updating static attribute data during user registration; a video full-feature extraction module for extracting deep video features and performing quantization encoding based on computer vision and natural language processing technologies; a demand preference calculation module for performing quantitative calculations of user preferences for various types of videos; a model dynamic update module for iteratively updating the weight parameters and attenuation coefficients of the calculation model in real time based on the user's subsequent video interaction behavior to adapt to changes in user behavior; an intelligent acquisition strategy generation module for generating differentiated video data acquisition strategies based on user demand preferences and terminal environment characteristics and distributing them to the terminal-side acquisition units; and a data storage module for storing user-related data, video-related data, model parameters, and iteration records.

[0010] This invention also provides the following technical solution: a data information acquisition method, comprising the following steps:

[0011] Step 1: Initial collection of user and video basic data. The user static attribute management module collects and stores static attribute data during the user registration phase and assigns a unique user ID to each user. The video full feature extraction module completes full feature collection and quantization encoding of all videos in the platform to generate video feature vectors. The video lightweight feature collection module collects basic video features simultaneously. The terminal environment perception module on the terminal side collects the initial environmental features of the terminal.

[0012] Step 2: Real-time collection of user video interaction dynamic behavior. During the video playback process, the user behavior perception module on the terminal side captures all user video interaction behaviors in real time with a minimum collection granularity of 100ms, records the behavior type, behavior occurrence timestamp, and associated video ID. After the edge preprocessing module removes abnormal behaviors, the data is pushed to the data receiving and fusion module.

[0013] Step 3: Multi-dimensional data collection, quantization, encoding, and fusion. The data receiving and fusion module merges the dynamic behavior data, terminal environment data, and lightweight video feature data collected by the terminal-side collection unit with the user static attribute data and full video feature data stored by the server-side processing unit. Through one-hot encoding or tag encoding embedding, the non-numerical data is converted into numerical features, generating a standardized user-video interaction data matrix.

[0014] Step 4: Quantitative calculation of user video demand preference based on the calculation model. The demand preference calculation module calls the user video demand comprehensive preference calculation model. With the user-video interaction data matrix as input, it calculates the user static attribute basic preference and the user dynamic behavior preference respectively. The dynamic behavior preference is corrected by the time decay factor and then fused to obtain the user video demand comprehensive preference.

[0015] Step 5: Calculate the dynamic iterative update of the model. After a user generates a new video interaction behavior, the server-side model dynamic update module uses the new behavior data as a feedback sample to update the behavior weight parameters of the dynamic behavior preference and the decay coefficient of the time decay factor. The model update cycle is dynamically adjusted according to the user's behavior activity.

[0016] Step 6: Execution of differentiated video data acquisition strategy based on preference. The intelligent acquisition strategy generation module generates and distributes differentiated acquisition strategies to the terminal acquisition unit based on the user's video needs, preferences, and terminal environment characteristics. The terminal then executes video data acquisition according to the strategy.

[0017] Preferably, the user video demand comprehensive preference calculation model adopts a three-layer quantization fusion architecture, and the calculation formula of the user video demand comprehensive preference calculation model is as follows:

[0018] ;

[0019] In the formula, P(i,j) represents the overall preference of user i for video j, and α is the static preference weight coefficient. Let i represent the user's basic preference for the static attributes of video j. Let λ(t) represent the dynamic behavior preference of user i for video j, and let λ(t) be the time decay correction factor.

[0020] In a further preferred embodiment, the static attribute base preference degree The calculation formula is:

[0021] ;

[0022] In the formula, n is the total number of matching dimensions between user static attributes and video features, and k is the k-th matching dimension. Let S(i,k) be the weight coefficient of the k-th matching dimension, S(i,k) be the standardized feature value of user i in the k-th matching dimension, and T(j,k) be the standardized feature value of video j in the k-th matching dimension.

[0023] In a further preferred embodiment, the dynamic behavior preference is... The calculation formula is:

[0024] ;

[0025] In the formula, m represents the total number of user interaction types with the video. Let B(i,j,m) be the weight coefficient for the m-th type of interaction behavior, and let B(i,j,m) be the standardized behavior value of user i's m-th type of interaction behavior with video j. The denominator is... This is the normalization factor.

[0026] In a further preferred embodiment, the time decay correction factor λ(t) is calculated using the following formula:

[0027] ;

[0028] In the formula, e is the natural constant, γ is the time decay coefficient, and Δt is the behavior time difference.

[0029] In a further preferred embodiment, the terminal-side acquisition unit and the server-side processing unit adopt a bandwidth-adaptive encrypted transmission protocol. When the bandwidth is below a threshold, only core behavior and feature data are transmitted; when the bandwidth is sufficient, all data is transmitted, and all data is processed with end-to-end encryption.

[0030] In a further preferred embodiment, the core execution rules of the differentiated video data acquisition strategy described in step 6 include: for video types ranking in the top 30% of user preference, increasing the feature acquisition frequency and adding deep feature acquisition dimensions for that type of video; for low-bandwidth and low-computing-power terminals, only collecting basic video features and core user behaviors; enabling real-time high-frequency acquisition during periods of high user video playback activity, and enabling low-frequency timed acquisition during periods of low activity activity; for video types with a user preference of 0, stopping active acquisition and retaining only passive acquisition capabilities.

[0031] In a further preferred embodiment, the standardization rules for the interactive behaviors are as follows: the completion rate is the ratio of the actual viewing time to the total video duration; for click, like, favorite, comment, and replay behaviors, the value is 1 if the behavior exists, and 0 if it does not exist; the fast-forward or skip behavior is the ratio of the fast-forward or skip duration to the total video duration; the swipe-away behavior is 1 if it exists, and 0 if it does not exist.

[0032] (III) Beneficial Effects

[0033] Compared with the prior art, the present invention provides a data information acquisition system and a data information acquisition method, which have the following beneficial effects:

[0034] In this invention, a quantization model that integrates static attributes, dynamic behavior, and time decay characteristics is used to transform fuzzy requirements into calculable preference degrees, solving the problem that traditional algorithms have difficulty quantifying behavioral weights and time effects, and improving the accuracy of capturing users' real video needs.

[0035] In this invention, the model's dynamic iterative update mechanism can optimize parameters in real time, adapting to users' short-term interest switching and long-term stable demand behavior characteristics, and avoiding demand deviations caused by the rigidity of traditional algorithm models.

[0036] In this invention, the matching dimensions, behavior types, and weight coefficients of the model can all be adjusted according to different scenarios such as short videos, long videos, and professional domain videos, making it widely applicable. Furthermore, the differentiated collection strategy, combined with terminal environment awareness, avoids indiscriminate full collection, reduces the resource consumption of the terminal and server, and balances the requirements of comprehensive collection with lightweight design. Attached Figure Description

[0037] Figure 1 This is an architecture diagram of the data information collection system according to the implementation plan;

[0038] Figure 2 This is a flowchart of the data collection method according to the implementation plan. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] See Figure 1 A data information acquisition system consists of a terminal-side acquisition unit and a server-side processing unit. The two interact with each other through a bandwidth-adaptive encrypted transmission protocol. The overall architecture balances lightweight acquisition with high-precision computing requirements.

[0041] The terminal-side acquisition unit is deployed on user video playback terminals (mobile phones, tablets, computers, smart TVs, etc.). With low resource consumption as its core objective, it completes front-end data acquisition and preliminary preprocessing, specifically including four modules:

[0042] User behavior awareness module: With a minimum collection granularity of 100ms, it captures the dynamic interaction behavior between users and videos in real time, records the behavior type (click, viewing duration, viewing completion rate, like, favorite, comment, fast forward or skip / replay, swipe away), the timestamp of the behavior, and the associated video identifier, to ensure the real-time and completeness of the behavior data.

[0043] Terminal environment perception module: Collects terminal hardware characteristics (device type, screen resolution, computing power), network environment characteristics (bandwidth, latency, network type), and playback environment characteristics (playback period, volume, full-screen status) to provide environmental basis for subsequent differentiated collection strategies.

[0044] Lightweight video feature acquisition module: Extracts basic video features that can be quickly obtained from the terminal side, including video ID, category label, duration, resolution, and release time, and standardizes the encoding of features to reduce data transmission overhead.

[0045] Edge preprocessing module: Deduplicatizes the collected multi-dimensional data, removes outliers (such as accidental clicks), and calibrates timestamps. The preprocessed data is pushed to the server at a fixed frequency, while short-term behavior data is cached locally to avoid data loss.

[0046] The server-side processing unit is deployed on a cloud server or edge cloud node to complete deep data processing, model calculation, storage, and policy generation. It specifically includes seven modules:

[0047] Data receiving and fusion module: Receives preprocessed data pushed from the terminal side and fuses it with user static attributes and full video feature data stored in the cloud. Through one-hot encoding or tag embedding technology, it converts non-numerical data (such as behavior type and video tag) into numerical features, generating a standardized user-video interaction data matrix to provide a unified input for model calculation.

[0048] User Static Attribute Management Module: Stores and updates static attribute data (age, gender, occupation, self-selected video interest tags, region) during the user registration phase, providing a basis for demand preference calculation, and also supports dynamic updates of user attributes.

[0049] Full Video Feature Extraction Module: Based on computer vision and natural language processing technologies, it extracts deep features from videos, including keywords in the video content, audio semantics, subtitle information, creator tags, and audience profiles. After quantization and encoding, these features are stored in association with the basic features.

[0050] Demand Preference Calculation Module: Equipped with a core user video demand comprehensive preference calculation model, it takes a data matrix as input and outputs the user's comprehensive preference for each video through three-layer calculation: static basic preference, dynamic behavioral preference, and time decay correction.

[0051] The model dynamic update module uses subsequent user interactions as feedback samples to iteratively update the model's behavior weight parameters and time decay coefficient in real time, adapting to changes in user behavior (i.e., increasing the decay coefficient when user behavior changes frequently and decreasing it when behavior is stable); at the same time, it adjusts the static preference weight coefficient based on the amount of user behavior data to optimize model adaptability.

[0052] Intelligent data acquisition strategy generation module: Based on the user's comprehensive preferences and terminal environment characteristics, it generates differentiated data acquisition strategies and distributes them to the terminal side, realizing dynamic adjustment of data acquisition frequency and feature dimensions.

[0053] Data storage module: Adopting a distributed storage architecture, it is divided into a user database (storing static attributes, historical behaviors, and preference results), a video database (storing basic features, deep features, and acquisition logs), and a computing model database (storing model parameters and iteration records), ensuring data security and scalability.

[0054] The terminal and server sides adopt a bandwidth-adaptive encrypted transmission protocol: when the bandwidth is below the threshold, only core behavior and basic feature data are transmitted to reduce resource consumption; when the bandwidth is sufficient, all data is transmitted; all data is encrypted end-to-end to ensure data transmission security.

[0055] See Figure 2 Based on the above system, the data acquisition method includes the following six steps:

[0056] Step 1: Initial collection of user and video basic data:

[0057] The server-side user static attribute management module collects user registration data and assigns a unique ID, storing it as static attribute data; the video full feature extraction module completes the full feature collection and encoding of all videos on the platform, generating video feature vectors, and the terminal side synchronously collects basic video features; the terminal environment perception module collects initial environmental features and completes basic data initialization.

[0058] Step 2: Real-time collection of user video interaction dynamic behavior:

[0059] When a user plays a video, the user behavior perception module on the terminal side captures various interactive behaviors in real time. After the edge preprocessing module removes abnormal data, the data is pushed to the server at a fixed frequency.

[0060] Step 3: Multi-dimensional data quantization, encoding, and fusion:

[0061] The server-side data receiving and fusion module integrates terminal dynamic behavior data, environmental data, and basic video features with cloud-based static attribute data and full video features, converting them into numerical features and generating a standardized user-video interaction data matrix.

[0062] Step 4: Calculate demand preferences based on the computational model:

[0063] The server calls the user video demand comprehensive preference calculation model, taking the user-video interaction data matrix as input, and completes three-layer quantization calculation to obtain the comprehensive preference. The calculation formula of the model is as follows:

[0064] ;

[0065] ;

[0066] ;

[0067] ;

[0068] In the overall formula of the model:

[0069] P(i,j) represents the overall preference of user i for video j;

[0070] α is the static preference weight coefficient, with a value of [0.2, 0.4], which is adaptively adjusted by the model's dynamic update module based on the amount of user behavior data (for example, α is 0.4 for new users and 0.2 for old users).

[0071] Let represent the basic static attribute preference of user i for video j, with a value range of [0,1]. The larger the value, the higher the matching degree between the user's inherent attributes and the video features.

[0072] Let P2(i,j)∈[0,1] represent the dynamic behavior preference of user i for video j. The larger the value, the more active the user's real-time interactive behavior towards the video and the higher the demand.

[0073] λ(t) is a time decay correction factor, used to correct the impact of long-term dynamic behavior on users' real-time needs. The closer the time of the user behavior, the stronger its ability to characterize the current needs, and the higher its weight; the further back in time, the lower its weight.

[0074] In the formula for calculating the basic preference of the static attribute of the model:

[0075] n is the total number of matching dimensions between user static attributes and video features (n≥5, including age, occupation, interest tags, etc.).

[0076] k is the k-th matching dimension;

[0077] The weight coefficient for the k-th matching dimension. The weights are set according to the degree of influence of the dimension on video demand (e.g., the weight of user-selected interest tags is higher than that of the geographic dimension).

[0078] S(i,k) is the standardized feature value of user i in the k-th matching dimension (value [0,1]), which is obtained by standardizing the user's static attributes after one-hot encoding or tag encoding embedding (e.g., if the user's interest tag is "technology", then the technology dimension S(i,k)=1, and other tag dimensions S(i,k)=0).

[0079] T(j,k) is the standardized feature value of video j in the k-th matching dimension (value [0,1]). The feature value is proportionally allocated to cross-class videos and is obtained by standardizing the video features after quantization encoding (e.g., if the category of video j is "technology", then the technology dimension T(j,k)=1, and the other category dimensions T(j,k)=0; if the video is cross-class, such as "technology + life", then the T(j,k) of both dimensions is 0.5).

[0080] In the formula for calculating dynamic behavioral preferences:

[0081] m represents the total number of user interaction behaviors with the video (m=8, including 8 types of behaviors such as click, completion rate, like, favorite, comment, replay, fast forward or skip, and swipe away).

[0082] The weight coefficients for the m-th type of interaction behavior (positive behavior) Positive values ​​indicate positive behavior. The value is negative, representing the collection behavior in positive behavior. The highest value is the swipe action in the negative behavior. (Minimum value)

[0083] B(i,j,m) is the standardized behavior value (value [0,1]) of user i's m-th interaction behavior with video j. The standardization rules for the interaction behavior are as follows: the completion rate is the ratio of the actual viewing time to the total video duration; for click, like, favorite, comment, and replay behaviors, the value is 1 if the behavior exists, and 0 if it does not; the fast-forward or skip behavior is the ratio of the fast-forward or skip duration to the total video duration; the swipe-away behavior is 1 if it exists, and 0 if it does not.

[0084] denominator This is a normalization factor to ensure that the final calculation result falls within the [0,1] interval, thus avoiding distortion of the result due to excessively small weights.

[0085] In the formula for calculating the time decay correction factor:

[0086] e is a natural constant;

[0087] γ is the time decay coefficient (value [0.001, 0.1]), which is adjusted by the model dynamic update module according to the frequency of user behavior changes. The faster the user behavior changes, the larger γ is and the faster the decay; the more stable the user behavior is, the smaller γ is and the slower the decay.

[0088] Δt is the action time difference, which is the difference between the current time t0 and the time t when the action occurs. If the action occurs at the current time, there is no decay.

[0089] Step 5: Dynamically iteratively update the model:

[0090] After a user generates a new interaction, the server updates the dynamic behavior weight parameters (increasing the weight of high-frequency behaviors and decreasing the weight of occasional behaviors) and the time decay coefficient; the update cycle is adjusted according to the user's activity level, for example, it can update once every 5 minutes for highly active users and once every hour for inactive users.

[0091] Step 6: Execution of differentiated data collection strategy:

[0092] The server generates and distributes the collection strategy, and the terminal executes the strategy accordingly. Specifically, the collection frequency and feature dimensions of the top 30% of video types with the highest preference are increased; only core data is collected for low-bandwidth / low-computing-power terminals; real-time high-frequency collection is carried out during high-activity periods, and low-frequency timed collection is carried out during low-activity periods; active collection of video types with a preference of 0 is stopped.

[0093] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance, unless otherwise expressly specified and limited. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data information acquisition system, characterized in that, Includes a terminal-side data acquisition unit and a server-side processing unit. The terminal-side acquisition unit is deployed on the user's video playback terminal and includes: The user behavior perception module is used to collect dynamic interaction data between users and videos in real time. The terminal environment perception module is used to collect terminal hardware characteristics, network environment characteristics, and playback environment characteristics. The lightweight video feature acquisition module is used to collect basic features that can be extracted from the video terminal and complete standardized encoding. The edge preprocessing module is used to deduplicatize, remove outliers, and calibrate timestamps on the collected multi-dimensional data. It also pushes the preprocessed data to the server processing unit at a fixed frequency and caches local short-term behavioral data. The server-side processing unit is deployed on a cloud server and includes: The data receiving and fusion module is used to receive preprocessed data from the acquisition units on each terminal side; The user static attribute management module is used to store and update static attribute data when a user registers. The full video feature extraction module is used to extract deep features from videos and complete quantization encoding based on computer vision and natural language processing technologies. The demand preference calculation module is used to perform quantitative calculations of user preferences for various types of videos; The model dynamic update module is used to iteratively update the weight parameters and decay coefficients of the calculation model in real time based on the user's subsequent video interaction behavior, so as to adapt to changes in user behavior. The intelligent acquisition strategy generation module is used to generate differentiated video data acquisition strategies based on user needs and preferences and terminal environment characteristics, and then distribute them to the terminal-side acquisition unit. The data storage module is used to store user-related data, video-related data, model parameters, and iteration records.

2. A data information acquisition method, based on the data information acquisition system of claim 1, characterized in that, Includes the following steps: Step 1: Initial collection of user and video basic data. The user static attribute management module collects and stores static attribute data during the user registration phase and assigns a unique user ID to each user. The video full feature extraction module completes full feature collection and quantization encoding of all videos in the platform to generate video feature vectors. The video lightweight feature collection module collects basic video features simultaneously. The terminal environment perception module on the terminal side collects the initial environmental features of the terminal. Step 2: Real-time collection of user video interaction dynamic behavior. During the video playback process, the user behavior perception module on the terminal side captures all user video interaction behaviors in real time with a minimum collection granularity of 100ms, records the behavior type, behavior occurrence timestamp, and associated video ID. After the edge preprocessing module removes abnormal behaviors, the data is pushed to the data receiving and fusion module. Step 3: Multi-dimensional data collection, quantization, encoding, and fusion. The data receiving and fusion module merges the dynamic behavior data, terminal environment data, and lightweight video feature data collected by the terminal-side collection unit with the user static attribute data and full video feature data stored by the server-side processing unit. Through one-hot encoding or tag encoding embedding, the non-numerical data is converted into numerical features, generating a standardized user-video interaction data matrix. Step 4: Quantitative calculation of user video demand preference based on the calculation model. The demand preference calculation module calls the user video demand comprehensive preference calculation model. With the user-video interaction data matrix as input, it calculates the user static attribute basic preference and the user dynamic behavior preference respectively. The dynamic behavior preference is corrected by the time decay factor and then fused to obtain the user video demand comprehensive preference. Step 5: Calculate the dynamic iterative update of the model. After a user generates a new video interaction behavior, the server-side model dynamic update module uses the new behavior data as a feedback sample to update the behavior weight parameters of the dynamic behavior preference and the decay coefficient of the time decay factor. The model update cycle is dynamically adjusted according to the user's behavior activity. Step 6: Execution of differentiated video data acquisition strategy based on preference. The intelligent acquisition strategy generation module generates and distributes differentiated acquisition strategies to the terminal acquisition unit based on the user's video needs, preferences, and terminal environment characteristics. The terminal then executes video data acquisition according to the strategy.

3. The data information acquisition method according to claim 2, characterized in that: The user video demand comprehensive preference calculation model adopts a three-layer quantization fusion architecture, and the calculation formula of the user video demand comprehensive preference calculation model is as follows: ； In the formula, P(i,j) is the overall preference of user i for video j, α is the static preference weight coefficient, P1(i,j) is the basic static attribute preference of user i for video j, P2(i,j) is the dynamic behavior preference of user i for video j, and λ(t) is the time decay correction factor.

4. The data information acquisition method according to claim 3, characterized in that: The formula for calculating the static attribute basic preference degree P1(i,j) is as follows: ； In the formula, n is the total number of matching dimensions between user static attributes and video features, and k is the k-th matching dimension. Let S(i,k) be the weight coefficient of the k-th matching dimension, S(i,k) be the standardized feature value of user i in the k-th matching dimension, and T(j,k) be the standardized feature value of video j in the k-th matching dimension.

5. The data information acquisition method according to claim 3, characterized in that: The aforementioned dynamic behavioral preference The calculation formula is: ； In the formula, m represents the total number of user interaction types with the video. Let B(i,j,m) be the weight coefficient for the m-th type of interaction behavior, and let B(i,j,m) be the standardized behavior value of user i's m-th type of interaction behavior with video j. The denominator is... This is the normalization factor.

6. The data information acquisition method according to claim 3, characterized in that: The formula for calculating the time decay correction factor λ(t) is as follows: ； In the formula, e is the natural constant, γ is the time decay coefficient, and Δt is the behavior time difference.

7. The data information acquisition method according to claim 2, characterized in that: The terminal-side acquisition unit and the server-side processing unit adopt a bandwidth-adaptive encrypted transmission protocol. When the bandwidth is lower than the threshold, only core behavior and feature data are transmitted; when the bandwidth is sufficient, all data is transmitted, and all data is processed with end-to-end encryption.

8. The data information acquisition method according to claim 2, characterized in that: The core execution rules of the differentiated video data acquisition strategy described in step 6 include: For the top 30% of video types in terms of user preference, increase the frequency of feature collection for this type of video and increase the dimensions of deep feature collection. For terminals with low bandwidth and low computing power, only basic video features and core user behaviors are collected; Enable real-time high-frequency data collection during peak user video playback periods, and enable low-frequency timed data collection during off-peak periods. For video types with zero user preference, stop actively capturing them and only retain the ability to passively capture them.

9. A data information acquisition method according to claim 2, characterized in that: The standardized rules for the interaction behavior include: The completion rate is calculated as the ratio of actual viewing time to the total video duration. Click, like, favorite, comment, replay actions are assigned a value of 1 if the action occurs, and a value of 0 if the action does not occur. Fast-forwarding or skipping is calculated as the ratio of the fast-forwarding or skipping duration to the total video duration. The value of the swipe action is 1 if it exists, and 0 if it does not.

Images