Negative feedback monitoring method and system applied to media data

Abstract

The invention is suitable for the technical field of data processing, and provides a negative feedback monitoring method applied to media data, which comprises the following steps: S1, media data acquisition and fusion: acquiring media data and user interaction behavior data from different channels, and constructing a unified user session sequence model, performing time sequence association on the multi-modal data in the single user session; and S2, multi-modal negative feedback feature fusion: processing the media data, extracting negative feedback features of text, vision, audio and user behavior modals, and analyzing an association relationship among the negative feedback features of different modals by using a cross-modal associator. According to the invention, the operation range of the user session sequence model is limited to be a single continuous access period, so that the problem of cross-session data fragmentation is effectively avoided.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a negative feedback monitoring method and system for media data. Background Technology

[0002] In the digital information dissemination environment, media formats such as video, live streaming, and social media content have become the core carriers of user interaction. Platform operators need to continuously monitor user feedback to optimize the content ecosystem and service experience.

[0003] Negative feedback behavior, as a key indicator of user satisfaction, encompasses explicit behaviors such as negative comments, reporting actions, quitting midway, and unfollowing, as well as implicit behaviors such as swiping to skip, muting playback, and quickly switching between functions. These behaviors collectively reflect users' dissatisfaction with the content or functions.

[0004] However, current media data negative feedback monitoring systems have significant shortcomings. The ability to dynamically track the propagation path of negative feedback is severely inadequate. Existing solutions mostly rely on static statistical models to generate post-event reports, preventing administrators from observing in real time how negative public opinion spreads among users through content recommendations, social sharing, and other functional modules, thus delaying the optimal window for risk intervention. At the prediction and control level, technical means are limited to simple threshold-triggered alarm mechanisms, failing to establish quantitative correlation models between implicit negative feedback behaviors and subsequent explicit negative events, resulting in weak predictability of future negative feedback trends. The formulation of control strategies relies on preset rules, lacking intelligent decision-making mechanisms dynamically generated based on real-time data. This prevents precise intervention targeting content entities, functional modules, or user profiles, ultimately leading to slow platform response and limited governance effectiveness. These problems highlight the systemic deficiencies of existing technologies in data integration, feature correlation, dynamic tracing, and intelligent response. Summary of the Invention

[0005] The purpose of this invention is to provide a negative feedback monitoring method and system for media data, aiming to solve the problem that in existing negative feedback monitoring methods for media data, the formulation of control strategies relies on preset rules and lacks an intelligent decision-making mechanism that is dynamically generated based on real-time data.

[0006] This invention is implemented as follows: a negative feedback monitoring method for media data, comprising the following steps:

[0007] Step S1: Collect media data and user interaction behavior data from different channels, and construct a unified user session sequence model to perform time-series correlation of multimodal data within a single user session;

[0008] Step S2: Process the media data, extract negative feedback features of text, visual, audio and user behavior modalities, and use a cross-modal correlator to analyze the correlation between negative feedback features of different modalities;

[0009] Step S3: Based on the fused negative feedback features, construct a dynamic negative feedback source graph, where the nodes of the graph include content entities, functional entities and user profile tags, and the edges of the graph represent the associations and association strengths between nodes caused by negative feedback.

[0010] Step S4: Based on historical negative feedback data and the negative feedback source map, use a prediction model to predict the negative feedback trend;

[0011] Step S5: Based on the negative feedback source map and prediction results, the policy rule engine automatically generates and executes control strategies to intervene in related content, functions and recommendation systems.

[0012] Furthermore, in step S1, the user session sequence model is used to associate and serialize the video viewing, comment posting, interactive operations, and mid-exit behavior generated by the user during a single continuous access on the timeline.

[0013] Furthermore, in step S2, when a preset negative feature is identified in the first mode, a simultaneous search is performed to determine whether a corresponding negative feedback feature exists in the second mode that is in a temporal proximity.

[0014] The correlation strength between the negative feedback features of the first mode and the second mode is calculated, and the correlation strength is used to verify the accuracy of feature recognition; wherein the calculation process includes statistical hypothesis testing and machine learning model.

[0015] Furthermore, in step S3, the negative feedback source graph is a dynamic graph, whose node and edge weights are updated in real time as new data flows in; the graph is used to visualize the propagation path of negative public opinion and locate the root cause of the content that leads to negative user feedback.

[0016] Furthermore, in step S4, the prediction model includes a time series model and a sequence learning model, used to establish a quantitative prediction relationship between implicit negative feedback and explicit negative feedback, specifically:

[0017] Using one or more implicit negative feedback data from users' cessation behavior, avoidance behavior, and silence behavior as input, predict the changing trends of the number of negative comments, reports, or unfollows within a specific future time window.

[0018] Furthermore, in step S5, the rules of the policy rule engine are dynamically generated based on the node relationships revealed by the negative feedback source graph, and the rules are bound to the causal relationship chains in the graph.

[0019] Furthermore, in step S5, the control strategy includes at least one of the following:

[0020] When the negative feedback source map shows that negative feedback is strongly associated with a specific content ID, it triggers automatic downgrading of the content or sends it to the manual review queue.

[0021] When the negative feedback source map shows a strong correlation between a large number of ad avoidance behaviors and a specific ad creative, an adjustment instruction is automatically sent to the ad delivery system.

[0022] When the predictive model outputs a high-risk warning with negative feedback, it automatically pauses the push of related controversial content and generates a pre-set response strategy report for the customer service team.

[0023] The present invention also provides a negative feedback monitoring system for media data, used to implement the aforementioned negative feedback monitoring method for media data, the system comprising:

[0024] The data acquisition module is used to collect media data and user interaction behavior data from different channels, and to build a unified user session sequence model to correlate multimodal data within a single user session in a time sequence.

[0025] The feature fusion module is used to process the media data, extract negative feedback features of text, visual, audio and user behavior modalities, and use a cross-modal correlator to analyze the correlation between negative feedback features of different modalities.

[0026] The source graph construction module is used to construct a dynamic negative feedback source graph based on the fused negative feedback features. The nodes of the graph include content entities, functional entities and user profile tags, and the edges of the graph represent the associations and association strengths between nodes caused by negative feedback.

[0027] The predictive insight module is used to predict negative feedback trends based on historical negative feedback data and the negative feedback source map using a predictive model.

[0028] The intelligent control module is used to automatically generate and execute control strategies based on the negative feedback source map and prediction results through the strategy rule engine, and to intervene in related content, functions and recommendation systems.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] This invention effectively avoids the problem of cross-session data fragmentation by limiting the operational scope of the user session sequence model to a single continuous access period. Based on this, the model serializes behaviors such as video viewing, commenting, interactive operations, and mid-session exit according to their actual chronological order. This temporal correlation mechanism particularly highlights the logic of generating negative experience chains based on behavioral sequences. For example, when the system detects that a user immediately exits a video after watching it and subsequently posts a negative comment, it can accurately capture the potential causal relationship reflected in this event sequence. Simultaneously, by unifying the modeling of explicit and implicit behaviors, it achieves accurate reconstruction of the complete path of the user's negative experience, providing the necessary temporal foundation for subsequent cross-modal feature fusion and source tracing analysis.

[0031] This invention solves the problems of static strategy execution and insufficient scenario adaptation in intelligent control, and realizes dynamic adaptation capability for different negative feedback scenarios. Based on the real-time correlation analysis of the negative feedback source map and the early warning signal of the prediction model, it can automatically trigger specific operation instructions such as content demotion, advertising adjustment or suspension of high-risk content, effectively curbing the spread of negative feedback in the content dissemination, advertising and controversial content push.

[0032] In this invention, the rules of the strategy rule engine can generate and bind causal relationship chains in real time based on the dynamic characteristics of the negative feedback source graph, thereby achieving precise intervention in the negative feedback propagation path. This design effectively solves the problem that traditional static rules cannot respond to real-time changes in a timely manner, significantly improving the pertinence and effectiveness of the control strategy. Attached Figure Description

[0033] Figure 1 A schematic diagram illustrating the steps of a negative feedback monitoring method applied to media data;

[0034] Figure 2 This is a schematic diagram of a negative feedback monitoring system applied to media data. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0036] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0037] like Figure 1 As shown, an embodiment of the present invention provides a negative feedback monitoring method for media data, comprising the following steps:

[0038] Step S1: Collect media data and user interaction behavior data from different channels, and construct a unified user session sequence model. The multimodal data within a single user session is temporally correlated. The user session sequence model is based on the user's continuous activities on the platform. When the interval between adjacent operations exceeds 30 minutes, it is considered that the session has ended. Video viewing, commenting, interactive operations, and mid-exit behavior within a single session are correlated and serialized on the timeline.

[0039] Step S1 achieves media data collection and fusion. By constructing a unified user session sequence model, multimodal data such as video viewing, comment posting, interactive operations, and mid-term exit are associated and serialized on the timeline. This ensures that implicit negative feedback behaviors and explicit behaviors are completely restored within a single user session, providing a time-consistent data foundation for subsequent analysis, thereby solving the problem of single data collection dimensions.

[0040] Step S2: Process the media data, extract negative feedback features of text, visual, audio and user behavior modalities, and use a cross-modal correlator to analyze the correlation between negative feedback features of different modalities;

[0041] Specifically, when a preset negative feature is identified in the first modality, the system simultaneously searches for the existence of a corresponding negative feedback feature in the second modality within a time window of ±T (T can be set according to the business scenario, such as ±5 seconds).

[0042] The association strength between different modal features is calculated using a cross-modal attention mechanism or a Transformer-based fusion model, and the significance of the association is verified using statistical tests (e.g., p-value < 0.05).

[0043] Step S2 processes the media data to extract negative feedback features from text, visual, audio, and user behavior modalities. It then uses a cross-modal correlator to analyze the correlation between negative feedback features from different modalities, enabling the negative feedback signals from different modalities to corroborate each other and accurately pinpoint the root cause of the negative feedback, thereby solving the problem of isolated multimodal analysis.

[0044] Step S3: Based on the fused negative feedback features, construct a dynamic negative feedback source graph. The nodes of the graph include content entities, functional entities, and user profile tags. The edges of the graph represent the associations and association strengths between nodes caused by negative feedback. The negative feedback source graph adopts a time decay-based weight update mechanism. When new data flows in, the weights of nodes and edges are updated in real time through an exponentially weighted moving average algorithm, with a decay factor α∈[0.8,0.95], to ensure that the graph reflects the latest situation while maintaining historical continuity.

[0045] Step S3 constructs a dynamic negative feedback source tracing graph based on the fused negative feedback features. Its nodes cover content entities, functional entities, and user profile tags. The edges represent the associations and association strengths between nodes caused by negative feedback, and the weights are updated in real time with new data. By visually displaying the path of negative public opinion dissemination, the root cause of content or functional defects can be located, thereby solving the problem of lack of dynamic source tracing.

[0046] Step S4: Based on historical negative feedback data and the negative feedback source map, use a prediction model to predict the negative feedback trend;

[0047] Step S4 uses historical negative feedback data and negative feedback source map to establish a quantitative prediction relationship between implicit and explicit negative feedback using a prediction model, thereby achieving a forward-looking insight into future negative feedback trends and solving the problem of insufficient prediction ability.

[0048] The implicit negative feedback data is quantified in the following way:

[0049] 1. Suspension behavior: The proportion of sessions with a completion rate of less than 20%;

[0050] 2. Avoidance behavior: The frequency of operations where the swipe skip speed is more than twice the normal value;

[0051] 3. Silent Behavior: The proportion of historically active users who suddenly stop interacting for more than 7 days.

[0052] Based on the above quantitative indicators, the LSTM-ARIMA hybrid model is used to predict the changing trends of the number of negative comments and reports in the next 24 hours.

[0053] Step S5: Based on the negative feedback source map and prediction results, the policy rule engine automatically generates and executes control strategies to intervene in related content, functions and recommendation systems.

[0054] The policy rule engine dynamically generates rules based on the node association strength in the negative feedback source graph, specifically including:

[0055] 1) When the node association strength exceeds the threshold θ1, the rule template filling mechanism is triggered;

[0056] 2) Extract causal relationship chain features through graph neural networks to generate regulation rules adapted to the current graph structure;

[0057] 3) All generated rules must pass a confidence test (confidence > 80%) before they can be executed.

[0058] The threshold θ1 was determined through historical data verification, with an initial value of 0.7, which can be adjusted according to business needs.

[0059] Step S5 automatically generates and executes control strategies based on the negative feedback source map and prediction results through the strategy rule engine. This intervenes in content, functions, and the recommendation system, dynamically binding rules to the causal relationship chains in the graph, ensuring timely and effective interventions, thus addressing the issue of static control strategies. Various technical features work closely together: the time-series correlation data in S1 provides input for S2; the cross-modal feature fusion in S2 supports the dynamic graph construction in S3; the negative feedback source map in S3, together with historical data, drives the prediction in S4; and the prediction results in S4, in conjunction with the graph in S3, guide the intelligent control in S5, forming a complete chain from data perception to action execution, thereby systematically solving the challenges of accuracy and real-time performance in negative feedback monitoring.

[0060] In practical applications, media data acquisition and fusion can be understood as the process of obtaining data from multiple independent sources and integrating it into a unified framework. For example, video playback records, comment data, and user actions can be collected using log file parsing tools, and timestamp alignment algorithms can be used to correlate these data along a timeline, thereby forming a complete user session sequence model. Its main purpose is to provide a consistent data foundation for subsequent analysis.

[0061] Furthermore, multimodal negative feedback feature fusion refers to extracting features that reflect users' negative emotions from multiple types of data and analyzing the potential relationships between these features using specific technical means. For example, natural language processing techniques can be used to identify negative emotional words in text, combined with image recognition algorithms to detect abnormal elements in visual content, while behavioral pattern analysis can be used to discover users' mid-game exit behavior. The correlations between these features can be verified through statistical methods or machine learning models to improve the accuracy of feature recognition.

[0062] Specifically, constructing a negative feedback source graph is a technical means of representing negative feedback information graphically. Nodes can include specific media content, platform functional modules, and user attribute tags, while edges represent the interactions and strengths between these nodes caused by negative feedback. For example, the propagation path of negative public opinion can be displayed in the form of a weighted directed graph, where the weights can be dynamically adjusted based on the frequency or scope of impact of negative feedback events.

[0063] Furthermore, negative feedback prediction and insight can be achieved through various modeling methods. For example, regression analysis can be used to establish a quantitative relationship between implicit and explicit negative feedback, or deep learning models can be used to capture the time-series characteristics of negative feedback trends. The aim is to predict potential future changes in negative feedback by learning from historical data.

[0064] The innovation of this application lies in its comprehensive design, which addresses the shortcomings of existing technologies, such as single data collection dimensions, isolated multimodal analysis, and lack of dynamic tracing capabilities. Specifically, it achieves temporal correlation of multimodal data by constructing a unified user session sequence model, improves the accuracy of negative feedback feature identification by utilizing a cross-modal correlator, and enables the visual tracking of negative public opinion dissemination paths through a dynamic negative feedback tracing graph. These improvements collectively constitute a closed-loop management system from data collection to intelligent control, effectively addressing the core challenges in media data negative feedback monitoring.

[0065] In a preferred embodiment of the present invention, in step S1, the user session sequence model is used to correlate and serialize the video viewing, comment posting, interactive operations, and mid-session exit behaviors generated by a user during a single continuous access period on a timeline. The user session sequence model is an analysis tool capable of temporally integrating user behavior data, which can be implemented using timestamp-based behavior log aggregation technology or an event-driven state machine model. The purpose of introducing this model is to ensure that multimodal behavior data can be accurately correlated on a unified timeline, thereby providing a reliable data foundation for subsequent negative feedback tracing.

[0066] In a preferred embodiment of the present invention, in step S2, when a preset negative feature is identified in the first mode, a simultaneous search is performed to determine whether a corresponding negative feedback feature exists in a second mode that is in a temporal proximity.

[0067] Specifically, the first modality refers to a specific modality in the media data, such as text, visual, or audio, which can be implemented using natural language processing, image recognition, or audio analysis techniques. The pre-defined negative features can be understood as pre-defined patterns or attributes that may trigger negative user feedback, aiming to accurately locate potential negative factors. Furthermore, the second modality refers to another modality different from the first, designed to enhance the reliability of feature recognition through cross-modal analysis. In addition, the time-proximity synchronous retrieval mechanism, based on the continuity of user behavior over time, limits the time range to ensure a close correlation between the features of the two modalities, thereby avoiding misjudgments caused by time misalignment.

[0068] The correlation strength between the negative feedback features of the first and second modalities is calculated using statistical hypothesis testing or machine learning models to verify the accuracy of feature recognition or to discover potential causal relationships. Specifically, statistical hypothesis testing can be a t-test or a chi-square test to assess the significance of the association; while machine learning models can employ algorithms such as deep neural networks or random forests to mine complex patterns from historical data, aiming to distinguish between true causal relationships and accidental associations.

[0069] In this embodiment, during the multimodal negative feedback feature fusion process, when a preset negative feature is detected in the first modality, the system immediately triggers a synchronous retrieval operation for the second modality. This process is based on the principle of temporal proximity, ensuring that the features of the two modalities are strongly correlated in the user session. For example, in a video viewing scenario, if the visual modality identifies uncomfortable visual features, the system will simultaneously check the behavioral modality within the same time window, such as exiting the operation midway, thereby capturing the user's immediate negative feedback reaction. Based on this, the correlation strength between the negative feedback features of the first and second modalities is quantitatively analyzed through statistical hypothesis testing or machine learning models, which can effectively verify the accuracy of feature recognition and reveal the hidden causal chain. This design not only enhances the timeliness of feature association but also restores the true temporal logic of the user's negative experience, providing a solid basis for tracing the source of negative feedback. At the same time, this scheme, combined with the aforementioned multimodal negative feedback feature fusion method, further improves the accuracy and reliability of cross-modal analysis, solves the problem of inaccurate feature recognition caused by time mismatch or lack of statistical verification, and thus achieves dynamic optimization of negative feedback tracing.

[0070] In a preferred embodiment of the present invention, in step S3, the negative feedback source graph is a dynamic graph, the weights of its nodes and edges are updated in real time as new data flows in; the graph is used to visualize the propagation path of negative public opinion and locate the root cause of the content or functional defects that lead to negative user feedback.

[0071] In this embodiment, in practical applications, a dynamic graph refers to a graphical model that continuously adjusts its internal structure and parameters based on changes in input data. It can be implemented using graph neural networks based on incremental learning or streaming data processing frameworks. The weights of nodes and edges are numerical indicators reflecting the strength of associations between content entities, functional entities, and user profile tags. These weights can be dynamically adjusted using sliding window statistical methods or exponentially weighted moving average algorithms to ensure the graph promptly reflects the latest negative feedback state. Visualization refers to the technical means of presenting complex data relationships to users through an intuitive graphical interface. This can be achieved using force-directed graph layout algorithms or timeline animation technology, aiming to help managers quickly understand the dynamics of negative public opinion dissemination.

[0072] Specifically, in the aforementioned negative feedback monitoring method applied to media data, real-time tracking of the negative feedback propagation path is achieved by constructing a dynamic graph. First, the system continuously receives media data and user interaction behavior data from different channels. After preprocessing, this data is transformed into incremental information that can be used to update the graph. When new data flows in, the system automatically triggers a weight update mechanism, dynamically adjusting the correlation strength by calculating the impact of the new data on existing nodes and edges. This real-time update mechanism ensures that the graph accurately reflects the current negative feedback state, avoiding analytical biases caused by data lag in traditional static graphs.

[0073] Building upon this foundation, a visual interface displays the propagation path of negative public opinion, allowing administrators to clearly observe how negative feedback spreads from initial content nodes to specific user groups. Simultaneously, the system can accurately pinpoint the specific content or functional module triggering negative user feedback based on real-time updated graph data. This design not only solves the problem of ambiguous root cause identification due to untimely graph updates but also provides accurate intervention targets for subsequent intelligent control. By integrating dynamic graphs with multimodal negative feedback characteristics, a complete solution is formed capable of continuously capturing the evolution of negative feedback, significantly improving the platform's response speed and processing efficiency to negative feedback events.

[0074] In a preferred embodiment of the present invention, in step S4, the prediction model is a time series model or a sequence learning model, used to establish a quantitative prediction relationship between implicit negative feedback and explicit negative feedback, specifically as follows:

[0075] Using one or more implicit negative feedback data from users' cessation behavior, avoidance behavior, and silence behavior as input, predict the changing trends of the number of negative comments, reports, or unfollows within a specific future time window.

[0076] In this embodiment, specifically, a time series model refers to a mathematical model that can capture the patterns of data changes over time. It can be implemented using ARIMA models, LSTM networks, or other algorithms suitable for processing time-series data. A sequence learning model, on the other hand, is a machine learning model specifically designed for analyzing serialized data. It can be implemented using structures such as the Transformer architecture and RNN variants. The purpose of introducing these models is to reveal the intrinsic relationship between implicit and explicit negative feedback by modeling the temporal dynamics of user behavior, thereby providing technical support for the forward-looking prediction of negative feedback trends.

[0077] In detail, the above scheme first extracts implicit negative feedback data, such as user abrupt termination, avoidance, and silence, by performing temporal correlation on multimodal data in the user session sequence model. This data is then fed into a time series model or sequence learning model, which predicts future negative comments, reports, or unfollows based on the learning results of historical data. Because time series models and sequence learning models can effectively capture the continuous evolution of user behavior, they can generate accurate trend predictions based on early implicit signals before users generate explicit negative feedback. Furthermore, this prediction mechanism not only solves the lag problem caused by relying on static thresholds in traditional methods but also provides timely and accurate decision-making basis for subsequent intelligent regulation by quantifying the relationship between implicit and explicit negative feedback. Based on this, combined with a dynamically updated negative feedback source map, the reliability and practicality of the prediction results are further enhanced, thereby achieving efficient intervention in negative feedback trends.

[0078] Through the above technical solutions, the system can identify the risk evolution path in advance based on negative experiences that users do not directly express, providing the platform with a forward-looking negative feedback management capability, and significantly improving the efficiency of user experience optimization and content risk management.

[0079] In a preferred embodiment of the present invention, in step S5, the rules of the strategy rule engine are dynamically generated based on the node association relationships revealed by the negative feedback tracing graph, and the rules are bound to the causal relationship chains in the graph.

[0080] In this embodiment, the strategy rule engine refers to an intelligent tool capable of automatically generating control strategies based on input data. It can be implemented using algorithms based on graph structure analysis, such as extracting key paths and node weight changes in the negative feedback source graph through graph traversal algorithms, thereby generating corresponding rules. Its purpose is to ensure that the control strategy can respond in real-time to changes in the negative feedback graph, avoiding intervention lag due to fixed rules.

[0081] Furthermore, a negative feedback source graph is a dynamically updated graph structure used to represent the negative feedback relationships and their strengths among content entities, functional entities, and user profile tags. It can be implemented as a weighted directed graph, where node weights and edge weights are adjusted in real time as new data flows in. The purpose of this design is to provide accurate data support for the policy rule engine, enabling its generated rules to reflect the latest negative feedback propagation paths and root causes.

[0082] In detail, during actual operation, the strategy rule engine first extracts the relationships between nodes and their dynamic changes from the negative feedback source graph. This information includes, but is not limited to, the causal chain between content entities and negative user behavior, and the strength of the association between functional entities and specific negative feedback patterns. Based on this information, the strategy rule engine dynamically generates control rules and binds the rules to specific causal chains in the graph. For example, when the negative feedback source graph shows a strong association between negative feedback and a certain content ID, the rule will automatically trigger a downgrade of that content or send it to a manual review queue. This mechanism ensures the targetedness and timeliness of the control strategy, avoiding the response delay problem caused by traditional static rules.

[0083] Furthermore, because the negative feedback source map can visually demonstrate the propagation path of negative public opinion and pinpoint the root causes of problems, the rules generated by the strategy rule engine can be directly applied to these paths and root causes. For example, when the predictive model outputs a high-risk warning for negative feedback, the rule will automatically pause the push of relevant controversial content and generate a pre-set response strategy report for the customer service team. This approach not only improves the efficiency of regulation but also enhances the system's intelligence level.

[0084] In a preferred embodiment of the present invention, in step S5, the control strategy includes at least one of the following:

[0085] When the negative feedback source map shows that negative feedback is strongly associated with a specific content ID, it triggers automatic downgrading of the content or sends it to the manual review queue.

[0086] When the negative feedback source map shows a strong correlation between a large number of ad avoidance behaviors and a specific ad creative, an adjustment instruction is automatically sent to the ad delivery system.

[0087] When the predictive model outputs a high-risk warning with negative feedback, it automatically pauses the push of related controversial content and generates a pre-set response strategy report for the customer service team.

[0088] In this embodiment, specifically, the negative feedback source graph refers to a dynamic graph constructed based on the fused negative feedback features. Its nodes include content entities, functional entities, and user profile tags, and edges represent the associations and their strengths between nodes caused by negative feedback. In practical applications, the negative feedback source graph can reflect the inflow of new data by updating the weights of nodes and edges in real time, thereby visually displaying the propagation path of negative public opinion and locating the root causes of negative user feedback or functional defects.

[0089] The prediction model can be a time series model or a sequence learning model, used to establish a quantitative predictive relationship between implicit and explicit negative feedback. In practice, it can use one or more implicit negative feedback data from user cessation behavior, avoidance behavior, and silence behavior as input to predict the changing trends of the number of negative comments, reports, or unfollows within a specific future time window.

[0090] In detail, during the intelligent control process, when the negative feedback source map reveals a strong correlation between negative feedback and a specific content ID, the content is automatically downgraded or manually reviewed based on the strength of this correlation. This avoids the generalization of problematic content, ensuring that intervention directly targets the root cause of the content entity and improves the accuracy of content governance. When the negative feedback source map detects a large number of ad avoidance behaviors that are highly correlated with specific ad creatives, it automatically sends adjustment instructions to the ad delivery system. This mechanism can quickly respond to implicit negative feedback from users regarding ad creatives, dynamically linking behavioral data with the ad system to optimize the real-time performance of ad delivery strategies. When the predictive model outputs a high-risk warning for negative feedback in advance, it automatically pauses the push of controversial content and generates a customer service response report. This achieves preventative intervention based on predicted trends, reducing the actual occurrence of negative feedback, while providing pre-set support solutions for customer service and strengthening the platform's proactive control capabilities over potential risks. Overall, these strategies, by combining the dynamic correlation analysis of the negative feedback source map with the forward-looking judgment of the predictive model, enable control measures to adapt to specific scenarios such as content management, ad optimization, and risk prevention, significantly improving the platform's response speed and processing accuracy to negative events.

[0091] like Figure 2 As shown, a negative feedback monitoring system for media data, applied to the negative feedback monitoring method for media data in the above embodiments, includes:

[0092] The media data acquisition and fusion module is used to collect media data and user interaction behavior data from different channels, and to build a unified user session sequence model to correlate multimodal data within a single user session in a time sequence.

[0093] The multimodal negative feedback feature fusion module is used to process the media data, extract negative feedback features of text, visual, audio and user behavior modalities, and use a cross-modal correlator to analyze the correlation between negative feedback features of different modalities.

[0094] The negative feedback source graph construction module is used to construct a dynamic negative feedback source graph based on the fused negative feedback features. The nodes of the graph include content entities, functional entities and user profile tags, and the edges of the graph represent the associations and association strengths between nodes caused by negative feedback.

[0095] The negative feedback prediction and insight module is used to predict negative feedback trends based on historical negative feedback data and the negative feedback source map using a prediction model.

[0096] The intelligent control module is used to automatically generate and execute control strategies based on the negative feedback source map and prediction results through the strategy rule engine, and to intervene in related content, functions and recommendation systems.

[0097] The core innovation of this application lies in combining the media data acquisition and fusion module with the multimodal negative feedback feature fusion module using a unified user session sequence model and cross-modal correlator, while introducing a dynamic negative feedback tracing graph and a prediction model. This solves the problems in existing technologies such as the ineffective acquisition of implicit negative feedback behavior, inaccurate cause localization due to isolated processing of multimodal data, lack of dynamic tracing capability for negative feedback, and lack of prediction mechanism. It achieves the effect of closed-loop management of the entire process from data acquisition to intelligent control.

[0098] Specifically, the system enhances its ability to capture implicit negative feedback behavior by integrating the temporal correlation of multimodal data, ensuring a complete reconstruction of the user's negative experience journey; it achieves precise localization of the root causes of negative feedback by analyzing the potential causal relationships between negative feedback characteristics of different modalities through cross-modal correlators; it supports the visual tracking of the negative public opinion dissemination path through dynamic negative feedback source mapping, providing technical support for real-time localization of content root causes or functional defects; it establishes a quantitative relationship between implicit and explicit negative feedback through predictive models, achieving forward-looking insights into future negative feedback trends; and it dynamically generates control strategies through a strategy rule engine, ensuring that intervention measures are closely bound to the causal relationship chain of negative feedback, thereby significantly improving the accuracy and real-time performance of negative feedback monitoring.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A negative feedback monitoring method for media data, characterized in that, Includes the following steps: Step S1: Collect media data and user interaction behavior data from different channels, and construct a unified user session sequence model to perform time-series correlation of multimodal data within a single user session; Step S2: Process the media data, extract negative feedback features of text, visual, audio and user behavior modalities, and use a cross-modal correlator to analyze the correlation between negative feedback features of different modalities; Step S3: Based on the fused negative feedback features, construct a dynamic negative feedback source graph, where the nodes of the graph include content entities, functional entities and user profile tags, and the edges of the graph represent the associations and association strengths between nodes caused by negative feedback. Step S4: Based on historical negative feedback data and the negative feedback source map, use a prediction model to predict the negative feedback trend; Step S5: Based on the negative feedback source map and prediction results, the policy rule engine automatically generates and executes control strategies to intervene in related content, functions and recommendation systems.

2. The negative feedback monitoring method for media data according to claim 1, characterized in that, In step S1, the user session sequence model is used to associate and serialize the video viewing, comment posting, interactive operations, and mid-exit behavior generated by the user during a single continuous access on the timeline.

3. The negative feedback monitoring method for media data according to claim 1, characterized in that, In step S2, when a preset negative feature is identified in the first mode, a simultaneous search is performed to determine whether a corresponding negative feedback feature exists in the second mode that is in a temporal proximity. The correlation strength between the negative feedback features of the first mode and the second mode is calculated, and the correlation strength is used to verify the accuracy of feature recognition; wherein the calculation process includes statistical hypothesis testing and machine learning model.

4. The negative feedback monitoring method for media data according to claim 1, characterized in that, In step S3, the negative feedback source graph is a dynamic graph whose node and edge weights are updated in real time as new data flows in; the graph is used to visualize the propagation path of negative public opinion and locate the root cause of the content that leads to negative user feedback.

5. The negative feedback monitoring method for media data according to claim 1, characterized in that, In step S4, the prediction model includes a time series model and a sequence learning model, used to establish a quantitative prediction relationship between implicit negative feedback and explicit negative feedback, specifically: Using one or more implicit negative feedback data from users' cessation behavior, avoidance behavior, and silence behavior as input, predict the changing trends of the number of negative comments, reports, or unfollows within a specific future time window.

6. The negative feedback monitoring method for media data according to claim 1, characterized in that, In step S5, the rules of the policy rule engine are dynamically generated based on the node relationships revealed by the negative feedback source graph, and the rules are bound to the causal relationship chains in the graph.

7. The negative feedback monitoring method for media data according to claim 1, characterized in that, In step S5, the control strategy includes at least one of the following: When the negative feedback source map shows that negative feedback is strongly associated with a specific content ID, it triggers automatic downgrading of the content or sends it to the manual review queue. When the negative feedback source map shows a strong correlation between a large number of ad avoidance behaviors and a specific ad creative, an adjustment instruction is automatically sent to the ad delivery system. When the predictive model outputs a high-risk warning with negative feedback, it automatically pauses the push of related controversial content and generates a pre-set response strategy report for the customer service team.

8. A negative feedback monitoring system for media data, used to implement the negative feedback monitoring method for media data as described in any one of claims 1-7, characterized in that, The system includes: The data acquisition module is used to collect media data and user interaction behavior data from different channels, and to build a unified user session sequence model to correlate multimodal data within a single user session in a time sequence. The feature fusion module is used to process the media data, extract negative feedback features of text, visual, audio and user behavior modalities, and use a cross-modal correlator to analyze the correlation between negative feedback features of different modalities. The source graph construction module is used to construct a dynamic negative feedback source graph based on the fused negative feedback features. The nodes of the graph include content entities, functional entities and user profile tags, and the edges of the graph represent the associations and association strengths between nodes caused by negative feedback. The predictive insight module is used to predict negative feedback trends based on historical negative feedback data and the negative feedback source map using a predictive model. The intelligent control module is used to automatically generate and execute control strategies based on the negative feedback source map and prediction results through the strategy rule engine, and to intervene in related content, functions and recommendation systems.

Images