A Real-time Analysis System and Method for Live-stream E-commerce Based on Multi-dimensional Data Fusion

Abstract

This invention discloses a real-time analysis system and method for live-streaming e-commerce based on multi-dimensional data fusion, relating to the field of big data analysis technology; including: Step 1: Multi-dimensional data collection: collecting audio and video dimension data, transaction dimension data, interaction dimension data, public opinion dimension data, logistics dimension data, and anchor physiological dimension data respectively; Step 2: Synchronizing and compensating for missing data in the multi-dimensional data; Step 3: Performing parallel feature inference on four tasks—GMV prediction, violation risk score, price, and anchor fatigue—based on a fast-loop incremental dual-tower XGBoost-LightGBM fusion model, and outputting a traffic-risk control joint strategy based on the inferred features using a slow-loop reinforcement learning scheduler, executing the strategy and monitoring in real time; Step 4: Performing hash storage of original events, model versions, and decision instructions in a consortium blockchain manner to complete the audit loop.

Description

Technical Field

[0001] This invention discloses a real-time analysis system and method for live-streaming e-commerce based on multi-dimensional data fusion, which relates to the field of big data analysis technology. Background Technology

[0002] Existing real-time analytics for live-streaming e-commerce has some shortcomings, such as: Data silos: The time base of six dimensions of data, including audio and video, transactions, interactions, public opinion, logistics, and the physiological state of broadcasters, is not unified, resulting in high latency in data integration.

[0003] Model rigidity: Offline training + scheduled deployment cannot capture sudden patterns such as minute-level promotions and flash sales.

[0004] Risk control is lagging behind: violations and price fraud rely on manual review, with an average processing time of more than 30 seconds, leading to large-scale public opinion incidents.

[0005] Traffic scheduling is simplistic: traffic is allocated based solely on the number of online users or the historical average GMV, ignoring real-time risks. Summary of the Invention

[0006] This invention addresses the shortcomings of existing technologies by providing a real-time analysis system and method for live-streaming e-commerce based on multi-dimensional data fusion. This solves the problems of data silos, rigid models, lagging risk control, and simplistic traffic scheduling in existing live-streaming real-time analysis, thereby improving the live-streaming effect.

[0007] The specific solution proposed in this invention is as follows: This invention also provides a real-time analysis method for live-streaming e-commerce based on multi-dimensional data fusion, including: Step 1: Conduct multi-dimensional data collection: Collect data in the following dimensions: audio and video, transaction, interaction, public opinion, logistics, and the broadcaster's physiological state. Step 2: Synchronize and compensate for missing data in the multi-dimensional data; Step 3: Based on the fast-loop incremental dual-tower XGBoost-LightGBM fusion model, parallel feature inference is performed on four tasks: GMV prediction, violation risk score, price, and streamer fatigue. Based on the slow-loop reinforcement learning scheduler, a traffic-risk control joint strategy is output according to the inferred features, the strategy is executed and monitored in real time. Step 4: Use a consortium blockchain to perform hash-based evidence storage of the original events, model versions, and decision instructions to complete the audit loop.

[0008] Furthermore, in step 1 of the aforementioned method for real-time analysis of live-stream e-commerce based on multi-dimensional data fusion, when collecting audio and video dimension data, redundant streaming is performed using three channels: RTMP / SRT / WebRTC at multiple bitrates; and edge FPGA is used for ROI calculation and audio transcription. When collecting transaction-related data, we comprehensively utilize platform APIs, self-built order gateways, payment callbacks, and logistics tracking to aggregate transaction-related data. When collecting interactive data, six types of events are collected: bullet comments, likes, gifts, red envelopes, product clicks, and inventory deductions. A 128 KB circular buffer is built into the SDK. When collecting public opinion data, the RoBERTa model is used to perform real-time sentiment polarity analysis through the interface of the interactive platform in order to obtain public opinion data. When collecting logistics data, data is collected based on three levels of nodes: warehouse outbound, express delivery pickup, and abnormal storage. When collecting physiological data from livestreamers, we use wearable heart rate monitors, camera micro-expressions, and voiceprint fatigue analysis to analyze their physiological state and obtain physiological data.

[0009] Furthermore, in step 2 of the aforementioned real-time analysis method for live-streaming e-commerce based on multi-dimensional data fusion, a dynamic window is used to synchronize the multi-dimensional data: an adaptive window length of 1s-60s is selected, and data synchronization is adjusted in real time according to the data arrival jitter variance. The Kalman filter algorithm is used to compensate for the missing sampled data.

[0010] Furthermore, in step 3 of the aforementioned real-time analysis method for live-streaming e-commerce based on multi-dimensional data fusion, the fast-loop incremental dual-tower XGBoost-LightGBM fusion model includes tower 1, which is the XGBoost framework used to handle high-dimensional sparse cross features; and tower 2, which is the LightGBM framework used to handle temporal aggregate features. Slow-loop reinforcement learning scheduler is used for: State: s = [Real-time GMV estimate, violation risk score, streamer fatigue level, logistics delay probability]; Action: a = [Traffic weight, risk control threshold, price circuit breaker]; Reward: r = ΔGMV – λ1 × violation exposure – λ2 × user complaints – λ3 × streamer fatigue; λ1-λ3 are weighting coefficients. The policy gradient algorithm PPO-Clip is adopted, and a dual-tower model is used as the feature extractor in both the Actor network and the Critic network to achieve end-to-end gradient backpropagation.

[0011] This invention also provides a real-time analysis system for live-streaming e-commerce based on multi-dimensional data fusion, including a six-dimensional acquisition array module, a synchronization and compensation module, a dual-loop decision-making module, and a trusted closed-loop evidence storage module. The six-dimensional acquisition array module performs multi-dimensional data acquisition: it collects data in the dimensions of audio and video, transactions, interaction, public opinion, logistics, and the physiological state of the broadcaster. The synchronization and compensation module synchronizes and compensates for missing data in the multi-dimensional data. The dual-loop decision module performs parallel feature inference on four tasks: GMV prediction, violation risk score, price, and streamer fatigue based on the fast-loop incremental dual-tower XGBoost-LightGBM fusion model. It also outputs a traffic-risk control joint strategy based on the inferred features using a slow-loop reinforcement learning scheduler, executes the strategy, and monitors it in real time. The trusted closed-loop evidence storage module uses a consortium blockchain to perform hash-based evidence storage of original events, model versions, and decision instructions, thus completing the audit closed loop.

[0012] Furthermore, when the six-dimensional acquisition array module of the live e-commerce real-time analysis system based on multi-dimensional data fusion acquires audio and video dimension data, it performs redundant streaming of RTMP / SRT / WebRTC at multiple bitrates; and utilizes edge FPGA for ROI calculation and audio transcription. When collecting transaction-related data, we comprehensively utilize platform APIs, self-built order gateways, payment callbacks, and logistics tracking to aggregate transaction-related data. When collecting interactive data, six types of events are collected: bullet comments, likes, gifts, red envelopes, product clicks, and inventory deductions. A 128 KB circular buffer is built into the SDK. When collecting public opinion data, the RoBERTa model is used to perform real-time sentiment polarity analysis through the interface of the interactive platform in order to obtain public opinion data. When collecting logistics data, data is collected based on three levels of nodes: warehouse outbound, express delivery pickup, and abnormal storage. When collecting physiological data from livestreamers, we use wearable heart rate monitors, camera micro-expressions, and voiceprint fatigue analysis to analyze their physiological state and obtain physiological data.

[0013] Furthermore, the synchronization and compensation module of the live-streaming e-commerce real-time analysis system based on multi-dimensional data fusion utilizes a dynamic window to synchronize the multi-dimensional data: selecting an adaptive window length of 1s-60s, and adjusting data synchronization in real time based on the data arrival jitter variance. The Kalman filter algorithm is used to compensate for the missing sampled data.

[0014] Furthermore, the fast-loop incremental dual-tower XGBoost-LightGBM fusion model of the dual-loop decision module of the live e-commerce real-time analysis system based on multi-dimensional data fusion includes tower 1, which is the XGBoost framework, used to process high-dimensional sparse cross features; and tower 2, which is the LightGBM framework, used to process temporal aggregate features. The dual-loop decision module, based on a slow-loop reinforcement learning scheduler, targets: State: s = [Real-time GMV estimate, violation risk score, streamer fatigue level, logistics delay probability]; Action: a = [Traffic weight, risk control threshold, price circuit breaker]; Reward: r = ΔGMV – λ1 × violation exposure – λ2 × user complaints – λ3 × streamer fatigue; λ1-λ3 are weighting coefficients. The policy gradient algorithm PPO-Clip is adopted, and a dual-tower model is used as the feature extractor in both the Actor network and the Critic network to achieve end-to-end gradient backpropagation.

[0015] The advantages of this invention are: This invention utilizes an incremental dual-tower XGBoost-LightGBM fusion framework and a reinforcement learning-based traffic-risk control coupled scheduling strategy to achieve an end-to-end intelligent closed loop for real-time GMV prediction, second-level interception of illegal content, immediate circuit breaking for price anomalies, quantifiable assessment of streamer health, and joint scheduling of platform traffic and risk control. It also employs blockchain notarization to ensure regulatory traceability and judicial evidence collection; significantly improving resource utilization and requiring a substantial reduction in the number of servers of the same scale. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention. Example

[0018] This invention also provides a real-time analysis method for live-streaming e-commerce based on multi-dimensional data fusion, including: Step 1: Collect multi-dimensional data: Collect data in the following dimensions: audio and video, transaction, interaction, public opinion, logistics, and the anchor's physiological state.

[0019] When acquiring audio and video data, redundant streaming is performed using three channels: RTMP / SRT / WebRTC at multiple bitrates; and edge FPGA is used for ROI calculation and audio transcription. When collecting transaction-related data, we comprehensively utilize platform APIs, self-built order gateways, payment callbacks, and logistics tracking to aggregate transaction-related data. When collecting interactive data, six types of events are collected: bullet comments, likes, gifts, red envelopes, product clicks, and inventory deductions. A 128 KB circular buffer is built into the SDK. When collecting public opinion data, the RoBERTa model is used to perform real-time sentiment polarity analysis through the interface of the interactive platform in order to obtain public opinion data. When collecting logistics data, data is collected based on three levels of nodes: warehouse outbound, express delivery pickup, and abnormal storage. When collecting physiological data from livestreamers, we use wearable heart rate monitors, camera micro-expressions, and voiceprint fatigue analysis to analyze their physiological state and obtain physiological data.

[0020] Step 2: Synchronize and compensate for missing data in the multi-dimensional data.

[0021] The multi-dimensional data is synchronized using a dynamic window: an adaptive window length of 1-60 seconds is selected, and data synchronization is adjusted in real time based on the data arrival jitter variance. The Kalman filter algorithm is used to compensate for the missing sampled data.

[0022] RoaringBitmap + Run-Length Encoding can be used to compress the state of hundreds of millions of users, with a single node memory usage of <2 GB.

[0023] Step 3: Based on the fast-loop incremental dual-tower XGBoost-LightGBM fusion model, parallel feature inference is performed on four tasks: GMV prediction, violation risk score, price, and streamer fatigue. Based on the slow-loop reinforcement learning scheduler, a traffic-risk control joint strategy is output according to the inferred features, the strategy is executed, and real-time monitoring is performed.

[0024] The fast-loop incremental dual-tower XGBoost-LightGBM fusion model includes Tower 1, the XGBoost framework, for handling high-dimensional sparse cross features; and Tower 2, the LightGBM framework, for handling temporal aggregated features. A 30-second model update can be achieved using Online-Bagging with a sliding window weight α(t) = exp(-Δt / τ).

[0025] Slow-loop reinforcement learning scheduler is used for: State: s = [Real-time GMV estimate, violation risk score, streamer fatigue level, logistics delay probability]; Action: a = [Traffic weight, risk control threshold, price circuit breaker]; Reward: r = ΔGMV – λ1 × violation exposure – λ2 × user complaints – λ3 × streamer fatigue; λ1-λ3 are weighting coefficients. The policy gradient algorithm PPO-Clip is adopted, and a dual-tower model is used as the feature extractor in both the Actor network and the Critic network to achieve end-to-end gradient backpropagation.

[0026] Step 4: Use a consortium blockchain to perform hash-based evidence storage of the original events, model versions, and decision instructions to complete the audit loop.

[0027] The consortium blockchain uses PBFT consensus, with a block time ≤ 500 ms, and writes the hashes to the original IPFS log. Specifically, it collects raw events over a period of time, packages them, stores them in IPFS to obtain hash H1, and writes H1 to the consortium blockchain. A specific model version is used to process the state and make decisions. The parameters and feature schema of model M are used to calculate hash H2, which is then written to the consortium blockchain; the decision instructions are used to calculate hash H3, which is also written to the consortium blockchain. This creates an immutable log on the consortium blockchain. Example

[0028] This invention also provides a real-time analysis system for live-streaming e-commerce based on multi-dimensional data fusion, including a six-dimensional acquisition array module, a synchronization and compensation module, a dual-loop decision-making module, and a trusted closed-loop evidence storage module. The six-dimensional acquisition array module performs multi-dimensional data acquisition: it collects data in the dimensions of audio and video, transactions, interaction, public opinion, logistics, and the physiological state of the broadcaster. The synchronization and compensation module synchronizes and compensates for missing data in the multi-dimensional data. The dual-loop decision module performs parallel feature inference on four tasks: GMV prediction, violation risk score, price, and streamer fatigue based on the fast-loop incremental dual-tower XGBoost-LightGBM fusion model. It also outputs a traffic-risk control joint strategy based on the inferred features using a slow-loop reinforcement learning scheduler, executes the strategy, and monitors it in real time. The trusted closed-loop evidence storage module uses a consortium blockchain to perform hash-based evidence storage of original events, model versions, and decision instructions, thus completing the audit closed loop.

[0029] The information interaction and execution process between the modules in the above system are based on the same concept as the method embodiment of the present invention, and the specific details can be found in the description in the method embodiment of the present invention, and will not be repeated here.

[0030] Similarly, this invention's system, based on the incremental dual-tower XGBoost-LightGBM fusion framework and a reinforcement learning-based traffic-risk control coupled scheduling strategy, achieves an end-to-end intelligent closed loop for real-time GMV prediction, second-level interception of illegal content, immediate circuit breaking for price anomalies, quantifiable assessment of streamer health, and joint scheduling of platform traffic and risk control. It also uses blockchain for evidence storage, satisfying regulatory traceability and judicial evidence collection requirements; significantly improving resource utilization and requiring significantly fewer servers of the same scale.

[0031] It should be noted that not all steps and modules in the above processes and system structures are mandatory; some steps or modules can be omitted as needed. The execution order of each step is not fixed and can be adjusted as required. The system structures described in the above embodiments can be physical or logical structures. That is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.

[0032] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A real-time analysis method for live-streaming e-commerce based on multi-dimensional data fusion, characterized by: include: Step 1: Conduct multi-dimensional data collection: Collect data in the following dimensions: audio and video, transaction, interaction, public opinion, logistics, and the broadcaster's physiological state. Step 2: Synchronize and compensate for missing data in the multi-dimensional data; Step 3: Based on the fast-loop incremental dual-tower XGBoost-LightGBM fusion model, parallel feature inference is performed on four tasks: GMV prediction, violation risk score, price, and streamer fatigue. Based on the slow-loop reinforcement learning scheduler, a traffic-risk control joint strategy is output according to the inferred features, the strategy is executed and monitored in real time. Step 4: Use a consortium blockchain to perform hash-based evidence storage of the original events, model versions, and decision instructions to complete the audit loop.

2. The real-time analysis method for live-streaming e-commerce based on multi-dimensional data fusion as described in claim 1, characterized in that: In step 1, when collecting audio and video data, redundant streaming is performed using three channels: RTMP, SRT, and WebRTC. ROI calculation and audio transcription are also performed using an edge FPGA. When collecting transaction-related data, we comprehensively utilize platform APIs, self-built order gateways, payment callbacks, and logistics tracking to aggregate transaction-related data. When collecting interactive data, six types of events are collected: bullet comments, likes, gifts, red envelopes, product clicks, and inventory deductions. A 128 KB circular buffer is built into the SDK. When collecting public opinion data, the RoBERTa model is used to perform real-time sentiment polarity analysis through the interface of the interactive platform in order to obtain public opinion data. When collecting logistics data, data is collected based on three levels of nodes: warehouse outbound, express delivery pickup, and abnormal storage. When collecting physiological data from livestreamers, we use wearable heart rate monitors, camera micro-expressions, and voiceprint fatigue analysis to analyze their physiological state and obtain physiological data.

3. The real-time analysis method for live-streaming e-commerce based on multi-dimensional data fusion as described in claim 1, characterized in that: Step 2 utilizes a dynamic window to synchronize the multi-dimensional data: an adaptive window length of 1-60 seconds is selected, and data synchronization is adjusted in real time based on the data arrival jitter variance. The Kalman filter algorithm is used to compensate for the missing sampled data.

4. The real-time analysis method for live-streaming e-commerce based on multi-dimensional data fusion as described in claim 1, Its characteristic is that the fast loop incremental dual-tower XGBoost-LightGBM fusion model in step 3 includes tower 1, which is the XGBoost framework, used to handle high-dimensional sparse cross features; and tower 2, which is the LightGBM framework, used to handle temporal aggregate features. Slow-loop reinforcement learning scheduler is used for: State: s = [Real-time GMV estimate, violation risk score, streamer fatigue level, logistics delay probability]; Action: a = [Traffic weight, risk control threshold, price circuit breaker]; Reward: r = ΔGMV – λ1 × violation exposure – λ2 × user complaints – λ3 × streamer fatigue; λ1-λ3 are weighting coefficients. The policy gradient algorithm PPO-Clip is adopted, and a dual-tower model is used as the feature extractor in both the Actor network and the Critic network to achieve end-to-end gradient backpropagation.

5. A real-time analysis system for live-streaming e-commerce based on multi-dimensional data fusion, characterized by: It includes a six-dimensional acquisition array module, a synchronization and compensation module, a dual-loop decision-making module, and a trusted closed-loop evidence storage module. The six-dimensional acquisition array module performs multi-dimensional data acquisition: it collects data in the dimensions of audio and video, transactions, interaction, public opinion, logistics, and the physiological state of the broadcaster. The synchronization and compensation module synchronizes and compensates for missing data in the multi-dimensional data. The dual-loop decision module performs parallel feature inference on four tasks: GMV prediction, violation risk score, price, and streamer fatigue based on the fast-loop incremental dual-tower XGBoost-LightGBM fusion model. It also outputs a traffic-risk control joint strategy based on the inferred features using a slow-loop reinforcement learning scheduler, executes the strategy, and monitors it in real time. The trusted closed-loop evidence storage module uses a consortium blockchain to perform hash-based evidence storage of original events, model versions, and decision instructions, thus completing the audit closed loop.

6. The real-time analysis system for live-streaming e-commerce based on multi-dimensional data fusion as described in claim 5, characterized in that: When the six-dimensional acquisition array module acquires audio and video dimension data, it performs redundant streaming of multiple bitrates RTMP / SRT / WebRTC; and utilizes edge FPGA for ROI calculation and audio transcription. When collecting transaction-related data, we comprehensively utilize platform APIs, self-built order gateways, payment callbacks, and logistics tracking to aggregate transaction-related data. When collecting interactive data, six types of events are collected: bullet comments, likes, gifts, red envelopes, product clicks, and inventory deductions. A 128 KB circular buffer is built into the SDK. When collecting public opinion data, the RoBERTa model is used to perform real-time sentiment polarity analysis through the interface of the interactive platform in order to obtain public opinion data. When collecting logistics data, data is collected based on three levels of nodes: warehouse outbound, express delivery pickup, and abnormal storage. When collecting physiological data from livestreamers, we use wearable heart rate monitors, camera micro-expressions, and voiceprint fatigue analysis to analyze their physiological state and obtain physiological data.

7. A real-time analysis system for live-streaming e-commerce based on multi-dimensional data fusion as described in claim 5, characterized in that: The synchronization and compensation module uses a dynamic window to synchronize the multi-dimensional data: selecting an adaptive window length of 1s-60s, and adjusting the data synchronization in real time based on the data arrival jitter variance. The Kalman filter algorithm is used to compensate for the missing sampled data.

8. A real-time analysis system for live-streaming e-commerce based on multi-dimensional data fusion as described in claim 5, characterized in that: The fast-loop incremental dual-tower XGBoost-LightGBM fusion model of the dual-loop decision module includes Tower 1, which is the XGBoost framework for handling high-dimensional sparse cross features; and Tower 2, which is the LightGBM framework for handling temporal aggregate features. The dual-loop decision module, based on a slow-loop reinforcement learning scheduler, targets: State: s = [Real-time GMV estimate, violation risk score, streamer fatigue level, logistics delay probability]; Action: a = [Traffic weight, risk control threshold, price circuit breaker]; Reward: r = ΔGMV – λ1 × violation exposure – λ2 × user complaints – λ3 × streamer fatigue; λ1-λ3 are weighting coefficients. The policy gradient algorithm PPO-Clip is adopted, and a dual-tower model is used as the feature extractor in both the Actor network and the Critic network to achieve end-to-end gradient backpropagation.

Images