Deep Behavioral Recognition Model Construction for Live Streaming Reward Fraud - Fusing Music Sentiment Analysis and User Interaction Timing Graph Neural Networks

Abstract

The rapid development of the network live broadcasting industry has given rise to the economic model of reward, but it also brings the problem of frequent fraud. Anchor through false identity packaging, acting on behalf of playing, software beauty and other means to mislead users to irrational reward, the platform lacks an effective regulatory mechanism, the user's rights and interests are seriously damaged. Traditional fraud detection methods are difficult to effectively identify complex user interaction behavior patterns and emotional manipulation, and there is an urgent need to build an intelligent detection model that integrates multi-dimensional information to protect consumer rights and interests. In view of the problem of identifying fraudulent behavior of network live streaming bounty, the research constructed a graph neural network deep recognition model that integrates music sentiment analysis and user interaction time sequence. The model extracts audio features of the live broadcast through the music sentiment analysis module, and uses the Mel frequency cepstrum coefficient for sentiment classification; constructs a multi-view graph neural network architecture, uses the cosine similarity and K-nearest neighbor algorithm to construct node relationships, and fuses the embedded vector information from different viewpoints through the attention mechanism; and designs a five-classification detection framework to identify the behavioral types of normal interactions, irrational consumption frauds, shoddy product frauds, false information frauds, and misconception frauds. fraud and misperception fraud and other behavior types. Experimental results on Yelp and Amazon datasets show that the model improves 0.1298, 0.2186, and 0.2364 in F1 metrics compared to GCN, GAT, and GraphSAGE, respectively, and achieves 98.95% recall and 99.14% precision in the five-classification detection of webcaster fraud. The model effectively solves the problems of sample imbalance and complex interaction relationship identification in live streaming reward fraud detection, and provides technical support for risk control of webcasting platforms.