Using multimodal features and user classification to evaluate the design effectiveness of online advertising

Abstract

Traditional online advertising evaluation indicators (e.g., click-through rate and conversion rate) are characterized by limitations in scope and delayed feedback, rendering them inadequate for real-time and comprehensive evaluation. The existing models have insufficient accuracy when fusing multiple features. Therefore, the research proposes an evaluation framework that integrates multimodal features (MF) with attribute integration for user classification and constructs a cascading model based on MF and timing features (TC). The classification of user types is achieved through the attribute integration method, which utilizes multi-source data, including cognitive styles and emotional states. Quantification of features is facilitated by employing membership functions. The cascade model involves combining Markov chains and the PrefixSpan algorithm to mine temporal patterns. The results showed that when the user type was 2, the cluster center distances were (1.3, 0.9), and the type differences were significant. Under the activation of ReLU, the Auc of the cascade model was 0.803 and the logloss was 0.45. Under tanh activation, Auc = 0.795 and logloss = 0.452. The evaluation accuracy rate of the cascade model from 18:00 to 22:00 can reach up to 0.99 at most. The studies show that this method improves the evaluation accuracy and provides technical references for the industry.