A late fusion framework with multiple optimization methods for media interestingness

The recent advancement in Multimedia Analytical, Computer Vision (CV), and Artificial Intelligence (AI) algorithms resulted in several interesting tools allowing an automatic analysis and retrieval of multimedia content of users’ interests. However, retrieving the content of interest generally involves analysis and extraction of semantic features, such as emotions and interestingness-level. The extraction of such meaningful information is a complex task and generally, the performance of individual algorithms is very low. One way to enhance the performance of the individual algorithms is to combine the predictive capabilities of multiple algorithms using fusion schemes. This allows the individual algorithms to complement each other, leading to improved performance. This paper proposes several fusion methods for the media interestingness score prediction task introduced in CLEF Fusion 2022. The proposed methods include both a naive fusion scheme, where all the inducers are treated equally and a merit-based fusion scheme where multiple weight optimization methods are employed to assign weights to the individual inducers. In total, we used six optimization methods including a Particle Swarm Optimization (PSO), a Genetic Algorithm (GA), Nelder-Mead, Trust Region Constrained (TRC), and Limited-memory Broyden–Fletcher–Goldfarb–Shanno Algorithm (LBFGSA), and Truncated Newton Algorithm (TNA). Overall better results are obtained with PSO and TNA achieving 0.109 mean average precision @10. The task is complex and generally, scores are low. We believe the presented analysis will provide a baseline for future research in the domain.