Video classification based on descriptions generated by large language models

The goal of the thesis is to design and implement an explainable approach to video classification using descriptions generated by multimodal large language models. A web system Media Feature Lab (React + Flask) was developed, which in a five-phase pipeline automatically proposes and extracts textual attributes from videos via LLM and trains a RuleKit model that produces human-interpretable rules. The pipeline was validated on the MediaEval Predicting Video Memorability dataset and additionally on a medical dataset for pneumonia detection. On the Movie Memorability dataset, no statistically significant improvement over reference baselinemodelswasdemonstrated;themaincontributionistheproof-of-conceptdemonstration of the complete end-to-end process and the interpretability of its outputs.