의미 확장 검색과 감성 기반 후보 선별을 통한 영화 추천 알고리즘

초록

This study proposes a personalized movie recommendation system that filters out negatively reviewed films based on sentiment analysis and recommends content using a hybrid approach that incorporates keyword and genre information. Sentiment classification is performed using the pre-trained Korean language model KcELECTRA, and movies with a negative review ratio exceeding a predefined threshold (e.g., 60%) are excluded from the recommendation pool. Keywords are expanded using Word2Vec embeddings and a domain-specific association dictionary, and the recommendation ranking is determined based on the density of related reviews. When keyword-based recommendation is infeasible, high-rated movies within the selected genre are provided as alternatives. Experimental results show that, across two conditions(noisy/clean), the proposed method achieved Precision 0.424, Recall 1.000, F1 0.582/0.564, and a Neg-Filter-Rate of 8.873%/9.224%.

키워드