<![CDATA[The rapid growth of the film industry and streaming platforms has led to information overload and filter bubbles that make it difficult for users to find content matching their narrative preferences. Prior content-based filtering approaches relying on word-frequency methods (TFIDF) suffer from a semantic gap and commonly depend on a single public dataset and a reference title (seed movie) as input. This study combines a public dataset with IMDb web-scraping results (a maximum population of 5,000 titles) and applies a Skip-gram Word2Vec model to represent movie synopses as 200-dimensional semantic vectors, paired with Cosine similarity to measure the closeness between a user's free-text keyword and movie synopses without requiring a seed movie. Data were split using an 80:20 Holdout method, and algorithm performance was evaluated on a Top-3 Recommendation window using Precision@K, Recall@K, and Mean Reciprocal Rank (MRR), with ground truth validated by two experts through Inter-Annotator Agreement. Testing on 25 queries produced a Precision@3 of 0.5333, Recall@3 of 0.7800, and MRR of 0.7300. These results indicate that integrating word embeddings with web scraping yields semantically relevant movie recommendations from free keyword input, though comparisons with baseline methods are needed for more definitive performance claims]]>
Copyrights © 2026
