영화 추천 알고리즘 대결 — 전통 협업 필터링 vs 비대칭 비율 기반 협업 필터링

Open Netflix and "Recommended for You" appears. YouTube endlessly suggests "Up Next." At the heart of these recommendations lies Collaborative Filtering (CFA) — a simple but powerful principle: "people with similar tastes will like similar things."

But traditional CFA has limitations. Data sparsity — most users rate a tiny fraction of available movies, making similarity calculations inaccurate. Cold start — no data exists for new users or new movies, making recommendations impossible. New approaches are emerging to address these issues.

This study proposed Dissymmetrical Percentage Collaborative Filtering (DSPCFA) and compared rating prediction accuracy against traditional CFA. DSPCFA's key innovation: accounting for rating distribution asymmetry when calculating user similarity. Some users are "generous raters" (consistently high scores), others are "strict raters" (consistently low). Ignoring this difference distorts similarity calculations.

Experimentally, DSPCFA reduced rating prediction error (MAE, RMSE) compared to traditional CFA. Improvement was most pronounced for users with fewer ratings — asymmetry correction's effect increased as data became sparser.

Recommendation systems are core technology for e-commerce, OTT, and music streaming. DSPCFA offers a practical improvement that boosts prediction accuracy without major infrastructure changes. Small algorithmic changes can transform millions of users' experiences.