Great question—and very common in early-stage platforms! Traditional collaborative filtering (like matrix factorization) relies heavily on explicit ratings (e.g., 1–5 stars), but *implicit feedback*—like watch completions, pause frequency, rewatch counts, and even time spent per minute—can absolutely power modern collaborative filtering. In fact, libraries like LightFM and implicit (Python) are designed exactly for this: they treat plays, likes, or session durations as confidence-weighted signals rather than hard ratings. For example, a full watch + rewatch might map to a confidence score of 20, while a 2-minute drop-off might be 2. You’d build a user-item interaction matrix where rows = users, columns = movies, and values = weighted engagement scores. That said, since you’re rating-free *and* likely facing a cold-start problem for new users/movies, consider a hybrid approach: start with implicit collaborative filtering for engaged users, then layer in content-based features (genre, director, cast embeddings from plot text or metadata) for newcomers. Also, don’t overlook simple baselines—'most watched this week' or 'similar to what you just finished' often outperform complex models early on. Would you like a quick sketch of how to structure that implicit interaction matrix using pandas and implicit?