content-based filtering

**Content-based filtering** recommends **items similar to what a user previously liked** — analyzing item features (genre, keywords, attributes) to suggest similar items, enabling personalized recommendations even for new items without user interaction history. **What Is Content-Based Filtering?** - **Definition**: Recommend items similar to user's past preferences. - **Method**: Match item features to user profile. - **Data**: Item attributes, user interaction history. - **Principle**: If you liked X, you'll like similar items. **How It Works** **1. Item Representation**: Extract features (genre, keywords, actors, ingredients, specifications). **2. User Profile**: Build profile from items user liked (aggregate features). **3. Similarity Matching**: Find items similar to user profile. **4. Ranking**: Score and rank candidate items. **Feature Types** **Structured**: Genre, price, size, color, brand, category. **Text**: Descriptions, reviews, tags, keywords. **Audio/Visual**: Image features, audio features, video content. **Metadata**: Author, director, artist, publisher, release date. **Similarity Measures** **Cosine Similarity**: Angle between feature vectors. **Euclidean Distance**: Geometric distance in feature space. **Jaccard Similarity**: Overlap of categorical features. **TF-IDF**: Text similarity based on term importance. **Advantages** - **No Cold Start for Items**: New items can be recommended immediately. - **Transparency**: Explainable ("Recommended because you liked X"). - **User Independence**: Doesn't need other users' data. - **Niche Items**: Can recommend unpopular items if features match. **Limitations** **Limited Diversity**: Only recommends similar items (filter bubble). **Feature Engineering**: Requires good item features. **New User Cold Start**: Still need user history. **Overspecialization**: Can't discover different types of items. **No Quality Signal**: Doesn't know if similar items are actually good. **Applications** - **News**: Recommend articles similar to what you read. - **Movies**: "If you liked this movie, try these similar films." - **Music**: Recommend songs with similar audio features. - **E-Commerce**: Products with similar specifications. - **Jobs**: Positions matching your skills and experience. **Tools**: scikit-learn (TF-IDF, cosine similarity), Gensim (doc2vec), sentence-transformers (embeddings).