precision at k

**Precision@K** measures **fraction of top-K results that are relevant** — evaluating what percentage of the first K results are actually useful, a simple and intuitive ranking metric. **What Is Precision@K?** - **Definition**: Percentage of top-K results that are relevant. - **Formula**: P@K = (# relevant in top-K) / K. - **Range**: 0 (no relevant results) to 1 (all top-K relevant). **Example** Top 10 results: 7 relevant, 3 not relevant. - Precision@10 = 7/10 = 0.7 (70% precision). **Why Precision@K?** - **User-Centric**: Users typically view only top-K results. - **Simple**: Easy to understand and explain. - **Practical**: Reflects real user experience. - **Actionable**: Clear target for improvement. **Common K Values** - **P@1**: Is top result relevant? (most critical). - **P@5**: Are top 5 results relevant? - **P@10**: Are top 10 results relevant? (common for search). - **P@20**: For longer result lists. **Limitations** - **Ignores Position**: Treats all top-K positions equally. - **Ignores Recall**: Doesn't consider relevant results beyond K. - **Binary**: Doesn't handle graded relevance. - **K-Dependent**: Different K values give different scores. **Precision@K vs. Other Metrics** **vs. Recall@K**: Precision = relevant retrieved / retrieved, Recall = relevant retrieved / total relevant. **vs. NDCG**: Precision@K binary, NDCG handles graded relevance and position. **vs. MAP**: Precision@K single cutoff, MAP averages precision at all relevant positions. **Applications**: Search evaluation, recommendation evaluation, information retrieval, any ranked list evaluation. **Tools**: scikit-learn, IR evaluation libraries, easy to implement. Precision@K is **the most intuitive ranking metric** — by measuring what fraction of top results are relevant, it directly captures user experience and is easy to understand and communicate.