Machine Learning — Recommendation System Comparison
This project compares Content-Based Filtering (CBF) and User-Based Collaborative Filtering (UBCF) for book recommendation under the same data preparation, train-test split, Top-N setting, and evaluation users. The goal is not only to identify which method performs better, but also to explain why the performance gap occurs.
The raw data consists of 271,360 book records, 1,149,780 rating records, and 278,858 user records. After removing implicit ratings (rating 0), aligning interactions at Book-Title level, and applying iterative filtering with at least 10 user interactions and 3 title interactions, the modeling data contains 4,196 users, 16,490 titles, and 133,907 interactions.
Data Understanding → Data Cleaning → Explicit Rating Filtering → User-Title Aggregation → Iterative K-Core Filtering → 80:20 Split per User → CBF with TF-IDF Metadata Similarity → UBCF with User-Item Matrix & User Similarity → Top-10 Recommendation → Precision, Recall, F1-Score & Diagnostic Evaluation
CBF achieved macro-average Precision of 0.0277 across 300 evaluated users.
CBF achieved macro-average Recall of 0.0840, higher than UBCF's 0.0073.
CBF achieved macro-average F1-Score of 0.0377, compared with 0.0042 for UBCF.
Both methods were evaluated on the same 300 users with Top-10 recommendation lists.
CBF produced recommendations for all evaluated users, giving 100% recommendation coverage in the diagnostic summary.
UBCF only produced recommendations for 75% of evaluated users, with an average recommendation count of 7.5 instead of 10.
The UBCF train user-item matrix is extremely sparse at about 99.85%, limiting overlapping ratings between users and weakening neighborhood-based recommendation opportunities.
CBF achieved a 19.67% hit rate in the diagnostic summary, while UBCF achieved 3.00%, showing that the observed performance gap is also related to recommendation availability and sparse interactions.
Relevant items in the test data are defined using Book-Rating ≥ 9. The filtered dataset is split per user into approximately 78.75% train and 21.25% test. Evaluation uses 300 users selected from 3,135 evaluable users, with Top-N fixed at 10 for both methods.
In this experimental setting, CBF performs better than UBCF on Precision, Recall, F1-Score, recommendation coverage, and hit rate. The result should be interpreted as specific to this dataset and evaluation design: CBF benefits from book metadata, while UBCF is strongly affected by the sparse user-item interaction matrix and limited user overlap.
