Yelp Recommendation System
The original recommender — and the starting point of the multi-GPU study
PySpark MinHash/LSH collaborative filtering · Sep.–Dec. 2020 · USC INF 553 / CSCI 596
A Yelp recommendation pipeline built on Spark RDDs: MinHash/LSH candidate generation, TF-IDF content-based recommendation, and item- and user-based collaborative filtering with Pearson similarity.
This is where the Multi-GPU Similarity Engine came from. Revisiting it years later from a systems perspective — separating the compute-intensive similarity evaluation from the application pipeline, then asking how the same workload maps onto different execution backends — turned a course project into a controlled study of CPU, distributed-memory, and multi-GPU execution. End-to-end RMSE 0.8652, against the archived Spark model’s 0.8657.