Instead of focusing on collaborative filtering then matrix factorization, we decided to implement both serial versions so that we can parallelize our work after the checkpoint. Here is a list of things we have already finished:

A Serial Implementation of Collaborative Filtering and Performance Analysis

• Code is available here at Github
• Algorithm explanation:
  1. We calculated pairwise user similarity using the Pearson Correlation
  2. To recommend to user $u$, for any other user $u'$ different from $u$, we look at the items $u'$ has rated, weight them by $u'$’s rating and the similarity between $u$ and $u'$, and use this as the predicted score for user $u$.
  3. We ran the algorithm on the $ml-100k$ dataset mentioned in the proposal
     • We ran gprof on this serial implementation of the code, the result is available here: gist

A CUDA Implementation of Collaborative Filtering and Performance Analysis

• Code is available here at Github

A Serial Implementation of SVD Matrix Factorization and Performance Analysis

• Code is available here at Github

API

Our API is available at Apiary. We designed this API so that even if we don’t have time to furnish the frontend, we still have a RESTful service available.

Server and Front-end

• Code is available here at Github

Issues and Concerns

• We decided not to work on the matrix factorization because we see a lot of possible optimization techniques on this one.

Goals and Deliverables

• We won’t be implementing the parallel matrix factorization algorithm. Instead, we will focus more on the collaborative filtering algorithm.

Future Plans

• We are going to try these techniques on the collborative filtering algorithm next:
  1. Data Compression
  2. Multi-GPU (we have two)
  3. CuBLAS
  4. Faster matrix multiplication so that the amortized cost of recommending multiple users is low
  5. Sorting optimization (Since we only need the $topn$ recommendations, we don’t have to sort everything)
  6. Use K-nearest Neighbor algorithm to find clusters of similar users. In this way, we can approximate the final result.