LLORMA: Local Low-Rank Matrix Approximation
Authors: Joonseok Lee, Seungyeon Kim, Guy Lebanon, Yoram Singer, Samy Bengio
JMLR 2016 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conducted several experiments with recommendation data. In Section 7.1, we compare LLORMA to SVD and other state-of-the-art techniques. We also examine in the section dependency of LLORMA on the rank r, the number of anchor points q, and the training set size. In Section 7.2, we compare the parallel and global versions of LLORMA. Section 7.3 introduces several anchor point selection schemes and compare them experimentally. We used four popular recommendation systems datasets. The Movie Lens1 dataset is one of the most popular datasets in the literature. We used all versions of Movie Lens dataset, namely: 100K (1K 2K with 105 observations), 1M (6K 4K with 106 observations), and 10M (70K 10K with 107 observations). We also tested LLORMA on the Netflix dataset which is of size 480K 18K with 108 observations and the Bookcrossing dataset (100K 300K with 106 observations). These two datasets are much larger than the Movie Lens dataset. We also report results on the Yelp dataset (40K 10K with 105 observations), which is a recent dataset that is part of the ACM Rec Sys 2013 challenge2. |
| Researcher Affiliation | Collaboration | Joonseok Lee EMAIL Seungyeon Kim EMAIL Google Research, Mountain View, CA USA Guy Lebanon EMAIL Linked In, Mountain View, CA USA Yoram Singer EMAIL Samy Bengio EMAIL Google Research, Mountain View, CA USA. Part of this work was done while the first and second authors were in Georgia Institute of Technology. |
| Pseudocode | Yes | Algorithm 1 The Parallel LLORMA Algorithm; Algorithm 2 The Global LLORMA Algorithm |
| Open Source Code | No | The paper discusses using the PREA toolkit (Lee et al., 2012c) for comparative studies, but there is no explicit statement or link provided within the paper indicating that the source code for the LLORMA methodology itself is open-source or publicly available. |
| Open Datasets | Yes | We used four popular recommendation systems datasets. The Movie Lens1 dataset is one of the most popular datasets in the literature. We used all versions of Movie Lens dataset, namely: 100K (1K 2K with 105 observations), 1M (6K 4K with 106 observations), and 10M (70K 10K with 107 observations). We also tested LLORMA on the Netflix dataset which is of size 480K 18K with 108 observations and the Bookcrossing dataset (100K 300K with 106 observations). These two datasets are much larger than the Movie Lens dataset. We also report results on the Yelp dataset (40K 10K with 105 observations), which is a recent dataset that is part of the ACM Rec Sys 2013 challenge2. Footnote 1: http://www.grouplens.org/. Footnote 2: http://recsys.acm.org/recsys13/recsys-2013-challenge-workshop/ |
| Dataset Splits | Yes | Unless stated otherwise, we randomly divided the available data into training and test sets such that the ratio of training set size to test set size was 9:1. We created five random partitions and report the average performance over the five partitions. The test set size was fixed to 10% of the Movie Lens 1M and the RMSE was averaged over five random train-test splits. |
| Hardware Specification | Yes | The experiment was conducted on a quad-core machine with 4 threads while suspending any other process. |
| Software Dependencies | No | The paper mentions a fixed stepsize for gradient descent and regularization values, which are hyperparameters. While it refers to using the 'PREA toolkit', it does not provide any specific version numbers for this or any other software components, which is required for reproducibility. |
| Experiment Setup | Yes | We used the Epanechnikov kernel with h1 = h2 = 0.8, a fixed stepsize for gradient descent of µ = 0.01, and a 2-norm regularization value of λU = λV = 0.001. These values were selected using cross-validation. We set and did not attempt to optimize the parameters T = 100 (maximum number of iterations), ϵ = 0.0001 (gradient descent convergence threshold), and q = 50 (number of anchor points). |