Kostadin Georgiev, VMware Bulgaria Preslav Nakov , Qatar Computing Research Institute

1. A non-IID Frameworkfor Collaborative Filtering with Restricted Boltzmann Machines Kostadin Georgiev, VMware Bulgaria PreslavNakov, Qatar Computing Research Institute ICML, June 17, 2013, Atlanta 5. Experiments & Evaluation 1. Overview • We propose a non-IID framework for collaborative filtering • based on Restricted Boltzmann Machines (RBM) • models both user-user and item-item correlations • uses real values in the visible layer (as opposed to multinomial) to model user-item ratings, thus taking advantage of the natural order between them • We further explore the potential of combining the original training data with data generated by the RBM-based model itself in a bootstrapping fashion • Evaluation: on two MovieLens datasets - with 100K and 1M user-item ratings, respectively • Results: Rival the state-of-the-art Data • Evaluation • Two MovieLens datasets: • 100k: • 1,682 movies assigned • 943 users • 100,000 ratings • sparseness: 93.7% • Each rating is between 1 (worst) and 5 (best) • Mean Absolute Error (MAE): • Cross-validation • 5-fold • 80%:20% training:testing data splits • 1M: • 3,952 movies • 6,040 users • 1 million ratings • sparseness: 95.8% - Item-based RBM model outperforms user-based, but not by much. - Hybrid item-based RBM + NB model is relatively insensitive to number of units. 2. User/Item-based RBM Model • The visible layer • represents either all ratings by a user or all rating for an item • units model ratings as real values (vs. multinomial) • noise-free reconstruction is better 3. Non-IID Hybrid RBM Model In the IID case, multinomial visible units are better than real-valued. In the non-IID case: real-valued visible units outperform multinomial. 6. Comparison to Previous Work MovieLens 1M • Remove the IID assumption for the training data • Topology: Unit vij is connected to two independent hidden layers: one user-based and another item-based. • Missing values (ratings): the generatedpredictions are used during testing, but are ignored during training • Training procedure: we average the predictions of the user-based and of the item-based RBM models MovieLens 100k 4. Neighborhood + I-RBM Use the I-RBM predictions from a neighborhood-based algorithm: 7. Conclusion & Future Work However, compute the averages from the original ratings: • Conclusion • proposed a non-IID RBM framework for collaborative filtering • rivals the best previous algorithms, which are more complex • Future work • add an additional layer to model higher-order correlations • add content-based features, e.g., demographic