Steffen Rendle, Research Scientist, Google
summary
Presents a single, generic model that replaces bespoke matrix/tensor factorizations: factorization machines that take arbitrary feature vectors and learn pairwise interactions with compact latent factors. Implements SGD, ALS/CD and MCMC inference with automatic hyperparameter selection, runs on Netflix Prize and KDDCup 2012 data, uses ~128 latent dimensions, and matches state‑of‑the‑art accuracy with simple preprocessing and no hand‑tuned regularization or learning rates. Shows runtime proportional to nonzeros × latent dim and examples of sequential/time features.
AI-written summary · about
chapters (19)
- 0:00Start
- 4:19Tensor Factorization
- 4:30Sequential Factorization Models
- 6:19Data and Variable Representation
- 7:41Linear Regression
- 9:42Polynomial Regression
- 10:59Representation: Matrix/ Tensor vs. Feature Vectors
- 15:54Application to Sparse Feature Vectors
- 18:17Factorization Machine (FM)
- 20:38Factorization Machines: Discussion
- 21:55Outline
- 22:24Matrix Factorization and Factorization Machines
- 23:02RDF-Triple Prediction with Factorization Machines
- 23:32Time with Factorization Machines
- 26:08Computation Complexity
- 30:34Stochastic Gradient Descent (SGD)
- 32:25libFM Software
- 33:12(Context-aware) Rating Prediction
- 33:40Netflix Prize
related talks
29mFurther enhancements of recommender systems using deep learning
38mRecommender systems with Tensorflow
42mAlex Smola, Professor in the Machine Learning Department, Carnegie Mellon University
01h 03mRecent Progress in High-Dimensional Learning
39mAnd Then There Are Algorithms
22mRonald Menich, Chief Data Scientist, Predictix, LLC
52mTensor Methods for Learning Latent Variable Models: Theory and Practice
56mTensor Decompositions for Estimating Latent Variable Models