Nonparametric Sparse Tensor Factorization with Hierarchical Gamma Processes
Authors: Conor Tillinghast, Zheng Wang, Shandian Zhe
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the advantage of our method in several benchmark datasets. For evaluation, we first ran simulations to confirm that our method can indeed generate increasingly sparser tensors. In three real-world applications, we examined the performance in predicting missing entry indices and entry values. In both tasks, our approach obtains the best performance, as compared with the state-of-the-art dense models and (Tillinghast & Zhe, 2021). |
| Researcher Affiliation | Academia | 1Department of Mathematics, University of Utah 2School of Computing, University of Utah. Correspondence to: Shandian Zhe <zhe@cs.utah.edu>. |
| Pseudocode | No | The paper describes algorithmic steps in text but does not include structured pseudocode or an algorithm block. |
| Open Source Code | No | We implemented our method with Py Torch (Paszke et al., 2019). CP-Bayes, GPTF, NEST-1 and NEST-2 were implemented with Tensor Flow (Abadi et al., 2016). Since all these methods used stochastic mini-batch optimization, we chose the learning rate from {10 4, 2 10 4, 5 10 4, 10 3, 5 10 3, 10 2}, and set the mini-batch size to 200 for Alog and Movie Lens, and 512 for SG. We ran 700 epochs on all the three datasets, which is enough for convergence. We used the original MATLAB implementation of CP-ALS and CPWOPT (http://www.tensortoolbox.org/), and C++ implementation of P-Tucker (https://github. com/sejoonoh/P-Tucker), and their default settings. No link for their code is provided. |
| Open Datasets | Yes | (1) Alog (Zhe et al., 2016b), extracted from a file access log, depicting the access frequency among users, actions, and resources, of size 200 100 200. (2) Movie Lens (https://grouplens.org/datasets/movielens/100k/), a three-mode (user, movie, time) tensor, of size 1000 1700 31. (3) SG (Li et al., 2015; Liu et al., 2019), extracted from Foursquare Singapore data, representing (user, location, point-of-interest) check-ins, of size 2321 5596 1600. |
| Dataset Splits | Yes | We randomly split the existent entries in each dataset into 80% for training and 20% for test. ... We ran an extra validation in the training set to select R1 and R2 (R1 + R2 = R) (In Fig. 5 of Appendix, we show how the performance of our method varies along with all possible combinations of R1 and R2). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instances) used for running its experiments. |
| Software Dependencies | No | The paper mentions implementing methods with PyTorch and TensorFlow and cites the original papers, but does not specify the version numbers for these software libraries used in their experiments. |
| Experiment Setup | Yes | We chose the learning rate from {10 4, 2 10 4, 5 10 4, 10 3, 5 10 3, 10 2}, and set the mini-batch size to 200 for Alog and Movie Lens, and 512 for SG. We ran 700 epochs on all the three datasets, which is enough for convergence. |