Nonparametric Factor Trajectory Learning for Dynamic Tensor Decomposition
Authors: Zheng Wang, Shandian Zhe
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluated NONFAT in three real-world applications. We compared with the state-of-the-art tensor decomposition methods that incorporate both continuous and discretized time information. In most cases, NONFAT outperforms the competing methods, often by a large margin. NONFAT also achieves much better test log-likelihood, showing superior posterior inference results. We showcase the learned factor trajectories, which exhibit interesting temporal patterns and extrapolate well to the non-training region. The entry value prediction by NONFAT also shows a more reasonable uncertainty estimate in both interpolation and extrapolation. |
| Researcher Affiliation | Academia | Zheng Wang 1 Shandian Zhe 1 1School of Computing, University of Utah. Correspondence to: Shandian Zhe <zhe@cs.utah.edu>. |
| Pseudocode | No | The paper describes algorithmic steps verbally, such as 'develop a nested stochastic mini-batch variational learning algorithm', but it does not include a formally structured pseudocode block or an algorithm figure labeled as such. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code for the described methodology, nor does it provide a link to a code repository. |
| Open Datasets | Yes | (1) Beijing Air Quality1, hourly concentration measurements... 1https://archive.ics.uci.edu/ml/datasets/ Beijing+Multi-Site+Air-Quality+Data (2) Mobile Ads2, a 10-day click-through-rate dataset... 2https://www.kaggle.com/c/ avazu-ctr-prediction (3) DBLP 3, the bibliographic records... 3https://dblp.uni-trier.de/xml/ |
| Dataset Splits | Yes | We followed the standard testing procedure as in (Xu et al., 2012; Kang et al., 2012; Zhe et al., 2016b) to randomly sample 80% observed tensor entries (and their timestamps) for training and tested the prediction accuracy on the remaining entries. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU or CPU models, processor types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper states 'We implemented all the methods with Py Torch (Paszke et al., 2019)' but does not provide a specific version number for PyTorch or any other software dependency. |
| Experiment Setup | Yes | For all the GP baselines, we used SE kernel and followed (Zhe et al., 2016a) to use sparse variational GP framework (Hensman et al., 2013b) for scalable posterior inference, with 100 pseudo inputs. For NONFAT, the number of pseudo inputs for both levels of GPs was set to 100. For NN baselines, we used three-layer neural networks, with 50 neurons per layer and tanh as the activation... All the models were trained with stochastic mini-batch optimization. We used ADAM (Kingma and Ba, 2014) for all the methods, and the mini-batch size is 100. The learning rate was chosen from {10 4, 5 10 4, 10 3, 5 10 3, 10 2}. To ensure convergence, we ran each methods for 10K epochs. We varied the number of latent factors R from {2, 3, 5, 7}. |