Streaming Bayesian Deep Tensor Factorization

Authors: Shikai Fang, Zheng Wang, Zhimeng Pan, Ji Liu, Shandian Zhe

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental For evaluation, we examined SBDT in four real-world large-scale applications, including both binary and continuous tensors. We compared with the state-of-the-art streaming tensor factorization algorithm (Du et al., 2018) based on a multilinear form, and streaming nonlinear factorization methods implemented with SVB. In both running and final predictive performance, our method consistently outperforms the competing approaches, mostly by a large margin. The running accuracy of SBDT is also much more stable and smooth than the SVB based methods.
Researcher Affiliation Collaboration 1University of Uath 2Kwai Inc 3University of Rochester.
Pseudocode Yes Algorithm 1 Streaming Bayesian Deep Tensor Factorization (SBDT)
Open Source Code No The paper mentions that it implemented its method with Theano and TensorFlow, and refers to a MATLAB implementation for a baseline method (POST) at a GitHub link. However, it does not provide an explicit statement or link to the open-source code for SBDT itself.
Open Datasets Yes (1) DBLP (Du et al., 2018)... (2) Anime(https: //www.kaggle.com/Cooper Union/ anime-recommendations-database)... (3) ACC (Du et al., 2018)... (4) Movie Len1M (https: //grouplens.org/datasets/movielens/)
Dataset Splits No The paper explicitly describes training and test splits, but it does not mention a separate validation set or its corresponding split.
Hardware Specification Yes We ran all the methods on a desktop machine with Intel i9-9900K CPU and 32GB memory. We did not use GPU acceleration to run SBDT for a fair comparison.
Software Dependencies No The paper mentions software used (Theano, TensorFlow, Matlab) but does not provide specific version numbers for these dependencies, which are necessary for reproducibility.
Experiment Setup Yes For SVB/SS-GPTF, we set the number of pseudo inputs to 128 in their sparse approximations. We used Adam (Kingma & Ba, 2014) for the stochastic optimization in SVB-DTF and SVB/SS-GPTF, where we set the number of epochs to 100 in processing each streaming batch and tuned the learning rate from {10 5, 5 10 5, 10 4, 3 10 4, 5 10 4, 10 3, 3 10 3, 5 10 3, 10 2}. For SBDT and SVB-DTF, We used a 3-layer NN, with 50 nodes in each hidden layer. We tested Re LU and tanh activations.