Scalable Bayesian Low-Rank Decomposition of Incomplete Multiway Tensors

Authors: Piyush Rai, Yingjian Wang, Shengbo Guo, Gary Chen, David Dunson, Lawrence Carin

ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform experiments with our model on both synthetic and real-world tensor datasets, and compare it with several baselines. The datasets used in our experiments span a wide range of application domains, such as chemometrics, multirelational social networks, brain-signal analysis (EEG), and image analysis.
Researcher Affiliation Collaboration Piyush Rai PIYUSH.RAI@DUKE.EDU Yingjian Wang YW65@DUKE.EDU Shengbo Guo S.GUO@SAMSUNG.COM Gary Chen GARY.CHEN@SAMSUNG.COM David Dunson DUNSON@DUKE.EDU Lawrence Carin LCARIN@DUKE.EDU Department of Electrical and Computer Engineering, Duke University, Durham, NC 27708, USA Samsung Research America Department of Statistical Science, Duke University, Durham, NC 27708, USA
Pseudocode No The paper describes the Gibbs sampling updates mathematically but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any statement or link regarding the public availability of its source code.
Open Datasets Yes Amino Acid data (Xu et al., 2013; Chu & Ghahramani, 2009), Flow Injection data (Xu et al., 2013; Chu & Ghahramani, 2009), Lazega-Lawyers multirelational social network data (Lazega, 2001), Kinship multirelational data (Nickel et al., 2011), Nation multirelational data (Nickel et al., 2011), benchmark Lena image.
Dataset Splits Yes For this task, we treat 50% of the data as missing and reconstruct it using the model learned on the remaining 50% data.For each dataset, except Kinship, we treat 90% of the entries as missing and predict them using the rest 10% data. For Kinship data we use the experimental setting of 90% training and 10% test data as done in other recent works (Nickel et al., 2011; Jenatton et al., 2012).Each experiment is repeated 10 times with different splits of observed and missing data.
Hardware Specification No The paper states it used an 'unoptimized MATLAB implementation' but does not provide any specific hardware details such as CPU, GPU, or memory specifications for running the experiments.
Software Dependencies No The paper mentions using an 'unoptimized MATLAB implementation' but does not specify any software versions or other dependencies.
Experiment Setup Yes For our model, we set µ(k) r to zero vector and Σ(k) r to identity matrix.We run the sampling based methods for 1500 iterations with 1000 burn-in iterations, collect samples every 5 iterations after the burnin phase, and report all results using the posterior sample based averages.