Noise-Contrastive Estimation for Multivariate Point Processes

Authors: Hongyuan Mei, Tom Wan, Jason Eisner

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On several synthetic and real-world datasets, our method shows benefits: for the model to achieve the same level of log-likelihood on held-out data, our method needs considerably fewer function evaluations and less wall-clock time. and 5 Experiments We evaluate our NCE method on several synthetic and real-world datasets, with comparison to MLE, Guo et al. (2018) (denoted as b-NCE), and least-squares estimation (LSE) (Eichler et al., 2017).
Researcher Affiliation Academia Hongyuan Mei Tom Wan Jason Eisner Department of Computer Science, Johns Hopkins University 3400 N. Charles Street, Baltimore, MD 21218 U.S.A {hmei,tom,jason}@cs.jhu.edu
Pseudocode Yes The full pseudocode is given in Algorithm 1 in the supplementary material.
Open Source Code Yes By describing this method and releasing code, we hope to facilitate probabilistic modeling of continuous-time sequential data in many domains.
Open Datasets Yes Euro Email (Paranjape et al., 2017). This dataset contains time-stamped emails between anonymized members of a European research institute. We work on a subset of 100 most active members and then end up with K = 10000 possible event types and 50000 training event tokens. and Bitcoin OTC (Kumar et al., 2016). This dataset contains time-stamped rating (positive/negative) records between anonymized users on the Bitcoin OTC trading platform. We work on a subset of 100 most active users and then end up with K = 19800 (self-rating not allowed) possible event types but only 1000 training event tokens: this is an extremely data-sparse setting. and Robo Cup (Chen & Mooney, 2008). and IPTV (Xu et al., 2018).
Dataset Splits Yes For all the datasets, we use the first 80% as the training set, the next 10% as the validation set, and the last 10% as the test set.
Hardware Specification Yes two Titan X Pascal GPUs donated by NVIDIA Corporation and compute cycles from the Maryland Advanced Research Computing Center.
Software Dependencies No The paper mentions software like PyTorch and the Adam optimizer, but does not provide specific version numbers for these or other software dependencies used in the experiments.
Experiment Setup Yes We always set the minibatch size B to exhaust the GPU capacity, so smaller ρ or M allows larger B. and The displayed ρ and M values are among the better ones that we found during hyperparameter search.