Learning Sample-Specific Models with Low-Rank Personalized Regression

Authors: Ben Lengerich, Bryon Aragam, Eric P. Xing

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare personalized regression (hereafter, PR) to four baselines: 1) Population linear or logistic regression, 2) A mixture regression (MR) model, 3) Varying coefficients (VC), 4) Deep neural networks (DNN). First, we evaluate each method s ability to recover the true parameters from simulated data. Then we present three real data case studies, each progressively more challenging than the previous: 1) Stock prediction using financial data, 2) Cancer diagnosis from mass spectrometry data, and 3) Electoral prediction using historical election data. The results are summarized in Table 1 for easy reference.
Researcher Affiliation Academia Benjamin Lengerich Carnegie Mellon University blengeri@cs.cmu.edu Bryon Aragam University of Chicago bryon@chicagobooth.edu Eric P. Xing Carnegie Mellon University epxing@cs.cmu.edu
Pseudocode Yes Algorithm 1 Personalized Estimation
Open Source Code Yes A Python implementation is available at http://www.github.com/blengerich/ personalized_regression.
Open Datasets Yes Here, we investigate the capacity of PR to distinguish malignant from benign skin lesions using a dataset of desorption electrospray ionization mass spectrometry imaging (DESI-MSI) of a common skin cancer, basal cell carcinoma (BCC) [22] (details in supplement).
Dataset Splits No The paper refers to 'test sets' and 'out-of-sample prediction results', implying a split for evaluation, but does not specify explicit percentages, counts, or a standard citation for train/validation/test dataset splits.
Hardware Specification Yes With these performance improvements, we are able to fit models to datasets with over 10,000 samples and 1000s of predictors on a Macbook Pro with 16GB RAM in under an hour.
Software Dependencies No The paper mentions 'A Python implementation' but does not specify specific software dependencies with version numbers.
Experiment Setup Yes A discussion of hyperparameter selection is contained in Section. B.3 of the supplement. and Each personalized estimator is endowed with a personalized learning rate (i) t = t/kb (i) t b (pop)k1, which scales the global learning rate t according to how far the estimator has traveled. and In our experiments, we use kn = 3.