selection-bias

Official Tensorflow implementation for Deep Generative Positive-Unlabeled Learning under Selection Bias (VAE-PU) in CIKM 2020.

Longform data analysis article arguing every “dataset” is actually three: Observed (captured rows), Missing (what should exist but doesn’t), and Excluded (what filters/joins/dropna removed). Includes dataset accounting, join-loss and missingness audits, segmentation checks, and practical templates to prevent biased KPIs and wrong conclusions.

Code for a paper on estimation and evaluation of penalized survival models with high dimensional left-truncated and right-censored (LTRC) survival data

ELISL: Early-Late Synthetic Lethality Prediction in Cancer by Tepeli YI, Seale C, Gonçalves JP (bioRxiv 2022, Bioinformatics 2023)

Code accompanying the thesis project: "Understanding and Correcting Selection Bias in the Sentiments derived from Flemish Tweets".

An R Package for Estimation under Nonignorable Nonresponse in Sample Surveys

Simulated data for different selection bias structures

R code for reproduce real data analysis in MRAPSS paper.

Metric-DST: Mitigating Selection Bias Through Diversity-Guided Semi-Supervised Metric Learning by Tepeli YI, de Wolf M, and Gonçalves JP (arXiv 2024)

Confidence interval simulations and hypothesis testing in R, with a practical study of selection and survivorship bias.

DCAST: Diverse Class-Aware Self-Training for Fairer Learning by Tepeli YI and Gonçalves JP (arXiv 2024)

SBSL: Selection Bias-resilient Synthetic Lethality prediction models by Seale CF, Tepeli YI, Gonçalves JP (Bioinformatics 2022)

Navigation index for four independent research repositories with manuscripts, code, configurations, experiments, and evidence.

EvalFloor: is your LLM eval improvement real? Trying k prompt variants and keeping the best scores points on noise alone — this computes how many.

A research lab for off-policy evaluation, exploration, and policy-generated bias in contextual-bandit recommendation systems.

Grades a hyperparameter sweep against a truth it cannot see, including one task whose labels come from a fair coin so every configuration is provably at chance. A sweep on it reports 0.5202 on average and as high as 0.5672, and 95.3 percent of the improvement it claimed was selection bias.

A loyalty campaign appeared to lift weekly spend by $37.89. 95% of that gap existed before the campaign started.

Are public quant datasets point-in-time safe? Reproducible look-ahead-bias screens: OSAP (Chen-Zimmermann), JKP Global Factor Data, raw SEC EDGAR as-filed panels, feature-store as-of joins.

How much of a backtested Sharpe ratio is search? A study of the deflated Sharpe correction against 576 archived trials — including every configuration that lost.

Bank marketing ML model (92% ROC-AUC) with XGBoost + Platt scaling. EDA-driven binning, handles 93% class imbalance, addresses data leakage & selection bias. 16 docs covering nuances & business impact.