Abstract:We develop a solution framework for mitigating algorithmic bias in machine-learning classifiers. We consider an interactive problem setting where Alice (e.g., the firm) proposes to use data records to build a target classifier, while Bob (e.g., the regulator) builds a bias-proof classifier to protect sensitive dimensions from being discriminated on. Specific measures of algorithm fairness and a priori definition of bias terms are exempted in this propose and review scenario. Our solution relies on the idea that the discriminating strength (i.e., the bias) in Alice's target classifier corresponds to the separating capability (i.e., the signal) of Bob's discrimination-proof classifier, thus the goal is to reduce the separability of Bob's classifier while preserving the separability of Alice's classifier, obviating the fairness-utility tradeoff by separating two classifications. We measure the pairwise distance between data attributes from indicating their differences in contributing to classifier's separability. Multidimensional scaling (MDS) on the distance matrix helps establish the mapping between attributes and inferred bias sources. This bias-concentration information helps establish a ranking on attributes that suggests their bias-proneness. We mitigate bias at Bob's classifier (i.e., limit the classifier's separability) by searching for appropriate transformation of data records at bias-prone attributes. We consider polynomial/re-binning transformation at numerical/categorical data attributes, greedily suppressing attributes' maximum bias concentration below a threshold. The framework is demonstrated with a detailed case study at a well-known classification dataset. We achieve good bias-mitigation performance at 10/12 machine classifiers; results are robust across multiple design choices and engineering options. We further showcase the framework's performance at a larger dataset having >100 attributes. Results document a clear example where classification's fairness and classifier's utility under certain definitions are both enhanced.

An Approach for Reducing the Numeric Rating Bias

Sentiment Classification for Chinese Reviews: a Comparison Between SVM and Semantic Approaches

Improving the performance of lexicon-based review sentiment analysis method by reducing additional introduced sentiment bias

Bias-Aware Design for Informed Decisions: Raising Awareness of Self-Selection Bias in User Ratings and Reviews

Online Evaluation Re-Scoring Based on Review Behavior Analysis

Combating Selection Biases in Recommender Systems with a Few Unbiased Ratings

A Novel Customer Review Analysis System Based on Balanced Deep Review and Rating Differences in User Preference

Sustained effects of brief daily stress (fighting) upon brain and adrenal catecholamines and adrenal spleen, and heart weights of mice.

Reducing Popularity Bias in Recommender Systems through AUC-Optimal Negative Sampling

Exploring and Mitigating Gender Bias in Recommender Systems with Explicit Feedback

Exploring and mitigating gender bias in book recommender systems with explicit feedback

User Bias Removal in Review Score Prediction

'Propose and Review': Interactive Bias Mitigation for Machine Classifiers

Combining review text content and reviewer-item rating matrix to predict review rating

The Influence of Herd Mentality on Rating Bias and Popularity Bias: A Bi-Process Debiasing Recommendation Model Based on Matrix Factorization

A Warning Approach to Mitigating Bandwagon Bias in Online Ratings: Theoretical Analysis and Experimental Investigations.

Topic Detection and Summarization of User Reviews

Estimating Feature Ratings Through an Effective Review Selection Approach

Balancing Unobserved Confounding with a Few Unbiased Ratings in Debiased Recommendations

Fusion learning of preference and bias from ratings and reviews for item recommendation

A HYBRID DEEP LEARNING APPROACH FOR SENTIMENT ANALYSIS IN PRODUCT REVIEWS