Machine learning has automated much of financial fraud detection, notifying firms of, or even blocking, questionable transactions instantly. However, data imbalance starves traditionally trained models of the content necessary to detect fraud. This study examines three separate factors of credit card fraud detection via machine learning. First, it assesses the potential for different sampling methods, undersampling and Synthetic Minority Oversampling Technique (SMOTE), to improve algorithm performance in data-starved environments. Additionally, five industry-practical machine learning algorithms are evaluated on total fraud cost savings in addition to traditional statistical metrics. Finally, an ensemble of individual models is trained with a genetic algorithm to attempt to generate higher cost efficiency than its components. Monte Carlo performance distributions discerned random undersampling outperformed SMOTE in lowering fraud costs, and that an ensemble was unable to outperform its individual parts. Most notably,the F-1 Score, a traditional metric often used to measure performance with imbalanced data, was uncorrelated with derived cost efficiency. Assuming a realistic cost structure can be derived, cost-based metrics provide an essential supplement to objective statistical evaluation.

This article is about my journey of developing an image segmentation model for self-driving cars. I did this only for my personal amusement - there is no commercial intention involved. Check out the notebook including all the code for this project on kaggle. Self-driving cars are the future of individual transport. Who does not like the idea of entering a car, setting the destination, laying back and relaxing while the computer does the rest.

Imagine an advanced fighter aircraft is patrolling a hostile conflict area and a bogie suddenly appears on radar accelerating aggressively at them. The pilot, with the assistance of an Artificial Intelligence co-pilot, has a fraction of a second to decide what action to take – ignore, avoid, flee, bluff, or attack. The costs associated with False Positive and False Negative are substantial – a wrong decision that could potentially provoke a war or lead to the death of the pilot. What is one to do…and why? No one less than the Defense Advanced Research Projects Agency (DARPA) and the Department of Defense (DoD) are interested in not only applying AI to decide what to do in hostile, unstable and rapidly devolving environments but also want to understand why an AI model recommended a particular action.

In critical decision-making scenarios, optimizing accuracy can lead to a biased classifier, hence past work recommends enforcing group-based fairness metrics in addition to maximizing accuracy. However, doing so exposes the classifier to another kind of bias called infra-marginality. This refers to individual-level bias where some individuals/subgroups can be worse off than under simply optimizing for accuracy. For instance, a classifier implementing race-based parity may significantly disadvantage women of the advantaged race. To quantify this bias, we propose a general notion of $\eta$-infra-marginality that can be used to evaluate the extent of this bias. We prove theoretically that, unlike other fairness metrics, infra-marginality does not have a trade-off with accuracy: high accuracy directly leads to low infra-marginality. This observation is confirmed through empirical analysis on multiple simulated and real-world datasets. Further, we find that maximizing group fairness often increases infra-marginality, suggesting the consideration of both group-level fairness and individual-level infra-marginality. However, measuring infra-marginality requires knowledge of the true distribution of individual-level outcomes correctly and explicitly. We propose a practical method to measure infra-marginality, and a simple algorithm to maximize group-wise accuracy and avoid infra-marginality.

Deep networks are well-known to be fragile to adversarial attacks. Using several standard image datasets and established attack mechanisms, we conduct an empirical analysis of deep representations under attack, and find that the attack causes the internal representation to shift closer to the "false" class. Motivated by this observation, we propose to regularize the representation space under attack with metric learning in order to produce more robust classifiers. By carefully sampling examples for metric learning, our learned representation not only increases robustness, but also can detect previously unseen adversarial samples. Quantitative experiments show improvement of robustness accuracy by up to 4\% and detection efficiency by up to 6\% according to Area Under Curve (AUC) score over baselines.

In this paper, we address the challenging problem of learning from imbalanced data using a Nearest-Neighbor (NN) algorithm. In this setting, the minority examples typically belong to the class of interest requiring the optimization of specific criteria, like the F-Measure. Based on simple geometrical ideas, we introduce an algorithm that reweights the distance between a query sample and any positive training example. This leads to a modification of the Voronoi regions and thus of the decision boundaries of the NN algorithm. We provide a theoretical justification about the weighting scheme needed to reduce the False Negative rate while controlling the number of False Positives. We perform an extensive experimental study on many public imbalanced datasets, but also on large scale non public data from the French Ministry of Economy and Finance on a tax fraud detection task, showing that our method is very effective and, interestingly, yields the best performance when combined with state of the art sampling methods.

All single ECBS scores are averaged to compare with the ensemble models.

Accurately predicting patients' risk of 30-day hospital readmission would enable hospitals to efficiently allocate resource-intensive interventions. We develop a new method, Categorical Co-Frequency Analysis (CoFA), for clustering diagnosis codes from the International Classification of Diseases (ICD) according to the similarity in relationships between covariates and readmission risk. CoFA measures the similarity between diagnoses by the frequency with which two diagnoses are split in the same direction versus split apart in random forests to predict readmission risk. Applying CoFA to de-identified data from Berkshire Medical Center, we identified three groups of diagnoses that vary in readmission risk. To evaluate CoFA, we compared readmission risk models using ICD majors and CoFA groups to a baseline model without diagnosis variables. We found substituting ICD majors for the CoFA-identified clusters simplified the model without compromising the accuracy of predictions. Fitting separate models for each ICD major and CoFA group did not improve predictions, suggesting that readmission risk may be more homogeneous that heterogeneous across diagnosis groups.

Progress in ML for health care to date has been limited by the lack of well-defined questions and a dearth of annotated datasets. Many ML researchers remain focused on questions for which annotations are readily available, without necessarily questioning the clinical relevance of the problems and their solutions. For example, a popular benchmark challenge in the community focuses on predicting in-hospital mortality on the basis of data collected during the first 48 hours after admission to the intensive care unit4. Clearly annotated data are publicly available, and in recent years, performance on this task has approached an area under the curve (Box 1) of 0.9 (ref. However, assessing clinical utility requires careful evaluation against the scenario in which the model will be used.

Algorithmic risk assessments are increasingly used to help humans make decisions in high-stakes settings, such as medicine, criminal justice and education. In each of these cases, the purpose of the risk assessment tool is to inform actions, such as medical treatments or release conditions, often with the aim of reducing the likelihood of an adverse event such as hospital readmission or recidivism. Problematically, most tools are trained and evaluated on historical data in which the outcomes observed depend on the historical decision-making policy. These tools thus reflect risk under the historical policy, rather than under the different decision options that the tool is intended to inform. Even when tools are constructed to predict risk under a specific decision, they are often improperly evaluated as predictors of the target outcome. Focusing on the evaluation task, in this paper we define counterfactual analogues of common predictive performance and algorithmic fairness metrics that we argue are better suited for the decision-making context. We introduce a new method for estimating the proposed metrics using doubly robust estimation. We provide theoretical results that show that only under strong conditions can fairness according to the standard metric and the counterfactual metric simultaneously hold. Consequently, fairness-promoting methods that target parity in a standard fairness metric may --- and as we show empirically, do --- induce greater imbalance in the counterfactual analogue. We provide empirical comparisons on both synthetic data and a real world child welfare dataset to demonstrate how the proposed method improves upon standard practice.