In this paper, we introduce a new concept of stability for cross-validation, called the $\left( \beta, \varpi \right)$-stability, and use it as a new perspective to build the general theory for cross-validation. The $\left( \beta, \varpi \right)$-stability mathematically connects the generalization ability and the stability of the cross-validated model via the Rademacher complexity. Our result reveals mathematically the effect of cross-validation from two sides: on one hand, cross-validation picks the model with the best empirical generalization ability by validating all the alternatives on test sets; on the other hand, cross-validation may compromise the stability of the model selection by causing subsampling error. Moreover, the difference between training and test errors in q\textsuperscript{th} round, sometimes referred to as the generalization error, might be autocorrelated on q. Guided by the ideas above, the $\left( \beta, \varpi \right)$-stability help us derivd a new class of Rademacher bounds, referred to as the one-round/convoluted Rademacher bounds, for the stability of cross-validation in both the i.i.d.\ and non-i.i.d.\ cases. For both light-tail and heavy-tail losses, the new bounds quantify the stability of the one-round/average test error of the cross-validated model in terms of its one-round/average training error, the sample sizes $n$, number of folds $K$, the tail property of the loss (encoded as Orlicz-$\Psi_\nu$ norms) and the Rademacher complexity of the model class $\Lambda$. The new class of bounds not only quantitatively reveals the stability of the generalization ability of the cross-validated model, it also shows empirically the optimal choice for number of folds $K$, at which the upper bound of the one-round/average test error is lowest, or, to put it in another way, where the test error is most stable.

A core challenge in the analysis of experimental data is that the impact of some intervention is often not entirely captured by a single, well-defined outcome. Instead there may be a large number of outcome variables that are potentially affected and of interest. In this paper, we propose a data-driven approach rooted in machine learning to the problem of testing effects on such groups of outcome variables. It is based on two simple observations. First, the 'false-positive' problem that a group of outcomes is similar to the concern of 'over-fitting,' which has been the focus of a large literature in statistics and computer science. We can thus leverage sample-splitting methods from the machine-learning playbook that are designed to control over-fitting to ensure that statistical models express generalizable insights about treatment effects. The second simple observation is that the question whether treatment affects a group of variables is equivalent to the question whether treatment is predictable from these variables better than some trivial benchmark (provided treatment is assigned randomly). This formulation allows us to leverage data-driven predictors from the machine-learning literature to flexibly mine for effects, rather than rely on more rigid approaches like multiple-testing corrections and pre-analysis plans. We formulate a specific methodology and present three kinds of results: first, our test is exactly sized for the null hypothesis of no effect; second, a specific version is asymptotically equivalent to a benchmark joint Wald test in a linear regression; and third, this methodology can guide inference on where an intervention has effects. Finally, we argue that our approach can naturally deal with typical features of real-world experiments, and be adapted to baseline balance checks.

We present a novel subset scan method to detect if a probabilistic binary classifier has statistically significant bias -- over or under predicting the risk -- for some subgroup, and identify the characteristics of this subgroup. This form of model checking and goodness-of-fit test provides a way to interpretably detect the presence of classifier bias or regions of poor classifier fit. This allows consideration of not just subgroups of a priori interest or small dimensions, but the space of all possible subgroups of features. To address the difficulty of considering these exponentially many possible subgroups, we use subset scan and parametric bootstrap-based methods. Extending this method, we can penalize the complexity of the detected subgroup and also identify subgroups with high classification errors. We demonstrate these methods and find interesting results on the COMPAS crime recidivism and credit delinquency data.

As machine learning algorithms become increasingly sophisticated to exploit subtle features of the data, they often become more dependent on simulations. This paper presents a new approach called weakly supervised classification in which class proportions are the only input into the machine learning algorithm. Using one of the most challenging binary classification tasks in high energy physics - quark versus gluon tagging - we show that weakly supervised classification can match the performance of fully supervised algorithms. Furthermore, by design, the new algorithm is insensitive to any mis-modeling of discriminating features in the data by the simulation. Weakly supervised classification is a general procedure that can be applied to a wide variety of learning problems to boost performance and robustness when detailed simulations are not reliable or not available.

Complex statistical machine learning models are increasingly being used or considered for use in high-stakes decision-making pipelines in domains such as financial services, health care, criminal justice and human services. These models are often investigated as possible improvements over more classical tools such as regression models or human judgement. While the modeling approach may be new, the practice of using some form of risk assessment to inform decisions is not. When determining whether a new model should be adopted, it is therefore essential to be able to compare the proposed model to the existing approach across a range of task-relevant accuracy and fairness metrics. Looking at overall performance metrics, however, may be misleading. Even when two models have comparable overall performance, they may nevertheless disagree in their classifications on a considerable fraction of cases. In this paper we introduce a model comparison framework for automatically identifying subgroups in which the differences between models are most pronounced. Our primary focus is on identifying subgroups where the models differ in terms of fairness-related quantities such as racial or gender disparities. We present experimental results from a recidivism prediction task and a hypothetical lending example.

Consider a binary decision making process where a single machine learning classifier replaces a multitude of humans. We raise questions about the resulting loss of diversity in the decision making process. We study the potential benefits of using random classifier ensembles instead of a single classifier in the context of fairness-aware learning and demonstrate various attractive properties: (i) an ensemble of fair classifiers is guaranteed to be fair, for several different measures of fairness, (ii) an ensemble of unfair classifiers can still achieve fair outcomes, and (iii) an ensemble of classifiers can achieve better accuracy-fairness trade-offs than a single classifier. Finally, we introduce notions of distributional fairness to characterize further potential benefits of random classifier ensembles.

In this paper we investigate the behavioural differences between mobile phone customers with prepaid and postpaid subscriptions. Our study reveals that (a) postpaid customers are more active in terms of service usage and (b) there are strong structural correlations in the mobile phone call network as connections between customers of the same subscription type are much more frequent than those between customers of different subscription types. Based on these observations we provide methods to detect the subscription type of customers by using information about their personal call statistics, and also their egocentric networks simultaneously. The key of our first approach is to cast this classification problem as a problem of graph labelling, which can be solved by max-flow min-cut algorithms. Our experiments show that, by using both user attributes and relationships, the proposed graph labelling approach is able to achieve a classification accuracy of $\sim 87\%$, which outperforms by $\sim 7\%$ supervised learning methods using only user attributes. In our second problem we aim to infer the subscription type of customers of external operators. We propose via approximate methods to solve this problem by using node attributes, and a two-ways indirect inference method based on observed homophilic structural correlations. Our results have straightforward applications in behavioural prediction and personal marketing.

Networking hardware giant Cisco Systems (NASDAQ:CSCO) has managed to maintain its dominance in the switching and routing markets despite significant shifts in the networking landscape. Software-defined networking and cloud computing have both threatened Cisco's business model of selling expensive, proprietary boxes in recent years. Even with these challenges, Cisco controlled 55.6% of the Ethernet switching market and 45% of the combined service provider and enterprise router market during the fourth quarter of 2016. This dominance doesn't mean that Cisco can sit on its hands. The company has been growing its software and services businesses, aiming to become a seller of solutions, not just hardware.

In the first part of this series, I introduced the Outbrain Click Prediction machine learning competition. That post described some preliminary and important data science tasks like exploratory data analysis and feature engineering performed for the competition, using a Spark cluster deployed on Google Dataproc. In this post, I describe the competition evaluation, the design of my cross-validation strategy and my baseline models using statistics and trees ensembles. In that competition, Kagglers were required to rank recommended ads by decreasing predicted likelihood of being clicked. Sponsored search advertising, contextual advertising, display advertising and real-time bidding auctions have all relied heavily on the ability of learned models to predict ad click–through rates (CTRs) accurately, quickly and reliably.

You know all too well at this point that all sorts of digital attacks are lurking on the internet. You could encounter ransomware, a virus, or a sketchy phish at any moment. Even creepier, though, some malicious code can actually hide inside other, benign software--and be programmed to jump out when you aren't expecting it. Hackers are increasingly using this technique, known as steganography, to trick internet users and smuggle malicious payloads past security scanners and firewalls. Unlike cryptography, which works to obscure content so it can't be understood, steganography's goal is to hide the fact that content exists at all by embedding it something else. And since steganography is a concept, not a specific method of clandestine data delivery, it can be used in all sorts of ingenious (and worrying) attacks.