Stochastic linear bandits are a natural and well-studied model for structured exploration/exploitation problems and are widely used in applications such as on-line marketing and recommendation. One of the main challenges faced by practitioners hoping to apply existing algorithms is that usually the feedback is randomly delayed and delays are only partially observable. For example, while a purchase is usually observable some time after the display, the decision of not buying is never explicitly sent to the system. In other words, the learner only observes delayed positive events. We formalize this problem as a novel stochastic delayed linear bandit and propose OTFLinUCB and OTFLinTS, two computationally efficient algorithms able to integrate new information as it becomes available and to deal with the permanently censored feedback. We prove optimal O(d\sqrt{T}) bounds on the regret of the first algorithm and study the dependency on delay-dependent parameters. Our model, assumptions and results are validated by experiments on simulated and real data.

We study regression problems in which an adversary can exercise some control over the data generation process. Learner and adversary have conflicting but not necessarily perfectly antagonistic objectives. We study the case in which the learner is not fully informed about the adversary’s objective; instead, any knowledge of the learner about parameters of the adversary’s goal may be reflected in a Bayesian prior. We model this problem as a Bayesian game, and characterize conditions under which a unique Bayesian equilibrium point exists. We experimentally compare the Bayesian equilibrium strategy to the Nash equilibrium strategy, the minimax strategy, and regular linear regression.

The standard assumption of identically distributed training and test data is violated when test data are generated in response to a predictive model. This becomes apparent, for example, in the context of email spam filtering, where an email service provider employs a spam filter and the spam sender can take this filter into account when generating new emails. We model the interaction between learner and data generator as a Stackelberg competition in which the learner plays the role of the leader and the data generator may react on the leader's move. We derive an optimization problem to determine the solution of this game and present several instances of the Stackelberg prediction game. We show that the Stackelberg prediction game generalizes existing prediction models. Finally, we explore properties of the discussed models empirically in the context of email spam filtering.

We study a setting in which Poisson processes generate sequences of decision-making events. The optimization goal is allowed to depend on the rate of decision outcomes; the rate may depend on a potentially long backlog of events and decisions. We model the problem as a Poisson process with a throttling policy that enforces a data-dependent rate limit and reduce the learning problem to a convex optimization problem that can be solved efficiently. This problem setting matches applications in which damage caused by an attacker grows as a function of the rate of unsuppressed hostile events. We report on experiments on abuse detection for an email service.

The standard assumption of identically distributed training and test data can be violated when an adversary can exercise some control over the generation of the test data. In a prediction game, a learner produces a predictive model while an adversary may alter the distribution of input data. We study single-shot prediction games in which the cost functions of learner and adversary are not necessarily antagonistic. We identify conditions under which the prediction game has a unique Nash equilibrium, and derive algorithms that will find the equilibrial prediction models. In a case study, we explore properties of Nash-equilibrial prediction models for email spam filtering empirically.

We address classification problems for which the training instances are governed by a distribution that is allowed to differ arbitrarily from the test distribution---problems also referred to as classification under covariate shift. We derive a solution that is purely discriminative: neither training nor test distribution are modeled explicitly. We formulate the general problem of learning under covariate shift as an integrated optimization problem. We derive a kernel logistic regression classifier for differing training and test distributions.

We present a new approach to find an optimal large margin classifier based on the p-center which was proposed by Moretti in 2003. Starting with the p-Center of a general polytope, we extend this definition to a polyhedral cone, and introduce an algorithm approximating the p-Center of the version space, which we call p-Center machine (PCM). In addition, we present a large-scale and a soft boundary version of the PCM, and compare their performance to the support vector machine and the Bayes point machine. It turns out that the p-Center is close to the Bayes point and is similar in performance to the support vector machine as well as the Bayes point machine. Additionally, the proposed algorithm is highly parallelizable and thus very efficient in terms of computational effort.

The kernel perceptron is one of the simplest and fastest kernel machines, its performance, however, is inferior to other well known kernel machines. We introduce an algorithm that combines several approaches, mainly Herbrich's large-scale Bayes point machine and the soft perceptron in order to improve the kernel perceptron. Our experiments, which were based on standard benchmark datasets, show that the performance of the perceptron can be improved significantly with similar computational effort.