Ofer Dekel

Large margin hierarchical classification

We present an algorithmic framework for supervised classification learning where the set of labels is organized in a predefined hierarchical structure. This structure is encoded by a rooted tree which induces a metric over the label set. Our approach combines ideas from large margin kernel methods and Bayesian analysis. Following the large margin principle, we associate a prototype with each label in the tree and formulate the learning task as an optimization problem with varying margin constraints. In the spirit of Bayesian methods, we impose similarity requirements between the prototypes corresponding to adjacent labels in the hierarchy. We describe new online and batch algorithms for solving the constrained optimization problem. We derive a worst case loss-bound for the online algorithm and provide generalization analysis for its batch counterpart. We demonstrate the merits of our approach with a series of experiments on synthetic, text and speech data.

The Forgetron: A Kernel-Based Perceptron on a Budget

The Perceptron algorithm, despite its simplicity, often performs well in online classification tasks. The Perceptron becomes especially effective when it is used in conjunction with kernel functions. However, a common difficulty encountered when implementing kernel-based online algorithms is the amount of memory required to store the online hypothesis, which may grow unboundedly as the algorithm progresses. Moreover, the running time of each online round grows linearly with the amount of memory used to store the hypothesis. In this paper, we present the Forgetron family of kernel-based online classification algorithms, which overcome this problem by restricting themselves to a predefined memory budget. We obtain different members of this family by modifying the kernel-based Perceptron in various ways. We also prove a unified mistake bound for all of the Forgetron algorithms. To our knowledge, this is the first online kernel-based learning paradigm which, on one hand, maintains a strict limit on the amount of memory it uses and, on the other hand, entertains a relative mistake bound. We conclude with experiments using real datasets, which underscore the merits of our approach.

Good learners for evil teachers

We consider a supervised machine learning scenario where labels are provided by a heterogeneous set of teachers, some of which are mediocre, incompetent, or perhaps even malicious. We present an algorithm, built on the SVM framework, that explicitly attempts to cope with low-quality and malicious teachers by decreasing their influence on the learning process. Our algorithm does not receive any prior information on the teachers, nor does it resort to repeated labeling (where each example is labeled by multiple teachers). We provide a theoretical analysis of our algorithm and demonstrate its merits empirically. Finally, we present a second algorithm with promising empirical results but without a formal analysis.

Incentive compatible regression learning

We initiate the study of incentives in a general machine learning framework. We focus on a game-theoretic regression learning setting where private information is elicited from multiple agents with different, possibly conflicting, views on how to label the points of an input space. This conflict potentially gives rise to untruthfulness on the part of the agents. In the restricted but important case when every agent cares about a single point, and under mild assumptions, we show that agents are motivated to tell the truth. In a more general setting, we study the power and limitations of mechanisms without payments. We finally establish that, in the general setting, the VCG mechanism goes a long way in guaranteeing truthfulness and economic efficiency.

Learning to classify with missing and corrupted features

A common assumption in supervised machine learning is that the training examples provided to the learning algorithm are statistically identical to the instances encountered later on, during the classification phase. This assumption is unrealistic in many real-world situations where machine learning techniques are used. We focus on the case where features of a binary classification problem, which were available during the training phase, are either deleted or become corrupted during the classification phase. We prepare for the worst by assuming that the subset of deleted and corrupted features is controlled by an adversary, and may vary from instance to instance. We design and analyze two novel learning algorithms that anticipate the actions of the adversary and account for them when training a classifier. Our first technique formulates the learning problem as a linear program. We discuss how the particular structure of this program can be exploited for computational efficiency and we prove statistical bounds on the risk of the resulting classifier. Our second technique addresses the robust learning problem by combining a modified version of the Perceptron algorithm with an online-to-batch conversion technique, and also comes with statistical generalization guarantees. We demonstrate the effectiveness of our approach with a set of experiments.

An online algorithm for hierarchical phoneme classification

We present an algorithmic framework for phoneme classification where the set of phonemes is organized in a predefined hierarchical structure. This structure is encoded via a rooted tree which induces a metric over the set of phonemes. Our approach combines techniques from large margin kernel methods and Bayesian analysis. Extending the notion of large margin to hierarchical classification, we associate a prototype with each individual phoneme and with each phonetic group which corresponds to a node in the tree. We then formulate the learning task as an optimization problem with margin constraints over the phoneme set. In the spirit of Bayesian methods, we impose similarity requirements between the prototypes corresponding to adjacent phonemes in the phonetic hierarchy. We describe a new online algorithm for solving the hierarchical classification problem and provide worst-case loss analysis for the algorithm. We demonstrate the merits of our approach by applying the algorithm to synthetic data and as well as speech data.

Smooth ε-Insensitive Regression by Loss Symmetrization

We describe new loss functions for regression problems along with an accompanying algorithmic framework which utilizes these functions. These loss functions are derived by symmetrization of margin-based losses commonly used in boosting algorithms, namely, the logistic loss and the exponential loss. The resulting symmetric logistic loss can be viewed as a smooth approximation to the e-insensitive hinge loss used in support vector regression. We describe and analyze two parametric families of batch learning algorithms for minimizing these symmetric losses. The first family employs an iterative log-additive update which can be viewed as a regression counterpart to recent boosting algorithms. The second family utilizes an iterative additive update step. We also describe and analyze online gradient descent (GD) and exponentiated gradient (EG) algorithms for the symmetric logistic loss. A byproduct of our work is a new simple form of regularization for boosting-based classification and regression algorithms. Our regression framework also has implications on classification algorithms, namely, a new additive update boosting algorithm for classification. We demonstrate the merits of our algorithms in a series of experiments.

Bandits with switching costs: T 2/3 regret

We study the adversarial multi-armed bandit problem in a setting where the player incurs a unit cost each time he switches actions. We prove that the players T-round minimax regret in this setting is [EQUATION], thereby closing a fundamental gap in our understanding of learning with bandit feedback. In the corresponding full-information version of the problem, the minimax regret is known to grow at a much slower rate of Θ(√T). The difference between these two rates provides the first indication that learning with bandit feedback can be significantly harder than learning with full information feedback (previous results only showed a different dependence on the number of actions, but not on T.) In addition to characterizing the inherent difficulty of the multi-armed bandit problem with switching costs, our results also resolve several other open problems in online learning. One direct implication is that learning with bandit feedback against bounded-memory adaptive adversaries has a minimax regret of [EQUATION]. Another implication is that the minimax regret of online learning in adversarial Markov decision processes (MDPs) is [EQUATION]. The key to all of our results is a new randomized construction of a multi-scale random walk, which is of independent interest and likely to prove useful in additional settings.

Individual Sequence Prediction Using Memory-Efficient Context Trees

Context trees are a popular and effective tool for tasks such as compression, sequential prediction, and language modeling. We present an algebraic perspective of context trees for the task of individual sequence prediction. Our approach stems from a generalization of the notion of margin used for linear predictors. By exporting the concept of margin to context trees, we are able to cast the individual sequence prediction problem as the task of finding a linear separator in a Hilbert space, and to apply techniques from machine learning and online optimization to this problem. Our main contribution is a memory efficient adaptation of the perceptron algorithm for individual sequence prediction. We name our algorithm the shallow perceptron and prove a shifting mistake bound, which relates its performance with the performance of any sequence of context trees. We also prove that the shallow perceptron grows a context tree at a rate that is upper bounded by its mistake rate, which imposes an upper bound on the size of the trees grown by our algorithm.