We study the problem of classifying data in a given taxonomy when classifications associated with multiple and/or partial paths are allowed. We introduce a new algorithm that incrementally learns a linear-threshold classifier for each node of the taxonomy. A hierarchical classification is obtained by evaluating the trained node classifiers in a top-down fashion. To evaluate classifiers in our multipath framework, we define a new hierarchical loss function, the H-loss, capturing the intuition that whenever a classification mistake is made on a node of the taxonomy, then no loss should be charged for any additional mistake occurring in the subtree of that node. Making no assumptions on the mechanism generating the data instances, and assuming a linear noise model for the labels, we bound the H-loss of our on-line algorithm in terms of the H-loss of a reference classifier knowing the true parameters of the label-generating process. We show that, in expectation, the excess cumulative H-loss grows at most logarithmically in the length of the data sequence. Furthermore, our analysis reveals the precise dependence of the rate of convergence on the eigenstructure of the data each node observes. Our theoretical results are complemented by a number of experiments on texual corpora. In these experiments we show that, after only one epoch of training, our algorithm performs much better than Perceptron-based hierarchical classifiers, and reasonably close to a hierarchical support vector machine.

We introduce confidence-weighted linear classifiers, which add parameter confidence information to linear classifiers. Online learners in this setting update both classifier parameters and the estimate of their confidence. The particular online algorithms we study here maintain a Gaussian distribution over parameter vectors and update the mean and covariance of the distribution with each instance. Empirical evaluation on a range of NLP tasks show that our algorithm improves over other state of the art online and batch methods, learns faster in the online setting, and lends itself to better classifier combination after parallel training. 1.

A selective sampling algorithm is a learning algorithm for classification that, based on the past observed data, decides whether to ask the label of each new instance to be classified. In this paper, we introduce a general technique for turning linear-threshold classification algorithms from the general additive family into randomized selective sampling algorithms. For the most popular algorithms in this family we derive mistake bounds that hold for individual sequences of examples. These bounds

In this paper we show that several known algorithms for sequential prediction problems (including Weighted Majority and the quasi-additive family of Grove, Littlestone, and Schuurmans), for playing iterated games (including Freund and Schapire's Hedge and MW, as well as the -strategies of Hart and Mas-Colell), and for boosting (including AdaBoost) are special cases of a general decision strategy based on the notion of potential. By analyzing this strategy we derive known performance bounds, as well as new bounds, as simple corollaries of a single general theorem. Besides offering a new and unified view on a large family of algorithms, we establish a connection between potential-based analysis in learning and their counterparts independently developed in game theory. By exploiting this connection, we show that certain learning problems are instances of more general game-theoretic problems. In particular, we describe a notion of generalized regret and show its applications in learning theory.

Abstract. In this paper we investigate selective sampling, a learning model where the learner observes a sequence of i.i.d. unlabeled instances each time deciding whether to query the label of the current instance. We assume that labels are binary and stochastically related to instances via a linear probabilistic function whose coefficients are arbitrary and unknown. We then introduce a new selective sampling rule and show that its expected regret (with respect to the classifier knowing the underlying linear function and observing the label realization after each prediction) grows not much faster than the number of sampled labels. Furthermore, under additional assumptions on the true margin distribution, we prove that the number of sampled labels grows only logarithmically in the number of observed instances. Experiments carried out on a text categorization problem show that: (1) our selective sampling algorithm performs better than the Perceptron algorithm even when the latter is given the true label after each classification; (2) when allowed to observe the true label after each classification, the performance of our algorithm remains the same. Finally, we note that by expressing our selective sampling rule in dual variables we can learn nonlinear probabilistic functions via the kernel machinery. 1

A large number of variants of the Perceptron algorithm have been proposed and partially evaluated in recent work. One type of algorithm aims for noise tolerance by replacing the last hypothesis of the perceptron with another hypothesis or a vote among hypotheses. Another type simply adds a margin term to the perceptron in order to increase robustness and accuracy, as done in support vector machines. A third type borrows further from support vector machines and constrains the update function of the perceptron in ways that mimic soft-margin techniques. The performance of these algorithms, and the potential for combining different techniques, has not been studied in depth. This paper provides such an experimental study and reveals some interesting facts about the algorithms. In particular the perceptron with margin is an effective method for tolerating noise and stabilizing the algorithm. This is surprising since the margin in itself is not designed or used for noise tolerance, and there are no known guarantees for such performance. In most cases, similar performance is obtained by the voted-perceptron which has the advantage that it does not require parameter selection. Techniques using soft margin ideas are run-time intensive and do not give additional performance benefits. The results also highlight the difficulty with automatic parameter selection which is required with some of these variants.

We present AROW, a new online learning algorithm that combines several useful properties: large margin training, confidence weighting, and the capacity to handle non-separable data. AROW performs adaptive regularization of the prediction function upon seeing each new instance, allowing it to perform especially well in the presence of label noise. We derive a mistake bound, similar in form to the second order perceptron bound, that does not assume separability. We also relate our algorithm to recent confidence-weighted online learning techniques and show empirically that AROW achieves state-of-the-art performance and notable robustness in the case of non-separable data. 1

Shifting bounds for on-line classification algorithms ensure good performance on any sequence of examples that is well predicted by a sequence of smoothly changing classifiers. When proving shifting bounds for kernel-based classifiers, one also faces the problem of storing a number of support vectors that can grow unboundedly, unless an eviction policy is used to keep this number under control. In this paper, we show that shifting and on-line learning on a budget can be combined surprisingly well. First, we introduce and analyze a shifting Perceptron algorithm achieving the best known shifting bounds while using an unlimited budget. Second, we show that by applying to the Perceptron algorithm the simplest possible eviction policy, which discards a random support vector each time a new one comes in, we achieve a shifting bound close to the one we obtained with no budget restrictions. More importantly, we show that our randomized algorithm strikes the optimal trade-off U = Θ ` √ B ´ between budget B and norm U of the largest classifier in the comparison sequence.

Confidence-weighted (CW) learning [6], an online learning method for linear classifiers, maintains a Gaussian distributions over weight vectors, with a covariance matrix that represents uncertainty about weights and correlations. Confidence constraints ensure that a weight vector drawn from the hypothesis distribution correctly classifies examples with a specified probability. Within this framework, we derive a new convex form of the constraint and analyze it in the mistake bound model. Empirical evaluation with both synthetic and text data shows our version of CW learning achieves lower cumulative and out-of-sample errors than commonly used first-order and second-order online methods. 1

We design and analyze interacting online algorithms for multitask classification that perform better than independent learners whenever the tasks are related in a certain sense. We formalize task relatedness in different ways, and derive formal guarantees on the performance advantage provided by interaction. Our online analysis gives new stimulating insights into previously known co-regularization techniques, such as the multitask kernels and the margin correlation analysis for multiview learning. In the last part we apply our approach to spectral co-regularization: we introduce a natural matrix extension of the quasiadditive algorithm for classification and prove bounds depending on certain unitarily invariant norms of the matrix of task coefficients. 1