Robust bounds for classification via selective sampling

Cesa-Bianchi, Nicolò; Gentile, Claudio; Orabona, Francesco

doi:10.1145/1553374.1553390

Cited by 37 publications

(40 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…. , K, a weight vector w i ∈ R d and a correlation matrix A i ∈ R d×d , and operates similarly to 2nd-order (or ridge regression)-like algorithms (Hoerl & Kennard, 1970;Azoury & Warmuth, 2001;CesaBianchi et al, 2005) (see also, e.g., (Strehl & Littman, 2008;Crammer et al, 2009a;Cesa-Bianchi et al, 2009;Dekel et al, 2010)). The weight vectors are initialized to zero, and the matrices A i to (1 + α) 2 times the identity matrix I of size d. For brevity, we denote by A a single matrix of size dK × dK defined to be the block-diagonal matrix A = diag(A 1 , A 2 , .…”

Section: The New Bandit Algorithmmentioning

confidence: 99%

Multiclass classification with bandit feedback using adaptive regularization

2012

Self Cite

View full text Add to dashboard Cite

We present a new multiclass algorithm in the bandit framework, where after making a prediction, the learning algorithm receives only partial feedback, i.e., a single bit of right-orwrong, rather then the true label. Our algorithm is based on the 2nd-order Perceptron, and uses upper-confidence bounds to trade off exploration and exploitation. We analyze this algorithm in a partial adversarial setting, where instances are chosen adversarially, while the labels are chosen according to a linear probabilistic model, which is also chosen adversarially. We show a regret of O( √ T log T ), which improves over the current best bounds of O(T 2/3 ) in the fully adversarial setting. We evaluate our algorithm on nine real-world text classification problems, obtaining state-of-the-art results, even compared with non-bandit online algorithms, especially when label noise is introduced.

show abstract

Section: The New Bandit Algorithmmentioning

confidence: 99%

Multiclass classification with bandit feedback using adaptive regularization

2012

Self Cite

View full text Add to dashboard Cite

show abstract

“…Specifically, there are two kinds of settings for online active learning, selective sampling setting (Cavallanti et al 2009;Cesa-Bianchi et al 2009;Dekel et al 2010;Orabona and CesaBianchi 2011) and label efficient learning setting. We summarize their differences in several aspects.…”

Section: Online Active Learningmentioning

confidence: 99%

Online Passive-Aggressive Active learning

2016

View full text Add to dashboard Cite

We investigate online active learning techniques for online classification tasks. Unlike traditional supervised learning approaches, either batch or online learning, which often require to request class labels of each incoming instance, online active learning queries only a subset of informative incoming instances to update the classification model, aiming to maximize classification performance with minimal human labelling effort during the entire online learning task. In this paper, we present a new family of online active learning algorithms called Passive-Aggressive Active (PAA) learning algorithms by adapting the Passive-Aggressive algorithms in online active learning settings. Unlike conventional Perceptron-based approaches that employ only the misclassified instances for updating the model, the proposed PAA learning algorithms not only use the misclassified instances to update the classifier, but also exploit correctly classified examples with low prediction confidence. Specifically, we propose several variants of PAA algorithms to tackle three types of online learning tasks: binary classification, multi-class classification, and cost-sensitive classification. We give the mistake bounds of the proposed algorithms in theory, and conduct extensive experiments to evaluate the empirical performance of our techniques on both standard and large-scale datasets, in which the encouraging results validate the empirical effectiveness of the proposed algorithms.

show abstract

“…In selective sampling 0 ≤ κ ≤ 1 is a parameter of the algorithm, n is the number of steps with a margin less than , and the bound holds for any for any 0 < < 1. -Bianchi et al (2009) analyze a learning setting which is complementary to the hybrid setting introduced in this paper. They consider the selective sampling problem in which inputs are arbitrarily generated by an adversary while labels a noisy observations of a linear hypothesis.…”

Section: Related Workmentioning

confidence: 99%

Learning with stochastic inputs and adversarial outputs

Lazaric¹,

Munos²

2012

Journal of Computer and System Sciences

View full text Add to dashboard Cite

Most of the research in online learning is focused either on the problem of adversarial classification (i.e., both inputs and labels are arbitrarily chosen by an adversary) or on the traditional supervised learning problem in which samples are independent and identically distributed according to a stationary probability distribution. Nonetheless, in a number of domains the relationship between inputs and outputs may be adversarial, whereas input instances are i.i.d. from a stationary distribution (e.g., user preferences). This scenario can be formalized as a learning problem with stochastic inputs and adversarial outputs. In this paper, we introduce this novel stochastic-adversarial learning setting and we analyze its learnability. In particular, we show that in binary classification, given a hypothesis space H with finite VC-dimension, it is possible to design an algorithm which incrementally builds a suitable finite set of hypotheses from H used as input for an exponentially weighted forecaster and achieves a cumulative regret of order O( nV C(H) log n) with overwhelming probability. This result shows that whenever inputs are i.i.d. it is possible to solve any binary classification problem using a finite VCdimension hypothesis space with a sub-linear regret independently from the way labels are generated (either stochastic or adversarial). We also discuss extensions to multi-label classification, regression, learning from experts and bandit settings with stochastic side information, and application to games.

show abstract

Robust bounds for classification via selective sampling

Cited by 37 publications

References 9 publications

Multiclass classification with bandit feedback using adaptive regularization

Multiclass classification with bandit feedback using adaptive regularization

Online Passive-Aggressive Active learning

Learning with stochastic inputs and adversarial outputs

Contact Info

Product

Resources

About