Sequential approaches for learning datum-wise sparse representations

Abstract

In supervised classification, data representation is usually considered at the dataset level: one looks for the "best" representation of data assuming it to be the same for all the data in the data space. We propose a different approach where the representations used for classification are tailored to each datum in the data space. One immediate goal is to obtain sparse datum-wise representations: our approach learns to build a representation specific to each datum that contains only a small subset of the features, thus allowing classification to be fast and efficient. This representation is obtained by way of a sequential decision process that sequentially chooses which features to acquire before classifying a particular point; this process is learned through algorithms based on Reinforcement Learning.The proposed method performs well on an ensemble of medium-sized sparse classification problems. It offers an alternative to global sparsity approaches, and is a natural framework for sequential classification problems. The method extends easily to a whole family of sparsity-related problem which would otherwise require developing specific solutions. This is the case in particular for cost-sensitive and limited-budget classification, where feature acquisition is costly and is often performed sequentially. Finally, our approach can handle non-differentiable loss functions or combinatorial optimization encountered in more complex feature selection problems.

Sequential Cost-Sensitive Feature Acquisition

“…This phenomenon is due to the presence of expensive features that clearly bring relevant information. A similar behaviour is observed with GreedyMiser and with B-REAM, but the latter seems more agile and able to better benefit from relevant expensive features 7 . We suppose that this is due to the use of reinforcement-learning in- Fig.…”

“…∇ γ,θ fγ,i(zt).ci (7) with a t sampled w.r.t. f γ (z t ), and f γ,i is the i-th component of the output of f γ .…”

“…Regarding reinforcement methods, algorithms to learn acquisition policies have been proposed for example using imitation policies [9], however this requires an oracle to guide learning. [7] presents a method where the "state" of the process is represented as a vector of the acquired features (built following a pre-defined heuristic), this representation state is then used to learn and follow the acquisition policy. Visual attention models such as [12], which often rely on policy-gradient, are also close to our work, while being specific to a particular type of inputs (images).…”

Early text classification: a Naïve solution

“…In [5] the authors extend the MDP proposed for sequential text classification to deal with any other type of data. The formulation is almost the same as in [4], although this time the MDP can decide what feature to sample from the instance under analysis (i.e., there is no sequential input).…”

“…This is not a surprising result if we notice that this problem is highly imbalanced: the imbalance ratio for training and test partitions is of 12.1 and 9.56, respectively. Furthermore, the reduction of the vocabulary may affect significantly this particular domain (the jargon used in chat conversations 5 Police officers acted as children, predators are real.…”