Shihao Ji

Multitask Compressive Sensing

Compressive sensing (CS) is a framework whereby one performs N nonadaptive measurements to constitute a vector v isin RN used to recover an approximation u isin RM desired signal u isin RM with N << M this is performed under the assumption that u is sparse in the basis represented by the matrix Psi RMtimesM. It has been demonstrated that with appropriate design of the compressive measurements used to define v, the decompressive mapping vrarru may be performed with error ||u-u||2 2 having asymptotic properties analogous to those of the best adaptive transform-coding algorithm applied in the basis Psi. The mapping vrarru constitutes an inverse problem, often solved using l1 regularization or related techniques. In most previous research, if L > 1 sets of compressive measurements {vi}i=1,L are performed, each of the associated {ui}i=1,Lare recovered one at a time, independently. In many applications the L ldquotasksrdquo defined by the mappings virarrui are not statistically independent, and it may be possible to improve the performance of the inversion if statistical interrelationships are exploited. In this paper, we address this problem within a multitask learning setting, wherein the mapping vrarru for each task corresponds to inferring the parameters (here, wavelet coefficients) associated with the desired signal vi, and a shared prior is placed across all of the L tasks. Under this hierarchical Bayesian modeling, data from all L tasks contribute toward inferring a posterior on the hyperparameters, and once the shared prior is thereby inferred, the data from each of the L individual tasks is then employed to estimate the task-dependent wavelet coefficients. An empirical Bayesian procedure for the estimation of hyperparameters is considered; two fast inference algorithms extending the relevance vector machine (RVM) are developed. Example results on several data sets demonstrate the effectiveness and robustness of the proposed algorithms.

Intent-based diversification of web search results: metrics and algorithms

We study the problem of web search result diversification in the case where intent based relevance scores are available. A diversified search result will hopefully satisfy the information need of user-L.s who may have different intents. In this context, we first analyze the properties of an intent-based metric, ERR-IA, to measure relevance and diversity altogether. We argue that this is a better metric than some previously proposed intent aware metrics and show that it has a better correlation with abandonment rate. We then propose an algorithm to rerank web search results based on optimizing an objective function corresponding to this metric and evaluate it on shopping related queries.

Cost-sensitive feature acquisition and classification

There are many sensing challenges for which one must balance the effectiveness of a given measurement with the associated sensing cost. For example, when performing a diagnosis a doctor must balance the cost and benefit of a given test (measurement), and the decision to stop sensing (stop performing tests) must account for the risk to the patient and doctor (malpractice) for a given diagnosis based on observed data. This motivates a cost-sensitive classification problem in which the features (sensing results) are not given a priori; the algorithm determines which features to acquire next, as well as when to stop sensing and make a classification decision based on previous observations (accounting for the costs of various types of errors, as well as the rewards of being correct). We formally define the cost-sensitive classification problem and solve it via a partially observable Markov decision process (POMDP). While the POMDP constitutes an intuitively appealing formulation, the intrinsic properties of classification tasks resist application of it to this problem. We circumvent the difficulties of the POMDP via a myopic approach, with an adaptive stopping criterion linked to the standard POMDP. The myopic algorithm is computationally feasible, easily handles continuous features, and seamlessly avoids repeated actions. Experiments with several benchmark data sets show that the proposed method yields state-of-the-art performance, and importantly our method uses only a small fraction of the features that are generally used in competitive approaches.

Variational Bayes for continuous hidden Markov models and its application to active learning

In this paper, we present a variational Bayes (VB) framework for learning continuous hidden Markov models (CHMMs), and we examine the VB framework within active learning. Unlike a maximum likelihood or maximum a posteriori training procedure, which yield a point estimate of the CHMM parameters, VB-based training yields an estimate of the full posterior of the model parameters. This is particularly important for small training sets since it gives a measure of confidence in the accuracy of the learned model. This is utilized within the context of active learning, for which we acquire labels for those feature vectors for which knowledge of the associated label would be most informative for reducing model-parameter uncertainty. Three active learning algorithms are considered in this paper: 1) query by committee (QBC), with the goal of selecting data for labeling that minimize the classification variance, 2) a maximum expected information gain method that seeks to label data with the goal of reducing the entropy of the model parameters, and 3) an error-reduction-based procedure that attempts to minimize classification error over the test data. The experimental results are presented for synthetic and measured data. We demonstrate that all of these active learning methods can significantly reduce the amount of required labeling, compared to random selection of samples for labeling.

Global ranking by exploiting user clicks

It is now widely recognized that user interactions with search results can provide substantial relevance information on the documents displayed in the search results. In this paper, we focus on extracting relevance information from one source of user interactions, i.e., user click data, which records the sequence of documents being clicked and not clicked in the result set during a user search session. We formulate the problem as a global ranking problem, emphasizing the importance of the sequential nature of user clicks, with the goal to predict the relevance labels of all the documents in a search session. This is distinct from conventional learning to rank methods that usually design a ranking model defined on a single document; in contrast, in our model the relational information among the documents as manifested by an aggregation of user clicks is exploited to rank all the documents jointly. In particular, we adapt several sequential supervised learning algorithms, including the conditional random field (CRF), the sliding window method and the recurrent sliding window method, to the global ranking problem. Experiments on the click data collected from a commercial search engine demonstrate that our methods can outperform the baseline models for search results re-ranking.

Nonmyopic Multiaspect Sensing With Partially Observable Markov Decision Processes

We consider the problem of sensing a concealed or distant target by interrogation from multiple sensors situated on a single platform. The available actions that may be taken are selection of the next relative target-platform orientation and the next sensor to be deployed. The target is modeled in terms of a set of states, each state representing a contiguous set of target-sensor orientations over which the scattering physics is relatively stationary. The sequence of states sampled at multiple target-sensor orientations may be modeled as a Markov process. The sensor only has access to the scattered fields, without knowledge of the particular state being sampled, and, therefore, the problem is modeled as a partially observable Markov decision process (POMDP). The POMDP yields a policy, in which the belief state at any point is mapped to a corresponding action. The nonmyopic policy is compared to an approximate myopic approach, with example results presented for measured underwater acoustic scattering data

Adaptive multiaspect target classification and detection with hidden Markov models

Target detection and classification are considered based on backscattered signals observed from a sequence of target-sensor orientations, with the measurements performed as a function of orientation (angle) at a fixed range. The theory of optimal experiments is applied to adaptively optimize the sequence of target-sensor orientations considered. This is motivated by the fact that if fewer, better-chosen measurements are used then targets can be recognized more accurately with less time and expense. Specifically, based on the previous sequence of observations O/sub t/={O/sub 1/,...,O/sub t/}, the technique determines what change in relative target-sensor orientation /spl Delta//spl theta//sub t+1/ is optimal for performing measurement t+1, to yield observation O/sub t+1/. The target is assumed distant or hidden, and, therefore, the absolute target-sensor orientation is unknown. We detail the adaptive-sensing algorithm, employing a hidden Markov model representation of the multiaspect scattered fields, and example classification and detection results are presented for underwater targets using acoustic scattering data.

Adaptive Multimodality Sensing of Landmines

The problem of adaptive multimodality sensing of landmines is considered based on electromagnetic induction (EMI) and ground-penetrating radar (GPR) sensors. Two formulations are considered based on a partially observable Markov decision process (POMDP) framework. In the first formulation, it is assumed that sufficient training data are available, and a POMDP model is designed based on physics-based features, with model selection performed via a variational Bayes analysis of several possible models. In the second approach, the training data are assumed absent or insufficient, and a lifelong-learning approach is considered, in which exploration and exploitation are integrated. We provide a detailed description of both formulations, with example results presented using measured EMI and GPR data, for buried mines and clutter

A general integrative genomic feature transcription factor binding site prediction method applied to analysis of USF1 binding in cardiovascular disease

Transcription factors are key mediators of human complex disease processes. Identifying the target genes of transcription factors will increase our understanding of the biological network leading to disease risk. The prediction of transcription factor binding sites (TFBSs) is one method to identify these target genes; however, current prediction methods need improvement. We chose the transcription factor upstream stimulatory factor l (USF1) to evaluate the performance of our novel TFBS prediction method because of its known genetic association with coronary artery disease (CAD) and the recent availability of USF1 chromatin immunoprecipitation microarray (ChIP-chip) results. The specific goals of our study were to develop a novel and accurate genome-scale method for predicting USF1 binding sites and associated target genes to aid in the study of CAD. Previously published USF1 ChIP-chip data for 1 per cent of the genome were used to develop and evaluate several kernel logistic regression prediction models. A combination of genomic features (phylogenetic conservation, regulatory potential, presence of a CpG island and DNaseI hypersensitivity), as well as position weight matrix (PWM) scores, were used as variables for these models. Our most accurate predictor achieved an area under the receiver operator characteristic curve of 0.827 during cross-validation experiments, significantly outperforming standard PWM-based prediction methods. When applied to the whole human genome, we predicted 24,010 USF1 binding sites within 5 kilobases upstream of the transcription start site of 9,721 genes. These predictions included 16 of 20 genes with strong evidence of USF1 regulation. Finally, in the spirit of genomic convergence, we integrated independent experimental CAD data with these USF1 binding site prediction results to develop a prioritised set of candidate genes for future CAD studies. We have shown that our novel prediction method, which employs genomic features related to the presence of regulatory elements, enables more accurate and efficient prediction of USF1 binding sites. This method can be extended to other transcription factors identified in human disease studies to help further our understanding of the biology of complex disease.

Multitask Classification by Learning the Task Relevance

We consider the problem of multitask learning (MTL), in which we simultaneously learn classifiers for multiple data sets (tasks), with sharing of intertask data as appropriate. We introduce a set of relevance parameters that control the degree to which data from other tasks are used in estimating the current tasks classifier parameters. The set of relevance parameters are learned by maximizing their posterior probability, yielding an expectation-maximization (EM) algorithm. We illustrate the effectiveness of our approach through experimental results on a practical data set.