Learning classification models from real-world data often requires substantial human effort devoted to instance annotation. As the instance-based annotating process can be very time-consuming and costly, we propose a novel active learning framework that builds classification models from human-annotatedregions. A region is defined by a set of conjunctive patterns that are formed by value ranges over the input features. A region label is a human assessment of the classproportionin the data population covered by the region. By leveraginglearning from label proportionsalgorithms, regions and their class proportions can be used to train instance-based classification models. However, the key challenge is that in practice, very few regions are defined already. Therefore, to identify regions important for model learning, we design ahierarchical active learning(HAL) framework, which actively builds a hierarchy of regions. Similar to the decision-tree learning process, our approach progressively divides the input data space into smaller sub-regions, solicits labels for the new regions, and retrains the base classification model with all the leaf regions. And we further develop amulti-hierarchy(forest) solution, which builds multiple shallower hierarchies that have more informative, diverse, and simpler regions. We evaluate our HAL framework on numerous impactful classification datasets as well as on a real user study - on the survival analysis of colorectal cancer patients. The results demonstrate that region-based active learning methods can learn high-quality classifiers from very few labeled regions. Hence, our framework is shown very effective in reducing the human annotation effort needed for building classification models.

Learning classification models in practice usually requires numerous labeled data for training. However, instance-based annotation can be inefficient for humans to perform. In this article, we propose and study a new type of human supervision that is fast to perform and useful for model learning. Instead of labeling individual instances, humans provide supervision to dataregions, which are subspaces of the input data space, representing subpopulations of data. Since labeling now is performed on a region level, 0/1 labeling becomes imprecise. Thus, we design the region label to be aqualitativeassessment of the class proportion, which coarsely preserves the labeling precision but is also easy for humans to do. To identify informative regions for labeling and learning, we further devise ahierarchical active learningprocess that recursively constructs a region hierarchy. This process is semisupervised in the sense that it is driven by both active learning strategies and human expertise, where humans can provide discriminative features. To evaluate our framework, we conducted extensive experiments on nine datasets as well as a real user study on a survival analysis of colorectal cancer patients. The results have clearly demonstrated the superiority of our region-based active learning framework against many instance-based active learning methods.

Continuous-time event sequences represent discrete events occurring in continuous time. Such sequences arise frequently in real-life. Usually we expect the sequences to follow some regular pattern over time. However, sometimes these patterns may be interrupted by unexpected absence or occurrences of events. Identification of these unexpected cases can be very important as they may point to abnormal situations that need human attention. In this work, we study and develop methods for detecting outliers in continuous-time event sequences, including unexpected absence and unexpected occurrences of events. Since the patterns that event sequences tend to follow may change in different contexts, we develop outlier detection methods based on point processes that can take context information into account. Our methods are based on Bayesian decision theory and hypothesis testing with theoretical guarantees. To test the performance of the methods, we conduct experiments on both synthetic data and real-world clinical data and show the effectiveness of the proposed methods.

Learning of classification models from real-world data often requires substantial human effort devoted to instance annotation. As this process can be very time-consuming and costly, finding effective ways to reduce the annotation cost becomes critical for building such models. To address this problem we explore a new type of human feedback - region-based feedback. Briefly, a region is defined as a hypercubic subspace of the input data space and represents a subpopulation of data instances; the region's label is a human assessment of the class proportion of the data subpopulation. By using learning from label proportions algorithms one can learn instance-based classifiers from such labeled regions. In general, the key challenge is that there can be infinite many regions one can define and query in a given data space. To minimize the number and complexity of region-based queries, we propose and develop a hierarchical active learning solution that aims at incrementally building a concise hierarchy of regions. Furthermore, to avoid building a possibly class-irrelevant region hierarchy, we further propose to grow multiple different hierarchies in parallel and expand those more informative hierarchies. Through experiments on numerous data sets, we demonstrate that methods using region-based feedback can learn very good classifiers from very few and simple queries, and hence are highly effective in reducing human annotation effort needed for building classification models.

In this paper, we study the problem of learning multi-class classification models from a limited set of labeled examples obtained from human annotator. We propose a new machine learning framework that learns multi-class classification models from ordered class sets the annotator may use to express not only her top class choice but also other competing classes still under consideration. Such ordered sets of competing classes are common, for example, in various diagnostic tasks. In this paper, we first develop strategies for learning multi-class classification models from examples associated with ordered class set information. After that we develop an active learning strategy that considers such a feedback. We evaluate the benefit of the framework on multiple datasets. We show that class-order feedback and active learning can reduce the annotation cost both individually and jointly.

The problems a patient suffers from can be summarized in terms of a list of patient diagnoses. The diagnoses are typically organized in a hierarchy (or a lattice structure) in which many different low-level diagnoses are covered by one or more diagnostic categories. An interesting machine learning problem is related to learning of a wide range of diagnostic models (at different levels of abstraction) that can automatically assign a diagnosis or a diagnostic category to a specific patient. While one can always approach this problem by learning models for each diagnostic task independently, an interesting open question is how one can leverage the knowledge of a diagnostic hierarchy to improve the classification and outperform independent diagnostic models. In this work, we study this problem by designing a new hierarchical classification learning framework in which multiple diagnostic classification targets are explicitly related via diagnostic hierarchy relations. By conducting experiments on MIMIC-III data and ICD-9 diagnosis hierarchy, we demonstrate that our framework leads to improved classification performance on individual diagnostic tasks when compared to independently learned diagnostic models. This improvement is stronger for diagnoses with a low prior and smaller number of positive training examples.

Real-world event sequences consist of complex mixtures of different types of events occurring in time. An event may depend on past events of the same type, as well as, the other types. Point processes define a general class of models for event sequences. ``Regressive point processes'' refer to point processes that directly model the dependency between an event and any past event, an example of which is a Hawkes process. In this work, we propose and develop a new nonparametric regressive point process model based on Gaussian processes. We show that our model can represent better many commonly observed real-world event sequences and capture the dependencies between events that are difficult to model using existing nonparametric Hawkes process variants. We demonstrate the improved predictive performance of our model against state-of-the-art baselines on multiple synthetic and real-world datasets.

Learning of classification models in practice often relies on human annotation effort in which humans assign class labels to data instances. As this process can be very time-consuming and costly, finding effective ways to reduce the annotation cost becomes critical for building such models. To solve this problem, instead of soliciting instance-based annotation we explore region-based annotation as the human feedback. A region is defined as a hyper-cubic subspace of the input space X and it covers a subpopulation of data instances that fall into this region. Each region is labeled with a number in [0,1] (in binary classification setting), representing a human estimate of the positive (or negative) class proportion in the subpopulation. To quickly discover pure regions (in terms of class proportion) in the data, we have developed a novel active learning framework that constructs regions in a hierarchical and adaptive way. Hierarchical means that regions are incrementally built into a hierarchical tree, which is done by repeatedly splitting the input space. Adaptive means that our framework can adaptively choose the best heuristic for each of the region splits. Through experiments on numerous datasets we demonstrate that our framework can identify pure regions in very few region queries. Thus our approach is shown to be effective in learning classification models from very limited human feedback.

Learning of classification models in practice often relies on nontrivial human annotation effort in which humans assign class labels to data instances. As this process can be very time consuming and costly, finding effective ways to reduce the annotation cost becomes critical for building such models. In this work we solve this problem by exploring a new approach that actively learns classification models from groups, which are subpopulations of instances, and human feedback on the groups. Each group is labeled with a number in [0,1] interval representing a human estimate of the proportion of instances with one of the class labels in this subpopulation. To form the groups to be annotated, we develop a hierarchical active learning framework that divides the whole population into smaller subpopulations, which allows us to gradually learn more refined models from the subpopulations and their class proportion labels. Our extensive experiments on numerous datasets show that our method is competitive and outperforms existing approaches for reducing the human annotation cost.

In this work, we focus on models and analysis of multivariate time series data that are organized in hierarchies. Such time series are referred to as hierarchical time series (HTS) and they are very common in business, management, energy consumption, social networks, or web traffic modeling and analysis. We propose a new flexible hierarchical forecasting framework, that takes advantage of the hierarchical relational structure to predict individual time series. Our new forecasting framework is able to (1) handle HTS modeling and forecasting problems; (2) make accurate forecasting for HTS with seasonal patterns; (3) incorporate various individual forecasting models and combine heuristics based on the HTS datasets' own characterization. The proposed framework is evaluated on a real-world web traffic data set. The results demonstrate that our approach is superior when applied to hierarchical web traffic prediction problems, and it outperforms alternative time series prediction models in terms of accuracy

A clinical decision support system and its components may malfunction due to different reasons. The objective of this work is to develop computational methods that can help us to monitor the system and assure its proper operation by promptly detecting and analyzing changes in its behavior. We develop a new change-point detection method using the Multi-Process Dynamic Linear Model. The experiments on real and simulated data show that our method outperforms existing change-point detection methods, leading to higher accuracy and shorter delay in the detection.

Building of an accurate predictive model of clinical time series for a patient is critical for understanding of the patient condition, its dynamics, and optimal patient management. Unfortunately, this process is not straightforward. First, patient-specific variations are typically large and population-based models derived or learned from many different patients are often unable to support accurate predictions for each individual patient. Moreover, time series observed for one patient at any point in time may be too short and insufficient to learn a high-quality patient-specific model just from the patient's own data. To address these problems we propose, develop and experiment with a new adaptive forecasting framework for building multivariate clinical time series models for a patient and for supporting patient-specific predictions. The framework relies on the adaptive model switching approach that at any point in time selects the most promising time series model out of the pool of many possible models, and consequently, combines advantages of the population, patient-specific and short-term individualized predictive models. We demonstrate that the adaptive model switching framework is very promising approach to support personalized time series prediction, and that it is able to outperform predictions based on pure population and patient-specific models, as well as, other patient-specific model adaptation strategies.

Annotation of classification data by humans can be a time-consuming and tedious process. Finding ways of reducing the annotation effort is critical for building the classification models in practice and for applying them to a variety of classification tasks. In this paper, we develop a new active learning framework that combines two strategies to reduce the annotation effort. First, it relies on label uncertainty information obtained from the human in terms of the Likert-scale feedback. Second, it uses active learning to annotate examples with the greatest expected change. We propose a Bayesian approach to calculate the expectation and an incremental SVM solver to reduce the time complexity of the solvers. We show the combination of our active learning strategy and the Likert-scale feedback can learn classification models more rapidly and with a smaller number of labeled instances than methods that rely on either Likert-scale labels or active learning alone.

This paper overviews and discusses our recent work on a multivariate conditional outlier detection framework for clinical applications.

Building accurate predictive models of clinical multivariate time series is crucial for understanding of the patient condition, the dynamics of a disease, and clinical decision making. A challenging aspect of this process is that the model should be flexible and adaptive to reflect well patient-specific temporal behaviors and this also in the case when the available patient-specific data are sparse and short span. To address this problem we propose and develop an adaptive two-stage forecasting approach for modeling multivariate, irregularly sampled clinical time series of varying lengths. The proposed model (1) learns the population trend from a collection of time series for past patients; (2) captures individual-specific short-term multivariate variability; and (3) adapts by automatically adjusting its predictions based on new observations. The proposed forecasting model is evaluated on a real-world clinical time series dataset. The results demonstrate the benefits of our approach on the prediction tasks for multivariate, irregularly sampled clinical time series, and show that it can outperform both the population based and patient-specific time series prediction models in terms of prediction accuracy.

We develop and test a new classification model learning algorithm that relies on the soft-label information and that is able to learn classification models more rapidly and with a smaller number of labeled instances than existing approaches.

The linear dynamical system (LDS) model is arguably the most commonly used time series model for real-world engineering and financial applications due to its relative simplicity, mathematically predictable behavior, and the fact that exact inference and predictions for the model can be done efficiently. In this work, we propose a new generalized LDS framework, gLDS, for learning LDS models from a collection of multivariate time series (MTS) data based on matrix factorization, which is different from traditional EM learning and spectral learning algorithms. In gLDS, each MTS sequence is factorized as a product of a shared emission matrix and a sequence-specific (hidden) state dynamics, where an individual hidden state sequence is represented with the help of a shared transition matrix. One advantage of our generalized formulation is that various types of constraints can be easily incorporated into the learning process. Furthermore, we propose a novel temporal smoothing regularization approach for learning the LDS model, which stabilizes the model, its learning algorithm and predictions it makes. Experiments on several real-world MTS data show that (1) regular LDS models learned from gLDS are able to achieve better time series predictive performance than other LDS learning algorithms; (2) constraints can be directly integrated into the learning process to achieve special properties such as stability, low-rankness; and (3) the proposed temporal smoothing regularization encourages more stable and accurate predictions.

This paper overviews the background, goals, past achievements and future directions of our research that aims to build a multivariate conditional anomaly detection framework for the clinical application.

Linear Dynamical System (LDS) is an elegant mathematical framework for modeling and learning Multivariate Time Series (MTS). However, in general, it is difficult to set the dimension of an LDS's hidden state space. A small number of hidden states may not be able to model the complexities of a MTS, while a large number of hidden states can lead to overfitting. In this paper, we study learning methods that impose various regularization penalties on the transition matrix of the LDS model and propose a regularized LDS learning framework (rLDS) which aims to (1) automatically shut down LDSs' spurious and unnecessary dimensions, and consequently, address the problem of choosing the optimal number of hidden states; (2) prevent the overfitting problem given a small amount of MTS data; and (3) support accurate MTS forecasting. To learn the regularized LDS from data we incorporate a second order cone program and a generalized gradient descent method into the Maximum a Posteriori framework and use Expectation Maximization to obtain a low-rank transition matrix of the LDS model. We propose two priors for modeling the matrix which lead to two instances of our rLDS. We show that our rLDS is able to recover well the intrinsic dimensionality of the time series dynamics and it improves the predictive performance when compared to baselines on both synthetic and real-world MTS datasets.

Learning probabilistic predictive models that are well calibrated is critical for many prediction and decision-making tasks in artificial intelligence. In this paper we present a new non-parametric calibration method called Bayesian Binning into Quantiles (BBQ) which addresses key limitations of existing calibration methods. The method post processes the output of a binary classification algorithm; thus, it can be readily combined with many existing classification algorithms. The method is computationally tractable, and empirically accurate, as evidenced by the set of experiments reported here on both real and simulated datasets.

that indicate how sure an expert is in the class labels she provides. If meaningful confidence information can be incorporated into a learning method, fewer patient instances may need to be labeled to learn an accurate model. In addition, while accuracy of predictions is important for any inference model, a model of patients must be interpretable so that clinicians can understand how the model is making decisions. To these ends, we develop a novel metric learning method called Confidence bAsed MEtric Learning (CAMEL) that supports inclusion of confidence labels, but also emphasizes interpretability in three ways. First, our method induces sparsity, thus producing simple models that use only a few features from patient EHRs. Second, CAMEL naturally produces confidence scores that can be taken into consideration when clinicians make treatment decisions. Third, the metrics learned by CAMEL induce multidimensional spaces where each dimension represents a different "factor" that clinicians can use to assess patients. In our experimental evaluation, we show on a real-world clinical data set that our CAMEL methods are able to learn models that are as or more accurate as other methods that use the same supervision. Furthermore, we show that when CAMEL uses confidence scores it is able to learn models as or more accurate as others we tested while using only 10% of the training instances. Finally, we perform qualitative assessments on the metrics learned by CAMEL and show that they identify and clearly articulate important factors in how the model performs inference.

Hierarchical time series (HTS) is a special class of multivariate time series where many related time series are organized in a hierarchical tree structure and they are consistent across hierarchy levels. HTS modeling is crucial and serves as the basis for business planning and management in many areas such as manufacturing inventory, energy and traffic management. However, due to machine failures, network disturbances or human maloperation, HTS data suffer from missing values across different hierarchical levels. In this paper, we study the missing value estimation problem under hierarchical web traffic settings, where the user-visit traffic are organized in various hierarchical structures, such as geographical structure and website structure. We develop an efficient algorithm, HTSImpute, to accurately estimate the missing value in multivariate noisy web traffic time series with specific hierarchical consistency in HTS settings. Our HTSImpute is able to (1) utilize the temporal dependence information within each individual time series, (2) exploit the intra-relations between time series through hierarchy, (3) guarantee the satisfaction of hierarchical consistency constraints. Results on three synthetic HTS datasets and three real-world hierarchical web traffic datasets demonstrate that our approach is able to provide more accurate and hierarchically consistent estimations than other baselines.

Learning a kernel matrix from relative comparison human feedback is an important problem with applications in collaborative filtering, object retrieval, and search. For learning a kernel over a large number of objects, existing methods face significant scalability issues inhibiting the application of these methods to settings where a kernel is learned in an online and timely fashion. In this paper we propose a novel framework called Efficient online Relative comparison Kernel LEarning (ERKLE), for efficiently learning the similarity of a large set of objects in an online manner. We learn a kernel from relative comparisons via stochastic gradient descent, one query response at a time, by taking advantage of the sparse and low-rank properties of the gradient to efficiently restrict the kernel to lie in the space of positive semidefinite matrices. In addition, we derive a passive-aggressive online update for minimally satisfying new relative comparisons as to not disrupt the influence of previously obtained comparisons. Experimentally, we demonstrate a considerable improvement in speed while obtaining improved or comparable accuracy compared to current methods in the online learning setting.

We develop a novel probabilistic ensemble framework for multi-label classification that is based on the mixtures-of-experts architecture. In this framework, we combine multi-label classification models in the classifier chains family that decompose the class posterior distribution P(Y1, …, Yd|X) using a product of posterior distributions over components of the output space. Our approach captures different input–output and output–output relations that tend to change across data. As a result, we can recover a rich set of dependency relations among inputs and outputs that a single multi-label classification model cannot capture due to its modeling simplifications. We develop and present algorithms for learning the mixtures-of-experts models from data and for performing multi-label predictions on unseen data instances. Experiments on multiple benchmark datasets demonstrate that our approach achieves highly competitive results and outperforms the existing state-of-the-art multi-label classification methods.

Learning probabilistic predictive models that are well calibrated is critical for many prediction and decision-making tasks in Data mining. This paper presents two new non-parametric methods for calibrating outputs of binary classification models: a method based on the Bayes optimal selection and a method based on the Bayesian model averaging. The advantage of these methods is that they are independent of the algorithm used to learn a predictive model, and they can be applied in a post-processing step, after the model is learned. This makes them applicable to a wide variety of machine learning models and methods. These calibration methods, as well as other methods, are tested on a variety of datasets in terms of both discrimination and calibration performance. The results show the methods either outperform or are comparable in performance to the state-of-the-art calibration methods.

We propose a new probabilistic approach for multi-label classification that aims to represent the class posterior distribution P(Y|X). Our approach uses a mixture of tree-structured Bayesian networks, which can leverage the computational advantages of conditional tree-structured models and the abilities of mixtures to compensate for tree-structured restrictions. We develop algorithms for learning the model from data and for performing multi-label predictions using the learned model. Experiments on multiple datasets demonstrate that our approach outperforms several state-of-the-art multi-label classification methods.

This paper studies multi-label classification problem in which data instances are associated with multiple, possibly high-dimensional, label vectors. This problem is especially challenging when labels are dependent and one cannot decompose the problem into a set of independent classification problems. To address the problem and properly represent label dependencies we propose and study a pairwise conditional random Field (CRF) model. We develop a new approach for learning the structure and parameters of the CRF from data. The approach maximizes the pseudo likelihood of observed labels and relies on the fast proximal gradient descend for learning the structure and limited memory BFGS for learning the parameters of the model. Empirical results on several datasets show that our approach outperforms several multi-label classification baselines, including recently published state-of-the-art methods.

OBJECTIVE: Learning of classification models in medicine often relies on data labeled by a human expert. Since labeling of clinical data may be time-consuming, finding ways of alleviating the labeling costs is critical for our ability to automatically learn such models. In this paper we propose a new machine learning approach that is able to learn improved binary classification models more efficiently by refining the binary class information in the training phase with soft labels that reflect how strongly the human expert feels about the original class labels. MATERIALS AND METHODS: Two types of methods that can learn improved binary classification models from soft labels are proposed. The first relies on probabilistic/numeric labels, the other on ordinal categorical labels. We study and demonstrate the benefits of these methods for learning an alerting model for heparin induced thrombocytopenia. The experiments are conducted on the data of 377 patient instances labeled by three different human experts. The methods are compared using the area under the receiver operating characteristic curve (AUC) score. RESULTS: Our AUC results show that the new approach is capable of learning classification models more efficiently compared to traditional learning methods. The improvement in AUC is most remarkable when the number of examples we learn from is small. CONCLUSIONS: A new classification learning framework that lets us learn from auxiliary soft-label information provided by a human expert is a promising new direction for learning classification models from expert labels, reducing the time and cost needed to label data.

The objective of multi-dimensional classification is to learn a function that accurately maps each data instance to a vector of class labels. Multi-dimensional classification appears in a wide range of applications including text categorization, gene functionality classification, semantic image labeling, etc. Usually, in such problems, the class variables are not independent, but rather exhibit conditional dependence relations among them. Hence, the key to the success of multi-dimensional classification is to effectively model such dependencies and use them to facilitate the learning. In this paper, we propose a new probabilistic approach that represents class conditional dependencies in an effective yet computationally efficient way. Our approach uses a special tree-structured Bayesian network model to represent the conditional joint distribution of the class variables given the feature variables. We develop and present efficient algorithms for learning the model from data and for performing exact probabilistic inferences on the model. Extensive experiments on multiple datasets demonstrate that our approach achieves highly competitive results when it is compared to existing state-of-the-art methods.

Development of accurate models of complex clinical time series data is critical for understanding the disease, its dynamics, and subsequently patient management and clinical decision making. Clinical time series differ from other time series applications mainly in that observations are often missing and made at irregular time intervals. In this work, we propose and test a new probabilistic approach for modeling clinical time series data that is optimized to handle irregularly sampled observations. Our model is defined by a sequence of Gaussian processes (GPs), each restricted to a window of a finite size, where dependencies among two consecutive Gaussian processes are represented using a linear dynamical system. We develop algorithms supporting both model learning and inference. Experiments on real-world clinical time series data show that our model is better for modeling clinical time series and that it outperforms or is close to alternative time series prediction models.

We study the problem of learning classification models from complex multivariate temporal data encountered in electronic health record systems. The challenge is to define a good set of features that are able to represent well the temporal aspect of the data. Our method relies on temporal abstractions and temporal pattern mining to extract the classification features. Temporal pattern mining usually returns a large number of temporal patterns, most of which may be irrelevant to the classification task. To address this problem, we present the Minimal Predictive Temporal Patterns framework to generate a small set of predictive and non-spurious patterns. We apply our approach to the real-world clinical task of predicting patients who are at risk of developing heparin induced thrombocytopenia. The results demonstrate the benefit of our approach in efficiently learning accurate classifiers, which is a key step for developing intelligent clinical monitoring systems.

In this paper, we present a new approach that lets us extract, and represent relations among terms (concepts) in the documents and uses these relations to support various document analysis applications. Our approach works by building a graph of local co-occurrence relations among terms that are extracted directly from text and by defining a global similarity metric among these terms and sets of terms using the graph and its connectivity. We demonstrate the benefit of the approach on the problem of MeSH keyword annotation of documents based on their abstracts.

Improving the performance of classifiers using pattern mining techniques has been an active topic of data mining research. In this work we introduce the recent temporal pattern mining framework for finding predictive patterns for monitoring and event detection problems in complex multivariate time series data. This framework first converts time series into time-interval sequences of temporal abstractions. It then constructs more complex temporal patterns backwards in time using temporal operators. We apply our framework to health care data of 13,558 diabetic patients and show its benefits by efficiently finding useful patterns for detecting and diagnosing adverse medical conditions that are associated with diabetes.

The focus of this paper is on how to select a small sample of examples for labeling that can help us to evaluate many different classification models unknown at the time of sampling. We are particularly interested in studying the sampling strategies for problems in which the prevalence of the two classes is highly biased toward one of the classes. The evaluation measures of interest we want to estimate as accurately as possible are those obtained from the contingency table. We provide a careful theoretical analysis on sensitivity, specificity, and precision and show how sampling strategies should be adapted to the rate of skewness in data in order to effectively compute the three aforementioned evaluation measures.

We study the problem of learning classification models from complex multivariate temporal data encountered in electronic health record systems. The challenge is to define a good set of features that are able to represent well the temporal aspect of the data. Our method relies on temporal abstractions and temporal pattern mining to extract the classification features. Temporal pattern mining usually returns a large number of temporal patterns, most of which may be irrelevant to the classification task. To address this problem, we present the minimal predictive temporal patterns framework to generate a small set of predictive and non-spurious patterns. We apply our approach to the real-world clinical task of predicting patients who are at risk of developing heparin induced thrombocytopenia. The results demonstrate the benefit of our approach in learning accurate classifiers, which is a key step for developing intelligent clinical monitoring systems.

Finding ways of incorporating auxiliary information or auxiliary data into the learning process has been the topic of active data mining and machine learning research in recent years. In this work we study and develop a new framework for classification learning problem in which, in addition to class labels, the learner is provided with an auxiliary (probabilistic) information that reflects how strong the expert feels about the class label. This approach can be extremely useful for many practical classification tasks that rely on subjective label assessment and where the cost of acquiring additional auxiliary information is negligible when compared to the cost of the example analysis and labelling. We develop classification algorithms capable of using the auxiliary information to make the learning process more efficient in terms of the sample complexity. We demonstrate the benefit of the approach on a number of synthetic and real world data sets by comparing it to the learning with class labels only.

In this paper, we consider the problem of conditional anomaly detection that aims to identify data instances with an unusual response or a class label. We develop a new non-parametric approach for conditional anomaly detection based on the soft harmonic solution, with which we estimate the confidence of the label to detect anomalous mislabeling. We further regularize the solution to avoid the detection of isolated examples and examples on the boundary of the distribution support. We demonstrate the efficacy of the proposed method on several synthetic and UCI ML datasets in detecting unusual labels when compared to several baseline approaches. We also evaluate the performance of our method on a real-world electronic health record dataset where we seek to identify unusual patient-management decisions.

Pairwise Markov Networks (PMN) are an important class of Markov networks which, due to their simplicity, are widely used in many applications such as image analysis, bioinformatics, sensor networks, etc. However, learning of Markov networks from data is a challenging task; there are many possible structures one must consider and each of these structures comes with its own parameters making it easy to overfit the model with limited data. To deal with the problem, recent learning methods build upon the L1 regularization to express the bias towards sparse network structures. In this paper, we propose a new and more flexible framework that let us bias the structure, that can, for example, encode the preference to networks with certain local substructures which as a whole exhibit some special global structure. We experiment with and show the benefit of our framework on two types of problems: learning of modular networks and learning of traffic networks models.

Choosing good features to represent objects can be crucial to the success of supervised machine learning methods. Recently, there has been a great interest in applying data mining techniques to construct new classification features. The rationale behind this approach is that patterns (feature-value combinations) could capture more underlying semantics than single features. Hence the inclusion of some patterns can improve the classification performance. Currently, most methods adopt a two-phases approach by generating all frequent patterns in the first phase and selecting the discriminative patterns in the second phase. However, this approach has limited success because it is usually very difficult to correctly identify important predictive patterns in a large set of highly correlated frequent patterns.

In this paper, we define a new query expansion method that relies on term similarity metric derived from the electric resistance network. This proposed metric lets us measure the mutual relevancy in between terms and between their groups. This paper shows how to define this metric automatically from the document collection, and then apply it in query expansion for document retrieval tasks. The experiments show this method can be used to find good expansion terms of search queries and improve document retrieval performance on two TREC genomic track datasets.

Whole-sample mass spectrometry (MS) proteomics allows for a parallel measurement of hundreds of proteins present in a variety of biospecimens. Unfortunately, the association between MS signals and these proteins is not straightforward. The need to interpret mass spectra demands the development of methods for accurate labeling of ion species in such profiles. To aid this process, we have developed a new peak-labeling procedure for associating protein and peptide labels with peaks. This computational method builds upon characteristics of proteins expected to be in the sample, such as the amino sequence, mass weight, and expected concentration within the sample. A new probabilistic score that incorporates this information is proposed. We evaluate and demonstrate our method's ability to label peaks first on simulated MS spectra and then on MS spectra from human serum with a spiked-in calibration mixture.

The increasing availability of digital documents in the last decade has prompted the development of machine learning techniques to automatically classify and organize text documents. The majority of text classification systems rely on the vector space model, which represents the documents as vectors in the term space. Each vector component is assigned a weight that reflects the importance of the term in the document. Typically, these weights are assigned using an information retrieval (IR) approach, such as the famous tf-idf function. In this work, we study two weighting schemes based on information gain and chi-square statistics. These schemes take advantage of the category label information to weight the terms according to their distributions across the different categories. We show that using these supervised weights instead of conventional unsupervised weights can greatly improve the performance of the k-nearest neighbor (KNN) classifier. Experimental evaluations, carried out on multiple text classification tasks, demonstrate the benefits of this approach in creating accurate text classifiers.

The increased availability of time series datasets prompts the development of new tools and methods that allow machine learning classifiers to better cope with time series data. Time series data are usually characterized by a high space dimensionality and a very strong correlation among features. This special nature makes the development of effective time series classifiers a challenging task. This work proposes and analyzes methods combining spectral decomposition and feature selection for time series classification problems and compares them against methods that work with original time series and time-dependent features. Briefly, our approach first applies discrete cosine transform (DCT) or discrete wavelet transform (DWT) on time series data. Then, it performs supervised feature selection/reduction by selecting only the most discriminative set of coefficients to represent the data. Experimental evaluations, carried out on multiple datasets, demonstrate the benefits of our approach in learning efficient and accurate time series classifiers.

Research articles typically introduce new results or findings and relate them to knowledge entities of immediate relevance. However, a large body of context knowledge related to the results is often not explicitly mentioned in the article. To overcome this limitation the state-of-the-art information retrieval approaches rely on the latent semantic analysis in which terms in articles are projected to a lower dimensional latent space and best possible matches in this space are identified. However, this approach may not perform well enough if the number of explicit knowledge entities in the articles is too small compared to the amount of knowledge in the domain. We address the problem by exploiting a domain knowledge layer, a rich network of relations among knowledge entities in the domain extracted from a large corpus of documents. The knowledge layer supplies the context knowledge that lets us relate different knowledge entities and hence improve the information retrieval performance. We develop and study a new framework for i) learning and aggregating the relations in the knowledge layer from the literature corpus; ii) and for exploiting these relations to improve the information-retrieval of relevant documents.

MOTIVATION: The 'reproducibility' of mass spectrometry proteomic profiling has become an intensely controversial topic. The mere mention of concern over the 'reproducibility' of data generated from any particular platform can lead to the anxiety over the generalizability of its results and its role in the future of discovery proteomics. In this study, we examine the reproducibility of proteomic profiles generated by surface-enhanced laser desorption/ionization time-of-flight mass spectrometry (SELDI-TOF-MS) across multiple data-generation sessions. We analyze the problem in terms of the reproducibility of signals, reproducibility of discriminative features and reproducibility of multivariate classification models on profiles for serum samples from early lung cancer and healthy control subjects. RESULTS: Proteomic profiles in individual data-generation sessions experience within-session variability. We show that combining data from multiple sessions introduces additional (inter-session) noise. While additional noise can affect the discriminative analysis, we show that its average effect on profiles in our study is relatively small. Moreover, for the purposes of prediction on future (previously unseen) data, classifiers trained on multi-session data are able to adapt to inter-session noise and improve their classification accuracy.

Efficient representations and solutions for large decision problems with continuous and discrete variables are among the most important challenges faced by the designers of automated decision support systems. In this paper, we describe a novel hybrid factored Markov decision process (MDP) model that allows for a compact representation of these problems, and a new hybrid approximate linear programming (HALP) framework that permits their efficient solutions. The central idea of HALP is to approximate the optimal value function by a linear combination of basis functions and optimize its weights by linear programming. We analyze both theoretical and computational aspects of this approach, and demonstrate its scale-up potential on several hybrid optimization problems.

We develop a new component analysis framework, the Noisy-Or Component Analyzer (NOCA), that targets high-dimensional binary data. NOCA is a probabilistic latent variable model that assumes the expression of observed high-dimensional binary data is driven by a small number of hidden binary sources combined via noisy-or units. The component analysis procedure is equivalent to learning of NOCA parameters. Since the classical EM formulation of the NOCA learning problem is intractable, we develop its variational approximation. We test the NOCA framework on two problems: (1) a synthetic image-decomposition problem and (2) a co-citation data analysis problem for thousands of CiteSeer documents. We demonstrate good performance of the new model on both problems. In addition, we contrast the model to two mixture-based latent-factor models: the probabilistic latent semantic analysis (PLSA) and latent Dirichlet allocation (LDA). Differing assumptions underlying these models cause them to discover different types of structure in co-citation data, thus illustrating the benefit of NOCA in building our understanding of high-dimensional data sets.

Latent factor models offer a very useful framework for modeling dependencies in high-dimensional multivariate data. In this work we investigate a class of latent factor models with hidden noisy-or units that let us decouple high dimensional vectors of observable binary random variables using a ‘small’ number of hidden binary factors. Since the problem of learning of such models from data is intractable, we develop its variational approximation. We analyze special properties of the optimization problem, in particular its “built-in” regularization effect and discuss its importance for model recovery. We test the noisy-or model on an image deconvolution problem and illustrate the ability of the variational method to succesfully learn the underlying image components. Finally, we apply the latent noisy-or model to analyze citations in a large collection of Statistical Machine Learning papers and show the benefit of the model and algorithms by discovering useful and semantically sound components characterizing the dataset.

Approximate linear programming (ALP) has emerged recently as one of the most promising methods for solving complex factored MDPs with (cid:2)nite state spaces. In this work we show that ALP solutions are not limited only to MDPs with (cid:2)nite state spaces, but that they can also be applied successfully to factored continuous-state MDPs (CMDPs). We show how one can build an ALP-based approximation for such a model and contrast it to existing solution methods. We argue that this approach offers a robust alternative for solving high dimensional continuous-state space problems. The point is supported by experiments on three CMDP problems with 24-25 continuous state factors.

Partially observable Markov decision processes (POMDPs) provide an elegant mathematical framework for modeling complex decision and planning problems in stochastic domains in which states of the system are observable only indirectly, via a set of imperfect or noisy observations. The modeling advantage of POMDPs, however, comes at a price -- exact methods for solving them are computationally very expensive and thus applicable in practice only to very simple problems. We focus on efficient approximation (heuristic) methods that attempt to alleviate the computational problem and trade off accuracy for speed. We have two objectives here. First, we survey various approximation methods, analyze their properties and relations and provide some new insights into their differences. Second, we present a number of new approximation methods and novel refinements of existing techniques. The theoretical results are supported by experiments on a problem from the agent navigation domain.