This paper tackles the problem of selecting among several linear estimators in non-parametric regression; this includes model selection for linear regression, the choice of a regularization parameter in kernel ridge regression, spline smoothing or locally weighted regression, and the choice of a kernel in multiple kernel learning. We propose a new algorithm which first estimates consistently the variance of the noise, based upon the concept of minimal penalty, which was previously introduced in the context of model selection. Then, plugging our variance estimate in Mallows' $C_L$ penalty is proved to lead to an algorithm satisfying an oracle inequality. Simulation experiments with kernel ridge regression and multiple kernel learning show that the proposed algorithm often improves significantly existing calibration procedures such as generalized cross-validation.

We propose a method for nonparametric density estimation that exhibits robustness to contamination of the training sample. This method achieves robustness by combining a traditional kernel density estimator (KDE) with ideas from classical $M$-estimation. We interpret the KDE based on a radial, positive semi-definite kernel as a sample mean in the associated reproducing kernel Hilbert space. Since the sample mean is sensitive to outliers, we estimate it robustly via $M$-estimation, yielding a robust kernel density estimator (RKDE). An RKDE can be computed efficiently via a kernelized iteratively re-weighted least squares (IRWLS) algorithm. Necessary and sufficient conditions are given for kernelized IRWLS to converge to the global minimizer of the $M$-estimator objective function. The robustness of the RKDE is demonstrated with a representer theorem, the influence function, and experimental results for density estimation and anomaly detection.

Ensemble methods for supervised machine learning have become popular due to their ability to accurately predict class labels with groups of simple, lightweight "base learners." While ensembles offer computationally efficient models that have good predictive capability they tend to be large and offer little insight into the patterns or structure in a dataset. We consider an ensemble technique that returns a model of ranked rules. The model accurately predicts class labels and has the advantage of indicating which parameter constraints are most useful for predicting those labels. An example of the rule ensemble method successfully ranking rules and selecting attributes is given with a dataset containing images of potential supernovas where the number of necessary features is reduced from 39 to 21. We also compare the rule ensemble method on a set of multi-class problems with boosting and bagging, which are two well known ensemble techniques that use decision trees as base learners, but do not have a rule ranking scheme.

Statistical relational learning techniques have been successfully applied in a wide range of relational domains. In most of these applications, the human designers capitalized on their background knowledge by following a trial-and-error trajectory, where relational features are manually defined by a human engineer, parameters are learned for those features on the training data, the resulting model is validated, and the cycle repeats as the engineer adjusts the set of features. This paper seeks to streamline application development in large relational domains by introducing a light-weight approach that efficiently evaluates relational features on pieces of the relational graph that are streamed to it one at a time. We evaluate our approach on two social media tasks and demonstrate that it leads to more accurate models that are learned faster.

Graphical Gaussian models have proven to be useful tools for exploring network structures based on multivariate data. Applications to studies of gene expression have generated substantial interest in these models, and resulting recent progress includes the development of fitting methodology involving penalization of the likelihood function. In this paper we advocate the use of multivariate $t$-distributions for more robust inference of graphs. In particular, we demonstrate that penalized likelihood inference combined with an application of the EM algorithm provides a computationally efficient approach to model selection in the $t$-distribution case. We consider two versions of multivariate $t$-distributions, one of which requires the use of approximation techniques. For this distribution, we describe a Markov chain Monte Carlo EM algorithm based on a Gibbs sampler as well as a simple variational approximation that makes the resulting method feasible in large problems.

We present techniques for gathering data that expose errors of automatic predictive models. In certain common settings, traditional methods for evaluating predictive models tend to miss rare-but-important errors---most importantly, rare cases for which the model is confident of its prediction (but wrong). In this paper we present a system that, in a game-like setting, asks humans to identify cases what will cause the predictive-model-based system to fail. Such techniques are valuable in discovering problematic cases that do not reveal themselves during the normal operation of the system, and may include cases that are rare but catastrophic. We describe the design of the system, including design iterations that did not quite work. In particular, the system incentivizes humans to provide examples that are difficult for the model to handle, by providing a reward proportional to the magnitude of the predictive model's error. The humans are asked to ``\emph{Beat the Machine}'' and find cases where the automatic model (``\emph{the Machine}'') is wrong. Experiments show that the humans using Beat the Machine identify more errors than traditional techniques for discovering errors in from predictive models, and indeed, they identify many more errors where the machine is confident it is correct. Further, the cases the humans identify seem to be not simply outliers, butcoherent areas missed completely by the model. Beat the machine identifies the ``unknown unknowns.''

This paper describes the development of a corpus of blog posts that are annotated for the presence of attempts to persuade and corresponding tactics employed in persuasive messages. We investigate the feasibility of classifying blog posts as persuasive or non-persuasive on the basis of lexical features in the text and the tactics (as provided by human annotators). Annotated tactics provide substantial assistance in classifying persuasion, particularly tactics indicating formal reasoning, deontic obligation, and discussions of possible outcomes, suggesting that learning to identify tactics may be an excellent first step to detecting attempts to persuade.

Analyzing chat traffic has important applications for both the military and the civilian world. This paper presents a case study of a real-world application of chat analysis in support of team training exercise in the military. It compares the results of an unsupervised learning approach with those of a supervised classification approach. The paper also discusses some of the specific challenges presented by this domain.

In this paper, we study a time-critical facet of context-awareness: context transitions, which we model as changes in specific context types over time, e.g., activity or location. We present results from a user-centred field study involving participant interviews and motion data capture from two mobile device sensors: the accelerometer and magnetic field sensor. The results show how the participants subjectively interpret their daily context transitions with variable granularity, and a comparison of these context transitions with mobile device motion data shows how the motion data poorly reflect the identified transitions. The results imply that care should be taken when representing and modelling users’ subjective interpretations of context, as well as the objective nature of context sensors. Furthermore, processing and usability trade-offs should be made if real-time on-device transition detection is to be implemented.

In 2009, an average American spent 3 hours per day watching TV. Recent advances in TV entertainment technologies, such as on-demand content, browsing the Internet, and 3D displays, have changed the traditional role of the TV and turned it into the center of home entertainment. Most of these technologies are personal and would benefit from seamless identification of who sits in front of the TV. In this work, we propose a practical and highly accurate solution to this problem. This solution uses a camera, which is mounted on a TV, to recognize faces of people in front of the TV. To make the approach practical, we employ online learning on graphs and show that we can learn highly accurate face models in difficult circumstances from as little as one labeled example. To evaluate our solutions, we collected a 10-hour long dataset of 8 people who watch TV. Our precision and recall are in the upper nineties, and show the promise of utilizing our approach in an embedded setting.