Multi-step ahead forecasting is still an open challenge in time series forecasting. Several approaches that deal with this complex problem have been proposed in the literature but an extensive comparison on a large number of tasks is still missing. This paper aims to fill this gap by reviewing existing strategies for multi-step ahead forecasting and comparing them in theoretical and practical terms. To attain such an objective, we performed a large scale comparison of these different strategies using a large experimental benchmark (namely the 111 series from the NN5 forecasting competition). In addition, we considered the effects of deseasonalization, input variable selection, and forecast combination on these strategies and on multi-step ahead forecasting at large. The following three findings appear to be consistently supported by the experimental results: Multiple-Output strategies are the best performing approaches, deseasonalization leads to uniformly improved forecast accuracy, and input selection is more effective when performed in conjunction with deseasonalization.

We generated a dataset of 200 GB with 10^9 features, to test our recent b-bit minwise hashing algorithms for training very large-scale logistic regression and SVM. The results confirm our prior work that, compared with the VW hashing algorithm (which has the same variance as random projections), b-bit minwise hashing is substantially more accurate at the same storage. For example, with merely 30 hashed values per data point, b-bit minwise hashing can achieve similar accuracies as VW with 2^14 hashed values per data point. We demonstrate that the preprocessing cost of b-bit minwise hashing is roughly on the same order of magnitude as the data loading time. Furthermore, by using a GPU, the preprocessing cost can be reduced to a small fraction of the data loading time. Minwise hashing has been widely used in industry, at least in the context of search. One reason for its popularity is that one can efficiently simulate permutations by (e.g.,) universal hashing. In other words, there is no need to store the permutation matrix. In this paper, we empirically verify this practice, by demonstrating that even using the simplest 2-universal hashing does not degrade the learning performance.

In data sets with many more features than observations, independent screening based on all univariate regression models leads to a computationally convenient variable selection method. Recent efforts have shown that in the case of generalized linear models, independent screening may suffice to capture all relevant features with high probability, even in ultra-high dimension. It is unclear whether this formal sure screening property is attainable when the response is a right-censored survival time. We propose a computationally very efficient independent screening method for survival data which can be viewed as the natural survival equivalent of correlation screening. We state conditions under which the method admits the sure screening property within a general class of single-index hazard rate models with ultra-high dimensional features. An iterative variant is also described which combines screening with penalized regression in order to handle more complex feature covariance structures. The methods are evaluated through simulation studies and through application to a real gene expression dataset.

Video games are increasingly producing huge datasets available for analysis resulting from players engaging in interactive environments. These datasets enable investigation of individual player behavior at a massive scale, which can lead to reduced production costs and improved player retention. We present an approach for modeling player retention in Madden NFL 11, a commercial football game. Our approach encodes gameplay patterns of specific players as feature vectors and models player retention as a regression problem. By building an accurate model of player retention, we are able to identify which gameplay elements are most influential in maintaining active players. The outcome of our tool is recommendations which will be used to influence the design of future titles in the Madden NFL series.

Variable selection problems are typically addressed under apenalized optimization framework. Nonconvex penalties such as the minimax concave plus (MCP) and smoothly clipped absolute deviation(SCAD), have been demonstrated to have the properties of sparsity practically and theoretically. In this paper we propose a new nonconvex penalty that we call exponential-type penalty. The exponential-type penalty is characterized by a positive parameter,which establishes a connection with the ell 0 and ell 1 penalties.We apply this new penalty to sparse supervised learning problems. To solve to resulting optimization problem, we resort to a reweighted ell 1 minimization method. Moreover, we devise an efficient method for the adaptive update of the tuning parameter. Our experimental results are encouraging. They show that the exponential-type penalty is competitive with MCP and SCAD.

Important ecological phenomena are often observed indirectly. Consequently, probabilistic latent variable models provide an important tool, because they can include explicit models of the ecological phenomenon of interest and the process by which it is observed. However, existing latent variable methods rely on hand-formulated parametric models, which are expensive to design and require extensive preprocessing of the data. Nonparametric methods (such as regression trees) automate these decisions and produce highly accurate models. However, existing tree methods learn direct mappings from inputs to outputs — they cannot be applied to latent variable models. This paper describes a methodology for integrating nonparametric tree methods into probabilistic latent variable models by extending functional gradient boosting. The approach is presented in the context of occupancy-detection (OD) modeling, where the goal is to model the distribution of a species from imperfect detections. Experiments on 12 real and 3 synthetic bird species compare standard and tree-boosted OD models (latent variable models) with standard and tree-boosted logistic regression models (without latent structure). All methods perform similarly when predicting the observed variables, but the OD models learn better representations of the latent process. Most importantly, tree-boosted OD models learn the best latent representations when nonlinearities and interactions are present.

Ordinal regression is an important research topic in machine learning. It aims to automatically determine the implied rating of a data item on a fixed, discrete rating scale. In this paper, we present a novel ordinal regression approach via manifold learning, which is capable of uncovering the embedded nonlinear structure of the data set according to the observations in the highdimensional feature space. By optimizing the order information of the observations and preserving the intrinsic geometry of the data set simultaneously, the proposed algorithm provides the faithful ordinal regression to the new coming data points. To offer more general solution to the data with natural tensor structure, we further introduce the multilinear extension of the proposed algorithm, which can support the ordinal regression of high order data like images. Experiments on various data sets validate the effectiveness of the proposed algorithm as well as its extension.

Applications that adapt to a particular end user often make inaccurate predictions during the early stages when training data is limited. Although an end user can improve the learning algorithm by labeling more training data, this process is time consuming and too ad hoc to target a particular area of inaccuracy. To solve this problem, we propose a new learning algorithm based on Locally Weighted Logistic Regression for feature labeling by end users, enabling them to point out which features are important for a class, rather than provide new training instances. In our user study, the first allowing ordinary end users to freely choose features to label directly from text documents, our algorithm was more effective than others at leveraging end users’ feature labels to improve the learning algorithm. Our results strongly suggest that allowing users to freely choose features to label is a promising method for allowing end users to improve learning algorithms effectively.

In order to better protect and conserve biodiversity, ecologists use machine learning and statistics to understand how species respond to their environment and to predict how they will respond to future climate change, habitat loss and other threats. A fundamental modeling task is to estimate the probability that a given species is present in (or uses) a site, conditional on environmental variables such as precipitation and temperature. For a limited number of species, survey data consisting of both presence and absence records are available, and can be used to fit a variety of conventional classification and regression models. For most species, however, the available data consist only of occurrence records --- locations where the species has been observed. In two closely-related but separate bodies of ecological literature, diverse special-purpose models have been developed that contrast occurrence data with a random sample of available environmental conditions. The most widespread statistical approaches involve either fitting an exponential model of species' conditional probability of presence, or fitting a naive logistic model in which the random sample of available conditions is treated as absence data; both approaches have well-known drawbacks, and do not necessarily produce valid probabilities. After summarizing existing methods, we overcome their drawbacks by introducing a new scaled binomial loss function for estimating an underlying logistic model of species presence/absence. Like the Expectation-Maximization approach of Ward et al. and the method of Steinberg and Cardell, our approach requires an estimate of population prevalence, $\Pr(y=1)$, since prevalence is not identifiable from occurrence data alone. In contrast to the latter two methods, our loss function is straightforward to integrate into a variety of existing modeling frameworks such as generalized linear and additive models and boosted regression trees. We also demonstrate that approaches by Lele and Keim and by Lancaster and Imbens that surmount the identifiability issue by making parametric data assumptions do not typically produce valid probability estimates.

In this paper we present a data-driven approach to modeling end user energy consumption in residential and commercial buildings. Our model is based upon a data set of monthly electricity and gas bills, collected by a utility over the course of several years, for approximately 6,500 buildings in Cambridge, MA. In addition, we use publicly available tax assessor records and geographical survey information to determine corresponding features for the buildings. Using both parametric and non-parametric learning methods, we learn models that predict distributions over energy usage based upon these features, and use these models to develop two end-user systems. For utilities or authorized institutions (those who may obtain access to the full data) we provide a system that visualizes energy consumption for each building in the city; this allows companies to quickly identify outliers (buildings which use much more energy than expected even after conditioning on the relevant predictors), for instance allowing them to target homes for potential retrofits or tiered pricing schemes. For other end users, we provide an interface for entering their own electricity and gas usage, along with basic information about their home, to determine how their consumption compares to that of similar buildings as predicted by our model. Merely allowing users to contextualize their consumption in this way, relating it to the consumption in similar buildings, can itself produce behavior changes to significantly reduce consumption.