Understanding the demographics of users of online social networks has important applications for health, marketing, and public messaging. In this paper, we predict the demographics of Twitter users based on whom they follow. Whereas most prior approaches rely on a supervised learning approach, in which individual users are labeled with demographics, we instead create a distantly labeled dataset by collecting audience measurement data for 1,500 websites (e.g., 50% of visitors to gizmodo.com are estimated to have a bachelor's degree). We then fit a regression model to predict these demographics using information about the followers of each website on Twitter. The resulting average held-out correlation is .77 across six different variables (gender, age, ethnicity, education, income, and child status). We additionally validate the model on a smaller set of Twitter users labeled individually for ethnicity and gender, finding performance that is surprisingly competitive with a fully supervised approach.

User-generated reviews are valuable resources for decision making. Identifying the aspect categories discussed in a given review sentence (e.g., “food” and “service” in restaurant reviews) is an important task of sentiment analysis and opinion mining. Given a predefined aspect category set, most previous researches leverage hand-crafted features and a classification algorithm to accomplish the task. The crucial step to achieve better performance is feature engineering which consumes much human effort and may be unstable when the product domain changes. In this paper, we propose a representation learning approach to automatically learn useful features for aspect category detection. Specifically, a semi-supervised word embedding algorithm is first proposed to obtain continuous word representations on a large set of reviews with noisy labels. Afterwards, we propose to generate deeper and hybrid features through neural networks stacked on the word vectors. A logistic regression classifier is finally trained with the hybrid features to predict the aspect category. The experiments are carried out on a benchmark dataset released by SemEval-2014. Our approach achieves the state-of-the-art performance and outperforms the best participating team as well as a few strong baselines.

Fast and efficient learning over large bodies of commonsense knowledge is a key requirement for cognitive systems. Semantic web knowledge bases provide an important new resource of ground facts from which plausible inferences can be learned. This paper applies structured logistic regression with analogical generalization (SLogAn) to make use of structural as well as statistical information to achieve rapid and robust learning. SLogAn achieves state-of-the-art performance in a standard triplet classification task on two data sets and, in addition, can provide understandable explanations for its answers.

The F-measure, which has originally been introduced in information retrieval, is nowadays routinely used as a performance metric for problems such as binary classification, multi-label classification, and structured output prediction. Optimizing this measure is a statistically and computationally challenging problem, since no closed-form solution exists. Adopting a decision-theoretic perspective, this article provides a formal and experimental analysis of different approaches for maximizing the F-measure. We start with a Bayes-risk analysis of related loss functions, such as Hamming loss and subset zero-one loss, showing that optimizing such losses as a surrogate of the F-measure leads to a high worst-case regret. Subsequently, we perform a similar type of analysis for F-measure maximizing algorithms, showing that such algorithms are approximate, while relying on additional assumptions regarding the statistical distribution of the binary response variables. Furthermore, we present a new algorithm which is not only computationally efficient but also Bayes-optimal, regardless of the underlying distribution. To this end, the algorithm requires only a quadratic (with respect to the number of binary responses) number of parameters of the joint distribution. We illustrate the practical performance of all analyzed methods by means of experiments with multi-label classification problems.

How can knowing the location of my friends be used to more accurately predict my location? This paper explores socially-aware location prediction under a particularly challenging setting where the underlying interactions and social network are unknown and must be inferred over continuous spatiotemporal data. Our method samples inferred network topology using a linear regression model to predict future individual locations. We present an in-depth empirical study comparing different network models and network sampling regimes under a bootstrapped sampling baseline. Furthermore, our qualitative analysis demonstrates the value of social information in population mobility modeling under our application’s challenges.

Generalized Additive Models (GAM) are a widely popular class of regression models to forecast electricity demand, due to their high accuracy, flexibility and interpretability. However, the residuals of the fitted GAM are typically heteroscedastic and leptokurtic caused by the nature of energy data. In this paper we propose a novel approach to estimate the time-varying conditional variance of the GAM residuals, which we call the GAM2 algorithm. It allows utility companies and network operators to assess the uncertainty of future electricity demand and incorporate it into their planning processes. The basic idea of our algorithm is to apply another GAM to the squared residuals to explain the dependence of uncertainty on exogenous variables. Empirical evidence shows that the residuals rescaled by the estimated conditional variance are approximately normal. We combine our modeling approach with online learning algorithms that adjust for dynamic changes in the distributions of demand. We illustrate our method by a case study on data from RTE, the operator of the French transmission grid.

Residential and commercial buildings are responsible for about 40% of primary energy consumption in the United States, hence improving their energy efficiency could have important sustainability benefits. The design of a building has tremendous effect on its energy profile, and recently there has been an increased interest in developing optimization methods that support the design of high performance buildings. Previous approaches are either based on simulation optimization or on training an accurate predictive model that is queried during the optimization. We propose a method that more tightly integrates the machine learning and optimization components, by employing active learning during optimization. In particular, we use a Gaussian Process (GP) model for the prediction and active learning and multi-objective genetic algorithm NSGA-II for the optimization. We develop a comprehensive and publicly available benchmark for building design optimization. We evaluate 5 machine learning approaches on our dataset, and show that the GP model is competitive, in addition to being well-suited for the active learning setting. We compare our optimization approach against the 2-stage approach and simulation optimization. Our results show that our approach produces solutions at the Pareto frontier compared to the other two approaches, while using only a fraction of the simulations and time.

Bike sharing systems consist of a fleet of bikes placed in a network of docking stations. These bikes can then be rented and returned to any of the docking stations after usage. Predicting unrealized bike demand at locations currently without bike stations is important for effectively designing and expanding bike sharing systems. We predict pairwise bike demand for New York City’s Citi Bike system. Since the system is driven by daily commuters we focus only on the morning rush hours between 7:00 AM to 11:00 AM during weekdays. We use taxi usage, weather and spatial variables as covariates to predict bike demand, and further analyze the influence of precipitation and day of week. We show that aggregating stations in neighborhoods can substantially improve predictions. The presented model can assist planners by predicting bike demand at a macroscopic level, between pairs of neighborhoods.

We aim to profile characteristics of areas of variant units across a district, city or a country. Studying attributes of areas can be very useful in several situations. In the past, research has focused mainly on studying specific char- acteristics of areas using a few selected attributes. In this paper we propose an alternative view on neighbourhood profiles. Instead of characterising a neighbourhood through a set of attributes such as those collected by the census, we propose use of a low-dimensional fea- ture representation, or embedding, created from one or more input sources. The purpose of the embeddings is having a generic representation for entities that can do well across several downstream tasks such as regression for attributes prediction.

In many estimation problems, e.g. linear and logistic regression, we wish to minimize an unknown objective given only unbiased samples of the objective function. Furthermore, we aim to achieve this using as few samples as possible. In the absence of computational constraints, the minimizer of a sample average of observed data -- commonly referred to as either the empirical risk minimizer (ERM) or the $M$-estimator -- is widely regarded as the estimation strategy of choice due to its desirable statistical convergence properties. Our goal in this work is to perform as well as the ERM, on every problem, while minimizing the use of computational resources such as running time and space usage. We provide a simple streaming algorithm which, under standard regularity assumptions on the underlying problem, enjoys the following properties: * The algorithm can be implemented in linear time with a single pass of the observed data, using space linear in the size of a single sample. * The algorithm achieves the same statistical rate of convergence as the empirical risk minimizer on every problem, even considering constant factors. * The algorithm's performance depends on the initial error at a rate that decreases super-polynomially. * The algorithm is easily parallelizable. Moreover, we quantify the (finite-sample) rate at which the algorithm becomes competitive with the ERM.