Selecting homes with energy-efficient infrastructure is important for renters, because infrastructure influences energy consumption more than in-home behavior.Personalized energy estimates can guide prospective tenants toward energy-efficient homes, but this information is not readily available. Utility estimates are not typically offered to house-hunters, and existing technologies like carbon calculators require users to answer (prohibitively) many questions that may require considerable research to answer. For the task of providing personalized utility estimates to prospective tenants, we present a cost-based model for feature selection at training time, where all features are available and costs assigned to each feature reflect the difficulty of acquisition. At test time, we have immediate access to some features but others are difficult to acquire (costly). In this limited-information setting, we strategically order questions we ask each user, tailored to previous information provided, to give the most accurate predictions while minimizing the cost to users. During the critical first 10 questions that our approach selects, prediction accuracy improves equally to fixed order approaches, but prediction certainty is higher.

With the advances in the field of medical informatics, automated clinical decision support systems are becoming the de facto standard in personalized diagnosis. In order to establish high accuracy and confidence in personalized diagnosis, massive amounts of distributed, heterogeneous, correlated and high-dimensional patient data from different sources such as wearable sensors, mobile applications, Electronic Health Record (EHR) databases etc. need to be processed. This requires learning both locally and globally due to privacy constraints and/or distributed nature of the multi-modal medical data. In the last decade, a large number of meta-learning techniques have been proposed in which local learners make online predictions based on their locally-collected data instances, and feed these predictions to an ensemble learner,which fuses them and issues a global prediction. However, most of these works do not provide performance guarantees or, when they do,these guarantees are asymptotic. None of these existing works provide confidence estimates about the issued predictions or rate of learning guarantees for the ensemble learner. In this paper, we provide a systematic ensemble learning method called Hedged Bandits, which comes with both long run (asymptotic) and short run (rate of learning) performance guarantees. Moreover, we show that our proposed method outperforms all existing ensemble learning techniques, even in the presence of concept drift.

Electricity disaggregation focuses on classification ofindividual appliances by monitoring aggregate electricalsignals. In this paper we present a novel algorithmto automatically correct labels, discard contaminatedtraining samples, and boost signal to noise ratio throughhigh frequency noise reduction. We also propose amethod for prioritized classification which classifies applianceswith the most intense signals first. When testedon four houses in Kaggles Belkin dataset, these methodsautomatically relabel over 77% of all training samplesand decrease error rate by an average of 45% in bothreal power and high frequency noise classification.

While an immense amount of work has gone into novel clustering algorithms, little work has focused on developing compact, domain-specific explanations for the results of the clustering algorithms. Attaching semantic meaning to a cluster has numerous benefits, including the ability for such a description to be both human and machine readable. In this paper, we assume that the clusters are given to us, and find the minimal set of features that can differentiate one cluster from the remaining set of samples. We formulate this problem as an integer linear program. By using samples not belonging to the cluster in the optimization formulation, the resulting description will be minimal and contain no false positives. The efficacy of this method is demonstrated on simulation data and real-world malware data run in a sandbox that collects behavioral characteristics. In the case of malware, once it has been clustered, it would have been sent to a reverse engineer who would have been tasked with creating the actual meaning of the clustering results and disseminating this information through signatures or indicators of compromise. This is a time-consuming process that can take hours to weeks depending on the complexity of the malware family. The methods presented in this paper automatically generate optimal signatures, which can then be quickly propagated to help contain the spread of a malware family.

Utilities use Demand Response (DR) to balance supply and demand in the electric grid by involving customers in efforts to reduce electricity consumption during peak periods. To implement and adapt DR under dynamically changing conditions of the grid, reliable prediction of reduced consumption is critical. However, despite the wealth of research on electricity consumption prediction and DR being long in practice, the problem of reduced consumption prediction remains largely un-addressed. In this paper, we identify unique computational challenges associated with the prediction of reduced consumption and contrast this to that of normal consumption and DR baseline prediction. We propose a novel ensemble model that leverages different sequences of daily electricity consumption on DR event days as well as contextual attributes for reduced consumption prediction. We demonstrate the success of our model on a large, real-world, high resolution dataset from a university microgrid comprising of over 950 DR events across a diverse set of 32 buildings. Our model achieves an average error of 13.5%, an 8.8% improvement over the baseline. Our work is particularly relevant for buildings where electricity consumption is not tied to strict schedules. Our results and insights should prove useful to the researchers and practitioners working in the sustainable energy domain.

In the past few years various government funding organizations such as the U.S. National Institutes of Health and the U.S.\ National Science Foundation have provided access to large publicly-available online databases documenting the grants that they have funded over the past few decades. These databases provide an excellent opportunity for the application of statistical text analysis techniques to infer useful quantitative information about how funding patterns have changed over time. In this paper we analyze data from the National Cancer Institute (part of National Institutes of Health) and show how text classification techniques provide a useful starting point for analyzing how funding for cancer research has evolved over the past 20 years in the United States.

Relation extraction, which learns semantic relations of concept pairs from text, is an approach for mining commonsense knowledge. This paper investigates an approach for relation extraction, which helps expand a commonsense knowledge base with little labor work. We proposed a framework that learns new pairs from Chinese corpora by adopting concept pairs in Chinese commonsense knowledge base as seeds. Multiple instance learning is utilized as the learning algorithm for predicting relation for unseen pairs. The performance of our system could be improved by learning multiple iterations. The results in each iteration are manually evaluated and processed to next iteration as seeds. Our experiments extracted new pairs for relations “AtLocation”, “CapableOf”, and “HasProperty”. This study showed that new pairs could be extracted from text without huge humans work.

We're hearing a lot about automated machine learning lately, inspired in part by growing demand and the shortage of data scientists. But like many innovations, automated machine learning did not simply appear out of the blue; it is the product of at least twenty years of development. Before Unica Software launched its successful suite of marketing automation software, the company's primary business was predictive analytics, with a particular focus on neural networks. In 1995, Unica introduced Pattern Recognition Workbench (PRW), a software package that used an automated grid search to optimize model tuning for neural networks. Three years later, Unica partnered with Group 1 Software (now owned by Pitney Bowes) to market Model 1, a tool that automated model selection over four different types of predictive models.

While I think nonconvex optimization is a very interesting area of research, I often feel like the community overstates its importance in ML. Finding a global optimum almost never happens in successful usage of neural nets (we use early-stopping with a validation set) and is not necessarily the best-idea for many applications. Rather, I feel that selecting a proper objective function, class of models, and regularization strategy are just as important considerations for ML. That said, much of the ML/statistical theory only holds when the empirical risk minimizer is actually found (M-estimation). One way to bridge this theoretical gap is of course to explicitly ensure you can find global optima via superior optimization or changing the model (eg.

Instructor: Patrick Winston In this lecture, we explore support vector machines in some mathematical detail. We use Lagrange multipliers to maximize the width of the street given certain constraints. If needed, we transform vectors into another space, using a kernel function.