A new algorithm is proposed which accelerates the mini-batch k-means algorithm of Sculley (2010) by using the distance bounding approach of Elkan (2003). We argue that, when incorporating distance bounds into a mini-batch algorithm, already used data should preferentially be reused. To this end we propose using nested mini-batches, whereby data in a mini-batch at iteration t is automatically reused at iteration t+1. Using nested mini-batches presents two difficulties. The first is that unbalanced use of data can bias estimates, which we resolve by ensuring that each data sample contributes exactly once to centroids. The second is in choosing mini-batch sizes, which we address by balancing premature fine-tuning of centroids with redundancy induced slow-down. Experiments show that the resulting nmbatch algorithm is very effective, often arriving within 1% of the empirical minimum 100 times earlier than the standard mini-batch algorithm.

Process mining is a research field focused on the analysis of event data with the aim of extracting insights in processes. Applying process mining techniques on data from smart home environments has the potential to provide valuable insights in (un)healthy habits and to contribute to ambient assisted living solutions. Finding the right event labels to enable application of process mining techniques is however far from trivial, as simply using the triggering sensor as the label for sensor events results in uninformative models that allow for too much behavior (overgeneralizing). Refinements of sensor level event labels suggested by domain experts have shown to enable discovery of more precise and insightful process models. However, there exist no automated approach to generate refinements of event labels in the context of process mining. In this paper we propose a framework for automated generation of label refinements based on the time attribute of events. We show on a case study with real life smart home event data that behaviorally more specific, and therefore more insightful, process models can be found by using automatically generated refined labels in process discovery.

In an era of ubiquitous large-scale streaming data, the availability of data far exceeds the capacity of expert human analysts. In many settings, such data is either discarded or stored unprocessed in datacenters. This paper proposes a method of online data thinning, in which large-scale streaming datasets are winnowed to preserve unique, anomalous, or salient elements for timely expert analysis. At the heart of this proposed approach is an online anomaly detection method based on dynamic, low-rank Gaussian mixture models. Specifically, the high-dimensional covariances matrices associated with the Gaussian components are associated with low-rank models. According to this model, most observations lie near a union of subspaces. The low-rank modeling mitigates the curse of dimensionality associated with anomaly detection for high-dimensional data, and recent advances in subspace clustering and subspace tracking allow the proposed method to adapt to dynamic environments. Furthermore, the proposed method allows subsampling, is robust to missing data, and uses a mini-batch online optimization approach. The resulting algorithms are scalable, efficient, and are capable of operating in real time. Experiments on wide-area motion imagery and e-mail databases illustrate the efficacy of the proposed approach.

Several recent deep neural networks experiments leverage the generalist-specialist paradigm for classification. However, no formal study compared the performance of different clustering algorithms for class assignment. In this paper we perform such a study, suggest slight modifications to the clustering procedures, and propose a novel algorithm designed to optimize the performance of of the specialist-generalist classification system. Our experiments on the CIFAR-10 and CIFAR-100 datasets allow us to investigate situations for varying number of classes on similar data. We find that our \emph{greedy pairs} clustering algorithm consistently outperforms other alternatives, while the choice of the confusion matrix has little impact on the final performance.

Creativity is a complex, multi-faceted concept encompassing a variety of related aspects, abilities, properties and behaviours. If we wish to study creativity scientifically, then a tractable and well-articulated model of creativity is required. Such a model would be of great value to researchers investigating the nature of creativity and in particular, those concerned with the evaluation of creative practice. This paper describes a unique approach to developing a suitable model of how creative behaviour emerges that is based on the words people use to describe the concept. Using techniques from the field of statistical natural language processing, we identify a collection of fourteen key components of creativity through an analysis of a corpus of academic papers on the topic. Words are identified which appear significantly often in connection with discussions of the concept. Using a measure of lexical similarity to help cluster these words, a number of distinct themes emerge, which collectively contribute to a comprehensive and multi-perspective model of creativity. The components provide an ontology of creativity: a set of building blocks which can be used to model creative practice in a variety of domains. The components have been employed in two case studies to evaluate the creativity of computational systems and have proven useful in articulating achievements of this work and directions for further research.

Data Envelopment Analysis, also known as DEA, is a non-parametric method for performing frontier analysis. It uses linear programming to estimate the efficiency of multiple decision-making units and it is commonly used in production, management and economics. The technique was first proposed by Charnes, Cooper and Rhodes in 1978 and since then it became a valuable tool for estimating production frontiers. Update: The Datumbox Machine Learning Framework is now open-source and free to download. When I first encountered the method 5-6 years ago, I was amazed by the originality of the algorithm, its simplicity and the cleverness of the ideas that it used. I was even more amazed to see that the technique worked well outside of its usual applications (financial, operation research etc) since it could be successfully applied in Online Marketing, Search Engine Ranking and for creating composite metrics.

The Best Paper award for this year's International Conference on Very Large Data Bases (VLDB) goes to "Compressed Linear Algebra for Large-Scale Machine Learning", authored by a PhD candidate at the University of Maryland and four senior researchers from IBM. Their method for compressing matrices for linear algebra operations promises to provide users significant increases in speed with less memory. In particular, the compression technology provides benefits at two different parts of the data science process. Before training a model, a data scientist typically goes through multiple iterations of feature engineering. Common feature engineering tasks include examining the data with descriptive statistics and transforming the values in columns to better suit the assumptions built into different types of machine learning models.

Here we study non-convex composite optimization: first, a finite-sum of smooth but non-convex functions, and second, a general function that admits a simple proximal mapping. Most research on stochastic methods for composite optimization assumes convexity or strong convexity of each function. In this paper, we extend this problem into the non-convex setting using variance reduction techniques, such as prox-SVRG and prox-SAGA. We prove that, with a constant step size, both prox-SVRG and prox-SAGA are suitable for non-convex composite optimization, and help the problem converge to a stationary point within $O(1/\epsilon)$ iterations. That is similar to the convergence rate seen with the state-of-the-art RSAG method and faster than stochastic gradient descent. Our analysis is also extended into the min-batch setting, which linearly accelerates the convergence. To the best of our knowledge, this is the first analysis of convergence rate of variance-reduced proximal stochastic gradient for non-convex composite optimization.

We describe a probabilistic PARAFAC/CANDECOMP (CP) factorization for multiway (i.e., tensor) data that incorporates auxiliary covariates, SupCP. SupCP generalizes the supervised singular value decomposition (SupSVD) for vector-valued observations, to allow for observations that have the form of a matrix or higher-order array. Such data are increasingly encountered in biomedical research and other fields. We describe a likelihood-based latent variable representation of the CP factorization, in which the latent variables are informed by additional covariates. We give conditions for identifiability, and develop an EM algorithm for simultaneous estimation of all model parameters. SupCP can be used for dimension reduction, capturing latent structures that are more accurate and interpretable due to covariate supervision. Moreover, SupCP specifies a full probability distribution for a multiway data observation with given covariate values, which can be used for predictive modeling. We conduct comprehensive simulations to evaluate the SupCP algorithm, and we apply it to a facial image database with facial descriptors (e.g., smiling / not smiling) as covariates. Software is available at https://github.com/lockEF/SupCP .

We study the relationship between online Gaussian process (GP) regression and kernel least mean squares (KLMS) algorithms. While the latter have no capacity of storing the entire posterior distribution during online learning, we discover that their operation corresponds to the assumption of a fixed posterior covariance that follows a simple parametric model. Interestingly, several well-known KLMS algorithms correspond to specific cases of this model. The probabilistic perspective allows us to understand how each of them handles uncertainty, which could explain some of their performance differences.