Real-world relational data are seldom stationary, yet traditional collaborative filtering algorithms generally rely on this assumption. Motivated by our sales prediction problem, we propose a factor-based algorithm that is able to take time into account. By introducing additional factors for time, we formalize this problem as a tensor factorization with a special constraint on the time dimension. Further, we provide a fully Bayesian treatment to avoid tuning parameters and achieve automatic model complexity control. To learn the model we develop an efficient sampling procedure that is capable of analyzing large-scale data sets. This new algorithm, called Bayesian Probabilistic Tensor Factorization (BPTF), is evaluated on several real-world problems including sales prediction and movie recommendation. Empirical results demonstrate the superiority of our temporal model.

No abstract

The widespread application of deep learning has changed the landscape of computation in the data center. In particular, personalized recommendation for content ranking is now largely accomplished leveraging deep neural networks. However, despite the importance of these models and the amount of compute cycles they consume, relatively little research attention has been devoted to systems for recommendation. To facilitate research and to advance the understanding of these workloads, this paper presents a set of real-world, productionscale DNNs for personalized recommendation coupled with relevant performance metrics for evaluation. In addition to releasing a set of open-source workloads, we conduct indepth analysis that underpins future system design and optimization for at-scale recommendation: Inference latency varies by 60% across three Intel server generations, batching and co-location of inferences can drastically improve latency-bounded throughput, and the diverse composition of recommendation models leads to different optimization strategies.Preprint. Under submission.

No abstract

No abstract

No abstract

We investigate the use of dimensionality reduction techniques for the classification of stellar spectra selected from the SDSS. Using local linear embedding (LLE), a technique that preserves the local (and possibly non-linear) structure within high dimensional data sets, we show that the majority of stellar spectra can be represented as a one dimensional sequence within a three dimensional space. The position along this sequence is highly correlated with spectral temperature. Deviations from this "stellar locus" are indicative of spectra with strong emission lines (including misclassified galaxies) or broad absorption lines (e.g. Carbon stars). Based on this analysis, we propose a hierarchical classification scheme using LLE that progressively identifies and classifies stellar spectra in a manner that requires no feature extraction and that can reproduce the classic MK classifications to an accuracy of one type. I. INTRODUCTIONAutomatic classification of stellar data is a problem as old as the use of computers in astronomy. Since survey projects have begun presenting us with spectral data from literally hundreds of thousands of sources, it has become untenable to classify all of them "by hand." Computer science provides several tools and algorithms available to help us alleviate the human experts' work load. Neural networks [Storrie-Lombardi et al. 1994, Singh et al. 1998] and Principal Component Analysis (PCA; Deeming 1964) are among the most popular computational tools presently used to tackle large sets of astronomical data. In this paper, we will consider a relatively new method: Local Linear Embedding (LLE) [Roweis and Saul 2000].At the heart of automated stellar classification is dimensionality reduction: taking a large number N of D 1 dimensional data (in this paper we consider N = 49, 529 stellar spectra sampled over D = 500 wavelength bins), and projecting the data onto a basis such that the first d D dimensions contain the bulk of the physical information encoded in the data. LLE attempts to reduce the dimensionality of the input data points while preserving the non-linear relationships between them. It does so by analyzing the data incrementally, in small neighborhoods, rather than all at once. This is particularly useful for astronomical classification, as we shall see below, since it can be simultaneously sensitive to continuum and line shapes. Thus, LLE ought to provide a more robust object classification from fewer projected dimensions than PCA. Vanderplas and Connolly (2009) explored LLE as a means of characterizing the spectral energy distributions of galaxies. They found the method very effective and, in some cases, more accurate than traditional tests at distinguishing different types of galaxies (broad-and narrow-line QSO's, emission line galaxies, quiescent galaxies, and absorption galaxies; see their Figure 2) without the need to identify and measure individual features in the galaxies' spectra. We use their code 1 to analyze 49,529 stellar spectra from the SDSS Data Release 7 (DR7) [Abazaji...

No abstract