Big data characterized by variety can be divided into 3 principal categories: numeric structured data, semi-structured data, and unstructured multimedia data involving audio, video, and text. Decision making requires multiple analytical engines suitable for each type of data, programming languages, algorithms, visualization tools, and user interfaces. More often than not, industrial analytics is conducted ad hoc by lashing together analytics components such as distributed data sources, analytics engines, and algorithms. This kind of piecemeal approach ignores scale, security, governance, reliability, model management and fault tolerance that are paramount for industrial strength analytics. A unified, versatile, and robust architecture that combines various components in a single integrated platform is the need of the hour. Teradata Vantage (TD Vantage) is such a platform for delivering production quality enterprise analytics at scale. In this paper, we outline the proposed TD Vantage (available in the market and under continuous development) that unifies data, engines, and algorithms operating in a seamless symphony. We will demonstrate its capabilities through three proofs of concept biz: image data using TensorFlow, text data using Spark, and transaction data using Aster (now renamed Machine Learning Engine or MLE), with Teradata orchestrating interactions among the various components.

Ease-of-use analytics at scale is the holy grail of industrial strength machine learning. In order to reap benefits from the mother-lode of business related data; tools, technologies, and analytical functions should operate in perpetual symphony to derive insightful business outcomes. While there have been advances in APIs, algorithms, and user interfaces, building an end to end workflow spanning data ingestion, data preparation, model training, model scoring, visualization and finally continuous improvement and model management received limited investment. In this paper we demonstrate an analytical workflow that integrates multiple analytical tools and techniques for image recognition wrapped in the model management framework. As analytics in industry is maturing, analytics implementations are no longer one-off, but are components of Analytics Operations known as AnalyticsOps. We introduce the notion of Model Quality Index (MQI) to track model performance. The MQI is similar to Process Capability Index (PCI) common in 6σ programs in manufacturing. Our solution combines relational databases (Teradata DB), Machine Learning (Teradata/Aster), Deep Learning (TensorFlow), Hadoop Distributed File System (HDFS), and user interface tools over a communication fabric (Teradata QueryGrid). In particular, we demonstrate a hand written word recognition use-case for an enterprise customer cast in a model management workflow for repeatable deployments across a range of businesses.

Ease-of-use analytics at scale is the holy grail of industrial strength machine learning. While there have been advances in APIs, algorithms, and user interfaces, building an end to end work flow involving data ingestion, data preparation, model training, model scoring, and visualization received limited investment and effort producing only marginal innovation. This paper outlines a proof of concept that demonstrates an analytical workflow that integrates multiple analytical tools and techniques for image recognition. The solution combines Relational Databases and Machine Learning (teradata), Deep Learning (TensorFlow), Distributed File System (HDFS), Graphical Processing Units, and user interface tools over a communication fabric (teradata QueryGrid). In particular, we demonstrate hand written word recognition through an application of Convolutional Neural Networks in TensorFlow and teradata's custom analytical functions to recognize first names and last names in the payee field section of financial negotiable instruments such as in a cheque. We hope that this paper will serve as a guide to successful implementation of analytical workflows in production.

The need for fast on-line algorithms to analyze high data-rate measurements is a vital element in production settings. Given the ever-increasing number of data sources coupled with increasing complexity of applications, and workload patterns, anomaly detection methods should be light-weight and must operate in real-time. In many modern applications, data arrive in a streaming fashion. Therefore, the underlying assumption of classical methods that the data is a sample from a stable distribution is not valid, and Gaussian and non-parametric based methods such as the control chart and boxplot are inadequate. Streaming data is an ever-changing superposition of distributions. Detection of such changes in real-time is one of the fundamental challenges. We propose low-complexity robust modifications to the conventional Tukey boxplot based on fast highly efficient robust estimates of scale. Results using synthetic as well as real-world data show that our methods outperform the Tukey boxplot and methods based on Gaussian limits.

The class imbalance problem is pervasive in machine learning. To accurately classify the minority class, current methods rely on sampling schemes to close the gap between classes, or on the application of error costs to create algorithms which favor the minority class. Since the sampling schemes and costs must be specified, these methods are highly dependent on the class distributions present in the training set. This makes them difficult to apply in settings where the level of imbalance changes, such as in online streaming data. Often they cannot handle multi-class problems. We present a novel single-class algorithm called Class Conditional Nearest Neighbor Distribution (CCNND), which mitigates the effects of class imbalance through local geometric structure in the data. Our algorithm can be applied seamlessly to problems with any level of imbalance or number of classes, and new examples are simply added to the training set. We show that it performs as well as or better than top sampling and cost-weighting methods on four imbalanced datasets from the UCI Machine Learning Repository, and then apply it to streaming data from the oil and gas industry alongside a modified nearest neighbor algorithm. Our algorithm's competitive performance relative to the state-of-the-art, coupled with its extremely simple implementation and automatic adjustment for minority classes, demonstrates that it is worth further study.

Data centers are growing in size and complexity driven by trends such as cloud computing and on-line services. Such large data centers pose several challenges for system management. Key among them is anomaly detection which is required to monitor and analyze metrics across several thousands servers and across multiple layers of abstractions to detect anomalous system behavior. In practice, multiple anomaly detection tools are used to continuously raise alarms across multiple metrics and servers. These alarms include both true positives and false alarms. Administrators and management tools act on these alarms for diagnosis and deeper root cause analysis and take appropriate management actions to mitigate the anomalous behaviors. Given the scale and scope of the system, the administrators and management tools are overwhelmed with the large number of alarms at any given instant, many of which are false alarms. It is therefore necessary to prioritize and rank these alarms, so as to take timely actions that maintain the service level agreements for the data center. Existing techniques for such ranking are ad-hoc and not scalable. We propose ranking windows of monitored metrics based on their probability of occurrence. We explain how these probabilities can be computed based either on the false positive rates for which the accompanying anomaly detectors were designed, or, when available, on the probability models underlying the false positive rates. In the simplest case, the ranking procedure reduces to computing the Z-score of the observed measurements and computing a statistic from a window of Z-scores to use as a basis for ranking. The proposed techniques are reliable, lightweight and easy to deploy in the modern data center. We have validated these techniques on synthetic data containing injected anomalies and on data acquired from production data centers.

Many practical data streams are typically composed of several states known as regimes. In this paper, we invoke phase space reconstruction methods from non-linear time series and dynamical systems for regime detection. But the data collected from sensors is normally noisy, does not have constant amplitude and is sometimes plagued by shifts in the mean. All these aspects make modeling even more difficult. We propose a representation of the time series in the phase space with a modified embedding, which is invariant to translation and scale. The features we use for regime detection are based on comparing trajectory segments in the modified embedding space with cross-correntropy, which is a generalized correlation function. We apply our algorithm to non-linear oscillations, and compare its performance with the standard time delay embedding.

Online anomaly detection is an important step in data center management, requiring light-weight techniques that provide sufficient accuracy for subsequent diagnosis and management actions. This paper presents statistical techniques based on the Tukey and Relative Entropy statistics, and applies them to data collected from a production environment and to data captured from a testbed for multi-tier web applications running on server class machines. The proposed techniques are lightweight and improve over standard Gaussian assumptions in terms of performance.

In streaming and sensor data applications, the problems of synopsis construction and outlier detection are important. Due to their low complexity, desirable properties and relative ease of understanding, wavelet based techniques are often used for both synopsis construction and anomaly detection. In streaming data literature, Mallat's algorithm is often used to achieve a Haar wavelet decomposition in O(n) time. However, there is one limitation to this popular technique, in that it leads to a dyadic decomposition of data. We demonstrate that the property of non-dyadicity is of considerable use in synopsis construction and anomaly detection. In this regard we present several application results, a synopsis data structure for streaming data that is an order of magnitude superior to the popular Haar based wavelet technique, a method for finding anomalies for sensor data over non-dyadic hierarchies, etc. In our work, we enable non-dyadicity by proposing a Mallat like construction for a wavelet system that admits non-dyadic basis. Our algorithm builds a non-dyadic hierarchical structure, and is more efficient than the state of the art construction. We prove the correctness of our construction by showing that our basis functions demonstrates the properties of a wavelet system.