Skip to main content

2017 | Buch

Outlier Analysis

insite
SUCHEN

Über dieses Buch

This book provides comprehensive coverage of the field of outlier analysis from a computer science point of view. It integrates methods from data mining, machine learning, and statistics within the computational framework and therefore appeals to multiple communities. The chapters of this book can be organized into three categories:Basic algorithms: Chapters 1 through 7 discuss the fundamental algorithms for outlier analysis, including probabilistic and statistical methods, linear methods, proximity-based methods, high-dimensional (subspace) methods, ensemble methods, and supervised methods.Domain-specific methods: Chapters 8 through 12 discuss outlier detection algorithms for various domains of data, such as text, categorical data, time-series data, discrete sequence data, spatial data, and network data.Applications: Chapter 13 is devoted to various applications of outlier analysis. Some guidance is also provided for the practitioner.The second edition of this book is more detailed and is written to appeal to both researchers and practitioners. Significant new material has been added on topics such as kernel methods, one-class support-vector machines, matrix factorization, neural networks, outlier ensembles, time-series methods, and subspace methods. It is written as a textbook and can be used for classroom teaching.

Inhaltsverzeichnis

Frontmatter
Chapter 1. An Introduction to Outlier Analysis
Abstract
Outliers are also referred to as abnormalities, discordants, deviants, or anomalies in the data mining and statistics literature. In most applications, the data is created by one or more generating processes, which could either reflect activity in the system or observations collected about entities. When the generating process behaves unusually, it results in the creation of outliers. Therefore, an outlier often contains useful information about abnormal characteristics of the systems and entities that impact the data generation process. The recognition of such unusual characteristics provides useful application-specific insights.
Charu C. Aggarwal
Chapter 2. Probabilistic and Statistical Models for Outlier Detection
Abstract
The earliest methods for outlier detection were rooted in probabilistic and statistical models and date back to the nineteenth century [180]. These methods were proposed well before the advent and popularization of computer technology and were therefore designed without much focus on practical issues such as data representation or computational efficiency. Nevertheless, the underlying mathematical models are extremely useful and have eventually been adapted to a variety of computational scenarios.
Charu C. Aggarwal
Chapter 3. Linear Models for Outlier Detection
Abstract
The attributes in real data are usually highly correlated. Such dependencies provide the ability to predict attributes from one another. The notions of prediction and anomaly detection are intimately related. Outliers are, after all, values that deviate from expected (or predicted) values on the basis of a particular model. Linear models focus on the use of interattribute dependencies to achieve this goal. In the classical statistics literature, this process is referred to as regression modeling.
Charu C. Aggarwal
Chapter 4. Proximity-Based Outlier Detection
Abstract
Proximity-based techniques define a data point as an outlier when its locality (or proximity) is sparsely populated. The proximity of a data point may be defined in a variety of ways, which are subtly different from one another but are similar enough to merit unified treatment within a single chapter.
Charu C. Aggarwal
Chapter 5. High-Dimensional Outlier Detection: The Subspace Method
Abstract
Many real data sets are very high dimensional. In some scenarios, real data sets may contain hundreds or thousands of dimensions. With increasing dimensionality, many of the conventional outlier detection methods do not work very effectively. This is an artifact of the well-known curse of dimensionality. In high-dimensional space, the data becomes sparse, and the true outliers become masked by the noise effects of multiple irrelevant dimensions, when analyzed in full dimensionality.
Charu C. Aggarwal
Chapter 6. Outlier Ensembles
Abstract
Ensemble analysis is a popular method used to improve the accuracy of various data mining algorithms. Ensemble methods combine the outputs of multiple algorithms or base detectors to create a unified output. The basic idea of the approach is that some algorithms will do well on a particular subset of points whereas other algorithms will do better on other subsets of points. However, the ensemble combination is often able to perform more robustly across the board because of its ability to combine the outputs of multiple algorithms. In this chapter, will use the terms base detector and component detector interchangeably to denote the individual algorithms whose outputs are combined to create the final result.
Charu C. Aggarwal
Chapter 7. Supervised Outlier Detection
Abstract
The discussions in the previous chapters focus on the problem of unsupervised outlier detection in which no prior information is available about the abnormalities in the data. In such scenarios, many of the anomalies found correspond to noise or other uninteresting phenomena. It has been observed [338, 374, 531] in diverse applications such as system anomaly detection, financial fraud, and Web robot detection that interesting anomalies are often highly specific to particular types of abnormal activity in the underlying application. In such cases, an unsupervised outlier detection method might discover noise, which is not specific to that activity, and therefore may not be of interest to an analyst. In many cases, different types of abnormal instances could be present, and it may be desirable to distinguish among them. For example, in an intrusion-detection scenario, different types of intrusion anomalies are possible, and the specific type of an intrusion is important information.
Charu C. Aggarwal
Chapter 8. Outlier Detection in Categorical, Text, and Mixed Attribute Data
Abstract
The discussion in the previous chapters has primarily focused on numerical data. However, the setting of numerical data represents a gross oversimplification because categorical attributes are ubiquitous in real-world data. For example, although demographic data may contain quantitative attributes such as the age, most other attributes such as gender, race, and ZIP code are categorical. Data collected from surveys may often contain responses to multiple-choice questions that are categorical. Similarly, many types of data such as the names of people and entities, IP-addresses, and URLs are inherently categorical. In many cases, categorical and numeric attributes are found in the same data set. Such mixed-attribute data are often challenging to machine-learning applications because of the difficulties in treating the various types of attributes in a homogeneous and consistent way.
Charu C. Aggarwal
Chapter 9. Time Series and Multidimensional Streaming Outlier Detection
Abstract
The temporal and streaming outlier-detection scenarios arise in the context of many applications such as sensor data, mechanical systems diagnosis, medical data, network intrusion data, newswire text posts, or financial posts. In such problem settings, the assumption of temporal continuity plays a critical role in identifying outliers. Temporal continuity refers to the fact that the patterns in the data are not expected to change abruptly, unless there are abnormal processes at work. It is worth noting that outlier analysis has diverse formulations in the context of temporal data, in some of which temporal continuity is more important than others. In time-series data, temporal continuity is immediate, and expected to be very strong. In multidimensional data with a temporal component (e.g., text streams), temporal continuity is much weaker, and is present only from the perspective of aggregate trends.
Charu C. Aggarwal
Chapter 10. Outlier Detection in Discrete Sequences
Abstract
The previous chapter discusses anomaly detection from the perspective of continuous time series. A related setting is one in which the individual elements at each time stamp are discrete-valued (i.e., categorical). Such discrete time-series are also referred to as sequences. Discrete-valued temporal scenarios arise in numerous systems diagnosis, intrusion detection, and biological applications. In some domains such as intrusion detection and systems diagnosis, the discrete sequences are caused by temporal ordering, whereas in other domains such as biological data, the discrete sequences are caused by physical ordering. Nevertheless, at a logical level, the differences in the problem definitions for the two cases are relatively minor. The primary difference is that temporal data often has a specific direction to the analysis in real scenarios (i.e., forward in time), whereas this may not be the case for data based on placement relationships. At the analytical level, the models for the two cases are different in minimal ways and typically have cross-applicability.
Charu C. Aggarwal
Chapter 11. Spatial Outlier Detection
Abstract
Spatial data shares a number of similarities with time-series data in being a contextual data type. In fact, it is often possible for the spatial and temporal attributes to occur in various combinations of behavioral and contextual attributes. Such data is also referred to as spatiotemporal data. For example, in some applications, such as hurricane tracking, the contextual attributes are both spatial and temporal.
Charu C. Aggarwal
Chapter 12. Outlier Detection in Graphs and Networks
Abstract
Graphs represent one of the most powerful and general forms of data representation. These structures are used to express diverse data, ranging from multidimensional entity-relation graphs, the Web, social networks, communication networks, and biological and chemical compounds.
Charu C. Aggarwal
Chapter 13. Applications of Outlier Analysis
Abstract
Outlier analysis has numerous applications in a wide variety of domains, such as the financial industry, quality control, fault diagnosis, intrusion detection, Web analytics, and medical diagnosis. The applications of outlier analysis are so diverse that it is impossible to exhaustively cover all possibilities in a single chapter. Therefore, the goal of this chapter is to cover many problem domains at a higher level and show how they map to the various techniques discussed in earlier chapters. The practical issues and challenges in the context of real data sets will also be discussed. This will provide a broader understanding of the issues involved in problem domain to technique mapping.
Charu C. Aggarwal
Backmatter
Metadaten
Titel
Outlier Analysis
verfasst von
Charu C. Aggarwal
Copyright-Jahr
2017
Electronic ISBN
978-3-319-47578-3
Print ISBN
978-3-319-47577-6
DOI
https://doi.org/10.1007/978-3-319-47578-3