Data Mining: The Textbook

Feature extraction and portability: The raw data is often in a form that is not suit-able for processing. Examples include raw logs, documents, semistructured data, and possibly other forms of heterogeneous data. In such cases, it may be desirable to derive meaningful features from the data. Generally, features with good semantic interpretability are more desirable because they simplify the ability of the analyst to understand intermediate results. Furthermore, they are usually better tied to the goals of the data mining application at hand. In some cases where the data is obtained from multiple sources, it needs to be integrated into a single database for processing. In addition, some algorithms may work only with a specific data type, whereas the data may contain heterogeneous types. In such cases, data type portability becomes

C. C. Aggarwal, Data Mining: The Textbook, DOI 10.1007/978-3-319-14142-8 2

27

c Springer International Publishing Switzerland 2015

28 CHAPTER 2. DATA PREPARATION

important where attributes of one type are transformed to another. This results in a more homogeneous data set that can be processed by existing algorithms.

2. 
   Data cleaning: In the data cleaning phase, missing, erroneous, and inconsistent entries are removed from the data. In addition, some missing entries may also be estimated by a process known as imputation.
   

2. 
   Data reduction, selection, and transformation: In this phase, the size of the data is reduced through data subset selection, feature subset selection, or data transforma-tion. The gains obtained in this phase are twofold. First, when the size of the data is reduced, the algorithms are generally more efficient. Second, if irrelevant features or irrelevant records are removed, the quality of the data mining process is improved. The first goal is achieved by generic sampling and dimensionality reduction techniques. To achieve the second goal, a highly problem-specific approach must be used for feature selection. For example, a feature selection approach that works well for clustering may not work well for classification.
   

Some forms of feature selection are tightly integrated with the problem at hand. Later chapters on specific problems such as clustering and classification will contain detailed discussions on feature selection.

This chapter is organized as follows. The feature extraction phase is discussed in Sect. 2.2. The data cleaning phase is covered in Sect. 2.3. The data reduction phase is explained in Sect. 2.4. A summary is given in Sect. 2.5.

2.2 Feature Extraction and Portability

The first phase of the data mining process is creating a set of features that the analyst can work with. In cases where the data is in raw and unstructured form (e.g., raw text, sensor signals), the relevant features need to be extracted for processing. In other cases where a heterogeneous mixture of features is available in different forms, an “off-the-shelf” analytical approach is often not available to process such data. In such cases, it may be desirable to transform the data into a uniform representation for processing. This is referred to as data type porting.

2.2.1 Feature Extraction

The first phase of feature extraction is a crucial one, though it is very application specific. In some cases, feature extraction is closely related to the concept of data type portability, where low-level features of one type may be transformed to higher-level features of another type. The nature of feature extraction depends on the domain from which the data is drawn:

1.