Data Mining Concepts and Instances

Input Representations

• Input data for DM methods comes in various forms, including concepts, instances, and attributes.
• A concept description is what's being learned.
• Defining a concept precisely is challenging, so a description that's understandable, debatable, and usable as a guide is the aim.

Instances

• Input data is a collection of instances.
• Instances are individual, independent examples of the concept to be learned.
• Sometimes, raw data can't be neatly separated into individual instances.
• Background/contextual knowledge might be part of the input.
• Raw data might be a unified mass and not easily fragmented.
• A single, continuous sequence of data (e.g., time series) might not be easily subdivided.
• Instances are characterized by attribute values that measure different aspects.
• Many schemes use numeric and nominal attributes (categories).
• Input data can be a text file.

Styles of Learning

• Classification: Learning a way to categorize (classify) new, unseen examples. Most examples in Chapter 1 are classification problems (e.g., weather, contact lens). Multilabelled instances can be tricky.
• Association: Identifying and examining structures within data for any "interesting" relations among features, instead of focusing on predicting a specific class.
• Clustering: Identifying naturally grouped items.
• Numeric prediction: Predicting a numerical outcome rather than a category.

ARFF Format

• ARFF is a file format used to store data examples which can include numeric, nominal, string, or date type attributes.
• The ARFF file is structured with distinct sections for specifying the relation name, attributes, and the data instances.

Sparse Data

• Examples of sparse data include market basket data and document word count data.
• In sparse data, unspecified values are represented as zeroes, not as missing values.

Normalization

• Data normalization can be carried out by the DM algorithm itself.
• Normalization involves fixed ranges (e.g. 0-1) by dividing by the maximum value, or dividing the range.
• Another method subtracts the mean and divides by the standard deviation.
• "Distance" between attribute values is often crucial for many DM algorithms.

Measuring Distance for Nominal Attributes

• To measure "distance" between nominal attributes, a value of 1 is assigned if the attributes are the same, and 0 if they are different.
• Several synthetic binary attributes are created for each nominal attribute.
• A genuine mapping between nominal values and numeric scales might exist (e.g., zip codes to latitude and longitude).
• Nominal values can sometimes be expressed as integers (e.g., employee codes).

Ordinal vs. Nominal Data

• Sometimes data can be represented either as ordinal or nominal (e.g., age group).
• This representation choice depends on context (e.g., in the contact lens dataset)

Missing Values

• Missing values are often indicated by out-of-range entries, blank entries or dashes in nominal attribute.
• A variety of missing values distinctions (unknown, unrecorded, irrelevant) might be coded using different numbers e.g., as '-1', '-2', '-3'.
• The reasoning behind missing values can be important and might need to be considered.
• Missing information itself could be relevant to the analysis (e.g. an earlier test determined that the value wasn't applicable).

Inaccurate Values

• It is important to determine if the DM file has invalid data or values.
• Data for analysis is often not purposely gathered and might contain defects like unrecorded or blank values.
• Data errors and omissions are often important details.
• Errors in data entry and typing errors are common and need attention.
• Duplicate information might exist and need to be removed or recognized.
• The quality and freshness of the data might also be a factor to consider.