1. A Unifying View on Instance Selection
Thomas Reinartz
DaimlerChrysler AG,
Research and Technology, Germany

2. Outlines Introduction Focusing Tasks
Evaluation Criteria for Instance Selection
Unifying Framework for Instance Selection
Evaluation
Conclusions

3. Introduction CRISP-DM (Cross-Industry Standard Process of Data Mining)
Business Understanding
Data Understanding
Data Preparation: data selection, cleaning, construction, integration, formatting
Modeling
Evaluation
Deployment
Data Selection: data shrink or data reduction is needed for huge data size, Focusing

4. Focusing Tasks (1) Data as a table (A,T) Focusing Specification
A : Attribute – characterized by name, type, domain of values
T : Tuple or Instance – a sequence of attribute values
Focusing Specification
Focusing Input : table, a component of a table
Focusing Output : a subset of Focusing Input
Simple subset
Constrained subset
Constructed subset
Focusing Criterion : the relation between input and output

5. Focusing Tasks (2) Focusing Context
Data Mining Goal: classification, prediction, description,
concept description, summarization, dependency analysis
Data Characteristics: simple statistics, information quality
Data Mining Algorithm
Instance Selection: a particular focusing task where input is a set of cases, output is a subset of input

6. Evaluation Criteria for Instance Selection
Different Evaluation Strategies
Filter and Wrapper Evaluation
Filter Approach: only considers data reduction
w.r.t. the mean, variance, distribution, joint distribution
Wrapper Approach: evaluate with specific data mining aspect
e.g. execution time, storage requirements, accuracy, complexity
Isolated and Comparative Evaluation (for Solutions)
Separated and Combined Evaluation (for Criteria)

7. Unifying Framework for Instance Selection
InputSamplingClusteringPrototypingOutput
Evaluations
Sampling : simple random sampling, systematic sampling, stratified sampling
Clustering
Prototyping
Order is not important

8. Evaluation
Generic Sampling (GENSAM) – implement unifying framework by additional preparation steps
Sorting: select by ordering the values of cases for the important attribute
Stratification: separate cases by the attribute intervals (continuous) or attribute values (discrete) in the order of attribute relevance
Intelligent Sampling: random sampling, stratified sampling, systematic sampling, leader sampling, similarity-driven sampling by combining the above methods
Experimental Setting
Goal: Classification,
Algorithm: C4.5, Instance-based learning (NN classifier)
Instance selection methods: simple random sampling (S), simple random sampling with stratification (RS), systematic sampling (S), systematic sampling with sorting (SS), leader sampling (L), leader sampling with sorting and stratification (LS)
Data: training set (80%), test set (20%)

10. Conclusions Evaluation Criteria More Intelligent Focusing Solutions
Analytical, experimental studies for different instant selection techniques