Discretization in Data Mining
Data discretization is the process of converting a large number of
continuous data values into a smaller number of intervals or groups. This
makes the data easier to analyze, understand, and manage.
In simple terms, discretization transforms continuous numerical data into a
finite set of ranges or categories while trying to keep the loss of
information as small as possible.
There are two main types of discretization:
1. Supervised Discretization
In supervised discretization, the class label (target variable) is used
while dividing the data into intervals.
This means the discretization process considers how the data is related to
the class values.
2. Unsupervised Discretization
In unsupervised discretization, class labels are not used.
The method only depends on the distribution of the data itself.
Two common strategies used here are:
- Top-down splitting – starting with a large interval and dividing it into smaller intervals.
- Bottom-up merging – starting with small intervals and combining them into larger ones.
Techniques of Data Discretization
Several techniques are used in data mining to perform discretization.
1. Histogram Analysis
A histogram is a graphical representation that shows the frequency
distribution of continuousdata.
It helps analysts understand the structure of the data, such as:
- Outliers
- Skewness
- Normal distribution
By observing the histogram, suitable intervals for discretization can be
created.
2. Binning
Binning is a data smoothing technique that groups a large number of
continuous values intosmaller intervals called bins.
This method helps:
- Reduce noise in data
- Simplify data representation
- Create concept hierarchies
Example:
Marks from 0–100 can be grouped into bins like:
0–40 → Low
41–70 → Medium
71–100 → High
3. Cluster Analysis
Cluster analysis is another technique used for discretization.
In this method:
- A clustering algorithm divides data into groups (clusters).
- Each cluster contains values that are similar to each other.
- These clusters can then be used as intervals for discretization.
4. Discretization using Decision Tree Analysis
Decision trees can also be used for discretization.
This method works using a top-down splitting approach and is usually
supervised.
Steps involved:
- Select the attribute with the lowest entropy.
- Apply a recursive process.
- Divide the attribute values into different intervals.
- Continue splitting until meaningful ranges are formed.
5. Discretization using Correlation Analysis
In this method, linear regression is used to find the best neighboring
intervals.
Steps include:
- Identify small intervals first.
- Merge them into larger overlapping intervals.
- Finally, create a set of optimal intervals.
- This is also a supervised method.
- Data Discretization and Concept Hierarchy Generation
A concept hierarchy represents data in a structured order from specific to
general concepts.
It organizes data into levels of abstraction.
For example:
City → Country → Continent
Example:
New Delhi → India → Asia
Concept hierarchies help in data summarization and data mining tasks.
Types of Hierarchy
1. Top-Down Mapping
Top-down mapping starts with general information and moves toward more
specificdetails.
Example:
Continent → Country → City
2. Bottom-Up Mapping
Bottom-up mapping starts with specific data and gradually moves toward more
generalinformation.
Example:
City → Country → Continent
Data Discretization and Binarization
Data Discretization converts continuous data into intervals or categories.
Data Binarization converts attributes into binary values (0 or 1).
Example of binarization
Temperature:
Above 30°C → 1 (Hot)
Below 30°C → 0 (Not Hot)
Importance of Discretization
Discretization is important in data mining because:
- It simplifies complex continuous data.
- It reduces noise in the dataset.
- It improves the signal-to-noise ratio.
- It makes many machine learning algorithms work more efficiently.
- It helps in better data interpretation and visualization.
