What is Noise in Data Mining?
In data mining, noise refers to unwanted,
incorrect, or meaningless data that does not represent the real
information.
In simple terms:
- Data = Useful information (signal) + Unwanted errors (noise)
Noisy data can confuse systems and lead to wrong
analysis or incorrect conclusions if not handled properly.
Why is Noisy Data a Problem?
- Reduces accuracy of results
- Increases storage space unnecessarily
- Makes analysis difficult
- Can lead to wrong decisions
Causes of Noisy Data
- Noise can come from many sources, such as:
- Hardware issues (sensor faults, device errors)
- Human errors (wrong data entry, spelling mistakes)
- Software errors (bugs, incorrect processing)
- OCR or speech recognition errors
- Use of slang, abbreviations, or unclear text
Types of Errors
- Measurement errors → caused by devices
- Random errors → caused during data collection or processing
Types of Noise in Data Mining
1. Class Noise (Label Noise)
This happens when data is assigned the wrong category or label.
Examples:
- Misclassification → A positive example labeled as negative
- Contradictory data → Same data but different labels
Causes:
- Human mistakes
- Lack of proper information
- Subjective labeling
2. Attribute Noise
This happens
when the feature values (data fields) are incorrect or missing.
Examples:
- Wrong values (e.g., age = 150)
- Missing values (?)
- Irrelevant values (data that doesn’t matter)
Impact:
Attribute noise often affects models
more than class noise.
Special Case: Outliers
Outliers are data
points that are very different from the rest.
Causes:
- Data entry mistakes
- System errors
- Genuine rare events
Important:
- Removing real outliers can harm results
- Keeping wrong outliers can also distort analysis
Other Source of Noise: Fraud
Sometimes, data is
intentionally changed to show better results.
This can make data look
cleaner but actually reduces reliability.
Types of Random Noise
White
noise → Completely random and unavoidable
Measured using Signal-to-Noise
Ratio (SNR)
Simulating Noise in Data
To test models, artificial
noise is added:
Based on:
- Where noise is added (input or output)
- Distribution (uniform or Gaussian)
- Amount of noise (percentage of affected data)
Common Noise Techniques
Class Noise:
- Uniform noise → Random labels are changed
- Pair-wise noise → Majority class mislabeled as second largest class
Attribute Noise:
- Uniform noise → Random values added
- Gaussian noise → Small variations added using normal distribution
How to Handle Noisy Data?
Removing noise is called data smoothing or data cleaning.
1. Binning
Data is sorted and divided into groups (bins)
Values are
replaced using:
- Mean
- Median
- Boundary values
2.
Regression
Fits a line or equation to the data
Helps smooth out
noise and predict correct values
3. Clustering
Groups similar data
together
Data far from clusters are treated as noise or outliers
4. Outlier Analysis
Detects unusual values:
- Univariate outliers → Single feature
- Multivariate outliers → Multiple features
- Point outliers → Single abnormal value
- Contextual outliers → Depends on situation
- Collective outliers → Group of unusual data
Importance of Data Cleaning
- Removes noise and missing values
- Improves data quality
- Ensures better analysis results
In real-world projects, data
preprocessing can take up to 90% of the total work
Final Note
Good
data leads to good results.
- More noise = Less accuracy
- Clean data = Better decisions
