FDA - Data Mining Methods, week 4

Data mining categories - Supervised method

Uses labeled data to predict the output for new, unseen data

→ Example: data= customer financial info, whether they defaulted on a loan

task= predict if new applicant is likely to default

Data mining categories - Unsupervised method

Uses unlabeled data to find hidden patterns, groupings or structures.

→ Example: data= shopping transactions

task= identify patterns like: ‘people who buy bread often buy butter’

Data mining categories - Global method

Creates information applicable to all data.

Data mining categories - Local method

Creates information only applicable to some of the data.

Data mining method - Linear regression, goal

• Supervised and global

• Goal= to create a linear model to relate an output variable to one or more input variables

Data mining method - Linear regression, SSD

SSD= Sum of Squared Deviations

• The lower the SSD of a model the better the model.

→ SSD expresses variation relative to the model. Small SSD shows you that the prediction is close to the real data

• Always look at the outcome of SSD in relation to the y-value to see if it is large or small

→ small: SSD= 0,8 and y-value= around 10

→ large: SSD= 3,6 and y-value= ranges 0-5

Data mining method - Linear regression, R²

Explains how well the regression line fits the data.

→ - R²= a slope down from left to right

→ + R²= a slope going up from left to right

The closer the R² to 0, the regression line has little to no predictive power.

→ -1 = perfect predictive power

→ -0.2 = little predictive power

→ 1 = perfect predictive power

→ 0.3 = little predictive power

Data mining method - Linear regression, residual plot

Used to check the linear regression plot.

→ Good plot= residual plot centered around 0, no pattern

→ Bad plot= residual plot with patterns, clusters and extreme plots(outliers)

Data mining method - Clustering, goal

• Unsupervised and global

• Goal= partition all locations in geographically meaningful groups.

Data mining method - Clustering, k-means

This is clustering by algorithms.

1. Pick random centroids of clusters.

2. Assign points to nearest centroid.

3. Recompute centroids until clusters are stable

→ The smaller the distance between centroids the better

Data mining method - Decision tree, goal

• Supervised and global

• Goal= Learn a tree to separate ‘quit’ and ‘non-quit’ cases.

Data mining method - Decision tree, structure

Branches= the lines in between the ovals and cubes

Nodes= the ovals, the ‘topics’

Tree’s leaves= the cubes, the eventual outcome

→ Branches connect the nodes and the tree’s leaves

Data mining method - Decision tree, confusion matrix

Used to check the quality of the decision tree.

Data mining method - Association rules, goal

• Unsupervised and local

• Goal= to find high-confidence associations between frequently occurring (exam outcomes)