G4G0-2/AI & Data Mining/Week 1/Lecture 2 - Input and Output.md

Attributes

• Each example described by fixed pre-defined set of features (attributes)
• Number of attributes may vary
  • ex. Transportation Vehicles
    • no. wheels not applicable to ships
    • no. masts not applicable to cars
  • Possible solution: "irrelevant value" flag
• Attributes may be dependent on other attributes

Taxonomy of Data Types

Nominal Attributes

• Distinct symbols
  • Serve as labels or names
• No relation implied among nominal values
• Only equality tests can be performed
  • ex. outlook = sunny

Sources of Missing Values

• Malfunctioning / Misconfigured Equipment
• Changes in design
• Collation of different datasets
• Data not collected for mining
  • Errors and omissions dont affect purpose of data
  • ex. Banks do not need to know age in banking datasets, DOB may contain missing values
• Missing value may have significance
  • ex. medical diagnoses can be made from tests a doctor decides, rather than the outcome.
  • Most DM algos assume this is not the case, hence "missing" may need to be coded as an additional nominal value.

Inaccurate Values

• Typographical errors in nominal attributes
• Typographical and measurement errors in numeric attributes
• Deliberate errors
  • ex. Incorrect ZIP codes, unsanitised inputs
• Duplicate examples

Weka and ARFF

Weather Dataset in ARFF

Getting to Know the Data

• First task, get to know data
• Simple visualisations useful:
  • Nominal: bar graph
  • Numeric: histograms
• 2D and 3D plots show dependencies
• Need to consult experts
• Too much data? Take sample.

Concept Descriptions

• Output of DM algorithm
• Many ways of representing:
  • Decision Trees
  • Rules
  • Linear Regression Functions

Decision Trees

• Divide-and-Conquer approach
• Trees drawn upside down
  • Node at top is root
• Edges are branches
• Rectangles represent leaves
• Leaves assign classification
• Nodes involve testing attribute

Decision Tree with Nominal Attributes

• Number of branches usually equal to number values
• Attribute not tested more than once.

Decision Tree with Numeric Attributes

• Test whether value is greater or less than constant
• Attribute may be tested multiple times

Decision Trees with Missing Values

• Not clear which branch should be taken when node tests attribute with missing value
• Does absence of a value have significance?
  • Yes => Treat as separate value during training
  • No => Treat in special way during testing
    • Assign sample to most popular branch

Classification Rules

• Popular alternative to decision tree
• Antecedent (pre-condition) - series of tests
• Tests usually logically ANDed together
• Consequent (conclusion) - usually a class
• Individual rules often logically ORed together

If-Then Rules for Contact Lenses

Nuggets

• Are rules independent
• Problem: Ignores process of executing rules
  • Ordered set (decision list)
    • Order important for interpretation
  • Unordered set
    • Rules may overlap and lead to different conclusions for the same example
    • Needs conflict resolution

Executing Rules

• What if \geq 2 rules conflict?
  • Give no conclusion?
  • Go with the rule that covers largest no. training samples?
• What is no rule applies to test example?
  • Give no conclusion?
  • Go with class that is most frequent?

Special Case: Boolean Classes

• Assumption: if example does not belong to class "yes", belongs to "no"
• Solution: only learn rules for class "yes", use default rule for "no"
• Order is important, no conflicts.