Question

Answer/discuss briefly the following questions.

(a) What is the motivation for using standardized lift in association rule learning?

(b) Using your own words, describe the appropriate process for evaluating and comparing multiple classifiers.

(c) Discuss and compare advantages and disadvantages of classification trees, random forests and boosting.

Answer 1

Answer

a) The lift value in an association rule idenotes the measure of importance of a rule. It is a major performance indicator of a target model to predict or classify data poinmts against a random choice target model.

Standardising lift is a technique which takes into account the range of values lift may take to increase the effectiveness of the association rule. Using standardised lift to rank association rules has the effect of ranking a rule depending on the relative position of its lift to the maximum and minimum potential values. This results in a natural and absolute method of ranking association rules.

b) Processes to evaluate multiple classifiers are as mentioned below: Confusion Matrix, ROC curve, Area under the curve (AUC), Concordance, discordance etc

The best and most relevant process which is used in Confusion matrix. Below is the basic layout of the same

		Truth
		+	-
Predicted	+	True Positives (TP)	False Positives (FP)
-	False Negatives (FN)	True Negatives (TN)

Many parameters can be calculated using confusion matrix:

1) Sensitivity / True Positive Rate

2) Specificity / False Positive Rate

3) Precision / Positive predictive value

4) Negative Predictive value

5) Many others like False discovery rate, False ommision rate etc

c) Below is the answer

	Advantages	Disadvantages
Classification Tree	It doesn’t require normalization and scaling of data	Sometimes the calculation in Classification and decision trees is complex as compared to other algorithms
Requires less effort at Data preparation stage	Due to high complexity, it may be expensive and time consuming
Self explanatory and easily explanable to technical teams and business leads	Not a good technique for regression and predicting continuous output
Random Forest	The prediction performance is one of the best in all the supervised learning techniques	Randon forest model is inherently less interpretable than an classification tree
Can be used for both classifictaion and regression problems	High computational costs
Robust to outliers	Training a large number of trees, may result in slow prediction
Boosting	Curbs Over-fitting of data	Sensitive to outliers
Easy to interpret algorithm	Time and computationally expensive