Anti-Discrimination In Decision Making

Abstract- Discrimination is action that denies social participation or human rights to categories of people based on prejudice. It includes unjust or unequal treatment of different groups of people, especially on the grounds of race, religion, age, or sex. Discrimination is one of the negative social perceptions about data mining. Automated decision making is the main aim of the data mining such as classification rule mining etc. Historical training dataset is used for creating decision models. If the training dataset is modified unfairly on the discriminatory attributes, discriminated decisions may occur. So the antidiscrimination techniques are introduced with the data mining. In proposed anti-discrimination system, discriminations are discovered

Direct Discrimination Discovery

Let r: D, B -> C be a PND classification rule, and let x = conf(D, B -> C) and A be a discriminatory item

Elb(x,y)= f(x)/y, f(x)>0 0,f(x)< α equation is used to convert each α-discriminative rule into α-protectiverule. The following inequality is applied for each α-discriminative rule r : A, B->C in α-discriminatory rules collection. Where A is a discriminative item set.

conf(r: A, B -> C) < α . conf(⌐A, B -> C)

To satisfy above inequality, confidence of α- discriminative rule (A, B->C) has to be decreased to a valueless than confidence of rule ⌐A, B->C, and also the confidence of ⌐A, B->C rule should not be changed ordecreased. To do that, transform the records ⌐A to A in thesubset of records which support the rule ⌐A, B->⌐C and have minimum impact on other rules. Similarly we can do,

Method 1: ⌐A, B->⌐C to A, B->⌐C

Method 2: A, B->C to A, B->⌐C

Method 3: ⌐A,B->⌐C to ⌐A, B->C

Algorithm 1: Rule Protection (Method 1)

Input : Original dataset, Freq Rule, PD rule, DIs, α

Output : Transformed dataset foreachpdrule in PD rules FreqRule = FreqRule – pdrule DSc = select all the records from original Dataset which support ⌐A, B⌐C foreach record in DSc

The below conditions are checked for rule generalization: Cond 1: conf(A,B->C) >= p . conf( D,B->C)

Cond 2: conf(r′′ : A,B-> D) >= p.

There are three cases arise based on the above twoconditions,

If both conditions are satisfied for atleast one non redlining rule then there is no need of transformation. If condition1 is not satisfied but condition2 is satisfied by atleast a rule, then minimum data transformationrequired to fulfill the condition 1. If no rule satisfies condition2, rule generalization method is not possible direct rule protection method required to transform the dataset.

To enforce the condition1, the above inequality incondition1 is rewritten as, conf(r′ : A, B->C) C to the value less than the righthand side of above equation and confidence of rule D, B->C should not modified. A possible solution to decrease thisconfidence is to transform the class item from C to ⌐C in therecords which satisfy the rule A, B,⌐D -> C and have minimum impact on other rules.

Algorithm 2 : Rule Generalization

Input: DB(dataset), FreqRule, p >=0:8, α,