Question

I am trying to implement Naive bayes algorithm on some real time data.I am aware of the rules of bayes but I am not sure how to implement on my data.My data looks like as below.There are total 2 labels in my data which are ok,fraud and testing data labelled as unkn.I need to classify all the unkn records as either ok or fraud by applying Naive Bayes Algorithm.How do I achieve this? Please some one help me.

1,v1,p1,182,1665,unkn
2,v2,p1,3072,8780,ok
3,v3,p1,20393,76990,ok
4,v4,p1,112,1100,fraud
5,v3,p1,6164,20260,unkn
6,v5,p2,104,1155,ok
7,v6,p2,350,5680,unkn
8,v7,p2,200,4010,ok
9,v8,p2,233,2855,unkn
10,v9,p2,118,1175,unkn

Bayes Rules:-

Posterior Probability of unkn being ok = Prior Probability of ok * Likelihood of unkn given ok.

Posterior Probability of unkn being fraud = Prior Probability of fraud * Likelihood of unkn given fraud.

Answer 1

I am assuming the row 1,v1,p1,182,1665,unkn is interpreted as:

• 1, v1 = some identifiers
• p1,182,1665 = features of your data point
• unkn = label, in this case unknown

With that notation in mind, your training data consists of all lines that have label ok or fraud, and your testing data is the rest. You have to calculate a priors and conditional likelihoods:

1. Prior probability of ok is the proportion of ok examples in the training data. The same applies for fraud
2. For each feature f, such as v1 or p1, its likelihood given ok is the proportion of ok examples in the training data which contain the feature. For instance, p1 is contained in 2 out of 4 ok examples, giving you a probability of 0.5.

For each example multiply together the probabilities you calculated for all of its features in step 2. Multiply the result by the probability in step 1 to obtain the (joint) probability of your example belonging to a particular class.

Caveats:

• Multiplying probabilities together will eventually result in underflow. You might want to add the logs of those probabilities instead.
• The algorithm I described works for discrete-valued features only. The continuous-valued features you appear to have above (e.g. 182) need to be converted to discrete (e.g. by binning) or you need to come up with some other way of estimating the conditional probability in step 2. Google for continuous-valued Naive Bayes