Python libs used sys, time, copy
Must run in this order:
- import_batch_input(...)
- feature_1(...)
- feature_2(...)
- feature_3(...)
As long as you just run the antifraud.py file, there shouldn't be any issues. The reason for running one after the other is because they are dependent on each other for optimization.
batch file is uploaded into dictionary graph with keys as id and list of user ids as value. For example: graph_batch = {id1: [id2,id3], id2: [id1], id3: [id1]}
feature 1 uses simple look up in dictionary graph
feature 2 takes two lists from each id key and takes intersection. If they intersect then it is trusted otherwise unverified. Note: feature 2 re-uses the output_list of trusted/unverified from feature 1.
feature 3 uses breach first bidirectional search to the 3rd degree on each side. Note: feature 3 re-uses the output_list of trusted/unverified from feature 2.
Here are my run times:
loading batch_payment.txt takes 25 seconds feature 1 takes 35 seconds feature 2 takes 38 seconds feature 3 takes 460 seconds
A couple of ideas for other fraud detection, but didn't implement:
- Check if payment exceeds certain amount. For example: people are more likely to transfer small amounts like for coffee, lunch, and movie tickets, etc. So, if any any amount exceeds above $80 could be labaled unverified.
- Detect unusual frequency in money transfers within specific time period. This would need a user object to store the last transaction timestamps. For exmaple: if there were 10 transactions within 4 hour period then it could be labeled unverified as people are most likely to transfer money once a day or a few times a week.
Zorigt