Fraud-Detection-of-NY-Property-Data

Applying fraud analysis in the NY Property Tax dataset with 1048575 observations and 30 variables

====

File name: NY_property_data Source

Url: https://data.cityofnewyork.us/Housing-Development/Property-Valuation-and-Assessment-Data/rgy2-tti8

Data Provided by: Department of Finance (DOF)

Dataset Owner: NYC OpenData

Category: Housing & Development

Date Created: September 2, 2011 Metadata Last Updated: September 5, 2014

Data volume: 1048575 Fields: 30 fields, 16 continuous, 12 categorical, 2 text

Details variables: RECORD, BBLE, BLOCK, LOT, EASEMENT, OWNER, BLDGCL, TAXCLASS, LTFRONT, LTDEPTH, STORIES, FULLVAL, AVLAND, AVTOT, EXLAND, EXTOT, EXCD1, STADDR, ZIP, EXMPTCL, BLDFRONT, BLDDEPTH, AVLAND2, EXLAND2, EXLAND2, EXTOT2, EXCD2, PERIOD, YEAR, VALTYPE.

Purpose of the project

• Build metrics and detect the potential fraud records

Outline of Approach

Step1: Data Cleaning

• Data Quality Report (DQR)
• Create 50 more insightful variables based on original variables
• Partition based on 7 key metrics
• Dealing with missing value, Z-scale

Step2: PCA (Principle Componenet Analysis)

• PCA (Visualization, Decide to use 13 PCs)

Step3: Set Fraud Score:

1. Euclidean Distance

• Based on projection of original features on the 13 PCs' directions
• Calculate the euclidean distance to the origin as the fraud score

2. Auto-encoder(Neural Network Algorithm)

• Use h2o package to implement auto-encoder
• Output results from the PCA serves as the input
• Set two hidden layers, both five features
• Calculate MSE of each record as the fraud score

3. Comparison

• 76% overlapping in the highest 10000 fraud scores between Euclidean Distance and Auto-encoder          

4. Discover insights

• Check the 30 records with top fraud scores and identify fraud patterns within them.

==== Keep updating...