This repository is a simple demo of using Personalized PageRank to identify provider features that may be useful for anomaly detection in healthcare payment data.

Authors:

• Ofer Mendelevitch, Hortonworks
• Jiwon Seo, Stanford University

Identifying fraud, waste and abuse is a critical application for any insurance company, and a big focus for healthcare insureres around the world. Various approaches have been applied to address this problem, including rules-based and machine-learning-based solutions.

Graph algorithms have potential to improve the accuracy of such systems, and are thus of high interest. In this demo we show an approach to identifying anomalies in a real-world healthcare payment dataset, the Medicare-B data set, using a variant of the personalized PageRank algorithm.

We use Apache Pig and SociaLite (open-source graph analysis platform) to preprocess the data and analyze the data. The details of the algorithm is described in Hortonworks blog (TBD link)

To try this demo on the Hortonworks Sandbox, follow these steps:

• Download the HDP Sandbox and install it on your machine, following the instructions here: http://hortonworks.com/products/hortonworks-sandbox/. We recommend at least 8GB to be allocated to the Sandbox. Whether you use VMWare, VirtualBox or Hyper-V is your choice, the demo should work equally well on each of these products.
• Login to the new sandbox as root
• Install ant: "yum install -y ant"
• Install some development package: "yum install -y python-devel gcc-c++"
• Install the python pandas library: "pip install pandas"
• login as user guest: "su - guest"
• Clone this github repository: "git clone https://github.com/ofermend/medicare-demo.git"
• Change folder to the medicare-demo folder: "cd medicare-demo"
• run setup script to download datasets: "source setup.sh". This downloads various datasets used in the demo and places them into the "ref_data" subfolder
• Configure SociaLite. SociaLite source code is already included in this repository, under the "socialite" folder. You will need to configure and compile SociaLite before using it, using the following steps:
  • cd socialite
  • Edit conf/socialite-env.sh and ensure settings are correct (or change if needed):
    • JAVA_HOME should point to your Java installation path (e.g. /usr/lib/jvm/jdk1.7.0_51/), or be commented out if this environment variable is already set.
    • HADOOP_HOME should point to your Hadoop installation path (e.g., /usr/hdp/current/hadoop-client), or be commented out if this environment variable is already set.
    • SOCIALITE_HEAPSIZE should be set to 6000 (6GB). In the repo it is set to 10000 (10GB) to enable processing the full graph, but with the sandbox we will work with a smaller sub-graph so that it fits in the more limited memory space.
  • Compile socialite using this command: "ant"

If you have a Hadoop cluster you can use, follow these steps:

• Login to one of the edge nodes on the cluster.
• Ensure you have python, pandas and ant installed on your edge node. You might need to ask your administrator to help you with those if they are not already installed.
• Clone this github repository: "git clone https://github.com/ofermend/medicare-demo.git"
• Change folder to the medicare-demo folder: "cd medicare-demo"
• run setup script to download datasets: "source setup.sh". This downloads various datasets used in the demo and places them into the "ref_data" subfolder
• Configure SociaLite. SociaLite source code is already included in this repository, under the "socialite" folder. You will need to configure and compile SociaLite before using it, using the following steps:
  • cd socialite
  • Edit conf/socialite-env.sh and ensure settings are correct (or change if needed):
    • JAVA_HOME should point to your Java installation path (e.g. /usr/lib/jvm/jdk1.7.0_51/), or be commented out if this environment variable is already set.
    • HADOOP_HOME should point to your Hadoop installation path (e.g., /usr/hdp/current/hadoop-client), or be commented out if this environment variable is already set.
    • SOCIALITE_HEAPSIZE should be set to 10000 (10GB).
  • Compile socialite using this command: "ant"

To run the demo code, follow these steps:

• run1.sh: perform pre-processing using Apache PIG scripts to clean the data and generate the provider similarity graph with 10% of the providers.

  • If running on the Sandbox, you should use "run1.sh 10". This executes the script but only on 10% of the providers. This limits the dataset to a smaller size, making it reasonable for a small sandbox environment

• "run2.sh": runs the personalized PageRank algorithm to find anomalies in the data set. The find-anomalies.py implements the algorithm. The output is printed on the screen and also stored in the results-100.txt so it's easier to view it later.

  • If running in the Sandbox, you should use "run2.sh 10" to target the script at the 10% reduced-size dataset. The output file will be named results-10.txt