|By Roger Barga, Avinash Joshi, Pravin Venugopal||
|June 27, 2014 10:15 AM EDT||
This article explores how to detect fraud among online banking customers in real-time by running an ensemble of statistical and machine learning algorithms on a dataset of customer transactions and demographic data. The algorithms, namely Logistic Regression, Self-Organizing Maps and Support Vector Machines, are operationalized using a multi-agent framework for real-time data analysis. This article also explores the cloud environment for real-time analytics by deploying the agent framework in a cloud environment that meets computational demands by letting users' provision virtual machines within managed data centers, freeing them from the worry of acquiring and setting up new hardware and networks.
Real-time decision making is becoming increasingly valuable with the advancement of data collection and analytics techniques. Due to the increase in the speed of processing, the classical data warehousing model is moving toward a real-time model. A platform that enables the rapid development and deployment of applications, reducing the lag between data acquisition and actionable insight has become of paramount importance in the corporate world. Such a system can be used for the classic case of deriving information from data collected in the past and also to have a real-time engine that reacts to events as they occur. Some examples of such applications include:
- A product company can get real-time feedback for their new releases using data from social media
- Algorithmic trading by reacting in real times to fluctuations in stock prices
- Real-time recommendations for food and entertainment based on a customer's location
- Traffic signal operations based on real-time information of volume of traffic
- E-commerce websites can detect a customer transaction being authentic or fraudulent in real-time
A cloud-based ecosystem enables users to build an application that detects, in real-time, fraudulent customers based on their demographic information and financial history. Multiple algorithms are utilized to detect fraud and the output is aggregated to improve prediction accuracy.
The dataset used to demonstrate this application comprises of various customer demographic variables and financial information such as age, residential address, office address, income type, income frequency, bankruptcy filing status, etc. The dependent variable (the variable to be predicted) is called "bad", which is a binary variable taking the value 0 (for not fraud) or 1 (for fraud).
Using Cloud for Effective Usage of Resources
A system that allows the development of applications capable of churning out results in real-time has multiple services running in tandem and is highly resource intensive. By deploying the system in the cloud, maintenance and load balancing of the system can be handled efficiently. It will also give the user more time to focus on application development. For the purpose of fraud detection, the active components, for example, include:
- Web services
This approach combines the strengths and synergies of both cloud computing and machine learning technologies, providing a small company or even a startup that is unlikely to have specialized staff and necessary infrastructure for what is a computationally intensive approach, the ability to build a system that make decisions based on historical transactions.
As multiple algorithms are to be run on the same data, a real-time agent paradigm is chosen to run these algorithms. An agent is an autonomous entity that may expect inputs and send outputs after performing a set of instructions. In a real-time system, these agents are wired together with directed connections to form an agency. An agent typically has two behaviors, cyclic and triggered. Cyclic agents, as the name suggests, run continuously in a loop and do not need any input. These are usually the first agents in an agency and are used for streaming data to the agency by connecting to an external real-time data source. A triggered agent runs every time it receives a message from a cyclic agent or another triggered agent. Once it consumes one message, it waits for the next message to arrive.
Figure 1: A simple agency with two agents
In Figure 1, Agent 1 is a cyclic agent while Agent 2 is a triggered agent. Agent 1 finishes its computation and sends a message to Agent 2, which uses the message as an input for further computation.
Feature Selection and Data Treatment
The dataset used for demonstrating fraud detection agency has 250 variables (features) pertaining to the demographic and financial history of the customers. To reduce the number of features, a Random Forest run was conducted on the dataset to obtain variable importance. Next, the top 30 variables were selected based on the variable importance. This reduced dataset was used for running a list of classification algorithms.
Algorithms for Fraud Detection
The fraud detection problem is a binary classification problem for which we have chosen three different algorithms to classify the input data into fraud (1) and not fraud (0). Each algorithm is configured as a triggered agent for our real-time system.
This is a probabilistic classification model where the dependent variable (the variable to be predicted) is a binary variable or a categorical variable. In case of binary dependent variables favorable outcomes are represented as 1 and non-favorable outcomes are represented as 0. Logistic regression models the probability of the dependent variable taking the value 0 or 1.
For the fraud detection problem, the dependent variable "bad" is modelled to give probabilities to each customer of being fraud or not. The equation takes multiple variables as input and returns a value between 0 & 1 which is the probability of "bad" being 0. If this value is greater than 0.7, then that customer is classified as not fraud.
Self-Organizing Maps (SOM)
This is an artificial neural network that uses unsupervised learning to represent the data in lower (typically two dimensions) dimensions. This representation of the input data in lower dimensions is called a map. Like most artificial neural networks, SOMs operate in two modes: training and mapping. "Training" builds the map using input examples, while "mapping" automatically classifies a new input vector.
For the fraud detection problem, the input space which is a fifty dimensional space is mapped to a two dimensional lattice of nodes. The training is done using data from the recent past and the new data is mapped using the trained model, which puts it either in the "fraud" cluster or "not - fraud" cluster.
Figure 2: x is an in-put vector in higher dimension, discretized in 2D using wij as the weight matrix
Image Source: http://www.lohninger.com/helpcsuite/kohonen_network_-_background_information.htm
Support Vector Machines (SVM)
This is a supervised learning technique used generally for classifying data. It needs a training dataset where the data is already classified into the required categories. It creates a hyperplane or set of hyperplanes that can be used for classification. The hyperplane is chosen such that it separates the different classes and the margin between the samples in the training set is widest.
For the fraud detection problem, SVM classifies the data points into two classes. The hyperplane is chosen by training the model over the past data. Using the variable "bad", the clusters are labeled as "0" (fraud) and "1" (not fraud). The new data points are classified using the hyperplane obtained while training.
Figure 3: Of the three hyperplanes which segment the data, H2 is the hyperplane which classifies the data accurately
Image Source: http://en.wikipedia.org/wiki/File:Svm_separating_hyperplanes.png
Fraud Detection Agency
A four-tier agency is created to build a workflow process for fraud detection.
Streamer Agent (Tier 1): This agent streams data in real-time to agents in Tier 2. It is the first agent in the agency and its behavior is cyclic. It connects to a real-time data source, pre-processes the data and sends it to the agents in the next layer.
Algorithm Agents (Tier 2): This tier has multiple agents running an ensemble of algorithms with one agent per algorithm. Each agent receives the message from the streamer agent and uses a pre-trained (trained on historical data) model for scoring.
Collator Agent (Tier 3): This agent receives scores from agents in Tier 2 and generates a single score by aggregating the scores. It then converts the score into an appropriate JSON format and sends it to an UI agent for consumption.
User Interface Agent (Tier 4): This agent pushes the messages it receives to a socket server. Any external socket client can be used to consume these messages.
Figure 4: The Fraud detection agency with agents in each layer. The final agent is mapped to a port to which a socket client can connect
Results and Model Validation
The models were trained on 70% of the data and the remaining 30% of the data was streamed to the above agency simulating a real-time data source.
Under-sample: The ratio of number of 0s to the number of 1s in the original dataset for the variable "bad" is 20:1. This would lead to biasing the models towards 0. To overcome this, we sample the training dataset by under-sampling the number of 0s to maintain the ration at 10:1.
The final output of the agency is the classification of the input as fraudulent or not. Since the value for the variable "bad" is already known for this data, it helps us gauge the accuracy of the aggregated model.
Figure 5: Accuracy for detecting fraud ("bad"=1) for different sampling ratio between no.of 0s and no. of 1s in the training dataset
Fraud detection can be improved by running an ensemble of algorithms in parallel and aggregating the predictions in real-time. This entire end-to-end application was designed and deployed in three working days. This shows the power of a system that enables easy deployment of real-time analytics applications. The work flow becomes inherently parallel as these agents run as separate processes communicating with each other. Deploying this in the cloud makes it horizontally scalable owing to effective load balancing and hardware maintenance. It also provides higher data security and makes the system fault tolerant by making processes mobile. This combination of a real-time application development system and a cloud-based computing enables even non-technical teams to rapidly deploy applications.
- Gravic Inc, "The Evolution of Real-Time Business Intelligence", "http://www.gravic.com/shadowbase/pdf/white-papers/Shadowbase-for-Real-Time-Business-Intelligence.pdf"
- Bernhard Schlkopf, Alexander J. Smola ( 2002), "Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond (Adaptive Computation and Machine Learning)", MIT Press
- Christopher Burges (1998), "A Tutorial on Support Vector Machines for Pattern Recognition", Data Mining and Knowledge Discovery, Kluwer Publishers
- Kohonen, T. (Sep 1990), "The self-organizing map", Proceedings of IEEE
- Samuel Kaski (1997), "Data Exploration Using Self-Organizing Maps", ACTA POLYTECHNICA SCANDINAVICA: MATHEMATICS, COMPUTING AND MANAGEMENT IN ENGINEERING SERIES NO. 82,
- Rokach, L. (2010). "Ensemble based classifiers". Artificial Intelligence Review
- Robin Genuer, Jean-Michel Poggi, Christine Tuleau-Malot, "Variable Selection using Random Forests", http://robin.genuer.fr/genuer-poggi-tuleau.varselect-rf.preprint.pdf