Blog article
See all stories »

Bayesian v Neural Networks

Payment fraud evolves constantly, changing and adapting to the systems, networks and infrastructure. Criminals are always on the lookout for a weak link in the payment cycle to exploit, altering their behaviour to extract maximum value from their fraud activities.

Responding to this ever-changing landscape is the responsibility of everyone involved in the payments industry, from banks and payment service providers (PSPs) to issuers and merchants. Fraud prevention models are the main tool, but not all are the same.

The two main types are Bayesian models and neural networks. These two differ in their approach, but the goal remains the same: use patterns in data to separate legitimate from fraudulent transactions.

Fraud detection is the process of classifying the transactions into classes of legitimate  or fraudulent. Usually the objective of fraud detection is to maximise correct predictions while managing incorrect predictions at an acceptable level of cost. Both Bayesian and neural networks seek to achieve this, but they work in very different ways.

Problems with fraud detection

Detecting and responding to fraud is not easy. Fraud detection systems need to be able to handle highly skewed distributions of data, since only a very small percentage of transactions are fraudulent. Also, noise, errors in data and poorly maintained records need to be filtered out and managed in such a way as to minimise the number of transactions that are assumed legitimate, when actually they are fraudulent.

Critically, these systems must be able to adapt themselves to new kinds of fraud. If reactions aren’t extremely agile, new frauds will slip past old defences and customer experience will suffer.

Neural problems

It is often wrongly assumed that neural networks are a fast, easy and reliable technique to obtain good results in different areas, one study from Vrije Universiteit Brussel noted. In practice, the researchers said, it is found that the great difficulty in applying neural networks resides in the choice of a good set of pre-processing operations and a trade off between the different parameters that have to be chosen.

Meanwhile a research paper published in the World of Computer Science and Information Technology Journal last year highlighted the key issues with neural networks.

The report, Electronic Payment Fraud Detection Techniques, listed a number of disadvantages, which include poor explanation capability, less efficiency in processing large data sets, difficulty setting up and operating, sensitivity to data format, and the fact that different data representations can produce different results. Moreover Neural networks “can never be exact”, according to the study.

However, it did note that they can be effective in dealing with noisy data, in predicting patterns, in solving complex problems, and in processing new instances.

Bayesian advantages

There are a number of key advantages to the Bayesian model. Firstly, it has the ability to adapt in real time, which makes it easy to update the underlying probability distributions as fraud analysts tag transactions as fraudulent or legitimate.

Transparency is also improved, as the reasoning behind each fraud score is made clear to the fraud officer. As Bayesian techniques are based on calculating probabilities of all the different fraud risks, it is clear which of them have contributed to the overall fraud score for a transaction.

Unreliable data is another worry for tackling fraud, as a handful of fraudulent transactions occur within a much larger number of legitimate payments. A University of Salford report highlighted how the number of fraudulent transactions is much less than the total number of transactions, meaning the system will have to handle skewed distributions of the data.

However, this is taken into account by Bayesian networks, as the calibration and training of models and sub-models is based on a data set that contains all legitimate and fraudulent transactions.

The Vrije Universiteit Brussel paper contained some key findings. Firstly, it noted that Bayesian networks could detect up to eight per cent more fraudulent transactions than neural networks. In addition, it noted that learning times for the Bayesian networks were far shorter. Bayesian networks, as the authors of the study proclaim, yield better results concerning fraud detection.

With fraud costing the industry vast sums, it is vital banks, merchants and PSPs look to the very best fraud prevention systems available. As the data shows, this means Bayesian networks.


Comments: (1)

A Finextra member
A Finextra member 23 October, 2013, 05:58Be the first to give this comment the thumbs up 0 likes I would just like to add a quick comment on this excellent blog to say that other techniques do exist which are used in abuse detection. One which I personally have used is that of fuzzy logic. I apologize for sounding like a bbc news reporter, but I thought it was worth pointing out. Bruce

Now hiring