Neural networks and the customer experience

Neural networks and the customer experience Andrew Blackman, Senior Product Manager at Cerillion, looks at the evolution of neural networks and their application within the communications industry.

When was the last time that your Communications Services Provider (CSP) contacted you to tell you there’d been a problem with your service and to offer you something as compensation? In my experience, this has never happened for any of my communications services. Only when I contact my CSP with a problem do they seem to take any notice of my customer experience and react to try and keep me as a customer.

Contrast this with say Amazon, where recently I pre-ordered a book for a certain price. When the book came out and was delivered, I then received proactive notification that the item had come down in price by £0.04 since my pre-order and I was refunded the difference. This must’ve cost Amazon more to process than the monetary saving to me, however needless to say it was a very positive customer experience.

Of course managing communications networks and service delivery is a much more complex business than the simple price change / refund scenario above, but the underlying principle is just the same – putting your customer first. And you would think that with the technologies available today, more CSPs would have the ability to monitor the actual customer experience, predict churn and take proactive steps to keep their customers and address the root causes of problems.

Having spent a large amount of time performing revenue assurance and retention functions in a previous company, I found the process not only to be mind numbingly boring but also very labour intensive. This is due to the fact that the amount of data that needs to be collated from multiple sources to identify a churn candidate can be vast and also very time consuming to collect and process. So quite often the time required to gain the intelligence may actually be longer than the time available to react and save the subscriber. Not only does this traditional approach involve a number of manual steps, but it is also limited by the skills and working hours of the staff analysing the data.

Churn management systems based on neural networks have attempted to solve this problem, with some solutions being around for over a decade now. However, these systems have achieved only limited success, partly due to timing – whilst markets are still in a high subscriber growth phase, the emphasis is on customer acquisition rather than retention; but also due to the nature of their implementation with most systems still operating ‘offline’ taking periodic feeds of data for processing in batch.

Artificial intelligence and neural networks would appear to most as being a relatively new and modern technology or one which is purely the stuff of science fiction. However, neural networks have actually been around in computing since 1950, with one of the first being developed by Frank Rosenblatt at Cornell Aeronautical Laboratories in 1957 called ‘Perceptron’.

The basic principle of a neuron within a biological entity is simple. At a given point, be it through some physical stimulation, such as the change between night and day, the neuron will ‘fire’. So logically a neuron firing can represent an event occurring, as previously the transition from darkness to daylight. A single neuron on its own doesn’t do a great deal. However, the connections between many neurons – the neural network – together allow us to create infinitely more complex decision making logic.

The other principle of a neural network is the learning capabilities of the network itself. Network learning is the process of supplying either inputs or outputs to the network to evaluate and derive the patterns which result in the output we are looking for. The learning process works through all connotations of the network and on each cycle removes the patterns which do not result in the desired output. During this process the network may adjust the threshold at which the neuron fires from that of its initial setting. This adjustment of the threshold is fundamentally learning from past patterns.

The first model of a neuron, the ‘MP neuron’, was limited in that it didn’t have learning capabilities. However through further development the ‘Perceptron’ overcame these initial limitations by implementing a learning network which had adjustable weighting values on the connections between neurons.

The neural networks of today have advanced greatly and resolved many of the early limitations and shortcomings. They are in use in our daily lives probably without our knowledge. A good example of this is facial recognition technology in the latest digital cameras. So neural networks sit firmly in software today solving problems which in the past could only be solved by humans. The technology has recently taken another giant leap forward with IBM announcing the development of the ‘Brain Chip’ microprocessor, which will be the closest we have ever been to having a computer that can mimic the functions of a human brain.

In the communications industry, the behaviour of a subscriber over its lifecycle supplies us with all the necessary data that we need to evaluate in a neural network. As the pattern of behaviour changes, it is these changes that we need to monitor to identify any subscriber at risk of churn. For example, a subscriber’s buying pattern is a critical variable that tells us a lot about their current mood. The reason being that as a subscriber’s brand loyalty reduces their buying frequency would decrease also. Add to this another critical variable regarding service quality, which can be tracked by direct network measurement and the number of faults a subscriber has raised, and the two together can be used as part of the churn prediction.

In context of a neural network, each data item we need to monitor would be assigned a specific neuron. These neurons would be allocated an initial threshold at which to fire. Once we have defined the neural network we then train it based on historical data and understood patterns. Once the network has been trained we then apply the network to the live customer base with the objective of identifying churn candidates and continuously learning new patterns of behaviour.

Neural networks are a thing of both our technology past and our future. Their key application is in the area of pattern matching, which is not limited to just churn management, but can be applied across many areas of a CSP’s business including fraud detection, fault monitoring and marketing campaign management. The market shift towards processing everything ‘online’ and the industry focus on improving the customer experience means that the time is now for neural networks to come into their own in the communications sector and deliver real value. And with the development of IBM’s Brain Chip, we are on the cusp of a computing revolution with the potential for ‘artificial intelligence’ to become a core element of the business and operational support systems of the future.