Neural networks allow the identification of the most important variables related to churn and [are] doing a very good job in pattern detection, writes Marc Valdivia at NPAW (Nice People At Work).
Artificial intelligence (AI) has been ever present in the technical departments of leading companies since it became a sensation in the ‘90s. It’s important to know that the function of AI that offers the best results is known as machine learning. Machine learning is the ability of computer systems to constantly learn by themselves. The only two things missing are data and computational capacity.
The importance of data and the power of calculation are fundamental to understanding the success of AI in the previous twenty years. This allows us not only to obtain information from users who make use of online platforms, but also to know the nature of their experience, the problems they experience and the navigation they perform on their devices, in order to build 360-degree, engaging end-user experiences.
Intelligent anomaly detection
Data allows us to find common patterns that trigger an action – such as user churn – or in the detection of anomalies, which occur when a particular behaviour is not within the expected pattern.
Precisely segmented content recommendation
Moreover, recommendation systems can be built to detect correlations between users (user-based recommendations) or use a person’s own previous experience to offer content (content-based recommendations).
Churn detection and prediction
The detection and prevention of user churn is the number one concern of video services, as it directly affects their income. For this, it is interesting to identify the main causes or influencers for churn, such as poor streaming quality, inefficient customer service, an unappealing content catalogue, unrelated content recommendations, advertising overload and unfriendly UI. It is also possible to predict and identify users at risk of unsubscribing from a video platform by analysing their behaviour and frequency of visits and then comparing it with the previous behaviours of users who left the platform.
Neural networks solve these types of problems and they do a very good job in pattern detection. For an isolated set of data, for example a single visualisation, one would advise using a simple neural network. For a sequence of events, one would opt for an RNN (recurrent neural network) which has a temporary memory of past events, such as a user’s visiting pattern.
Intelligent alert system
Anomaly detection services allow content and technical teams to detect when any of the quality metrics are not in their confidence values and which combination of dimensions is causing such failure. For example, if a particular combination of city and CDN gets a local buffer value that is too high, the confidence values for each metric and dimension cannot be fixed – they are variable in time and the system must adapt (note that a constant problem is not an anomaly). In addition, each possible combination must be treated independently. So-called autoencoders are a type of neural network proving to be very useful in solving these problems, despite existing for a different purpose. They consist of two parts: the first stage of coding and the second stage of decoding. When encoding and decoding ‘normal’ values, the result is the same, but when an anomaly is introduced, the decoding varies with respect to the initial input. There we have an alert.
The market trend is towards incorporating artificial intelligence in more areas of the business, creating automatic marketing campaigns customised for each user, or even models that regulate traffic on platforms so that errors and waiting times are reduced. Exciting new AI applications are to follow in the video sphere, and anyone with a video service will experience them.