NAGRA Insight at the Applied Machine Learning Days 2020
By Pietro Berkes, Principal Data Scientist at NAGRA Insight
The Insight team recently attended the EPFL’s Applied Machine Learning Days 2020, an interesting event placed at the intersection of Machine Learning (ML) and business.
It was a great opportunity to find out more about today’s hot topics in the ML sphere, such as the issue of data privacy, the impact of Artificial Intelligence (AI) in the medical field, and how ML can help improve consumer behavior prediction models.
Different views on the future of privacy
Among the hot topics was the discussion of the past, present and future of privacy with three exceptional experts in the field. Michal Kosinski from Stanford, famous for showing that psychological traits can be predicted by a handful of Facebook likes, Max Tegmark from MIT, a renowned futurist, and Edward Snowden, who certainly does not need any introduction.
The picture of the future they gave is one where privacy is going to disappear due to the Internet and artificial intelligence (AI), although their opinions varied considerably regarding the consequences of that. Kosinski and Tegmark stressed the opportunities of AI and pointed out that humanity has never been more connected, informed, and generally well-off than today, while Snowden on the other hand had a more pessimistic look on the future. Although opinions might differ, all experts agree: there is a need to define privacy rules adapted to the new data-driven world we live in.
AI to disrupt the medical field?
AI applications in the medical field were also much discussed this year: as drug discovery and development is extremely long and costly, companies are heavily investing in ML to make this process more efficient and lower the costs. For instance, ML can help discover the relation between compounds, proteins and diseases, which would open up the possibility to predict new targets for existing drugs, or to predict side-effects for drug combinations.
Novartis and Microsoft jointly announced at AMLD a new collaboration in the form of an innovation lab to transform the pharma business through AI. This follows some recent announcements of major investments in this sector by some of the most AI-savvy companies such as Microsoft and Google Health, and I am curious to see whether AI will deliver on expectations.
Improving the prediction model performance in the telecom business
Another topic which is being shared at the moment in the ML community and which is of particular interest for the industry is the issue of creating consumer behavior prediction models with higher performance levels.
This topic particularly interests us at NAGRA Insight as we create machine learning workflows for our customers in the TV and telecom industry to help them make impactful business decisions. Let’s take a closer look.
How to deal with annual offer expiration
Subscription-based businesses are increasingly dependent on Machine Learning to create personalized offers for upselling new products, maximizing subscription revenue and minimizing the loss of subscribers. However, they often use black-box models that fail to capture fine-grained, low-probability behaviors that have strong business impact.
To address this issue, NAGRA Insight has a unique approach to predicting the impact of business actions on telecom subscribers.
I presented this case at AMLD using the example of the price evolution in the telecom industry, where new subscribers are often acquired with a promotional, low-price offer that matures into the final, full-price offer over 1 to 4 years.
Taking the path from acquisition offer to full price offer right is a tricky balance between meeting revenue objectives and not alienating (and eventually losing) customers.
One of the crucial aspects is to take the business process into account in the modelling in order to improve model performance and interpretability. In the case that I presented, the business model involved has the granularity of a single subscriber and takes into account the various possible actions of the subscriber when the price increase occurs.
Modelling the “uplift”
The second important requirement is to model the differential effect of actions, called the “uplift”. Most of the predictions that are done in the industry right now are not useful for taking actions. In order to be relevant, it is necessary to be able to model the causal relation between actions. In other words, one needs to be able to predict for each subscriber what difference an action is going to make (the uplift of the action). Revenue lift alone is not sufficient to make a decision.
However, it is not easy to model personalized uplift. The combination of ML and uplift modelling is recent and still a matter of active research. We have developed an advanced framework for that in our Insight product and implement it with our customers. The objective is to enable operators to achieve optimized impact at each decision point.
And finally, it is essential to make sure that clients are in control of the business objectives and decisions. The products that we deliver are merely the tools that allow that to happen quickly and at scale.
The topic and the presentation of this business case triggered many questions and a lot of interest from data scientists eager to exchange ideas. We will be happy to continue this conversation so please do not hesitate to get in touch with us!
