The big challenge in using artificial intelligence and machine learning in small language areas and markets is that there is less learning data available. MittMedia’s Magnus Engström attended Qvik’s Business & Beers night to tell us how his company uses data for personalization and selling subscriptions in Sweden, a market roughly twice the size of Finland. Qvik’s passionate data maturity enthusiast Tuukka Puumala liked MittMedia’s open-minded attitude to experimentation.
“MittMedia collects location data from its readers and compares it with the news they are reading to test whether people are really interested in news happening nearby. This is precisely the right way to go about testing.”
MittMedia has its own data team with both developers and data scientists, which makes development projects and experiments quick and easy to implement. But how could you tap into data-based content testing without an in-house team? What would Qvik do?
“We’d start by experimenting with ways of encouraging people to spend more time on the website. After a one- or two-month test, we would then see whether those who stay longer are more likely to buy something. We would measure, make some changes and see if we couldn’t nudge the indicators in the direction we want. That’s what working with data is about.”
In its data tests, MittMedia noticed things like a direct correlation between subscriptions and the number of stories read. The more stories users read on the website, the more likely they are to subscribe. But then, something strange happens: it’s as if the customer runs out of reading material, and new customers actually read less. This has a negative impact on subscription renewals.
“It is difficult to personalize the reading material and offer interesting content, because you need such a huge amount of it. Even AI will have trouble finding something to read for you if the content isn’t there,” Puumala says, summing up the problem.
Solutions tailored to the US won’t cut it in Finland’s small market
According to Ahti Ahde, one of Qvik’s heavy-weight data crunchers, the low amount of data is a special problem for small language areas and markets.
“Finnish news sites cannot compete with German or English sites in the amount of users and content, so machine learning runs into issues with the volume and quality of learning data. This is why the quality of Google’s or Amazon’s Finnish-language cloud services, for example, is nowhere near the major world languages.”
Even though AI and machine-learning tools are currently designed for large markets, that won’t stop Qvik from creating solutions for Finland.
“We are working on things like adding user-based session analysis to collaborative filtering methods, semi-supervised topic modeling solutions and the further refinement of unsupervised methods with, for example, the interleaving method used by Netflix. These methods seek to identify users better on every device and offer them the perfect content at the perfect time, for the right device.”
For the press and media, we are developing tools for making work like background research, studying a new topic or bringing in new perspectives from successful stories easier.
“When people think about data, they often worry about the costs of computing power. But the costs are normally not the issue. Rather, the problem is finding projects of the right size to avoid the issues with learning data, i.e. being able to make meaningful analyses and reach your financial or operational goals,” Ahde concludes.