What is it about?

News recommenders predict a user's interest in a news item, typically based on their reading history. The currently most successful algorithms for this purpose are deep learning based, and are difficult to interpret. We do not know for which reasons such a recommender proposes a certain news item to a user. In this work, we found a way to open this black box. Our method reveals which news in a user's reading history have an influence on a particular recommendation. This way, we can check what the system learned and whether it actually attends to reasonable parts of reading histories to compute its recommendations.

Featured Image

Why is it important?

News recommender systems are deployed in many commercial applications. They can help us handle the large quantity of incoming news every day and point out the ones that are actually relevant to us. At the same time news consumption is an important factor in our opinion formation and recommenders may affect it. A blackbox system may not function correctly, point out the wrong news or even bias its users. Therefore, it is very important to get a better understanding of what is going on in such systems. Our method can verify whether individual recommendations are reasonable regarding what the user has read in the past, and we can also use it to identify bad recommendations that do not come from any meaningful correspondence to previously read news.

Perspectives

Our results have two faces. On the one side they show that in many cases news recommenders do base their predictions on related news and topics in a user's reading history. So here they function as we would expect. However, we also see cases where they strongly recommend certain news but users have actually not read anything in the past that is related to this recommendation. This is concerning because it shows that the system does not only learn what we expect. When digging deeper we find that the issue is in the user representations that the recommender learns internally. It leads to some news being recommended to effectively all users no matter what they have read before. This appears to be related to the training task that is typically used for such news recommenders. They are taught to predict clicks on news based on what someone has read before. However, we click many news precisely because we have not read about them before, or may not click them because we have already read so much about them. Typical recommenders are not equipped to differentiate between such cases and we believe that addressing this issue will be important to make news recommenders more reliable and safer in the future.

Lucas Möller
Universitat Stuttgart

Read the Original

This page is a summary of: Explaining Neural News Recommendation with Attributions onto Reading Histories, ACM Transactions on Intelligent Systems and Technology, June 2024, ACM (Association for Computing Machinery),
DOI: 10.1145/3673233.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page