5. Recommender Systems - Understanding Implicit Feedback
Recommender systems rely on two different types of feedback (input). The most convenient is the high quality explicit feedback, that consists of an explicit preference given by a user for a given item. Often explicit feedback is collected via a star rating system. However, this kind of (precious) feedback is not always available (only few user give explicit feedback for items they have purchased) so recommender systems have to infer user preferences from the far more abundant implicit feedback. In an implicit model, once the user has given consent to collecting usage data, no additional explicit feedback action (e.g. ratings) is required on the user's side.
Implicit feedback includes purchase history, browsing history, search patterns, etc. For example, a user that purchased many books by the same author probably likes that author.
However it is crucial to identify and understand the most important characteristics of implicit feedback, that are:
1. No negative feedback
When observing users behaviour, we can infer only which items they probably like. It is difficult to reliably infer which items a user didn't like. For example, a user that didn't purchase an item might have done so because she did not like the item or just because she did not know about that item or the item was not available. This asymmetry does not exist in explicit feedback where users tell us both what they like and what they dislike. Recommender systems that rely on explicit feedback, therefore, tend to focus on the collected information, user-item pairs for which a rating exists. All the other user-item relationships, which constitute by far the majority of the data are treated as missing data. With recommender systems that rely on implicit feedback, we cannot only concentrate on the gathered information because it will leave us only with positive feedback, significantly misrepresenting the full user profile. It is crucial to address the missing data which is where the negative feedback is expected to be found.
2. Implicit feedback is noisy
When tracking users behaviour we can only guess their preference and true motives. For example we may collect a purchase action for a user, but this does not necessary imply a positive feedback. The item could have been purchased as a gift, or it could also be that the user was disappointed with the item.
3. Preference vs Confidence
The numerical value of explicit feedback indicates a preference, whereas the numerical value of implicit feedback indicates a confidence. Systems based on explicit feedback let the user express their level of preference, e.g. a star rating between 1 (total dislike) and 5 (really like). On the other hand, numerical values of implicit feedback describe the frequency of actions. A larger value doesn't indicate a high preference, but the numerical value of the feedback tells us about the confidence that we have in certain observation. A one time event might be caused by various reasons that may have nothing to do with the user preferences, but a recurring event is more likely to reflect the user's opinion.
So, to recap. Implicit feedback is data we gather from the users behaviour, with no ratings or specific actions is needed. It could be what items a user purchased, browsing history, search patterns, how many times they played a song or watched a movie, how long they’ve spent reading a specific article etc. The upside is that we have a lot more of this data, the downside is that it’s more noisy and not always apparent what it means. Working with implicit feedback is harder than working with explicit feedback, but it's something a recommender system cannot (should not) avoid.
This article is written by Riccardo Saccomandi, Co-founder and CTO of Kickdynamic.