Full text available: proceedings | direct link.
D. Kotkov,
S. Wang, and J. Veijalainen. Cross-domain recommendations with overlapping
items. In Proceedings of the 12th International Conference on Web Information
Systems and Technologies, pages 131-138, 2016.
Cross-Domain Recommendations with Overlapping Items
If you are
not familiar with recommender systems, you might want to read definitions
provided here.
Suppose you
would like to run your own music recommendation service. You have audio
recordings and a few users. You can suggest recordings to users based on
attributes of recordings. For example, if a user likes Madonna, then you can
suggest more songs of Madonna. However, this way to suggest recordings (Content-based
filtering) has quite low performance. If you had user ratings, you would use collaborative
filtering, but you do not have them in the beginning. What you could do is to
collect ratings from another service like Last.fm. In this paper, we wanted to
find out whether ratings from another service would help improve recommendation
performance.
It might
seem obvious that more data should help us recommend better, but that might not
be the case. For example, the service you want to collect ratings from might
have users with completely different behavior from the behavior on your
service.
To answer
this question, we collected the dataset from Russian online social network
vk.com. We then matched this dataset with ratings collected from Last.fm (not
all the recordings were matched). The results are rather straightforward. More
data improves the results. We also discovered that the more items have ratings from
both systems (overlapping), the better improvement. Low number of overlapping
items creates a bias in the data, which might decrease the recommendation
performance.
In a nutshell
- The combination of sources improves the recommendation accuracy.
- The improvement increases with the growth of items that have ratings from both systems.
No comments:
Post a Comment