It has been a while that I've been studying about recommendation engines and how can they be applied on mobile apps. More and more data is exchanged between those platforms. Great examples of mobile apps that are using recommender engines are Google Hotpot and Foursquare. They not only connect people to each other, but help users discover places around them.
|Bizzy: A mobile recommender for Places|
However, they only scratch the surface in this field, which is considered a novel research area in recommender systems. There are several topics to be explored such as:
- The location-based recommenders suggest items based on how far away we are from them (sometimes this can be manually changed). This could be a problem, if you consider the distance as the main factor for your recommender. Let me explain with an example. Imagine that you receive music concerts recommendations from your app around your in radius of 2km. When you're looking for live music, there's a band playing 1 km which will be recommended, but your favorite band, which is playing 2.1 km aways, will be out of the final list. And worst, if my favorite band will play tomorrow and I am at home looking for recommendations, it won't be suggested in this case because of the distance. It is necessary those systems to consider our habits in order to provide recommendations based on the most checked-in places that I visit (one possibility).
- The mobile recommenders must consider the context where the users are inserted. However the recommenders currently must receive what people are looking for, before receiving the suggestions. Wouldn't be interesting to consider the time events or even the historical habits ? Are those factors enough ?
- The information about the location and place (content) is also important in the recommendation computation. Imagine you exploring places around you at Foursquare and there are trending places around you (lot's of people there). This recommendation will be received considering only distance ? It is necessary to consider the temporal information associated to the place.
- Just because I've been many times at the Nipon's Sushi , it doesn't mean that I don't want to receive the recommendation again. The process of discovery and re-discovery is important either. The current systems, in general, don't consider the user's familiarity with the locations the user frequent.
- Venues are venues. Events are the "main" item interested by the user, not the venue. Ok, I like receiving a recommendation of a place, but sometimes I'd like to know what's happening there. Furthermore, the current check-ins and rating systems reflect what's happening now, but not what I am planning in the future. The discovery process and prediction decision are main important issues when you're designing a mobile recommender, specially dealing with temporal short-life like events.
- Noisy data. Recommendations of "my apartment", "my mother house". Recommendations can suffer with those types of places.
- How do we collect the data ? It will be by check-ins, 5 *scale ? Or passive using GPS ? This brings issues about the recommendation interface and how the data will be influenced by the social-signalling noise.
- Some systems use the user's search history to recommend places. This can be noisy, considering that sometimes that what the user searches online often do not match what the user needs when he wants to go out. Just because I looked for information about the Recife airport, it does mean that I'd like to receive airport recommendations. The point is: the relevance of places that you search for online doesn't match places that you would like to discover in the real world.
Those are some of several challenges faced by mobile recommenders. Mobile recommender systems are still growing and there's lots of research around it. But one important observation to make is that the best techniques for recommenders in the web sometimes are not suited for mobile recommenders. It's required that the recommendation designer be able to balance between the distance and the preferences from the user and the related items, understand the context where the user is inserted and help him to discover and re-find places and events hidden in his city!
There will be a workshop during the ACM Recommender Systems 2011 about mobile recommenders. Unfortunately this year I won't be attending , but it is on my plans! By the way, you still can submit your paper until July's 25th!
PS: I've found this great blog about mobile recommendations and mobile data mining : Urban Mining. I recommend!
PS2: Recommended reading about why people check-in. An interesting research about why people are interested in checking-in mobile applications.
I hope you enjoyed this article,