1. Information about the paper
Do, Trinh Minh Tri, and Daniel Gatica-Perez. “Contextual conditional models for smartphone-based human mobility prediction.” In Proceedings of the 2012 ACM conference on ubiquitous computing, pp. 163-172. ACM, 2012.
Human behavior is often complex and context-dependent. This paper presents a general technique to exploit this “multidimensional” contextual variable for human mobility prediction. We use an ensemble method, in which we extract different mobility patterns with multiple models and then combine these models under a probabilistic framework. The key idea lies in the assumption that human mobility can be explained by several mobility patterns that depend on a subset of the contextual variables and these can be learned by a simple model. We showed how this idea can be applied to two specific online prediction tasks: what is the next place a user will visit? and how long will he stay in the current place?. Using smartphone data collected from 153 users during 17 months, we show the potential of our method in predicting human mobility in real life.
2. My review of the paper
This paper tried to address two fundamental tasks for mobility prediction: what is the next place a user will visit?; how long will he stay in the current place?. The input for this approach is the user context: partial path, current location, time. And the output is the variable we want to predict: next destination, arrival time. It also uses additional contextual information as input, such as density of bluetooth devices nearby and various application logs in the phone.
The approach is: collecting and extracting the input using previous work on mobile sensing, then predict the output by combining previous works’ prediction model to achieve better result. It has successfully demonstrated a good prediction result based on its proposed approach.
This paper offers new method which combines previous works into one single framework. Therefore, the contribution is not really new.
- This paper introduces new framework for predicting mobility pattern.
- It has 153 users in 17 months data collection, which is quite large training data.
- It has been cited by 24 following works, which indicates it is a good foundation for following works.
- Some limitations are already stated in the paper:
- Low rate of trusted observations, could be addressed in future work, using maximized sensing techniques or exploiting/predicting missing observations.
- Some future directions are already stated in the paper:
- improving predictive result with more contextual variable and other mobility pattern.
- applying the idea into other human behavior prediction, not just place and time prediction.
- Compute prediction result whenever the context changes, not just at arrival or leaving a place.
- Using other contextual variables or using non-parametric methods.
- Other interesting future direction from the following work which turns out to be from the same author of this paper is: predicting what apps will likely be used by the user in the next place.
- I am thinking of future work that incorporates social networking. For example if two users or more are using this same prediction framework, then we can infer more relevant prediction based on their relationship and their meeting occurences. The can be spouses, friends, or coworkers.
- I am also thinking that this prediction will only be useful to the user if the application gives suggestion or recommendation according to the next place predicted. Because just predicting the next place a user will be is not useful to the user. He has already known that. For example, if a user is predicted to go to a particular restaurant he usually go for lunch, the app could show the current menu in the restaurant, and with further coordination with restaurants, it could directly order what he wants. And by the time he gets there, his order is ready to eat. Also, the better part is, to incorporate this feature into smart-assistant like Siri (iPhone), Cortana (Windows Phone), and Google Now (Android). Some subset of this idea has already been implemented in these smart-assistants. Or, the user can ask to open Yelp or other recommendation platform (or a new social recommendation platform altogether based on this approach), whenever he wants to try something new for lunch. And the app should suggest the new restaurant based on restaurants he liked in the past.
- Interaction with user is also a good approach. Instead of blindly predicting, we could use the user’s help to personalize the prediction training. And if the app will suggest or recommend things based on prediction that will actually be useful, I think users will want to cooperate. We can use pandora.com as an analogy. It is a music streaming platform that uses learning algorithm to predict the next song we like to hear. Whenever it plays a song, it provides two icons: thumbs up when we like the predicted song, and thumb down when we do not like the predicted song. Then the algorithm will learn more and gives better prediction the next time, based on this thumb input. We can leverage this approach to improve this mobility prediction further.