1. Information about the paper

Citation

Shin, Choonsung, Jin-Hyuk Hong, and Anind K. Dey. “Understanding and prediction of mobile application usage for smart phones.” In Proceedings of the 2012 ACM Conference on Ubiquitous Computing, pp. 173-182. ACM, 2012.

Abstract

It is becoming harder to find an app on one’s smart phone due to the increasing number of apps available and installed on smart phones today. We collect sensory data including app use from smart phones, to perform a comprehensive analysis of the context related to mobile app use, and build prediction models that calculate the probability of an app in the current context. Based on these models, we developed a dynamic home screen application that presents icons for the most probable apps on the main screen of the phone and highlights the most probable one. Our models outperformed other strategies, and, in particular, improved prediction accuracy by 8% over Most Frequently Used from 79.8% to 87.8% (for 9 candidate apps). Also, we found that the dynamic home screen improved accessibility to apps on the phone, compared to the conventional static home screen in terms of accuracy, required touch input and app selection time.

paper-review-12.1

 

paper-review-12.2

 

 

2. My review of the paper

Summary:

This paper tried to infer the likelihood of an app to be launched next. Then, based on that information, it will create dynamic homescreen in the user’s smartphone. The user will then benefit from easier and faster access to his next app which most likely to be used.

The inference method used in this paper is by collecting three types of sensors data:

  1. User-related: GPS, cellular network location, 3D accelerometer, personal schedule, calls and SMSs
  2. Environment-related: Illumination, carrier, Wi-Fi, Bluetooth, screen status, battery status, setting status, and device status
  3. App-related: running apps, active app, app status

They ended up with 37 features from those sensors. The prediction model is based on the probability that a corresponding app will most likely be launched for a given context from those sensors features. They used naïve bayes classifier combined with Greedy Thick Thinning feature selection method.

As it turns out, most useful features for the inference method are: cell ID, hour of day, and previously used apps. The paper successfully demonstrated an 8% accuracy improvement over previous work that calculates Most Frequently Used apps.

Comments

Pros:

  1. It is a good idea to capture the data from power hungry sensors only when the sensors are already on. Then it will save unnecessary battery drain. It might lower prediction accuracy by not turning them on. However, since the user itself does not activate those sensors, they should not be significantly useful either if we turn them on.
  2. It has been cited by 34 following works in two years which is quite good and proves that this paper is useful.

Cons:

  1. It might be useful to mention what Android phones were used (since there are many) and which version of Android version used. There might be different features and different compatibility with different phones and different Android OS version. It might help people when they try to do further research.

Discussions and Future Directions:

  1. Since there was no statistical difference in preference between static home and dynamic home (4.1 vs. 3.7 on a 5 point Likert scale) in terms of satisfaction; and some participants said that they lose control over their home screen; I would suggest to put most likely to be frequently used apps in first row, and then make their icons static; then provides one other row of the most likely to be used apps with less probability to be frequently used.
  2. As already stated in the paper, it should accommodate newly installed app. Furthermore, we can suggest what app should be installed next, based on previously used apps and context from sensors. For example, games related to previously played games. Moreover, we can add this recommendation as the third row.
  3. It can also be developed further by also preload the suggested apps instead of just providing the icons. Then the user will also benefit from faster launch of the app. We can further predict not just an app, but an app category. Then suggest related highly reviewed (not yet installed) app within the same category.
  4. To some extent, this paper might not be significantly useful. Other than there was no statistical difference in satisfaction of using static and dynamic homescreen (especially when the user has tweaked his static homescreen to really fit his app usage pattern); one following work found that some apps might not be launched from homescreen, and launched from notifications or from other apps instead.
  5. Other feature that should be used as features is notification response. Furthermore, we can offer filtering notification using the result from this paper. For example, we can only show the user notifications from their most likely next app to be used in their home screen. We can further generate dynamic notification system that show relevant notification based on context. For example, we can delay a notification that is predicted to be more useful one hour later.
  6. Other feature that should be used is calendar, email, search queries. For example, if the user is still at home and should be at a meeting that should be reached by driving, we can suggest Google Maps app or Waze app with traffic information. From email, if detected that the user just bought a plane ticket, we can suggest information related to the destination and airport.
  7. As also stated in the paper, we should consider the possibility of a change in user’s apps usage pattern.
  8. We can also further develop useful features by incorporating this features into the smart assistants which are: Siri, Google Now, or Cortana for each respective platforms.
  9. We can also further provide a way to save the result in the cloud. Then if a user change their device, they can have this personalization instantly.