This is a short summary or overview I wrote after reading a conference paper from the ACM Digital Library. The original paper can be found here: Click here
Diversity in Smartphone Usage
A study was conducted to analyse the variation in smartphone use across a variety of users. Detailed usage traces were collected from 255 users, grouped into two datasets.
Dataset 1 – Consisted of 33 Android users out of which 17 were research workers and 16 were students. A custom logging utility was deployed on HTC Dream smartphones with unlimited plans. The logger collected detailed data such as the state of the screen, start and end times of calls, application interaction time, network traffic, and battery level. 7-21 weeks of data was gathered per user, with the average being 9 weeks. The logger ran in the background, keeping data records in a local SQLite database on the phone, and uploading them only when the phone was being charged.
Dataset 2 – Consisted of 222 Windows Mobile users across different demographic and geographic locations. The demographic categories were: SC (social communicators; required voice and text), LPU (life power users; needed a multi-function device), BPU (business power users; needed advanced phone for business), OP (organiser practicals; needed a simple device for management). 8-28 weeks of data was gathered per user by a third party, with the average being 16 weeks. Traces were collected using a logger that recorded start and end times of applications using API calls.
User Interactions – The logger kept track of sessions. For dataset 1, a session meant the screen was on or a call was active. For dataset 2, a session was when an application was in the foreground.
Application Use – For dataset 1 (Android), usage counters were updated by the OS, using a timer that ran from onStart(), onResume(), or onRestart() till the time that onPause(), onStop(), or onDestroy() were called in the activity lifecycle. For dataset 2 (Windows), timestamped records of start and end times of the foreground applications were made by the OS.
Network Traffic – Data was available only for dataset 1. All data sent or received by the phone was tracked, including 3G radio and 802.11 wireless link, except that exchanged over USB cable.
Energy Consumption – Amount of energy drain was estimated based on remaining battery indicator. If the battery percentage goes down by X% in a time period for a battery with capacity Y mAh, the approximate energy drain in that period is calculated as XY mAh.
Analysis of Result
Interaction Time – The time that users spent interacting with smartphones ranged between 30-500 minutes a day. There was a lack of cluster formation in the range, and users covered the entire range between the two extremes. This shows that it is more effective for a phone manager to learn a user’s behaviour dynamically than to map a user to a predefined category. It is seen that dataset 1 users had longer interaction sessions, while the number of sessions was similar for both the datasets.
Interaction Sessions – There were between 10-200 sessions a day on average, and the session lengths were around 10-250 seconds. Some applications had longer sessions, and even for the same application, different users had different session lengths. The distribution of a single user’s sessions as well as the time between sessions had a skewed distribution (most of them short, some very long). This showed that the longer a user has been in an ‘off period’, the greater is the chance that the user will continue in that state.
Diurnal Patterns – Diurnal patterns refers to the variation in the usage pattern of the smartphone during the day and night. Daytime use was seen to be much higher than nighttime use. Roughly 70% of users in each dataset had a peak hour usage greater than twice their mean usage. The diurnal ratio (ratio of mean usage during peak hour to mean usage across all hours) was calculated and it is seen that heavy users tend to use their phone more consistently during the day, while light users tend to have concentrated use during certain hours of the day.
Number of Applications – Users used around 10-90 applications each overall. This shows that there is high smartphone application usage on Android and Windows phones as well, apart from iPhones.
Application Popularity – Applications belong to one of several categories – communication, browsing, media, productivity, system, games, maps, and other. It is seen that users devote most of their attention to a subset of applications of their choice. High school students used communication and games applications more and researchers used productivity applications more, but the difference was minimal and not statistically significant. Some applications, such as messaging, were used more during the day than at night, indicating that there are diurnal patterns in application use.
Application Sessions – Close to 90% of the users used only one application in one session. It is seen that the popularity of games is twice as high when there are high session lengths.
Traffic Per Day – 1-1000 MB of data was received and 0.3-100 MB sent in a day in general. Users who consumed more traffic tended to use communication applications more, which shows that most of the network traffic is generated by communication applications.
Interactive Traffic – Interactive traffic refers to traffic generated while the screen is on, as well as for a short time window just before the screen is turned on. For 90% users, over 50% traffic is interactive, and for the rest, almost none is interactive. This implies that energy saving mechanisms which reschedule network activity so that bundling of traffic can be done will have varying impact.
Diurnal Patterns – 80% users generate twice their average amount of traffic in their peak hour. This is because there is a high proportion of interactive traffic, and interactions have a diurnal pattern.
Energy drain depends on user interactions and platform hardware and software. Energy drain magnitudes were highly varied, giving a battery life of 4-120 hours for different users using a 1200 mAh battery. There were also diurnal variations, as more energy was consumed during the day. This result shows that user activities contribute heavily to energy drain, with a greater impact than platform hardware and software.