Frequency-based detection of task switches Rahul Nair Yahoo! Research Berkeley [email_address] Steve Voida & Elizabeth Mynatt Georgia Institute of Technology {svoida, mynatt}@cc.gatech.edu

A Question How did you know that this talk was beginning?

How did you know that this talk was beginning?

A Question How did you know that this talk was beginning? Increase in movement Fiddling with the projector New presenter looking around

How did you know that this talk was beginning?

Increase in movement

Fiddling with the projector

New presenter looking around

A Question How did you know that this talk was beginning? Increase in movement Fiddling with the projector New presenter looking around The amount of activity in the room changed

How did you know that this talk was beginning?

Increase in movement

Fiddling with the projector

New presenter looking around

The amount of activity in the room changed

Frequency-based detection of task switches Rahul Nair Yahoo! Research Berkeley [email_address] Steve Voida & Elizabeth Mynatt Georgia Institute of Technology {svoida, mynatt}@cc.gatech.edu

People are particular about Displays Users have a “working set” of windows related to each task (Henderson 1986) There are distinct differences in window layouts for each task (Hutchins 2004) As interested in hiding windows as they are in displaying task windows (Hutchins 2004)

Users have a “working set” of windows related to each task (Henderson 1986)

There are distinct differences in window layouts for each task (Hutchins 2004)

As interested in hiding windows as they are in displaying task windows (Hutchins 2004)

A frequency-based approach There is a shift in the window interaction frequencies as users “setup” for each task The shift could be either an increase or decrease in the average time between interactions Rearranging windows is as significant as opening or closing them

There is a shift in the window interaction frequencies as users “setup” for each task

The shift could be either an increase or decrease in the average time between interactions

Rearranging windows is as significant as opening or closing them

Software design Tracks the low level window manipulation Uses a sliding window algorithm to compare current interaction speed with session average If the window average exceeds the thresholds of the session average a switch is logged Best results with a 7 interaction window and a thresholds of 0.67 and 1.5

Tracks the low level window manipulation

Uses a sliding window algorithm to compare current interaction speed with session average

If the window average exceeds the thresholds of the session average a switch is logged

Best results with a 7 interaction window and a thresholds of 0.67 and 1.5

Software design If a switch is detected the software asks the user for confirmation Uses an IM style popup window Ignored popups are considered to be false positives 5 minute timeout between successive popups

If a switch is detected the software asks the user for confirmation

Uses an IM style popup window

Ignored popups are considered to be false positives

5 minute timeout between successive popups

Study Design 6 participants 2 Professors, 3 grad students, 1 IT professional Installed on their primary work machine 2 week study Followed by questionnaires and interviews

6 participants

2 Professors, 3 grad students, 1 IT professional

Installed on their primary work machine

2 week study

Followed by questionnaires and interviews

Results 21.33 48 225 IT Professional (IP) 40.66 37 91 Graduate Student 3 (G3) 53.92 117 217 Graduate Student 2 (G2) 33.51 123 367 Graduate Student 1 (G1) 56.58 43 76 Professor 2 (P2) 94.74 54 57 Professor 1 (P1) Accuracy (%) Switches confirmed Switches detected Subject

User Feedback P1 was very appreciative and said that almost all her tasks were detected P2 felt that short tasks were not always detected Rapid task switching behavior IP had the lowest accuracy but was the most enthusiastic co-opted out log files to fill out project time cards Noticed all switches but was over sensitive

P1 was very appreciative and said that almost all her tasks were detected

P2 felt that short tasks were not always detected

Rapid task switching behavior

IP had the lowest accuracy but was the most enthusiastic

co-opted out log files to fill out project time cards

Noticed all switches but was over sensitive

The Instant Messaging (IM) Effect Some of the variance can be explained by IM usage High accuracy subjects did not use IM while low accuracy subjects were regular IM users The software detected IM usage as a task switch while the users felt that IM was not a new task Users felt that IM was a side channel of information and were irritated when it was detected as a new task

Some of the variance can be explained by IM usage

High accuracy subjects did not use IM while low accuracy subjects were regular IM users

The software detected IM usage as a task switch while the users felt that IM was not a new task

Users felt that IM was a side channel of information and were irritated when it was detected as a new task

Advantages Can have extremely high accuracy Low computational cost Fewer privacy concerns since actual document and activity data is not being tagged Can spot a task switch without requiring identification of the task itself Time management applications

Can have extremely high accuracy

Low computational cost

Fewer privacy concerns since actual document and activity data is not being tagged

Can spot a task switch without requiring identification of the task itself

Time management applications

Future Work Integrating web browser URL information Dynamically adapting to users by adjusting window sizes, threshold values and popup timeouts Allow users to explicitly ignore windows like IM, etc…

Integrating web browser URL information

Dynamically adapting to users by adjusting window sizes, threshold values and popup timeouts

Allow users to explicitly ignore windows like IM, etc…

Questions? Rahul Nair [email_address] www.rahulnair.net

Questions?

Rahul Nair

[email_address]

www.rahulnair.net

Consider a situation where the algorithm is evaluating the n+1 th window event. t i is the time between the i th and (i-1) th event Task average, = Moving average, = Ratio = Algorithm

Consider a situation where the algorithm is evaluating the n+1 th window event. t i is the time between the i th and (i-1) th event

Task average, =

Moving average, =

Ratio =