Responsiveness in Instant Messaging: Predictive Models
Supporting Inter-Personal Communication
Daniel Avrahami and Scott E. Hudson
Human Computer Interaction Institute
Carnegie Mellon University, Pittsburgh, PA 15213
{ nx6, scott.hudson }@cs.cmu.edu

ABSTRACT
For the majority of us, inter-personal communication is an
essential part of our daily lives. Instant Messaging, or IM,
has been growing in popularity for personal and work-
related communication. The low cost of sending a message,
combined with the limited awareness provided by current
IM systems result in messages often arriving at
inconvenient or disruptive times. In a step towards solving
this problem, we created statistical models that successfully
predict responsiveness to incoming instant messages –
simply put: whether the receiver is likely to respond to a
message within a certain time period. These models were
constructed using a large corpus of real IM interaction
collected from 16 participants, including over 90,000
messages. The models we present can predict, with
accuracy as high as 90.1%, whether a message sent to begin
a new session of communication would get a response
within 30 seconds, 1, 2, 5, and 10 minutes. This type of
prediction can be used, for example, to drive online-status
indicators, or in services aimed at finding potential
communicators.
Author Keywords
Statistical models of human activity, Responsiveness,
Interruptibility, Availability, Awareness.
ACM Classification Keywords
H5.2. Information interfaces and presentation: User Interfaces;
H1.2. Models and Principles: User/Machine Systems.
INTRODUCTION
Inter-personal communication through Instant Messaging,
or IM, is gaining increasing popularity in the work place
and elsewhere. IM programs, or clients, facilitate one-on-
one communication between a user and their list of
contacts, commonly referred to as buddies, by allowing
them to send and receive short textual messages (“instant
messages”).
Unlike face-to-face communication, users of IM cannot
easily detect whether a buddy is available for
communication or not. As the use of IM is growing, and in
particular in the work place, the inability to detect a
buddy’s state can often result in communication
breakdowns with negative effects on both communication
partners. For the receiver, communication at the wrong time
might be disruptive to their ongoing work. If, on the other
hand, receivers simply decide to ignore communication, the
initiator’s productivity might suffer as they are left waiting
for a piece of information needed for their work.
If, however, we were able to accurately predict whether a
user was likely to respond to a message within a certain
period of time, then some of these breakdowns could be
prevented. For example, models could be used to
automatically provide different "traditional" online-status
indicators to different buddies depending on predicted
responsiveness. Alternatively, models can be used to
increase the salience of incoming messages that may
deserve immediate attention if responsiveness is predicted
to be low. One could also imagine a system whose role is to
allow its users to locate others who are available for
conversation (for example, to find other users who can
provide them with help or support) while hiding those who
aren’t. This would benefit users looking for help, whose
messages would be more likely to get a response, as well as
busy users who would be able to stay on task uninterrupted.
The work presented in this paper describes the creation of
accurate statistical models that are capable of predicting a
user’s responsiveness to incoming messages – simply put:
whether the receiver is likely to respond to a message
within a certain period of time. For example, of the models
presented in this paper, one was able to predict with 89.4%
accuracy whether a user will reply to a message within 5
minutes and another with 90.1% accuracy a response within
10 minutes (Figure 1).
Background
A number of benefits of using IM have contributed to its
increasing popularity. With its near-synchronous nature, IM
is positioned somewhere between synchronous
communication channels (such as phone or face-to-face)
and asynchronous communication channels (such as email,
newsgroups, and online forums). This near-synchronous
nature allows conversations to range from a rapid exchange

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise,
or republish, to post on servers or to redistribute to lists, requires prior
specific permission and/or a fee.
CHI 2006, April 22–27, 2006, Montréal, Québec, Canada.
Copyright 2006 ACM 1-59593-178-3/06/0004...$5.00.

731
CHI 2006 Proceedings • Using Knowledge to Predict & Manage April 22-27, 2006 • Montréal, Québec, Canada

of messages, to hours or even days passing between
messages in the same conversation. Since IM is inherently
asynchronous, users can choose when or whether to
respond to an incoming message. As noted by [ 25], users
welcome the ability to use “plausible deniability” when
electing not to respond to messages. IM is thus often
regarded as less disruptive than other synchronous
communication channels. In fact, IM is sometimes used for
communication even between users who share the same
physical work-space in an attempt not to disrupt one
another’s work. This asynchrony means that messages often
arrive when a user is engaged in other tasks. Indeed,
research shows that users often multitask when using IM
[ 14, 20, 25]. Particularly in the work place, messages may
thus arrive when a user is engaged in important and
potentially urgent work.
This means that while it is convenient and desirable for the
sender to initiate a conversation, it may be undesirable and
often inconvenient for the receiver. The receiver must then
choose between staying on task and engaging in
conversation. Staying on task and not responding may come
at a cost to the initiator, who may need some information
from the receiver. The receiver herself may incur a social
cost from being portrayed as unresponsive. Engaging in
conversation, on the other hand, will often come at a cost to
the receiver’s ongoing work [ 27].
One of the most important features of IM clients is the
ability to provide some awareness of presence. IM clients
typically provide this information by indicating whether a
user is online and whether the user is currently active or
idle (often referred to as the user’s “Online Status”). Most
IM clients also allow users to set additional indicators to
signal whether they are busy or away from the computer.
Those, however, are often insufficient as they require users
to remember to set and reset them [ 23]. Begole et al.
presented a system that was able to predict a person’s
presence based on observed patterns [ 5].
As noted in [ 4] and [ 11], knowing whether a person is
present, however, does not necessarily provide an
indication of whether or not that person is available for
communication. A user who is not present (typically
indicated as ‘offline’ or ‘idle’) is indeed not available for
communication. On the other hand, a user engaged in an
important task and unavailable for communication will be
indicated by an IM client as present (unless they
remembered to manually set their status to ‘Busy’).
Since the content or topic of an incoming message is
typically unknown to the user before it arrives, users
generally have to attend to all messages. While the tool
presented in [ 2] increases alerts to some messages based on
their content, it does not prevent default alerts from taking
place. As a result, users will sometimes elect to turn their
IM client off when they are busy, refusing incoming
messages altogether [ 25]. As Isaacs et al note, however,
most IM conversations held in the workplace are work-
related [ 20]. This makes closing the IM client a less
desirable strategy. Similar to the use of Caller ID in phones,
a user can typically also see who the sender of the message
is before attending to the message. However, even this brief
interruption can, in and of itself, be disruptive [ 13]. Results
from [ 1] and [ 8] suggest that, given information about the
receiver, senders would be able, and willing, to time their
messages to accommodate for the receiver’s state.
Interruptions and Interruptibility
Incoming instant messages join an ever growing number of
interruptions a person is exposed to. Those include
interruptions external to the computer, such as telephone
calls or people stopping by to ask a question, as well as
interruptions from various computer applications, including
alerts of incoming email, calendar notifications, or
notifications of new items from RSS feeds. Unlike face-to-
face interaction, most computer-generated or computer-
mediated interruptions occur entirely without regard to
whether the receiver is ready to accept them.
A number of studies have been performed showing the
negative effect of interruptions on people’s performance.
[ 13], for example, showed that even a very short
interruption can be disruptive, while [ 7] showed that even
an ignored interruption can have a negative effect. Field
studies on the effects of interruptions in the workplace
observed that, while interruptions can be beneficial to
people’s work [ 26], some perceive them to be such a
problem that they will physically move away from their
computer or even offices to avoid them [ 18]. In the
particular case of IM, we observed a number of managers
who refused to use IM for fear of being interrupted.
In previous work [ 19] we have demonstrated the ability to
create statistical models that predicted, with relatively high
accuracy, time periods reported by participants as highly
non-interruptible. [ 17], for example, presented statistical
models that were able to predict whether a user is “Busy” or
“Not Busy” with accuracy as high as 87%.
0
10
20
30
40
50
60
70
80
90
100
30sec 1min 2min 5min 10min
Predict response within
%

A
c
c
u
r
a
t
e
Figure 1. Accuracy of models predicting response to Session
Initiation Attempts (SIA-5) within 30 seconds, 1, 2, 5, and 10
minutes. Baseline prior probability is shown with the black
lines
732
CHI 2006 Proceedings • Using Knowledge to Predict & Manage April 22-27, 2006 • Montréal, Québec, Canada