Predicting Selective Availability for Instant Messaging

Abstract

Instant messaging (IM) systems allow users to spontaneously communicate over distance, yet they bear the risk for disruption of the recipient. In order to reduce disruption, novel approaches for detecting and presenting mutual availability are needed. In this paper we show how fine-grained IM availability predictions can be made for nomadic users solely based on sensors installed on a laptop computer. Our approach provides comparable accuracies to previous work, while it eliminates the need for augmenting the offices or the users with further sensors. We performed a user study to collect sensor data. Alongside with labels collected by means of Experience Sampling, the data allow for creating probabilistic models for predicting selective availability. This way, we demonstrate how the required effort involved in proactively managing one's availability selectively towards a variety of recipients can be reduced by automatic adaptation, and give insights in the lessons learned.

LiLoLe—A Framework for Lifelong Learning from Sensor Data Streams for Predictive User Modelling

“…Further, as recent work has shown, such general models are not feasible in more complex settings. For example, when Fetter et al [10] examined the performance of general vs. personal models in the context of predicting Instant Messaging availability of nomadic users based on a number of sensors in mobile settings, they found several aspects that degraded the performance of general models such as: the learned concept (i.e., availability) is highly personalised, and the variation between users was accordingly high; some features that had high predictive power for an individual user had a limited predictive power for other users, and thus were discarded; and some features that worked well in the individual models even had contradictory information for other users when used in general models. Our approach accounts for these challenges.…”

“…For example, if several people share the preference for muting their work phone when at home, the sensor data that allows localising a person as at home, is individually different (e.g., BSSID of the home network) [10].…”

“…In GUI based systems, an experience sampling based approach [10,16] is feasible, as exemplified for our scenario in Fig. 2.…”

“…The input arrives from a plethora of heterogeneous sensors and the adaptation decisions often imply personalised aspects and individual preferences that also can change over time. An illustrative example is the scenario of mediating interruptions by predicting the interruptibility of users [10,11,17], where the data of numerous different sensors is brought together with personal interruptibility reports to predict the availability of a user for communication. In such scenarios, general models do not offer enough accuracy and thus fail to offer adequate adaptation support for various reasons, as reflected on by Fetter et al [10].…”

“…An illustrative example is the scenario of mediating interruptions by predicting the interruptibility of users [10,11,17], where the data of numerous different sensors is brought together with personal interruptibility reports to predict the availability of a user for communication. In such scenarios, general models do not offer enough accuracy and thus fail to offer adequate adaptation support for various reasons, as reflected on by Fetter et al [10]. In order to overcome such challenges the idea of lifelong learning systems for user modelling [18] was introduced.…”