Johnson Apacible

Learning and reasoning about interruption

We present methods for inferring the cost of interrupting users based on multiple streams of events including information generated by interactions with computing devices, visual and acoustical analyses, and data drawn from online calendars. Following a review of prior work on techniques for deliberating about the cost of interruption associated with notifications, we introduce methods for learning models from data that can be used to compute the expected cost of interruption for a user. We describe the Interruption Workbench, a set of event-capture and modeling tools. Finally, we review experiments that characterize the accuracy of the models for predicting interruption cost and discuss research directions.

BusyBody: creating and fielding personalized models of the cost of interruption

Interest has been growing in opportunities to build and deploy statistical models that can infer a computer users current interruptability from computer activity and relevant contextual information. We describe a system that intermittently asks users to assess their perceived interruptability during a training phase and that builds decision-theoretic models with the ability to predict the cost of interrupting the user. The models are used at run-time to compute the expected cost of interruptions, providing a mediator for incoming notifications, based on a consideration of a users current and recent history of computer activity, meeting status, location, time of day, and whether a conversation is detected.

Bayesphone: precomputation of context-sensitive policies for inquiry and action in mobile devices

Inference and decision making with probabilistic user models may be infeasible on portable devices such as cell phones. We highlight the opportunity for storing and using precomputed inferences about ideal actions for future situations, based on offline learning and reasoning with the user models. As a motivating example, we focus on the use precomputation of call-handling policies for cell phones. The methods hinge on the learning of Bayesian user models for predicting whether users will attend meetings on their calendar and the cost of being interrupted by incoming calls should a meeting be attended.

Balancing awareness and interruption: investigation of notification deferral policies

We review experiments with bounded deferral, a method aimed at reducing the disruptiveness of incoming messages and alerts in return for bounded delays in receiving information. Bounded deferral provides users with a means for balancing awareness about potentially urgent information with the cost of interruption.

Identifying salient entities in web pages

We propose a system that determines the salience of entities within web documents. Many recent advances in commercial search engines leverage the identification of entities in web pages. However, for many pages, only a small subset of entities are central to the document, which can lead to degraded relevance for entity triggered experiences. We address this problem by devising a system that scores each entity on a web page according to its centrality to the page content. We propose salience classification functions that incorporate various cues from document content, web search logs, and a large web graph. To cost-effectively train the models, we introduce a soft labeling methodology that generates a set of annotations based on user behaviors observed in web search logs. We evaluate several variations of our model via a large-scale empirical study conducted over a test set, which we release publicly to the research community. We demonstrate that our methods significantly outperform competitive baselines and the previous state of the art, while keeping the human annotation cost to a minimum.

Answering math queries with search engines

Conventional search engines such as Bing and Google provide a user with a short answer to some queries as well as a ranked list of documents, in order to better meet her information needs. In this paper we study a class of such queries that we call math. Calculations (e.g. 12% of 24

Clustering sequential data into hierarchical patterns

, square root of 120), unit conversions (e.g. convert 10 meter to feet), and symbolic computations (e.g. plot x^2+x+1) are examples of math queries. Among the queries that should be answered, math queries are special because of the infinite combinations of numbers and symbols, and rather few keywords that form them. Answering math queries must be done through real time computations rather than keyword searches or database look ups. The lack of a formal definition for the entire range of math queries makes it hard to automatically identify them all. We propose a novel approach for recognizing and classifying math queries using large scale search logs, and investigate its accuracy through empirical experiments and statistical analysis. It allows us to discover classes of math queries even if we do not know their structures in advance. It also helps to identify queries that are not math even though they might look like math queries. We also evaluate the usefulness of math answers based on the implicit feedback from users. Traditional approaches for evaluating the quality of search results mostly rely on the click information and interpret a click on a link as a sign of satisfaction. Answers to math queries do not contain links, therefore such metrics are not applicable to them. In this paper we describe two evaluation metrics that can be applied for math queries, and present the results on a large collection of math queries taken from Bings search logs.

Thematic response to a computer user's context, such as by a wearable personal computer

Sequential data, i.e. text string, is a common yet important data type. Automatically discovering patterns for sequential data is useful but challenging. In this paper, we address this task by clustering strings into hierarchical patterns. Such pattern hierarchy is particularly helpful for users to discover meaningful patterns as well as to interpret the encapsulated knowledge. We present the clustering algorithm in details and evaluate it on a large, real dataset of street addresses. The experiments demonstrate the effectiveness of our approach, making it a useful tool for analyzing and interpreting sequential data.