Shreya Havaldar, Hamidreza Alvari, John Palowitch, Mohammad Javad Hosseini, Senaka Buthpitiya, Alex Fabrikant. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

We introduce Zebra regression, a novel streaming regression task. A stream of periodically-created point particles travels along a finite, 1-dimensional trajectory, and we observe a stream of sporadic, noisy reports that estimate the location of some particle, but without exact information on which particle is being observed. We seek a streaming algorithm that maintains accurate estimates of the number and locations of the particles that are currently en route. Our motivation comes from locating "particles" that are public transit vehicles serving "frequency based" routes --- routes whose schedules announce an approximate frequency of departures, rather than explicit per-trip timings. We expect similar problems to also arise in other settings where observations are scarce and stochastic, and object identity is not readily available.

We present BusTr, a machine-learned model for translating road traffic forecasts into predictions of bus delays, used by Google Maps to serve the majority of the world's public transit systems where no official real-time bus tracking is provided. We demonstrate that our neural sequence model improves over DeepTTE, the state-of-the-art baseline, both in performance (-30% MAPE) and training stability. We also demonstrate significant generalization gains over simpler models, evaluated on longitudinal data to cope with a constantly evolving world.

In this paper we develop a model capable of classifying drivers from their driving behaviors sensed by only low level sensors. The sensing platform consists of data available from the diagnostic outlet (OBD) of the car and smartphone sensors. We develop a window based support vector machine model to classify drivers. We test our model with two datasets collected under both controlled and naturalistic conditions. Furthermore, we evaluate the model using each sensor source (car and phone) independently and combining both the sensors. The average classification accuracies attained with data collected from three different cars shared between couples in a naturalistic environment were 75.83%, 85.83% and 86.67% using only phone sensors, only cars sensors and combined car and phone sensors respectively.

Recent mobile phones integrate fingerprint scanners to authenticate users biometrically and replace passwords, making authentication more convenient for users. However, due to their cost, capacitive fingerprint scanners have been limited to top-of-the-line phones, a result of the required resolution and quality of the sensor. We present Bodyprint, a biometric authentication system that detects users' biometric features using the same type of capacitive sensing, but uses the touchscreen as the image sensor instead. While the input resolution of a touchscreen is ~6 dpi, the surface area is larger, allowing the touch sensor to scan users' body parts, such as ears, fingers, fists, and palms by pressing them against the display. Bodyprint compensates for the low input resolution with an increased false rejection rate, but does not compromise on authentication precision: In our evaluation with 12 participants, Bodyprint classified body parts with 99.98% accuracy and identifies users with 99.52% accuracy with a retry likelihood of 26.82% to prevent false positives, thereby bringing reliable biometric user authentication to a vast number of commodity devices.

As mobile context-aware services gain mainstream popularity, there is increased interest in developing techniques that can detect anomalous activities for applications such as user authentication, adaptive assist technologies and remote elder-care monitoring. Existing approaches have limited applicability as they regularly poll power-hungry sensors (e.g., accelerometer, GPS) reducing the availability of devices to perform anomaly detection. This paper present SALCS (Soft Authentication with Low-Cost Signatures), an approach for anomaly detection on a user's routine comprised of a collection of anomaly detection techniques utilizing soft-sensor data (e.g., call-logs, messages) and radio channel information (e.g., GSM cell IDs), all of which are available as part of a phone's routine usage. Using these information sources we model aspects of a person's routine, such as movement, messaging and conversation patterns. We present extensive evaluations of the individual anomaly detection techniques, compare the collection SALCS to an existing power-hungry approach showing SALCS has a 7.6% higher detection rate and gives 5x better coverage throughout the day.

Virtual world conferences have been shown to give users an increased sense of presence in a collaboration as opposed to teleconferences, video-conferences and web-conferences. Such telepresence encourages remote participants to engage in the collaboration. Current virtual world collaboration applications rely on mouse/keyboard interfaces to create pure-virtual collaborations. In this paper we propose HyPhIVE, a system to address hybrid collaboration betweenthe physical world and virtual worlds. In hybrid collaboration scenarios, a group of people collaborate in the real world and others join them remotely via a virtual world. HyPhIVE uses non-intrusive mobile sensors to detect real world users’ collaboration context such as their position, direction of gaze, gestures and voice. HyPhIVE projects the sensed real world collaboration into a virtual world in a way that collaboration patterns are preserved. Remote users join the collaboration using virtual world clients and interact with other users’ avatars. User studies have shown that HyPhIVE effectively projects real world collaborations into a virtual world and it improves users’ experience of remote collaboration.