Deepak Pai

Padati: A Robust Pedestrian Dead Reckoning System on Smartphones

Location based services are proving to be the next driving factors for growth in smartphones. While GPS solves the problem of accurate localization in outdoor environments, indoor localization is still an area of active research. Emergence of new generation smartphones with low cost sensors, have provided an effective way of indoor localization by pedestrian dead reckoning (PDR). We propose a robust mechanism for detecting the step of a person and estimating his step length. Our system is independent of the location and orientation of the device. Our system is shown to perform 45% better than the traditional PDR systems proposed in prior-art. Another important problem in PDR system is determining the orientation of the mobile device and the direction of user motion. Many systems assume the device to be oriented in the direction of the user motion. Some of the recent systems use accelerometer, magnetometer patterns and PCA to detect the direction of user orientation. We propose a system which uses map matching and particle filtering to determine the direction of user motion. We tabulate our findings on the feasibility of such a system.

Modelling Visit Similarity Using Click-Stream Data: A Supervised Approach

Identifying and targeting visitors on e-commerce website with personalized content in real-time is extremely important to marketers. Although such targeting exists today, it is based on demographic attributes of the visitors. We show that dynamic visitor attributes extracted from their click-stream provide much better predictive capabilities of visitor intent. In this work, we propose a mechanism for identifying similar visitor sessions on a website based on their click-streams. Novel techniques for extracting features from visitor clicks are employed. Large margin nearest neighbour (LMNN) algorithm is used to learn a similarity metric between any two sessions. Further the sessions are classified into purchasers and non-purchasers using k-nearest neighbour (kNN) classification. Experimental results showing significant improvements over baseline algorithms based on Hidden Markov Model(HMM), support vector machine (SVM) and random forest are presented on two large real-world data sets.

Wally: crowd powered image matching on tablets

In this paper we propose a crowd sourced approach for solving large scale object retrieval. We have built a tablet application which displays a query image and a database image. The crowd provides their input to indicate, if there is a match between the query and database image or not. We test our application on a crowd of low-income individuals. We observe that our target crowd had a very high accuracy on the considered dataset. We observe significant improvement as compared to vision based image matching algorithms available in prior-art. We also observe that with simplistic interfaces, even low literacy and low income people could participate in the crowdsourcing tasks. This provides them a significant income opportunity. We have validated our claims on two publicly available University of Kentucky datasets and ORL Face recognition dataset.

EnTwine: Feature Analysis and Candidate Selection for Social User Identity Aggregation

Organizations measure their social audience based on the number of users, fans, and followers on social media. Every social media platform has its user identity and a single user is present across varied platforms. Due to the disconnected user profiles, identifying duplicate users across media is non-trivial. There is a need to create a complete view of a user for various applications such as targeting and user profile construction. This view is not easily available due to the individual identities. In this work, we explore the feature space across social media that can be leveraged for intelligent user identity aggregation. Further, we present a two-phased unified identity creation process using our feature analysis, unsupervised candidate selection, and supervised user matching algorithms on four different social networks.

Real-time visualization and targeting of online visitors

Identifying and targeting visitors on an e-commerce website with personalized content in real-time is extremely important to marketers. Although such targeting exists today, it is based on demographic attributes of the visitors. We show that dynamic visitor attributes extracted from their click-stream provide much better predictive capabilities of visitor intent. In this demonstration, we showcase an interactive real-time user interface for marketers to visualize and target visitor segments. Our dashboard not only provides the marketers understanding of their visitor click patterns, but also lets them target individual or group of visitors with offers and promotions.