Projects
Research/Industry projects
- SketchSimRank - SimRank for Graph Streams (Theoretical Guarantees) (Advisor - Prof. Soumen Chakrabarti, IIT Bombay)
Original SimRank proposed for static graphs computes the similarity ranking in a recursive fashion.
Here we propose simrank with approximate guarantees for graph streams (dynamic graphs) using only fixed space (hashing) More to be updated in description
- Recommendation Systems towards better medical predictions (Advisor - Amit Sharma, Microsoft Research India)
Developed user feature embeddings based on user responses logging and telemetry logging for building recommendation models for recommending different therapeutic activities and sections of the app as interventions to the user.
Started with initial collaborative and content based filtering for recommendations, and further developed a causal recommendation model where each micro-intervention was indicated as treatment to the user.
- Real Time Anomaly Detection in Enterprise environment - NetApp Collaboration (Advisor - Niloy Ganguly, Bivas Mitra, Mainack Mondal, Indian Institute of Technology, Kharagpur)
- Proposed a novel anomaly detection along with failure prediction in an enterprise setting with modules/microservices interacting with each other. Studied how module/microservice interaction changes w.r.t time within the normal period and the anomaly period.
- For online phase, based on logging intervals proposed a thresholding mechanism to indicate if the corresponding next logging interval is anomalous based on previous behaviour. Presented our method to researchers at NetApp for different enterprise datasets like Hadoop, IBM Bluegene.
- Theoretical Analysis of Metric Learning Algorithms (Advisor - Ashish Ghosh, Indian Statistical Institute)
- Performed theoretical study on different Metric Learning algorithms to learn similarity metric from data distribution.
- Did an empirical analysis as well as evaluation of metric learning methodologies w.r.t different datasets like Iris, Wine Dataset, thus showcasing performance \& limitations across various data distribution.
- Group activity Recognition (*Advisor - Bivas Mitra, Sandip Chakraborty)
- Performed a theoretical study on traditional group activity recognition models.
- Based on temporal sensor data distribution, estimate missing data through Expectation Maximization algorithm.
- Estimated Group formation over temporal sensor data like SSID-WiFI values
- Studied the correlation effect and casuality for different user features like GPA,locationData w.r.t group activity and formation
- Contributed to the theoretical analysis of the model (initial working of GroupSense) that got acknowledged in Paper.
Course projects/Initiatives