Structural Causal Model(SCM)-based Data Augmentation for Robust Session-based Recommendation, TCS Research

It is well known that DNNs-based models for classification ideally require a large amount of balanced data, i.e., each class should have a sufficient and similar number of data instances. However, real- world data is often skewed and suffers from several biases.Popularity bias in Recommendation Systems pose serious problems when less popular/new items are present. Users interact with a small number of items, which results in a skewed distribution of items clicked/ordered by users and makes the observed dataset extremely sparse. A neural network based temporal SCM framework was proposed to generate the high-quality counterfactual data to enhance the performance of SR models.

Mtech Thesis: Counterfactual Inference and Survival Analysis, IIIT Delhi

Thesis Advisor: Dr. Ranjitha Prasad

In order to answer question such as ‘Will changing the treatment regimen delay the onset of diabetes?’ requires predicting counterfactual using survival analysis. Proposed a novel SurvCI framework which consists of a counterfactual inference algorithm for data with time-to-event outcomes. The proposed approach incorporates a personalised parametric mixture density to model survival functions and outperforms the baselines on synthetic and semi-synthetic dataset.

Speaker Diarization, IIIT Delhi

Advisor: Dr. Pravesh Biyani

Devloped the Speaker Diarization pipelines to find who spoke when in an audio and total number of speakers in an audio recording of Hindi Dataset. Reduced the diarization error rate (DER) to 12.1% using Mean Shift Clustering.