Adversarial Attacks on Recommendation System Models [Link to Thesis]

  • Using hyperparameter tuning techniques to tune the graph data’s structure in order to drop the recommender’s performance
  • Attacking the user-item graph using meta-learning techniques, for reducing the generalization performance of collaborative filtering systems

Multimodal Transformers for Image Captioning

  • Image captioning formulated as a multimodal translation task — input image features are extracted using various encoder models and used for caption generation.
  • Used multimodal transformer to capture both intra-modal and inter-modal interactions in a unified attention block.

Multi-class Text Classification

  • Compared C-LSTMs (a combination of CNN and LSTM), transformers and deep averaging networks for classifying text documents.
  • Used self-attention mechanism at output, dynamic meta-embeddings at input and experimented with encoder blocks, positional embeddings, bi-gram embeddings to improve model performance.

Dynamic Meta-Embeddings for Improved Sentence Representations

  • Used attention weights in models to learn the importance of each type of embedding (GloVe, Word2Vec and FastText) for a given task.
  • Linear combinations of multiple word-embeddings used to perform sentence classification.

Textual Entailment Recognition

  • Problem formulated as the task of deciding whether a given text entails a given hypothesis.
  • Experimented with various architectures for performing textual entailment, and trained on SNLI corpus.

Duckworth Lewis Method

  • Implemented the Duckworth-Lewis method for predicting target runs in rain-affected cricket games using non-linear regression methods.
  • Performed extensive analysis of cricket matches data from 1999 to 2011.

Modelling the spread of COVID-19 in Karnataka, India

  • Implemented the SEIR and SEIRV (with vaccinations) algorithms to model the spread of COVID-19 in Karnataka, India using data from Jan 2021 to Sep 2021.
  • Analysed the effect of immunity waning, contact rate, vaccine efficacy, and parameters like mean incubation period, mean recovery period on the rate of new infections.

Movie Recommendation using Collaborative Filtering

  • Constructed and analysed the user-item interaction graph from the Movielens dataset.
  • Implemented matrix factorization, content-based recommender, collaborative filtering and neural collaborative filtering to recommend relevant movies to users.