Sequential Recommendation using both Self-attention & Knowledge Graph This is a research I have currently been doing on Recommender System. Evaluated on Steam reviews, Amazon Beauty/Games, MovieLens dataset using metrics like Hit@k, NDCG@k, AUC. Written in Python, Pytorch. Implemented Translation-based Recommender using Numpy, compared with baseline model: Bayesian Personalized Ranking. Implemented Self-attentive Deep Neural Networks (sasRec) using ideas from Transformer, achieved state-of-the-art performance. Introduced hidden variable to sasRec to capture heterogeneous relationships among items. Generated knowledge graph (KG) embeddings using KB4REC, implemented graph convolutional network for learning structural proximity among entities in KG and sequential recommendation. Implement Tranformer as graph convolution network for learning structural proximity and long-term dependency in one network (undergoing).

Visual Question Answering & Image Captioning Reimplemented and Improved VQA 2017 & 18 champion's work. Evaluated on BLEU1 & BLEU4 score. Written in Pytorch Image captioning: Implemented pre-trained ResNet50 as encoder and LSTM as decoder, used teacher forcing for training, experiemnted with generation techniques: deterministic, stochastic, beam search VQA17: Implemented faster-RCNN to extract bottom-up attention, implemented GRU to generate top-down attention weights, used stochastic teacher forcing for training. Improvements: Experimented with pre-trained word embeddings: Word2Vec, Fastext and GloVe. Fine-tuned bottom-up attention, used Adamax for optimization, applied dropout & gradient clipping, experimented with different fusion methods. [PDF], [CODE]