Vishesh Mittal

Vishesh Mittal

Student at University of Southern California

Projects

Spatiotemporal Forecasting of Traffic Flow Data using GNN (Ongoing)

In recent times, the development of transport infrastructure has had a great boom, however, traffic problems continue to spread due to an increase in the population in urban areas which ultimately increases pressure on transport networks. Thus, we develop a GNN model (SST-GNN currently) which provides forecasts of traffic flow data such as velocity, etc. Additionally, we introduce DIVE dataset created by collecting traffic data over 20 nodes within the Old Panvel, Maharashtra of the Indian Province. The model achieved an RMSE of 1.76, 1.69 and 1.69 on the DIVE dataset at 15, 30 and 45 mins intervals for prediction.

Classification of ISL using Pose and Object Detection based Techniques

We present two approaches for the classification of Indian Sign Language: (a) Pose-based approach utilizes an LSTM model which takes the skeletal pose landmarks from Mediapipe for a sequence of frames as an input to infer and predict the action. (b) Object detection-based approach utilizes a model built on Scaled-YOLOv4 architecture which performs a frame-by-frame inference. LSTM model achieved an accuracy of around 98% for 8 classes. However, it was found from experimentation that this approach is not very scalable due to the drastic fall in model performance with increase in number of classes. The object detection route allows us to train far more number of classes (about thrice) on the Scaled-YOLOv4 Architecture with only little impact to performance with rise in number of classes. This provides a better and more scalable solution to Sign-to-Text translation with the ability to infer on images, videos and live-cam. The final accuracy of our YOLO model was 95.9% for 25 classes. The notebooks utilized for the project are available on github (link).

Customer Review Sentiment Analysis

The project aims to analyse the sentiment of amazon reviews with the help of Machine Learning & Deep Learning. Sentiment analysis (or opinion mining) uses natural language processing and machine learning to interpret and classify emotions in subjective data. This helps to further understand the overall sentiment of customers as well as predict the sentiments of incoming reviews. It was observed that DT has the lowest scores (0.78) while the rest of the models had a similar accuracy (0.86 - 0.87). Also, it can be seen that the neural network models appear to have similar accuracy (0.85) as that of the simpler ML counterparts however, it must be noted that this accuracy was achieved with half the data than that feeded to ML models. As evident, the neural network models outperform the ML models. Additionally, the models were deployed to Heroku and an application was developed using Flutter to predict sentiment of any input sentence using API calls. The dataset utilized is available on github (link). Our codebase for model creation (link), deployment (link) and flutter application (link) has been made public.