Publications

2023

1. 

   HRescue: A Modern ML Approach for Employee Attrition Prediction

   Rudresh Veerkhare, Parshwa Shah, Jiten Sidhpura, and 1 more author

   In Machine Learning and Big Data Analytics, 2023

   Abs PDF

   The biggest strength for any organization is its employees. Companies invest a lot of money and time in employees to retain them. A lot of opportunities are available for employees in this competitive technological world. Employees are leaving the organization for various reasons leading to a high attrition rate. It is not only the employees but also their talent, skills, and values leave the organization. Machine learning models can assist companies in attrition prediction. In this study, we have used the IBM HR Analytics dataset. Being an imbalanced dataset, we used ADASYN to balance it. Later, we used state-of-the-art gradient boosting algorithms such as XGBoost and CatBoost on original and balanced datasets. We also implemented a new privacy-based approach with the help of federated learning that will help different companies to train a single robust model. Finally, with the help of SHAP values, we analyzed our model predictions. For model evaluation, we used F1-Score and accuracy as our core metrics. We obtained an F1-Score of 0.69 and 0.94 on the imbalanced and balanced dataset, respectively. This solution can help companies to foresee employees who can resign so that they can take appropriate measures.

2. 

   FedSpam: Privacy Preserving SMS Spam Prediction

   Jiten Sidhpura, Parshwa Shah, Rudresh Veerkhare, and 1 more author

   In Neural Information Processing, 2023

   Abs PDF

   SMSes are a great way to communicate via short text. But, some people take advantage of this service by spamming innocent people. Deep learning models need more SMS data to be adaptive and accurate. But, SMS data being sensitive, should not leave the device premises. Therefore, we have proposed a federated learning approach in our research study. Initially, a distilBERT model having validation accuracy of 98% is transported to mobile clients. Mobile clients train this local model via the SMSes received and send their local model weights to server for aggregation. The process is done iteratively making the model robust and resistant to latest spam techniques. Model prediction analysis is done at server side using global model to check which words in message influence spam and ham. On-device training experiment is conducted on a client and it is observed that the losses of the global model converge after every iteration.

2022

1. 

   Face To BMI: A Deep Learning Based Approach for Computing BMI from Face

   Jiten Sidhpura, Rudresh Veerkhare, Parshwa Shah, and 1 more author

   In 2022 International Conference on Innovative Trends in Information Technology (ICITIIT), 2022

   PDF