Aim of Research:
To propose an adaptable conversational AI chatbot system that “naturally” interact with learners and generate answers with different levels of details based on learners’ profiles and acquired knowledge from previous conversation sessions. 

Proposed Solution:

• Train a Deep Reinforcement Learning and Deep NLP techniques (text summarization, embedding and transformers) based conversational chatbot to improve the future path of conversations and generate answers that improve the learning experience.
• Integrate learner profiles in a continual learning loop and mutually update the learner’s profile and tailored responses with the global award of increasing the quality of conversations.
• Proof of concept: experiment the chatbot and apply it to a specific domain such as “Deep Learning” to answer a given question with different level of details.

Team: Haruka George’23 (MS Data analytics)
Lifespan: 15/8/2022 – 15/5/2023

Aim of Research:
Social media serves as a platform to outsource all kinds of information from politics to entertainment, from health industry to the country’s administration. Some of the posts on these platforms are reliable, but most of them have some proportion of misinformation. The diversity of posts, such as the use of different languages, abbreviating words and messages with hidden meanings, make it more complex to identify the authenticity of the published information. One of the main challenges in detecting misinformation is that false claims can appear in countless variations and scarcity of ground truth labels to build trustworthy classifiers over time.

Proposed Solution:
This research will focus on identifying an end-to-end framework to find misinformation in social media postings using a NLP integrated hybrid approach, combining semi-supervised and deep supervised learning. More precisely, we are examining the following research question on “How effective are the semi-supervised algorithms in identifying misinformation in online social media postings (twitters)? “ The objectives for evaluating this research question are summarized as follows:

• Determine how labels can be propagated within a dataset that is largely unlabelled
• Implement hybrid models for classifying whether or not tweets on Twitter include misinformation (true for facts, false for misinformation) •
• Evaluate the goodness of manual labelling using probabilistic fact-checkers.
• Proof of concept: experiment the framework and apply it to tweets of potential health misinformation related to Covid-19, collected from 03/25/2020 to 10/30/2020 and benchmarks datasets from the literature. 

Team: Deeksha Joshi’23 (MS Data Analytics)
Lifespan: 15/8/2022 – 15/5/2023

Aim of Research: 
The rise of Massive Open Online Courses (MOOC) has propelled the rise of e-learning with all-time availability of learning content in the recent times. This scenario demands a similar perpetual tutor which can solve domain related doubts of the users and increase interaction with the learning content. The traditional dialogue systems lack personalization and fail to interact in an effective natural way to the learner. Tutoring Systems (ITS) lack personalization and fail to effectively adapt to newer and sophisticated neural approaches resulting in reduced effectiveness of knowledge acquisition. Therefore, it is vital to design an authoring tool that publishes a knowledge-rich domain- specific tutoring assistant which can help the user converse about a particular topic to solve queries/doubts related to a topic the student is learning on the MOOC’s platform.

Proposed Solution:
We proposed a methodology and algorithms to generate a knowledge graph-driven intelligent tutoring system, which pedagogically utilizes Natural Language Understanding (NLU) techniques to improve the quality of interactions of the tutoring system and the learner. The proposed domain-specific open knowledge graph built from scraping content from wikipedia and Medium websites, and contains a total of ~190,000 entities for specific domains such Machine Learning, statistics, and AI. The graph is used to train and tune a Dual Intent and Entity Transformer (DIET) and achieves accuracies, F-1, recall, and precision values in the range of 90-95% in classifying intents and entities in user messages. 

A minimum viable prototype, called Impulso, is under development with Graken (renamed to TypeDB) and RASA by a startup. 

Team: Atharva Mungeei’21 (MS Data Analytics)
Lifespan: 15/8/2020 – 15/5/2021

Aim of Research:
Automation is the new future of the world, which is changing complete the lifestyle of people by integrating technology into their daily activities. Automating building operations and conversion of a building into a “Smart Building” is one of the significant sectors gaining interest in recent years. It is challenging to implement such a strategy due to multiple constraints and factors which contributes to unpredictable variations of the temperature inside buildings (external weather, occupancy, day/night, seasons, appliances etc.). Most of existing solutions rely on supervised learning techniques to build models, which require historical training datasets. However, these models often become obsolete when the environments changes (new features, new appliances, etc.).

Proposed Solution:
This research aims to propose a deep reinforcement learning based solutions/scenarios that allow controlling the temperature according to user preferences while reducing energy costs . Deep reinforcement learning based energy-efficient heating controllers experiments different rewards: the comfort-driven reward (setpoint temperature vs indoor temperature) and integrated cost and comfort driven reward, and different settings: one building – one controller, multiple buildings-one controller, and multiple buildings-multiple controllers. 

Team: Anchal Gupta ’19 (MS Data analytics)
Lifespan: 15/8/2018 – 12/15/2019

Publication:

Anchal Gupta, Youakim Badr, Ashkan Negahban, Robin Qiu, “Energy-efficient heating control for smart buildings with deep reinforcement learning,” Journal of Building Engineering, Vol 34, Article number 1101739, 2021 (2021Impact Factor: 7.1)