MohammadJavad Mehditabar

Iran University of Science and Technology

• GPA : 18.60/20 (3.94/4)
• Supervisior : Prof. Sauleh Eetemadi [Email] [Google Scholar]

MBTI Personality Prediction Approach on Persian Twitter [Paper] [Poster]

We collected our Persian dataset by crawling Twitter user BIOs and tweets. Subsequently, we introduced the BiLSTM + Attention architecture, which outperformed the baseline model. Finally, we shifted our focus towards data analysis and model interpretability.

MBTI Sentiment Analysis of Long Persian Corpus using Hierarchical BERT

• Proposed a novel approach that leveraged two consecutive BERT models(Hierarchical BERT) surpassing the performance of two other implemented model Attention + BiLSTM and Sentence Aggregation with BERT.
• Peformed fine-tuning with Masked Language Modeling method on ParsBERT.
• Resolved GPU memory issue in spite of fine-tuning on both BERT models.
• Calculated TF-IDF and RNF (Relative Normalized frequency) to determine impact of words.
• Explored BERT explainability and intend to implement it.

Personality Detection Based On Tweets Of Users [GitHub]

• Utilized GPT-3.5 for data augmentation and also one-shot and zero-shot learning.
• Implemented sentence generation based on each label from GPT2 after fine-tuning our data on them. Furthermore, trained custom tokenizer based on sentence piece tokenizer.
• Proposed three different architecture BERT into BERT, truncated BERT and Word2Vec + ParsBERT embeddings for classification.
• Generate Word2Vec, FastText and BERT vector after training our dataset on them.
• Examinied papers with various model implementation, including SVM with leveraging LIWC and ConceptNet, LSTM and BERT.
• Studied and implemented NRC, TF-IDF approaches, and Resolved memory issue using Keras Data Generator.

Long Document Classification Methods On Transformers [GitHub]

• Implemented Hierarchical BERT, Sliding Window and Sentence Aggregation
• Investigated on LongFormer, docBERT, Unlimiformer and LongNet as a single transformer with longer input length, also analyzed ToBERT (Transformer over BERT) RoBERT (Recurrence over BERT) methods with effect of key sentence extraction.

Applied DataScience and Machine Learning

• Initially focused on data analyzing and visualization, Then on text mining and cleaning with Regex, Next concentrated on investigating classic machine learning models such as KNN, Logistic Regression, SVM, Random Forest and etc. Following this, I conducted an in-depth study of neural network approaches such as ANN, CNN, RNN. Furthermore, experimented methods of data normalizing, regularizing, augmenting, and hyperparameter tuning to enhance performance. Finally, evaluated transfer learning and few-shot learning with pre-trained models.

Personality Detection (NLP Course Project) [GitHub]

• Introduced novel model architectures for classification and incorporated additional non-text features to enhance performance.
• Implemented sentence generation with language modeling on GPT model.
• Automated the process of crawling, cleaning, and analyzing data, including word count histograms, TF-IDF, RNF, and more.

Visual Question Answering [GitHub]

• Built a custom transformer-based MiniVQA model from scratch. It involved extracting image embeddings, obtaining question embeddings from pretrained BERT embeddings, passing the question embedding through encoder and decoder layers, and generating answers by concatenating them and applying a linear layer at the end.
• Worked on ResNet as a pretrained image feature model and utilized a sentence transformer with Distil-BERT as the base model.

Covid Detection on X-Ray Chest Images (Deep Learning Course Project) [GitHub]

• Initially, data augmentation techniques, including rotation, flipping, and noise addition, were employed. The SqueezeNet model was selected for transfer learning, with the addition of two layers, Conv2d and Adaptive pooling, at the end. During training, the initial layers were frozen, and fine-tuning was conducted on the last two layers. Evaluation incorporated new criteria such as sensitivity and specificity, along with the generation of a confusion matrix and ROC curve.

Computational Intelligence [GitHub 1, 2, 3 ]

• Designed a Kohonen (SOFM) network to recognize and cluster MNIST dataset numbers based on similarity. Additionally, used the Kohonen network to approximate solutions to NP-Hard TSP problems, providing results close to the actual solutions.
• Implemented a Hopfield model which could remove noise from images.
• Implemented a neural network from scratch containing all forward and backward computation as well as custom hyperparameters.

Artificial Intelligence [GitHub]

• This course project consists of CS188 Berekely project. DFS, BFS, A* were implemented in the first phase. In the second phase adversarial methods such as multiagent minimax, expectimax algorithms were implemented. Finally, in the last phase implemented Reinforcement Learning methods such as value function, Q Learning, and Approximate Q learning.

Saku (Software Engineering Course Project) [GitHub]

• Developed a web-based application as a front-end developer using ReactJS. The application primarily functions as an auction platform and includes features such as in-app chat, auction trading, and bid proposal.

Dadmatech [Linkedin]

• Initially, developed a questionnaire web app using Django. Subsequently, retrieved user Twitter data through web crawling, followed by data cleaning using regex commands and data manipulation with pandas. Finally, focused on SQL queries for data extraction.