Introduction
Our project aims to revolutionize the way Myers-Briggs Type Indicator (MBTI) assessments are conducted. Traditional MBTI assessments, with their lengthy questionnaires, can be time-consuming and daunting for users.
By training the model on our dataset containing various users’ post sentences along with their corresponding MBTI types, the aim is to create a system that can accurately classify posts text inputs into one of the 16 MBTI personality types. Our solution would leverage advanced Transformer-based natural language processing (NLP) technology to analyze and interpret users' textual responses swiftly and accurately. By integrating this cutting-edge AI approach, we propose a streamlined, efficient, and user-friendly platform for personality assessment. Nowadays, MBTI personality test definitely is a trending topic. Motivated by the growing interest in understanding how language reflects personality traits and the potential applications in various fields such as social media analysis, marketing, and personalized recommendation systems, this project aims to explore the potential application of MBTI by analyzing user’s MBTI based on their social media posts. Thus, the decision to undertake this project was made. This project is a multi-classification problem. Here, our target variable includes 16 classes each representing one specific MBTI personality. Later on, in our preprocessing and modeling part, we would also consider transforming our problem into 4 binary classification problems where we would perform binary classification on each key dichotomies of MBTI personality code.
Challenges
The hardest part of the project so far has been . Whether it's issues with data collection, preprocessing complexities, or difficulties in model implementation, elaborate on the specific aspects that have been particularly challenging and why.
Hyperparameter tuning has been the most challenging aspect of our project so far. We have set up the initial frameworks for our models, which include a CNN and a Transformer, each designed to perform one of four binary classification tasks. These tasks involve predicting the dichotomy keys of the MBTI personality code. As we iterate through various configurations, including trying different types and numbers of layers, tuning learning rates for optimizers, and consider introducing regularizations based on validation scores, we find that adjustments to the hyperparameters yield only modest improvements in the metrics. Additionally, we frequently face issues with overfitting.
Insights
At this stage, we have observed some noteworthy results such as the CNN model’s overall accuracy has improved from 50% to 70%. We have also solved some of the overfitting issues observed through low validation scores. However, we have realized that the accuracy is closely related to the imbalances of the target variable, which indicates that further adjustment is needed. To address the challenges mentioned in the above section, we have revisited our entire process, from exploratory data analysis (EDA) and preprocessing to model construction. We are now focusing on addressing the imbalanced nature of our dataset during the preprocessing phase to see if this adjustment enhances model performance.
At this stage, the transformer is improving quite slowly due to several reasons. In our transformer model, after training 3 epochs, we reach an accuracy of 66% on the training set but only 57% on the validation set. Also, one current barrier is that the training cost using GPU is high and takes a lot of time and memory, which makes further tunning of our model even harder. We are now still working on optimizing our model structure and see if a more efficient method is available
Plan
We are on track with our project timeline. We have already implemented three approaches for our project: MLP as the baseline model, a CNN model, and a Transformer-based model. Moving forward, we need to dedicate more time to improving model accuracy and model interpreting. Currently, the result for the Transformer model is not as good as expected so we hope to modify the model architecture and fine-tune it to further improve the results. We are considering adding additional preprocessing methods to help us achieve our goals. After obtaining satisfying results, we plan to apply more metrics to explore the performance of our model.
Log in or sign up for Devpost to join the conversation.