Portfolio

Table of Contents

• Data Science & Machine Learning
  • Credit Card Market Segmentation and Cluster Prediction
  • Restuarants Review Rating Classification
• Artificial Intelligence (AI)
  • Facial Expression Recognition System
  • Brain Tumor Detection and Localization
• Machine Learning for NLP
  • Aspect Category and Polarity Classification
  • Next Word Prediction
  • Name Entity Recognition (NER) for Thai Language
  • Business Idea Generator App using Langchain and Large Language Model (LLM)
• Research Paper
  • Enhancing GPT-3.5 for Zero-Shot Thai Intent Classification via Cross-Lingual Prompts, Chain-of-Thought, and Self-Consistency
• Data Analysis and visualization
  • Netflix Top 10 and Financial Data Analysis
  • My Tableau Dashboard
  • Google Spreadsheet / Excel
• Project Repositories
• Data Engineer Workshop

Data Science & Machine Learning Projects

• Credit Card Market Segmentation and Cluster Prediction - Explore Project

  • Python, Exploratory Data Analysis (EDA), Pandas, Clustering, PCA, Logistic Regression
    

  

  

• Restuarants Rating Classification - Explore Project

  • Python, BERT, Logistic Regression, Pandas
    

  

---

Artificial Intelligence (AI) Projects

• Facial Expression Recognition System - Explore Project

  • Python, Exploratory Data Analysis (EDA), Pandas, OpenCV, Image augmentation, Data normalization, Tensorflow, CNNs, RESNET
    

  

  Overview

  • A system that automatically monitors people's emotions and expressions based on facial images
    • The dataset comprises 2000 images with facial key-point annotations and 20,000 facial images, each labeled with facial expression categories.
    • The tasks include detecting facial key points and categorizing each face into one of five emotion categories.
  • Tasks:
    • Perform image visualizations to understand the dataset.
    • Perform image augmentation to increase dataset diversity.
    • Conduct data normalization and prepare training data for model training.
    • Build deep Convolutional Neural Networks (CNNs) and residual neural network (RESNET) models for facial key points detection.
    • Save the trained model for deployment.

• Brain Tumor Detection and Localization - Explore Project

  • Python, Exploratory Data Analysis (EDA), Pandas, scikit-learn, OpenCV, Image Segmentation, Tensorflow, ResNet50, ResUNet
    
  
  Overview

  • Improve the speed and accuracy of brain tumors detection and localization based in MRI scans
    • The data comprises 3929 Brain MRI scans with brain tumor location from https://www.kaggle.com/mateuszbuda/lgg-mri-segmentation
    • The tasks include classification to detect if tumor exists or not and localizing the tumor if exists/li>
  • Tasks:
    • Perform data visualizations to understand the dataset.
    • Training classifier model to detect tumor
    • Train a segmentation ResUNet model to localize tumor if exist

---

Machine Learning for NLP Projects

• Business Idea Generator App (BizGen) using Langchain and Large Language Model (LLM) - Explore Project

  • Python, Langchain, OpenAI, LLM
    

  

• Aspect Category and Polarity Classification - Explore Project

  • Python, Pandas, nltk toolkit, Spacy, Logistic Regression, DAN, CNNs
    

  

  Overview

  • The dataset contains 3156 rows. The text is drawn from restaurant reviews, tokenized using nltk.word_tokenize and non-English alphabet symbols were cleaned out using regular expression.
  • The tasks include categorizing each text review into one of five aspect categories and into one of four sentiments.
  • Tasks :
    • Bag-of-word logistic regression model as a baseline for both sentiment and aspect classification. The features are created from the cleaned text.
    • Perform oversampling by multiplying the number of conflict label data in the training set to increase dataset diversity.
    • Trained both multi-class and multi-label logistic regression models for aspect classification.
    • For multi-label, used a binary logistic regression model to train each aspect model separately, and combine the end result prediction.
    • For Deep Learning Models, tried both pre-trained GloVe 300-dimensional word embeddings from stanford.edu and Word2Vec.
    • Build Deep Averaging Network (DAN) and Convolutional Neural Network (CNN).
    • Tuned Hyperparameters using grid search.

• Next word Prediction - Explore Project

  • Python, Tensorflow, Pandas
    

    Overview

    • Predicting next word based on the first letter
      • The training set is drawn from https://huggingface.co/datasets/gigaword
      • The development set provided for evaluation contains 94,825 rows with 3 columns:
      • 'context' column, 'first letter' column, and 'prediction' column.
      • The 'first letter' column provides the initial letter of the word to be predicted for each context, while the 'prediction' column contains the actual word that is to be generated.
    • Tasks :
      • For trigram model, a counter dictionary is used to count the number of occurrences of each trigram in the training data.
      • The model is trained in small batches, with a batch size of 2048, to accommodate the large size of the training set.
      • Once the model has been trained, the probability of each word is computed based on its frequency in the training data.
      • The trigram model is then used to predict the next word in a given context by selecting the word with the highest probability.
      • For kenlm (pre-trained 5-gram model), the next word in a given context is generated by looping over each word in the model's vocabulary and selecting the word with the highest probability.
      
      

• Name Entitiy Recognition (NER) for Thai Language - Explore Project

  • Python, Scikit-learn, Pandas, pythainlp
    

    Overview

    • Name Entity Recognition (NER) for Thai Language
      • The training and development data for this project were in Thai, and were first tokenized and separated by '|' using the pythainlp library (newmm dictionary).
      • The resulting text was then tagged with entity types, including 'ORG', 'PER', 'MEA', 'LOC', 'TTL', 'DTM', 'NUM', 'DES', 'MISC', 'TRM', and 'BRN', using 'B_' before each tag.
      • Each word and tag were separated by '\t', while sentences were separated by '\n'.
      • To preprocess the training data for the models, each word and tag in the dataset were split and stored in two separate lists: one for token sequences and one for label sequences.
    • Tasks :
      • Implemented Conditional Random Fields (CRF) with only the word, the previous word, and the next word as features as baseline.
      • Added conjunctive features to the model, which took the form of {word i-1 – word i – word i+1}, resembling bigram and trigram features to capture more contextual information about the words.
      • Explored the use of conjunctive part- of-speech (POS) tags as a feature to recognize named entities based on grammatical context, using the pythainlp pos_tag (orchid_ud).
      
      

---

Research Paper

• Enhancing GPT-3.5 for Thai Intent Classification via Cross-Lingual Prompts, Chain-of-Thought, and Self-Consistency - Read Paper

  
  • proposes a method using Large Language Models (LLMs) and cross-lingual techniques.
  • enhance classification performance by prompting GPT 3.5 in English rather than Thai.
  • Our Cross- Lingual Chain-of-Thought Prompt template (XCoT) improves LLM performance, integrating role-assigning and cross-lingual steps, surpassing standard prompts.

---

Projects Repositories

Data Engineer Workshop