Katrin-Misel Ponomarjova

Projects

Assessing credit default risk with supervised learning > link

• Data preprocessing and feature engineering of data from the Home Credit Default Risk dataset.
• Evaluating, comparing and tuning the performance of several supervised learning models.
• Models tested : DecisionTreeClassifier, RandomForestClassifier, GradientBoostingClassifier, XGBClassifier, LGBMClassifier, LinearSVC, RidgeClassifier
• Use of oversampling and undersampling techniques, cross validation and curve plotting.

E-commerce customer segmentation with unsupervised learning > link

• Data exploration, analysis, visualization and preprocessing : notebook link
• Feature engineering
• Choosing the amount of clusters with KMeans and determining client cluster
• Testing the temporal stability of the model

Natural Language Processing and Computer Vision for the Yelp reviews database > link

Natural Language Processing

1. Filtering out negative reviews from a sample (5000 reviews)
2. Preprocessing reviews to a format compatible with the NLP model
3. Extraction of topics from negative reviews with Gensim's Latent Dirichlet Allocation
4. Results analysis and visualization

Computer Vision: Image Classification

1. Equalizing the histograms for each photo in the sample (100 photos per label, 500 photos total)
2. Testing ORB for feature extraction
3. Dimensionality reduction and KMeans clustering
4. Using transfer learning with VGG16 for feature extraction
5. Dimensionality reduction and KMeans clustering
6. Visualizing and analyzing results
7. Analyzing some examples of mislabeled photos

A synthesis of the project can be found here : https://katrinmisel.github.io/project_synthesis.html

Comparing deep learning model performances on binary sentiment classification > link

• Exploratory data analysis of Tweet data
• Evaluating two embedding dictionaries: wiki2vec and GloVe
• Evaluating the performance of several models:
• Simple neural network vs. bidirectional LSTM
• GloVe vs. Wiki2Vec embedding
• Tweet preprocessor library vs. simple text prep function vs. no text prep
• Deploying two apps to the cloud:
  Simple Keras sequential model: a GloVe embedding layer, a bidirectional LSTM layer, a dense layer > link
  DistilBERT transformer from Huggingface > link

Image segmentation for autonomous vehicles > link

• Creating a custom data generator with Keras Sequence method
• Image augmentation with the Albumentations library
• Benchmarking multiple architectures, backbones and metrics:
  • Mini U-net as baseline > architecture
  • Architectures tested: U-net, PSPnet, Linknet
  • Backbones tested: VGG16, ResNet34
  • Loss functions: Categorical Focal Dice Loss, Categorical Focal Jaccard Loss
• Performance obtained for best model (U-net with ResNet34 backbone, Categorical Focal Dice Loss, augmented data):
  • MeanIoU = 0.73
  • Loss = 0.20
• Creating a webapp with Flask and Streamlit, deploying baseline > link

Deploying a content recommendation app with Azure > link

• Exploring recommendation algorithm options:
  • Content-Based Filtering with article embeddings
  • Collaborative Filtering, comparing two librairies: Surprise and Implicit
• Deploying a webapp using Azure functions (HTTP triggered) to Streamlit > link

Chatbot, tools:

• Microsoft Bot Framework SDK v4 for Python
• Azure cognitive service LUIS
• Web App with Azure
• Bot Framework Emulator