Data Science Portfolio

This repository contains portfolio of my data science projects presented in the form of Jupyter Notebooks.

Contents

Hotel images object detection

Problem type: Computer Vision. Supervised Object detection/Image classification problem.

Goal of the project is to predict hotel amenities (for example: swimming pool, desk, kitchen, coffeemaker, TV, bathtub, balcony, garden) based on the hotel photos. Images processing is handled the best by deep learning technologies. In this project following pre-trained models were tested on real hotel images:

• object detection model YOLOv3 (with ImageAI library);
• image classification model ResNet50 (with keras.applications module);
• object detection model YOLOv5 (using Ultralytics git repository).

Custom object detection model was trained using Open Images Dataset v4 to detect class Swimming pool. Following results were achieved on the test set:

• Precision: 96.3%;
• Recall: 90.8%;
• mAP@0.5: 94.7%;
• mAP@0.05:0.95: 77.4%.

Custom model was able to detect all swimming pools on the sample of real hotel photos. In the future it’s possible to train custom model to detect all necessary hotel amenities.

Challenges: Researching of ready-to-use models which contain all needed classes; setting-up environment for pre-trained models on Windows 10; finding annotated data for custom training; need to annotate real examples of hotel photos for model evaluation.

Used BMW cars price prediction

Problem type: Supervised regression problem.

The goals of this project are:

• explore what factors affect a sale price of used BMW cars and which characteristics are the most important to determine a car value;
• build a prototype of model which predicts price of used BMW cars.

Following results were achieved with CatBoostRegressor on the testing set:

• explained_variance: 0.9631,
• mean_absolute_error: 1435.5959,
• root_mean_squared_error: 2202.7952,
• mean_squared_log_error: 0.008453.

Then prototype of price predictor for used BMW cars was implemented and applied to predict cars price on real examples of cars deals from autotrader.co.uk with following results:

• explained_variance: 0.9341,
• mean_absolute_error: 628.75,
• root_mean_squared_error: 961.897,
• mean_squared_log_error: 0.005877.

Challenges: Data cleaning; detailed exploration of relationship of independent and target variables and feature engendering; testing of 12 algorithms; hyper-parameter tuning of the best performing algorithm; exploring feature selection.

Smartphone reviews sentiment prediction

Problem type: NLP. Supervised text classification problem.

In this project custom web-scraping script was implemented to scrape smartphone review from e-commerce marketplace rozetka.com. The goal of this project is to implement predictive model to classify reviews without rating by sentiment.

For this NLP problem BERT model was applied. Achieved results on the testing set:

• Accuracy: 98%.
• Area Under the Receiver Operating Characteristic Curve: 0.97, so there is a 97.7% chance that the model will be able to distinguish between positive and negative class for reviews.
• Predictions for positive reviews are: almost 97% of all predicted reviews are actually positive and 94% of all positive reviews were detected correctly.
• Predictions for negative reviews are: 98% of all predicted reviews are actually negative and 99% of all negative reviews were detected correctly.

Challenges: Low quality of reviews text (grammatical mistakes, Ukrainian/Russian languages and local non-official dialects, not clear and confusing text); reviews contain different types of content: opinion about product, feedback about shop service, questions; class imbalance in target variable.

Hotels rates change prediction

Problem type: Supervised classification problem with time-series data.

The goal of this project is to predict if at given date hotel changes price for a room (for any stay date of 365 days ahead) so that number of rate shopping iterations can be reduced. It’s important to predict all changes in rates (recall metric) and in the same time we need to measure precision, as low precision will increase updates iterations.

The version of the CatBoostClassifier model with increased weight of the positive class (to capture more rate changes) provided following results:

• Lisbon rooms: 90.77% of positive class observations are predicted correctly with 90.77% precision,
• London rooms: 93.97% of positive class observations are predicted correctly with 93.52% precision,
• Barcelona rooms: 98.38% of positive class observations are predicted correctly with 70.27% precision.

Challenges: Big amount of the initial data (6.77 millions of records); a lot of noise in the data; necessity to achieve high recall score; project goal is to reduce number of rate shopping iterations and when achieved, as the result we will have less new data gathered on rate updates history, which we will need for model weights updating.