Data science beginner projects

Description

Study projects developed during data science courses.

module_0

The function guesses the number and prints the number of attempts.

module_1

Studying the provided data using pandas.

module_2

EDA, prepare the data for the machine learning

• Filter outliers
• Perform correlation analysis in quantitative data
• Perform analysis of the nominative variables
• Select columns for the machine learning step.

module_3

Predict tripadvisor restaurant rating.

• Data cleaning
• Filling NA
• Outlier removing
• Feature Engineering
• EDA
• Using ML first time with default parameters
  First whole data preprocessing with eda and feature engineering.

module_4

Bank score prediction project

• Data cleaning
• Filling NA
• Outlier removing
• Feature Engineering
• EDA
• ML
• Naive model
• PCA, SVD to reduce the matrix size
• Hyperparameter tuning

module_8

Predict car classes from the pictures using deep learning

• 6 types of augmentation
• Different sizes of images starting from 512 to 224
• Different number of epochs
• Different batch sizes
• All model types that are presented in tf.keras.applications
• Fine-tuning and transfer learning
• LR were optimized using ReduceLROnPlateau
• Different optimizers
• Batch Normalization
• Different callback Keras functions
• TTA
• Different head architecture

new_module_2

Analysis of vacancies from HeadHunter using SQL query in jupyter notebook

diplom

Property price prediction
The data have a lot of outliers, mistakes, input errors, slang abbreviations, that's why the project was split into 2 parts data_cleaning.ipynb and eda_ml.ipynb

• Data cleaning
• Data Enrichment
• EDA
• Feature Engineering
• ML
• Outlier removal using different models: IsolationForest, EllipticEnvelope, LocalOutlierFactor
• Feature selection using different methods: RFE, SelectFromModel, FeatureImportance
• Testing of linear models. Baseline.
• Testing of 5 different advanced models: Random Forest, CatBoost, Gradient Boosting, XGBoost, LightGBM. Bagging and stacking have also been tested.
• Hyperparameter tuning