📌 About This Repository

This repository contains a collection of laboratory works completed during my university studies.
Each project demonstrates practical application of data analysis techniques, machine learning models, and business intelligence approaches.

Key skills demonstrated:

• Data cleaning & preprocessing (pandas, numpy)
• Exploratory Data Analysis (EDA) with visualizations (matplotlib, seaborn)
• Feature engineering & dimensionality reduction
• Supervised learning (regression, classification)
• NLP & text vectorization (TF-IDF)
• Model evaluation & comparison
• Business insight generation

---

📁 Projects Overview

1. Pharmaceutical Sales Forecasting (Lab 4)

File: Lab 4.ipynb

Task: Predict sales volume of drug categories based on seasonality.

Approach:

• Merged data from 4 sources (sales, products, categories, stores)
• Created seasonal feature from dates
• Built ElasticNet model with cross-validation → improved with Poisson regression

Results:

Metric	ElasticNet	Poisson	Improvement
SMAPE	43.9%	16.5%	-62%

Key insight: Poison regression significantly outperforms ElasticNet for count-based prediction tasks.

Skills: Data merging, EDA, time-based feature engineering, regression, model comparison.

---

2. Advertising Effectiveness Analysis (Lab 9)

File: Lab 9.ipynb

Task: Determine which advertising channels (TV, radio, newspaper) most influence product sales.

Approach:

• Polynomial regression (up to degree 4) to capture non-linear interactions
• Correlation analysis & feature interaction detection

Results:

• TV and radio are primary sales drivers
• Newspaper advertising had near-zero impact
• Polynomial degree 4 gave best predictive quality

Key insight: TV and radio have a synergistic effect — using both together produces additional lift.

Skills: Polynomial features, interaction effects, regression, visualization.

---

3. Author Attribution from Poetry Texts (Lab 5)

File: Lab 5.ipynb

Task: Classify poems by author based on text content.

Approach:

• Text vectorization (TF-IDF)
• Compared 4 models: Naive Bayes, Logistic Regression, SVM (RBF), Linear SVM

Results:

Model	Accuracy
Logistic Regression	81%
Linear SVM	80%
SVM (RBF)	77%
Naive Bayes	69%

Key insight: Logistic regression performed best for this multi-class text classification task.

Skills: NLP, TF-IDF, multi-class classification, model comparison.

---

4. Customer Churn Analysis & Revenue Calculation (Lab 7)

File: Lab 7.ipynb

Task: Analyze telecom customer behavior, calculate revenue per tariff plan, and identify churn patterns.

Approach:

• Merged user, call, message, and internet session data
• Calculated actual revenue per user based on tariff overages
• Visualized spending distributions by tariff type

Results:

• Users on cheaper tariff (smart) paid higher average bills than premium tariff users
• smart tariff contributed ~3.2x more total revenue than ultra
• Only 7.6% of customers churned during the period

Key insight: The cheaper tariff generates more revenue due to overage charges — a potential pricing optimization opportunity.

Skills: Data merging, revenue modeling, overage calculation, business metrics.

---

🛠️ Technologies Used

Category	Tools
Languages	Python
Data Manipulation	pandas, numpy
Visualization	matplotlib, seaborn
Machine Learning	scikit-learn
NLP	TF-IDF vectorization
Environment	Jupyter Notebook

---