Skip to content

README.md

Latest commit

 

History

History
55 lines (41 loc) · 3.64 KB

README.md

File metadata and controls

55 lines (41 loc) · 3.64 KB
Raw

Predictive Modeling Pipelines using Tidymodels in R

This repository contains two comprehensive predictive modeling tasks developed using the modern tidymodels framework in R. The projects demonstrate best practices in machine learning, including stratified data splitting, robust feature engineering, ensemble modeling, parallel processing, and automated hyperparameter tuning.

Repository Structure

  • validation task/
    • .gitignore
    • README.md
    • 01_biomass_walidacja_modeli.qmd -> Task 1: Validation and Model Comparison
    • 02_hotel_rates_tuning.qmd -> Task 2: Feature Engineering and Simulated Annealing

Project Overview

Task 1: Biomass High Heating Value (HHV) Prediction

Objective: Compare the performance of four distinct machine learning architectures to predict the High Heating Value (HHV) of biomass samples based on their elemental composition.

  • Dataset: biomass (from the modeldata package).
  • Key Engineering Steps:
    • Removal of high-cardinality structural identifiers (sample and dataset) to eliminate target leakage.
    • Stratified data splitting (80/20 train/test ratio) to maintain uniform distribution across segments.
    • Custom Recipe Strategies:
      • Linear Regression: Addressed skewed features using a Yeo-Johnson transformation followed by standard normalization.
      • Support Vector Machines (SVM): Applied Principal Component Analysis (PCA) retaining 95% of total variance to resolve predictor collinearity.
      • Tree-Based Models (MARS & Random Forest): Modeled using robust scale-invariant raw representations.
  • Validation & Statistical Rigor:
    • Evaluated models across 5-fold cross-validation with 2 repeats.
    • Conducted a Paired t-test ($p = 0.025$) comparing the top-performing architectures (Random Forest vs. MARS) to statistically justify the final model selection.

Task 2: Advanced Hotel Room Rates Forecasting & Optimization

Objective: Predict average daily hotel room prices (avg_price_per_room) while managing complex temporal indicators and sparse categorical groupings.

  • Dataset: hotel_rates (from the modeldata package), downsampled to 1,500 samples for computational efficiency during global heuristic searches.
  • Advanced Feature Engineering:
    • Deconstructed date-based entries (arrival_date) into explicit numeric and factor vectors (Day of Week, Month, Year).
    • Grouped rare categorical levels below a 2% threshold into a unified "Other" bin via step_other to prevent sparse matrix bloat.
    • Engineered a new domain-specific interaction term: total_nights (combining weekday and weekend night stays).
  • Screening and Heuristic Tuning:
    • Screening Phase: Evaluated Elastic Net, Random Forest, and XGBoost baselines over shared folds using a workflow_set.
    • Global Optimization: Selected the best architecture (XGBoost) and optimized its hyperparameters (trees, tree_depth, and learn_rate) utilizing a Simulated Annealing algorithm (finetune::tune_sim_anneal).
    • Performance Acceleration: Enabled asynchronous multi-core parallel processing (doParallel) to distribute resampling routines across available CPU threads.
  • Overfitting Assessment: Discussed the generalization gaps by closely mapping cross-validation metrics against final performance metrics on the unseen test holdout set.

Dependencies & Setup

Ensure you have R and the required library stack installed before attempting to render the Quarto documents:

install.packages(c("tidymodels", "tidyverse", "finetune", "doParallel", "vip", "earth", "kernlab", "ranger", "xgboost", "glmnet"))