EarthquakeDamage_B.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Introduction"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Within the United States alone, earthquakes destroy nearly $4.4B of economic value yearly. Our team will be delving into the 2015 Nepal Earthquake Open Data Portal. This data, which was collected using mobile devices following the Gorkha Earthquake of April 2015, details the level of destruction brought upon more than 200,000 buildings in the area. By utilizing various features reported by Nepali citizens (such as building size, purpose, and construction material), we will construct a machine learning classifier capable of determining the extent to which a building would be damaged. We will also run unsupervised clustering algorithms to search for features that best represent the damage classifications. This project, combined with technologies such as the CV-based city scanners following earthquakes (Ji, 2018) will ultimately provide a better understanding of susceptibility to earthquake-induced damage, valuable information that can be leveraged by city planners."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Data Exploration"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Write about initial filtered data to figure out approaches, results that tell us what to do"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Expected Results"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Using supervised machine learning algorithms we hope to identify which factors affect the level of damage to a building from an earthquake. We’ll compare each of the results by micro averaged F1 score, which will balance precision and recall modified to gauge accuracy for classification into 3 categories (DrivenData). We can also use dimensionality reduction to reduce the number of features from 38 to a more manageable amount by seeing which features are correlated to each other."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Methods"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The data has been collected for us, but we plan to spend a significant amount of time preparing (cleaning, encoding, etc) the data for the model. We will also perform meaningful visualizations to better understand the relationships between our features. The classifier will be trained using supervised learning algorithms such as Support Vector Regressors (Asim, 2018), Logistic Regression, and Hybrid Neural Networks (Asim, 2018) and unsupervised learning algorithms such as KMeans clustering. The hyperparameters will be tuned using the GridSearchCV process. The model performances will be compared using our test set and the one that ​generalizes t​ he best will be chosen.\n",
    "\n",
    "- Multiple Logistic regression\n",
    "- Support vector regression\n",
    "- Kmeans\n",
    "- DBSCAN\n",
    "- PCA\n",
    "- LDA\n",
    "- Cross validation\n",
    "- Artificial Neural Networks\n",
    "- Hybrid Neural Network\n",
    "- Decision Tree\n",
    "https://www.datasciencecentral.com/profiles/blogs/decision-tree-vs-random-forest-vs-boosted-trees-explained\n",
    "- Random Forest Decision Tree\n",
    "- XGBoost Decision Tree\n",
    " (add plots under section option)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Decision Trees"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "There were 4 different types of decision trees that were evaluated including normal decision trees, random forest, extreme gradient boosting, and lite gradient boosting. Using the one-hot encoded data to convert categorical data into numerical data and 5-fold cross validation, a small version of each type of tree was created to compare their effectiveness by the F1 score metric. \n",
    "\n",
    "<img src=\"source_images/baseline_tree_comp.png\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The results from the initial F1 score comparison demonstrate that random forest, extreme gradient boosting, and lite gradient boosting are more promising than the normal decision tree. The normal decision tree had already started overfitting the data since its train accuracy is significantly higher than its validation accuracy. Due to this overfitting, normal decision trees were not explored farther."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Random Forest\n",
    "\n",
    "Random forest decision trees reduce the variance in decision trees since they specify random feature subsets and build and combine many small trees. To make a prediction, random forest trees will either take the prediciton with the highest probability or average the predictions made by its smaller trees. By increasing the number of trees in the forest, both the training and validation accuracy increase due to the smaller tree averages converging on the true trends in the data. This is the idea of ensemble learning which is revisited later in the decision tree exploration."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Extreme Gradient Boosting\n",
    "\n",
    "Extreme gradient boosting also is composed of a set of decision trees but is built by training a tree to add to the forest. Instead of using an average or max of the probabilities from its trees while making predictions, extreme gradient boosting combines the tree prediciton results sequentially (Glen). Increasing the number of trees in the forest also improves the accuracy."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Lite Gradient Boosting\n",
    "\n",
    "Lite gradient boosting is also a gradient boosting framework, however it builds its trees by adding to leaves instead of adding levels.\n",
    "\n",
    "<img src=\"source_images/leafwisegrowth.png\">\n",
    "(Mandot)\n",
    "<img src=\"source_images/levelwisegrowth.png\">\n",
    "(Mandot)\n",
    "\n",
    "This architecture is highly dependent on the number of leaves it expands. More leaves allow it to fit data better, but also can cause overfitting. This architecture can also be used with ensemble learning, and adding more trees will increase the accuracy. The results from the baseline accuracy tree test indicate that the tree should be made to fit the training data better since the train and validation accuracies were very similar. A comparison of the F1 score validation accuracies with varying the number of leaves is shown below.\n",
    "\n",
    "<img src=\"source_images/num_leaves_analysis.png\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The comparison of number of leaves above reveals that the optimal number of trees is 80. While evaluating the optimal number of trees, the F1 score constantly increases with diminishing reward. However, as the number of trees increases, the training time increases. 3,000 was the final number of trees selected."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The overall results from tuning are shown in the plot below.\n",
    "\n",
    "<img src=\"source_images/tuned_tree_comp.png\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This plot shows that the lite gradient boosting was the best technique at the end of the tuning done in this analysis."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Results"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Synthesize information from Methods"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Discussion"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This project can benefit architects, engineers, and city planners by using the classification model to extrapolate and predict types of buildings that are likely to suffer from earthquake damage. Buildings with attributes similar to those that were more damaged can be reinforced. Both the visualization and classification models can be used in conjunction with earthquake prediction research (Rouet-Leduc, 2017) to provide advance humanitarian aid so buildings can be reinforced to take significantly less damage.\n",
    "\n",
    "Future plans, etc."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# References"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Add references"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Glen, Stephanie. “Decision Tree vs Random Forest vs Gradient Boosting Machines: Explained Simply.” Data Science Central, 28 July 2019, www.datasciencecentral.com/profiles/blogs/decision-tree-vs-random-forest-vs-boosted-trees-explained.\n",
    "\n",
    "Mandot, Pushkar. “What Is LightGBM, How to Implement It? How to Fine Tune the Parameters?” Medium, Medium, 1 Dec. 2018, medium.com/@pushkarmandot/https-medium-com-pushkarmandot-what-is-lightgbm-how-to-implement-it-how-to-fine-tune-the-parameters-60347819b7fc."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}