ASTERLAYSE_hackathon

Table of contents

1. Set up and data frame creation
   • Unzipping
   • Reading data
2. Data Cleaning and Preparation
   • Column Descriptions
   • Imputation
   • Dropping Null values
3. Data Visualization
   • Feature transformation
   • Correlation Visualization
     • Dropping Correlated columns
   • Histogram Plotting
   • Box Plots
4. Modeling
   • train test split of training data
   • linear regression model
   • Gradient boosting regression
   • Hist Gradient Boosting regression
5. model saving
6. Loading actual Test data
   • Test data preparation
   • Prediction

Asteroid Dataset Analysis Report

1. Set up and Data Frame Creation

a. Unzipping

• Description: Unzipping the asteroid dataset.
• Status: Complete

b. Reading Data

• Description: Reading the asteroid dataset.
• Status: Completed

2. Data Cleaning and Preprocessing

Column Description

Column Description	Null values	Column Status	Drop condition
Object Full Name	654128	Dropped	Due to high null values
a (Semi-Major Axis)	2	Used	-
e (Eccentricity)	0	Used	-
i (Inclination)	0	Used	-
om (Longitude Ascending Node)	0	Used	-
w (Argument of Perihelion)	0	Used	-
q (Perihelion Distance)	0	Used	-
ad (Aphelion Distance)	6	Dropped	Correlated with a
per_y (Orbital Period in Years)	1	Dropped	Correlated with a
data_arc (Data Arc Span)	12290	Used	-
Condition Code (Orbit Condition)	703	Used	-
n_obs_used (No of Observations)	0	Used	-
H (Absolute Magnitude)	2100	Used	-
NEO (Near Earth Object)	6	Used	-
PHA (Potentially Hazardous Asteroid)	13089	Used	-
Diameter	561590	Dropped	Due to high null values
Extent	671754	Dropped	Due to high null values
Albedo (Geometric Albedo)	562541	Dropped	Due to high null values
Rot_Per (Rotation Period)	656740	Dropped	Due to high null values
GM (Standard Gravitational Parameter)	671758	Dropped	Due to high null values
BV (Color Index B-V)	670951	Dropped	Due to high null values
UB (Color Index U-B)	670983	Dropped	Due to high null values
IR (Color Index I-R)	671770	Dropped	Due to high null values
Spec_B (Spectral Taxonomic Type SMASSII)	670420	Dropped	Due to high null values
Spec_T (Spectral Taxonomic Type Tholen)	670984	Dropped	Due to high null values
G (Magnitude Slope)	671668	Dropped	Due to high null values
Class (Asteroid Orbit Class)	0	Used	-
n (Mean Motion)	2	Used	-
per (Orbital Period in Days)	6	Dropped	Correlated with a
ma (Mean Anomaly)	6	Used	-
MOID (Earth Minimum Orbit Intersection Distance)	13089	Label	-

Imputation

• In the data imputation phase, KNN (K-Nearest Neighbors) imputation was applied to fill missing values in specific features that had a low number of null values
• Using KNN imputation for selected features with low null values
  • data_arc
  • condition_code
  • H
  • pha

Dropping Null Values

• Dropping columns with high null values: name, diameter, extent, albedo, rot_per, GM, BV, UB, IR, spec_B, spec_T, G

In summary, the data imputation using KNN and the removal of columns with a high number of null values were crucial steps in preparing the dataset for analysis. These actions ensured that the dataset contained relevant and complete information for further exploration and modeling.

3. Data Preparation

Feature Transformation

• Transformed categorical features to numerical counts with LabelEncoder
  • Class
  • Neo
  • Pha
  • Conditional_code

Correlation Visualization

Visualized data with Heatmap and plots b/w

• a vs per
• a vs per_y
• a vs ad

• Dropping Correlated Columns
  • per
  • per_y
  • ad

Histogram Plotting

Box plots

4. Modeling

Train-Test Split of Training Data

• 20% test split
• Training data shape: (526937, 15)
• Test data shape: (131735, 15)

Linear Regression Model

• Linear Regression, a simple yet powerful regression model, was applied to predict a continuous target variable using the training data features.

• Performance Metrics:

  • Mean Squared Error: 0.0013448835921428656
  • Root Mean Squared Error: 0.036672654555443156
  • Mean Absolute Error: 0.020814241770678144
  • R-squared (Coefficient of Determination): 0.9997452537976232

Gradient Boosting Regression

• Gradient Boosting Regression, an ensemble learning technique, was used to predict the target variable, leveraging multiple decision trees for complex relationship modeling.

• Performance Metrics:

  • Mean Squared Error: 0.0006807442439919126
  • Root Mean Squared Error: 0.026091075945462897
  • Mean Absolute Error: 0.014509631039392745
  • R-squared: 0.999871054259298

Hist Gradient Boosting Regression

• Hist Gradient Boosting Regression, optimized for large datasets and missing data handling, was employed to predict the target variable.

• Performance Metrics:

  • Mean Squared Error: 0.05544783520303665
  • Root Mean Squared Error: 0.23547364014478703
  • Mean Absolute Error: 0.030612144613337926
  • R-squared: 0.9894971389862222

5. Model Saving via Pickle

Pickle is a convenient way to save and load machine learning models in Python, it's essential for reducing repeated training time, especially when dealing with untrusted large data sources.

6. Actual Test Data

Upon loading the test dataset, it became evident that, like the training data, it also contained null values. To ensure that the data was in a suitable state for analysis, the following preprocessing steps were taken:

Test Data Preparation

• Dropping Columns with Null Values : Columns in the test dataset that contained a high number of null values were removed. This helped in cleaning the dataset and improved the quality of the data used for predictions.
• Dropping Correlated Columns : As in the training data, columns that were highly correlated with each other were dropped. This step was necessary to mitigate multicollinearity, which can lead to unstable predictions in certain models.

Prediction

In the prediction phase, two machine learning models were utilized to make predictions on the test data. These models were selected based on their suitability for handling different scenarios within the test dataset:

• Gradient Boosting Regressor : This model was applied to the test data for rows that did not contain null values. It's a powerful model for regression tasks and is well-suited for making predictions on data with complete information.
• Hist Gradient Boosting Regressor : For rows with null values, the Hist Gradient Boosting Regressor was employed. This model has the capability to handle missing data effectively and produce reliable predictions even when there are gaps in the input data.

The use of these two models allowed for a comprehensive and robust approach to prediction, ensuring that both complete and incomplete information in the test data could be processed effectively, providing valuable insights and predictions for your analysis.