MAPPING THE SKIES
Have you ever looked up to the sky in a clear night?
Meet the objects
-
STAR:
"A ball of gas burning ________ of miles away"
PUMBA
The Lion King (1994)
-
GALAXY:
A group of _________ of stars.
billions
Please, insert some absurdly huge number here
-
QUASAR:
A galaxy with an extremely luminous core, about _______ regular galaxies, located __________ light-years away.
Please insert some other completely insane number here
billions
What if ?
Earth was the size of a grain of sand (~1mm)
SUN
10 cm
12 m
SOLAR SYSTEM
~1 Km
LIGHT-YEAR
~750 Km
Who is who?
900 LY
9,000,000,000 LY
~100,000,000 LY
Earth to Moon and back
Earth to Pluto and back
Almost a REAL Light Year
five times
Dataset
SLOAN DIGITAL SKY SURVEY
MAP THE ENTIRE UNIVERSE
10,6
13,2
15,4
9,9
Dataset
1%
500,000 objects
5 light features
Pipeline Model
Logistic Regression
Stratified Train Test Split
F1 Macro
Training 10 models with default settings
10,000 objects
Test (2%)
EDA
Data Cleaning
Pre-
processing
Transform Columns
Standard Scaler
Stochastic Gradient Descent
K-Nearest Neighbors
Naive-Bayes
Decision Tree
ADA Boost
Bagging Classifier
Extra Trees
Gradient Boost
Random Forest
Fine-tuning
3 best models
Final
Training
Model
Random Forest
F1 Macro = 0.99
97.27% QUASARS
Classifies correctly
99.57% GALAXIES
99.97% STARS
The secret?
Redshift: 52% feature importance
Conclusions
-
Universe is BIG
-
Light from objects can tell us a lot, even from extreme distances
-
Model is accurate, but always can be improved
Thank
You!!
What is a Light-year?
The distance that light travels in a year
9,461,000,000,000 KM
Pipeline Model
- Logistic Regression
- Stochastic Gradient Descent
- K-Nearest Neighbors
- Naive-Bayes
- Decision Tree
- Stratified Train Test Split: 10,000 objects to Test (2%)
- Exploratory Data Analysis
- Data Cleaning
- Preprocessing: Transform columns, Standard Scaling
- Training 10 models with default settings
- Fine-tuning 3 chosen models
- Final training
F1 Macro
- ADA Boost
- Bagging Classifier
- Extra Trees
- Gradient Boost
- Random Forest