Documentation Index
Fetch the complete documentation index at: https://mintlify.com/LauraSilRu/exoplanet-profiler/llms.txt
Use this file to discover all available pages before exploring further.
02_eda.ipynb performs deep exploratory analysis on the 12 candidate features identified in notebook 01. It characterizes distributions, quantifies correlations, audits potential outliers using IQR, and builds the scientific rationale for keeping all 12 variables.
EDA Objectives
- Confirm
pl_nameas the unique, non-null row identifier (zero nulls, zero duplicates across all 910 rows) - Visualize distributions of all 12 features (histograms, box plots)
- Compute and visualize the correlation matrix
- Audit outliers using IQR as a descriptive tool (not for deletion)
- Identify skewed variables that will need log transformation
- Confirm that all 12 variables add distinct information
Key Insights
The EDA reveals several important properties of the TESS dataset that directly inform the preprocessing strategy applied in notebook 03. Kepler correlation. There is a strong positive correlation betweenpl_orbper (orbital period) and pl_orbsmax (semi-major axis). This is expected: Kepler’s third law links orbital period to orbital distance. The redundancy is intentional — PCA is designed to summarize correlated variables rather than discard them.
Right-skewed variables. pl_orbper, pl_orbsmax, pl_bmasse, and pl_insol are highly right-skewed due to the presence of extreme objects (hot Jupiters, distant wide-orbit planets). These variables are candidates for log1p transformation in notebook 03.
Bounded variable. pl_orbeccen is bounded between 0 and 1, so it does not require a log transform regardless of its distribution shape.
Already-logarithmic variable. st_logg represents the base-10 logarithm of stellar surface gravity and is already on a logarithmic scale — no additional transform is needed.
Correlation justifies PCA. The presence of redundancy across orbital and stellar variables confirms that PCA will effectively compress the feature space. However, no variable is dropped at this stage: correlated variables can still carry complementary signal across different PCA components.
Outlier Policy
IQR analysis is used descriptively in this notebook — it quantifies the extent
of extreme values but does not trigger any deletion. In astronomy, an outlier
is often a real, physically interesting object. A massive Hot Jupiter or a
planet on a highly eccentric orbit is not a measurement error; removing it
would distort the true distribution of exoplanet diversity.Mitigation (not deletion) is applied downstream:
log1p transforms compress
the right tail, and RobustScaler (median + IQR) limits the influence of
extremes on distance calculations without destroying any observation.Running the Notebook
Run after notebook 01. No files are written to
data/processed/ — all output
consists of exploratory visualizations (distribution plots, correlation
heatmaps, box plots). The notebook’s deliverable is the documented scientific
rationale for the preprocessing choices applied in notebook 03.