探索性数据分析与优质预测变量选择：非预处理场景下特征选择及双变量分析作用问询

Great questions—EDA is such a foundational step that often gets overlooked for its role in feature selection, so let's break this down clearly.

At its core, EDA is all about getting to know your data before you start building models. When it comes to picking top-tier predictor variables, it acts as your "data detective" in a few key ways:

• Filters out irrelevant variables upfront: EDA helps you quickly spot variables that have no logical or statistical link to your target. For example, if you're predicting customer churn, a variable like "customer's favorite color" (unless tied to product preferences) is obviously useless—you can cross it off your list immediately instead of wasting time on it later.
• Flags low-information or noisy variables: Using summary statistics (like df.describe() in Python) or visualizations (histograms, boxplots), you can identify variables with near-zero variance (e.g., a "membership tier" where 98% of users are in the basic tier) or variables with so many missing values they can't contribute meaningfully. These are dead weight for your model.
• Uncovers hidden predictive signals: EDA reveals patterns that hint at a variable's potential. For instance, a skewed "total purchase amount" variable might not look impressive at first glance, but a log transformation could uncover a strong linear relationship with your target (like customer lifetime value). Outliers might also indicate subgroups where a variable has extra predictive power.

While EDA is critical for preprocessing (cleaning missing values, fixing outliers), it does much more to guide feature selection. Here's the breakdown:

First, even univariate analysis (looking at single variables in isolation) helps:

• You can drop categorical variables with extreme class imbalance (e.g., a "promotion used" flag where only 1% of users said yes) since they can't explain meaningful variation in the target.
• Numeric variables with a single value (zero variance) get cut immediately—they offer no predictive value whatsoever.

But the real workhorse for feature selection here is bivariate analysis: examining the direct relationship between each predictor and your target variable. This is completely feasible (and recommended!) for most datasets, and it’s invaluable for narrowing down your features:

For regression problems (numeric target): #

• Use scatter plots paired with correlation coefficients (Pearson for linear relationships, Spearman for monotonic non-linear ones) to measure how tightly a predictor is linked to the target. A high absolute correlation (e.g., 0.7 or -0.6) means the variable is a strong candidate to keep. A correlation near 0? It’s probably not worth including.
• Example: Predicting house prices—scatter plots of "square footage" vs. "price" will show a clear positive correlation, making it a top predictor. A variable like "number of street lights nearby" might show no correlation at all, so you can safely drop it.

For classification problems (categorical target): #

• Use boxplots to compare the distribution of a numeric predictor across target classes (e.g., "debt-to-income ratio" vs. "loan default status"). If the distributions look drastically different, that variable has predictive power. For categorical predictors, use bar charts to check if class proportions of the target vary across predictor categories (e.g., "customer segment" vs. "churn rate").
• You can back this up with statistical tests: chi-squared tests for categorical-categorical pairs, or t-tests/ANOVA for numeric-categorical pairs, to quantify the strength of the association.
• Example: Predicting email spam—bar charts might show that emails with "free" in the subject line have a 70% spam rate, while those without have a 5% rate. That makes "subject line contains 'free'" a key predictor.

Why this works for feature selection: #

• It eliminates redundant or irrelevant variables, simplifying your model and reducing the risk of overfitting.
• It highlights variables that need transformation (e.g., a non-linear relationship that becomes linear after log scaling) to unlock their full predictive potential.
• It keeps you grounded in data-driven decisions instead of relying on assumptions about which variables "should" matter.

Finally, multivariate EDA (looking at relationships between multiple predictors) helps too: correlation heatmaps can spot multicollinearity (e.g., "square footage" and "number of bedrooms" being highly correlated). In these cases, you can pick one variable instead of both to avoid model instability.

内容的提问来源于stack exchange，提问作者Gale