Data Science

Tool Preferences

Rules

• Never use inplace=True in pandas -- assign to new variable
• Always set random_state or seed for reproducibility
• Use .copy() when creating DataFrame subsets to avoid SettingWithCopyWarning
• Prefer pd.concat() over DataFrame.append()
• Use category dtype for low-cardinality string columns
• Profile data before modeling: df.describe(), df.info(), null counts
• All quantitative claims must be backed by computed results, not assumptions
• Use logging, not print, for production pipelines
• Pin package versions in requirements for reproducibility

Anti-Patterns to Avoid

• Training on test data (always split first)
• Dropping nulls without understanding why they exist
• Using accuracy as sole metric for imbalanced classes
• Scaling/encoding before train-test split (data leakage)
• Correlation does not imply causation -- be explicit about this