AI Prompts for ChatGPT for Data Science

Write a Python EDA script for a dataset with these columns and types: [PASTE DF.DTYPES OUTPUT OR DESCRIBE COLUMNS]. The script should systematically: (1) check shape, data types, and memory usage, (2) identify missing values with percentages per column, (3) compute descriptive statistics separating numeric and categorical features, (4) identify potential duplicate rows, (5) check cardinality of categorical columns, (6) detect obvious data quality issues (negative values where impossible, dates out of range, impossible category combinations), (7) generate a profile report summary. Use pandas and print clear section headers. Do not plot yet.

Write a Python visualization suite for this dataset: [DESCRIBE DATASET: WHAT IT IS, KEY COLUMNS, WHAT I AM TRYING TO UNDERSTAND]. Use matplotlib and seaborn. Include: (1) distribution plots for all numeric features (histograms + KDE), (2) count plots for all categorical features, (3) a correlation heatmap, (4) box plots showing the target variable distribution across key categorical features, (5) a pairplot for the 5 most correlated features, (6) time series plot if there is a date column. Use a consistent style and add titles and axis labels to every plot. Arrange all plots in a logical narrative order.

I have a dataset with missing values. Here is the missing value summary: [PASTE DF.ISNULL().SUM() OUTPUT]. The target variable is: [DESCRIBE]. For each column with missing data, analyze: (1) is the missing data MCAR, MAR, or MNAR (and why), (2) what imputation strategy is appropriate (mean/median/mode, KNN imputation, model-based, forward fill, indicator variable), (3) the risk of each strategy for downstream modeling. Then write the pandas/sklearn code to implement the recommended imputation strategy. Use ColumnTransformer to keep training and test imputation separate.

Write Python code to detect and investigate outliers in this dataset. Columns of interest: [LIST NUMERIC COLUMNS]. Implement: (1) IQR method with customizable fence multiplier, (2) Z-score method, (3) isolation forest for multivariate outliers, (4) visualizations showing outliers in context (box plots, scatter plots). For each outlier detected, output the row data so I can inspect it. Then ask me to decide: remove, cap (winsorize), or flag with an indicator variable. Include a function that applies my chosen strategy and documents how many rows were affected.

Analyze the relationship between features and the target variable in my dataset. Target: [DESCRIBE TARGET VARIABLE, CLASSIFICATION OR REGRESSION]. Features: [LIST KEY FEATURES]. Write code to: (1) compute correlation (Pearson + Spearman) with the target for numeric features, (2) compute Cramér's V or chi-square for categorical features vs. target, (3) compute point-biserial correlation if target is binary, (4) plot the top 10 most correlated features sorted by absolute correlation, (5) flag any features with near-zero correlation that are candidates for removal, (6) identify any features correlated with each other above 0.85 (collinearity risk).

Write a scikit-learn preprocessing pipeline for a dataset with these characteristics: [DESCRIBE: WHICH COLUMNS ARE NUMERIC, WHICH ARE CATEGORICAL, WHICH NEED SPECIAL TREATMENT, WHAT THE TARGET IS]. The pipeline should use ColumnTransformer and Pipeline to: (1) impute missing values appropriately for each column type, (2) scale numeric features (StandardScaler or RobustScaler if outliers present), (3) encode categorical features (OrdinalEncoder vs OneHotEncoder vs TargetEncoder based on cardinality), (4) create any interaction or polynomial features, (5) handle any datetime columns by extracting meaningful components. Make the pipeline serializable with joblib.

Write Python code to extract useful features from these datetime columns in my dataset: [LIST DATETIME COLUMNS AND DESCRIBE WHAT THEY REPRESENT]. For each datetime column, extract: (1) year, month, day, hour, day of week, week of year where meaningful, (2) cyclical encoding of periodic features (sin/cos transformation for hour, month, day of week), (3) time since a reference date, (4) is_weekend, is_holiday (using a holiday calendar), (5) lag features if this is time series data [DESCRIBE TIME SERIES STRUCTURE IF APPLICABLE], (6) rolling window statistics if applicable. Explain which features are most likely to be predictive for my use case.

Implement target encoding for these high-cardinality categorical columns: [LIST COLUMNS WITH CARDINALITIES]. The target is [DESCRIBE: BINARY/MULTICLASS/CONTINUOUS]. Implement: (1) leave-one-out target encoding to prevent data leakage, (2) smoothing with prior to handle low-count categories, (3) cross-validation-safe implementation that fits on training folds and transforms test folds, (4) handling of unseen categories at inference time, (5) integration into a scikit-learn Pipeline. Explain the data leakage risk of naive target encoding and how this implementation prevents it.

Write Python code to select the most important features from a dataset with [X] features for a [CLASSIFICATION/REGRESSION] problem. Implement: (1) filter methods (correlation, chi-square, mutual information), (2) wrapper method (recursive feature elimination with cross-validation), (3) embedded method (feature importance from a tree model), (4) a consensus voting approach that combines all three methods, (5) a plot showing feature importance from each method. Return the final recommended feature set and explain how to validate that this selection does not hurt model performance.

My classification dataset is imbalanced: [DESCRIBE CLASS DISTRIBUTION, E.G., 95% NEGATIVE, 5% POSITIVE]. The task is to predict [DESCRIBE]. Implement and compare these approaches for handling imbalance: (1) class_weight parameter in the model, (2) SMOTE oversampling with imbalanced-learn, (3) undersampling the majority class, (4) threshold tuning on probabilities after training. For each approach: show the implementation, explain the trade-offs, and show how to evaluate performance using the right metrics (precision-recall curve, F1, ROC-AUC) rather than accuracy. Recommend the best approach for my specific use case.

Help me choose the right model for this problem: Data: [DESCRIBE SHAPE, FEATURE TYPES, MISSING DATA SITUATION]. Target: [CLASSIFICATION/REGRESSION, CLASS DISTRIBUTION]. Constraints: [DESCRIBE: NEED INTERPRETABILITY, INFERENCE LATENCY REQUIREMENT, TRAINING TIME BUDGET, DEPLOYABILITY]. Generate a model selection recommendation covering: (1) which model families are appropriate and why, (2) which to try first and in what order, (3) the baseline model I must beat to justify ML complexity, (4) what metric to optimize (and why not accuracy if classification), (5) any models that are inappropriate for this problem and why.

Write a proper cross-validation setup for this problem: [DESCRIBE: DATASET SIZE, WHETHER DATA IS I.I.D. OR TIME SERIES, ANY GROUP STRUCTURE LIKE MULTIPLE ROWS PER USER, TARGET CLASS BALANCE]. Choose between KFold, StratifiedKFold, GroupKFold, TimeSeriesSplit, or a custom splitter. Implement: (1) the correct CV splitter for my data structure, (2) a cross_val_score or cross_validate call with appropriate scoring metrics, (3) handling of the preprocessing pipeline inside the cross-validation (not outside, which would cause data leakage), (4) a results table with mean and std for each metric, (5) a brief explanation of why this CV strategy is correct for my data.

Write hyperparameter tuning code for a [DESCRIBE MODEL: XGBOOST/LIGHTGBM/RANDOM FOREST/NEURAL NETWORK] for [DESCRIBE PROBLEM]. Implement: (1) an Optuna study with a logical search space for each hyperparameter, (2) pruning via MedianPruner for early stopping of bad trials, (3) the objective function with proper cross-validation inside, (4) visualization of the optimization history and parameter importances, (5) extraction of the best parameters and retraining on the full training set, (6) comparison of default vs. tuned model performance. Explain which hyperparameters matter most for this model family.

Write a comprehensive model evaluation for a [BINARY CLASSIFICATION/MULTICLASS/REGRESSION] model. Implement: (1) the appropriate metrics (for classification: confusion matrix, precision, recall, F1, ROC-AUC, PR-AUC; for regression: RMSE, MAE, R², MAPE), (2) calibration plot if the model outputs probabilities, (3) learning curves (training and validation score vs. training size) to diagnose bias/variance, (4) feature importance plot, (5) error analysis on the worst predictions (cases where the model was most wrong), (6) if classification: threshold analysis to find the optimal classification threshold for my business objective.

Write Python code to interpret and explain predictions from a trained [MODEL TYPE] for [DESCRIBE PROBLEM]. Use SHAP: (1) compute SHAP values for the entire test set, (2) global feature importance (SHAP summary plot), (3) feature interaction effects (SHAP dependence plot for the top 3 features), (4) local explanation for 3 specific predictions: one true positive, one false positive, one false negative, with waterfall plots, (5) a plain-English summary of the 3 most important factors the model uses and what direction they push predictions. Also compare SHAP values to the model's native feature importance.

Help me write a summary of my model's performance for a non-technical stakeholder audience. My model does: [DESCRIBE WHAT IT PREDICTS]. Performance: [PASTE KEY METRICS]. Business context: [DESCRIBE HOW THE MODEL WILL BE USED AND WHAT DECISIONS IT WILL SUPPORT]. Write: (1) a one-paragraph plain-English summary of what the model does, (2) an explanation of the key metrics in business terms (not "precision" but "of the customers we flag as high-risk, X% actually are"), (3) what the model gets wrong and the likely business impact, (4) a comparison to the current process the model replaces. Avoid all ML jargon.

Format these model experiment results into a clean comparison table for a team meeting or report: [PASTE RESULTS FROM MULTIPLE MODEL RUNS OR CONFIGURATIONS]. Include: model name/config, key metrics, training time, inference time if relevant. Highlight the best model in each metric. Write a 200-word "Experiment Conclusions" section below the table covering: which model won and why, what we tried that did not help, the one surprising finding, and the recommended next steps. Keep technical jargon to a minimum for a mixed audience.

Interpret the results of an A/B test for my ML model: [DESCRIBE THE TEST SETUP, WHAT WAS THE CONTROL VS. TREATMENT, WHAT METRICS WERE MEASURED, AND PASTE THE RESULTS]. Help me: (1) confirm statistical significance was properly established (sample size, p-value, confidence intervals), (2) identify any practical significance vs. statistical significance gap, (3) check for any novelty effects or external confounds that could explain the results, (4) calculate the estimated business impact at full rollout, and (5) write a 250-word recommendation memo for the decision to roll out or not.

Write the markdown narrative cells for a Jupyter notebook about this analysis: [DESCRIBE THE ANALYSIS AND PASTE OR DESCRIBE THE CODE AND OUTPUT CELLS]. Each narrative cell should: introduce what the next code cell does and why, explain the output in plain language, highlight the most important insight from that cell, and connect it to the broader analytical story. Write narrative cells for: (1) the introduction and problem statement, (2) data overview, (3) key EDA findings, (4) modeling approach rationale, (5) results interpretation, and (6) conclusions and next steps.

Write a README for a data science project repository: [DESCRIBE THE PROJECT: WHAT PROBLEM IT SOLVES, WHAT DATA IT USES, WHAT MODEL WAS BUILT, WHAT THE KEY RESULTS ARE]. Include: (1) Project Overview (what it does and why it matters), (2) Key Results (the headline metric with context), (3) Project Structure (directory tree with one-line descriptions), (4) Setup Instructions (conda/pip environment, data download steps), (5) Usage (how to run training, evaluation, and inference), (6) Results Summary (model performance table), (7) Limitations and Next Steps. Keep it under 600 words. Every section should help a new team member get oriented quickly.