senior-data-scientist

name	senior-data-scientist
description	Use when the user needs ML pipelines, statistical analysis, data preprocessing, feature engineering, model selection, experiment tracking, or data visualization. Triggers: dataset exploration, model training, feature engineering, hyperparameter tuning, experiment tracking setup, statistical hypothesis testing, visualization creation.

Senior Data Scientist

Overview

Build end-to-end data science workflows from data exploration through model deployment. This skill covers data preprocessing, feature engineering, model selection, hyperparameter tuning, cross-validation, experiment tracking with MLflow/W&B, statistical testing, visualization with matplotlib/seaborn/plotly, and Jupyter notebook best practices.

Announce at start: "I'm using the senior-data-scientist skill for data science workflow."

---

Phase 1: Data Understanding

Goal: Profile the dataset and establish a baseline before any modeling.

Actions

1. Load and profile the dataset (shape, types, distributions)
2. Identify missing values, outliers, and data quality issues
3. Perform exploratory data analysis (EDA)
4. Define the target variable and success metrics
5. Establish baseline performance

Baseline Models (Always Start Here)

Task	Baseline Model	Why
Classification	Majority class classifier	Lower bound for accuracy
Classification	Logistic regression	Simple, interpretable
Regression	Mean predictor	Lower bound for RMSE
Regression	Linear regression	Simple, interpretable
Time series	Naive forecast (previous value)	Lower bound for MAE
Time series	Seasonal naive	Captures basic seasonality

STOP — Do NOT proceed to Phase 2 until:

• Dataset is profiled (shape, types, distributions)
• Missing values and outliers are documented
• Target variable is defined
• Success metrics are chosen
• Baseline performance is established

---

Phase 2: Feature Engineering

Goal: Transform raw data into features that improve model performance.

Actions

1. Handle missing values (imputation strategy)
2. Encode categorical variables
3. Scale/normalize numerical features
4. Create derived features
5. Feature selection (remove redundant/irrelevant)

Missing Value Strategy Decision Table

Strategy	When to Use	Implementation
Drop rows	< 5% missing, MCAR	df.dropna()
Mean/Median	Numerical, no outliers	SimpleImputer(strategy='median')
Mode	Categorical	SimpleImputer(strategy='most_frequent')
KNN Imputer	Structured missing patterns	KNNImputer(n_neighbors=5)
Iterative	Complex relationships	IterativeImputer()
Flag + Impute	Missingness is informative	Add is_missing column + impute

Categorical Encoding Decision Table

Method	When	Cardinality
One-Hot	Nominal, low cardinality	< 10 categories
Label/Ordinal	Ordinal features	Any
Target Encoding	High cardinality nominal	> 10 categories
Frequency Encoding	When frequency matters	Any
Binary Encoding	Very high cardinality	> 50 categories

Scaling Decision Table

Scaler	When	Robust to Outliers?
StandardScaler	Default choice (mean=0, std=1)	No
RobustScaler	Outliers present (median/IQR)	Yes
MinMaxScaler	Neural networks, distance-based [0,1]	No

Feature Types and Engineering

Feature Type	Techniques
Numerical	Log transform, polynomial, binning, interactions (A*B, A/B)
Temporal	Hour, day-of-week, is_weekend, time_since_event, cyclical (sin/cos), lags
Text	TF-IDF, word count, sentiment scores, named entities, embeddings
Categorical	Encoding (above), interaction with numerical features

Feature Selection Decision Table

Method	Type	Use When
Correlation matrix	Filter	Initial exploration
Mutual information	Filter	Non-linear relationships
Recursive Feature Elimination	Wrapper	Model-specific selection
L1 Regularization	Embedded	Linear models
Feature importance	Embedded	Tree-based models
Permutation importance	Model-agnostic	Final validation

STOP — Do NOT proceed to Phase 3 until:

• Missing values are handled with justified strategy
• Categorical variables are encoded appropriately
• Numerical features are scaled
• Feature engineering is done BEFORE train/test split on training data only
• Feature selection has reduced dimensionality if needed

---

Phase 3: Modeling

Goal: Select, train, and evaluate candidate models.

Actions

1. Select candidate algorithms
2. Set up cross-validation strategy
3. Train and evaluate candidates
4. Hyperparameter tuning
5. Final model selection and evaluation

Algorithm Decision Table

Data Characteristics	Try First	Also Consider
Tabular, < 10K rows	Random Forest, XGBoost	Logistic/Linear Regression
Tabular, > 10K rows	XGBoost, LightGBM	CatBoost, Neural Network
High dimensionality	Lasso/Ridge, SVM	Random Forest with selection
Time series	Prophet, ARIMA	LSTM, XGBoost with lag features
Text classification	Fine-tuned transformer	TF-IDF + Logistic Regression
Image classification	Pre-trained CNN (ResNet, EfficientNet)	Vision Transformer
Regression	XGBoost, Random Forest	Linear Regression, Neural Network
Anomaly detection	Isolation Forest	LOF, Autoencoder

Cross-Validation Strategy Decision Table

Strategy	When	Code
K-Fold (k=5)	Default, balanced data	KFold(n_splits=5)
Stratified K-Fold	Classification, imbalanced	StratifiedKFold(n_splits=5)
Time Series Split	Temporal data	TimeSeriesSplit(n_splits=5)
Group K-Fold	Grouped observations	GroupKFold(n_splits=5)
Leave-One-Out	Very small datasets	LeaveOneOut()

Evaluation Metrics Decision Table

Task	Primary Metric	Secondary Metrics
Binary Classification	AUC-ROC	F1, Precision, Recall, AP
Multiclass	Macro F1	Accuracy, Confusion Matrix
Regression	RMSE	MAE, R-squared, MAPE
Ranking	NDCG	MAP, MRR
Anomaly Detection	F1, AP	Precision@K, Recall@K

Hyperparameter Tuning Decision Table

Method	Compute Budget	Search Space	Implementation
Grid Search	Low (< 100 combos)	Small, known ranges	GridSearchCV
Random Search	Medium	Large, uncertain	RandomizedSearchCV
Bayesian (Optuna)	Any	Large, expensive	optuna.create_study()
Successive Halving	Large	Many candidates	HalvingRandomSearchCV

Common Hyperparameters (XGBoost/LightGBM)

param_space = {
    'n_estimators': [100, 300, 500, 1000],
    'max_depth': [3, 5, 7, 9],
    'learning_rate': [0.01, 0.05, 0.1],
    'subsample': [0.7, 0.8, 0.9],
    'colsample_bytree': [0.7, 0.8, 0.9],
    'min_child_weight': [1, 3, 5],
}

STOP — Do NOT proceed to Phase 4 until:

• At least 2 candidate models are evaluated
• Cross-validation is used (not just train/test split)
• Results beat the baseline from Phase 1
• Best model is selected with justification
• Overfitting is checked (train vs validation gap)

---

Phase 4: Deployment

Goal: Serialize, serve, and monitor the model in production.

Actions

1. Serialize model and preprocessing pipeline
2. Create prediction API or batch pipeline
3. Set up monitoring for data drift and model degradation
4. Document model card (inputs, outputs, limitations, biases)

STOP — Deployment complete when:

• Model is serialized with preprocessing pipeline
• Prediction API or batch pipeline works end-to-end
• Monitoring is configured for data drift
• Model card is documented

---

Experiment Tracking

MLflow Pattern

import mlflow

mlflow.set_experiment("customer-churn-prediction")

with mlflow.start_run(run_name="xgboost-v2"):
    mlflow.log_params(params)
    mlflow.log_metrics({"auc": auc_score, "f1": f1_score})
    mlflow.log_artifact("confusion_matrix.png")
    mlflow.sklearn.log_model(pipeline, "model")
    mlflow.set_tag("version", "2.1")

What to Track

Category	Items
Parameters	All hyperparameters, random seed
Metrics	Train and validation metrics
Data	Data version/hash, feature list
Artifacts	Plots, reports, model files
Metadata	Training duration, model size

---

Statistical Tests Decision Table

Question	Test	Assumption
Two group means different?	t-test (independent)	Normal distribution
Two groups (non-normal)?	Mann-Whitney U	None
Paired measurements?	Paired t-test	Normal differences
3+ group means?	ANOVA	Normal, equal variance
Categorical association?	Chi-squared	Expected freq > 5
Distribution normal?	Shapiro-Wilk	n < 5000
Two distributions different?	Kolmogorov-Smirnov	Continuous data

P-Value Guidelines

• p < 0.05: statistically significant (conventional)
• Always report effect size alongside p-value
• Adjust for multiple comparisons (Bonferroni, FDR)
• Statistical significance is not practical significance

---

Visualization Decision Table

Data Type	Plot	Library
Distribution	Histogram, KDE, Box plot	seaborn
Comparison	Bar chart, Grouped bar	matplotlib
Correlation	Scatter, Heatmap	seaborn
Trend	Line chart	matplotlib/plotly
Composition	Stacked bar, Pie (max 5 slices)	matplotlib
Interactive	Scatter, Line, Dashboard	plotly

Visualization Rules

• Title every plot descriptively
• Label axes with units
• Use colorblind-safe palettes (seaborn: colorblind)
• Start y-axis at 0 for bar charts
• Annotate key findings directly on plots

---

Jupyter Notebook Structure

1. ## Setup (imports, configuration)
2. ## Data Loading
3. ## Exploratory Data Analysis
4. ## Data Preprocessing
5. ## Feature Engineering
6. ## Modeling
7. ## Evaluation
8. ## Conclusions

Notebook Best Practices

• Restart and run all before sharing
• Keep cells focused and sequential
• Use markdown cells for explanations
• Extract reusable code to .py modules
• Version control with nbstripout
• Pin all dependency versions

---

Anti-Patterns / Common Mistakes

Anti-Pattern	Why It Is Wrong	Correct Approach
Training on test data	Data leakage, inflated metrics	Strict train/test separation
Feature engineering before split	Leaks test information into features	Engineer on training data only
Reporting training metrics	Not generalizable	Report validation/test metrics
Accuracy on imbalanced data	Misleading (majority class wins)	Use F1, AUC-ROC, or AP
Tuning on test set	Overfitting to test data	Use validation set for tuning
No baseline comparison	Cannot measure improvement	Always establish baseline first
Cherry-picking evaluation examples	Selection bias	Report on full evaluation set
Deploying without drift monitoring	Silent model degradation	Monitor input distributions

---

Integration Points

Skill	Relationship
senior-prompt-engineer	Prompt evaluation uses statistical testing methods
testing-strategy	ML testing follows the evaluation methodology
performance-optimization	Model inference optimization follows measurement cycle
acceptance-testing	Model performance thresholds become acceptance criteria
llm-as-judge	Subjective output evaluation uses LLM-as-judge
code-review	Notebook and pipeline code reviewed for quality

---

Skill Type

FLEXIBLE — Adapt preprocessing, modeling, and evaluation approaches to the specific data characteristics, business requirements, and compute constraints. The four-phase process and experiment tracking are strongly recommended. Always establish a baseline before modeling.