ML Pipeline
Machine learning for scientific research. Venv: source /Users/zhangmingda/clawd/.venv/bin/activate
Pipeline Overview
Data → Clean → Features → Split → Train → Evaluate → Interpret → Report
Model Selection Guide
| Task |
Data Size |
Interpretability Need |
Recommended |
| Classification (small) |
< 10K |
High |
Logistic Regression, Decision Tree |
| Classification (medium) |
10K-100K |
Medium |
Random Forest, XGBoost |
| Classification (large) |
> 100K |
Low OK |
Neural Network, XGBoost |
| Regression (linear) |
Any |
High |
Linear/Ridge/Lasso |
| Regression (nonlinear) |
Medium+ |
Medium |
Random Forest, Gradient Boosting |
| Clustering |
Any |
Medium |
K-Means, DBSCAN, Hierarchical |
| Dimensionality reduction |
Any |
Medium |
PCA, t-SNE, UMAP |
| Anomaly detection |
Any |
Medium |
Isolation Forest, LOF |
| Time series |
Any |
Varies |
ARIMA, Prophet, LSTM |
Standard Pipeline
import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split, cross_val_score, GridSearchCV
from sklearn.preprocessing import StandardScaler, LabelEncoder
from sklearn.metrics import classification_report, confusion_matrix, roc_auc_score
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import Pipeline
# 1. Preprocessing
X = df.drop('target', axis=1)
y = df['target']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)
# 2. Pipeline with scaling
pipe = Pipeline([
('scaler', StandardScaler()),
('model', RandomForestClassifier(random_state=42))
])
# 3. Cross-validation
scores = cross_val_score(pipe, X_train, y_train, cv=5, scoring='roc_auc')
print(f"CV AUC: {scores.mean():.3f} ± {scores.std():.3f}")
# 4. Hyperparameter tuning
param_grid = {
'model__n_estimators': [100, 300, 500],
'model__max_depth': [5, 10, None],
'model__min_samples_leaf': [1, 5, 10]
}
grid = GridSearchCV(pipe, param_grid, cv=5, scoring='roc_auc', n_jobs=-1)
grid.fit(X_train, y_train)
# 5. Evaluation
y_pred = grid.predict(X_test)
print(classification_report(y_test, y_pred))
print(f"Test AUC: {roc_auc_score(y_test, grid.predict_proba(X_test)[:,1]):.3f}")
Feature Importance & Explainability
# Built-in importance (tree models)
importances = grid.best_estimator_.named_steps['model'].feature_importances_
feat_imp = pd.Series(importances, index=X.columns).sort_values(ascending=False)
# SHAP values (model-agnostic)
# pip install shap
import shap
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_test)
shap.summary_plot(shap_values, X_test)
Unsupervised Learning
from sklearn.cluster import KMeans, DBSCAN
from sklearn.decomposition import PCA
from sklearn.manifold import TSNE
# PCA
pca = PCA(n_components=0.95) # retain 95% variance
X_pca = pca.fit_transform(X_scaled)
print(f"Components: {pca.n_components_}, Explained variance: {pca.explained_variance_ratio_.cumsum()[-1]:.3f}")
# K-Means with elbow method
inertias = [KMeans(n_clusters=k, random_state=42).fit(X_scaled).inertia_ for k in range(2, 11)]
# t-SNE visualization
X_tsne = TSNE(n_components=2, random_state=42, perplexity=30).fit_transform(X_scaled)
Reporting ML Results in Papers
Always include:
- Dataset description (size, features, class balance)
- Preprocessing steps
- Model selection rationale
- Cross-validation strategy (k-fold, stratified, leave-one-out)
- Hyperparameter search space and method
- Multiple metrics (accuracy, precision, recall, F1, AUC)
- Comparison with baselines
- Feature importance / model interpretation
- Confidence intervals or statistical tests on performance
- Code/data availability statement
Tips
- Always use stratified splits for imbalanced data
- Report multiple metrics, not just accuracy
- Compare against simple baselines (majority class, mean prediction)
- Use nested CV for unbiased performance estimation
- Check for data leakage (especially with time series)
- Document random seeds for reproducibility
1---2name: ml-pipeline3description: Machine learning pipeline for scientific research including data preprocessing, feature engineering, model selection, training, evaluation, and interpretation. Covers supervised/unsupervised learning, deep learning, cross-validation, hyperparameter tuning, and model explainability. Use when user asks to build a predictive model, classify data, cluster samples, do feature selection, or apply ML to research data. Triggers on "machine learning", "classification", "clustering", "random forest", "neural network", "deep learning", "predict", "feature selection", "cross-validation", "train model".4---5
6# ML Pipeline
7
8Machine learning for scientific research. Venv: `source /Users/zhangmingda/clawd/.venv/bin/activate`
9
10## Pipeline Overview
11
12```
13Data → Clean → Features → Split → Train → Evaluate → Interpret → Report
14```
15
16## Model Selection Guide
17
18| Task | Data Size | Interpretability Need | Recommended |
19|------|-----------|----------------------|-------------|
20| Classification (small) | < 10K | High | Logistic Regression, Decision Tree |
21| Classification (medium) | 10K-100K | Medium | Random Forest, XGBoost |
22| Classification (large) | > 100K | Low OK | Neural Network, XGBoost |
23| Regression (linear) | Any | High | Linear/Ridge/Lasso |
24| Regression (nonlinear) | Medium+ | Medium | Random Forest, Gradient Boosting |
25| Clustering | Any | Medium | K-Means, DBSCAN, Hierarchical |
26| Dimensionality reduction | Any | Medium | PCA, t-SNE, UMAP |
27| Anomaly detection | Any | Medium | Isolation Forest, LOF |
28| Time series | Any | Varies | ARIMA, Prophet, LSTM |
29
30## Standard Pipeline
31
32```python
33import numpy as np
34import pandas as pd
35from sklearn.model_selection import train_test_split, cross_val_score, GridSearchCV
36from sklearn.preprocessing import StandardScaler, LabelEncoder
37from sklearn.metrics import classification_report, confusion_matrix, roc_auc_score
38from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
39from sklearn.linear_model import LogisticRegression
40from sklearn.pipeline import Pipeline
41
42# 1. Preprocessing
43X = df.drop('target', axis=1)
44y = df['target']
45X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)
46
47# 2. Pipeline with scaling
48pipe = Pipeline([
49 ('scaler', StandardScaler()),
50 ('model', RandomForestClassifier(random_state=42))
51])
52
53# 3. Cross-validation
54scores = cross_val_score(pipe, X_train, y_train, cv=5, scoring='roc_auc')
55print(f"CV AUC: {scores.mean():.3f} ± {scores.std():.3f}")
56
57# 4. Hyperparameter tuning
58param_grid = {
59 'model__n_estimators': [100, 300, 500],
60 'model__max_depth': [5, 10, None],
61 'model__min_samples_leaf': [1, 5, 10]
62}
63grid = GridSearchCV(pipe, param_grid, cv=5, scoring='roc_auc', n_jobs=-1)
64grid.fit(X_train, y_train)
65
66# 5. Evaluation
67y_pred = grid.predict(X_test)
68print(classification_report(y_test, y_pred))
69print(f"Test AUC: {roc_auc_score(y_test, grid.predict_proba(X_test)[:,1]):.3f}")
70```
71
72## Feature Importance & Explainability
73
74```python
75# Built-in importance (tree models)
76importances = grid.best_estimator_.named_steps['model'].feature_importances_
77feat_imp = pd.Series(importances, index=X.columns).sort_values(ascending=False)
78
79# SHAP values (model-agnostic)
80# pip install shap
81import shap
82explainer = shap.TreeExplainer(model)
83shap_values = explainer.shap_values(X_test)
84shap.summary_plot(shap_values, X_test)
85```
86
87## Unsupervised Learning
88
89```python
90from sklearn.cluster import KMeans, DBSCAN
91from sklearn.decomposition import PCA
92from sklearn.manifold import TSNE
93
94# PCA
95pca = PCA(n_components=0.95) # retain 95% variance
96X_pca = pca.fit_transform(X_scaled)
97print(f"Components: {pca.n_components_}, Explained variance: {pca.explained_variance_ratio_.cumsum()[-1]:.3f}")
98
99# K-Means with elbow method
100inertias = [KMeans(n_clusters=k, random_state=42).fit(X_scaled).inertia_ for k in range(2, 11)]
101
102# t-SNE visualization
103X_tsne = TSNE(n_components=2, random_state=42, perplexity=30).fit_transform(X_scaled)
104```
105
106## Reporting ML Results in Papers
107
108Always include:
1091. Dataset description (size, features, class balance)
1102. Preprocessing steps
1113. Model selection rationale
1124. Cross-validation strategy (k-fold, stratified, leave-one-out)
1135. Hyperparameter search space and method
1146. Multiple metrics (accuracy, precision, recall, F1, AUC)
1157. Comparison with baselines
1168. Feature importance / model interpretation
1179. Confidence intervals or statistical tests on performance
11810. Code/data availability statement
119
120## Tips
121- Always use stratified splits for imbalanced data
122- Report multiple metrics, not just accuracy
123- Compare against simple baselines (majority class, mean prediction)
124- Use nested CV for unbiased performance estimation
125- Check for data leakage (especially with time series)
126- Document random seeds for reproducibility
beita6969