machine-learning

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: machine-learning
description: >-
  Machine Learning provides end-to-end ML pipeline construction with PyTorch and scikit-learn, covering model selection, training, evaluation, interpretability, hyperparameter tuning, and experiment tracking.
---

# Machine Learning

Part of [Agent Skills™](https://github.com/itallstartedwithaidea/agent-skills) by [googleadsagent.ai™](https://googleadsagent.ai)

## Description

Machine Learning provides end-to-end ML pipeline construction with PyTorch and scikit-learn, covering model selection, training, evaluation, interpretability, hyperparameter tuning, and experiment tracking. The agent builds reproducible ML workflows that follow software engineering best practices: version-controlled experiments, deterministic training, and interpretable results.

The gap between a working notebook and a production ML pipeline is enormous. This skill bridges that gap by enforcing structured experiment management, proper train/validation/test splits, stratified cross-validation, learning curve analysis, and systematic hyperparameter optimization. The agent tracks every experiment with its configuration, metrics, and artifacts, making it possible to reproduce any result months later.

Model interpretability is treated as a first-class requirement, not an optional post-hoc analysis. Every model comes with SHAP values, feature importance rankings, and partial dependence plots that explain what the model learned and why it makes specific predictions. Black-box predictions without explanations are insufficient for scientific and business-critical applications.

## Use When

- Building classification or regression models
- Tuning hyperparameters systematically
- Explaining model predictions with SHAP or feature importance
- Setting up experiment tracking for ML projects
- Evaluating model performance with proper cross-validation
- Training PyTorch models with structured training loops

## How It Works

```mermaid
graph TD
    A[Dataset] --> B[Train/Val/Test Split]
    B --> C[Feature Engineering]
    C --> D[Model Selection]
    D --> E[Hyperparameter Tuning: Optuna]
    E --> F[Cross-Validation]
    F --> G[Best Model Training]
    G --> H[Evaluation on Test Set]
    H --> I[Interpretability: SHAP]
    I --> J[Experiment Logging]
    J --> K[Model Registry]
```

The pipeline enforces a strict separation between tuning (using validation data) and final evaluation (using held-out test data). The test set is touched exactly once, preventing information leakage from repeated evaluation.

## Implementation

```python
import torch
import torch.nn as nn
from torch.utils.data import DataLoader, TensorDataset
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import classification_report, roc_auc_score
import optuna
import shap
import numpy as np

class Classifier(nn.Module):
    def __init__(self, input_dim: int, hidden_dim: int, dropout: float):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(input_dim, hidden_dim),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, hidden_dim // 2),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim // 2, 1),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.net(x)

def train_epoch(model, loader, optimizer, criterion, device):
    model.train()
    total_loss = 0
    for X_batch, y_batch in loader:
        X_batch, y_batch = X_batch.to(device), y_batch.to(device)
        optimizer.zero_grad()
        pred = model(X_batch).squeeze()
        loss = criterion(pred, y_batch.float())
        loss.backward()
        optimizer.step()
        total_loss += loss.item() * len(X_batch)
    return total_loss / len(loader.dataset)

def hyperparameter_search(X: np.ndarray, y: np.ndarray, n_trials: int = 50) -> dict:
    def objective(trial):
        hidden = trial.suggest_int("hidden_dim", 32, 256)
        lr = trial.su