47  Module 14 Cheat Sheet

Key concepts, definitions, and code from Chapters 33–34

A quick-reference summary of the essential ideas from Module 14. This module is a map rather than a toolkit — the goal is knowing what exists and where to go next.


Gradient Boosting

Gradient boosting builds trees sequentially, each one trained to correct the errors of the ensemble so far. A random forest builds trees in parallel and averages them. Boosting usually wins on tabular data, at the cost of more careful tuning.

Random Forest Gradient Boosting
Trees built In parallel, independently Sequentially, each fixing the last
Base trees Deep, low bias Shallow “stumps”, high bias
Combines by Averaging / voting Weighted sum
Overfits if Rarely — more trees is safe Yes — too many trees overfits
Tuning Forgiving Sensitive, especially learning rate
from sklearn.ensemble import GradientBoostingClassifier, GradientBoostingRegressor
from xgboost import XGBClassifier, XGBRegressor

model = XGBRegressor(
    n_estimators=500,
    learning_rate=0.05,   # smaller needs more trees but generalizes better
    max_depth=4,          # shallow by design
    subsample=0.8,
    random_state=123,
)
model.fit(X_train, y_train)

n_estimators and learning_rate trade off directly: halve the rate, roughly double the trees.


Neural Networks

A neural network stacks layers of simple units. Each unit computes a weighted sum of its inputs and applies a non-linear activation function; stacking these lets the network approximate very complex relationships. It learns by backpropagation — comparing predictions to truth and pushing the error backward to adjust weights.

Term Meaning
Layer A group of neurons; depth is what makes it “deep” learning
Activation Non-linearity (ReLU, sigmoid) — without it the network collapses to linear
Backpropagation How errors propagate backward to update weights
Epoch One full pass over the training data
Learning rate Step size for each weight update
Architecture Built for
Feedforward (MLP) General tabular data
CNN Images and spatial data
RNN / LSTM Sequences and time series
Transformer Language, and increasingly everything else

Choosing an Algorithm

Situation Reach for
Tabular data, need explanation Linear / logistic regression, a single tree
Tabular data, need accuracy Random forest, then gradient boosting
Images, audio, or text Deep learning, usually a pretrained model
Small dataset (< 1,000 rows) Simpler models; deep learning will overfit
Unlabeled data Clustering, PCA

On tabular business data, a tuned gradient boosting model is usually the strongest option, and neural networks rarely justify their complexity. Deep learning’s advantage shows up on unstructured data.

The interpretability–performance spectrum runs from linear models (fully explainable, less flexible) through trees and forests to deep learning (most flexible, hardest to explain). Regulated decisions — credit, hiring, medical — often require the explainable end regardless of accuracy.


Foundation Models and Transfer Learning

A foundation model is trained once at enormous scale, then adapted to specific tasks. Transfer learning means starting from those learned representations instead of from scratch — which is why modern practice often begins with a pretrained model and a small amount of task-specific data.


Where to Go Next

Skills beyond modeling — the ones that separate a course project from production work:

Area Why it matters
SQL and data engineering Most real work is getting and shaping the data
Version control and testing Analyses must be reproducible by someone else
Deployment and monitoring A model in a notebook creates no value
Communication The best analysis unexplained changes nothing
Experiment design A/B testing and causal inference answer “did it work?”

Choosing a direction:

If you want to Study next
Go deeper on prediction Gradient boosting, feature engineering at scale, time series
Work with text or images Deep learning, then transformers and pretrained models
Drive business decisions Causal inference, experiment design, optimization
Build and ship systems Data engineering, MLOps, cloud platforms

See Chapter 34 for the full roadmap and curated resources for each path.