Machine Learning

AI is the broad field of making machines simulate human intelligence. ML is a subset where machines learn from data. DL uses deep neural networks for complex patterns.

Understanding the hierarchy helps you choose the right approach — rule-based AI, statistical ML, or deep neural nets — based on your data size, interpretability needs, and compute budget.

ML is categorized by the type of feedback a model receives during learning — labeled data (supervised), no labels (unsupervised), or rewards from actions (reinforcement).

Choosing the wrong type means your model can't learn at all. A supervised model needs labels — if you don't have them, you need a different approach entirely.

Data can be categorical (labels/groups) or numerical (measured quantities). Each subtype — nominal, ordinal, discrete, continuous — determines how you can process and model it.

Data type determines which encoding to apply, which algorithms work, and which statistical tests are valid. Treating ordinal as nominal (or vice versa) leads to wrong models.

Pandas is Python's most powerful data manipulation library. It provides DataFrame and Series structures for loading, cleaning, transforming, and exploring tabular data.

80% of ML work is data preparation. Pandas is how you inspect nulls, filter rows, encode features, merge datasets, and create new columns — all before training begins.

Data preprocessing step where missing values (nulls/NaN) and duplicate rows are detected and handled before any analysis or modeling. Dirty data = wrong models. Clean data = reliable results.

Most ML algorithms cannot handle NaN values — they either crash or produce silently wrong results. Duplicate rows bias your model by over-representing certain patterns and inflating evaluation metrics.

EDA is the process of systematically investigating your dataset before building any model. You examine shape, distributions, relationships, missing values, and anomalies using statistics and visualizations.

Without EDA, you're modeling blindly. EDA prevents you from feeding garbage to your model — you catch outliers, skewed distributions, data leakage, encoding issues, and multicollinearity early.

Analyzing one variable at a time. You look at how values of a single column are distributed — their shape, center, spread, and extreme values.

Before comparing features, you need to understand each individually. A bimodal distribution, heavy skew, or extreme outlier in a single column can break models if not addressed.

Identifying data points that deviate significantly from the rest. Methods include IQR fences (box plot whiskers) and Z-score thresholds based on standard deviations from the mean.

Outliers distort model coefficients, inflate RMSE, and skew correlations. A single extreme value can move a regression line significantly. Must detect and decide: fix or remove.

Analyzing the relationship between two variables simultaneously — how they move together (numerical vs numerical), differ across groups (categorical vs numerical), or associate (categorical vs categorical).

Feature selection, correlation detection, and hypothesis testing all start here. You find which features actually relate to the target — and which are noise.

Testing whether a numerical variable (e.g. salary, score) significantly differs across categories (e.g. department, gender). We set up a null hypothesis H₀ and use a test statistic to decide whether to reject it.

Tells you whether a categorical feature has real predictive power for a numerical target. If salary doesn't differ by department, that feature is noise.

Testing whether two categorical variables are associated or independent. We compare observed data to what we would expect if the variables had absolutely no relationship.

Tells you if a categorical feature is related to a categorical target. If Gender and Movie Preference are independent, Gender won't help predict preferences.

Examining 3 or more variables simultaneously. The key tool is the correlation heatmap — a matrix where each cell shows Pearson's r between two features, color-coded by strength and direction.

Reveals multicollinearity (features too similar to each other), which bloats variance and hurts models. Also shows which features correlate with the target.

Mathematical transformations applied to a column's values to reduce skewness. They compress large values, making the distribution closer to normal so models work better.

Many ML algorithms assume normally distributed features. Highly skewed data gives disproportionate weight to extreme values. Log and Box-Cox transformations fix this without removing any rows.

Principal Component Analysis transforms many correlated features into fewer uncorrelated "principal components" that capture the maximum variance in the data.

Too many correlated features (multicollinearity) cause unstable model weights and slow training. PCA compresses information into independent components — directly eliminating multicollinearity.

Synthetic Minority Over-sampling Technique creates new artificial minority samples by interpolating between existing real minority samples — making class sizes more equal.

When 99% of samples are class 0 and 1% are class 1, a model can achieve 99% accuracy by just predicting 0 always — useless! SMOTE forces the model to actually learn the minority class.

A structured overview of the major ML model families — what problem each solves, what data it expects, and which real-world use cases it's best suited for.

Choosing the wrong model type is the single most common ML mistake. The model is a tool — match the tool to the job.

Dividing your dataset into separate subsets so the model is trained on one portion and evaluated on data it has never seen before. This simulates real-world performance.

If you evaluate on training data, your model looks perfect — it just memorized the answers. A separate test set reveals if it actually learned the pattern or just memorized noise.

The two fundamental problems in model training. Underfitting = model is too simple to learn the pattern. Overfitting = model memorizes training noise instead of learning the real pattern. Regularization is the fix for overfitting.

A model that performs perfectly on training data but fails on new data is useless. The goal is generalization — learning the real underlying pattern. This is the core challenge of ML.

Quantitative measures that tell you how well your model performs. Different metrics are appropriate for regression vs classification vs forecasting models.

Accuracy alone is misleading (class imbalance). RMSE penalizes large errors more than MAE. R² tells you proportion of variance explained. Choosing the right metric defines what "good" means for your problem.

A consolidated checklist of the most important habits, rules, and decisions that separate a model that actually works in production from one that only looks good on paper.

Knowing individual techniques is not enough. How you combine them, in what order, and what mistakes to avoid determines whether your model generalizes or fails silently on real data.