Pattern Recognition with Python in 14 Days

Fourteen days. One library, prkit, checked line by line against scikit-learn. By Day 14 it aligns two sequences into a complete recognizer.

A classifier can look correct on paper and still be wrong in practice, because one accuracy number hides which mistakes it makes and what each one costs. Most courses teach the formulas and stop short of checking that a hand-rolled routine agrees with a trusted library.

Day 1 draws the nearest-mean boundary on the Iris dataset, the simplest rule prkit builds. Day 5 fits a Gaussian's own mean and covariance straight from data instead of assuming them. Day 8 finds the widest-margin boundary, then bends it with a kernel where no line will separate the classes. Day 14 aligns two sequences by dynamic time warping into a complete recognition system.

Every number here was measured, never assumed. 308 examples across 14 days, each with its captured output and a figure, and every routine in prkit is checked against scikit-learn or SciPy before it is trusted. All 14 chapters close with a review-and-practice drill, so a formula is tested twice: once by the library, once by the reader. Nothing in prkit is asserted where it can instead be run.

YOUR 14-DAY PATH

Day 1 - Patterns, Features and Decisions: the feature vector, the decision rule, and counting mistakes.
Day 2 - Probability for Decisions: priors, likelihoods, and the posterior rule.
Day 3 - Decisions That Cost Different Amounts: loss matrices, expected risk, and the option to decide nothing.
Day 4 - The Gaussian and Its Boundaries: covariance, Mahalanobis distance, and the discriminants it produces.
Day 5 - Learning the Density from Data: maximum likelihood, shrinkage, and naive Bayes.
Day 6 - Densities Without a Formula: histograms, Parzen windows, and k nearest neighbours.
Day 7 - Drawing the Boundary Directly: the perceptron, least squares, and logistic regression.
Day 8 - The Widest Margin: margins, the hinge loss, and the kernel trick.
Day 9 - Changing the Features: principal components, whitening, and Fisher's direction.
Day 10 - Choosing the Features: filters, mutual information, and wrappers that leak.
Day 11 - Measuring What Was Built: cross-validation, confusion matrices, and ROC curves.
Day 12 - Many Classifiers, One Answer: voting, bagging, random forests, and boosting.
Day 13 - Patterns Without Labels: k-means, mixtures, and hierarchical clustering.
Day 14 - Sequences, and a Whole System: dynamic time warping, hidden states, and one measured pipeline.

WHO IT IS FOR. Python developers who want to know what a classifier is deciding, not just call one from a library. You need working Python and NumPy; scikit-learn and SciPy are used only to check prkit's own answers, never to replace them. No deep learning framework and no GPU are used anywhere in the book. It does not cover neural networks, which stay outside its 14 days.


Download book source code from: {GitHub}/{mohammadnsaqib}

428 pages, black and white, every output run and checked against a library.