Kursus
Dasar-Dasar Machine Learning
Kebiasaan yang menjaga model tetap jujur dan algoritma klasik, masing-masing dengan contoh scikit-learn yang bisa dijalankan.
- Pelajaran
- 30
- Waktu baca
- 6 jam 9 menit
- Level
- Pemula
- Bahasa pengantar
- Inggris
- Format
- Artikel dengan kode yang bisa dijalankan
Kursus
Yang akan Anda pelajari
Tiga puluh pelajaran artikel. Fondasi: membagi dataset, data leakage, dan trade-off bias–variance. Fitur: scaling, transformasi, dan rekayasa fitur. Evaluasi: metrik klasifikasi dan regresi serta data tidak seimbang. Algoritma: regresi linear dan logistik, k-nearest neighbours, support vector machine, decision tree, random forest, ensemble, dan gradient boosting. Lalu tuning hyperparameter, regularisasi, seleksi fitur, feature importance, SHAP dan partial dependence plot, clustering, reduksi dimensi, deteksi anomali, dan pipeline dari awal hingga akhir. Setiap pelajaran disertai kode yang bisa dijalankan beserta output-nya.
- Machine learning foundations
- Preparing features
- Measuring performance
- Linear and distance-based models
- Trees and ensembles
- Tuning and regularisation
- Feature selection
- Model interpretation
- Unsupervised learning
- Putting it together
Sebelum mulai
Bisa membaca dan menjalankan Python, dan terbiasa dengan array setingkat NumPy — keduanya dibahas di kursus Pemrograman Python untuk Data. Matematika yang dibutuhkan tidak lebih dari aljabar sekolah; setiap rumus dijelaskan dengan kata-kata.
Silabus
Silabus
Setiap pelajaran menjelaskan satu gagasan, lalu menjalankannya di scikit-learn beserta output-nya. Pelajaran pertama terbuka untuk semua orang.
01 · Machine learning foundations
- 01Dataset splittingTrain, validation and test sets, k-fold cross-validation, and splits that respect classes, groups and time.Pratinjau gratis12 mnt
- 02Data leakageHow information from the test set or the future sneaks into training, inflates your score, and how pipelines stop it.10 mnt
- 03The bias–variance trade-offWhy models underfit or overfit, how to tell from learning and validation curves, and what moves each kind of error.12 mnt
02 · Preparing features
- 04Feature scalingPut features on comparable scales for distance- and gradient-based models: standard, min-max and robust scaling, fitted on train only.11 mnt
- 05Feature transformationReshape skewed numbers, encode categories, bin, impute and expand features, each on the right columns with ColumnTransformer.11 mnt
- 06Feature engineeringBuild new features from domain knowledge — ratios, aggregates, dates and cycles — without leaking the answer.12 mnt
03 · Measuring performance
- 07Classification metricsAccuracy and why it lies, the confusion matrix, precision, recall and F1, ROC and precision–recall curves, and log loss.14 mnt
- 08Regression metricsMAE, RMSE, R² and MAPE: what each one punishes, how to read residual plots, and which to report for which decision.15 mnt
- 09Imbalanced dataWhen one class is rare: stratified splits, class weights, resampling, threshold moving and metrics that do not hide the minority.14 mnt
04 · Linear and distance-based models
- 10Linear regressionFit a straight line to predict a number: the squared-error loss, how the weights are found, regularisation and reading the coefficients.10 mnt
- 11Logistic regressionTurn a linear score into a probability with the sigmoid, train it with log loss, and choose the threshold that suits the decision.10 mnt
- 12k-nearest neighboursPredict from the closest examples: distance metrics, choosing k, why scaling matters and where the curse of dimensionality bites.11 mnt
- 13Support vector machinesFind the widest margin between classes, trade it against mistakes with C, and bend the boundary with kernels.12 mnt
05 · Trees and ensembles
- 14Decision treesSplit the data with yes/no questions: Gini impurity, entropy, reading a fitted tree and stopping it before it memorises.10 mnt
- 15Random forestsAverage many decorrelated trees: bootstrap samples, random feature subsets and the out-of-bag score.10 mnt
- 16Ensemble learningCombine models so their errors cancel: bagging, boosting, voting and stacking, and when each one helps.10 mnt
- 17Gradient boostingAdd small trees that each fix the last one’s residuals: learning rate, number of trees, early stopping and HistGradientBoosting.10 mnt
06 · Tuning and regularisation
- 18Hyperparameter tuningSearch settings with grid and random search on cross-validation, use log-scale ranges, and get an honest estimate with nested CV.11 mnt
- 19Regularisation: ridge, lasso and elastic netPenalise large weights to stop a linear model overfitting: L2 shrinks, L1 zeroes, elastic net mixes both, and alpha sets the strength.14 mnt
- 20Regularisation beyond linear modelsThe same idea in other models: tree depth and leaf limits, learning rate and subsampling in boosting, and early stopping.14 mnt
07 · Feature selection
- 21Feature selection: filter methodsDrop features before modelling with variance thresholds, correlation, mutual information and univariate tests, inside a pipeline.14 mnt
- 22Feature selection: wrapper and embedded methodsLet the model choose: recursive feature elimination, sequential selection, L1 and tree-based SelectFromModel, and their costs.14 mnt
08 · Model interpretation
- 23Feature importanceFind what a model relies on: coefficients, impurity and permutation importance, partial dependence and SHAP, and where each misleads.13 mnt
- 24SHAP valuesSplit one prediction into each feature’s contribution with Shapley values: waterfall, beeswarm and dependence plots, and their limits.15 mnt
- 25Partial dependence and ICE plotsSee how the prediction moves as one feature changes, on average and row by row, and spot interactions and correlated-feature traps.15 mnt
09 · Unsupervised learning
- 26k-means clusteringGroup rows with no labels: centroids and inertia, choosing k with the elbow and silhouette, and why scaling and initialisation matter.12 mnt
- 27Hierarchical clustering and DBSCANClusters of any shape: dendrograms and linkage, density-based clustering that finds noise, and choosing between them.12 mnt
- 28Dimensionality reduction: PCA, t-SNE and UMAPCompress many columns into a few: principal components and explained variance, and non-linear maps for plotting high-dimensional data.13 mnt
- 29Anomaly detectionFind the rows that do not fit: z-scores, isolation forest, local outlier factor and one-class SVM, and how to judge them without labels.14 mnt
10 · Putting it together
Masukan
Bantu kami menyempurnakan kursus ini
Usulkan pelajaran, tunjukkan bagian yang kurang jelas, atau laporkan kesalahan pada kode. Setiap pesan dibaca oleh tim kami dan dibalas lewat email.