---

name: unsupervised-learning description: "Run and debug ML-From-Scratch unsupervised PCA, clustering, association mining, GMM, genetic search, and RBM workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

ML-From-Scratch Unsupervised Learning

Use this sub-skill when a task asks for ML-From-Scratch workflows in clustering, dimensionality reduction, density modeling, association mining, transaction rules, string-target genetic search, or RBM reconstruction. It assumes the mlfromscratch package is importable in a CPU Python environment and keeps all runtime guidance self-contained.

Route here

• PCA dimensionality reduction with PCA().transform(X, n_components).
• Clustering with KMeans, DBSCAN, PAM, or GaussianMixtureModel.
• Frequent itemset and rule mining with Apriori or FPGrowth.
• Toy optimization/search with GeneticAlgorithm targeting a string.
• Advanced unsupervised reconstruction with RBM.

Route elsewhere

• Supervised regressors, classifiers, trees, ensembles, Bayes, LDA, SVM, boosting, and supervised neural estimators: ../supervised-learning/.
• Low-level neural-network layers, losses, optimizers, CNN/RNN/MLP, GAN, DCGAN, and autoencoder assembly details: ../deep-learning/.
• CartPole, DQN, Gym reset/step compatibility, replay memory, and rendering: ../reinforcement-learning/.
• Cross-cutting package install/import issues: ../../references/troubleshooting.md.

Start with these references

• references/algorithm-catalog.md for constructors, methods, return values, state, and known API quirks.
• references/workflows.md for safe recipes: clustering/PCA, association mining with string normalization, genetic search, and RBM reconstruction.
• references/troubleshooting.md for scaling, DBSCAN density parameters, KMeans empty clusters, transaction formatting, support/confidence, RBM speed, and headless plotting.
• ../../references/package-overview.md for package-level install context.
• ../../references/shared-utilities.md for shared helpers such as normalize, standardize, euclidean_distance, and plotting helpers.

Fast bundled checks

Run these from this sub-skill directory or adapt their in-memory patterns:

python scripts/run_clustering_smoke.py --help
python scripts/run_clustering_smoke.py
python scripts/run_association_smoke.py
python scripts/run_optimization_smoke.py --iterations 5

The bundled scripts avoid network access, credentials, destructive writes, and interactive plotting. They are smoke checks for import/API usability, not benchmark or quality claims.

Operating reminders

• These classes use repo-specific APIs, not scikit-learn estimators. Most clustering classes expose predict(X) directly; PCA exposes transform(X, n_components); there is often no separate fit call.
• Cluster labels are arbitrary IDs. Compare partitions by counts, separation, or downstream use rather than expecting fixed label numbers.
• Set numpy.random.seed(...) before KMeans, PAM, GMM, or GeneticAlgorithm when a deterministic result is needed.
• Scale or standardize numeric features before distance- or covariance-based algorithms unless all features are already comparable.
• Normalize string transactions before Apriori; use count-style support for FPGrowth in this implementation.
• Treat RBM as an advanced, slower reconstruction workflow. Start with tiny batches and few iterations before scaling up.