Machine Learning Labs

My submitted lab work for a university machine learning / data science course, done in early 2022. Nine R Markdown files, one per lab, covering base R, OLS from scratch, the perceptron and SVM, KNN, ggplot2, Rcpp, trees / bagging / random forests, data.table wrangling, and probability estimation with logistic regression and forward stepwise selection.

Attribution — read this first

The prose in these files is not mine. It is the instructor's.

These labs were handed out as .Rmd templates from the class repository. Everything you see in normal text — the problem statements, the explanations, the ROxygen function specs (@param, @return), the scaffolding chunks that set up data or plot a result, and the #TO-DO markers — was written by the instructor. My contribution is only:

• the R code I wrote inside the chunks, and
• the plain-text answers I typed under the prompts that asked for a written response (usually the lines beginning with #).

The course is Math 342W (with a parallel 650 masters section — several exercises are marked as required for masters students and extra credit for undergraduates). Labs 7-9 use the instructor's own R package, YARF.

I am publishing this as a record of coursework, not as original work. If you want to see code that is entirely mine, this is the wrong repository.

What each lab covers

Lab	Topic
1	Base R: vectors, factors, matrices, apply/lapply/split, sampling, missingness
2	my_reverse, flip_matrix, binary classification on iris, perceptron, SVM via e1071, 1-NN
3	my_simple_ols from scratch, estimation error vs. n, Galton height data, my_ols
4	Design matrix and hat matrix by hand, projections, SST = SSR + SSE
5	Diamonds: log transforms, nested model comparison A-F, in-sample vs. oos error, K-fold CV
6	ggplot2 on the GSS vocabulary dataset
7	Rcpp (all_angles, Fibonacci), trees, bagging, random forests, mtry tuning
8	Bagged OLS, RF imputation of missing data, data.table wrangling on storms, table joins
9	Bills/payments/discounts join, classification tree, asymmetric costs, ROC/AUC, forward stepwise

Known incomplete and non-running parts

Listing these because a reviewer will find them and should know they were known rather than missed. A good number are exercises the assignment itself marked as extra credit or as masters-only; the rest are things I ran out of time on.

Empty exercises (function documented, body left as #TO-DO):

• Lab 2, line 474 — linear_svm_learning_algorithm. The version above it (line 450) has my pseudocode in comments, which is what was required of 342W students; the real implementation below it is empty. The chunk that plots its line will therefore error.
• Lab 2, line 368 — maximum-margin perceptron (extra credit).
• Lab 2, line 567 — KNN with a k argument (masters / extra credit).
• Lab 3, line 298 — dataset with R^2 near 1 and arbitrarily high RMSE (masters / extra credit).
• Lab 4, line 58 — the X-perpendicular matrix (masters).
• Lab 5, line 324 — computing s_e_s_F over 200-observation slices. The ggplot call under it references s_e_s_F, so the chunk errors.
• Lab 7, lines 294, 372, 441 — bias-variance decomposition, and two plots of oob / bootstrap error by number of trees.
• Lab 8, lines 40, 46, 52 — punching holes in a matrix, building Xmiss, and the random forest imputation loop.
• Lab 8, lines 244, 250 — the two storms aggregation exercises.
• Lab 8, from line 508 onward ("Everything below here is due with lab 9") — error metrics, asymmetric costs, logistic regression, ROC, AUC, and the stepwise loop are all empty in Lab 8. Most of that material is answered in Lab 9 instead.
• Lab 9, line 169 — assigning asymmetric costs.

Code that would not run as written:

• Lab 2, flip_matrix — the parameter is x but the body refers to X. It only works if a matrix named X happens to exist in the global environment.
• Lab 2, second nn_algorithm_predict (line 545) — the spec asked me to add a distance argument d, but I never added it to the signature, and the body calls Xstar which is never defined. The simpler one-argument version above it (line 506) does work.
• Lab 3, my_simple_ols — divides by n taken from the global environment instead of length(x). It gives correct answers only when the global n matches the input length. Later in the same lab a chunk runs rm(list = ls()) and then calls the function again, which is why my own comment at line 280 says the RMSE and R^2 came out different than expected and I could not explain it. That is the reason.
• Lab 8, random_bagged_ols — sample(p_se, ...) is a typo for p_seq, and the function never fits or returns any models.
• Lab 9, the forward stepwise repeat loop — passes type = "binomial" where glm wants family = "binomial", fits yselect against the training design matrix, and stores the chosen predictor index in in_sample_brier_by_iteration instead of a Brier score. The plot at the end is drawn from that, so it does not show what it claims to.

Missing data and external dependencies:

• Labs 8 and 9 read bills_dataset/bills.csv.bz2, payments.csv.bz2 and discounts.csv.bz2. Those files were course-supplied and are not in this repository, so those sections cannot be re-run here.
• Lab 8, line 105 calls source() on a raw githubusercontent URL to pull in a pmean helper. That executes third-party code fetched over the network at knit time. I would not write it that way now — the right move is to vendor the one function or use rowMeans(..., na.rm = TRUE).
• Labs 7-9 need a JDK, rJava, and the YARF package installed from GitHub.

Running these

R with pacman installed. Each lab installs what it needs via pacman::p_load. Open a .Rmd in RStudio and knit, or run chunk by chunk. Given the items above, several labs will not knit end to end without edits.

License

Deliberately none. The assignment text and scaffolding in these files are the instructor's work and are not mine to license. The code I wrote inside the chunks is free for anyone to read or reuse.