R - Unsupervised Learning

K mean

kmeans(x 2 dimensional array, center = ….  how many groups, nstart = ….. how many time to run k mean)

$cluster contains the number of the cluster an observation belongs to

plot(x, col = $cluster, main = “k-means clusters”, xlab = “”, ylab = “”)
main = title, xlab,ylab = the labels

Total within SS (Sum of Squares), the lower the better

par(mfrow = c(2, 3)) Set up a 2×3 plotting grid, displaying 6 plots in 1 grid

Hierarchy clustering

hclust(dist(x), method = “complete/average/single/centroid”)

plot(hclust)

cutree(hclust, h = 6 or k = 2), cut by height(max distance of observations) or number of clusters

Complete and Average produce more balanced tree. Single and Centroid produce less balanced trees, this can be used to detect outliers!

Scaling

apply(x, 1 or 2, sd) find sd for row(1) or column(2)

scale(x) convert to normal distribution with mean = 0 and sd = 1

Dimensionality Reduction
Principal Component Analysis PCA

prcomp(x data, scale = FALSE normal distribution?, center = TRUE centered around 0 or not)

2018-11-05 13_56_00-unsupervised ch3.pdf - Adobe Acrobat Pro.png

92.46% variability of the data is retained with 1 components
97.769% variability of the data is retained with 2 components

Visualizing

Biplot, same direction vector showing correlation of variables

biplot(pr)

 

Scree plot, either by marginal or cumulative variance explained

var explained pve = pvar/sum(pvar), pvar = psd^2

summary(pr) can also shows the cumulative var explained

2018-11-05 14_09_17-Visualize variance explained _ R.png

 

Combining PCA and a clustering technique can improve the model!

Comments

Post a Comment

Popular posts from this blog

Jonas - Javascript

  variables – var abc = ”, – data types   Section 3 – How JS works Scoping When refer to a var that is not available in the current scope, the program will look to the outer scope (or even outer outer scope)   Section 4 – DOM Manipulation   classList, .add .remove classes .toggle, turn on/off the “active” class   Section 5 – Objects and Functions Everything is an object, check with  console.log () for __proto__ when we try to access properties or methods of an object it will look in its own scope first, then its proto parents, and so on. call, apply, bind, to change the  this  variable. And to pre-set some arguments of a function  
Read more

R - Supervised Learning

Image
  knn k nearest neighbors, k library(class) knn(training data, test data, labels), extra args: k = 2. prob = TRUE enables probability of each guess. Primitive bayes nrows(df), count rows subset(df, conditions/ col > 0 etc.), nrows this subset then do divisions Naive bayes model <- naive_bayes(y ~ x1 + x2 + … , data = df) y as explained by x1, x2 etc. predict(model , test_df) + args type = “prob” show posterior probability (instead of priori overall prob) Laplace correction:  add 1 to prevent the 0% probability affects all estimation Logistic Regression model <- glm(y ~ x1 + x2 …, data = df, family = “binomial”) predict(model, test_df, type = “response”) ROC curve to measure how well the model performs, AUC area under curve value from 0 – 1, 1 is best. library(pROC) ROC <- roc(actual var, predicted var) plot(ROC), auc(ROC) Decision Tree library(rpart) model <- rpart(y ~ x1 + x2 + …, data = df, method = “class”, control = rpart.control(cp = 0)) predict(m...
Read more

Consulting Interview

  Numerical skills – rather simple tips on quick estimation, rounding Top down approach is preferred Hypothesis – Always start with a hypothesis, then reason with the info to either prove it or disprove it. Don’t present info and a bunch of rationales before the conclusion, it’s just less clear to the listeners, especially C-level executives. Synthesis – Let’s do A. Because of B, C and D.
Read more