Classification, Clustering, And Feature Learning

Question 1

When is classification used?

Answer 1

When relating categorical response variables to predictor variables.

Question 2

What is a supervised process?

Answer 2

Processes that require a training dataset that shows how particular attributes match up to an outcome of interest.

E.g a set of data where we know for sure that individuals either have or do not have a disease, and we have various attributes measured for each individual.

This can be made into a model and tested on some test data.

Question 3

What is an unsupervised process?

Answer 3

A process that doesn’t require a training dataset or any pre defined outcomes

Question 4

What are the methods of classification?

Answer 4

1) Logistic regression
2) Random forest classification
3) Support vector machines

Question 5

What does the logistic function do?

Answer 5

Transforms linear function allowing us to visualise a binary y scale where y varies from 0 to 1.

This is useful in a binary classification problem when things you want to predict can either take a value of 0 or 1 (yes or no)

This allows for a continuous predictor variable on the X axis and a probability of achieving classification 1 or 0 on the y axis.

Question 6

What does the X axis of a logistic regression curve show?

Answer 6

Shows the continuous predictor variable

Question 7

What does the y axis of a logistic regression curve show?

Answer 7

The probability of achieving classification of 1 or 0.

The predictions (on the y axis) for each value of the predictor variable (X axis) give a probability for observing the outcome we have coded as “1” for that value of the predictor variable.

Question 8

How do support vector machines achieve classification?

Answer 8

They achieve classification by mapping the training data points to points in k dimensional space.

Where k is the number of attributes of the data.

Once data is plotted in k dimensional space a hyperplane separates known categories whilst achieving maximum separation between categories.

Question 9

What is a hyperplane?

Answer 9

The hyperplane is one dimensional order lower than the order of k dimensional space. Allowing the hyperplane to slice through k dimensional space therefore separating categories

Question 10

What does random forest classification do?

Answer 10

The algorithm takes the training dataset and uses it to build a set of decision trees, each based on a subset of the training data.

When we want to classify a new data point the data point is classified by every decision tree in the forest.

Each tree then votes for how the data point should be classified.

Giving us a measure of the probability that the data point belongs in a particular class.

Question 11

What is an advantage of random forest classification?

Answer 11

Random forest algorithms also offer embedded feature selection.

Therefore we can look at the measures it uses to determine the importance of each factor.

Question 12

What is mean decrease accuracy?

Answer 12

R permutes each variable of interest and measures how the accuracy of classification is affected.

The greater the decrease in accuracy when a variable is randomly permuted the more important it is.

Question 13

What is mean decrease Gini?

Answer 13

Gini is a way of measuring if something is homogeneous or heterogeneous.

Each node in each decision tree is associated with a decrease in Gini impurity score for descendent nodes compared to the parent node.

Adding up all the reductions every time a node decision tree is split based on a particular variable gives a measure of importance for that variable.

Question 14

What is a test data set? And what can it be used for in classification?

Answer 14

A test dataset is some data that has been kept where you know the classification of each data point and test a model against it to see how well the model makes predictions compared to the tire classifications.

Question 15

Draw a confusion matrix?

Answer 15

Drawing

Question 16

What is sensitivity?

Answer 16

What proportion of actually true readings did the test manage to detect.

True positives/(true positives + false negatives)

Question 17

What is specificity?

Answer 17

The proportion of actually false readings did the test manage to detect.

True negatives/(true negatives + false positives)

Question 18

What are ROC curves used for?

Answer 18

To study the predictive performance of a binary classifier across different values of its discrimination thershold

Question 19

What is a discrimination threshold?

Answer 19

The threshold at which the classification algorithm deems a data point to belong to one category or another

Question 20

What are the axis of a ROC curve?

Answer 20

Sensitivity Vs 1-specificity

Question 21

What is the AUC of a perfect classifier?

Answer 21

1

Question 22

What is the AUC of a useless classifier?

Answer 22

0.5

Question 23

Is clustering supervised or unsupervised?

Answer 23

Unsupervised

Question 24

Is classification supervised or unsupervised?

Answer 24

Supervised

Question 25

Why would you want to cluster?

Answer 25

When we believe there might be groups within the dataset but we are not sure how many groups or what their characteristics might be.

Question 26

What are the methods of clustering?

Answer 26

1) K-Means clustering

2) Hierarchical clustering

Question 27

What is the goal of k means clustering?

Answer 27

To assign n data points to k clusters

To minimise the within clusters sum of squares (the amount of variation within each cluster).

Question 28

What is the goal of heirarchical clustering and how is it done?

Answer 28

1) it builds a tree of data items based on a predefined distance metric between data points. Things that are closer to eachother appear closer in the tree.

2)Each data point starts off in its own cluster.
The algorithm iteratively merges the closest pairs of clusters untill you arrive at the never of clusters you wanted.

Question 29

How do you decide the right number of clusters?

Answer 29

1) Elbow Method

2) BIC and AIC (preferred model ( one with most appropriate clusters) is the one with minimum BIC/AIC).

Question 30

What does the elbow method plot?

Answer 30

Number of clusters on X axis

Within clusters sum of squares on y axis

Question 31

How does the elbow method work?

Answer 31

There comes a point where increasing the number of clusters doesn’t give you any more explanatory power.

This is the point on the graph where there is hardly any more decrease in within clusters sum of squares.

Question 32

How can you tell if clustering algorithm matches known categories?

Answer 32

Use Adjusted Rand Index.

Question 33

What is the adjusted Rand Index?

Answer 33

a+b/a+b+c+d

a= # pairs of items in the same cluster in both partitions
b= # pairs of items in different clusters in both partitions
c= # pairs in the same cluster in partition X but different clusters in partition Y

d= # pairs in different clusters in partition X but same clusters in partition Y

Question 34

What is the best score for an adjusted Rand Index?

Answer 34

The higher the better 1 is the best

Question 35

What is feature learning?

Answer 35

Feature learning transforms raw data into learned features.

Drawing elements out of all features into a learned feature which has combined elements of the features that were measured.

Question 36

How can feature learning be done?

Answer 36

Feature learning can be done through PCA

Question 37

What is PCA

Answer 37

The data is plotted on its principal components. Which are lines in space that account for the data’s variation .

This therefore can incorporate multiple features at the same time.