Midsemester Exam

Question 1

What’s the difference between Active and Passive learning?

Answer 1

Passive learning is where an expert tells you all the features that help you classify the data and you memorise them. Active learning is where an expert classifies a dataset for you and you discover the features for yourself.

Question 2

What is the process of Classification?

Answer 2

1. Get a training set of data
2. Apply a learning algorithm on the training set to create a model.
3. Apply the model to more of the training set and classify it

Question 3

What is one way you can classify data?

Answer 3

Using a decision tree

Question 4

Why don’t you explicitly tell the model the characteristics of data?

Answer 4

This can be difficult to do. For example, how do you determine that an email is spam? All the features of a spam email? Much easier to just tell the model which ones are spam and which are not.

Question 5

What is the difference between Classification and Clustering?

Answer 5

Classification is supervised learning. You are the expert who supervises the model. Clustering is unsupervised learning.

Question 6

What are the two steps in the classification process?

Answer 6

1. Model construction. 2. Model Usuage

Question 7

Explain the model construction step in the classification process.

Answer 7

To construct a model we need to

• get a training set
• Determine class labels for the training set. E.g Salmon, not salmon
• Determine how to represent the model: E.g using Decision trees, math formulas or rules.

Question 8

What do we do in the model usage step?

Answer 8

Check the accuracy of the model. How? By checking that the test sample’s class labels are correctly classified.

Question 9

What’s the difference between training data and testing data?

Answer 9

Training data and testing data are both partitions. The testing data class label is not known by the model. The mode’s job is accurately predict it.

Question 10

What is one way of determining the accuracy of a model?

Answer 10

Use a confusion matrix to determine the accuracy of the model

Question 11

How do you use a confusion matrix to determine the accuracy of a model?

Answer 11

General idea: The number of correctly classified tuples divided by all the tuples.

T-Positive + T-Negative ____________________________________

T-Positive + F-Positive + T-Negative + F-Negative

Question 12

What is a disadvantage of the Confusion Matrix method of testing model accuracy?

Answer 12

In an example of 10,000 fish that are to be classified. If there are only 10 Salmon that are classified. The accuracy calculation will be very high 99.9. But given the few number of Salmon in the dataset, this could be because the model isn’t detecting things correctly.

Question 13

What is precision and recall?

Answer 13

Precision: exactness – what % of tuples that the classifier labeled as positive are actually positive?

Recall: completeness – what % of positive tuples did the classifier correctly label as positive?

Question 14

How do we calculate precision?

Answer 14

Precision =

True Positive

________________________________

True Positive + False Positive

Question 15

How do we calculate recall?

Answer 15

Recall = True Positive / True Positive + False Negative

Question 16

What is the holdout method of estimation?

Answer 16

Take 70% as the training data and 30% as the testing data. Do this a couple of times.

Question 17

What is the cross validation method of estimation?

Answer 17

Divide the dataset into subsets. Check each subset as a test set and use the other subsets as training set. Do this for all subsets Example: Subset 1 - > test set Subset 2 - 10 -> training set Next step: Subset 2 -> test set Subset 1, Subset 3-10 -> training set

Question 18

What are the three main algorithms for classifying data?

Answer 18

Nearest Neighbour, Bayes, and Decision Tree

Question 19

What’s an instance-based classifier?

Answer 19

Rote learner model of classification. It has to match exactly for the classification to work.

Question 20

What are the three things you need to get K-neartest neighbour to work?

Answer 20

You need three things. 1. The stored records 2. Compute the distance to neighbours. 3. Choose how many neighbours to compare to.

Question 21

What are some things to consider for classification algorithms?

Answer 21

Accuracy - Does it correctly classify class labels Speed - time to build the model, time to use the model Robustness - handling signal and noise Scalability - handling large amounts of data Interpretability - understand the insight from the model

Question 22

What are similarity/dissimilarity in the context of classification algorithms?

Answer 22

Similarity is how alike two data objects are. 1 means the same.

Question 23

What measure do we use to compute similarity between data objects?

Answer 23

Use Euclidean distance. Distance equals the square root of the sum of the squared distance. Example: John(45yo , $10 000), Kelly(34yo, $15 000) Distance = sqrt[(45-34)^2 + (10 000 - 15 000)^2]

Question 24

What are the steps of the K-nearest algorithm/classification?

Answer 24

1. Compute the euclidean distance from the data point to other data points
2. Get the class label from the neighest neighbour. How? Majority vote based off the weight factor
3. What’s the weight factor? Weight = 1/distance^2

Question 25

Why is k-nearest neighbour considered lazy?

Answer 25

It doesn’t build a model. It remembers all the training data and has to compute distance to neighbours every time it is run.

Question 26

What are the issues with choosing a k-value for nearest neighbour?

Answer 26

If k is too small it is sensitive to noise. If k is too large it may pick up data from other class labels.

Question 27

How do we stop one attribute from dominating the k-means algorithm?

Answer 27

Normalize the data so one attribute doesn’t dramatically affect the k-means plots.

Question 28

What’s an obvious disadvantage of k-means algorithm?

Answer 28

Because it doesn’t build a model it has to compute the distance every time it classifies a new record. This is a relatively expensive operation.

Question 29

What’s a Naive Bayes Classifier?

Answer 29

A Naive Bayes Classifier uses Bayes theorem to compute the probability of a class label.

Question 30

What is an advantage of using a Naive Bayes Classifier?

Answer 30

It’s easy to implement and generally gets good results

Question 31

What’s a disadvantage of using a Naive Bayes Classifier?

Answer 31

Assumes attributes are independent. This is not likely in the real world for some cases. Usually there are dependencies in data.

Question 32

What is general idea for a decision tree classification?

Answer 32

Make a tree. Split the tree based on class attributes. The internal node is a class attribute. Make decisions at this point. The leaf node is the class label.

Question 33

What are some issues that you need to consider when developing a decision tree?

Answer 33

How to split the data for decision-making. How to optimize the decision-making part.

Question 34

What’s the difference between a binary split and a multi-way split?

Answer 34

Binary split is like a yes/no split. Two possible outcomes. Multiway split is where there is more than 2 possible outcomes.

Question 35

How do you determine the best split when making a decision tree?

Answer 35

Start with a low gini index.

Question 36

What is the Gini index?

Answer 36

The gini index is a measure of impurity of data used in determinings splits of data in a decision tree.

Question 37

What is the issue of overfitting?

Answer 37

Overfitting in the context of a decision tree is where your decision tree matches the data too precisely. The decision tree may be capturing noise.

Question 38

What’s splitting based on classification error?

Answer 38

TODO

Question 39

What is underfitting?

Answer 39

Underfitting is when a model is too simple, both training and test errors are large

Question 40

How do we fix the issue of overfitting?

Answer 40

Stop the algorithm before it makes a fully grown tree. 1. Stop if all instances belong to the same class. 2. Stop if all the attribute values are the same. Other notes Stop if it doesn’t improve the Gini index.

Question 41

What is an advantage of decision tree based classification?

Answer 41

1. It’s inexpensive to construct. 2. Extremely fast at classifying unknown records 3. Easy to interpret for small-sized trees 4. Accuracy is comparable to other classification techniques for many simple data sets

Question 42

What is cluster analysis?

Answer 42

Cluster analysis is finding objects that belong to a particular. You want the distance between objects in a group to be small. You want the distance between groups to be large. In other words, you want groups of data to look distanct in the groups they belong to.

Question 43

What are some real-world examples of when we use cluster analysis?

Answer 43

Grouping of stocks, websites, genes and proteins. When we want to summarize large datasets.

Question 44

Clustering is used as a ______________ tool.

Answer 44

Clustering is often used as a pre-processing tool before performing a classification or recommendation.

Question 45

How do you define good clustering?

Answer 45

High similarity within clusters. You want it to be cohesive within a cluster. Low similarity with other clusters. You want it to be distinct between other clusters.

Question 46

Why is clustering ambiguous?

Answer 46

Clustering is ambiguous because determining a cluster can be subjective. Where to divide? How many clusters?

Question 47

What’s the difference between hierarchical and partition clustering?

Answer 47

Hierarchical clustering is overlapping that’s organized as a tree. Partition clustering is organized as separate groups with no overlaps.

Question 48

What’s the general idea for K-means clustering?

Answer 48

Each cluster has a centroid. It’s a center point which you measure data points to. The number of clusters is determined at the beginning.

Question 49

What is the process of K-means clustering algorithm?

Answer 49

1. Set a centroid point,
2. Assign data points to the nearest centroid.
3. Recompute the centroid as an average of all the points in the cluster. AKA The centroid moves.
4. Continue 2-3 until the centroid doesn’t move. AKA the “right” clustering has happened.

Note: sometimes we change step 4 until they only move very little.

Question 50

How do you validate the K-means cluster?

Answer 50

Use the Sum of Squared Error formula The errror is the distance from a point to the nearest cluster centroid.

Question 51

Why is picking initial centroids important?

Answer 51

Depending on the data spread, the initial centroids may converge and not actually represent the data clusters.

Question 52

What is the solution to the initial centroid problem?

Answer 52

Do multiple runs and see if there’s a difference.

Question 53

What is post-processing for k-means clustering?

Answer 53

Eliminate ‘small’ clusters that may represent outliers Split ‘loose’ clusters (clusters with relatively high SSE) Merge ‘close’ clusters (clusters with relatively low SSE)

Question 54

What are some limitations of K-means clustering?

Answer 54

• Difference sizes -
• Different densities
• Different shapes (non-circular)

Question 55

How do we determine do k-means algorithm for complex data types?

Answer 55

Use Minkwoski distance,

distance = Sum of (datapoint Pattribute - datapoint Qattribute)^r to the power of (1/r)

r is 1 = manhattan distance

r is 2 = euclidean distance

r is infinity is Max

Question 56

What’s the difference betwen L₁(1, norm), L₂(2, norm) and L_infinity(X,Y) ?

Answer 56

L₁(1, norm) is manhattan distance.

L₂(2, norm) is euclidean distance.

L_infinity(X,Y) is the max of these.

Question 57

What’s the formula for L1 (1-norm)?

Answer 57

Question 58

What’s the formula for L2(2-norm)?

Answer 58

Question 59

What’s the formula for L_{infinity ?}

L∞ (X,Y) ?

Answer 59

Question 60

Can you calculate the L1, L2 and L_infinity?

Answer 60

Question 61

What’s Jaccard Distance?

Answer 61

Jaccard distance is 1 minus the Jaccard Coefficient

Question 62

What’s the Jaccard Coefficient?

Answer 62

Question 63

What’s cosine distance/similarity?

Answer 63

It’s between 0 and 1.

A = {1, 2, 3}

B = {4, 5, 6}

||A|| = 1*1 + 2*2 +3*3

||B||= 4*4 + 5*5 + 6*6

|A . B| = 1*4 + 2*5 +3*6

Question 64

What is Edit distance? How do we measure similarity?

Answer 64

The number of changes that need to be made to transform one string to another.

“Ben” -> “Jen” = 1 edit distance

“Good” -> “Evil” = 4 edit distance.

1/ (1 + edit distance)

Question 65

What is hierarchical clustering?

Answer 65

Hierarchical clustering is when cluster partitions belong to parent clusters.

Question 66

What are the strengths of hierarchical clustering?

Answer 66

No assumption about the number of clusters.

Can be useful for taxonomy. Monkeys -> primates

Question 67

What is agglomerative and divisive hierarchical clustering?

Answer 67

Agglomerative, you start with each point as a single cluster and start to merge them together.

Divisive, you start with one cluster and divide as you go.

Question 68

What are the four ways we can cluster data in agglomerative clustering?

Answer 68

1. Min (single linkage)
2. Max (complete linkage)
3. Group Average
4. Centroid distance

Question 69

What is min/single linkage?

Answer 69

Measure from two closest points in each cluster.

Question 70

What is max/complete linkage?

Answer 70

The measurement from the furthest point in each cluster.

Question 71

What is a limitation of min/single linkage?

Answer 71

It is sensitive noise

Question 72

What is a limitation of max/complete linkage?

Answer 72

It can break large clusters

Question 73

What is group average cluster similarity?

Answer 73

The average distance between each point and all the other points in the other cluster. Summed up. Divided by the size of each cluster multiplied together

Question 74

What is inter-cluster analysis centroid ammogerative clustering?

Answer 74

Calculate the similarity between clusters using the centroids of each cluster

Question 75

How do you calculate F-Measure?

Answer 75

F = 2 * (Precision x Recall)

___________________________

Precision + Recall

Question 76

If you have the precision and recall of two different models. How do you determine which one is better?

Answer 76

Use the F-measure

F-Measure = 2 * Precision x Recall

__________________________________

Precision + Recall

The closer to 1, the better.