Lecture 2

Question 1

What is a learning example

Answer 1

• decompose objects into features (from all characteristics pick important characteristics of instances)
• decompse inputs into features (?dunno what he means)
• a feature is a measurable aspect of an object/instance
• features are extracted before learning
• some LA can extract features from some types of inputs (e.g. image or text)

Question 2

a feature

Answer 2

a measurable apsect of an object

Question 3

feature transformation

Answer 3

new features of X often include transformations features. These transformations are part of pre-processing, but can make a problem much easier.

PCA, Neural Networks, scaling and normalisation

example: from catesian coordinates to Polar coordinates

Question 4

name some classifiers

Answer 4

logistic regression
kNN
decision tree

Question 5

Decision boundaries logistic regression

In short:

• DB can be linear function or polynomial function (both with one or more variables)
• iterative

Answer 5

• regression coefficients decision boundary usually estimated using maximum likelihood estimation.
• g ( f (x ) ) * = decisions boundary is linear or multiple linear equation,
  CAN ALSO BE polynomial function (but classifier is always 2 classes, Y hat = 0 or Y hat = 1 )
• you can not find a closed-form solution like LR
• ITERATIVE PROCESS untill process has converged
• function composition is an operation that takes two functions f and g and produces a function h such that h(x) = g(f(x)). In this operation, the function g is applied to the result of applying the function f to x.

Question 6

logistic regression

Answer 6

Y variable is binary (nominaal) and X variables can be categorical or numeric.

Question 7

K-nearest Neighbor

Answer 7

Simple idea: similarity (distance, input can be numeric or categorical)

Give a new example Xj

We look for the most similar example in the TrS

Predict the same target for Xj

The key component of kNN is the distance function. Depending on how you define distance you can get very different classifiers / performance.

Question 8

Measurement of / defining similarity

• small distance ; similar object
• large distance ; dissimilar object

Answer 8

Distance functions

Numeric (5):

• General Lp-metric (Minkowski)
• Euclidean distance (p=2)
• Manhattan distance (p=1)
• Maximum metric (Chebyshev) (p = infinite)
• Cosine Distance

https://www.sciencedirect.com/topics/computer-science/minkowski-distance –> see Figure 2.23.

Categorical (2):

• Hamming distance (p=0)
• Levenshtein distance

Question 9

Hamming distance

Answer 9

Looks at each attribute; are they equal or not. Count the different attributes.

1. KAROLIN and KATHRIN hamming distance is 3, because the ROL are different, this is THR in the second one.
2. 1 0 1 1 1 0 1 and 1 0 0 1 0 0 1 hamming distance is 2

Question 10

Chebyshev distance

Answer 10

generalisation of the Minikowski distance for h –> infinity .

Let’s use two objects, x1 = (1, 2) and x2 = (3, 5). The second attribute gives the greatest difference between values for the objects, which is 5 − 2 = 3. This is the Chebyshev distance.

Question 11

Manhatten distance

Answer 11

You can only go sideways and up.

Let x1 = (1, 2) and x2 = (3, 5) represent two objects. The Manhattan distance between the two is 2 + 3 = 5.

2 = 3 - 1 / 3 = 5 - 2

Question 12

Euclidean distance

Answer 12

In 2 dimensions:

Let x1 = (1, 2) and x2 = (3, 5) represent two objects. The Euclidean distance between the two is wortel van (2^2 + 3^2) = 3.61

2 = 3 - 1 / 3 = 5 - 2

Question 13

Finding the nearest neighbor

Answer 13

Learning as MEMORIZATION

Given a test point, measure the distances to all the training points and pick the k nearest ones

Their labels define the estimated label of the test point (the label corresponding with features)

Question 14

Choose right value of k

Answer 14

if you pick to LARGE –> UNDERFITTING
- everything is classified as most probable class

if you pick to SMALL –> OVERFITTING (variability, unstable decision boundaries)

Question 15

units in kNN

Answer 15

Units do matter in kNN

Suppose you had a dataset (m “examples” by n “features”) and all but one feature dimension had values strictly between 0 and 1, while a single feature dimension had values that range from -1000000 to 1000000. When taking the euclidean distance between pairs of “examples”, the values of the feature dimensions that range between 0 and 1 may become uninformative and the algorithm would essentially rely on the single dimension whose values are substantially larger

You have to transform features to standardized units –> z-score dimensions (just standardizing ==> z = ( X - mean of X) / s.d. of X

Question 16

Problems kNN

Answer 16

some dimensions may be more informative about the class than others

Can we take this into account in the k-NN algorithm?

Do we need to ‘‘forget” training examples, with a growing (training) dataset?

Question 17

Advantages kNN

Answer 17

fast learner (since there is no abstraction)

fast classification possible (using smart indexing structures like k-d-trees)

directly provides illustrative examples (–> the k-nearest neighbors)

Question 18

Cosine distance

Answer 18

Cosine similarity is generally used as a metric for measuring distance when the magnitude of the vectors does not matter.

Question 19

Decision tree classifier

Answer 19

a classification scheme which generates a tree and a set of rules from given dataset.

It take one feature at the time and test a binary condition. Each node test a condition on a feature (if else statement).

The order of the nodes are important. The first question maximizes the information gained (entropy) from the answer.

Based on information theory.

Normalisation does not help, no need to do that

The decision boundaries are PERPENDICULAR to the instance-space-axes

PRUNE/PRUNING is a technique that reduces the size of decision trees by removing sections of the tree that provide little power to classify instances –> REDUCE OVERFIT

Question 20

i3D algorithm (decision tree)

Answer 20

Basic algorithm

• all (training) instances are assigned to the root
• the next attribute (test) is seleected - splitting strategy
• the training set is partitioned using the split attribute
• proceed for all partitions recursively –> locally optimizing algorithm

Stopping criterion

• no more splitting attributes
• all instances of the node belong to exacly to one class.

Question 21

Entropy

Answer 21

Tells us how pure or impure a subset is. Number between 0 and 1 (bits).

If the entropy is 1, you are totally uncertain, you have 50 percent change it is yes or no.

If the entropy is 0, you are totally certain, 100% sure what the outcome label is going to be, it is always yes or always no.

Question 22

Information Gain

Answer 22

Expected drop in entropy after the split.

If I plit on this attribute, how much more certain am I going to be after te split, compared to before the split.

You want entropy to be low (entropy of 0 = 100% certain, pure) and iniformation gain to be high (so entropy goes down).

Question 23

Complexity of induced model DT

Answer 23

The complexity of the model induced by a decision tree is determined by the depth of the tree

Increasing dept of the tree increases the number of decision boundaries

All DB are PERPENDICULAR to the feature axes, because at each node a decision is make about a single feature.

Question 24

Advantages of DT

Answer 24

simple to understand and interpret

work with relatively little data

help to find which feature is most important for classification

rule base

Question 25

Multiclass Classification - One-vs-all algorithm classifier

Answer 25

e.g. if you have 3 classes, you turn your dataset in 3 seperate binary classification problems. You have 3 LINEAR decision boundaries (even if it is perpendicular, the slope is undefined). Datapoints could be in 0 or more classes. Train a logistic regression classifier for eacht class i to predict the probability that y = i. On a new input x, to make a prediction, we run the 3 classifiers on input x and pick the class that is most confident that it is the right class.

Question 26

Underfit (you can only tell by comparing a val/train set error with a test set error) High bias, because your test data fit is far away from the data points, but low variance, because it is more likely that future, different datasets will fit better whit this line. https://www.youtube.com/watch?v=fDQkUN9yw44

Answer 26

- High bias (and low variance) ; ** High training error ** High CV error; approximately as high as training error

Question 27

Overfit (you can only tell by comparing a val/train set error with a test set error) Low bias, because the data points in the test set fits perfectly, but it is unlikely that is fits exactly as good with future datasets. So very unlikely that this describes the dataset. We can not generalize well beyond training data. https://www.youtube.com/watch?v=fDQkUN9yw44

Answer 27

- fit training data very closely - able to predict very well on training data - does not generalize wel on unseen data (test set) - High variance (and low bias) ; * * Low training error * * High CV error; CV error >> training error Common causes of overfit: a) Too complex model (to high degree of polynomials and/or too much features) b) Data has noise i.e. like there are outliers and errors in data c) Size of data used for training may not be enough

Question 28

Variance

Answer 28

Difference in fit between different test datasets --> distance between test dataset and fit line

Question 29

Hyperparameter

Answer 29

variable part of the model which is not set during learning on training data needs to be tuned on validation set

Question 30

Model Paramater

Answer 30

- required by the model when making predictions - value define the skill of the model on your problem (we need to use some data to estimate paramaters e.g. for logistic regression we have to use some data to estimate the shape of the S-curve - Are learned from data

Question 31

Model Paramater

Answer 31

- required by the model when making predictions - value define the skill of the model on your problem. - Are learned from data - Often not set manually by the practitioner - Often saved as part of the learned model - We need to use some data to estimate paramaters e.g. for logistic regression we have to use some data to estimate the shape of the S-curve. Estimating parameters is called 'training the algorithm'.

Question 32

Model Hyperparameter

Answer 32

- Often used in processes to help estimating model parameters - Often specified by the practitioner - Can be set using heuristics - Often tuned for a given predictive modeling problem e. g. the k, neighbor weights, distance metric (3) in K-nearest neighbors. You can control the fit in kNN by changing K. Smaller K - more fit, larger K - less fit.

Question 33

Grid search: hyperparameter optimization

Answer 33

Systematic search for best hyperparameter settings Manually choose values to test for each hyperparameter. Check validation accuracy/error for all combinations Number of parameters expand, but often behave parallel (relatively independent) Lots of computation

Question 34

Evaluating performance regression task

Answer 34

Coefficient of determination (R2) --> nested Root Mean Squared error --> is sensitive for outliers (because you square) Mean Absolute Error --> absolute value, not sensitive to outliers

Question 35

R2

Answer 35

How well the model predicts targets relative to the mean Equivalent to PROPORTION OF VARIANCE EXPLAINED BY THE MODEL The mean is not always a suitable baseline

Question 36

Evaluating performance classification task

Answer 36

Confusion matrix Accuracy, precision, recall, F1 score Area Under the ROC curve --> can inspect outcome for different TRESHOLDS (criterion)

Question 37

Accuracy

Answer 37

TN + TP / all values The proportion that your model predicted correctly. The predicted negatives and predicted positives were actually labeled as negative or positive

Question 38

False positive

Answer 38

The model predicted those datapoints to be positive, but the true label was negative You have some - values in your + prediction class

Question 39

False negative

Answer 39

The model predicted those datapoints to be negative, but the true label is positive You have some + values in your - prediction class

Question 40

Recall

Answer 40

TP / TP + FN fraction of all true positives that are correclty identified as positive Evaluation matrix if you want model that rarely fail to detect the true value (cancer, terrrorist)

Question 41

Precision

Answer 41

TP / TP + FP fraction of positives that the model correclty predicted as true positive Evaluation matrix when avoiding false positives is important. We are fine with cases that are not all TP instances are detected, but from the positves we want to be sure that they are truly positive.

Question 42

Specificity

Answer 42

FP / FP + TN fraction of true negatives that the classifier indentify incorrectly as postive (since a false positive that is a value that actually negative)

Question 43

Recall-oriented ML tasks

Answer 43

Consequense of not correctly identifying a positive case is high - legal discovery - tumor detection/earthquakes - often paired with human expert to filter the false postives

Question 44

Precision-oriented ML tasks

Answer 44

Consequense of not detecting false postive is high (e.g. predicting that someone likes something, but they doe not like it, youtube recommendations). You dont want to recommend something that someone does not like, people are not happy with that.

Question 45

Overfit in a confusion matrix

Answer 45

https://datascience.stackexchange.com/ questions/28426/train-accuracy-vs-test-accuracy-vs-confusion-matrix - Difference in train and test accuracy tells us IF model is overfitting. Train accuracy is higher than test accuracy = overfit Train accuracy is lower than test accuracy = underfit - Confusion matrix tells us HOW MUCH it is overfitting

Question 46

Cross validation

Answer 46

To estimate performance on the learned model from available data using one algorithm. --> estimating the parameters for machine learning methods (evaluating how well we trained the model by estimating parameter) To compare the performance of two ore more different algorithms and find the best algorithm for the available data. --> evaluating how well the machine learning methods work (compared to each other)

Question 47

CV methods

Answer 47

Hold out K-fold cross validation Leave one out

Question 48

Downsampling: unballanced classes

Answer 48

``` creates a balanced dataset by matching the number of samples in the minority class with a random sample from the majority class ```

Question 49

Upsampling: unballanced classes

Answer 49

``` matches the number of samples in the majority class with resampling from the minority class ```

Question 50

CV Time from high to low

Answer 50

leave one out k fold hold out

Question 51

CV dataset from large to small

Answer 51

hold out k fold leave one out bootstrapping

Question 52

Generalization

Answer 52

Ability to give accurate predictions for new, previously unseen data

Question 53

Assumptions for generalization

Answer 53

- Future unseen data will have the same properties (probability distribution, spread, correlation e.d.) as current training set - So if accurate on the train DS = accurate on test DS - But this may not happen if trained model is tuned too specifically to trained set

Question 54

Overfit (AMLwithP)

Answer 54

Can capture complex patterns by being great at predicting lots and lots of specific data samples or areas of local variation, but it often misses the global pattern in the training set that would help generalize in test set.

Question 55

Clustering Types of clustering methods https://youtu.be/CiA3ca7W7Eg

Answer 55

- Discover the underlying structure of the data. Identify a finite set of categories, classes or groups. - Objects in the same cluster should be as similar as possible. Objects in different clusters should be as dissimilar as possible. What sub-populations, groups exist in the data? - How many are there? - What are their sizes? - Do elements in a sub-population have any common properties? - Are sub-populations cohesive? Can they be further split up? - Are there outliers?

Question 56

Types of clustering (C) methods https://youtu.be/CiA3ca7W7Eg

Answer 56

• Partitioning methods • Hyperparameter: k, distance function • Goal: partition into k C's with minimal costs • Placing k centroids at random places in space • For each point i find nearest centroid and assing the point to cluster j • update each centroid (e.g. compute mean, as in k-means clustering) • stop when none of C assignments change • Hierarchical methods • HP: Distance function for points and C's • Determines hierarchy of C's, combines respective most similar C's * Density-based methods * Parameter: minimum points per C, distance ================================= • Categorize by goal: - - Monothetic: members have some common property (e.g. all are males aged 15-25) - - Polythetic: cluster members are similar to each other (distance defines membership) • Categorize by overlap: - - Hard clustering (elements either belong to a cluster or not) - - Soft clustering (clusters may overlap) • Flat & hierarchical

Question 57

Defining similarity with distance (D) functions

Answer 57

Requirements for distance (D) functions: - The D between two points is non-negative. - The D from a point to the point itself is equal to 0. - The distance from point i to point j is exactly equal to the distance from point j to point i (symmetry). - The distance from point i to point j via point k is always longer than the direct distance from i to j (triangle inequality).

Question 58

``` Distance functions (See lecture 2) ```

Answer 58

The similarity measure is the measure of how much alike two data objects are. Similarity measure in a data mining context is a distance with dimensions representing features of the objects. If this distance is small, it will be the high degree of similarity where large distance will be the low degree of similarity.

Question 59

Distance functions

Answer 59

Numeric (5): - General Lp-metric (Minkowski) - Euclidean distance (p=2) - Manhattan distance (p=1) - Maximum metric (Chebyshev) (p = infinite) - Cosine Distance https://www.sciencedirect.com/topics/computer-science/minkowski-distance --> see Figure 2.23. Categorical (2): - Hamming distance (p=0) - Levenshtein distance * Generalization of the Euclidean and - Manhattan distances. (If q = 2, d = Euclidean distance and if q = 1 d = Manhattan)

Question 60

Similarity functions

Answer 60

Although no single definition of a similarity measure exists, usually such measures are in some sense the inverse of distance metrics. - Cosine similarity - Pearsons r - OLS coefficient https://brenocon.com/blog/2012/03/cosine-similarity-pearson-correlation-and-ols-coefficients/

Question 61

Cosine Similarity

Answer 61

- Similar: angle is near 0 degrees and the cosine is near 1 (i.e. 100 %) * - Unrelated: angle is near 90 degrees and the cosine is near 0 (i.e. 0 %) - Opposite: angle is near 90 degrees and the cosine is near -1 (i.e. - 100 %) Cosine distance ignores the magnitude of the vectors * See table for degree and cosine clarification https: //socratic.org/questions/how-do-you-tell-whether-the-value-of-tan-90-degrees-is-positive-negative-zero-or

Question 62

K means clustering - Polythetic - Hard boundaries - Flat - K is hyperparameter

Answer 62

* Data partitioned into K sub-populations. K is hyperparameter. Needs to be specified. * Associates each point with a 'centroid'. A centroid = attribute-value representation of a cluster, sort prototypical individual in the sub-population, * Objective: - Minimize the paiswise squared deviations within C (inter cluster) - Maximize the sum of squared deviations between different C's

Question 63

Silhouet coefficient - Measure how well clustering has performed - Helps pick the right value of K - Where s(i) is closest to 1

Answer 63

How well clustering has performed - Pick range of candidates of values for K - Calculate the silhouette coefficient s(i) for each point in the dataset s(i) = b(i) - a(i) / {max b(i) & max a(i) } a(i) = find the average distance of point i to other points in the same cluster (inner cluster distance)--> these points needs to be as similar as possible, near 0 b(i) = average distance of point i to all the other points in the cluster that is the closest (between cluster distance) --> needs to be as dissimilar as possible (so distance as far as possible, near infinity) * ideally a(i) << b(i), * Ideal value of silhouette = 1 and worst possible value for silhouette = -1 - Plot silhouettes for each value of K to identify outliers * if a(i) > b(i) likely that the point is missclassified. Outliers/miss-classified instances have s(i) less than zero, (this is on the left side of the plots). - Pick the K where average silhouette is closest to 1