Download presentation

Presentation is loading. Please wait.

Published byFiona Warne Modified over 2 years ago

1
Clustering and Probability (Chap 7)

2
Review from Last Lecture Defined the K-means problem for formalizing the notion of clustering. Discussed the K-means algorithm. Noted that the K-means algorithm was “quite good” in discovering “concepts” from data (based on features). Noted the important distinction between “attributes” and “features”.

3
Example of K-means -1 Measure 1Measure 2 Patient 111 Patient 221 Patient 334 Patient 445 Let initial centroids be C1 = (1,1) and C2 = (2,1)

4
Example of K-means-2 Measure 1 Measure 2 Dist to C1 (1,1) Dist to C2 (2,1) nearest Patient 11101C1 Patient 22110C2 Patient C2 Patient C2 C1: = (1,1); C2 = ((2+3+4)/3, (1+4+5)/3)) = (3,3.33)

5
Example of K-means-3 Measure1Measure2Dist to C1 (1,1) Dist to C2 (3,3.33) nearest Patient C1 Patient C1 Patient C2 Patient C2 C1: = ((1+2)/2, (1+1)/2) = (1.5,1); C2 = ((3 +4)/2), (4+5)/2) = (3.5, 4.5)

6
Example of K-means-4 Measure1Measure2Dist to C1 (1.5,1) Dist to C2 ( ) nearest Patient C1 Patient C1 Patient C2 Patient C2 C1: = ((1+2)/2, (1+1)/2) = (1.5,1); C2 = ((3 +4)/2), (4+5)/2) = (3.5, 4.5)

7
Example: 2 Clusters c c c c c c c c A(-1,2) B(1,2) C(-1,-2) D(1,-2) (0,0) K-means Problem: Solution is (0,2) and (0,-2) and the clusters are {A,B} and {C,D} K-means Algorithm: Suppose the initial centroids are (-1,0) and (1,0) then {A,C} and {B,D} end up as the two clusters. 4 2

8
Several other issues regarding clustering How do you select the initial centroids? How do you select the right number of clusters ? How do you deal with non-Euclidean distance/similarity measures ? Other approaches (like hierarchical, spectral etc.) Curse of high-dimensionality.

9
Question S-LengthS-WidthP-LengthP-WidthFlower SmallMediumSmallMediumA (SetosA) Medium Large O(Versicolor) MediumSmall LargeI (Virginica) Large MediumSmallA LargeSmallMediumSmall? What should the “prediction” be for the flower ?

10
Prediction and Probability When we make predictions we should assign “probabilities” with the prediction. Examples: 20% chance it will rain tomorrow. 50% chance that the tumor is malignant. 60% chance that the stock market will fall by the end of the week. 30% that the next president of the United States will be a Democrat. 0.1% chance that the user will click on a banner-ad. How do we assign probabilities to complex events.. using smart data algorithms…and counting.

11
Probability Basics Probability is a deep topic…..but for most cases the rules are straightforward to apply.. Terminology Experiment Sample Space Events Probability Rules of probability Conditional Probability Bayes Rule

12
Probability: Sample Space Consider an experiment and let S be the space of possible outcomes. Example: Experiment is tossing a coin; S={h,t} Experiment is rolling a pair of dice: S={(1,1),(1,2),…(6,6)} Experiment is a race consisting of three cars: 1,2 and 3. The sample space is {(1,2,3),(1,3,2),(2,1,3),(2,3,1),(3,1,2),(3,2,1)}

13
Probabilities Let Sample Space S = {1,2,…m} Consider numbers p i is the probability that the outcome of the experiment is i. Suppose we toss a fair coin. Sample space is S={h,t}. Then p h = 0.5 and p t = 0.5.

14
Probability Experiment: Will it rain or not in Sydney : S = {rain, no-rain} P rain = 138/365 =0.38; P no-rain = 227/365 Assigning (or rather how to) probabilities is a deep philosophical problem. What is the probability that the “green object standing outside my house is a burglar dressed in green.”

15
Probability An Event A is a set of possible outcomes of the experiment. Thus A is a subset of S. Let A be the event of getting a seven when we roll a pair of dice. A = {(1,6),(6,1),(2,5),(5,2),(4,3),(3,4) } P(A) = 6/36 = 1/6 In general

16
Probability The sample space S and events are “sets”. P(S) = 1; P(Φ) = 0 Addition: Often Complement:

17
Example Suppose the probability of raining today is 0.4 and tomorrow is also 0.4 and on both days is 0.1. What is the probability it does not rain on either day. S={(R,N), (R,R),(N,N),(N,R)} Let A be the event that it will rain today and B it will rain tomorrow. Then A ={(R,N), (R,R)} ; B={(N,R),(R,R)} Rain at least today or tomorrow: Will not rain on either day: 1 – 0.7 = 0.3

18
Conditional Probability One of the most important concepts in all of Data Mining and Machine Learning P(A|B) = P(AB)/P(B)..assuming P(B) not equal 0. Conditional probability of A given B has occurred. Probability it will rain tomorrow given it has rained today. P(A|B) = P(AB)/(B) = 0.1/0.4 = ¼ = 0.25 In general P(A|B) is not equal to P(B|A)

19
We need conditional probability to answer…. S-LengthS-WidthP-LengthP-WidthFlower SmallMediumSmallMediumA (SetosA) Medium Large O(Versicolor) MediumSmall LargeI (Virginica) Large MediumSmallA LargeSmallMediumSmall? What should the “prediction” be for the flower ?

20
Bayes Rule P(A|B) = P(AB)/P(B); P(B|A) = P(BA)|P(A) Now P(AB) = P(BA) Thus P(A|B)P(B) = P(B|A)P(A) Thus P(A|B) = [P(B|A)P(A)]/[P(B)] This is called Bayes Rule Basis of almost all prediction Latest theories hypothesize that human memory and action is Bayes rule in action.

21
Bayes Rule Prior Posterior

22
Bayes Rule: Example The ASX market goes up 60% of the days of a year. 40% of the time it stays the same or goes down. The day the ASX is up, there is a 50% chance that the Shanghai Index is up. On other days there is 30% chance that Shanghai goes up. Suppose The Shanghai market is up. What is the probability that ASX was up. Define A1 as “ASX is up”; A2 is “ASX is not up” Define S1 as “Shanghai is up”; S2 is “Shanghai is not up” We want to calculate P(A1|S1) ? P(A1) = 0.6; P(A2) = 0.4; P(S1|A1) = 0.5; P(S1|A2) = 0.3 P(S2|A1) = 1 – P(S1|A1) = 0.5; P(S2|A2) = 1 –P(S1|A2) = 0.7;

23
Bayes Rule: Example We want to calculate P(A1|S1) ? P(A1) = 0.6; P(A2) = 0.4; P(S1|A1) = 0.5; P(S1|A2) = 0.3 P(S2|A1) = 1 – P(S1|A1) = 0.5; P(S2|A2) = 1 –P(S1|A2) = 0.7; P(A1|S1) = P(S1|A1)P(A1)/(P(S1)) How do we calculate P(S1) ?

24
Bayes Rule: Example P(S1) = P(S1,A1) + P(S1,A2) [Key Step] = P(S1|A1)P(A1) + P(S1|A2)P(A2) = 0.5 x x 0.4 = 0.42 Finally, P(A1|S1) = P(S1|A1)P(A1)/P(S1) = (0.5 x 0.6)/0.42 = 0.71

25
Example: Iris Flower F=Flower; SL=Sepal Length; SW = Sepal Width; PL=Petal Length; PW =Petal Width Data LargeSmallMediumSmall? choose the maximum

26
Example: Iris Flower So how do we compute: P(Data|F=A) ? This is a a non-trivial question…[subject to much research] How many times does “Data” appear in the “database” when F=A. In this case “Data” is a 4-dimensional “data vector.” Each component takes 3 values (small, medium, large). Thus number of combinations 3^4 = 81.

27
Example: Iris Flower Conditional Independence P(Data|F=A) = P(SL=Large,SW=Small,PL=Medium,PW=Small|F=A) ~= P(SL=Large|F=A)P(SW=Small|F=A)P(PL=Medium|A)P(PW=Small|A) The above is an assumption to make the “computation easier.” Surprisingly evidence suggest that it works reasonably well in practice. This prediction method (which exploits conditional independence) is called “Naïve Bayes Classifier.”

Similar presentations

© 2017 SlidePlayer.com Inc.

All rights reserved.

Ads by Google