CS246: Mining Massive Datasets Jure Leskovec, Stanford University
http://cs246.stanford.edu
Given a set of points, with a notion of distance between points, group the points into some number of clusters, so that Usually:
Members of a cluster are close/similar to each other Members of different clusters are dissimilar Points are in a high-dimensional space Similarity is defined using a distance measure
Euclidean, Cosine, Jaccard, edit distance,
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
x x x x x x x x x xx x x x x x x xx x x x x x x x xx x x
x x x x x x x x x x
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
Clustering in two dimensions looks easy. Clustering small amounts of data looks easy. And in most cases, looks are not deceiving. Many applications involve not 2, but 10 or 10,000 dimensions. High-dimensional spaces look different: almost all pairs of points are at about the same distance
Jure Leskovec, Stanford C246: Mining Massive Datasets 4
1/29/2012
Assume random points within a bounding box In 2 dimensions:
e.g., values between 0 and 1 in each dimension A variety of distances between 0 and 1.41.
In 10,000 dimensions: The difference in any one dimension is distributed as a triangle The law of large numbers applies Actual distance between two random points is the sqrt of the sum of squares of essentially the same set of differences
Jure Leskovec, Stanford C246: Mining Massive Datasets 5
1/29/2012
A catalog of 2 billion sky objects represents objects by their radiation in 7 dimensions (frequency bands) Problem: Cluster into similar objects, e.g., galaxies, nearby stars, quasars, etc. Sloan Sky Survey is a newer, better version of this
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
Intuitively: Music divides into categories, and customers prefer a few categories Represent a CD by the customers who bought it Similar CDs have similar sets of customers, and vice-versa
But what are categories really?
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
Space of all CDs: Think of a space with one dimension for each customer
Values in a dimension may be 0 or 1 only A movie is a point in this space is (x1, x2,, xk), where xi = 1 iff the i th customer bought the CD
For Amazon, the dimension is tens of millions An alternative: use minhashing/LSH to get Jaccard similarity between close CDs 1 minus Jaccard similarity can serve as a distance
Jure Leskovec, Stanford C246: Mining Massive Datasets 8
Compare with boolean matrix: rows = customers; cols. = CDs
1/29/2012
Represent a document by a vector (x1, x2,, xk), where xi = 1 iff the i th word (in some order) appears in the document Documents with similar sets of words may be about the same topic
It actually doesnt matter if k is infinite; i.e., we dont limit the set of words
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
As with CDs we have a choice when we think of documents as sets of words or shingles:
Sets as vectors: measure similarity by the cosine distance. Sets as sets: measure similarity by the Jaccard distance. Sets as points: measure similarity by Euclidean distance.
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
10
Hierarchical:
Agglomerative (bottom up):
Initially, each point is a cluster Repeatedly combine the two nearest clusters into one.
Divisive (top down):
Start with one cluster and recursively split it
Point assignment:
Maintain a set of clusters Points belong to nearest cluster
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
11
Key operation: Repeatedly combine two nearest clusters Three important questions:
How do you represent a cluster of more than one point? How do you determine the nearness of clusters? When to stop combining clusters?
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
12
Key problem: As you build clusters, how do you represent the location of each cluster, to tell which pair of clusters is closest? Euclidean case: each cluster has a centroid = average of its points
Measure cluster distances by distances of centroids
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
13
(5,3) o (1,2) o x (1.5,1.5) x (1,1) o (2,1) o (0,0) o (4,1) x (4.5,0.5) o (5,0) x (4.7,1.3)
Data: o datapoint x centroid
1/29/2012 Jure Leskovec, Stanford C246: Mining Massive Datasets
Dendrogram
14
The only locations we can talk about are the points themselves
i.e., there is no average of two points
Approach 1: clustroid = point closest to other points
Treat clustroid as if it were centroid, when computing intercluster distances
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
15
Possible meanings of closest:
Smallest maximum distance to the other points. Smallest average distance to other points. Smallest sum of squares of distances to other points.
For distance metric d clustroid c of cluster C is:
min d ( x, c)
c xC
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
16
Approach 2: Intercluster distance = minimum of the distances between any two points, one from each cluster. Approach 3: Pick a notion of cohesion of clusters, e.g., maximum distance from the clustroid
Merge clusters whose union is most cohesive
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
17
Approach 1: Use the diameter of the merged cluster = maximum distance between points in the cluster Approach 2: Use the average distance between points in the cluster Approach 3: Use a density-based approach
Take the diameter or avg. distance, e.g., and divide by the number of points in the cluster Perhaps raise the number of points to a power first, e.g., square-root.
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
18
Nave implementation of hierarchical clustering:
At each step, compute pairwise distances between all pairs of clusters, then merge O(N3)
Careful implementation using priority queue can reduce time to O(N2 log N)
Still too expensive for really big datasets that do not fit in memory
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
19
Assumes Euclidean space/distance Start by picking k, the number of clusters Initialize clusters by picking one point per cluster.
Example: pick one point at random, then k-1 other points, each as far away as possible from the previous points
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
21
For each point, place it in the cluster whose current centroid it is nearest After all points are assigned, fix the centroids of the k clusters. Optional: reassign all points to their closest centroid
Sometimes moves points between clusters
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
22
Reassigned points 4 x 6 7 5 x 3 1 8
Clusters after first round
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
23
How to select k? Try different k, looking at the change in the average distance to centroid, as k increases. Average falls rapidly until right k, then changes little
Best value of k Average distance to centroid
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
24
Too few; many long distances to centroid.
x x x x x x x x x xx x x x x x x xx x x x x x x x xx x x
x x x x x x x x x x
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
25
Just right; distances rather short.
x x x x x x x x x xx x x x x x x xx x x x x x x x xx x x
x x x x x x x x x x
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
26
Too many; little improvement in average distance.
x x x x x x x x x xx x x x x x x xx x x x x x x x xx x x
x x x x x x x x x x
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
27
BFR [Bradley-Fayyad-Reina] is a variant of k-means designed to handle very large (disk-resident) data sets It assumes that clusters are normally distributed around a centroid in a Euclidean space
Standard deviations in different dimensions may vary
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
28
Points are read one main-memory-full at a time. Most points from previous memory loads are summarized by simple statistics To begin, from the initial load we select the initial k centroids by some sensible approach
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
29
Possible initialization strategies of the k cluster centers:
Take a small random sample and cluster optimally Take a sample; pick a random point, and then k1 more points, each as far from the previously selected points as possible (As you will learn in HW2, picking random set of k points does not work too well)
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
30
Discard set (DS): Compression set (CS):
Points close enough to a centroid to be summarized. Groups of points that are close together but not close to any centroid. These points are summarized, but not assigned to a cluster Isolated points
Retained set (RS):
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
31
Points in the RS
Compressed sets. Their points are in the CS.
A cluster. Its points are in the DS.
The centroid
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
32
For each cluster, the discard set is summarized by: The number of points, N The vector SUM, whose ith component is the sum of the coordinates of the points in the ith dimension The vector SUMSQ: ith component = sum of squares of coordinates in ith dimension
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
33
2d + 1 values represent any size cluster Averages in each dimension (the centroid) can be calculated as SUMi /N Variance of a clusters discard set in dimension i is: (SUMSQi /N ) (SUMi /N )2
And standard deviation is the square root of that SUMi = i th component of SUM d = number of dimensions
Q: Why use this representation of clusters?
Jure Leskovec, Stanford C246: Mining Massive Datasets 34
1/29/2012
Processing the Memory-Load of points: Find those points that are sufficiently close to a cluster centroid; Add those points to that cluster and the DS
Use any main-memory clustering algorithm to cluster the remaining points and the old RS
Clusters go to the CS; outlying points to the RS
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
35
Processing the Memory-Load of points (2): Adjust statistics of the clusters to account for the new points.
Add Ns, SUMs, SUMSQs
Consider merging compressed sets in the CS If this is the last round, merge all compressed sets in the CS and all RS points into their nearest cluster
Jure Leskovec, Stanford C246: Mining Massive Datasets 36
1/29/2012
How do we decide if a point is close enough to a cluster that we will add the point to that cluster? How do we decide whether two compressed sets deserve to be combined into one?
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
37
We need a way to decide whether to put a new point into a cluster BFR suggest two ways:
The Mahalanobis distance is less than a threshold Low likelihood of the currently nearest centroid changing
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
38
Normalized Euclidean distance from centroid For point (x1,,xd) and centroid (c1,,cd)
1. Normalize in each dimension: yi = (xi - ci)/i 2. Take sum of the squares of the yi 3. Take the square root , =
2
i standard deviation of points in the cluster in the ith dimension
39
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
If clusters are normally distributed in d dimensions, then after transformation, one standard deviation = Accept a point for a cluster if its M.D. is < some threshold, e.g. 4 standard deviations
i.e., 68% of the points of the cluster will have a Mahalanobis distance <
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
40
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
41
Should 2 CS subclusters be combined? Compute the variance of the combined subcluster Combine if the variance is below some threshold Many alternatives: treat dimensions differently, consider density.
Jure Leskovec, Stanford C246: Mining Massive Datasets 42
N, SUM, and SUMSQ allow us to make that calculation quickly
1/29/2012
Problem with BFR/k-means:
Assumes clusters are normally distributed in each dimension And axes are fixed ellipses at an angle are not OK
Vs.
CURE:
Assumes a Euclidean distance Allows clusters to assume any shape
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
43
h h e e e salary e h h h e e e h h h h e h e e
age
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
44
Pick a random sample of points that fit in main memory 1) Initial clusters: 2) Pick disperse points:
Cluster these points hierarchically group nearest points/clusters For each cluster, pick a sample of points, as dispersed as possible From the sample, pick representatives by moving them (say) 20% toward the centroid of the cluster
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
45
h h e e e salary h e h h h e e e h h h h e h e e
age
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
46
h h e e e salary h e h h h e e e h h h h e h e e
Pick (say) 4 remote points for each cluster.
age
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
47
h h e e e salary h e h h h e e e h h h h e h e e
Move points (say) 20% toward the centroid.
age
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
48
Now, visit each point p in the data set Place it in the closest cluster
Normal definition of closest: that cluster with the closest (to p) among all the sample points of all the clusters.
1/29/2012
Jure Leskovec, Stanford C246: Mining Massive Datasets
49
