Mean shift and Hierarchical clustering
4 likes4,474 views
Mean shift clustering finds clusters by locating peaks in the probability density function of the data. It iteratively moves data points to the mean of nearby points until convergence. Hierarchical clustering builds clusters gradually by either merging or splitting clusters at each step. There are two types: divisive which splits clusters, and agglomerative which merges clusters. Agglomerative clustering starts with each point as a cluster and iteratively merges the closest pair of clusters until all are merged based on a chosen linkage method like complete or average linkage. The choice of distance metric and linkage method impacts the resulting clusters.
Mean shift and Hierarchical clustering
- 1. Clustering for new discovery in data Mean shift clustering Hierarchical clustering - Kunal Parmar Houston Machine Learning Meetup 1/21/2017
- 2. Clustering : A world without labels • Finding hidden structure in data when we don’t have labels/classes for the data • We group data together based on some notion of similarity in the feature space
- 3. Clustering approaches covered in previous lecture • k-means clustering o Iterative partitioning into k clusters based on proximity of an observation to the cluster mean
- 4. Clustering approaches covered in previous lecture • DBSCAN o Partition the feature space based on density
- 5. In this segment, Mean shift clustering Hierarchical clustering
- 6. Mean shift clustering • Mean shift clustering is a non-parametric iterative mode-based clustering technique based on kernel density estimation. • It is very commonly used in the field of computer vision because of it’s high efficiency in image segmentation.
- 7. Mean shift clustering • It assumes that our data is sampled from an underlying probability distribution • The algorithm finds out the modes(peaks) of the probability distribution. The underlying kernel distribution at the mode corresponds to a cluster
- 8. Kernel density estimation Set of points KDE surface
- 9. Algorithm: Mean shift 1. Define a window (bandwidth of the kernel to be used for estimation) and place the window on a data point 2. Calculate mean of all the points within the window 3. Move the window to the location of the mean 4. Repeat step 2-3 until convergence • On convergence, all data points within that window form a cluster.
- 14. Types of kernels • Generally, a Gaussian kernel is used for probability estimation in mean shift clustering. • However, other kinds of kernels that can be used are, o Rectangular kernel o Flat kernel, etc. • The choice of kernel affects the clustering result
- 15. Types of kernels • The choice of the bandwidth of the kernel(window) will also impact the clustering result o Small kernels will result in lots of clusters, some even being individual data points o Big kernels will result in one or two huge clusters
- 16. Pros and cons : Mean Shift • Pros o Model-free, doesn’t assume predefined shape of clusters o Only relies on one parameter: kernel bandwidth h o Robust to outliers • Cons o The selection of window size is not trivial o Computationally expensive; O(𝑛2 ) o Sensitive to selection of kernel bandwidth; small h will slow down convergence, large h speeds it up but might merge two modes
- 17. Applications : Mean Shift • Clustering and segmentation • dfsn
- 18. Applications : Mean Shift • Clustering and Segmentation
- 19. Hierarchical Clustering • Hierarchical clustering creates clusters that have a predetermined ordering from top to bottom. • There are two types of hierarchical clustering: o Divisive • Top to bottom approach o Agglomerative • Bottom to top approach
- 20. Algorithm: Hierarchical agglomerative clustering 1. Place each data point in it’s own singleton group 2. Iteratively merge the two closest groups 3. Repeat step 2 until all the data points are merged into a single cluster • We obtain a dendogram(tree-like structure) at the final step. We cut the dendogram at a certain level to obtain the final set of clusters.
- 21. Cluster similarity or dissimilarity • Distance metric o Euclidean distance o Manhattan distance o Jaccard index, etc. • Linkage criteria o Single linkage o Complete linkage o Average linkage
- 22. Linkage criteria • It is the quantification of the distance between sets of observations/intermediate clusters formed in the agglomeration process
- 23. Single linkage • Distance between two clusters is the shortest distance between two points in each cluster
- 24. Complete linkage • Distance between two clusters is the longest distance between two points in each cluster
- 25. Average linkage • Distance between clusters is the average distance between each point in one cluster to every point in other cluster
- 26. Example: Hierarchical clustering • We consider a small dataset with seven samples; o (A, B, C, D, E, F, G) • Metrics used in this example o Distance metric: Jaccard index o Linkage criteria: Complete linkage
- 27. Example: Hierarchical clustering • We construct a dissimilarity matrix based on Jaccard index. • B and F are merged in this step as they have the lowest dissimilarity
- 28. Example: Hierarchical clustering • How do we calculate distance of (B,F) with other clusters? o This is where the choice of linkage criteria comes in o Since we are using complete linkage, we use the maximum distance between two clusters o So, • Dissimilarity(B, A) : 0.5000 • Dissimilarity(F, A) : 0.6250 • Hence, Dissimilarity((B,F), A) : 0.6250
- 29. Example: Hierarchical clustering • We iteratively merge clusters at each step until all the data points are covered, i. merge two clusters with lowest dissimilarity ii. update the dissimilarity matrix based on merged clusters o sfs
- 30. Dendogram • At the end of the agglomeration process, we obtain a dendogram that looks like this, • sfdafdfsdfsd
- 31. Cutting the tree • We cut the dendogram at a level where there is a jump in the clustering levels/dissimilarities
- 32. Cutting the tree • If we cut the tree at 0.5, then we can say that within each cluster the samples have more than 50% similarity • So our final set of clusters is, i. (B,F), ii. (A,E,C,G) and iii. (D)
- 33. Final set of clusters
- 34. Impact of metrics • The metrics chosen for hierarchical clustering can lead to vastly different clusters. • Distance metric o In a 2-dimensional space, the distance between the point (1,0) and the origin (0,0) can be 2 under Manhattan distance, 2 under Euclidean distance. • Linkage criteria o Distance between two clusters can be different based on linkage criteria used
- 35. Linkage criteria • Complete linkage is the most popular metric used for hierarchical clustering. It is less sensitive to outliers. • Single linkage can handle non-elliptical shapes. But, single linkage can lead to clusters that are quite heterogeneous internally and it more sensitive to outliers and noise
- 36. Pros and Cons : Hierarchical Clustering • Pros o No assumption of a particular number of clusters o May correspond to meaningful taxonomies • Cons o Once a decision is made to combine two clusters, it can’t be undone o Too slow for large data sets, O(𝑛2 log(𝑛))
- 37. References i. https://spin.atomicobject.com/2015/05/26/mean- shift-clustering/ ii. http://vision.stanford.edu/teaching/cs131_fall1314 _nope/lectures/lecture13_kmeans_cs131.pdf iii. http://84.89.132.1/~michael/stanford/maeb7.pdf
- 38. Thank you!