Incremental concpetual clustering - reading group discussion
-
Upload
kalpa-gunaratna -
Category
Education
-
view
355 -
download
0
Transcript of Incremental concpetual clustering - reading group discussion
Incremental Conceptual Clustering
Kalpa Gunaratna
Reading group discussions @Kno.e.sis
Based on Fisher’s Cobweb algorithm
Clustering *
• Clustering is the unsupervised classification of patterns into groups.
* Jain, Anil K., M. Narasimha Murty, and Patrick J. Flynn. "Data clustering: a review." ACM computing surveys (CSUR) 31, no. 3 (1999): 264-323.2
3
Focus on hierarchical clustering
• Single link clusteringThe distance between two clusters is the minimum of the distances between all pairs of patterns drawn from the two clusters.
In other words, evaluates dissimilarity between two clusters as the dissimilarity of the nearest patterns, one from each cluster.
• Complete link clusteringThe distance between two clusters is the maximum of all pairs between the two clusters.
In other words, evaluates dissimilarity between two clusters as the greatest distance between any two patterns, one from each cluster.
• Produces compact clusters.
4
• Single link algorithm can extract concentric clusters as shown below whereas complete link cannot.
5
• But single link algorithm suffers from chaining effect as shown below whereas complete link does not have this effect. Therefore, researchers believe complete link gives more useful clusters in real problems.
6
• Dendrogram
7
Our focus – Incremental Conceptual Clustering (Cobweb) 1, 2
Given a set of observations, humans acquire concepts that organize those observations and use them in classifying future experiences. This type of concept formation can occur in the absence of a tutor and it can take place despite irrelevant and incomplete information.
81. Fisher, Douglas H. "Knowledge acquisition via incremental conceptual clustering." Machine learning 2, no. 2 (1987): 139-172.2. Gennari, John H., Pat Langley, and Doug Fisher. "Models of incremental concept formation." Artificial intelligence 40, no. 1 (1989): 11-61.
• Cobweb• Uses a hill climbing search strategy having operators enabling bi-directional
travel in the space.• Hill climbing is a classic AI search method in which one applies all operator instantiations,
compares the resulting states using an evaluation function, selects the best state, and iterates until no more progress can be made.
• Has a function called Category Utility to decide on what action to take in the hill climbing search. • Computes similarity within clusters and dissimilarity between clusters.
9
• Category utility function
• Intra-class similarity is measured by P(Ai=Vij/Ck). - predictability• The larger this probability, the greater the proportion of class members sharing the value
and the more predictable the value is of class members.
• Inter-class similarity is measured by P(Ck/Ai=Vij). - predictiveness• The larger this probability, the fewer the objects in contrasting classes that share this
value and the more predictive the value is of the class.
10
𝑘
𝑖
𝑗
𝑃 𝐴𝑖 = 𝑉𝑖𝑗 𝑃 𝐶𝑘/𝐴𝑖 = 𝑉𝑖𝑗 𝑃 𝐴𝑖 = 𝑉𝑖𝑗/𝐶𝑘
Using Bayes’ theorem
𝑘
𝑃(𝐶𝑘)
𝑖
𝑗
𝑃 𝐴𝑖 = 𝑉𝑖𝑗/𝐶𝑘2
This is the expected number of attribute values that one can correctly guess for an arbitrary member of class Ck.
11
• They further went on to say that CU as the increase in the expected number of attribute values that can be correctly guessed, given a set of n categories, over the expected number of correct guesses without such knowledge.
• Divided by K so that merging, splitting, or adding nodes is taken care of (will discuss now).
12
• There are four main operators in creating the hierarchy.• Classify into an existing class.
• Create a new class.
• Combine two classes into one (merging).
• Divide a class into several classes (splitting).
• Because of the last two operations, this is normally not sensitive to the order of items to be clustered.
13
• Merging
14
• Splitting
15
16
17
• Positive points about Incremental Conceptual Clustering (as I see)• Unsupervised
• Input order does not matter
• Efficient – does not compute similarity/dissimilarity between all pairs/combinations
• Good for dynamic environments
• Bi-directional search space walk in the hierarchy construction
• Try to mimic human categorization behavior
• Clustering is based on probability – not just a similarity score
18