Exploring a clustering technique using mathematical group theory

Step I: Identification of ICM

[caption id="attachment_20428" align="aligncenter" width="596" caption="A scatter plot"][/caption]

The x-axis, identified by the base marker, represents numerical values from 1 to 100 (integers). The y-axis, identified by the decimal marker, represents numerical values from 0 to 1 (decimal, 2 significant digits).

From all the data, two points are selected at random, marked in the image by a black rectangle around them. They are labelled as the initial cluster markers (ICM).

Step II: Geometric tagging

The Euclidean distance between each point in the scatter plot and the ICM is calculated. In this case, there are 48 points (after excluding the ICM): x₁, x_2, . . ., x₄₈. Therefore, the distances between x₁ and ICM₁ and ICM₂ are calculated; between x₂ and ICM₁ and ICM₂, and so on until x₄₈ and ICM₁ and ICM₂.

If the distance between some x_i and ICM_a is lower than x_i and ICM_b, then x_i is grouped with ICM_a and tagged as x_i^a. This gives rise to a clear demarcation; the data set is now split amongst x_i^a and x_j^b. Each section is identified as S{x_i,j^a,b}.

Step III: Averaging

Before any average is computed, it must be determined as to which value is to be considered. Three cases are presented below.

1. x-value: if the x-value of each data-point is to be averaged in each cluster, then the computed average will lie along a straight line x_i = µ_i


2. y-value: if the y-value of each data-point is to be averaged in each cluster, then the computed average will lie along a straight line y_i = µ_i


3. Alien value: consider the following table.

Base marker (1-50)	Decimal marker (0-1)	Anonymous marker (1-100)
4	0.71	20
13	0.61	38
14	0.98	7
18	0.03	4
11	0.68	55
37	0.26	67
30	0.04	46
21	0.22	57
37	0.48	90
47	0.94	43

The x- and y-axes represent the values in the first two columns while the third column shows a set of values not considered in the construction of the scatter plot. Since each row in this table is denoted as a point in the plot, each such point is also associated with a certain “anonymous value” as shown in the table.

These can be considered in the averaging process, whereby all the “anonymous” values of all the points in each cluster are averaged separately:

S(x_i^a): A(x_i^a)

S(x_j^b): A(x_j^b)

These new points, A_a and A_b, are plotted in their respective clusters.

Step IV: Convergence

A_a and A_b become the new ICM, and steps I to III are repeated.

The iterations stop when the new A_a and A_b converge with the A_a and A_b from the previous iteration.

Result

After the values have converged, the two S{x_i,j^a,b} now presented by the machine are two distinct groups of data, the emergence of which also signals that the machine has “learnt”. If a very large database is supplied as an input to the machine, there is an advantage as well as a disadvantage.

• Advantage: the machine learns better


• Disadvantage: data may not converge at all due to improper choice of clusters

In order to prevent the second possibility, a suitable number of clusters has to be determined for n, the number of data-points. As a rule of thumb,

k = (n / 2)^1/2

Here, ‘k’ is the number of clusters.

If a dataset of 50 points is input, then k = 5 clusters are suggested. Once the iterations have been performed and convergence has been attained in five separate clusters, the centroids of each cluster (or, the average value of the data-points in each cluster) can now be used to arrive at 2 super-clusters.

Exploring a clustering technique using mathematical group theory

Step I: Identification of ICM

[caption id="attachment_20428" align="aligncenter" width="596" caption="A scatter plot"][/caption]

The x-axis, identified by the base marker, represents numerical values from 1 to 100 (integers). The y-axis, identified by the decimal marker, represents numerical values from 0 to 1 (decimal, 2 significant digits).

From all the data, two points are selected at random, marked in the image by a black rectangle around them. They are labelled as the initial cluster markers (ICM).

Step II: Geometric tagging

The Euclidean distance between each point in the scatter plot and the ICM is calculated. In this case, there are 48 points (after excluding the ICM): x₁, x_2, . . ., x₄₈. Therefore, the distances between x₁ and ICM₁ and ICM₂ are calculated; between x₂ and ICM₁ and ICM₂, and so on until x₄₈ and ICM₁ and ICM₂.

If the distance between some x_i and ICM_a is lower than x_i and ICM_b, then x_i is grouped with ICM_a and tagged as x_i^a. This gives rise to a clear demarcation; the data set is now split amongst x_i^a and x_j^b. Each section is identified as S{x_i,j^a,b}.

Step III: Averaging

Before any average is computed, it must be determined as to which value is to be considered. Three cases are presented below.

1. x-value: if the x-value of each data-point is to be averaged in each cluster, then the computed average will lie along a straight line x_i = µ_i


2. y-value: if the y-value of each data-point is to be averaged in each cluster, then the computed average will lie along a straight line y_i = µ_i


3. Alien value: consider the following table.

Base marker (1-50)	Decimal marker (0-1)	Anonymous marker (1-100)
4	0.71	20
13	0.61	38
14	0.98	7
18	0.03	4
11	0.68	55
37	0.26	67
30	0.04	46
21	0.22	57
37	0.48	90
47	0.94	43

The x- and y-axes represent the values in the first two columns while the third column shows a set of values not considered in the construction of the scatter plot. Since each row in this table is denoted as a point in the plot, each such point is also associated with a certain “anonymous value” as shown in the table.

These can be considered in the averaging process, whereby all the “anonymous” values of all the points in each cluster are averaged separately:

S(x_i^a): A(x_i^a)

S(x_j^b): A(x_j^b)

These new points, A_a and A_b, are plotted in their respective clusters.

Step IV: Convergence

A_a and A_b become the new ICM, and steps I to III are repeated.

The iterations stop when the new A_a and A_b converge with the A_a and A_b from the previous iteration.

Result

After the values have converged, the two S{x_i,j^a,b} now presented by the machine are two distinct groups of data, the emergence of which also signals that the machine has “learnt”. If a very large database is supplied as an input to the machine, there is an advantage as well as a disadvantage.

• Advantage: the machine learns better


• Disadvantage: data may not converge at all due to improper choice of clusters

In order to prevent the second possibility, a suitable number of clusters has to be determined for n, the number of data-points. As a rule of thumb,

k = (n / 2)^1/2

Here, ‘k’ is the number of clusters.

If a dataset of 50 points is input, then k = 5 clusters are suggested. Once the iterations have been performed and convergence has been attained in five separate clusters, the centroids of each cluster (or, the average value of the data-points in each cluster) can now be used to arrive at 2 super-clusters.

A simple connect between feedback systems and fuzzy logic controllers

Click on the image for a larger view.

The following picture details the three ways to depict the transformation from fuzzy logic to crisp logic. The processual representation (the flowchart in the middle) shows the positive feedback loop that the fuzzy logic controller uses to make the dataset crisp ("p.c." in the chart stands for "position controller").



 

The algorithm, shown leftmost, delineates the logical statements that constitute the following flowchart. The purpose of the positive feedback loop is served by the nested if-then clusters that, going by the graphics on the first row, ensure that intuitive conclusions (as represented by the fuzzy set) are brought as close to the real thing (as represented by the crisp set) as possible using statistical data. The equation on the rightmost determines the mathematical convergence, i.e., the minimum number of line segments that would have to be perfectly aligned for any following segments to just fall in place after them (ref: German tank problem).

A simple connect between feedback systems and fuzzy logic controllers

Click on the image for a larger view.

The following picture details the three ways to depict the transformation from fuzzy logic to crisp logic. The processual representation (the flowchart in the middle) shows the positive feedback loop that the fuzzy logic controller uses to make the dataset crisp ("p.c." in the chart stands for "position controller").



 

The algorithm, shown leftmost, delineates the logical statements that constitute the following flowchart. The purpose of the positive feedback loop is served by the nested if-then clusters that, going by the graphics on the first row, ensure that intuitive conclusions (as represented by the fuzzy set) are brought as close to the real thing (as represented by the crisp set) as possible using statistical data. The equation on the rightmost determines the mathematical convergence, i.e., the minimum number of line segments that would have to be perfectly aligned for any following segments to just fall in place after them (ref: German tank problem).

BSODYOUFUCKINGPUTA!

Microsoft, Inc., should be shut down simply because they continue to manufacture an operating system that fails for no particular reason right when I'm in the middle of something important. With annual investments as well as turnover in the billions of dollars, why have they not been able to eliminate any possibility of the Blue Screen Of Death (BSOD)?

It doesn't help that MS Office 2007-onward has an auto-save feature: if they have to deploy something like that, they've to do it right. If the mechanism has been included for convenience's sake, then I don't want to be shown multiple copies of the file and asked which one I want to view or delete depending on the timestamp. If it has been included as a compensation against the BSOD, then the auto-save function is valuable only if it automatically saves the once every minute and not once every three minutes. With a typing speed of 75 words per minute, I lose 150 words right there.

If anything, they should just dump the auto-save functionality and focus on making an OS that's as big but not as critically inconsistent! I have always taken the romantic path to computing experience and have consciously rejected the Mac at all points of time for its oversimplification of concepts, but the BSOD is not something I'm willing to tolerate because it has a direct and frustrating impact on my work.

Many people like the OS and some can make do with the absence of a sensible design philosophy, but this algorithmic vortex is bound to lose MS some customers when, as competition builds, all that stands between one OS and the other is the BSO-fucking-D.

BSODYOUFUCKINGPUTA!

Microsoft, Inc., should be shut down simply because they continue to manufacture an operating system that fails for no particular reason right when I'm in the middle of something important. With annual investments as well as turnover in the billions of dollars, why have they not been able to eliminate any possibility of the Blue Screen Of Death (BSOD)?

It doesn't help that MS Office 2007-onward has an auto-save feature: if they have to deploy something like that, they've to do it right. If the mechanism has been included for convenience's sake, then I don't want to be shown multiple copies of the file and asked which one I want to view or delete depending on the timestamp. If it has been included as a compensation against the BSOD, then the auto-save function is valuable only if it automatically saves the once every minute and not once every three minutes. With a typing speed of 75 words per minute, I lose 150 words right there.

If anything, they should just dump the auto-save functionality and focus on making an OS that's as big but not as critically inconsistent! I have always taken the romantic path to computing experience and have consciously rejected the Mac at all points of time for its oversimplification of concepts, but the BSOD is not something I'm willing to tolerate because it has a direct and frustrating impact on my work.

Many people like the OS and some can make do with the absence of a sensible design philosophy, but this algorithmic vortex is bound to lose MS some customers when, as competition builds, all that stands between one OS and the other is the BSO-fucking-D.