Ramesh Oswal

Clustering:

Fully Unsupervisd Algo

Automatically partition unlabeled data points  into groups of data points

Useful for

              Organizin data

              Understanding Hidden Structure

 Appliations:

              Cluster NewsArticles, Web Pages or search results by topic.

              Cluster Protein sequences by func or genes according to expression profile.

              Cluster user of Social network based on intrest(Commuity detection) - Ad target

Optimization Background:

              Coordinate Descent:

                             Goal - Minimize a function J(theta)

                                            eg. theta = argmin(J(theta))

                             Idea: Pick one dimension, and minimize along that dimension.

                                                                Algorithm:          

1.      Choosing Initial Point heta

2.      Repeat until stopping criteria is reached

a.      Theta1 = argmin J(theta1,theta2,….thetan)

b.      Theta2 = argmin J(theta1,theta2,….thetan)

c.      Theta1 = argmin J(theta1,theta2,….thetan)

.

.

d.      Thetan = argmin J(theta1,theta2,….thetan)

Note: Steps abc..,d are exact line search alo ng same axis.

                              In some cases this algo can get stuck and start oscillating along two points. In this scenario we can use something called as Block Coordinate Descent.

              Block Coordinate Descent:

                                                           Here:  An eg. With two block alpha and beta where theta = [alpha, beta]

                                                           Goal: alpha,beta = argmin J(alpha,beta) where alpha in Ra and beta in Rb

                                                                               Idea: Minimize over an entire group of variables at a time.

                                                           Algo:

1.)    Choose alpha and beta

2.)    Repeat until stopping criteria is reached

a.      Alpha – argmin (J(alpha,beta))

b.      Beta – argmin (J(alpha,beta))              

   Clustering:

Left figure - Distance between two points in same cluster is less than different cluster

Right Figure - Distance between two points in same cluster is more than different cluster

             Now we define a Object function to minimize.

What is clustering? Goal is to partition unlabeled instances into group of similar points.

Input: Unlabeld data D = {X1,X2,X3….Xn}, X(i) belongs Rm

              *We do not know the labels!

Output:

View#1: Labeled Instances {(X(1),Z(1)), (X(2),Z(2)), (X(3),Z(3))….. (X(n),Z(n))}

                                                                                         Where Z(1) E {1,2,3,….,K} k is # of clusters

View#2: Clusterings: C1,C2,…..Ck where k is number of clusters Ci = {X(i): Z(i)=j} pints in jth partition

Questions:

1.)    How many clusters are there?

2.)    How do we define similarity between points? Eg. Eucledian distance

Object based clusteting:

              Eg K-Means Objective -    

                             Input: D = {Xi}i=1N

                             Cluster Centers: {c1,c2,…,ck} = c

Decision Rule: Assign point X(i) to its nearest cluster center cj

Objective:

              C= argmin sumi=1N minj in {1,..,k}||X(i)-cj||2

Graphical Representation used for EDA.(List is incomplete)

Recently, Me along with my friends from CMU participated in Steel Hacks 2017. We build a mobile application which identifies user activity using Android Accelerator and plays song from Spotify based on activity and user preferences. 

I worked on the User Activity Recognition part where I used Android’s Google Accelerator api to receive the co-ordinates of Phone Movement. Based on this we calculated the speed with which the user is moving and categorized it into different modes like: Still, Slow Walking Brisk Walking, Jogging and Sprint.

Now Based on these different modes we pulled the Users Spotify details and played the song based on the activity and his preferences.

For more details please go to the below dev post link.