Multi-stage programming model with K-means, EM, LinearReg, and LogisticReg

[kijungs]

Related issues: #14, #16 , #17

Other algorithmic improvements:

• Users can specify the number of clusters
• Initial centroids are assigned randomly
• The amount of data broadcast and gathered is reduced

Result of the correctness test:

1. K-Means and EM

• Data: sample_cluster
• Actual number of clusters: 3
• Actual means of clusters: {1,3,5}, {6,4,2}, {3,5,7}
• Actual (diagonal) covariances of clusters: {1,2,3}, {3,2,1}, {2,2,2}
• Means obtained by K-means: {1.09,2.55,4.70}, {6.33,3.90,2.00} , {2.89,5.23,7.17}
• Means obtained by EM: {1.028, 2.565, 4.839}, {6.145, 3.864, 2.000}, {2.863, 5.088, 7.015}
• Covariances obtained by EM: {0.809, 1.683, 2.951}, {3.443, 1.566, 0.972}, {2.472, 1.874, 1.869}

2. Logistic Regression

• Data: sample_classification
• Actual model: 0.5 - 3x1 + 2x2 -1.5x3
• Actual accuracy: 90%
• model obtained by Logistic Regression: 0.0007-0.006x1+0.0036x2-0.0027x3 (only ratio matters)
• accuracy obtained by Logistic Regression: 87%

3. Linear Regression

• Data: sample_regression
• Actual model: 0.5 - 3x1 + 2x2 -1.5x3 + gaussian random error with 1.0 standard deviation
• model obtained by Linear Regression: 0.5-2.6x1+1.7x2-1.2x3

[bgchun]

Is an init stage a BSP algorithm? This sounds strange.

Later we have an asynchronous execution engine, we need to separate out the BSP-specific part from the general part.

[bgchun]

@kijungshin @jsjason The code looks nicer than before.

I left non-trivial amount of comments.
There are several issues to be addressed. It's not desirable to address all of them in this pull request.
I suggest the following. Add proper license headers, year, comments, and address coding style comments (e.g., line spacing) in this pull request.
Address the comments about code that can be easily addressable.
For the rest of comments that need to be addressed, we create issues and keep track of them.

Let's address the ones you can do quickly and merge this gigantic pull request ASAP.

[kijungs]

@bgchun @jsjason I will address trivial comments and create issues for non-trivial comments.

[kijungs]

I removed unnecessary empty lines in every code according to the sun java coding convention.
I also added a license header to every code.

[bgchun]

@jsjason @kijungshin If you agree that minor comments are addressed and major comments are registered as issues, please merge this pull request. This has been long overdue.

[jsjason]

@bgchun We agree that this is taking much longer than expected. Although @kijungshin has addressed most of the comments, there is still a few more he'd like to address; we decided to create a separate issue regarding the remaining comments, since almost all of them are about lacking explanations, renaming classes, etc. We'll get the job done shortly.

[jsjason]

All comments have been either addressed or marked as a separate issue. I'll merge this.