Keith's Ph.D. Research
About
Documents
Experiments
·
·
·
                                          
Printer Friendly Version
References

Intro ] [ Stuart Floyd MS Thesis Final.pdf ] [ Stuart Floyd Thesis Presentation.pdf ] [ Stuart Floyd Thesis Presentation Notes ]

Up: Documents ]

2007-09-27
----------

General note:

While reading this presentation it occurred to me that the 
discretization of attributes might be made less sensitive
if bins were allowed to overlap. that is, let a value be
a member of more than one bin. 1-6 5-10 5 and 6 would belong
to both bins.

page 4
Contains best diagrams of Whipple Procedure I've found to date.

page 5
input data listed includes Medical History, Diagnostic Tests, and 
Treatment. All these things occur in time. I would be surprised to
find no temporal information of use. Looks like the main obstacle
would be getting dates for that data.

page 6
This page states there are 60 patients in the data set. These are
the patients for whom the total life expectancy is known. That is
to say those who are known to be deceased. It might be worth adding
the ~30 other patients who may still be alive with a "at least"
measurement of life expectancy using the last known date of contact.

In the current data set there are classifications of imaging 
results. Would there be anything to gain by inspecting the images
directly using machine vision techniques? Carolina doesn't think
so and would prefer to not try. 

page 13
Things to add to the list of things I have to learn about:
   Machine Learning Algorithms: Support Vector Machines
   Evaluation Techniques: Attribute Selected Classifier, ROC Curves
   Feature Selection: Gain Ratio, Principle Components, ReliefF, 
      Support Vector Machines
   Meta Learning: Model Selector by Stuart Floyd

page 35
I wonder why Support Vector Machines needed 10 more attributes
than Logistic Regression to achieve similar results. Is this
evidence that those ~10 attributes have useful information
that Logistic Regression does not use to its advantage? I have
to see what attributes in each model where used to see where
the overlap was and was not.

page 53
I suppose after I look into predicting survival time I'll attempt
quality of life as suggested by the future work.

I have to see what data is contained in the larger national databases
and see if any of it can be leveraged.

I wonder if UMass has added patients to their database since this 
work.
Intro
A Comparison Of Unsupervised Dimension Reduction Algorithms For Classification Notes
Beck AT 1961.pdf
Carolina Ruiz Bagging Boosting.ppt
Distance Preserving Dimension Reduction For Manifold Learning Notes
FeatureSelectionForMachineLearning-ComparingACorrelation-basedFilterApproachToTheWrapper.pdf
Hayward Thesis.pdf
Hayward Thesis Notes
Matrices Vector Spaces And Information Retrieval Notes
Shivin Misra thesis final.pdf
Stuart Floyd MS Thesis Final.pdf
Stuart Floyd Thesis Presentation.pdf
Stuart Floyd Thesis Presentation Notes

by: Keith A. Pray
Last Modified: October 11, 2007 11:46 PM
© 2007 - 2007 Keith A. Pray.
All rights reserved.

Current Theme: 

Kapowee Hosted | Kapow Generated in 0.008 second | XHTML | CSS | Section 508