top of page

FastPole Algorithm

Data Science Techniques for Discovering Dynamic Brain Activity

Project Abstract

fmri-scanss_1024.jpg

In this project, we use data science techniques to discover linear correlation strength between more than two time series fMRI data. Julia language is used to perform the data mining process and MATLAB is used to visualize result data. Based on Clique Based Multipole Search algorithm, a new algorithm named FastPole is proposed. FastPole is proved to promote the calculation efficiency and reduce the repetitive linear characteristic computations. Further, FastPole shows significant advantages on stability and accuracy when dealing with large size fMRI data.

Previous Method Review

Saurabh Agrawal and etc. developed a method called Clique Based Multipole Search (CoMEt).

CoMEt searches for multipoles by restricting the limit on the linear gain as well as dependence for promising candidates.  

 

The idea of Clique Based Multipole Search is to find promising candidates using correlation graph, and then check all the promising candidates if they satisfy the threshold, therefore find all multipoles and store them in a set U. After finding all of the multipoles, the CoMEt algorithm removes the duplicates and non-maximal multipoles.

However, this may result in long-time calculation if the set size is large. It is because the algorithm takes every promising candidate into consideration and judgement even though the only dependence itself is less than the threshold; and then calculate the linear dependence and linear gain each time before the verification.

 

This also results in repetitive computation on subsets’ linear gain as well as dependence, and on sets that their family sets already satisfy the threshold.

Proposed Solution

FastPole finds multipoles from maximal-cliques and sub-cliques. The key point of FastPole is that dependences of all sets are collected in tuples and utilized for different purposes. The dependence of a sub-clique can be used for calculating gain of its supersets, and it can also be used for determining whether the sub-clique is a multipole. Moreover, in FastPole, we don’t consider subsets of a multipole and thus reduce amount of calculation. 

Proposed preliminary features

  •  Flow path of multipole selection is optimized in FastPole since the redundant evaluation process is removed.

  • Computing methods for linear gain is improved in FastPole since sorted linear dependence sequence is utilized.

  • Calculation is significantly reduced in FastPole , especially for situations of high complexity.

Team members

个人照3.jpg

Electrical Engineering

Zhiqing Wang

Responsible for algorithm compiling, website making, testing and optimization

c2.jpg

Electrical Engineering

Siyi Chen

Responsible for algorithm compiling, weekly report, data visualization, team management and poster making

l.jpg

Electrical Engineering

Qian Liu

Responsible for algorithm compiling, weekly report, poster making and algorithm structure

a.jpg

Project Advisor

Atluri Gowtham

Providing feedback and guidance 

zhiyuan.jpg

Ph.D. Student

Zhiyuan Li

Providing idea and guidance on the project

bottom of page