Technical Journal - Spectral Clustering for Matlab

       After our meeting with Ethan, Dr. Hassibi's graduate student, we were tasked to implement spectral clustering on the adjacency matrices dataset. In order to determine the algorithm's integrity, we have to use false positives and false negatives. As mentioned, the first step was to clean the data and use spectral clustering on it. We soon realized that the spectral clustering function would only work on R2019b. We downloaded R2019b and it recognized spectralcluster as a function. During the weekend, I did a bit of self-study and tried to teach myself MATLAB syntax which I found similar to Python 3.8. Below is a video that I found helpful.

The error displayed in the console when we tried to use spectralcluster on R2019a



       The first step is to load in the dataset which is set as a struct, a block of memory with physically grouped variables. Assign observed and rawAdj to CAdj and Adj respectively to then entrywise multiply the two to Adj1. We use the eye function to get an n-by-n matrix with ones on major diagonals and zeroes elsewhere. To get a fullstorage matrix, the full function must be used. Then, spectralcluster is used on the Laplacian of the graph-form of gAdj and includes 3 for the number of classes there are. To get our results, we use a confusionmat with the ground truth.

Code for spectral clustering

       The program did not work. By that I mean it had a very poor accuracy score. For some reason, in contrast to the time when we used spectral for the Karate Club dataset, it did a horrendous job, leading me to believe that some information was lost in the process and threw everything off. As mentioned before, we used a confusion matrix to show results. The thought that "false positives and false negatives works well for binary classification problems" seemed reasonable but the confusion matrix can quantify the results of an infinite amount of clusters. The rows are the true class and the columns are the predicted class. Ideally, only the diagonals should have the number of nodes classified as such but in our case, all of them were classified as 1. As a result, we had poor results.
Confusion map used to show results

       In conclusion, while spectral clustering had been unsuccessfully implemented, we found out about our new goals for the next few weeks. As I mentioned before, minimum cut could have some potential, despite the dataset being super large. It is worth the shot if we cannot get good results from spectral. I am also curious as to what Dr. Ramya, the Ph.D. student who provided the dataset, did to correctly predict the dataset. Below is a video of minimum cut:


Comments

Post a Comment

Popular posts from this blog

Technical Journal - Research Papers

Image
On the Efficiency of Data Collection for Crowdsourced Classification :        The paper aims to explain the accuracy gaps between popular collection policies (state-action mapping) on crowdsourced data. The quality of crowdsourced data is often highly variable so it is suspected that it is the cause. However, studies show that the policies used to collect such data have a strong impact on the accuracy of the system. The first theoretical explanation of the accuracy gaps is that the non-adaptive uniform allocation, and the adaptive uncertainty sampling and information gain maximization. Done with the representation of the collection process in terms of random walks in the log-odds (log of odds of guessing correctly/incorrectly) domain, the author derives lower and upper bounds on the accuracy of the policies. The bounds accurately tell us by how much do two adaptive policies trump a non-adaptive one. It is believed that techniques used by them are applicable to addi...
Read more

Journal

Image
On Tuesday, March 24th, we had a meeting with Professor Hassibi over Zoom. Prior to our talk, we had issues with our code regarding the fact that it didn't produce the results expected. He told us to find and graph the eigenvalues of the graph Laplacian, as that would indicate the number of clusters found and value you can refer other datapoints to. He noted that graphing the corresponding eigenvector would indicate whether a cluster had been isolated or not. On that note, it was up to us to tinker with values for a bit and try different algorithms like K-means. Below is a video of the meeting with Professor Hassibi: This week has been very interesting, to say the least. As you may know, COVID-19 has just taken the U.S. for a turn and has put a pause on daily life, forcing businesses to adapt, overcome, or fail in some cases. For us high schoolers, classes have migrated to the Zoom platform. With it followed teachers with bad internet service, leading to indecipherable dialect...
Read more

First Month

Image
The first four weeks of this Caltech Stem class have been interesting, to say the least. Much of the material we went into have been all over the place with various objectives. We started off with teamwork exercises, watching college students present their research, replicating one for ourselves, writing a concept map, and going to Caltech to meet Dr. Hassibi. We had two, teamwork workshops for the first two weeks. One of which was the dominoes problem where each student was given two dominoes and we were instructed to get three dominoes per person by exchanging dominoes for chips and chips for dominoes. Instructions were very vague other than that everyone must have a domino to "survive" a round. We struggled in the first few attempts but through several days of discussion and trial & error, we were able to find a way to accomplish our goal. Some key underlying goals of this exercise were the importance of leadership, communication, and open-mindedness...
Read more