SMLG (Statistical Machine Learning Group) Discussion Forum

Probabilistic Graphical Models, Fall 2021. Please post your weekly proposed outcome here every Friday.

Here is the script for Step One and Two from my project. So basically, the data is cleaned and normalized. Top 10 results related to ZRvsZD10days recovered will be uploaded into the BN network.

Hi Dr. Yoo,

Here is the merged and cleaned data I have for now. It has 3882 obs and 139 variables. Thank you.

Update for Project

The data has been discretized and the model network has been constructed ( Naive Bayesian Network).
Step 3 and 4 have been modifed from the proposal.
Step 3 : Data was discretized using -1 (downregulated), not significant (0), 1( upregulated). All datapoints had a logfold change greater than |2| and adjusted p-vale of less than 0.05.
Step 4: I did not use this step as it is necessary for bigger networks because this network only has the top 15 genes, it is not necessary. I completed step 5 instead which was to create a model network. I created a Naive Bayes Network around the gene Dzip3.

Here is an image of the proposed network

This week I log transformed the data and also discretized it. Now I have both continuous and discrete data ready for analysis. I tried a network with fast.iamb and it returned a network with some nodes having no arcs. I will keep on improving the network next week. Thank you.

Attached is the metadata and OTU table for the 16S rRNA data, both downloaded from the Human Microbiome Project (https://ibdmdb.org/tunnel/public/summary.html)

I have started looking into my outcome variable (symptoms of inflammatory bowel disease) and the various sleep measures available in the data (8 measures). The attached PowerPoint summarizes these data, as well as lists decisions that have been made and items that I plan to work on in the upcoming week.

This week I created more models and call equations. I also meaure arc strengh and and calculated each models network score.

Next week I will run the both exact and aproximate inferance, which keeps me on target for the proposal.

This week I created two average networks with Hillclimbing and tabu, and then simplified them with threshold = 0.85. Two networks are slightly different. It is interesting to see that the arcs are mostly from disease to metals, not from metals to disease. I am gonna fit parameters.

Attached is my updated proposal, an updated PPT, and a GeNIe file for my initial BN. I learned the BN structure using the hill-climbing algorithm based on the log likelihood score and setting the maximum number of parents per node to 3. I plan to look into increasing the maximum number of parents per node to see if this improves the model.