Identifying Similar Models

[kuncao0228]

In Phase 1 for the NLP project, we identified key concepts from a range of models of epidemiology and grouped them. For a method of identifying similar models, I am proposing that we create graph representations of model to key concepts that relate only to that specific model, for all the models in our code base. We can then compare vector space similarity of the concepts between models. The user in this case would input (model that exist in code base, K value for KNN) to obtain a list of similar models.

The idea here is that given an example of a set of markov models, the entities that get created from referenced text would ideally be : probability, states, action, random, other model specific attributes. The models would then be grouped based on these entity attributes that they share.

Information from extracted code can be used to eliminate extraneous concepts to a certain degree, but key concepts like "random" in the markov example that is unlikely to be mentioned in code would need to be preserved.

[jpfairbanks]

So this would take the olog representation of the formulation and map each box to a concept in the knowledge graph? The mapping is based on the similarity of the textual names using a deep NLP similarity model.

Then we could look at the sub graphs of overlapping concepts and compare them as graphs. I think this would be a pretty good way of comparing ologs.