Performing Inference and Parameter Learning

[rodrigoazs]

Hi,

I intend to learn weights for given possible paths and perform inference with these paths. Does the code provide that? If not, do I need to create my own feature matrix and fill its values using PRA or SFE random walks?

In addition, I'm quite confused about the metrics used in the paper (MAP and MRR) and what relevant entities would be in this context. Metrics as log-likelihood can't be calculated?

Thank you.

[matt-gardner]

You'll have to write your own entry point in the code to do this, but the PRA code will let you find random walk probabilities for a given set of paths. For SFE, there's no mechanism for doing a search given a set of paths, it just does a single BFS to find all paths of a given length. You could feasibly just filter out the feature matrix that SFE computes to only include the paths that you want, though that'd be inefficient, as you're doing a full BFS just to throw away most of it. Writing code to search a graph for a given set of paths isn't very hard; you might be better off just writing your own code for that.

For MAP and MRR, we're evaluating a ranked list of answers, where some answers are correct and some are incorrect. These are standard metrics for evaluating ranked lists. I'm not sure what you even mean by "log-likelihood" as a metric.