Stabilise learning to predict relations

[jmsfltchr]

Problem to Solve

During relation prediction we see that the focus of the algorithm is to correctly classify all nodes and edges. Of these there are typically far more pre-existing nodes and edges that we don't care about than that we do.

Current Workaround

None

Proposed Solution

Possible options:

• Update the loss function of the learner
• Propagate preexistence through the neural network to reduce the burden on the network
• Filter out Unnecessary nodes in the decoder stage of the network, prior to calculating loss
• Reduce the number of output classes from [preexists, does not exist, does exist] to [does not exist, does exist]

[jmsfltchr]

Solved by #101. Note that the number of output classes is still 3, although the loss function now ignores preexists