Numerical Literal Prediction

[Demirrr]

Interface

from dicee import KGE
# (1) Upload a trained KGE model.
pre_trained_kge = KGE(path='..')
# (2) Train a KGE model on numerical literals
pre_trained_kge.train_literal(path="")

def train_literal(path=""):
      # (3) Read a knowledge graph where all tails are numerical literals
      # (4) For each unique relation (e.g. dataproperty), we introduce a regression problem.
      # (5) Tackle each regression problem individually.
      # ....

Lets say we have a knowledge graph containing only data properties

A height 185
A age 35
B height 180
B age 19

A and B embeddings are learned at step (2). Let's assume A = [0.1, -1.5] and B = [0.9, -0.5]

Lets assume that the embeddings of data properties are defined as such height=[0.0, 0.0] and age=[0.0, 0.0]

yhat = sum( [0.1, -1.5] * [0.0, 0.0] )
y = 185
# compute loss w.r.t.
single_loss = pow( yhat - y, 2)

# Compute the gradient of the loss w.r.t. weights.

Ideally, we should be able to observe

print(pre_trained_kge.predict_literal(h="A", r="height")) # 180
print(pre_trained_kge.predict_literal(h="Demir", r="age")) # 35