Deep Learning-based Ensemble Method for Temporal Knowledge Graph Embedding in Link Prediction.

Three experiments are conducted in this project.

• Each individual TKGE model is re-predicted based on the same dataset as other experiments, namely DE-SimplE, DE-DistMult, DE-TransE, ATiSE, and TERO.
• To demonstrate that the neural network can learn better weights than other methods, conducted the grid search experiment as the baseline method.
• Conducted two neural network experiments with different numbers of inputs. For 5 inputs, it took the simulated score of the top 1 prediction for each TKGE model, and for 25 inputs, it took the simulated score of the top 5 predictions for each TKGE model.

TKGE Models

• DE-SimplE (including DE-SimplE, DE-DistMult, and DE-TransE)
• TERO
• ATiSE

TKGE Dataset

• ICEWS14

Neural Network Models

• 25input
• 5input

Neural Network Dataset

• 25t_5w has 25 inputs corresponding to the top 5 predictions of each TKGE model, and target values are the rank of the correct answer of each TKGE model.
• 5p_5w has 5 inputs corresponding to 5 TKGE models, 5 predictions, and target values are the rank of the correct answer of each TKGE model.