functional Embeddings: Embedding Knowledge Graphs in function spaces

Knowledge graph embedding is successful when using division algebras ($\mathbb{R}$, $\mathbb{C}$, $\mathbb{Q}$, etc.) as these spaces help model complex relations and patterns in a KG dataset. So far, only vectors have been used to compute the embeddings of KGs. This repository aims to extend this idea and consider another alternative: function space. So we compute and represent the embeddings of entities and relations as functions. First with polynomial functions, then trigonometric and neural network functions. Here we implemented three functional embedding models. The first, called PolyMult, embeds using polynomial functions. The second is called LFMult1 embed, which uses a trigonometric function, and the third is called LFMult embed, which uses neural networks. Technical details can be found here

Installation

First, make sure you have Anaconda installed

Click me!

Installation from Source

conda create -n decal python=3.10.13 --no-default-packages && conda activate decal && cd functionnal-embeddings &&
pip3 install -e .

Download Knowledge Graphs

wget https://files.dice-research.org/datasets/dice-embeddings/KGs.zip --no-check-certificate && unzip KGs.zip

Knowledge Graph Embedding Models

To see available Models

1. Translative models: TransE, TransR
2. Multiplicative models: Distmult, ComplEx, QMult, OMult
3. Convolutional models: ConEx, ConvO, ConvQ
4. Functional models: PolyMult, LFMult, FMult, LFMult1

How to use this repo?

Install all the necessary packages using:

 pip install -r requirements.txt 

Embedding with polynomials:

To get the results obtained in the paper for the UMLS data, do:

python3 run.py --model PolyMult --eval_model "train_val_test" --scoring_technique NegSample --degree 1 --lr 0.02 --embedding_dim 32 --num_epochs 500 --neg_ratio 50 --optim Adam --batch_size 1024

Embedding with trigonometric function:

Click me!

python3 run.py --model LFMult1 --eval_model "train_val_test" --scoring_technique NegSample --degree 1 --lr 0.02 --embedding_dim 32 --num_epochs 500 --neg_ratio 50 --optim Adam --batch_size 1024

Embedding with Neural Networks:

Click me!

python3 run.py --model LFMult1 --eval_model "train_val_test" --scoring_technique NegSample --degree 1 --lr 0.02 --embedding_dim 32 --num_epochs 500 --neg_ratio 50 --optim Adam --batch_size 1024