TINY-KE / link-prediction

GNN-based prediction of connections in a spatial configuration (floor plan).

link-prediction

AI-based autocompletion of graph-based architectural spatial configurations using link prediction with graph neural networks (GNN).

The main goal of the approach is to estimate the probability of relations between the rooms of the spatial configuration graph using the relevant semantic information.

The approach is published in the context of the paper "Autocompletion of Design Data in Semantic Building Models using Link Prediction and Graph Neural Networks" submitted and presented @ eCAADe 2022.

Requirements

Following packages are required to run link prediction:

python3

python3-pip

Deep Graph Library (DGL)

CUDA

Matplotlib

Run evaluation

Run ./train.sh in console.

You shoud get results similar to the evaluation in the paper (see linked paper).

The evaluation was tested on Ubuntu 22.04 LTS with Python 3.10 and CUDA 11.6.

Run API (lpapi)

Run ./lpapi.sh in console.