TreeRNN

TreeRNN is a small experimental (unregistered) package implementing variants of recurrent neural networks (RNNs) for tree-structured data. The input tree-structured data are assumed to be in the JSON format. The package was used to compute the tree RNNs in the following paper: Papež M, Rektoris M, Šmídl V, Pevný T. Sum-Product-Set Networks: Deep Tractable Models for Tree-Structured Graphs. In The 12th International Conference on Learning Representations (ICLR2024).

The package is built on top of JSONGrinder.jl and Mill.jl.

To use the package, execute the following steps.

1. Clone this repository.

git clone https://github.com/mlnpapez/TreeRNN.jl TreeRNN.jl

2. Go to the TreeRNN.jl repository.

cd TreeRNN.jl

3. Open the Julia 1.10.2 console and write:

using Pkg
Pkg.activate(".")
Pkg.instantiate()

4. Use the package:

using TreeRNN

The repository contains 12 datasets in JSON format (listed in the table below). All these datasets were downloaded from CTU Prague Relational Learning Repository and converted into the JSON format.

Dataset	# of instances	# of classes
mutagenesis	188	2
genes	862	15
cora	2708	7
citeseer	3312	6
webkp	877	5
world	239	7
chess	295	3
uw_cse	278	4
hepatitis	500	2
ftp	30000	3
ptc	343	2
dallas	219	7

The repository currently contains three models (listed in the table below).

Model	Description	Reference
TreeGRU	tree-structured gated reccurent unit	https://arxiv.org/abs/1610.02806
TreeLSTM	tree-structured long short-term memory	https://arxiv.org/abs/1503.00075
TreeMLP	tree-structured multilayer perceptron	-

Execute the following steps to train one of these models on a selected dataset.

1. Go to the TreeRNN.jl repository.

cd TreeRNN.jl

2. Open the Julia 1.10.2 console and write:

using Pkg
Pkg.activate("scripts")
Pkg.instantiate()
Pkg.develop(path=".")

3. Use the package:

include("scripts/train.jl")