Inductive Logical Query Answering in Knowledge Graphs

Authors: Michael Galkin, Zhaocheng Zhu, Hongyu Ren, Jian Tang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, we show that inductive models are able to perform logical reasoning at inference time over unseen nodes generalizing to graphs up to 500% larger than training ones. Exploring the efficiency effectiveness trade-off, we find the inductive relational structure representation method generally achieves higher performance, while the inductive node representation method is able to answer complex queries in the inference-only regime without any training on queries and scales to graphs of millions of nodes. Code is available at https://github.com/Deep Graph Learning/Inductive QE.
Researcher Affiliation Academia Mikhail Galkin Mila, Mc Gill University mikhail.galkin@mila.quebec Zhaocheng Zhu Mila, Université de Montréal zhuzhaoc@mila.quebec Hongyu Ren Stanford University hyren@stanford.edu Jian Tang Mila, HEC Montréal, CIFAR AI Chair jian.tang@hec.ca
Pseudocode No The paper describes algorithms and methods in prose but does not include any explicit pseudocode blocks or sections labeled 'Algorithm'.
Open Source Code Yes Code is available at https://github.com/Deep Graph Learning/Inductive QE.
Open Datasets Yes We create a novel suite of datasets based on FB15k-237 [29] (open license) and following the query generation process of Beta E [24]... Finally, we perform a scalability experiment evaluating complex query answering in the inductive mode on a new large dataset Wiki KG-QE constructed from OGB Wiki KG 2 [16] (CC0 license).
Dataset Splits Yes For validation and test graphs, we split the remaining set of entities into two non-overlapping sets each with 1 r / 2 |E| nodes. We then merge training and unseen nodes into the inference set of nodes Einf and induce inference graphs for validation and test from those sets, respectively, i.e., Eval inf = Etrain [ Eval and Etest inf = Etrain [ Etest. ...Training queries are sampled from the training graph Gtrain, validation and test queries are sampled from their respective inference graphs Ginf...
Hardware Specification Yes Node Piece-QE models were pre-trained and evaluated on a single Tesla V100 32 GB GPU whereas GNN-QE models were trained and evaluated on 4 Tesla V100 16GB.
Software Dependencies No The paper states: 'Both Node Piece-QE and GNN-QE are implemented4 with Py Torch [22] and trained with the Adam [18] optimizer.' However, it does not provide specific version numbers for PyTorch, Adam, or any other software libraries.
Experiment Setup Yes The non-parametric CQD-Beam [5] decoder for answering complex queries is tuned for each query type based on the validation set of queries, most of the setups employ a product t-norm, sigmoid entity score normalization, and beam size of 32. Following the literature, the GNN-QE models [40] were trained on 10 query patterns (1p/2p/3p/2i/3i/2in/3in/inp/pin/pni) where ip/pi/2u/up are only seen at inference time. Each model employs a 4-layer NBFNet [42] as a trainable projection operator with Dist Mult [33] composition function and PNA [10] aggregation.