ProjE: Embedding Projection for Knowledge Graph Completion

Authors: Baoxu Shi, Tim Weninger

AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental we show that simple changes in the architecture of the underlying model can outperform state-of-the-art models without the need for complex feature engineering. In this work, we present a shared variable neural network model called Proj E that fills-in missing information in a knowledge graph by learning joint embeddings of the knowledge graph s entities and edges, and through subtle, but important, changes to the standard loss function. In doing so, Proj E has a parameter size that is smaller than 11 out of 15 existing methods while performing 37% better than the current-best method on standard datasets. We also show, via a new fact checking task, that Proj E is capable of accurately determining the veracity of many declarative statements. We evaluate the Proj E model with entity prediction and relationship prediction tasks, and compare the performance against several existing methods using experimental procedures, datasets, and metrics established in the related work.
Researcher Affiliation Academia Baoxu Shi and Tim Weninger University of Notre Dame Email: {bshi, tweninge}@nd.edu
Pseudocode No A detailed algorithm description can be found in the Supplementary Material.
Open Source Code Yes Proj E is implemented in Python using Tensor Flow (Abadi et al. 2016); the code and data are available at https://github.com/nddsg/Proj E.
Open Datasets Yes The FB15K dataset is a 15,000-entity subset of Freebase; the Semantic MEDLINE Database (Sem Med DB) is a KG extracted from all of Pub Med (Kilicoglu et al. 2012); and DBpedia is KG extracted from Wikipedia infoboxes (Lehmann, Isele, and Jakob 2014).
Dataset Splits No The paper refers to a 'training phrase' and mentions using 'experimental procedures, datasets, and metrics established in the related work' and 'experiment settings in Trans E (Bordes et al. 2013) and PTrans E (Lin, Liu, and Sun 2015)', but it does not explicitly state specific train/validation/test split percentages or sample counts within the paper itself.
Hardware Specification No The paper states that 'Proj E is implemented in Python using Tensor Flow', but does not provide any specific details about the hardware used for experiments, such as GPU models, CPU types, or memory specifications.
Software Dependencies No The paper mentions 'Python' and 'Tensor Flow (Abadi et al. 2016)' and 'Adam (Kingma and Ba 2014)' as software components, but it does not provide specific version numbers for any of these.
Experiment Setup Yes For both entity and relationship prediction tasks, we use Adam (Kingma and Ba 2014) as the stochastic optimizer with default hyper-parameter settings: β1 = 0.9, β2 = 0.999, and ϵ = 1e 8. During the training phrase, we apply an L1 regularizer to all parameters in Proj E and a dropout layer on top of the combination operator to prevent over-fitting. The hyper-parameters in Proj E are the learning rate lr, embedding size k, mini-batch size b, regularizer weight α, dropout probability pd, and success probability for negative candidate sampling py. We set lr = 0.01, b = 200, α = 1e 5, and pd = 0.5 for both tasks, k = 200, py = 0.5 for the entity prediction task and k = 100, py = 0.75 for the relationship prediction task. For all tasks, Proj E was trained for at most 100 iterations, and all parameters were initialized from a uniform distribution U[ 6/ k] as suggested by Trans E (Bordes et al. 2013).