Knowledge Graph Completion with Adaptive Sparse Transfer Matrix
Authors: Guoliang Ji, Kang Liu, Shizhu He, Jun Zhao
AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments, we design structured and unstructured sparse patterns for transfer matrices and analyze their advantages and disadvantages. We evaluate our approach on triplet classification and link prediction tasks. Experimental results show that Tran Sparse outperforms Trans(E, H, R, and D) significantly, and achieves state-of-the-art performance. |
| Researcher Affiliation | Academia | Guoliang Ji, Kang Liu, Shizhu He, Jun Zhao National Laboratory of Pattern Recognition (NLPR) Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China {guoliang.ji, kliu, shizhu.he, jzhao}@nlpr.ia.ac.cn |
| Pseudocode | Yes | Algorithm 1 Learning Tran Sparse(separate). |
| Open Source Code | No | The paper does not provide an explicit statement or link to open-source code for the described methodology. |
| Open Datasets | Yes | Datasets We do triplet classification and link prediction tasks on Word Net (Miller 1995) and Freebase (Bollacker et al., 2008). ... In this paper, we use two subsets of Word Net: WN11 (Socher et al., 2013) and WN18 (Bordes et al., 2014). Freebase is a large collaborative knowledge base which consists of large amounts of facts. We also use two subsets of Freebase: FB15k (Bordes et al., 2014) and FB13 (Socher et al., 2013). |
| Dataset Splits | Yes | Table 2 lists the statistics of the four datasets. [Table 2 includes a '#Valid' column with specific counts for each dataset, e.g., 'WN11 ... 2,609', 'FB15k ... 50,000'] For triplet classification, we set a threshold δr for each relation r. δr is obtained by maximizing the classification accuracies on the valid set. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper does not specify software dependencies with version numbers (e.g., specific Python versions, library versions like TensorFlow/PyTorch versions). |
| Experiment Setup | Yes | In this experiments, we select the margin γ among {1, 1.5, 2, 4, 10}, the learning rate λ for SGD (Duchi, Hazan, and Singer 2011) among {0.1, 0.01, 0.001}, the minimum sparse degree θmin among {0.0, 0.1, 0.3, 0.7, 0.9}, the dimension of vectors n amon g {20, 50, 80, 100}, and the mini-batch size B among {100, 200, 1000, 4800}. The best configuration obtained by valid set are: γ = 4, λ = 0.001, θmin = 0.7, n = 20, B = 1000 and taking L1 as dissimilarity on WN11; γ = 1, λ = 0.001, θmin = 0.9, n = 100, B = 200 and taking L2 as dissimilarity on FB13; γ = 1.5, λ = 0.001, θmin = 0.0, n = 100, B = 4800 and taking L1 as dissimilarity on FB15k. |