Parameter-free Dynamic Graph Embedding for Link Prediction

Authors: Jiahao Liu, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, Ning Gu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on two link prediction tasks show that Free GEM can outperform the state-of-the-art methods in accuracy while achieving over 36X improvement in efficiency. All code and datasets can be found in https://github.com/Fudan CISL/Free GEM.
Researcher Affiliation Collaboration Jiahao Liu1,2, Dongsheng Li3, Hansu Gu4B, Tun Lu1,2B, Peng Zhang1,2, Ning Gu1,2 1School of Computer Science, Fudan University, Shanghai, China 2Shanghai Key Laboratory of Data Science, Fudan University, Shanghai, China 3Microsoft Research Asia, Shanghai, China 4Seattle, United States jiahaoliu21@m.fudan.edu.cn, dongsli@microsoft.com, hansug@acm.org {lutun, zhangpeng_, ninggu}@fudan.edu.cn
Pseudocode No The paper describes methods in text and figures (e.g., Figure 1: The high-level design of the proposed Free GEM method) but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes All code and datasets can be found in https://github.com/Fudan CISL/Free GEM.
Open Datasets Yes Datasets For the future item recommendation task, we use the Amazon Video, Amazon Game [9], Movie Lens-1M (ML-1M) and Movie Lens-100K (ML-100K) [8]. For the next interaction prediction task, we use Wikipedia and Last FM [14]. For all datasets, we use the first 80% interactions as training set, the following 10% interactions as validation set, and the last 10% interactions as test set.
Dataset Splits Yes For all datasets, we use the first 80% interactions as training set, the following 10% interactions as validation set, and the last 10% interactions as test set.
Hardware Specification No The paper discusses running time and computational efficiency ('Free GEM is at least 36X faster than JODIE and at least 370X faster than Co PE') but does not specify the hardware (e.g., CPU, GPU models, memory) used for conducting the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers, such as programming languages, libraries, or frameworks (e.g., 'Python 3.x', 'PyTorch 1.x').
Experiment Setup Yes The paper mentions specific hyperparameters: 'R = D α U RD α I , where α > 0 is a hyperparameter'; 'introduce a hyperparameter γ to control the ratio of high-frequency signals to low-frequency signals as follows: EU = USγ, EI = V Sγ (γ < 0.5)'; 'αi is the hyperparameter that controls the weight of the i-th path'; and 'decay the historical interaction score as exp{β(t/Ti 1)}, where Ti is the beginning time of the i-th stage and β > 0 is the decay coefficient.'