Hierarchical Diffusion Attention Network

Authors: Zhitao Wang, Wenjie Li

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed model against state-of-the-art sequential diffusion prediction models on three real diffusion datasets. The significantly better performance demonstrates the effectiveness of our model.
Researcher Affiliation Academia Zhitao Wang and Wenjie Li Department of Computing, The Hong Kong Polytechnic University, Hong Kong {csztwang, cswjli}@comp.polyu.edu.hk
Pseudocode No The paper describes the model architecture and equations but does not include a clearly labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code No The paper does not contain any explicit statement about releasing source code or provide a link to a code repository.
Open Datasets Yes Memes [Leskovec et al., 2009]: This dataset contains articles from mainstream news websites or blogs. ... Weibo [Zhang et al., 2013]: This dataset consists of content reposting logs crawled from Sina Weibo, a Chinese microblogging site. ... Twitter [Weng et al., 2013]: This dataset records the diffusion processes of hash-tags in Twitter.
Dataset Splits Yes We randomly sample 80% of cascades for training and the rest for validating and testing with an even split.
Hardware Specification Yes All models except CYAN-RNN1 are implemented with Tensorflow and trained on the same GTX1080Ti graphic card with the same batch size.
Software Dependencies No The paper mentions 'Tensorflow' as the implementation framework and 'Adam optimizer' but does not specify their version numbers or other software dependencies with versions.
Experiment Setup Yes For the proposed models, the dimension size of d is also 64, the learning rate is 0.001, the max observation time Tmax is 120 hours, the number of splitting time interval T is 40, and the non-linear activation functions fx, fu are selected as Exponential Linear Unit (ELU) [Clevert et al., 2016]. We also apply the Dropout [Srivastava et al., 2014] with the keep probability 0.8 and the L2 regularization on parameters to avoid over-fitting.