Deep Graph Representation Learning and Optimization for Influence Maximization
Authors: Chen Ling, Junji Jiang, Junxiang Wang, My T. Thai, Renhao Xue, James Song, Meikang Qiu, Liang Zhao
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive analyses are conducted over both synthetic and real-world datasets to demonstrate the overall performance of Deep IM. In this section, we compare the performance of our proposed Deep IM framework across six real networks in maximizing the influence under various settings, following a case study to qualitatively demonstrate the performance of Deep IM. We evaluate the performance of Deep IM in maximizing the influence against other approaches under various IM application schemes. Each model selects 1%, 5%, 10%, and 20% nodes in each dataset as seed nodes, and we allow each diffusion model to simulate until the diffusion process stops and record the average influence spread of 100 rounds. We report the percentage of final infected nodes (i.e., the number of infected nodes/the total number of nodes). |
| Researcher Affiliation | Collaboration | 1Emory University, Atlanta, GA 2Fudan University, Shanghai, CHINA 3NEC Labs America, Princeton, NJ 4University of Florida, Gainesville, FL 5Dakota State University, Madison, SD. |
| Pseudocode | Yes | Algorithm 1 Deep IM Prediction Framework |
| Open Source Code | Yes | The code and data are available at: https://github.com/triplej0079/Deep IM. |
| Open Datasets | Yes | The proposed Deep IM is compared with other approaches over six real-world datasets, including Cora-ML, Network Science, Power Grid, Jazz, Digg, and Weibo. Jazz (Rossi & Ahmed, 2015). Cora-ML (Mc Callum et al., 2000). Power Grid (Rossi & Ahmed, 2015). Network Science (Rossi & Ahmed, 2015). Digg (Panagopoulos et al., 2020). Weibo (Panagopoulos et al., 2020). |
| Dataset Splits | No | The paper mentions generating a training set from sampled seed nodes and influence spread data, but it does not specify explicit training/validation/test splits or cross-validation setup for its experiments. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used to run its experiments, such as GPU/CPU models, memory, or cloud instance types. |
| Software Dependencies | No | The paper mentions the use of 'GAT-structured diffusion estimation model' and 'Adam' optimizer, but it does not provide specific version numbers for any software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | For the configuration of each diffusion model, we use a weighted cascade version of the IC model, i.e., the propagation probability pu,v = 1/din v (din v denotes the in-degree of node v) for each edge e = (u, v) on graph G; For LT model, the threshold θ is set to be uniformly sampled from [0.3, 0.6] for each node v; the infection probability and recovery probability are set to be 0.001 in the SIS model. For Deep IM, the 2-layer GAT-structured diffusion estimation model that each layer contains 4 attention heads and the dimension of each attention channel is 64. Both encoder and decoder are symmetric 4-layer MLP with hidden size 512, 1024, 1024, and 1024 for each layer, respectively. We choose Adam with learning rates 0.001 and 0.0001 for optimizing both Eq. (6) and Eq. (7), respectively. |