kage08/graph_sample_rl

Influence maximization in unknown social networks: Learning Policies for Effective Graph Sampling (official code repository)

Influence maximization in unknown social networks: Learning Policies for Effective Graph Sampling

Official code for the paper Influence maximization in unknown social networks: Learning Policies for Effective Graph Sampling (to appear at AAMAS 2020).

A serious challenge when finding influential actors in real-worldsocial networks, to enable efficient community-wide interventions, is the lack of knowledge about the structure of the underlying network. Current state-of-the-art methods rely on handcrafted sampling algorithms; these methods sample nodes and their neighbours in a carefully constructed order and choose opinion leaders from this discovered network to maximize influence spread in the(unknown) complete network.In this work, we propose a reinforcement learning frameworkto discover effective network sampling heuristics by leveraging automatically learnt node and graph representations that encodeimportant structural properties of the network. At training time,the method identifies portions of the network such that the nodesselected from this sampled subgraph can effectively influence nodes in the complete network. The output of this training is a transferable, adaptive policy that identifies an effective sequence of nodes toquery on unseen graphs. The success of this policy is underpinned by a set of careful choices for embedding local and global infor-mation about the graph, and providing appropriate reward signals during training. We experiment with real-world social networksfrom four different domains and show that the policies learnedby our RL agent provide a 7-23% improvement over the currentstate-of-the-art method.

Paper link: https://arxiv.org/pdf/1907.11625.pdf

graph_sample_rl's Issues

Question about the network architecture in your work

Hi, I found that you adopt DQN in your paper. But I also found some keywords like actor_critic in the code, is it means the method you used is a combination of DQN and actor-critic method? Maybe, it seems like the DDPG method?

Thank you in advance.

code

Hello, the work you have done is really meaningful.
When looking at the code, I would like to know how you got the influence score in your paper. It is not shown in the code,
I have made relevant improvements, but I don't know how to evaluate him.
Thank you for letting me know.

Steps to reproduce results in the paper

Can you please provide the list of hyperparameters and other required steps to reproduce the results in the paper. I tried the default mentioned parameters as well as tried tuning the parameters and training multiple models but i can not get the values close to the one provided in the paper for twitter dataset. Also the models trained are not stable so wanted to check what I have been missing.

Thanks!