some questions about StATIK

[dadaer]

Hi,sorry for bothering.I am a master's student and my research direction is knowledge graph.Recently I read the paper about your statik work.After reading it I have the following question: Statik use a MPNN network to leverage structural information.It firstly needs to sample neighbors.However,you claim that statik doesn't need a support set in test procedure.My question is that If there isn't a support set,how can unseen entities sample neighbors and how does MPNN work? Best wishes and waiting for your reply.

[Elanmarkowitz]

Hi there, thanks for the interest in StATIK. By support set we mean edges connecting the inductive entity to entities in the training set. We do use such edges when they exist but don’t require it. Meaning, StATIK works if none of the edges connect to entities in the training graph (transfer setting), however it does rely on edges existing.

Potentially StATIK could work with no edges at all (I.e. MPNN is not used) but we don’t evaluate in that scenario.

Let me know if this clarifies things.

Best, Elan

[dadaer]

Hi,thanks for clarifing.You claim that you do use(rely on) the edges connecting unseen entity to entities in the traning set(edges in support set or auxiliary set).However,in your experiment section,the dateset you used doesn't has a supprot set or auxiliary set like work LAN (Wang et al.,2019) and InvTransE (Dai et al., 2021)used.So, how does the MPNN in StATIK work in test procedure. Best wishes.

[Elanmarkowitz]

"Specifically, for a given evaluation triple, Ti = (h, r, t), do head and tail prediction on the graph G − Ti. This means that given a query of the form (h, r, ?) or (?, r, t), rank all possible tail or head candidates (targets) such that the real triple is ranked as highly as possible." G is the graph at inference time.

The following closed issue goes through the inference/evaluation procedure. https://github.com/Elanmarkowitz/StATIK/issues/5