Link Prediction on Sequential Graph Data

I am new at DGL or Graph Neural Networks. I have some experience in NN and Graph Data but not in DGL.
I have some data includes; “user” ,“location”(not spatial only Id) and datetime. Users check in in these locations. So I think this data can be represented as Graph. “user1”–>checked_in–>“location1”.A user can check in same location at different times. For trial purposes for now I did not add user or location features. But edges have time feature.
I want to predict “at which location will the next check_in of a specific user?” to send mail about deals around to that location.
So, what model should i use? or, Is there any solution for this problem in Graph Neural Networks world?

1 Like

Definitely a great question. I’ve been thinking about the same kind of problem. Hope to hear some suggestions!


This is a challenging open question. If we ignore the temporal information, the problem is a standard recommendation problem. That user ---check in---> location is similar to user ---buy/like---> items. There are pretty much work has tried graph network algorithm to gain better results (i.e. GCMC).

However, in your problem (actually also same to recommendation problem), temporal information is really important. We are also seeking for this kind of work but there’s less research in this area (may be due to lack of temporal recommendation dataset). Our team has some exploratory try in this direction but didn’t get much result yet. One thing I could suggest is some variants of PinSage, that you can change the sampling method to incorporate the time locality information.

1 Like

Thank you for your reply and suggestions.
I will try this and I will be following the team.

I think the problem could be perhaps rephrased in many-graphs manner: one graph per each user, tracing sequentially their locations as nodes, time-since-the-prior-location as the node property. The problem then becomes graph classification - “which node likely comes next” (similarly I think web session recommenders are modeled), and graph regression - “time to the next location”.

1 Like