2018 - arXiv -Semi-supervised User Geolocation via Graph Convolutional Networks

[karan96]

Title: - Semi-supervised User Geolocation via Graph Convolutional Networks Year: - 2018 Venue: - arXiv

Main Problem The authors addresses the problem of inaccurate prediction of user geo-location. The author base this problem on the unavailability of sometimes text and sometimes network information which leads to researchers using either of the two data but not both for their take on predicting user geo-location.

Related Work The author divides the related work section into three different approaches: - text-based, network-based, multi-view based approaches. The author cites related papers to demonstrate that the previous work faces many shortcomings while addressing the above mentioned problem. Text-based approaches often times ignore non-geographical references and vernacular use of language, do not scale up to the magnitude of social media, or often times face with scarcity of data in a real world scenario. Network based approaches often failed to geolocate users as a result of disconnectedness graphs. The author also boldly claims that none of the multi-view based approaches with the exception of one effectively uses unlabelled data in the text view and use only the unlabelled information of the network view via the user-user graph.

Input A user's text view involving all of his social text and a user-user interaction network view is taken as inputs to the proposed model.

Output User's predicted location is given as output from the model.

Proposed Method Authors propose a transductive multiview geolocation model using Graph Convolutional networks. The authors also propose two multiview baseines based on the concatenation of text and network plus another model called DCCA based on Deep Canonical Correlation Analysis.

Experimentation

Dataset Used

1. GEOText(Supplied Link Not Working)
2. TWITTER-US(Need to dig deeper to get the link of the dataset, Will do it)
3. TWOTTER-WORLD

Input - 9k, 449k and 1.3m users, sourced from the above three mentioned datasets respectively.

Result - Prediction of geo-location of user.

Code https://github.com/afshinrahimi/geographconv

Pros/Cons

1. The authors report the performance timing for the models proposed on different datasets which is a nice way to benchmark the proposed methods.
2. However, only calculates the effect of labelling data using median error as the preferred metric.

[hosseinfani]

I know @afshinrahimi from back home. @afshinrahimi How are you? very long time. small-world, eh?