Source-linked AI summary
Semi-supervised User Geolocation via Graph Convolutional Networks
Afshin Rahimi, Trevor Cohn, Timothy Baldwin
TL;DR
Social media user geolocation matters for applications, but text and network views provide complementary information while only a small fraction of tweets is geotagged. The paper proposes multiview, transductive semi-supervised models, showing strong performance with minimal supervision and competitive performance with sufficient supervision.
Problem
User geolocation supports applications such as public health monitoring, but effective models must combine complementary text and network information amid scarce geotagged data.
Method
The paper proposes GCN, DCCA, and MLP-TXT+NET, multiview transductive semi-supervised models that jointly use text and network information to infer user locations.
Results
GCN and DCCA perform well with minimal supervision, while GCN is competitive with state-of-the-art models given sufficient supervision; highway gates control useful neighbourhood smoothing.
Takeaways & Limitations
Jointly modelling text and network information and using unlabelled data supports geolocation when supervision is limited.
Takeaways & Limitations
Performance varies by location because non-uniform local discussions make geolocation harder in less densely populated areas.
Abstract
from arXiv · showhide
Social media user geolocation is vital to many applications such as event detection. In this paper, we propose GCN, a multiview geolocation model based on Graph Convolutional Networks, that uses both text and network context. We compare GCN to the state-of-the-art, and to two baselines we propose, and show that our model achieves or is competitive with the state- of-the-art over three benchmark geolocation datasets when sufficient supervision is available. We also evaluate GCN under a minimal supervision scenario, and show it outperforms baselines. We find that highway network gates are essential for controlling the amount of useful neighbourhood expansion in GCN.
1 Introduction
User geolocation supports applications from public-health monitoring to emergency awareness, but public location data are scarce and noisy. The paper therefore motivates jointly using text and network information with unlabeled data, proposing GCN for minimal-supervision geolocation.
- User geolocation supports public-health monitoring, regional sentiment studies, and real-time emergency awareness systems.
- Third-party providers often lack direct access to IP, WiFi, and GPS signals, while profile locations are noisy and geotagged tweets cover only 1% of tweets.
- Predictive geolocation must therefore use publicly available signals such as tweet text and social interaction data.
- Prior work typically uses either supervised text models or graph-based semi-supervised label propagation based on location homophily.
- Text and network views are complementary, yet existing multiview methods incompletely exploit unlabeled data despite the need to perform well with minimal supervision.
- GCN jointly learns from text and network information, outperforming two strong baselines under minimal supervision and remaining competitive with state-of-the-art models when supervision is sufficient.
2 Model
The paper proposes GCN, a transductive multiview geolocation model that combines user text with network structure through graph convolution. Its layers expand neighbourhood information while highway gates regulate how much of that information is passed forward; DCCA provides a contrasting correlation-based multiview representation.
- GCN: GCN jointly learns user locations from a bag-of-words text view and an @-mention network view in a transductive setting.The model infers unlabelled locations from labelled users while representing each user’s content and interactions as separate views.
- GCN: Each GCN layer averages a user with labelled and unlabelled neighbours, then applies trainable linear parameters and a nonlinear activation.The normalized adjacency is formed from the interaction matrix plus a self-weight controlled by λ.
- GCN: Adding layers extends smoothing to farther neighbours, with a 3-layer GCN reaching users up to 3 hops away.This expansion can exploit location homophily across multiple hops, but also increases exposure to noisy neighbourhood information.
- Highway GCN: Layer-wise highway gates balance transformed neighbourhood information against the incoming representation.The gates use learned weights and biases with elementwise mixing, controlling how much neighbourhood expansion reaches each layer’s output.
- DCCA: DCCA separately transforms the text and network views to maximize their correlation, then concatenates the resulting projections as a multiview representation.The compressed representations are intended to reduce uncorrelated noise before being used by another task-specific model.
3 Experiments
The experiments evaluate GCN and multiview baselines on three established Twitter geolocation datasets using text and collapsed @-mention network views. The setup includes graph-based and concatenation baselines, with dataset-specific architectures and neighbourhood controls.
- Data: Experiments use GEOTEXT, TWITTER-US, and TWITTER-WORLD, containing approximately 9k, 449k, and 1.3m users, respectively.The datasets cover the continental US or the whole world and provide predefined training, development, and test partitions.
- Baselines: DCCA processes text and network views through separate neural networks, concatenates their outputs, and feeds them to a supervised location predictor.Its unsupervised components maximize output correlation before supervised prediction.
- Constructing the Views: The text view uses binary term frequency, inverse document frequency, and l2-normalised samples, while the network view comes from a collapsed @-mention graph.Users are connected when one mentions another or when they co-mention another user.
- Model Configuration: GCN uses highway layers to regulate neighbourhood information and dataset-specific depths and widths, with its final graph-convolutional softmax averaging over radius 4.The reported GCN layer sizes are 300, 600, and 900 for GEOTEXT, TWITTER-US, and TWITTER-WORLD.
- Baselines: GCN-LP replaces text features with one-hot neighbour encodings and applies graph convolutions over k-hop neighbours with a gating mechanism.This baseline parallels label propagation while adding learned layer parameters and control over the smoothing radius.
- Baselines: MLP-TXT+NET concatenates the bag-of-words text view with the network vector and classifies users using a single hidden-layer multilayer perceptron.The model is a simple transductive supervised baseline with dataset-specific hidden-layer sizes.
4 Results and Analysis
Across representation, supervision, benchmark, and error analyses, GCN benefits from jointly using text and network information, especially under limited supervision. Highway gates regulate neighbourhood expansion, while performance remains sensitive to labelled-data quantity and geographic imbalance.
- Representation: GCN substantially improves t-SNE representations over concatenation and slightly improves on Deep CCA, with further convolution clustering samples more tightly.The visualisations compare concatenated network features, Deep CCA representations, and one or two graph-convolution applications.
- Labelled Data Size: When more than 20% of samples are labelled, GCN and MLP-TXT+NET clearly outperform DCCA; with nearly all data labelled on larger datasets, MLP-TXT+NET outperforms GCN.At 1% labelled data on GEOTEXT, DCCA outperforms GCN, attributed to its fewer parameters and the limited supervision available for training GCN.
- Neighbourhood Expansion: Adding GCN layers can reduce accuracy without gates as neighbourhood expansion introduces noise, whereas highway gates improve performance through three layers and peak at four layers.The four-layer peak is compatible with the observed shortest-path distribution in GEOTEXT’s graph.
- Performance: Joint text-and-network modelling outperforms text-only and network-only approaches, although MLP-TXT+NET exceeds GCN on the largest datasets when labelled data is abundant.The comparison includes the proposed joint models, GCN-LP, and prior supervised and hybrid systems.
- Performance: Under minimal supervision, DCCA and GCN outperform MLP-TXT+NET, while GCN generally outperforms DCCA and achieves strong results against text-only, network-only, and hybrid models.The minimal-supervision comparison uses 1% labelled data; an exception is Acc@161 on GEOTEXT, where the labelled sample count is very low.
- Error Analysis: With only 1% labelled data, GCN’s errors are higher for smaller states and often concentrate in larger or neighbouring states such as California, New York, Texas, and Ohio.The analysis links these errors to insufficient labelled samples in smaller states and reports cross-state confusions involving geographically or socially connected regions.
5 Related Work
Prior user-geolocation work spans text-based, network-based, and multiview approaches, but important gaps remain in jointly modelling views and exploiting unlabelled information. The paper addresses these gaps with multiview semi-supervised models.
- Text-based methods infer location from geographical language variation, while network-based methods rely on location homophily in user interactions.
- GCN detects local terms from unlabelled data through graph convolutions, extending the vocabulary beyond terms present in labelled data.
- Network-based models can fail for disconnected users, with one prior system unable to geolocate 37% of users.
- Hybrid approaches incorporate text into graphs, ensemble separate models, or combine text and network information, but most do not effectively exploit unlabelled text data.
- Three shortcomings motivate this work: limited joint exploitation of text and network information, ineffective use of unlabelled data in both views, and inadequate modelling of their interaction.
6 Conclusion
The paper proposes three multiview, transductive, semi-supervised models that jointly use text and network information for user geolocation. Joint modelling outperforms single-view and hybrid approaches, while GCN and DCCA perform well with minimal supervision.
- The paper proposes GCN, DCCA, and MLP-TXT+NET as multiview, transductive, semi-supervised models using text and network information jointly.
- Joint modelling of text and network information outperforms network-only, text-only, and hybrid geolocation models.
- GCN and DCCA perform well under minimal supervision by effectively using unlabelled data.
- The models assume all social connections exhibit location homophily and ignore the context in which users interact.