![](https://csdnimg.cn/release/download_crawler_static/88935641/bg3.jpg)
Information Fusion 81 (2022) 1–13
3
F. Zhou et al.
language characteristics efficiently without manual location feature en-
gineering, a recent study [34] finds that TF-IDF is consistently superior
to doc2vec due to the location-indicative words captured in TF-IDF.
Our present work enables better location-awareness than the exist-
ing literature and, in particular, HGNN distinguishes the crowd effects
from different geographic regions.
2.2. Network-based methods
Online social relationships are also important indicators for user
geolocation under the homophily assumption [35–38], i.e., people
prefer to interact with others in nearby areas. Backstrom et al. [35]
examine the relationship between users’ geographical proximity and
online friendships on Facebook, and find that the likelihood of relations
between any user pair drops monotonically as a function of distance.
Rather than solely relying on friendships, more and more works uti-
lize various types of connections, such as the co-mention tags and
mentions between non-friends, to construct closer social interactions
beyond friendships [2,31]. In this way, similar interests among users
can be retrieved from such implicit networks to improve geolocation
accuracy [30,39,40]. Moreover, researchers also identify some noisy
interaction factors that may degrade the prediction performance. For
example, social influence of celebrities is a distracting factor that
may confuse the prediction and thus is removed from the built user
network [30,41].
Although the existing approaches have tackled the aspect of ex-
plicitly modeling location dependency between social connected users,
some challenges have not been properly addressed — namely, the
sparsity of geo-tagged users and the inaccurate label propagation. More
importantly, friends’ locations are usually contradicting each other,
which hinders the practical applicability of these works. In contrast, our
HGNN learns both region-level and node-level features and aggregates
them in a manner that provides better intepretability.
2.3. Multi-information fusion based models
Recent efforts have leveraged deep graph learning methods to
model user interaction networks by fusing user-generated contents
and various meta-data, such as user profiles, tweeting time, and user
timezone. For example, MENET [7] exploits node2vec [13] to learn
user representations, combined with text representation learned by
doc2vec, for predicting users’ locations. Another work [9] employs
GCNs [15] for learning network structures with the graph convolution
and pooling operations, which has achieved state-of-the-art geolocation
performance. A recent work [34] investigate several graph embedding
methods and found that NetMF [42] performs better than node2vec and
GraphSAGE [17] on user geolocation task, but does not show superior
performance than GCN-based models [9,34].
It is worth noting that some works make use of various meta-data
(e.g., self-declared location in profile and timezone information) for
improving the prediction performance. For example, user timezone,
as well as UTC offset and country noun, have been used for user
geolocation [7,8,28,31,43]. While such auxiliary information is a strong
indicator for regularizing the locations the model predicted, a majority
of users are not willing to open this privacy information, which is
sometimes camouflaged or posted casually. We further note that there
is another line of efforts [36,44–47] studying the Twitter message
geolocation problem which tries to identify the tweeting locations
rather than the Twitter user location discussed in this work.
Despite the promising results on improving geolocation perfor-
mance, existing state-of-the-art methods fail to identify the importance
of individual users that we addressed in this work. Arguably, while
various graph embedding techniques can be utilized for network rep-
resentation in user geolocation, understanding the influence of user
connections is more important for interpreting the behavior of the
geolocation models and therefore benefits downstream decision mak-
ings. In this spirit, we initiate the attempt to analyze theoretically
and experimentally how the properties of graph structures influence
the geolocation performance. This not only demystifies and interprets
the predictions made by the model but outlines the underlying con-
straints of existing approaches, which, in turn, should be taken into
consideration in modeling and predicting user geolocation.
2.4. Graph neural networks
Graph neural networks are effective methods models for analyzing
and learning from data on graphs, and have been successfully applied
to a variety of domains including image processing [48], social net-
works [49], transportation systems [50], etc. Existing GNN models vary
from each others on message passing mechanisms, while most of them
rely on flat information aggregations [15]. There are several hierar-
chical GNN frameworks that gradually coarsen the original graph with
pooling operation for graph classification [51,52] and image recogni-
tion [48,53]. The main difference with our work is how HGNN model
defines the graph hierarchy for clusters and exploits the geographic
information. Directly applying GraphPool [51] or HGP-SL [52] for UG
task is problematic since both of them fail to consider the relative
location of nodes w.r.t. other nodes/clusters and cannot cluster the
unlabeled nodes. Another related work is PGNN [54], recently proposed
to learn the relative position of nodes. However, it does not leverage
nodes’ geographic information that is critical for UG. More importantly,
all these methods are suitable for fully connected graph learning, while
our HGNN model is capable of incorporating unlabeled and isolated
nodes and thus is more suitable for UG task.
Despite the promising performance gains on many graph tasks,
most GNNs are still black-box models without human-understandable
model behaviors and explanations. Although GAT [16] can learn the
importance of edges and thus, to some extent, explain the node
aggregation behaviors via attention mechanism, it is limited to spe-
cific architectures and fails to provide single-instance explanations.
To adaptively adjust the influence of each node, a learnt exploitation
of information from neighborhoods of differing locality and selective
combining of different aggregations was proposed in [55]. Though
their method can automatically discover the importance of each node
in a GNN, it is not specifically designed for explaining model predic-
tions. GNNExplainer [56] was proposed to explain the predictions of
model-agnostic GNNs. It interprets the GNN models by maximizing the
mutual information between a subgraph (or a subset of node features)
and the predictions for the original graph. Another work [57] uses
image interpretation methods, such as sensitivity analysis, guided back-
propagation, and layer-wise relevance propagation (LRP), to explain
the node-level predictions. GraphLIME [58] is a local interpretable
method that captures the nonlinear dependency between features and
predictions. It then considers the perturbation near a node and uses
a linear explanation model to find features as explanations for GNNs.
X-GNN [59] proposes to find the graph patterns that maximize a
particular prediction through graph generation, which is formulated as
a reinforcement learning problem and trained with a policy gradient
method. GNN-LRP [60] is a theoretically founded XAI method for inter-
preting GNN predictions, which is derived from the higher-order Taylor
expansions based on LRP. A recent work [61] systematically reviews
existing explainable GNN methods, and proposes to enable information
fusion for multi-modal causability using interpretable GNNs.
What separates our work from the existing GNN-based approaches is
that we propose a learning model which incorporates the geolocations
and distances and we provide a greater extent of explainability.