Abstract:
With the prevalence of advanced geo-positioning techniques and the popularity of locationbased
social networks, a large amount of geo-tagged data is available. Point-of-interest (POI)
recommendation is one of the key research problems in mining location-based social network
data. Traditional recommendation methods, such as collaborative filtering and matrix factorization,
have been applied to POI recommendations. However, they can only model simple interactions
between users and items with linear models. Given that linear models are limited to
capturing linear user-item interactions, recent research proposes to use deep neural network for
modeling the non-linearity user-item interactions. The existing neural network models focus on
modeling the interactions between user and item. Socio-geographical context information such
as friends and geographical regions are important to a successful recommendation algorithm.
For example, one can recommend a restaurant for a user to visit based on the region where
the user often visits in addition to his/her preferences on POIs. Socio-geographical contexts
can be represented by a network structure with heterogeneous information, such as asymmetric
social relationships and POI-region relationships. However, the methods mentioned above are
more suitable for working on Euclidean feature space and cannot grasp the interactions among
various key factors in the socio-geographical context.
In this work, we design a recommendation framework based on heterogeneous graph neural
network structure, to address the above limitations. First of all, we model the location-based
social network as a heterogeneous graph consisting of three types of nodes, namely users, POIs
and regions. The heterogeneous graph covers various contextual information by different types
of edges connecting the three types of nodes: (1) the social context via the user-user connections;
(2) the user preferences via user-poi interactions, and (3) the geographical context
information via the Region-POI relations. Then, we extract key information for POI recommendation
based on several carefully designed meta-paths. Last, we follow the idea of graph
attention networks to learn graph node representations from the information extracted from
the heterogeneous graph. The learned node representations encode the rich socio-geographical
context information for POI recommendations. Based on the node representations, we develop
two recommendation methods. Our first method models the POI recommendation problem as
a multi-label classification problem. Precisely, by transforming the original classification problem
into ranking based recommendation, we take the user’s preference for POIs as the clas-
sification label. Our second method applies a multi-layer perceptron to transform the learned
user and POI representations to predict how likely a user will check-in a POI. Experiments on
real-world datasets demonstrate that the proposed methods outperform the state-of-the-art POI
recommendation algorithms by capturing the socio-geographical context.