MIGNN: A Multiple Intention-aware Graph Convolutional Neural
Network for POI Recommendation
Jingping Feng
, Mingxin Gan*,
and Yingxue Ma
School of Economics and Management, University of Science and Technology Beijing, Xueyuan Road, Beijing, China
School of Economics and Management, University of Science and Technology Beijing, Beijing, China
Keywords: Recommendation Systems, Deep Learning, Neural Network, Point-Of-Interest Recommendation.
Abstract: In recent years, the rapid development of urbanization and mobile communication motivated the point-of-
interest (POI) recommendation systems. Many models adopt deep learning methods to learn user and POI
embedding, and achieved some improvements. However, existing models seldom pay attention to user check-
in intentions. To address this problem, we propose a novel deep learning model to extract user check-in
intentions using graph network, namely Multiple Intention-aware Graph Convolutional Neural network
(MIGCN). We set four key modules in the proposed model, the embedding module for data preparation, the
intention decomposition module for intention learning, the intention integrating module for intention
embedding generating and the prediction module for future check-in prediction. We carried out a series of
experiments on two real-world datasets. The experimental results verified the superiority of the proposed
model compared with several the state-of-the-art methods.
The rapid development of mobile phones and
wearable devices motivated the development of
point-of-interest (POI) recommender systems. In POI
recommendation systems, following the
collaborative filtering (CF) algorithm (Zeng 2021),
the user and the POI are represented by a dense
vector, and the future check-in is predicted by the
matching score of user and POI embeddings.
Recently, researchers attempted to leverage some
context information to improve recommendation
performance, such as geographical location, check-in
time, POI category (Rahmani 2019) and social
network (WANG 2021). Nowadays, many
researchers adopted the graph neural network (GNN)
(Wang 2018) in recommendation systems, to extract
hidden features to generate better user/POI
embedding (Hamilton 2017, Kim 2018). Several
research found that grasping intention features help to
enhance the representations of users (Chang 2020,
Chang 2020). Hence, some researchers proposed
different methods of feature disentangling to discover
user intentions, such as auto-encoders (Tang 2021)
and generative models (Yang 2020), which enhance
the interpretability of user preference (Trieu 2021).
However, the difference between multiple user
intentions were ignored (Chen 2019). Specifically, in
POI recommendation systems, the real-world check-
in is influenced by different intentions. Therefore, it
is necessary to extract different intention features
from user-POI interactions. Similar tasks adopted
heterogeneous graphs to analyse various user
intentions (Guo 2020). For example, the neural graph
collaborative filtering model (NGCF) (Wang 2019),
was proposed in recommendation systems to
decouple user intentions. However, it leaves the work
of distinguish various intentions not well explored,
resulting in the loss of potential available
information. To overcome this problem, we proposed
a multiple intention-aware graph convolutional
neural network (MIGCN) to simultaneously extract
features of various users’ intentions and describe the
weight of different intention embedding.
2.1 Overview
As shown in Figure 1, the proposed MIGCN model
contains four modules: embedding module, intention
Feng, J., Gan, M. and Ma, Y.
MIGNN: A Multiple Intention-aware Graph Convolutional Neural Network for POI Recommendation.
DOI: 10.5220/0011173400003440
In Proceedings of the International Conference on Big Data Economy and Digital Management (BDEDM 2022), pages 256-261
ISBN: 978-989-758-593-7
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
decomposition module, intention integration module
and prediction module.
In Embedding Module, we construct
homogeneous graphs for both users and POIs and
generate embedding for users, POIs, and check-ins
via GCN in the embedding module. First, we generate
user embedding and POI embedding in embedding
module for data preparation.
Then, we extract the features of multiple user
intentions by decomposing the user embedding in
user graph constructed by user-POI interactions.
As for the Intention Extraction Module, we
extract the features of user intentions by
disentangling the check-in embedding in the user-
POI graph. By calculating the weight of each
intention, we achieved the fusion of the embedding of
both users and POIs by feeding them into a GCN via
an adjacent matrix. Then, we obtained all the
embedding by disentangling by a multi-head
attention mechanism and input them into the
prediction module. By calculating the weight of each
intention, we fuse the intention features in integration
Finally, we put the intention embedding and POI
embedding into prediction module, to get the
predicted probability of future check-ins. In
Prediction Module, we added an attention mechanism
to form the historical POIs embedding of the target
users, and calculated the weights of historical POIs
for the user. The prediction probability was obtained
by weighting the auxiliary reference value of the
predicted POI (the weight was adjusted as a
hyperparameter). Finally, according to the calculated
prediction probability, we ranked the POIs in
descending order as a recommendation for each user.
2.2 Embedding Module
At the beginning of the model, we use two embedding
layers to conduct the data preparation for intention
feature learning. The user id and POI id are mapped
into two dense vectors by user embedding layer and
POI embedding layer.
user idu W
POI idp W
represent the weight
matrix in user embedding and POI embedding layers.
Figure 1: Overview of MIGCN model.
2.3 Intention Decomposition Module
As for intention extraction, we first build a intention
decomposition module to extract different weights of
various intentions. Since user intentions existed in
user and POI embedding, we try to explore more
information of the users’ intentions from the
interactions of users and POIs via disentangling the
embedding of the user-POI pair
. Then, we
build two graphs for users and POIs respectively. In
the user-POI graph,
denote the
message propagation mechanisms for users and POIs
in l-th layer of the GCN.
represents the
presentation of the node and its l-1 order neighbors.
is a parameter matrix, whose parameters are
shared in the same order.
() () ( 1) () ( 1) ( 1)
mei eueiWW
we obtain the updated embedding of users in l-th
layer in the user-POI graph using GCN, as shown in
Equation (4):
() () ()
uu u p
eu LeakyRelu m m
MIGNN: A Multiple Intention-aware Graph Convolutional Neural Network for POI Recommendation
The interaction of check-ins is represented by the
Laplace matrix 𝚪
, where R is the interaction
matrix between a user and a POI, and D is the degree
matrix. In matrix R, if the user checked-in at the POI,
we define the corresponding value as 1:
As shown in Equation (6) (7), after the
propagation in the
l-th layer, we obtain the updated
user embedding
. We also update the POI
embedding, namely
* ()*(1) ()*(1)*(1)
ll ll l
eu LeakyRelu eu eu euΓ+I W ΓW
* ()*(1) ()*(1)*(1)
ll ll l
ei LeakyRelu ei ei eiΓ+I W ΓW
After the last layer, and are updated as follows:
=eu eu
=ei ei
2.4 Intention Integration
In intention integration module, we use the user-POI
embedding after the message propagation of the
disentangling mechanism to generate the integrated
multiple intention vector
, which was composed of
in l-th layer, as shown in Equation (10). Where
W represents the weight matrix. The weight matrix is
obtained via multi-head attention mechanism, as
shown in Equation (11)-(12).
() (1) (1) (1)
exp( )
exp( )
(1) ()
1..., 1
|| ( )
ikK ijj
represents the coefficient of multi-
head attention.
Figure 2: The explanation of intention decomposition and
The final intention decoupling is shown in Figure
2. Through decoupling the intention to “accompany
friends”, “need a dress”, “kill time”, and so on of
aspect 1, and “hungry”, “cheap price”, and so on of
aspect 2, we can obtain a more adequate
representation of features.
2.5 Prediction
In order to make the difference between the positive
sample and the negative sample as large as possible,
we use the loss function of BPR. We regard check-in
samples as positive samples
(,)eu ei
, and regarded
no check-in samples as negative samples
eu ei
[50]. To facilitate data processing, we randomly
choose a negative sample to match a positive one. The
formula derivation is shown in Eq. (13).
Loss ln ( ) || ||
eu ei
eu ei ei O
eu ei
Further, we joined the regularization 𝜆
reduce over-fitting. Here, 𝛿 represents the function
of sigmoid. 𝑦
indicates that the probability user
will check-in at POI 𝑖 , and 𝑦
that the predicted user will not check-in at POI 𝑖
3.1 Experiment Settings
Table 1: Experimental datasets.
Datasets Users POIs Check-ins Density
Foursquare 3,705 53,383 836,280 7.2*10
Gowalla 29,858 40,981 1,027,370 6.4*10
To evaluate the performance of MIGCN model, we
choose two real-world check-in datasets, Foursquare
and Gowalla, in our experiments. The datasets cover
user id, POI id and check-in timestamp. The statistic
of the datasets is shown in Table 1. The dataset was
divided into training and test sets in chronological
order, the first 70% check-ins are included in training
set, and the last 30% check-ins are included in test set.
The negative samples are generated by random
We adopt two widely used metrics, Recall and
NDCG at top 10 to access the recommendation
performance. We compare the proposed model with
the following baselines: MF (
Lian 2020) is a typical
recommendation method. Both user and POI are
mapped into a hidden space by the matrix
factorization algorithm. NGCF (
Wang 2019) is a
BDEDM 2022 - The International Conference on Big Data Economy and Digital Management
combination model of graph neural network and
collaborative filtering. The encoding of user-POI
interaction is explicitly embedded in the embedding
representation. DGCF (
Wang 2020) is a model based
on graph neural network for feature disentangling,
which combines a new neighbor routing mechanism
into a message propagation mechanism.
3.2 Performance Comparison
We compared the proposed MIGCN model with the
baselines on two datasets in terms of Recall and
The experimental results are shown in Table 2:
The proposed model outperforms the traditional
MF model. For example, in terms of Recall, MIGCN
achieves ~45.11% improvement compared with MF.
In terms of NDCG, MIGCN achieves ~43.82%
improvement compared with MF.
As for the advance graph neural network model,
such as NGCF and DGCF, MIGCN also achieves
significant improvement. For example, MIGCN
achieves ~19.79% and ~14.29% improvements than
NGCF and DGCF models in terms of Recall, and
achieves ~25.38% and ~19.27% improvements than
NGCF and DGCF models in terms of NDCG.
These results illustrate the MIGCN model was
better than MF, NGCF, DGCF, and verify the
superiority of the weighted intention decoupling and
intention combination in the proposed model.
Table 2: Performance of models.
Dataset Metric MF NGCF
Recall@10 0.3137 0.3800
Recall@20 0.4632 0.5200
NDCG@10 0.3400 0.3900
NDCG@20 0.4900 0.5250
Recall@10 0.3240 0.3890
Recall @20 0.4770 0.5380
NDCG@10 0.3500 0.3920
NDCG@20 0.5200 0.5960
Dataset Metric DGCF MIGCN
Recall@10 0.3983 0.4552
Recall@20 0.5410 0.6152
NDCG@10 0.4100 0.4890
NDCG@20 0.5310 0.5860
0.4150 0.4452
Recall @20 0.5760 0.6350
NDCG@10 0.4130 0.4991
NDCG@20 0.6030 0.6130
3.3 Visualization of Check-in based on
Figure 3: Visualization of Check-in based on MIGCN.
In order to explore the ability of the proposed
model in depicting different user intentions, we
visualize the intention vectors of several selected
users. as shown in Figure 3, we find that user
intentions in different aspects have different weights,
which determines user's check-in decision in real-
world. The differences in the colour of the four
intention aspects show that the proposed model
indeed distinguish different user intentions.
Figure 4: Visualization of weights in multi-head attention.
Further, we visualize the weights in multi-head
attention module to verify the positive effect of
attention mechanism on intention feature learning. As
shown in Figure 4. The dimensions of x axis and y
axis represented the six dimensions of user intention
feature, and different shades of colour represent
different weights of attention. From Figure 3 (a), we
see the red area on the (5, 5) head, which illustrated
that the intention represented by the 5-th head is the
most important one for predicting user check-ins.
However, the 0-th head has tiny contribution on
future check-in prediction, as the area around the 0-th
head is filled with the lightest colour. As shown in
Figure 3 (b), the (3, 5) head is coloured in dark blue,
it illustrated that the 3-th head had the great
importance on MIGCN.
MIGNN: A Multiple Intention-aware Graph Convolutional Neural Network for POI Recommendation
There may be multiple intentions that motivate users
to check-in in real life. We think different intentions
have different influence on user’s check-in decision.
Only a few existing studies address the learning of
multiple intentions. However, using the features of
multiple intention in recommendation algorithm is
conducive to the understanding of user preference,
and then to improve the recommendation
performance. In this paper, we aim to design an
intention representation model to enrich the
characterization of users and POIs for
recommendation. To make up for the deficiency of
GNN, we used the multi-head attention mechanism
and self-attention mechanism to focus on more
important POIs. The designs of the intention
extraction module and prediction module can capture
complex relationships between users and POIs,
hence, we can learn features and obtain more accurate
recommendations. Furthermore, the intention that is
extracted from our model has the ability to explain
the user’s check-in. This helps alleviate the problem
of insufficient features. We conducted a series of
experiments on two datasets to verify the
effectiveness of the proposed model. The comparison
results show that the proposed model outperforms the
state-of-the-art recommendation models.
For discussion, we attribute the effective
performance of the proposed model into the
following aspects:
(1) The proposed multiple intention graph neural
network model not only effectively describes user’s
multiple intention, but also calculates the weights of
different intentions from different views when
integrating it.
(2) The proposed method conducts an exhaustive
mining from user-POI interactions, it aggregated and
updated the embedding vectors of users and POIs.
Through the analysis of the datasets, we found that
users generally have different check-ins under
different intentions.
(3) The proposed model adopted an attention
mechanism to capture user intention in different
layers. We added multi-head attention mechanism in
the proposed model to integrate the multiple intention
features, serving to future prediction.
(4) The proposed model combines different
intention features with the model of historical check-
in interactions. The greater the weight, the more
important the corresponding intention feature plays in
future check-in prediction.
However, we do not consider other information,
such as comments, time, etc., which results in single-
dimensional information and thus cannot
dynamically capture dynamic user preference. In
future work, we will introduce additional auxiliary
information to capture dynamic changes for user
intention and further interpret dynamic intention
representations, such as in conjunction with the
temporal information of check-in data, which will
make feature representation more complete.
This research was supported by the National Natural
Science Foundation of China (Nos. 71871019,
71471016, 71729001).
Chang, B., Jang, G., Kim, S., & Kang, J., 2020. Learning
graph-based geographical latent representation for
point-of-interest recommendation. In Proceedings of
the 29th ACM International Conference on Information
& Knowledge Management. 135-144.
Chang, J., Gao, C., He, X., Jin, D., & Li, Y., 2020. Bundle
recommendation with graph convolutional networks. In
Proceedings of the 43rd International ACM SIGIR
Conference on Research and Development in
Information Retrieval. 1673-1676.
Chen, T., Yin, H., Chen, H., Yan, R., Nguyen, Q. V. H., &
Li, X., 2019. Air: Attentional intention-aware
recommender systems. In 2019 IEEE 35th International
Conference on Data Engineering (ICDE). 304-315.
Guo, X., Shi, C., & Liu, C., 2020. Intention Modeling from
Ordered and Unordered Facets for Sequential
Recommendation. In Proceedings of The Web
Conference 2020. 1127-1137.
Hamilton, W. L., Ying, R., & Leskovec, J., 2017. Inductive
representation learning on large graphs. In Proceedings
of the 31st International Conference on Neural
Information Processing Systems. 1025-1035.
Kim, H., & Mnih, A., 2018. Disentangling by factorising.
In International Conference on Machine Learning.
Lian, D., Wu, Y., Ge, Y., Xie, X., & Chen, E., 2020.
Geography-aware sequential location recommendation.
In Proceedings of the 26th ACM SIGKDD
International Conference on Knowledge Discovery &
Data Mining. 2009-2019.
Rahmani, H. A., Aliannejadi, M., Mirzaei Zadeh, R.,
Baratchi, M., Afsharchi, M., & Crestani, F., 2019.
Category-aware location embedding for point-of-
interest recommendation. In Proceedings of the 2019
ACM SIGIR International Conference on Theory of
Information Retrieval. 173-176.
BDEDM 2022 - The International Conference on Big Data Economy and Digital Management
Tang, P., Peng, K., & Dong, J. (2021). Nonlinear quality-
related fault detection using combined deep variational
information bottleneck and variational autoencoder. J.
ISA transactions, 114: 444-454.
Thanh Trieu, N., Pottier, B., Rodin, V., & Xuan Huynh, H.,
2021. Interpretable Machine Learning for
Meteorological Data. In 2021 The 5th International
Conference on Machine Learning and Soft Computing.
& CHEN, E. (2021). HyperSoRec: Exploiting
Hyperbolic User and Item Representations with
Multiple Aspects for Social-aware Recommendation. J.
ACM Transactions on Information Systems, 40, 1-28.
Wang, H., Zhang, F., Wang, J., Zhao, M., Li, W., Xie, X.,
& Guo, M., 2018. Ripplenet: Propagating user
preferences on the knowledge graph for recommender
systems. In Proceedings of the 27th ACM International
Conference on Information and Knowledge
Management. 417-426.
Wang, X., He, X., Wang, M., Feng, F., & Chua, T. S., 2019.
Neural graph collaborative filtering. In Proceedings of
the 42nd international ACM SIGIR conference on
Research and development in Information Retrieval.
Wang, X., Jin, H., Zhang, A., He, X., Xu, T., & Chua, T.
S., 2020. Disentangled Graph Collaborative Filtering.
In Proceedings of the 43rd International ACM SIGIR
Conference on Research and Development in
Information Retrieval. 1001-1010.
Yang, Y., Qiu, J., Song, M., Tao, D., & Wang, X., 2020.
Learning propagation rules for attribution map
generation. In European Conference on Computer
Vision., 2020, 672-688.
Zeng, J., Tang, H., Zhao, Y., Gao, M., & Wen, J. (2021).
PR-RCUC: A POI Recommendation Model Using
Region-Based Collaborative Filtering and User-Based
Mobile Context. J. Mobile Networks and Applications,
MIGNN: A Multiple Intention-aware Graph Convolutional Neural Network for POI Recommendation