Article Preview
TopIntroduction
With the rapid development of artificial intelligence (Guebli & Belkhir, 2021; Jiao et al., 2022; Tan et al., 2022), the Internet of Things (Chamra & Harmanani, 2020; Madhu et al., 2022), big data (Thirumalaisamy et al., 2022), cloud-fog computing (Thoumi & Haraty, 2022; Vijayakumar et al., 2022), smart communication (Almomani et al., 2022; Dwivedi, 2022; Ling & Hao, 2022; Samir et al., 2020), and other fields, intelligent mobile devices have become a necessity for people's public lives, and the demand for location-based services is showing explosive growth. Location-based social networks (LBSNs) have attracted a large number of users (Werneck et al., 2021). Users share their travel photos or check-in data through social networks to record access history and share life experiences, thus accumulating a large number of access footprints or check-in record data with geographical markers (Lai & Zeng, 2023; Liu, 2018; Xing et al., 2019). The historical access data of these users provide an opportunity to get insights into people's behavior and can be effectively utilized for personalized point-of-interest suggestion via social networks (Chakraborty et al., 2020; Mishra et al., 2020; Xing et al., 2018; Yang et al., 2019).
In recent years, deep-learning techniques have effectively improved the performance of POI recommendations. The widely used neural networks in POI recommendation include convolutional neural networks (CNNs) and recurrent neural networks (RNNs) (Wang et al., 2022). Among them, RNNs can model short-term check-in behavior of users, but there are problems with gradient vanishing and exploding. Therefore, gated recurrent units (GRUs) are used to better capture the long-term dependencies of user check-in behavior (Jia, 2023). Transformer is used to better extract semantic features from user check-in data (Halder et al., 2023; Wang et al., 2022; Yang et al., 2022). However, many existing methods often struggle to balance time series features, contextual semantic information, and spatial features.
To overcome the above issues, a personalized POI-recommendation method (TFCAGRU) utilizing the Transformer encoder and CAGRU is suggested. The innovations of the suggested method are as follows:
- (1)
The embedding layer and Transformer encoder are combined to capture POI semantic and spatial features.
- (2)
With the context information of time and space, the different intentions of user behavior can be better discovered by slicing the time interval and space interval in different dimensions; multilayer attention is introduced to enhance the recommendation model’s performance.
- (3)
When the context information is integrated into the gating structure, CAGRU can efficiently extract the spatiotemporal features of POI, which can better mine users' personalized interests.