Efficient Trajectory Prediction Using Check-In Patterns in Location-Based Social Network

Abstract

Location-based social networks (LBSNs) leverage geo-location technologies to connect users with places, events, and other users nearby. Using GPS data, platforms like Foursquare enable users to check into locations, share their locations, and receive location-based recommendations. A significant research gap in LBSNs lies in the limited exploration of users’ tendencies to withhold certain location data. While existing studies primarily focus on the locations users choose to disclose and the activities they attend, there is a lack of research on the hidden or intentionally omitted locations. Understanding these concealed patterns and integrating them into predictive models could enhance the accuracy and depth of location prediction, offering a more comprehensive view of user mobility behavior. This paper solves this gap by proposing an Associative Hidden Location Trajectory Prediction model (AHLTP) that leverages user trajectories to infer unchecked locations. The FP-growth mining technique is used in AHLTP to extract frequent patterns of check-in locations, combined with machine-learning methods such as K-nearest-neighbor, gradient-boosted-trees, and deep learning to classify hidden locations. Moreover, AHLTP uses association rule mining to derive the frequency of successive check-in pairs for the purpose of hidden location prediction. The proposed AHLTP integrated with the machine-learning models classifies the data effectively, with the KNN attaining the highest accuracy at 98%, followed by gradient-boosted trees at 96% and deep learning at 92%. Comparative study using a real-world dataset demonstrates the model’s superior accuracy compared to state-of-the-art approaches.

1. Introduction

The rapid growth of mobile technologies has attracted research attention to studies focusing on location-based social network (LBSN) data. Numerous specialized areas, like the study of people’s behavior in festivals, shopping centers, dining venues, tourism, and many more, have used LBSN data for analysis. These types of data contain heterogeneous features about customers from various venues; in order to perform more specialized studies, researchers must filter out the data related to particular venues. The dataset frequently contains thousands or millions of records before any data analysis, and it is necessary to manually filter out data pertinent to particular venue classifications. This is a laborious and time-consuming task for this kind of research [1,2,3].

Thus, in order to organize data without requiring manual labor, machine-learning techniques that can categorize the data according to certain features are needed. Using check-ins from diverse venues, such data offers a sample of different features of human actions and characteristics whilst engaging with LBSN throughout an array of activities. Information about broad demographic trends can be gathered from the analysis of such actions, and such information can be used to organize and develop celebrations, playgrounds, eateries, retail centers, and, eventually, smart cities. Additionally, the LBSN data have been utilized in more specific studies that have been shown to be extremely beneficial in various domains, such as determining the factors that influence restaurant appeal, the importance of parks, and tourism actions, among many others [4]. However, within the vast number of records, it is crucial for these specialist studies to take into account the data pertinent only to these venues and manually classify the unique data for each individual research project. Large volumes of data are one of the benefits of using LBSN data to study human behavior, but classifying the data to discover pertinent information can frequently be more challenging and time-consuming [2].

In location-based social networks, users and places are the two primary elements that engage with one other. This results in three distinct types of networks constituting an LBSN: (1) user–user social network, (2) user–location network, and (3) location–location spatial network. Figure 1 is a schematic illustration of these three networks and their interactions. Users are interconnected by relationships with family, friends, and colleagues, much like a traditional social network. However, it is also possible to specify spatial network connections by providing locations [5]. Subsequently, LBSN’s user–location network is a crucial element that links individuals and places.

With the help of LBSN’s check-in feature, users can engage in a variety of activities, maintain social connections with other users, and communicate spatiotemporal information (position and time). Since check-ins occur in real time, the trajectories derived from them depict a user’s actual physical movement. Consequently, check-in services in LBSNs fill the gap between the virtual and real worlds of social media. In recent years, the location prediction problem has gained much popularity [7,8]. It has been discovered, meanwhile, that people are hesitant to disclose places they have visited in private in their publicly available check-in trajectories. Advertising can more effectively target consumers by identifying unchecked places.

A large spatiotemporal dataset analysis reveals that check-ins are typically clustered around a small number of important points of interest [9]. This challenge is discussed in Section 5. The location prediction task uses these locations’ geographical coordinates and their semantic meaning as inputs. Furthermore, comparable check-ins made by people with similar social connections may be helpful in location prediction [10].

This paper primarily focuses on hidden locations for prediction, which involves locations that users privately visit but do not publicly check into. Scientifically, this can be identified as a sparsity problem in check-in data. An illustrative example, considering a user’s daily commute: First, the user leaves their home, passing several streets that are often tracked. On their way, they stop at a park for a break. The park’s location is known to locals, but the user prefers not to reveal that they visit it. Afterwards, the user continues to their office, where they check in publicly. The park visit remains hidden, presenting a challenge for location-based prediction models.

Given a published trajectory (l_a → l_b → l_c → l_d), the challenge is to infer unrecorded locations between successive check-ins, such as predicting a missing location between (l_b → l_c), which is known as the hidden location prediction problem. Unlike the next-location prediction problem, hidden location prediction aims to uncover unshared visits using historical check-in data.

Trajectory prediction plays a crucial role in developing applications for smart cities in a variety of fields, such as intelligent transportation systems [11]. It seeks to anticipate each entity’s future location over a specified period. Recently, trajectory prediction has attracted significant attention from a variety of academic fields, including machine learning, web and information retrieval, mobile computing, and others [12]. Basically, a trajectory can be shown as a series of nodes, where each node represents a place at a certain time interval. Trajectory data typically consist of spatial and temporal information, such as the sequence of locations or positions a user visits over time.

The key challenges of this research as mentioned above are the difficulty of predicting unchecked locations, along with the issue of data sparsity, which implies a substantial spatiotemporal dataset wherein check-ins are generally concentrated around a limited number of significant points of interest.

In this paper, we propose a model named Associative Hidden Location Trajectory Prediction (AHLTP), which employs a data mining technique to infer hidden locations from users’ publicly available trajectories. Addressing this challenge will unlock a more accurate, efficient location prediction model. AHLTP utilizes association rule mining to extract the frequency of consecutive check-in pairs. Mining the user check-in sequences, the model enhances the identification of implicit mobility patterns. Eventually, by integrating the preprocessed and resolved locations with the classifier, the AHLTP model significantly improves overall trajectory prediction, offering a more comprehensive and precise understanding of user movement behavior and dynamics. To further assess the efficiency of the AHLTP model, a comparison study is made with recent previous studies that have employed the same dataset.

The key contributions of the paper can be outlined as follows:

• Machine Learning for LBSN Data Analysis: Machine learning techniques are utilized to categorize LBSN data into popular location categories and predict user and resident mobility patterns.
• Hidden Location Prediction: A method is developed for predicting hidden or privately visited locations that are not explicitly disclosed in a user’s publicly available trajectory, thereby reconstructing a more complete movement history.
• Association Rule Mining for Hidden Locations: Consecutive check-in pairings are identified and leveraged to infer hidden locations, hence enhancing trajectory prediction accuracy.
• Semantic Context Analysis: Implicit attributes are considered for evaluation, such as the semantic aspects of locations (e.g., location types), to improve the understanding of user mobility behavior.
• Comprehensive Real-World Evaluation: Extensive experiments on real-world LBSN datasets are conducted, demonstrating that the proposed model outperforms state-of-the-art techniques in trajectory prediction.

Location prediction in LBSNs is an exciting yet challenging task. One major challenge is handling the study of users’ behavior to conceal information about specific hidden locations and investigate these hidden locations. Addressing this challenge will unlock a more accurate, efficient location prediction model. We will further discuss this research challenge in the following sections and how the proposed AHLTP model considers it.

The rest of the paper is organized as follows. Section 2 surveys the related work, while Section 3 states the empirical data analysis. The problem definition and proposed model are presented in depth in Section 4. Section 5 reviews the experimental evaluation, and Section 6 discusses the important key findings. Finally, the paper is concluded in Section 7 with a discussion of future work.

2. Related Work

Numerous analyses of user behavior in social networks have been made possible by the simplicity of spatiotemporal data collection technologies. These fall under several categories, such as location prediction, locating hidden social relationships, locating similar users, and spatiotemporal data analysis [1]. Spatiotemporal data can be acquired via the trajectories of moving objects captured by GPS devices, social events (e.g., postings, microblogs) that include location tags and timestamps, and environmental monitoring (e.g., forecasts, weather), etc. Concealed social connections in online social networks pertain to recognizing relationships among individuals that lack direct links within the social graph [13].

Zhao et al. [14] presented a spatiotemporal gated network, named STGN, for next points of interest (POI) recommendation by enhancing a long short-term memory network. Experiments were conducted to evaluate the performance of STGN on Foursquare, Gowalla and brightkite. Si et al. [15] introduced the APRA-SA, an adaptive point of interest recommendation method that incorporates user behavior and location data. Experiments were applied on Foursquare and Gowalla datasets. Zhao et al. [16] examined the emotional characteristics of locales by integrating user similarities with location data, alongside the POI recommendation method. They worked on Sina Weibo as the dataset. Neither of these works accounted for the time effect or user classification.

A heterogeneous graph embedding technique is used in [17] to model the trust-aware concept for LBSN, representing the components of LBSN and their relationships (user–location, user–user). In [18], the authors presented another graph convolutional network (PPR_IGCN) model that integrates social impact and cooperative influence into POI recommendations. It extracted the potential features of users and POIs, the results were conducted on Foursquare and Yelp datasets. Based on the anchoring effect, a latent Dirichlet allocation (LDA) model for the POI recommendation system was introduced in [19]. This model emphasized the critical role of initial check-in data and was evaluated on the Gowalla dataset. These studies, however, disregarded the temporal sequential influence of check-ins.

The impact of temporal sequential pattern has been discussed in [20], For frequent pattern mining of time-informed sequences, the authors presented the TimeInformed Pattern Mining (TiPam) technique. This technique is based on actual data that was gathered in Kampala, Uganda.

The investigation of data patterns related to venues is one of the main areas of inquiry in the study of LBSN [21]. Authors in [22] introduced a method for large-scale trajectory prediction called Relationship-aware Adaptive Hierarchical Graph Learning, or REAHG. They worked on hidden relationships between users and applied an evaluation based on a real-world Foursquare dataset and other taxis datasets. Another RNN approach utilizing spatiotemporal contexts to model sparse user mobility traces is introduced in [23]; the model was evaluated on Gowalla and Foursquare datasets. They claimed that their results outperform state-of-the art spatiotemporal RNNs by 15.9% when tackling the next location prediction task. In [24], location categories were identified utilizing various machine-learning algorithms for the prediction of venue classes. They registered SinaWeibo (or Weibo), one of the most extensively utilized location-based social networks in China. Four different classifiers were applied on a Chinese dataset, including the generalized linear model, logistic regression, deep learning, and gradient-boosted trees. The results showed the deep learning with 99% accuracy, gradient-boosted trees with 93%, logistic regression with 91%, and generalized linear model with 85% accuracy. Xu, Shuai, et al. [25] introduced a Venue2Vec model that makes it possible to predict a user’s next check-in location, as well as their future check-in location. The Foursquare real-world dataset was used for research in three separate cities: NYC, CA, and TKY.

The authors of [26] provided a framework for user identity linkage that helps locate people across networks and formalizes the relationship between geolocations and texts to help in location prediction. Two datasets are used in the experiment: the trajectories data from Foursquare and Dianping3, a Chinese dataset. According to their model, they claimed that the prediction accuracy was 89% and 96% for Foursquare and the Chinese dataset, respectively. In [27], the authors introduced self-ensembled contextual Thompson sampling (SECTS). The strategy sought to address the issues of cold-start users and data sparsity within the target domain. For experimentation, they utilized two real-world datasets, Gowalla and Foursquare, achieving an accuracy of 65% for point-of-interest recommendations.

Some related works considered the time context. As an illustration, Mazumdar, Pramit, et al. [28], when analyzing users’ movement habits, separated time into workdays and weekends. In other words, they were limited to making predictions in two time modes. In contrast, Cao et al. [29] analyzed users’ day-mode and week-mode mobility patterns to predict users’ future check-in positions in any fine-grained period.

Table 1. Summary of different cross-domain recommendation methods in related work.

	Dataset	Methodology Used	Model	Implicit (Hidden)/Explicit	Challenges	Association Rule Mining
Liu, J.; Yi et al., 2023 [18]	Yelp Foursquare	collaborative embedding vector between user and POI	graph convolutional network (PPR_IGCN)	Explicit data	Extract the potential features of users and POIs Data sparsity	Not applied
Yan, Hua, et al., 2023 [22]	Taxis in San Francisco Taxis in Beijing Foursquare	hierarchical graph pooling mechanism using spatial-temporal sequence representation	Relationship-aware Adaptive Hierarchical Graph Learning, or REAHG	Hidden data	Most of the graph convolutional networks rely only on local information, such as nearby neighbors	Not applied
Khan, Naimat Ullah, et al., 2023 [24]	Chinese dataset “Weibo”	machine learning classification and deep learning	Prediction of User Activities Using Machine Learning Models	Explicit data	Analyzing user activities and behavior	Not applied
Shao, Jiangli, et al., 2021 [26]	Foursquare Chinese dataset “Dianping3”	Multi-Layer Perception (MLP)	User Identity Linkage model	Explicit data	Utilizing asymmetric information for user identity linkage	Not applied
Acharya et al., 2025 [27]	Gowalla Foursquare	self-ensembled domain adaptation (SEDA) technique, which is a combination of transfer learning and reinforcement learning,	self-ensembled contextual Thompson sampling SECTS	Explicit data	Identifying cold-start users Data sparsity in the target domain	Not applied
Ghanaati, F., et al., 2023 [30]	Gowalla Foursquare	matrix factorization method (MF) in collaborative filtering CF-based approaches	extended attention gated recurrent unit (EAGRU)	Hidden data	Considering the efficacy of contextual information similarly	Not applied

summarizes the related studies and their methodologies for location prediction.

Although previous research proposed several effective methods, there are still some issues that need to be addressed to enhance the prediction of hidden location trajectories. First, while temporal sequential patterns for LBSNs have been developed to reflect user actions and behaviors, there has been limited research on understanding the preferences of users when checking into POIs and how to categorize these preferences. Moreover, the association relationships between checked-in locations are often not considered when making predictions. Besides, identifying hidden locations that are not displayed in a user’s publicly available trajectory.

3. Problem Definition

This section outlines how the proposed model was inspired by findings of hidden (implicit) relationships and dynamic relationships for sequential patterns. The section begins by describing the problem statement, followed by an explanation of the proposed AHLTP model.

Problem of Hidden Relationships

A sequential pattern is the behavior or routine in which people, groups, or objects move. It refers to approaches working with sequential patterns normally considering the interactions between people and concentrates on building relationships using geographical features and manually created rules. For example, two individuals connected to each other and walking together have similar trajectories. However, it is possible for different people who are not connected to each other to share similar movement patterns and the same interests. As illustrated in Figure 2, for instance, users A and B are not connected to each other; however, during weekdays, their patterns of mobility are similar: They always leave their residence to go to work, then to an event, then to eat lunch or dinner, and return to their residence. Such comparable connections are concealed in people’s actions and are not observable by rules that are created by hand.

In order to simplify and accurately present the problem definition, the temporal information is left out, and the checked-in places will be ordered in chronological order of occurrence within the sequence trajectory. Consequently, T_i = {c₁→ c₂ → c₃ → c₄ → … c_n} represents the published check-ins of an active user n. Notably, a location may show up more than once in a series. The proposed solution predicts the hidden location through the four steps:

• Build the user trajectory Ti, which can be represented as a set of nodes {c₁ → c₂ → c₃ → … c_n}, each of which stands for a location at a specific point in time.
• Choose a series of check-ins as possible candidates for predicting a hidden location.
• Predict a collection of unchecked or hidden locations for every pair of successive check-ins that are chosen.
• Sort the anticipated locations according to the chance of occurrence.
• The trajectory prediction indicates the prediction of the whole check-in trajectory whether it includes explicit locations or hidden locations. In other words, the hidden location prediction is a part of trajectory prediction.

The proposed AHLTP model utilizes the association rule mining technique to deduce relationships among the submitted check-ins to predict the next location of a user based on historical movement patterns.

4. Proposed AHLTP Model and Methodology

The associative hidden location trajectory prediction (AHLTP) model consists of four key processing steps, as illustrated in Figure 3. Data pre-processing involves cleaning and structuring raw check-in data to remove noise and standardize formats for further analysis. Feature extraction is performed on check-ins that are transformed into trajectories, and the trajectories are built as a series of POI categories. Associated locations are applied by employing association rule mining to uncover hidden relationships between visited locations. Finally, classification into several types of venues categorizes the inferred locations into predefined venue types based on extracted features, improving the model’s accuracy in predicting user movement patterns.

4.1. Data Pre-Processing

To transform the raw check-in data into a form that is both intelligible and helpful, pre-processing is an essential step. Missing, weakly managed, and out-of-range values are common problems with data. This is done in an effort to improve space performance and accuracy. As the amount of missing data is minimal and does not introduce significant bias, rows with missing values may simply be dropped. In order to determine the future location, trajectory pre-processing includes data cleaning, as the users with particularly low check-in records are excluded. To assess behavioral changes, it is essential to define discrete time intervals: one week (including weekdays and weekends), which will serve as an atomic unit for further temporal analysis.

4.2. Feature Extraction

Data splitting and feature selection: Using the location’s semantic properties (location type) is one way to take into account one of the implicit attributes. As shown in

Table 2. A sample of features in Foursquare check-in dataset.

UserID	VenueID	VenueCatID	Latitude	Longitude	UTC Time
470	49bbd6c0f964a520f4531fe3	4bf58dd8d48988d127951735	40.71981	−74.0026	Tue Apr 03 18:00:09 +0000 2012
979	4a43c0aef964a520c6a61fe3	4bf58dd8d48988d1df941735	40.60679958	−74.04416981	Tue Apr 03 18:00:25 +0000 2012
69	4c5cc7b485a1e21e00d35711	4bf58dd8d48988d103941735	40.71616168	−73.88307006	Tue Apr 03 18:02:24 +0000 2012
395	4bc7086715a7ef3bef9878da	4bf58dd8d48988d104941735	40.7451638	−73.98251878	Tue Apr 03 18:02:41 +0000 2012

, the dataset includes the features as extracted during the data acquisition. Using the venueCategoryID, the crawling process was implemented on the dataset to add the location’s type. The location’s type feature has an important impact on our AHLTP model, as we build the trajectory by using it.

After the dataset is crawled to extract the primary category_name variable, discrete time intervals are defined as follows: one week (including weekends and weekdays). Eventually, the trajectories are stored as a separate dataset, where each trajectory is represented by a series of nodes, each denoting a place at a particular moment in time.

4.3. Inferring Associated Locations

The AHLTP model investigates association rule mining [31] to determine the relationship between locations based on social network users’ publicly available check-in trajectories. Association rule mining is predominantly utilized on market basket data to forecast user behaviors. It can also yield substantial insights into user mobility from public trajectory data. Utilizing association rules on check-in data will help uncover correlations among the check-ins.

In AHLTP, frequent patterns of check-in locations are mined using the Frequent Pattern growth technique. Both the number of candidate location sets and the overall quantity of database crawls needed can be diminished with FP growth. Support, Confidence, and Lift are essential metrics that guide the discovery of meaningful patterns from data.

To formulate the association rules, we initially compile the users’ location datasets along with those of associated users, which will be analyzed to filter and organize location elements based on the established support and confidence levels. Next, the lift value is computed using confidence and support. Confidence is a direct measure of the usefulness of a rule for predicting the next step in a trajectory. Higher confidence means more reliable predictions of future locations in the trajectory. To exclude the least representative POI and decrease the number of candidates, locations with a confidence level below 0.8 are excluded from the trajectories. Concurrently, we gather the sets of frequently visited locations and organize the objects based on their respective lift values, with the most frequented location positioned at the top. The steps for inferring associated locations are demonstrated in Figure 4.

4.4. Classification into Several Types of Venues

For classifying the venues’ types, the following three machine learning techniques: deep learning, KNN, and gradient boosted tree, are suggested. Each of these classifiers is well-suited for LBSN prediction due to their strengths in handling spatial-temporal data and complex feature interactions [1,18,24].

4.4.1. K-Nearest Neighbor

One of the most basic machine learning methods for prediction is the K-nearest neighbor classifier. With cluster members serving as the objects and the clustering criterion being determined by their distance from one another, the K-nearest neighbor classifier creates clusters. KNN remains one of the initial algorithms learned in data science because of its simplicity and precision. As the dataset expands, KNN’s efficiency diminishes, adversely affecting overall model performance. It is frequently employed for basic recommendation systems, pattern recognition, data mining, financial market forecasting, intrusion detection, and other applications [32].

The Euclidean distance is frequently employed. The following formula is used to calculate the Euclidean distance for two entities having n-dimensional feature vectors, p = {p₁, p₂, …, p_n} and q = {q₁, q₂, …, q_n}:

$$ED(p,q)=\sqrt{{\sum }_{i=1}^n{{(p}_i-q_i)}^2}$$

(1)

The anticipated entity is then allocated to the cluster that contains the majority of its K-nearest neighbors. Lastly, it is anticipated that the entity will produce the same results as the entities in its cluster. The interval, nominal, and ordinal data types are features that the K-nearest neighbor classifier handles. Depending on the dataset being utilized, preparation may occasionally be required for data manipulations [33].

4.4.2. Deep Learning

This neural network-based technique is employed using the H2O framework along with utilizing layered data information to uncover relevant patterns and classes. Based on the data at hand, each neuron is taught via alteration and functions as a classifier by combining predictions for the output. It functions in an adaptive way, reducing the effort and time needed for categorization by naturally optimizing the neurons by developing with no human involvement. The design of feed-forward artificial neural networks serves as the foundation for this deep learning technique. The ability to understand intricate correlations connecting both input and output factors is made possible by the hidden layers. It employs supervised learning methods, and in order to develop the algorithm using our earlier research, designated training information is needed. By maximizing a loss function, the algorithm discovers throughout development the associations connecting the variables being used and the target classes. The difference across the true and expected outputs is measured by the loss function.

In deep learning models, the Rectified Linear Unit is the most often utilized activation function due to its simpler mathematical procedures, as ReLU requires less computing power than other activation functions like tanh and sigmoid. Because just a small number of neurons are active at once, the network is sparse, which makes computation simple and effective [12]. The ReLU function returns 0 if it receives any negative input, but for any positive value x, it returns that value back. So it can be written as f(x) = max(0, x).

4.4.3. Gradient Boosted Trees

Gradient Boosted Tree (GBT) is an ensemble learning method that builds multiple decision trees sequentially, where each new tree corrects the errors of the previous ones. This iterative approach helps capture complex non-linear relationships in location-based social network (LBSN) prediction, making GBT robust for predicting user mobility, check-in behaviors, and venue preferences [14]. The effectiveness of using a gradient-boosted tree-based model for the classification of LBSN data can be observed in the results section.

To optimize the model’s performance, the following GBT hyperparameters are tuned:

• Learning Rate (eta)—Controls the contribution of each tree to the final model. Default: 0.1, but in LBSN, a lower value (e.g., 0.05) can prevent overfitting, improving generalization.
• Number of Trees (n_estimators)—Defines how many trees are built. A higher number (50–300 trees) can improve accuracy, but too many may lead to overfitting.
• Maximum Depth (max_depth)—Determines how deep each tree can grow. Shallow trees (depth = 4–6) prevent overfitting while still capturing complex relationships.

5. Empirical Data Analysis

This section presents an overview of the primary approaches for evaluating work performed in location-based social networks (LBSNs) and the associated public LBSN datasets. Researchers utilize several LBSNs to examine the challenge of point-of-interest (POI) recommender systems, with four notable platforms: Foursquare, Gowalla, Brightkite, and Yelp. Owing to privacy concerns associated with publicizing such data, real-world LBSN datasets are few. Additionally, service providers are reluctant to share researchers’ huge datasets since they view these datasets as vital when it comes to offering a competitive product. The main characteristics of publicly accessible datasets that are frequently utilized by the LBSN research community are listed in

Table 3. Real-World LBSN datasets that are openly accessible [8].

Dataset	#Users	#Locations	#Checkins	#Links	Period
Gowalla	319 K	2.8 M	36 M	4.4 M	20 mo
BrightKite	58 K	917 K	4.49 M	214 K	30 mo
Foursquare	2.7 M	11.1 M	90 M	0	5 mo
Yelp	1.00 M	144 K	4.10 M	0	36 mo

.

An empirical data analysis was performed on Foursquare dataset to identify the LBSNs’ check-in trends with regard to spatial and semantic parameters, making use of the worldwide check-in data gathered in [34]. The Foursquare dataset includes check-in data venues in NYC collected from Foursquare from 24 October 2011 to 20 February 2012. It contains 227,428 check-ins. The data source used in the current study is acquired from Foursquare, containing the following attributes:

• User ID (anonymized)
• Venue ID (Foursquare)
• Venue Category ID (Foursquare)
• Latitude
• Longitude
• UTC time

As has been outlined in the previous section, recent research methods on POI recommendation focused on explicit interactions of LBSN objects such as users’ check-ins on POIs and social relationships, while neglecting implicit attributes (e.g., hidden locations and the association relations between locations) that cannot be directly observed but may notably contribute to the POI recommendation.

In this paper, a recommendation for considering one of the implicit attributes is to use the semantic properties of the location (location type). The main category of the location can be added by crawling the Foursquare dataset.

As seen in Figure 5, the chosen check-ins are plotted over an actual geographic map. The analysis of the dataset shows that the users’ check-in frequency is distributed unevenly. Figure 5 addresses the challenge of a large spatiotemporal dataset that shows check-ins are typically clustered around a small number of important points of interest.

As shown in the statistics count of each venue type in Figure 6, this dataset includes check-in, tip, and tag data of mainly food venues in NYC and nine other venue activity types (Travel & Transport, Shop & Service, Residence, Outdoors & Recreation, Professional & Other Places, Nightlife Spot, Arts & Entertainment, College & University, and Event).

6. Experimental Evaluation and Results

For the AHLTP model evaluation, the performance of classifiers is evaluated using two measures: Accuracy and Cohen’s kappa.

• To compute the accuracy measure, the confusion matrix is calculated. The accuracy is calculated by dividing the number of correct predictions by the total number of predictions made by the model.

Accuracy = Correct Predictions/All Predictions

(2)

• Cohen’s kappa ($\kappa$) is a statistical measure used to assess the agreement between two raters (or classifiers) who each classify $N$ items into $C$ mutually exclusive categories. It accounts for the possibility of an agreement occurring by chance.

$$\kappa ={\displaystyle \frac{P_o-P_e}{1-P_e}}$$

(3)

where:

• $P_o$ (Observed Agreement) is the proportion of times the two raters agree.

$$P_o={\displaystyle \frac{\sum _{i=1}^C{n}_{ii}}{N}}$$

(4)

where $n_{ii}$is the number of instances both raters assigned to the same category $i$ and $N$ is the total number of instances.

• $P_e$ (Expected Agreement by Chance) is the proportion of agreement that is expected if both raters are classifying randomly.

$$P_e={\displaystyle \frac{\sum _{i=1}^C{n}_{i1}{n}_{i2}}{N^2}}$$

(5)

where $n_{i1}$ is the proportion of instances that Rater 1 assigns to category $i$ and $n_{i2}$ is the proportion of instances that Rater 2 assigns to category $i$.

The values of Cohen’s kappa (κ) range from 0 to 1. κ = 1 implies perfect agreement, whereas κ = 0 means that the agreement is purely by chance.

All the machine learning models and evaluation criteria are developed via RapidMiner, a well-known machine learning tool. The testing is conducted using the 10-fold cross-validation method, ensuring a robust and reliable evaluation, where the dataset is split into 10 subsets. In each iteration, 80% of the data is used for training, while the remaining 20% is used for testing. This approach minimizes bias, enhances generalization, and provides a comprehensive assessment of the model’s performance across different data partitions.

This section presents the findings, accompanied by a comparison of the proposed model against related work. The proposed model is assessed through three phases. The first phase, A, is Machine Learning-Based Venue Classification. It aims to categorize location types by analyzing spatial, temporal, and contextual attributes, and hence assesses the improvement in location-based recommendations and user behavior modeling. The second phase, B, is building a sequential trajectory pattern. It assesses the improvement in classification accuracy through the incorporating of sequential movement patterns. The final phase, C, focuses on handling the prediction of hidden and missing locations in a user’s trajectory. The assessment evaluates the leveraging of both venue classification and sequential trajectory patterns after such handling.

6.1. Machine Learning-Based Venue Classification

The initial step (phase A) entails predicting the location after applying the preprocessing process and crawling the Foursquare to extract the location semantic attributes. Subsequently, three different classifiers were applied.

Figure 7 displays the accuracy performance measure after the preprocessing stage is completed and the semantic attributes for each location’s category are added.

The figure demonstrates the high accuracy of deep learning for classification of the LBSN data into predefined categories. The KNN and GBT were among the other models that did exceptionally well in classifying the LBSN data.

6.2. Enhanced ML by Integrating Sequential Trajectory Pattern for Check-Ins

Phase B of our approach involves constructing trajectories, represented as a collection of nodes, each signifying a place at a particular moment in time.

In particular, users with exceptionally low check-in records are separated out. For each individual check-in, the trajectories are shown as a series of POI categories; on average, each trajectory has 6.65 POI categories.

The segmentation of the check-ins is determined by temporal sequence for weekdays and weekends during a week. Figure 8 demonstrates improved accuracy values for the various classifiers, which indicates 8% accuracy enhancement for KNN, 6% enhancement for GBT, and 2% enhancement for deep learning.

The confusion matrix for each constructed model is presented to illustrate their performance on the testing data.

Table 4. Confusion matrix for KNN in phase B.

	True	Food	Arts & Entertainment	Nightlife Spot	Travel & Transport	Outdoors & Recreation	Shop & Service	Professional & Other Places	Event	Residence	College & University	Class Precision
Predicted
Food		9006	172	509	303	330	677	233	312	121	45	76.92%
Arts & Entertainment		96	730	38	32	39	45	32	49	8	3	68.10%
Nightlife Spot		354	71	2376	85	69	101	44	87	34	10	73.54%
Travel & Transport		286	46	75	5245	129	151	83	166	72	8	83.77%
Outdoors & Recreation		238	64	79	140	3302	136	65	133	52	15	78.17%
Shop & Service		553	54	115	132	136	3902	81	152	58	22	74.97%
Professional &Other Places		255	48	52	79	66	94	3375	67	19	11	83.01%
Event		328	63	82	109	107	120	99	2637	34	21	73.25%
Residence		144	20	36	84	78	82	47	57	4475	7	88.97%
College & University		45	11	8	12	21	13	9	28	11	930	85.48%
class recall		79.66%	57.08%	70.50%	84.31%	77.20%	73.33%	82.96%	71.50%	91.63%	86.75%

,

Table 5. Confusion matrix for deep learning in phase B.

	True	Food	Arts & Entertainment	Nightlife Spot	Travel & Transport	Outdoors & Recreation	Shop & Service	Professional & Other Places	Event	Residence	College & University	Class Precision
Predicted
Food		9425	130	409	248	243	533	177	250	83	31	81.75%
Arts & Entertainment		98	808	44	24	47	39	33	45	6	6	70.26%
Nightlife Spot		337	63	2550	65	57	87	45	76	22	12	76.95%
Travel & Transport		211	46	61	5452	91	112	62	119	40	9	87.89%
Outdoors & Recreation		207	54	60	109	3502	108	51	132	52	7	81.78%
Shop & Service		486	48	110	102	117	4183	62	140	43	14	78.85%
Professional & Other Places		161	43	42	62	47	76	3519	54	15	6	87.43%
Event		265	60	74	93	100	119	77	2823	34	21	77.00%
Residence		80	16	16	52	56	53	34	38	4582	3	92.94%
College & University		35	11	4	14	17	11	8	11	7	963	89.08%
class recall		83.37%	63.17%	75.67%	87.64%	81.88%	78.61%	86.50%	76.55%	93.82%	89.83%

and

Table 6. Confusion matrix for gradient boosted tree in phase B.

	True	Food	Arts & Entertainment	Nightlife Spot	Travel & Transport	Outdoors & Recreation	Shop & Service	Professional & Other Places	Event	Residence	College & University	Class Precision
Predicted
Food		10,313	211	554	514	437	981	504	752	206	132	70.62%
Arts & Entertainment		14	914	7	3	2	6	7	9	1	0	94.91%
Nightlife Spot		93	11	2694	23	18	15	8	19	12	1	93.09%
Travel & Transport		187	39	24	5451	83	126	72	98	44	7	88.91%
Outdoors & Recreation		110	26	12	56	3559	50	28	51	25	11	90.61%
Shop & Service		283	25	27	65	59	3954	75	104	27	14	85.36%
Professional & Other Places		172	29	35	58	46	99	3299	70	15	5	86.18%
Event		55	11	5	15	33	23	33	2538	7	6	93.10%
Residence		64	5	11	34	31	62	29	28	4540	5	94.41%
College & University		14	8	1	2	9	1	13	19	7	891	92.33%
class recall		91.23%	71.46%	79.94%	87.62%	83.21%	74.38%	81.10%	68.82%	92.96%	83.12%

show the confusion matrix for KNN, deep learning, and gradient boosted tree, respectively.

The confusion matrix helps analyze the classification performance of a model by showing actual vs. predicted values. In Table 4, the model correctly classified 9006 Food, 730 Arts & Entertainment, 2376 Nightlife Spot, 5245 Travel & Transport, 3302 Outdoors & Recreation, 3902 Shop & Service, 3375 Professional & Other Places, 2637 Event, 4475 Residence, and 930 College & University venues. This indicates that the model is operating effectively, especially for Food (9006 correct classifications). In Table 5 and Table 6, the model works effectively in the food area as well. This can be explained by the fact that, according to the statistics in Figure 3, food has the highest check-in count.

6.3. Enhanced Inference by AHLTP Model

The final phase C focuses on predicting hidden or missing locations in a user’s trajectory. By leveraging both venue classifications and sequential trajectory patterns, the model infers potentially unrecorded check-ins and predicts future locations, enhancing the accuracy of mobility forecasting in LBSN applications. During this phase, we employed the FP-Growth algorithm to identify frequent location sets and generate association rules. Improved accuracy values for the different classifiers are shown in Figure 9. Results in this phase have been significantly improved compared to phase B. KNN recorded accuracy of 98% with 18% enhancement compared to experiments in phase B. Deep learning had accuracy of 92% with 6% enhancement, while GBT recorded 96% accuracy with 12% enhancement compared to experiments in phase B.

To examine the quality of the inference of association rules, the comparison of each categorical variable is plotted through the computation of the Kappa measure. Figure 10 shows the significance of each category variable. The two primary kappa measures—for the Food and College & University categories—have the highest value with a 0.97 kappa measure. Subsequently, they are employed in the analysis of the lift threshold. Figure 11 and Figure 12 illustrate the kappa measure for Food and College & University, respectively. The threshold is standardized to fall within the interval [1,2,3,4,5,6,7,8,9,10]. The KNN has the highest and most stable kappa results compared with the other classifiers, followed by GBT. GBT achieves high accuracy and strong generalization, making it ideal for predicting user mobility patterns and personalizing location recommendations.

The dataset in this study consists of structured check-in data with clear categorical labels. Since KNN is a straightforward non-parametric technique, it directly leverages the spatial proximity of locations without requiring feature transformation.

It effectively manages the local structures present in geolocation data, where neighboring points, like locations or check-ins, frequently exhibit similar characteristics or behaviors. In other words, the check-in dataset exhibits clustered patterns where users frequently visit specific locations. KNN excels in such cases by assigning labels based on nearest neighbors (by using distance-based metrics) rather than requiring deep hierarchical feature extraction. Moreover, the AHLTP model employs association rule mining to infer hidden locations. KNN benefits from these pre-extracted relationships by making direct nearest-neighbor comparisons.

On the other hand, deep learning models often need high-dimensional features to develop hierarchical representations, which may not be fully leveraged with geolocation data, particularly if the data size is not large or lacks complex, high-dimensional relationships. Additionally, deep learning might struggle with sparse data where certain check-ins are underrepresented, leading to potential overfitting or suboptimal generalization.

To reinforce performance claims, we conducted a statistical significance test using a pairwise t-test to compare the performance of three classifiers. The pairwise t-test helps to assess whether the observed differences in classifier performance are statistically significant or not.

Table 7. Pairwise t-test performance for the three classifiers.

Classifiers	p-Value
KNN vs. DL	0.977
KNN vs. GBT	0.758
DL vs. GBT	0.959

shows the p-values for the three classifiers KNN, deep learning, and gradient boosted-tree integrated with the proposed AHLTP model, measured at a confidence level of 5%. In our analysis, performance is measured by accuracy. The highest p-value of 0.977 obtained for the KNN supports the results obtained, indicating KNN outperformed the other two classifiers.

Figure 13 presents a comparison of the accuracy of previous models and the proposed AHLTP method for trajectory prediction. Figure 13 compares models from related studies [25,26,27,29]. In [25], the authors provided a model named Venue2Vec, which incorporates temporal-spatial context, semantic information for textual analysis, and sequential relationships. The authors in [26] provided a methodology for linking user identities by comparing texts and geolocations across two social media platforms, using information inputs that are asymmetric, while the authors in [29] examined users’ check-in patterns using certain variables related to time periodicity and personal preference, which were then integrated into a classification model and a supervised scoring model. To solve the problem of cold-start users, the authors in [27] offered a model called SECTS that identifies POIs with a high probability for users with the same source and target domains.

As the predictions are made without taking into account the associations between checked-in locations, the proposed AHLTP model concentrates on this aspect. The result is that it outperforms the other models in Figure 13 with an accuracy of 98% for the KNN classifier. This comparison shows how successful the suggested method is in enhancing trajectory prediction accuracy.

7. Conclusions and Future Work

In this paper, an associative hidden location trajectory prediction (AHLTP) model is proposed to predict related locations to solve the sparsity problem in check-in data (hidden locations). The proposed AHLTP model explicitly utilizes spatiotemporal contexts to identify historical hidden states with significant predictive value. AHLTP illustrates the effectiveness of trajectory mining in location-based social networks (LBSN) for revealing hidden locations through association rule mining algorithms, including FP-Growth, while also considering implicit features by leveraging the semantic characteristics of the location. Based on the input data, the KNN had a high accuracy measure of 98%, which indicates that it can correctly predict a user’s location category. Additionally, the gradient-boosted tree model’s excellent accuracy of 96% shows that it can produce accurate predictions. Deep learning showed strong results in location prediction but requires large datasets to achieve high classification accuracy. With accuracy values of up to 98%, the experimental evaluation was conducted on a real-word dataset from Foursquare, demonstrating the efficacy of the suggested strategy and outperforming some of the most representative approaches in the field.

Future work will incorporate neighboring user movement data for better location prediction, in addition to expanding some real-world applications and datasets for smart city development and business intelligence.