A Knowledge Graph-Enhanced Hidden Markov Model for Personalized Travel Routing: Integrating Spatial and Semantic Data in Urban Environments

Highlights:

What are the main findings?

• We introduced a novel knowledge graph-based Hidden Markov Model (KHMM) that significantly enhances personalized route recommendations by effectively integrating spatial and semantic data from POIs.
• By leveraging knowledge graphs, the KHMM expands the traditional Hidden Markov Model’s state space to capture multi-dimensional and higher-order relationships between POIs.

What are the implications of the main findings?

• Improving adaptability to fluctuating user preferences and real-time travel conditions is crucial for developing intelligent transportation systems that can respond to the evolving needs of urban travelers, thereby fostering smarter cities.
• Providing insights into factors influencing travel behavior deepens the understanding of how spatial and semantic relationships shape route selection, having far-reaching implications for transportation planners and urban policymakers implementing tailored solutions to enhance urban mobility and efficiency.

Abstract

Personalized urban services are becoming increasingly significant in smart city systems. This shift from intelligent transportation to smart cities broadens the scope of personalized services, encompassing not just travel but a wide range of urban activities and needs. This study proposes a knowledge graph-based Hidden Markov Model (KHMM) to improve personalized route recommendations by incorporating both spatial and semantic relationships between Points of Interest (POIs) in a unified decision-making framework. The KHMM expands the state space of the traditional Hidden Markov Model using a knowledge graph, enabling the integration of multi-dimensional POI information and higher-order relationships. This approach reflects the spatial complexity of urban environments while addressing user-specific preferences. The model’s empirical evaluation, focused on Changsha, China, examined how temporal variations in public attention to POIs influence route selection. The results show that incorporating dynamic temporal and spatial data significantly enhances the model’s adaptability to changing user behaviors, supporting real-time, personalized route recommendations. By bridging individual preferences and road network structures, this research provides key insights into the factors shaping travel behavior and contributes to the development of adaptive and responsive urban transportation systems. These findings highlight the potential of the KHMM to advance intelligent travel services, offering improved spatial accuracy and personalized route planning.

1. Introduction

Route recommendation has emerged as a pivotal research area in smart city systems, with increasing attention directed toward developing intelligent, personalized solutions [1,2]. This research area finds critical applications in next POI recommendation techniques [3,4,5,6,7], predicted destination identification [8,9,10], travel planning [11,12], and personalized route recommendations [13,14,15], providing both convenience and efficiency in daily life. With the ongoing advancements in intelligent technologies and big data analytics, the scope and utility of route recommendation systems are poised to expand significantly.

Developing travel routes that balance universal applicability with personalized customization requires leveraging advanced information technologies, real-time data, and intelligent analytical frameworks. The primary challenge lies in reconciling the heterogeneous preferences of individuals with the general requirements of a diverse user base. Achieving this balance necessitates sophisticated personalization methods capable of catering to diverse user preferences and contexts. Moreover, the increasing complexity of user needs has elevated the significance of understanding subjective emotions and preferences in personalized route recommendation tasks [16,17,18]. Research has demonstrated that extracting the complex, abstract patterns of user behavior and preferences from extensive datasets enables a deeper understanding of user interests, thereby enhancing the connection between users and POIs [19,20,21]. This connection underpins the development of personalized routes tailored to individual users [22]. Incorporating the emotional attributes of POIs alongside geographic properties provides additional insights into user sentiments, further improving the quality and relevance of personalized recommendations [23].

With the diversification of travel purposes and the expansion of road networks, traditional route planning algorithms, such as Dijkstra, A*, and Floyd–Warshall, face significant limitations in their computational efficiency and adaptability to dynamic contexts. These algorithms rely heavily on static data and predefined heuristics, rendering them inadequate for capturing the dynamic interplay of user preferences, trip purposes, and real-time road conditions. Moreover, their lack of a user-centric experiential design often results in less interpretable recommendations. Consequently, researchers have shifted toward more comprehensive approaches that prioritize the holistic travel experiences of users. Numerous methods have been proposed for generating personalized route recommendations [24,25,26]. Among these, recurrent neural networks (RNNs) have exhibited strong performance in location prediction tasks, with studies focusing on enhancing or hybridizing RNN models to improve POI prediction [27,28,29]. For instance, [30] proposed a personalized recommendation model leveraging semantic features extracted from tourists’ trajectories and travel contexts. Similarly, Hidden Markov Models (HMMs) have proven effective for modeling check-in sequences and addressing new POI recommendation challenges [31]. Deep learning techniques, with their capacity to integrate and analyze multi-faceted features, have further bolstered travel recommendation systems [32,33,34]. Additionally, multi-objective optimization strategies, including hybrid tabu search algorithms [35], greedy genetic algorithms [36], and Pareto-based approaches [37], have been actively explored to meet diverse user demands.

Despite these advances, several critical gaps remain in personalized route recommendation research. Existing studies predominantly emphasize the correlation between neighboring POIs, often neglecting longer temporal dependencies and spatiotemporal correlations when providing continuous POI recommendations. Additionally, these models frequently overlook the influence of temporal and environmental interactions on travel behavior. Integrating static and dynamic data more effectively is crucial for improving the quality and adaptability of travel recommendations. Knowledge representation and organization play a pivotal role in addressing these challenges. Knowledge graphs (KGs) emerge as a promising semantic network framework, capable of integrating and representing vast amounts of structured knowledge, including geographic information, POI features, user preferences, and other related attributes. By incorporating dynamic updates and rich semantic associations, KGs can overcome the limitations of traditional static data-driven algorithms, thereby enhancing the adaptability and real-time applicability of route planning methods.

Moreover, KGs facilitate the integration of diverse data sources, enabling the identification of latent correlations among POIs and fostering stronger knowledge relevance and information expansion. These capabilities allow for the more comprehensive modeling of user preferences and travel needs, thereby providing diverse and tailored recommendation options. As auxiliary information, KGs address challenges such as cold starts and data sparsity, significantly improving the performance and interpretability of recommendation systems [38]. Recent applications of KGs in recommendation systems underscore their potential to enrich the understanding of relationships among entities, such as geographic locations, POI attributes, user profiles, and traffic networks. This integration aids in analyzing urban travel problems, including POI recommendations [39,40] and personalized travel planning [41]. Additionally, KGs effectively capture user preferences by analyzing interactions across multiple types of entities and relationships, thereby enhancing user modeling for POI recommendation tasks [42]. For example, [43] utilized KGs to construct a preference-based knowledge graph, associating users with POI attributes and deriving similarity matrices for POI recommendations. Similarly, attention mechanisms have been integrated with KGs to improve sequential POI recommendations [44,45]. Recent studies have also leveraged KGs to optimize urban transportation networks [46,47] and enhance the performance of tourism recommendation systems [48,49].

However, existing studies largely focus on the relationship between POIs and users, often overlooking the multi-dimensional attributes of POIs, including both spatial and non-spatial factors, that are critical for addressing path-finding challenges in route recommendation tasks. These deficiencies may reduce the adaptability of recommendation algorithms to diverse scenarios in practical applications and affect the accuracy and effectiveness of personalized route recommendation. To address these limitations, we propose a novel KHMM framework that integrates both spatial and semantic relationships between POIs into a unified decision-making model for personalized route recommendation.

The proposed KHMM framework expands the state space of traditional Hidden Markov Models by incorporating knowledge graphs, enabling the integration of multi-dimensional POI information and higher-order relationships. Specifically, the framework leverages the associations among multi-dimensional POI attributes to dynamically map crowdsourced spatiotemporal data onto real-world road networks, thereby recommending optimal routes tailored to user preferences. The key contributions of this study are as follows:

• We propose the KHMM, a personalized route recommendation method that combines knowledge graphs with an enhanced Hidden Markov Model to incorporate spatial and semantic relationships between POIs into a unified decision-making framework.
• We constructed a POI-KG that captures the multi-dimensional attribute features of POIs, enhancing entity information, improving entity connectivity, and addressing the path-finding challenge through an enhanced correlation between spatial and non-spatial attributes.
• We introduce a POI popularity index that accounts for public attention volumes across different time periods. This index integrates bidirectional weights within the KG, combining static and dynamic data while considering the impact of environmental interactions on travel patterns. This integration yields more accurate and context-aware route recommendations.

The rest of this paper is organized as follows: Section 2 introduces the KHMM, presenting the model’s architecture and working principles in detail. Section 3 reports experimental results for a real-world road network. Finally, Section 4 provides a discussion and conclusions.

2. Methodology

This study proposes a KHMM for generating practical and personalized travel route recommendations. As illustrated in Figure 1, the KHMM framework integrates multivariate data, including POIs, public recommendation data, road network data, and POI attribute data, to establish a crowdsourced geospatial information base. Geospatial associations are reconstructed using topological and spatial connections, while POIs are categorized into distinct groups based on their attributes to construct a travel KG. The KG calculates weights based on POI attributes to establish attribute-level associations among POIs. Finally, a two-layer HMM is employed to identify candidate points within the actual road network, enabling the generation of dynamic and personalized travel routes. The proposed method is detailed in the following subsections.

2.1. Problem Statement

We first introduce several notations related to personalized route recommendations. Considering a set of subjects, S, a person, u, is interested in a small subset of S, denoted as $S_u=\left\{s_u^1, s_u^2, \dots , s_u^{n_u}\right\}\subset S\setminus \left\{u\right\}$, where $n_u$ represents the number of subjects of interest to u. For each subject, $s_u^i\in S_u$, there exists a set of $m_i$ POIs, denoted as $P_{s_u^i}=\left\{p_{s_u^i}^1, p_{s_u^i}^2, \dots , p_{s_u^i}^{m_i}\right\}\subset P\setminus \left\{s_u^i\right\}$, where P is the universal set of POIs. As previously mentioned, the public attention to each element in $P_{s_u^i}$ may vary significantly. Among these, $k^i$ elements in $P_{s_u^i}$ receive major public recommendations and are termed core POIs, denoted as ${\widehat{P}}_{s_u^i}=\left\{{\widehat{p}}_{s_u^i}^1, {\widehat{p}}_{s_u^i}^2, \dots , {\widehat{p}}_{s_u^i}^{k^i}\right\}\subset P_{s_u^i}$. Personalized route recommendations for u should prioritize traversing as many core POIs as possible, ordered by interest, i.e., $R_u\setminus \left\{S_u\right\}$. Inspired by previous studies on route recommendations [50], practical routes are also constrained by the road network structures connecting core POIs across various values of ${\widehat{P}}_{s_u}$, namely ${\widehat{P}}_{s_u}$’s road constraints. Therefore, we mathematically define the recommended routes as

$$R_u\triangleq \left\{{\widehat{p}}_{s_u^1}, {\widehat{p}}_{s_u^2}, \dots , {\widehat{p}}_{s_u^{n_u}}|{\widehat{p}}_{s_u^i}\in P_{s_u^i}\right\}$$

(1)

The goal of the personalized route recommendation problem is to maximize user satisfaction. To achieve this, we aimed to maximize the overall public recommendation score of POIs while minimizing the route distance, following the user’s preferred subject order. In this context, public recommendations act as hard negative samples that contrast with personal interests, adding complexity to the task. Our model learns to extract the public’s cognitive knowledge of POIs, predicting which POIs in $S_u$ are likely to be valued without requiring any information about future target POIs from the user. Consequently, although the ground truth of core POIs is constructed using historical public comments, our model remains effective in satisfying personal interests.

2.2. Definition of Foundations in Route Planning

Definition 1 (POI Subject).

Each POI is associated with a set of characteristics, referred to as feature words. These characteristics represent POI tag information, which describes the textual data associated with the POI. POIs sharing identical features and characteristics are grouped into the same category. For example, POIs with the attribute label “higher education institution” are categorized as a “University.” Consequently, POIs with this type of attribute label are classified under the “University” category. Let S denote the set of subjects. A subset of subjects for a user, u, is denoted as $S_u=\left\{s_u^1, s_u^2, \dots , s_u^{n_u}\right\}\subset S\setminus \left\{u\right\}$.

Definition 2 (Subject Access Sequence).

A subject access sequence is a set of POI subjects representing the sequence of POI subjects accessed by a user based on their personal preferences. Formally, it is defined as a set of user-selected subject sequences, $S_u$ = $\left\{s_u^1, s_u^2, \dots , s_u^i\right\}$. Once a user completes the subject selection process, the access sequence is considered to have been determined. The relationship between the subjects and their corresponding POIs is defined as $s_u^i\to \left\{p_{s_u^i}^1, p_{s_u^i}^2, \dots , p_{s_u^i}^{m_i}\right\}$.

Definition 3 (Projection Point).

POIs are typically not located directly on roads. During route calculation, each POI is projected onto the nearest road segment [51]. The projected position of a POI, $p_{s_u^i}^{m_i}$, in an adjacent road section is referred to as a projection point, denoted as $x_{m_i,j}$. Since a POI may have multiple adjacent road sections, it can generate multiple projection points. The projection point $proj(p_{s_u^i}^{m_i},r_j)$ is calculated by determining the shortest distance from the POI to the road segment. The relationship between a POI and its projection points is defined as $p_{s_u^i}^{m_i}=(x_{1,j}, x_{2,j}, \dots , x_{m_i,j})$. A distance threshold is established based on the spatial relationship between the POI and the road segment, and projection points are calculated only for adjacent road sections within this threshold. Figure 2 illustrates the projection of two POIs onto a road network. For instance, $p_{s_u^i}^1$ has two projection points in adjacent road sections, while $p_{s_u^i}^2$ has three projection points.

2.3. Construction of Knowledge Graph

Integrating KGs with POIs enables the discovery of rich semantic relationships among POIs, facilitating the exploration of their connections. Consequently, KGs can significantly enhance the quality of personalized recommendations. POIs inherently possess extensive spatiotemporal information and exhibit strong spatial relevance. Analyzing the correlations between POIs using their spatiotemporal relationships is crucial for generating effective recommendations. This subsection elaborates on the construction of the POI-KG. A knowledge network is established by building a POI knowledge base and investigating the correlations between POIs. The degree of association between POIs is more comprehensively represented as the amount of POI tag information in the KG increases, enabling a deeper understanding of the underlying information. The correlation analysis evaluates the proximity between POIs and their attributes, leading to the construction of a weighted, directed KG. Finally, the A* algorithm is employed to identify the shortest route within the KG.

2.3.1. The Fundamentals of KGs

We can construct a POI knowledge base using POI entities, their attributes, and the relationships between them, such as their category, address, and time. As semantic networks, KGs exhibit significant expressive power and modeling flexibility. They can effectively model real-world entities, concepts, attributes, and relationships. A KG is a comprehensive knowledge network composed of nodes and edges. Nodes represent concepts or entities, while edges indicate semantic relationships between entities and concepts. The fundamental structure of a KG is represented as $G=(E, R, A)$, which encodes the triples of ‘entity-relationship-entity’ and ‘entity-attribute-attribute value’ [52]. E denotes entities, the basic elements in the KG, and primarily refers to subjects and POIs. A represents attributes, which describe the properties, characteristics, and features of POIs, such as addresses and categories. R denotes relationships, which connect two entities or entities and their properties, capturing the associations between them.

The entity and attribute knowledge in a POI-KG is extracted from diverse sources and structures, primarily sourced from encyclopedic web pages, such as Baidu Encyclopedia and Wikipedia, as well as professional travel websites. POI entities and attributes were identified by analyzing related studies.

Table 1. POI properties of websites.

POI Properties
Name	Category	Address	Recommended Duration of Visit
Ticket	Characteristic	Opening Hours	Collection of Cultural Relics
Relevant People	Famous Landmarks	Attraction Level	Suitable Season for Visiting
Relevant POIs	Surrounding POIs	Construction Time	Surrounding Transportation Facilities

provides detailed information about POI attributes, which include the unique POI identifier (PID), name, subject, address, ticket price, construction time, surrounding POIs, opening hours, and other relevant attributes. Additionally, the attribute data corresponding to different POIs and subjects may vary. This study selected Changsha City in Hunan Province as a case study to construct the POI attribute database.

2.3.2. Building POI-KG

Figure 3 illustrates the flowchart of POI-KG construction. The process begins with knowledge extraction from the captured data sources, which includes entity extraction, relationship extraction, and attribute extraction. The second step involves defining the POI ontology, a semantic data model that specifies the types of entities within a domain and describes their properties. The POI ontology serves as a framework for incorporating information about individual POIs, enabling the creation of instances for each ontology relationship. Subsequently, the KG employs a top-down approach to identify POI entities, achieving entity definition. Attributes are categorized into primary and secondary attributes based on their association with POIs. Primary attributes include categories that reflect the relevance of entities to each other, such as the subject, construction time, and surrounding POIs. Secondary attributes describe the unique features of POIs, such as their address and name. The attribute labels are further refined based on the unique characteristics of each POI to highlight distinctions among various POI types. Next, the attributes and relationships between entities and attributes are established. Finally, as knowledge accumulates, these entities and their interconnections are integrated into a comprehensive POI-KG.

Figure 4a depicts a unidirectional POI-KG, illustrating the semantic relationships between POIs and their attributes. The attributes of a POI include its name, subject, category, address, attraction level, construction time, opening hours, characteristics, ticket price, surrounding POIs, relevant POIs, surrounding transportation facilities, the recommended duration of a visit, suitable seasons for visiting, and contact information. There are three sizes of nodes: the largest nodes represent subjects, the medium-sized nodes represent POIs, and the smallest nodes represent attribute entities. Additionally, different colors in the diagram denote various entities and attributes, while edges represent semantic relationships between entities and attributes. Entities can be linked through shared attributes, establishing associations between different entities. A POI-KG strengthens the connections between POIs, weaving the knowledge web together more tightly and enabling knowledge path queries between entities.

2.3.3. Calculation of Edge Weight

The primary objective of constructing a KG is to analyze the associations between POIs to generate personalized recommendations. To achieve this, we introduced POI recommendation degrees and valid data derived from survey questionnaires to determine the edge weights in the POI-KG. These surveys were designed to capture individuals’ preferences for POIs and their selection criteria based on various attributes. The collected indices were then utilized to construct a bidirectional edge weight function, which quantified the weight of each edge. These edge weights reflected the strength of the interrelationships between the nodes connected by each edge.

When constructing a KG, certain POI attributes are exclusive and closely tied to specific POIs, while others may connect with surrounding nodes through shared attributes or relationships. To deepen these connections, we assigned weights to the edges, quantifying the similarity and strength of associations between POIs. This weight design enhanced the KG’s semantic richness and reasoning capabilities. In Figure 4b, the weight relationships between a POI and its attributes are depicted, with larger values indicating stronger correlations. For instance, the probability that the scenic area connects to Orange Island is 1.9810, signifying that Orange Island has a popularity index of 1.9810. Since Orange Island is classified as a scenic spot, the probability of it connecting to the scenic spot category is 1. Other relationships involved attributes such as the name, subject category, attraction level, construction time, surrounding POIs, and related POIs.

A questionnaire was introduced to determine the weights of POIs and attribute nodes. The questionnaire included standard socio-demographic variables such as gender, age, income, and work or study status [53]. Respondents from diverse groups were asked about the attributes of attractions they considered when planning a tour. Based on the questionnaire results, the average score for each attribute was calculated and normalized to assign weights to each attribute. The following analysis examines the impact of various attributes on POI selection:

• Popularity index of a POI: Users exhibited varying levels of interest and search activity for different POIs, leading to the generation of a POI popularity index. In this study, we utilized the 360 Trends Index [54] to calculate the popularity index of POIs. To ensure comparability across datasets, each theme was treated as a separate dataset, and the data were standardized using z-scores, enabling cross-dataset comparison and analysis. Additionally, considering the spatial autocorrelation of the area and the spatial correlations among POIs within a region, we employed the Kriging interpolation method to estimate values for POIs lacking index data. Finally, a constant related to the data range was added to all values to ensure they remained positive. The formula for calculating the POI popularity index is defined as follows:

  $$R\left(P_{s_u^i}\right)={\displaystyle \frac{H{\widehat{p}}_{s_u^i}^{k^i}-\mu }{\sigma }}+f$$

  (2)

  where $H{\widehat{p}}_{s_u^i}^{k^i}$ represents the value of the 360 Trends Index. $R\left(P_{s_u^i}\right)$ denotes the transformed value. $\mu$ is the mean score of the entire sample space. $\sigma$ is the standard deviation of the entire sample space. f is a constant related to the data range, ensuring all values remain positive.
• The rank of a POI. The rank of a POI typically represents its attraction level, signifying the overall quality of the location. Users often consider this rank when visiting a site. We calculated the weight of POI rankings based on the analysis of data from the collected questionnaires. The formula for calculating the rank of a POI is defined as follows:

  $$R_{rank}={\displaystyle \frac{n_{rank}}{N}}$$

  (3)

  where $n_{rank}$ represents the number of people who selected the indicator. N represents the total number of participants. According to the questionnaire survey data, the value of $R_{rank}$ was 0.27.
• Construction time. Each POI has a corresponding time of construction. Different POIs may share the same construction time, establishing a correlation between them. Based on the analysis of data from the questionnaire, the weight of the construction time $R_{con}$ was calculated to be 0.24.
• Opening hours. POIs and scenic spots have specific operating hours. Some are open all day, while others operate within fixed periods. For instance, museums generally close on Mondays and have established daily opening and closing times. POIs with different themes may share the same opening hours, suggesting a correlation between them. Based on the results of the questionnaire survey, the weight of the opening hours $R_{open}$ in the POI KG was 0.18.
• Surrounding POIs. Users often want to learn about other nearby POIs when visiting a particular one. Therefore, POIs that feature a wide range of nearby attractions are more likely to capture users’ interest. In the POI-KG, the correlation between POIs was strengthened by introducing the consideration of attribute relationships among nearby POIs. This attribute was calculated by considering a POI as the center of a circle, setting a reasonable radius as the threshold for buffer analysis, and taking into account the spatial location relationship of the POI. The edge weights of surrounding POIs were calculated using the POI popularity index as the weight value.
• Related POIs. When individuals search for a POI online, related POIs will appear. These POIs may be thematically relevant or geographically related to the target location, reflecting specific needs or interests at a given geographical point. The edge weights of these relevant POIs were calculated using the POI popularity index as the weight value.

2.3.4. Knowledge Graph Path Planning Based on A*

The KG constructed represents a weighted directed graph, which can be converted into an adjacency matrix and stored as arrays containing node and edge information. The A* (A Star) algorithm, renowned for its effectiveness in path-finding and graph traversal, was employed due to its notable performance and accuracy. Especially in directed graphs with weighted edges, A* can excel at efficiently computing the shortest path among multiple nodes. Consequently, we utilized the A* algorithm to search for the shortest path within the KG of POIs.

Based on the definition of entity–attribute relationships, the relationships were considered edges, and the above formula was used to calculate the weight coefficients of each edge. A weighted directed KG was constructed, and the shortest distance between nodes could be found according to the weight coefficients of the edges. We used a weighted adjacency matrix to represent the topology. The adjacency matrix was a two-dimensional array. It was based on using nodes and attributes in the KG to form the set vertex (n), and the relationship between them formed the edge (r). If there was a relationship between entities, the weights were set for the edges. If there was no edge, the weight was infinite. Where $G_r=(V, E)$ is a directed graph, $V=1, 2, \dots , N$ with $N>=1$, and let $c:E\left(G_r\right)\to R$ be an arbitrary weight function on the edges of $G_r$. $G_r$ is stored as an adjacency matrix with n vertices; therefore, the adjacency matrix A is an $n\ast n$ square matrix defined as

$$arc\left[i\right]\left[j\right]=\left\{\begin{array}{c}{W}_{ij}, (v_i, v_j)\in E\hfill \\ 0, i=j\hfill \\ \infty \hfill \end{array}\right.$$

(4)

The user-selected subject access sequence is taken as the input target, and the sequence is followed to find the shortest path sequentially. The algorithm will first confirm an intermediate node for given target nodes. It then judges whether the shortest distance between the two target nodes is less than the sum of the distances from the first target node to the intermediate node and then to the second target node. If the condition is met, the intermediate node is on the shortest path and the route is updated. Finally, the calculation of the shortest path is complete. The final output path is represented as $subject-POI-attribute-\dots -POI-subject$, which contains the POIs and attributes. The POIs, which are candidate points, will be used as hidden states in the two-layer HMM described in Section 2.4. The optimal route is calculated by combining these results with data on the road network.

2.4. Two-Layer Hidden Markov Model

A two-layer HMM is the core framework of the KHMM, with its state space comprising the observation state, hidden state, observation probability, and transition probability. It involves a dual stochastic process, effectively modeling random jump systems [55]. Consequently, we utilized an improved two-layer HMM method, combined with a KG, to model the route planning process. Figure 5 illustrates the KHMM. This model consists of three state space layers, differing from the traditional model through the incorporation of a KG between the observed and hidden states. Initially, the observed state space consists of user-determined access sequences, $O_u^i$, represented by the class of POIs. The observed states find the shortest path in the KG, where the output POI becomes the candidate point. The hidden state space is divided into two layers: the first layer includes POIs outputted from the KG, and the second layer comprises projection points. Transition probabilities between the hidden states are calculated in the second layer, while emission probabilities are determined between the observed and hidden states.

2.4.1. Construction of KHMM

• Observations. The observation state space comprises several POI subjects. Users choose a sequence of POI subjects to visit based on their preferences. This sequence is a set of observation sequences ($O_u^1, O_u^2, O_u^3, \dots , O_u^i$) where each observable state, $O_u^i$, represents a subject.
• Hidden States. The hidden state space of this model is a structure with two layers. The first layer comprises the POI nodes output from the KG, forming a POI set. These POIs are treated as candidate points, denoted as $(p_{s_u^i}^1, p_{s_u^i}^2, \dots , p_{s_u^i}^{m_i})$. The second layer is made up of the projection points $x_{m_i,j}$ of the candidate points, denoted as $(x_{1,j}, x_{2,j}, \dots , x_{m_i,j})$.
• Emission Probabilities. The emission probability is generated between the hidden and observed states, representing the probability distribution in this system state. In a real road network, each POI, $p_{s_u^i}^{m_i}$, generates an emission probability, $p\left(p_{s_u^i}^{m_i}\right|x_{m_i,j})$, in the adjacent road segment $r_j$. The projection point is expressed as $x_{m_i,j}$. The distance between the POI and the projection is calculated using the great circle distance, denoted as $\left|\right|p_{s_u^i}^{m_i}-x_{m_i,j}{\left|\right|}_{gc}$. The emission probability is defined by the following formula [50]:

  $$p\left(p_{s_u^i}^{m_i}\right|x_{m_i,j})=p\left(p_{s_u^i}^{m_i}\right){\displaystyle \frac{1}{\sqrt{2\pi }{\delta }_c}}{e}^{-0.5{\left({\displaystyle \frac{{∥p_{s_u^i}^{m_i}-x_{m_i,j}∥}_{gc}}{{\delta }_c}}\right)}^2}$$

  (5)
• Transition Probabilities. The transition probability matrix indicates the chances of a state changing from one to another. It is represented as the probability of projected points in a hidden state undergoing a state transfer at different times. Specifically, the projection point of a POI, $p_{s_u^i}^{m_i}$, on a road segment is indicated as $x_{m_i,j}$. The projection point of the next POI $p_{s_u^{i+1}}^{m_{i+1}}$ is denoted by $x_{m_{i+1},j}$. The “route distance” is defined as $\left|\right|x_{m_i,j}-x_{m_{i+1},j}{\left|\right|}_{route}$, which is the shortest path between the two projection positions. The distance between two candidate points is determined using the great circle distance, indicated as $\left|\right|p_{s_u^i}^{m_i}-p_{s_u^{i+1}}^{m_{i+1}}{\left|\right|}_{gc}$. The route distance between the projected points is compared to the great circle distance between POIs. In our experiments, a shorter road path distance between a pair of matched projected points closely aligned with the distance between the POIs [56]. The trend in the absolute value of the difference between the POI’s great circle distance and the path distance of the projection points was calculated, conforming to the exponential probability distribution given by Equation (6).

  $$p\left(dt\right)={\displaystyle \frac{1}{\gamma }}{e}^{-dt/\gamma }$$

  (6)

  $$dt={\displaystyle \frac{\left|{∥p_{s_u^i}^{m_i}-p_{s_u^{i+1}}^{m_{i+1}}∥}_{gc}-{∥x_{m_i,j}-x_{m_{i+1},j}∥}_{route}\right|}{timeinterval\ast timeinterval}}$$

  (7)

  $$timeinterval={\displaystyle \frac{{∥x_{m_i,j}-x_{m_{i+1},j}∥}_{route}}{velocity}}$$

  (8)
  The $\gamma$ parameter accounts for the difference between the route distance and the great circle distance. The $timeinterval$ is generally obtained using the ratio of the route distance $\left|\right|x_{m_i,j}-x_{m_{i+1},j}{\left|\right|}_{route}$ to the standard speed limit of urban roads, $velocity$. A vehicle speed of 40km/h was used as the reference value to perform the calculation. The parameters are explained in detail in the Experimental Section.

2.4.2. Generation of Recommended Routes

The KHMM generates routes using the path-finding principle of the Viterbi algorithm [57]. This algorithm identifies the most likely sequence of hidden states through dynamic programming, thereby revealing the path with the highest probability. A key characteristic of the Viterbi algorithm is that it ensures the sub-routes between each node are optimal. The maximum probability of the path between each state node is calculated recursively as a product of probabilities. After determining the path at each step, the optimal route is identified through backtracking.

In our work, the model was provided with a set of observations to identify the observed states. The POI in the KG output formed the first hidden layer, which served as the starting point for the Viterbi algorithm. The first hidden layer then determined the elements of the second layer’s projection points. Identifying the most probable sequence of hidden states was a critical aspect of this process. The second layer calculated the probability of a transition between observed and hidden states for each observed state. This involved identifying the maximum-probability path using the Viterbi algorithm. The path with the highest probability represented the optimal route, and the output node corresponded to the best projection point.

3. Experiments

This section presents experiments conducted to evaluate the effectiveness of the KHMM for personalized route recommendation using real-world datasets. First, we describe the experimental dataset, data preprocessing steps, and parameter configurations. Next, we analyze the results obtained using the KHMM approach. Finally, we compare the performance of the KHMM method with that of traditional methods and provide a detailed analysis of the findings. The experiments were conducted in a controlled environment using Python 3.8, ArcGIS 10.8, and QGIS 3.10.0.

3.1. Experimental Data

This study selected Changsha, Hunan Province, as the study area and obtained spatiotemporal data from crowdsourced sources, including POIs, road networks, OpenStreetMap (OSM), POI popularity indices, and POI attributes. A total of 300 POIs were selected as target sites to ensure the authenticity and validity of the experimental results. These POIs, primarily city tourism sites, were acquired using Baidu Map and OSM, with positional information extracted for road network analysis. To avoid data duplication, the collected POI data were processed to remove redundancies. Large scenic areas, which often contain multiple attractions, were treated as single POI clusters or, in the case of well-known landmarks, as individual POIs. Based on their attributes, the POIs were categorized into 21 themes: Museums/Exhibition Halls, City Parks/Squares, Urban Landmarks, Universities, Zoological and Botanical Gardens, scenic spots, Shopping Centers, Exhibition Centers, Sports Centers, Memorials, Transportation Hub Stations, Churches, Theaters, Technology/Art Galleries, Commercial Plazas, Commercial and Cultural Streets, Temples, Libraries, Historical Sites, Amusement Parks, and Secondary Schools.

The road network for Changsha was obtained using the OSM platform. OSM is a free and collaborative mapping service that allows users to modify, update, and improve map data. The road network consisted of multiple road segments: $G=(r_1,r_2,\dots ,r_j)$. Each road segment included a unique Feature ID (FID) value, start and end points, length, and starting coordinates.

The POI popularity index was derived from the 360 Trends Index, a comprehensive big data platform that monitors internet users’ browsing, search, and social media activity. This index reflects users’ search behavior and interest levels in specific keywords over time, providing insights into trending topics, user needs, and the demographic attributes of keyword searchers. Additionally, the 360 Trends Index is updated in real time, enabling access to search popularity data for different periods and facilitating a timely understanding of changing user interests. The monthly values from the 360 Trends Index for selected POIs in 2022 are presented in

Table 2. A partial example of POI data based on the 360 Trend Index.

Name	Jan	Feb	Mar	Apr	May	June	July	Aug	Sep	Oct	Nov	Dec
Hunan Provincial Museum	171	1469	1435	362	366	446	586	433	348	217	140	113
Changsha Museum	25	37	38	30	32	39	99	84	48	26	19	9
Orange Island	80	74	123	123	178	184	305	299	302	186	122	112
Yuelu Mountain	81	87	130	120	236	217	732	323	295	242	202	160

, illustrating their variation over time.

An online questionnaire was designed to gather user preferences for POIs and their choice of POI attributes. The questionnaire comprised 56 questions divided into three groups to collect specific and balanced information without favoring any particular POI or theme. To ensure data validity and maintain a consistent response rate, 180 questionnaires were distributed, resulting in 147 valid responses. The reliability of the survey participants and content was ensured by the completion of the questionnaire within a short time frame.

The data sources for the POI-KG included encyclopedic web pages such as Baidu Encyclopedia, Wikipedia, Qunar, Ctrip, and attraction-related websites. The attributes of the POIs encompassed their name, subject, category, address, attraction level, construction time, opening hours, characteristics, ticket price, surrounding POIs, relevant POIs, nearby transportation facilities, the recommended duration of a visit, suitable seasons for visiting, and contact information. The POI knowledge graph was constructed using Neo4j [58] for storing and visualizing nodes and attributes. By uncovering entities and relationships between POIs and their attributes, the constructed POI knowledge base included 3724 node labels and 9300 relationships.

3.2. Parameter Settings

In our route recommendation approach, ${\delta }_c$ and $\gamma$ are key parameters [56], determined by the positional measurements between POIs and their projection points. The parameter ${\delta }_c$ is related to the emission probability, with its value derived from the difference in the great circle distance between the POI and the projected point. A higher value of ${\delta }_c$ indicates lower confidence in the positional measurement. Conversely, $\gamma$ represents the difference between the great circle distance of POIs and the route distance of the projected points. The value of $\gamma$ varies across different groups of experimental subjects and is influenced by the data collected from these subjects. We employed the median absolute deviation (MAD) to estimate these parameters, as defined by the following equations:

$${\delta }_c=1.4826\times median{t}_t\left({∥p_{s_u^i}^{m_i}-x_{m_i,j}∥}_{gc}\right)$$

(9)

$$\gamma ={\displaystyle \frac{media{n}_t\left|{∥p_{s_u^i}^{m_i}-p_{s_u^{i+1}}^{m_{i+1}}∥}_{gc}-{∥x_{m_i,j}-x_{m_{i+1},j}∥}_{route}\right|}{ln2}}$$

(10)

These parameters were estimated directly from the experimental data in this study. Based on the experimental data, the value of ${\delta }_c$ was determined to be $35.95$ m. The value of $\gamma$, however, was specific to each group of experimental subjects.

3.3. Validity Analysis

Volunteers were recruited to participate in the development of personalized travel itineraries. Figure 6 shows the user-friendly interface, enabling volunteers to easily select travel subjects of interest and determine visit sequences based on their preferences. Three representative user plans created by these volunteers were selected to test the applicability and feasibility of the KHMM model.

We also examined the impact of POI recommendation indices from different periods on route recommendations. The POI popularity indices from February, May, August, and November were selected as parameter indicators for the calculations. First, the recommendation index was used as the edge weight, and the A* algorithm was applied to find the shortest path in the KG based on the visit sequences. Second, the output POIs represented the first hidden state, determining the second hidden element in the KHMM. Finally, the user obtained a complete optimal path within the road network.

Based on the subjects and visit sequences from the first group of experiments, the shortest route was output in the POI-KG. As illustrated in Figure 7, large pink nodes represented subjects while small nodes represented POIs, with different colors indicating different POI categories. Edge weights in the KG were based on the popularity indices of POIs from different months, reflecting variations in the POI outputs for different subjects across the months. Figure 8 depicts the personalized route recommendation output from the KHMM. The POIs from the POI-KG served as the first hidden state in the KHMM, which were then mapped to the actual road network to calculate the optimal projection position. Finally, a complete optimal route was recommended based on the user’s visit sequence.

Figure 8a shows the optimal recommended route for February, with numbers indicating the sequence of projection points. Identifying the projection points illustrated the importance of and priorities in route selection. For example, IFS had three projection positions in the road network: IFS 1, IFS 2, and IFS 3. Combining the location analysis of the previous IFS POI and the next POI, IFS 2 was selected as the optimal projection point. This selection was based not only on the geographical appropriateness but also on the convenience of the route. The distances between sequential projection points also influenced the selection of projection points. Figure 8b shows the recommended route for May, c for August, and d for December.

Figure 9 and Figure 10 show the optimal path results for experiment 2 and experiment 3. As is evident from the results, there were distinctions between the recommended routes in different months. This reflects the dynamic adjustment of POIs in routes over time. Such variations occur because each POI receives different levels of public attention at different times, and its search volume and popularity fluctuate. For example, some seasonal attractions may be popular in certain months and relatively quiet in others. The existence of such time sensitivity necessitates the consideration of seasonal and time-sensitive factors when planning itineraries to ensure that the recommended routes are both reasonable and attractive. Moreover, different environments and access sequences also influence the selection of POIs. The construction of travel routes is not simply a matter of stacking POIs but must account for the connections between them. When faced with different subject access sequences, intermediate nodes adapt to changes in the front and back nodes, optimizing the efficiency and attractiveness of the route. This adaptation responds to changing user needs and provides more personalized and dynamic POI recommendations.

On the other hand, the POI popularity index may have small values in one month but increase rapidly in another month. An increase in the popularity index raises the recommendation rate. However, the most popular POI does not necessarily represent the best choice for the subject. Distance factors and other additional influencing factors are also considered to calculate the best route. Most candidates correspond to the maximum recommendation value of POIs within the subject. This clearly demonstrates that the POI popularity index is a significant influencing factor in route recommendation.

This method reflects the dynamic changes in routes and captures their variation over time by comprehensively considering the recommendation degree values of POIs over different time scales. In practical applications, people’s interest in and demand for locations change over time. As an important indicator, the POI popularity index intuitively reflects this trend. Therefore, by combining the POI popularity index and POI attribute information for correlation analysis, users can receive more personalized and timely route recommendation services. This method improves the user’s travel experience, enhances the accuracy and practicality of route recommendations, and provides better route options suited to the user’s interests and current time context.

The POI-KG structure enhances the connections between POIs by leveraging POI attributes to reveal public preferences and construct a preference matrix. This structure not only captures the static attributes of POIs but also dynamically reflects the changing trends in public interest. The KG helps identify relationships between the public, users, and POIs. Integrating the POI popularity index with a KG increases the relevance of the data. By capturing dynamic changes in POI indices and their broader context within the KG, we gain insights into the factors influencing user preferences and behaviors, leading to more contextually meaningful recommendations.

Users can further define their preferences based on attributes such as the POI type, ticket requirements, and attraction level, rather than just selecting topics of interest. For example, if a user prefers to visit free POIs, the system will avoid suggesting POIs that require a ticket. In the POI-KG, nodes that require a ticket to be purchased will be skipped. In our initial experiments, routing results based on May and November indicated that Yuelu Academy and Tianxin Pavilion required tickets. The personalized route recommendations were re-routed to satisfy the user’s preferences, resulting in recommendations for Changsha Museum, Yuelu Mountain, IFS, and Wenheyou.

3.4. Comparative Experimental Analysis

The proposed KHMM model was compared with three baseline methods: a HMM, the A* algorithm, and the Top POI Popularity approach.

• HMM: A raw HMM relies on an initial state probability, a transition probability matrix, and an observation probability matrix. It uses POIs’ geographic locations to model data features, mapping positions to road segments to estimate POI states and predict sequences.
• A* Algorithm: A* is a widely used path-finding and graph traversal algorithm that identifies the optimal path from a start node to a goal node in a graph or network. It minimizes the total cost for paths with multiple nodes.
• Top POI Popularity: This method selects POIs with the highest popularity index within a subject area, using these as the basis for route planning.

The performance of these methods was evaluated using heuristic-based metrics [32], including the following:

• POI Popularity: The sum of popularity indices for POIs along the recommended route.
• Number of POI Similarities: The number of POIs consistent with the highest-ranked POIs in terms of popularity.
• Route Distance: The total distance traveled for the recommended itinerary.
• POI Recommendation Number: The total number of POIs recommended to the user during the itinerary.

Figure 11, Figure 12 and Figure 13 illustrate the routes generated by the KHMM and the baseline methods. In these figures, (a) shows the route generated by the KHMM, (b) shows the route generated by the HMM, (c) shows the route generated by A*, and (d) shows the route based on the highest POI popularity index. The results for the KHMM, HMM, A*, and Top POI Popularity were evaluated using the aforementioned metrics.

Figure 14 displays the total popularity indices of POIs for paths recommended across different months. The KHMM method consistently achieved high recommendation indices, with most suggested POIs ranking well within the overall index. In contrast, the HMM and A* yielded lower popularity indices, indicating that these methods overlook the degree of recommendation of POIs during path planning, potentially missing popular and worthwhile locations.

Figure 15 shows that the KHMM excelled in maintaining consistency with the most highly recommended POIs for each topic. Its recommendations shared over 50% similarity with the Top POI Popularity results, better matching top-ranked attractions. This demonstrates the KHMM’s ability to meet user expectations for highly popular destinations. In contrast, the HMM and A* yielded lower satisfaction levels.

Figure 16 presents the total route distances generated by the four methods. The HMM consistently produced the longest distances, likely because it prioritized state transition probabilities over distance optimization. The Top POI Popularity also failed to optimize the journey length effectively. While A* generated the shortest routes, the KHMM struck a balance, offering reasonable distances that considered user needs and the overall experience.

Figure 17 shows that the KHMM recommended more POIs than the standard method, which aligned POI counts with subject counts. For example, in the third experiment, the KHMM recommended six POIs per subject but output twelve POIs based on the POI-KG’s popularity index. By leveraging the KG’s rich information, the KHMM extended recommendations beyond specific visit topics, enabling users to discover related and potentially connected destinations. This enhanced the accuracy, diversity, and personalization of travel recommendations.

4. Discussion and Conclusions

This paper introduces a novel KHMM framework that integrates spatial and semantic relationships between POIs into a unified decision-making model for personalized route recommendations. By expanding the state space of a traditional HMM through the incorporation of knowledge graphs, the KHMM enables the integration of multi-dimensional POI information and higher-order relationships. The construction of a POI-KG establishes finer-grained relationships between POI entities and their attributes, enriching the dimensionality of POI data and enhancing their interconnectivity. Additionally, the introduction of a POI popularity index provides insights into public behavior and attention levels at different times. By combining POI attributes with user behavior data, the KHMM offers a deeper understanding of the connections between POIs and their relationships with users.

4.1. Embedding the Popularity Index Enhances Route Recommendations

Our case study in Changsha demonstrates the effectiveness of embedding the POI popularity index into the recommendation framework. As a quantitative measure of public attention to specific locations over time, the index reveals trends in POI popularity and their temporal variations. In the experiments, we collected data across different months to investigate the impact of time scales on the recommendation results. The findings indicate that POI recommendation values and suggested routes vary with time, underscoring the critical role of temporal dynamics in recommendation systems. This provides empirical support for optimizing future algorithms.

In practical applications, the user interest and demand for locations evolve over time. The POI popularity index captures these trends, enabling the KHMM to outperform traditional methods by aligning recommendations with public preferences. By incorporating dynamic temporal data, the KHMM enhances user satisfaction and ensures the timeliness of recommended routes.

4.2. Context-Aware Route Recommendations

The KHMM embeds a knowledge graph into a two-layer HMM, expanding the state space to leverage associations between multi-dimensional POI attributes. Crowdsourced spatiotemporal data are dynamically mapped onto the road network, enabling personalized route recommendations based on user preferences. The results demonstrate that incorporating dynamic temporal and spatial data improves the model’s adaptability to changing user behaviors, supporting real-time, context-aware recommendations. By bridging individual preferences and road network structures, our research offers key insights into the factors shaping travel behavior and contributes to the development of adaptive urban transportation systems.

Furthermore, the POI popularity index integrates bidirectional weights within the knowledge graph, combining static and dynamic data while accounting for environmental interactions. This integration yields more accurate and context-aware recommendations, highlighting the KHMM’s potential to advance intelligent travel services with improved spatial accuracy and personalized planning.

While the proposed KHMM model demonstrates satisfactory performance, further research is needed to incorporate additional factors. Future work will explore a broader range of entity data and relationships between entities and attributes. With the diversification of social platforms and user needs, new data sources for evaluating POIs and travel factors—such as transportation modes, travel times, and budgets—have emerged. Incorporating these factors will enhance route flexibility and recommendation accuracy, paving the way for more robust and adaptive travel planning systems.