Big Data-Driven Dynamic Analysis of Tourist Behavioral Trajectories and Intelligent Service Strategies in Tourist Attractions

In the context of the big data era, the traditional tourism model has gradually been replaced by a new, big data-led tourism model, and the new tourism model has led to a series of changes in the function and value of the traditional tourism model [1-2]. First, it has prompted tourism enterprises to deepen their connotation and extension. The traditional tourism industry is based on the coordination of tourism element information through the grouping of tours, providing tourism routes and tourism products, organizing tourists, and dredging upstream and downstream, and the data analysis also mainly comes from the internal records of the enterprise. Under the popularity of big data technology, traditional tourism enterprises have also gradually changed their mode of thinking, and on the basis of the traditional value chain, online marketing platforms such as Where to Go, Ctrip Travel, Touniu, etc., have emerged to provide better-personalized services for tourists based on the analysis of big data on tourists [3-5]. Secondly, the transformation of tourists’ tourism mode, tourists are the main body in tourism activities and are the most directly affected group in the whole tourism chain industry. With the development of information technology, tourists only use a cell phone, which can be from a search engine, social platforms, popular reviews, the various portals, to choose a simpler, more diversified, and convenient way to travel, including the destination of word of mouth, routes, attractions, human history and other types of information, so as to choose the group tour, self-driving tour, companion tour, and other travel modes suitable for their own [6-8].

For a long time, given the technical limitations, it has been impossible to accurately obtain the spatiotemporal behavior trajectory of tourists in tourist attractions and the spatiotemporal behavior of tourists within the city, thus making the quantitative study of the spatiotemporal change law of tourists face great challenges. However, in recent years, with the rapid development of communication technology, the emergence of mobile Internet, Internet of Things, sensing devices, etc., the current technology has been able to realize the minute-level aggregation and calculation of data [9-11]. Tourist’s cell phones, from the start of boot access to the network, began to non-stop and communication equipment and systems for signaling interaction, the network in real-time to grasp the location of the tourist’s area code and cellular cell information. The billing system reads real-time calls, SMS, and data service information. The location of the access network is included in these call orders. The contracted user server stores user information, including name, ID number, address, occupation, and contracted information. A simple call and Internet access will generate hundreds of signaling data, and even after filtering and screening, the amount of data is still very large. Signaling data, billing system, HSS server, convergence layer massive data flow, a little convergence to the big data platform, so that the user’s various behavioral trajectory and browsing information to achieve accurate monitoring, geographic information big data makes it possible to study the spatial and temporal laws of tourists [12-15].

In this paper, based on the existing theories related to trajectory analysis, we propose a trajectory analysis method for tourist behavior based on their long and short-term interaction preferences. The filtering mechanism obtained from the user trajectory characteristics is used to filter the target user, and then the user’s points of interest are analyzed according to the user’s long and short-term interaction preference. The combination of spatio-temporal attributes of behavioral trajectory and user preference model is applied to the point-of-interest recommendation scenario. Through the long and short-term memory model vectorization, the user’s trajectory information and user preference information are input into the recurrent neural network so as to capture the user’s spatio-temporal characteristics and long and short-term preferences in order to prepare for the point-of-interest recommendation. Finally, this paper conducts experiments on public real datasets to obtain the best recommendation effect of tourism points of interest, and conducts clustering analysis on tourists’ portraits and tourists’ differential preferences of points of interest during tourism.

2

Overview

1)

Analysis of tourist behavior trajectory in tourist attractions

The study of tourists’ spatial behavior is very important for the planning and management of tourist attractions, the improvement of tourists’ satisfaction, and the diversion of congestion guidance. Literature [16] proposes a grid-based open GPS trajectory data processing framework to take the tourist data of Yuanmingyuan Park scenic spot as an example to conduct the dynamic analysis of tourist behavior trajectory, and the experimental results found that there is a close relationship between the tourist access paths and the stay time, the stay time and the photo taking, and there is a strong consistency of the behavior of tourists in different seasons. Literature [17] takes Huashan, China, as the study site and uses open GPS trajectory data analysis, Markov chain, and cluster analysis methods to show the spatio-temporal movement behaviors of tourists within the study site, aiming to provide important insights for destination planning, marketing and resource management by destination management departments (DMDs). Literature [18] proposed a conceptual model based on a behavioral perspective based on the theory of spatio-temporal tourist behavior (STTB) as a comprehensive analytical framework to facilitate the accurate tracking of spatio-temporal travel behaviors in tourist destinations. Literature [19] used Ocean Park in Hong Kong as an experimental site to collect spatiotemporal behavioral information of tourists using handheld GPS tracking devices and proposed a geographic information system (GIS)-based visualization and clustering method in order to facilitate the analysis of spatiotemporal behavioral patterns of tourists at the micro-scale, which in turn provides reference value for theme park attraction management and tourists’ experience enhancement.

2)

Intelligent service strategy for tourist attractions

With the rapid development of the tourism service industry, the traditional tourism service does not meet the needs of users. Literature [20] mainly analyzes the tourism management strategy under the intelligent tourism Internet of Things service platform and proposes an energy consumption balanced business combination method based on the weighted regression mathematical method, ant colony algorithm, and cumulative aggregation quality of service, which is aimed at guaranteeing the life cycle and stability of the entire tourism service combination workflow. Literature [21] constructs a comprehensive tourism information service platform with the support of genetic algorithms and other big data technologies, which can improve the efficiency of tourism management and provide reference data for the construction of tourism information intelligent service. Literature [22] designed a conceptual framework of customer knowledge management consistent with smart tourism and consisting of eight processes, aiming to achieve intelligent management of tourism experience by destination management organizations. Literature [23] improves structural equation modeling based on the UTAUT2 model to examine the influencing factors of scenic area users’ willingness to use the smart tourism service system, and the results show that four factors, namely, convenience, social influence, performance expectation, and effort expectation, indirectly or directly affect users’ willingness to use the scenic area’s smart tourism service system, and the study provides sustainable and efficient development of the smart tourism service system with a Theoretical basis.

3

Method

3.1

Big Data-driven Tourist Behavior Trajectory Dynamic Analysis Technology

3.1.1

Behavioral Trajectory Preprocessing Techniques

Due to the influence of factors such as the precision limitation and storage mechanism of the dynamic collection equipment of tourists’ behavioral trajectory in tourist attractions driven by big data, the spatiotemporal trajectory data obtained from the collection is not accurate. Generally speaking, the more trajectory points are collected, the more accurate the trajectory accuracy is. However, due to the implicit feedback of tourists’ activities, the collected trajectory data may be insufficient. The data collection of trajectory points for moving objects with high frequency will have problems with data storage, transmission and processing. Therefore, we need to pre-process the trajectory data to improve the accuracy and sparsity of the data. Location-based service applications will obtain a lot of trajectory data, but some of these data will contain noisy data. Therefore, people must perform sparse processing and noise filtering on these trajectory data, which plays an important role in data cleaning, transmission, and storage of location-based services. The motion of moving objects is continuous, but the collected trajectory data has a certain discrete nature. Reducing the data acquisition frequency can solve this problem to a certain extent. However, reducing the acquisition frequency, in turn, will make a large part of the collected data may not be the data required by the application, and even most of them cannot represent the motion characteristics of the moving object. For the study of trajectory data, this paper uses the common mean filter trajectory noise filtering technique.

Mean value filtering [24-25], also known as linear filtering, is a method to reduce the noise of smoothed data with the following formula: (1) $X_{i} = \frac{1}{n} \sum_{j = i - n + 1}^{i} z_{j}$ \[{{X}_{i}}=\frac{1}{n}\sum\limits_{j=i-n+1}^{i}{{{z}_{j}}}\]

From the above equation, it can be seen that the moving smoothing window filtering approach can be based on known trajectory points. Meanwhile, the method is sensitive to discrete points. Further, a more effective mean value filtering method can be proposed for GPS point z with the following equation: (2) $X_{i} = m e d i a n {z_{i - n = 1}, z_{i - n = 2}, ..., z_{i - 1}, z_{i}}$ \[{{X}_{i}}=median\{{{z}_{i-n=1}},{{z}_{i-n=2}},...,{{z}_{i-1}},{{z}_{i}}\}\]

When the trajectory points are sparse, and the sliding window is large, other filtering methods should be used because the error of mean value filtering will be larger at this time.

3.1.2

Trajectory clustering

Trajectory clustering is to further divide similar trajectory segments or trajectories together on the basis of similarity analysis of trajectory data. The common trajectory clustering method is to represent the trajectories or trajectory segments as word vectors or segment vectors, and to complete the clustering analysis by calculating the distance between the vectors. Existing trajectory clustering techniques are divided into two categories: global trajectory clustering and sub-trajectory clustering. Most of the existing clustering methods consider the spatial location information or temporal information of the trajectories separately in the sub-trajectory similarity metric without considering the spatial and temporal properties of the trajectories jointly. This makes it inevitable that some trajectory segments are assigned to the wrong clusters.

Trajectory clustering algorithms usually represent trajectories in terms of feature vectors. Each dimension of the feature vector is associated with an attribute of the trajectory, so the number of dimensions is equal to the number of attributes. The similarity between trajectories is measured using the distance between feature vectors; the shorter the distance, the greater the similarity between trajectories. It should be noted that the dimensionality of different feature vectors may be different, which makes the existing trajectory clustering algorithms unavailable due to the fact that trajectories can contain many attributes such as length, shape, sampling rate, and so on of the trajectories. Therefore, a comprehensive comparison should be made based on different trajectory segments. A trajectory clustering method based on deep representation learning can re-examine the trajectory clustering problem by learning low-dimensional trajectory representations. In particular, a set of motion behavior features is extracted using a sliding window to capture the spatio-temporal invariant features in the motion trajectories, and then each trajectory is converted into a sequence of features describing the target motion. Finally, the learned trajectory representations are clustered using the K-means clustering algorithm.

3.1.3

Attention model

Attention modeling [26] is a method to disperse different attention strategies at different levels, which can simulate the way human beings observe things and allocate different attention to different parts of things by observing the whole thing. Among the known sequence modeling methods, the RNN model obtains the global information of the thing through iterative recursion, but the iterative steps are usually too redundant, whereas the CNN model can increase the observation angle of the thing by superimposing multiple windows, which can only focus on the local information of the thing, but can be enough to feel the overall information of the thing. Comparatively speaking, the attention model is more direct, it can directly obtain the global information of the thing by focusing on different components of the sequence.

The following equation can define the basic attention model: (3) $A t t e n t i o n (Q, K, V) = s o f t m a x (\frac{Q K^{t}}{\sqrt{d_{k}}}) V$ \[Attention(Q,K,V)=softmax\left( \frac{Q{{K}^{t}}}{\sqrt{{{d}_{k}}}} \right)V\] where Q ∈ R^n×d_k ,K ∈ R^m×d_k ,V ∈ R^m×d_v . Is essentially the multiplication of three matrices n×d_k ,m×d_k ,m×d_v to get a matrix of n×d_v, so the attention mechanism can be thought of as encoding a sequence of n×d_k text mapped into a new sequence of n×d_v text. Looking at this for each specific vector, the following formula can be derived: (4) $A t t e n t i o n (Q, K, V) = \sum_{s = 1}^{m} \frac{1}{z} \exp (\frac{q_{t}, k_{s}}{\sqrt{d_{k}}}) V$ \[Attention(Q,K,V)=\sum\limits_{s=1}^{m}{\frac{1}{z}}\exp \left( \frac{{{q}_{t}},{{k}_{s}}}{\sqrt{{{d}_{k}}}} \right)V\]

In the above formula, variable Z denotes the normalization factor, q,k,v denotes query, key, and value respectively, and K and V in the formula are one-to-one correspondence with key and value. From the formula we can see that q_t and vector k_s are multiplied by inner product and the similarity value between q_t and v_s is obtained by using softmax function. Finally, a new vector d_k is obtained by weighted summation, and the parameter ${\sqrt{d}}_{k}$ ${{\sqrt{d}}_{k}}$ in the formula mainly plays a regulating role in order to make the inner product not too large. Otherwise, it will be easy to be categorized as 0 or 1 in the Softmax layer classification.

In general, the attention mechanism can be expressed as a mapping, which is derived after computing a query by key-value pairs. Its main computational steps are expressed as follows: first, the corresponding weights of each query and each key can be computed by calculating the similarity, which is usually calculated by splicing, dot product, and perceptual machine; after calculating the similarity, the weights can be normalized by using the related Softmax function; finally, the required attention value can be obtained by summing up the weights of The final attention value can be obtained by weighting and summing the corresponding value and the previously obtained weights.

After h linear mapping of QKV through the parameter matrix, the multi-head attention mechanism uses the attention mechanism to splice the h results obtained as expressed in the following specific formula: (5) $h e a d_{i} = A t t e n t i o n (Q W_{i}^{q}, K W_{i}^{k}, V W_{i}^{ν})$ \[hea{{d}_{i}}=Attention(QW_{i}^{q},KW_{i}^{k},VW_{i}^{\nu })\]

Eq. $W_{i}^{q} \in R^{d_{q} \times {\tilde{d}}_{q}}, W_{i}^{k} \in R^{d_{k} \times {\tilde{d}}_{k}}, W_{i}^{ν} \in R^{d_{ν} \times {\tilde{d}}_{ν}}$ $W_{i}^{q}\in {{R}^{{{d}_{q}}\times {{{\tilde{d}}}_{q}}}},W_{i}^{k}\in {{R}^{{{d}_{k}}\times {{{\tilde{d}}}_{k}}}},W_{i}^{\nu }\in {{R}^{{{d}_{\nu }}\times {{{\tilde{d}}}_{\nu }}}}$ After that, the above formula is poly-spliced to arrive at the desired formula for the value of attention: (6) $M u l t i H e a d A t t e n t i o n (Q, K, V) = C o n c a t (h e a d_{1}, h e a d_{2}, ..., h e a d_{i})$ \[MultiHeadAttention(Q,K,V)=Concat(hea{{d}_{1}},hea{{d}_{2}},...,hea{{d}_{i}})\]

3.2

Intelligent Service Recommendation Model Based on Tourist Behavioral Trajectory Dynamics

In order to improve the accuracy of tourists’ next point of interest recommendation and the experience of smart service recommendation, this paper proposes a next POI point of interest recommendation model (MALS) based on multi-level attention mechanism and long and short-term preference modeling by combining deep learning and attention mechanism for the sparse characteristics of tourists’ check-ins, while considering contextual information such as time and geographic location.

3.2.1

Dynamic long-term preference module for tourists’ behavioral trajectories

1)

Long-term preference module

The tourist’s historical check-in sequence is input into the LSTM [27] by taking into account the six influencing factors of check-in, namely POI number, time, geographic location, number of Sundays, distance, and time difference of the tourist’s check-in at the same time to form an embedding vector representation of the tourist’s long-term check-in, which is then input into the deep learning model. The temporary embedding vector aggregation ${\tilde{x}}_{t_{k}}^{l}$ $\tilde{x}_{{{t}_{k}}}^{l}$ formed by the embedding layer for each moment is denoted as: (7) ${\tilde{x}}_{t_{k}}^{l} = W_{v} v_{t_{k}}^{u} + W_{l} l_{v} + W_{t} t_{k} + W_{w} w_{t_{k}} + W_{d} d_{t_{k}} + W_{t d} t d_{t_{k}} + {\tilde{b}}^{l}$ \[\tilde{x}_{{{t}_{k}}}^{l}={{W}_{v}}v_{{{t}_{k}}}^{u}+{{W}_{l}}{{l}_{v}}+{{W}_{t}}{{t}_{k}}+{{W}_{w}}{{w}_{{{t}_{k}}}}+{{W}_{d}}{{d}_{{{t}_{k}}}}+{{W}_{td}}t{{d}_{{{t}_{k}}}}+{{\tilde{b}}^{l}}\]

where W_v ∈ℝ^d×d ,W_l ∈ℝ^d×d ,W_t ∈ℝ^d×d ,W_w ∈ℝ^d×d ,W_d ∈ℝ^d×d ,W_td ∈ℝ^d×d , is the weight matrix and ${\tilde{b}}^{l} \in ℝ^{d}$ ${{\tilde{b}}^{l}}\in {{\mathbb{R}}^{d}}$ is the bias vector. $v_{t_{k}}^{u}$ $v_{{{t}_{k}}}^{u}$ is the embedding vector representation of POI number $v_{t_{k}}^{u}$ $v_{{{t}_{k}}}^{u}$ , t_v is the embedding vector representation of POI geographic location l_v , t_k is the embedding vector representation of check-in timestamp t_k , w_{t_k} is the embedding vector representation of number of days in a week w_{t_k} , d_{t_k} is the embedding vector representation of geographic distance between the current check-in and the last check-in d_{t_k} , and td_{t_k} is the embedding vector representation of time difference between the current check-in and the last check-in td_{t_k} .

2)

Deep learning network using LSTM as a long-term preference module

The temporary embedding vector aggregation ${\tilde{x}}_{t_{k}}^{l}$ $\tilde{x}_{{{t}_{k}}}^{l}$ at t_k moment generated in the embedding layer as shown in the following equation is fed into the LSTM, and together with the hidden state $h_{t_{k - 1}}^{l}$ $h_{{{t}_{k-1}}}^{l}$ at t_k–1 moment in the LSTM undergoes a series of operations to obtain the hidden state $h_{t_{k}}^{l}$ $h_{{{t}_{k}}}^{l}$ at t_k moment as: (8) $h_{t_{k}}^{l} = L S T M ({\tilde{x}}_{t_{k}}^{l}, h_{t_{k - 1}}^{l})$ \[h_{{{t}_{k}}}^{l}=LSTM(\tilde{x}_{{{t}_{k}}}^{l},h_{{{t}_{k-1}}}^{l})\]

where LSTM(·) is the set of formulas of the LSTM algorithm that can eventually obtain all the hidden states $H^{l} = {h_{t}_{1}^{l}, \dots h_{t_{k}^{l}}^{l}, \dots h_{t_{N}}^{l}}$ \[{{H}^{l}}=\{{{h}_{t}}_{1}^{l},\cdots h_{t_{k}^{l}}^{l},\cdots h_{{{t}_{N}}}^{l}\}\] of T = {t₁,…t_k,…t_N}.

3)

Contextual attention mechanism

Let $\tilde{x} (i, t_{k})$ $\tilde{x}(i,{{t}_{k}})$ denote the i th feature factor of the k th history check-in, I is the total number of feature factors affecting the decision of check-in, then 1 ≤ i ≤ I , I = 6 , and let ρ(i,t_k) denote the corresponding influence weight of the i th feature factor of the k th history check-in in the k th history check-in, and use the softmax function for the normalization, and there are: (9) $\tilde{ρ} (i, t_{k}) = \tanh (W_{i} [h_{t_{k - 1}}^{l}; c_{t_{k - 1}}^{l}] + W_{i}^{\tilde{x}} \tilde{x} (i, t_{k}) + b_{i})$ \[\tilde{\rho }(i,{{t}_{k}})=\tanh ({{W}_{i}}[h_{{{t}_{k-1}}}^{l};c_{{{t}_{k-1}}}^{l}]+W_{i}^{{\tilde{x}}}\tilde{x}(i,{{t}_{k}})+{{b}_{i}})\] (10) $ρ (i, t_{k}) = \frac{e x p (\tilde{ρ} (i, t_{k}))}{\sum_{i = 1}^{l} e x p (\tilde{ρ} (i, t_{k}))}, 1 \leq i \leq l$ \[\rho (i,{{t}_{k}})=\frac{exp(\tilde{\rho }(i,{{t}_{k}}))}{\sum\limits_{i=1}^{l}{e}xp(\tilde{\rho }(i,{{t}_{k}}))},1\le i\le l\] (11) $e x p (x) = e^{x}$ \[exp(x)={{e}^{x}}\]

Where, $W_{i} \in ℝ^{d \times 2 d}, W_{i}^{\tilde{x}} \in ℝ^{d \times d}$ ${{W}_{i}}\in {{\mathbb{R}}^{d\times 2d}},W_{i}^{{\tilde{x}}}\in {{\mathbb{R}}^{d\times d}}$ is the weight matrix, b_i ∈ℝ^d is the bias vector, $c_{t}_{k - 1}^{l}$ ${{c}_{t}}_{k-1}^{l}$ is the state of the cell at the t_k–1 moment in the LSTM network, $\tilde{ρ} (i, t_{k})$ $\tilde{\rho }(i,{{t}_{k}})$ is the ad hoc representation of the weights accounted for by the check-in factors, and ρ(i,t_k) denotes the weights accounted for by the i th factor in this check-in. The exact embedding vector of the tourist check-in activity is denoted as: (12) $\hat{x} (i, t_{k}) = \tilde{x} (i, t_{k}) \times ρ (i, t_{k})$ \[\hat{x}(i,{{t}_{k}})=\tilde{x}(i,{{t}_{k}})\times \rho (i,{{t}_{k}})\] (13) ${\hat{x}}_{t_{k}}^{l} = \sum_{i = 1}^{l} W (i) \hat{x} (i, t_{k}) + b$ \[\hat{x}_{{{t}_{k}}}^{l}=\sum\limits_{i=1}^{l}{W}(i)\hat{x}(i,{{t}_{k}})+b\]

where W(i) ∈ℝ^d ×d is the weight matrix of the i nd feature factor and b ∈ℝ^d is the bias vector. $\hat{x} (i, t_{k})$ $\hat{x}(i,{{t}_{k}})$ is the embedding vector representation of the i th feature factor of the updated k th history check-in. The final expression $x_{t_{k}}^{l}$ $x_{{{t}_{k}}}^{l}$ for the long-term check-in behavior of tourists is: (14) $x_{t_{k}}^{l} = {\hat{x}}_{t_{k}}^{l} + λ_{c}^{l} r_{c}^{u}$ \[x_{{{t}_{k}}}^{l}=\hat{x}_{{{t}_{k}}}^{l}+\lambda _{c}^{l}r_{c}^{u}\]

where $λ_{c}^{l}$ $\lambda _{c}^{l}$ is the share of category preferences of visitor u in the long-term preference module in the current check-in behavior, and $x_{t_{k}}^{l}$ $x_{{{t}_{k}}}^{l}$ is the final embedding vector representation of the hidden state. There: (15) $h_{t_{k}}^{l} = L S T M (x_{t_{k}}^{l}, h_{t_{k - 1}}^{l})$ \[h_{{{t}_{k}}}^{l}=LSTM(x_{{{t}_{k}}}^{l},h_{{{t}_{k-1}}}^{l})\]

Eventually, all the check-in activity embedding vector representations are used as input data into LSTM to get all the hidden states $H^{l} = {h_{t_{1}}^{l}, \dots, h_{t_{k}}^{l}, \dots, h_{t_{N}}^{l}}$ ${{H}^{l}}=\{h_{{{t}_{1}}}^{l},\cdots ,h_{{{t}_{k}}}^{l},\cdots ,h_{{{t}_{N}}}^{l}\}$ of T = {t₁,…,t_k,…,t_N} .

4)

Construct the embedding vector [28] representation of tourists’ long-term preferences

Let H^l ∈ℝ^d×N be the matrix consisting of all hidden state vectors ${h_{t_{1}}^{l}, \dots, h_{t_{k}}^{l}, \dots, h_{t_{N}}^{l}}$ $\{h_{{{t}_{1}}}^{l},\cdots ,h_{{{t}_{k}}}^{l},\cdots ,h_{{{t}_{N}}}^{l}\}$ in the long-term preference module, where d is the dimension of the hidden state vectors, each hidden state is a d -dimensional vector, and N is the length of the long-term sequence. Then there are: (16) $μ_{k}^{l} = \frac{\exp (g (h_{t_{k}}^{l}, q_{u}^{l}))}{\sum_{i = 1}^{N} e x p (g (h_{t_{k}}^{l}, q_{u}^{l}))}$ \[\mu _{k}^{l}=\frac{\exp (g(h_{{{t}_{k}}}^{l},q_{u}^{l}))}{\sum\limits_{i=1}^{N}{e}xp(g(h_{{{t}_{k}}}^{l},q_{u}^{l}))}\] (17) $g (h_{t_{k}}^{l}, q_{u}^{l}) = \frac{h_{t_{k}}^{l} {(q_{u}^{l})}^{T}}{\sqrt{d}}$ \[g(h_{{{t}_{k}}}^{l},q_{u}^{l})=\frac{h_{{{t}_{k}}}^{l}{{(q_{u}^{l})}^{T}}}{\sqrt{d}}\]

where $q_{u}^{l}$ $q_{u}^{l}$ is the query information for the long-term check-in sequence in the temporal attention mechanism. The long-term preferences of visitors at the fine-grained POI level $r_{l}^{u}$ $r_{l}^{u}$ are: (18) $r_{l}^{u} = \sum_{k = 1}^{N} μ_{k}^{l} h_{t_{k}}^{l}$ \[r_{l}^{u}=\sum\limits_{k=1}^{N}{\mu _{k}^{l}}\;h_{{{t}_{k}}}^{l}\]

where k denotes the visitor’s k rd check-in at time t_k , and N is the total number of check-ins in the long-term check-in sequence.

3.2.2

Dynamic short-term preference module for tourists’ behavioral trajectories

1) The embedding layer of the short-term preference module is used to obtain the embedding vector representation of the short-term check-in

In this paper, we use the last S check-ins in the historical check-in sequence as the tourist’s short-term check-in sequence, denoted as $S_{u} = {A_{t_{N - s + 1}}^{u}, A_{t_{N - s + 2}^{u}}^{u}, \dots, A_{t_{m}}^{u}, \dots, A_{t_{N}}^{u}}$ ${{S}_{u}}=\{A_{{{t}_{N-s+1}}}^{u},A_{t_{N-s+2}^{u}}^{u},\cdots ,A_{{{t}_{m}}}^{u},\cdots ,A_{{{t}_{N}}}^{u}\}$ , where $A_{t_{N - s + 1}^{u}}$ ${{A}_{t_{N-s+1}^{u}}}$ denotes the tourist’s first check-in in the short-term, S is the total number of short-term check-ins of the tourist, and an item in the set of short-term check-in sequences is denoted as $A_{t_{m}}^{u} = (u, v_{t_{m}}^{u}, l_{v}, c_{v}, t_{m}, w_{t_{m}})$ $A_{{{t}_{m}}}^{u}=(u,v_{{{t}_{m}}}^{u},{{l}_{v}},{{c}_{v}},{{t}_{m}},{{w}_{{{t}_{m}}}})$.

The temporary embedding vector aggregation ${\tilde{x}}_{t_{m}}^{s}$ \[\tilde{x}_{{{t}_{m}}}^{s}\] formed by the embedding layer for each moment is denoted as Eq: (19) ${\tilde{x}}_{t_{m}}^{s} = W_{v}^{s} v_{t_{m}}^{u} + W_{l}^{s} l_{v} + W_{t}^{s} t_{m} + W_{w}^{s} w_{t_{m}} + W_{d}^{s} d_{t_{m}} + W_{t d}^{s} t d_{t_{m}} + {\tilde{b}}^{s}$ $\tilde{x}_{{{t}_{m}}}^{s}=W_{v}^{s}v_{{{t}_{m}}}^{u}+W_{l}^{s}{{l}_{v}}+W_{t}^{s}{{t}_{m}}+W_{w}^{s}{{w}_{{{t}_{m}}}}+W_{d}^{s}{{d}_{{{t}_{m}}}}+W_{td}^{s}t{{d}_{{{t}_{m}}}}+{{\tilde{b}}^{s}}$

$W_{v}^{s} \in ℝ^{d \times d}, W_{l}^{s} \in ℝ^{d \times d}, W_{t}^{s} \in ℝ^{d \times d}, W_{w}^{s} \in ℝ^{d \times d}, W_{d}^{s} \in ℝ^{d \times d}, W_{t d}^{s} \in ℝ^{d \times d}$ $W_{v}^{s}\in {{\mathbb{R}}^{d\times d}},W_{l}^{s}\in {{\mathbb{R}}^{d\times d}},W_{t}^{s}\in {{\mathbb{R}}^{d\times d}},W_{w}^{s}\in {{\mathbb{R}}^{d\times d}},W_{d}^{s}\in {{\mathbb{R}}^{d\times d}},W_{td}^{s}\in {{\mathbb{R}}^{d\times d}}$ is the weight matrix, ${\tilde{b}}^{s} \in ℝ^{d}$ ${{\widetilde{b}}^{s}}\in {{\mathbb{R}}^{d}}$ is the bias vector. $v_{t}_{m}^{u}$ ${{v}_{t}}_{m}^{u}$ is the embedding vector representation of POI number $v_{t}_{m}^{u}$ ${{v}_{t}}_{m}^{u}$ , l_v is the embedding vector representation of POI geographic location l_v , t_m is the embedding vector representation of check-in timestamp t_m , w_{t_m} is the embedding vector representation of number of days in a week w_{t_m} , d_{t_m} is the embedding vector representation of geographic distance between the current check-in and the last check-in d_{t_m} , and td_{t_m} is the embedding vector representation of time difference between the current check-in and the last check-in td_{t_m} . The expression $x_{t_{m}}^{s}$ $x_{{{t}_{m}}}^{\text{s}}$ for the short-term check-in behavior of tourists is: (20) $x_{t_{m}}^{s} = {\tilde{x}}_{t_{m}}^{s} + λ_{c}^{s} r_{c}^{u}$ \[x_{{{t}_{m}}}^{s}=\tilde{x}_{{{t}_{m}}}^{s}+\lambda _{c}^{s}r_{c}^{u}\]

where $λ_{c}^{s}$ $\lambda _{c}^{s}$ is the category preference of visitor u in the current check-in behavior, and $x_{t_{m}}^{s}$ $x_{{{t}_{m}}}^{s}$ is the final embedding vector representation of the hidden state.

2) Deep learning network using RNN as a short-term preference module

The embedding vector representation $x_{t_{m}}^{s}$ $x_{{{t}_{m}}}^{s}$ of t_m moments generated in the embedding layer is inputted into the RNN and, together with the hidden state $h_{t}_{m - 1}^{s}$ ${{h}_{t}}_{m-1}^{s}$ of t_m−1 moments in the RNN, undergoes a series of operations to obtain the hidden state $h_{t}_{m}^{s}$ ${{h}_{t}}_{m}^{s}$ of t_m moments as: (21) $h_{t_{m}}^{s} = R N N (x_{t_{m}}^{s}, h_{t_{m - 1}}^{s})$ \[h_{{{t}_{m}}}^{s}=RNN(x_{{{t}_{m}}}^{s},h_{{{t}_{m-1}}}^{s})\]

where RNN(·) is the set of formulas for the RNN algorithm as follows. Eventually, all hidden states $H^{s} = {h_{t_{1}}^{s}, \dots, h_{t_{m}}^{s}, \dots, h_{t_{S}}^{s}}$ ${{H}^{s}}=\{h_{{{t}_{1}}}^{s},\cdots ,h_{{{t}_{m}}}^{s},\cdots ,h_{{{t}_{S}}}^{s}\}$ of T = {t₁,…,t_m,…,t_S} can be obtained. Then: (22) $h_{m} = U x_{m} + W s_{m - 1}$ \[{{h}_{m}}=U{{x}_{m}}+W{{s}_{m-1}}\] (23) $s_{m} = σ (h_{m})$ \[{{s}_{m}}=\sigma ({{h}_{m}})\] (24) $o_{m} = s o f t m a x (V s_{m})$ \[{{o}_{m}}=softmax(V{{s}_{m}})\]

3) Accurately construct the embedding vector representation of tourists’ short-term preferences

Let H^s ∈ℝ^d×s be the matrix consisting of all hidden state vectors ${h_{t_{1}}^{s}, \dots, h_{t_{m}}^{s}, \dots, h_{t_{s}}^{s}}$ $\{h_{{{t}_{1}}}^{s},\cdots ,h_{{{t}_{m}}}^{s},\cdots ,h_{{{t}_{s}}}^{s}\}$ in the short-term preference module, where d is the dimension of the hidden state vectors and S is the length of the short-term sequence. The weight vector μ_s refers to the weight $μ_{m}^{s}$ $\mu _{m}^{s}$ measured by the weight corresponding to each hidden state vector $h_{t_{m}}^{s}$ $h_{{{t}_{m}}}^{s}$ in the next check-in for the m th time. Namely: (25) $μ_{m}^{s} = \frac{\exp (g (h_{t_{m}}^{s}, q_{u}^{s}))}{\sum_{i = 1}^{N} e x p (g (h_{t_{m}}^{s}, q_{u}^{s}))}$ \[\mu _{m}^{s}=\frac{\exp (g(h_{{{t}_{m}}}^{s},q_{u}^{s}))}{\sum\limits_{i=1}^{N}{e}xp(g(h_{{{t}_{m}}}^{s},q_{u}^{s}))}\] (26) $g (h_{t_{m}}^{s}, q_{u}^{s}) = \frac{h_{t_{m}}^{s} {(q_{u}^{s})}^{T}}{\sqrt{d}}$ \[g(h_{{{t}_{m}}}^{s},q_{u}^{s})=\frac{h_{{{t}_{m}}}^{s}{{(q_{u}^{s})}^{T}}}{\sqrt{d}}\]

where $q_{u}^{s}$ $q_{u}^{s}$ is the query information for the short-term check-in sequence in the temporal attention mechanism. The short-term preferences of tourists at the fine-grained POI level $r_{s}^{u}$ $r_{s}^{u}$ are: (27) $r_{s}^{u} = \sum_{m = 1}^{s} μ_{m}^{s} h_{t_{m}}^{s}$ \[r_{s}^{u}=\sum\limits_{m=1}^{s}{\mu _{m}^{s}}h_{{{t}_{m}}}^{s}\]

where m denotes the visitor’s m rd check-in at time t_m , and S is the total number of check-ins in the short-term check-in sequence.

3.2.3

Output modules

In this paper, the long-term preference of tourists obtained in the long-term preference module and the short-term preference of tourists obtained in the short-term preference module are weighted and fused to obtain the final tourists’ preference as: (28) $r^{u} = α r_{l}^{u} + β r_{s}^{u}$ \[{{r}^{u}}=\alpha r_{l}^{u}+\beta r_{s}^{u}\]

where α and β are the weights of the long-term and short-term preferences to be learned, respectively. The next POI recommendation probability after normalization by the softmax function is defined as: (29) $o_{t_{N + 1}, v_{k}}^{u} = \frac{e x p (r^{u} v_{k})}{\sum_{k = 1}^{P} e x p (r^{u} v_{k})}$ \[o_{{{t}_{N+1}},{{v}_{k}}}^{u}=\frac{exp({{r}^{u}}{{v}_{k}})}{\sum\limits_{k=1}^{P}{e}xp({{r}^{u}}{{v}_{k}})}\]

where v_k is an embedding vector representation of the candidate POIs v_k filtered by the POI filter, and P is the total number of candidate POIs obtained by the POI filter.

3.3

Model training

The loss function for the long-term preference module is given in Eq: (30) $l l = \sum_{(v > v^{'}) \in Ω_{l}} l n (1 + e^{- (o_{t}^{v} - o_{t}^{v^{'}})})$ $ll=\underset{(v>{{v}^{\prime }})\in {{\text{ }\!\!\Omega\!\!\text{ }}_{l}}}{\mathop{\sum }}\,ln\left( 1+{{e}^{-\left( o_{t}^{v}-o_{t}^{{{v}^{\prime }}} \right)}} \right)$

Where v^' is the negative sample of POI v in the long-term preference module, Ω_l is the training sample consisting of u,v,v^' in the long term preference module, $o_{t}^{v}$ $o_{t}^{v}$ is the predicted probability of a tourist u visiting POI v^' 8 at time t , and $o_{t}^{v}$ $o_{t}^{v}$ is the predicted probability of a tourist u visiting POI v^' at time t .

The loss function for the short-term preference module is as in Eq: (31) $l s = \sum_{(v_{s} > v_{s}^{'}) \in Ω_{s}} l n (1 + e^{- (o_{t}^{v_{s}} - o_{t}^{v_{s}^{'}})})$ $ls=\underset{\left( {{v}_{s}}>v_{s}^{'} \right)\in {{\text{ }\!\!\Omega\!\!\text{ }}_{s}}}{\mathop{\sum }}\,ln\left( 1+{{e}^{-\left( o_{t}^{{{v}_{s}}}-o_{t}^{v_{s}^{'}} \right)}} \right)$

Where $v_{s}^{'}$ $v_{s}^{\prime }$ is the negative sample of POI v_s in the short-term preference module, Ω_s is the training sample consisting of $u, v_{s}, v_{s}^{'}$ $u,{{v}_{s}},v_{\text{s}}^{\prime }$ in the short-term preference module, $o_{t}^{ν_{s}}$ $o_{t}^{{{\nu }_{s}}}$ is the predicted probability of tourist u visiting POI v_s at time t , and $o_{t}^{ν_{s}}$ $o_{t}^{{{\nu }_{s}}}$ is the predicted probability of tourist u visiting POI v^' at time t .

In this paper, the total loss function of the MALS model algorithm is designed by combining the loss functions and regularization terms of the category module, the long-term preference module, and the short-term preference module: (32) $l = l c + l l + l s + \frac{ε}{2} | | θ^{2} | |$ \[l=lc+ll+ls+\frac{\varepsilon }{2}||{{\theta }^{2}}||\]

where ε is the regularization factor and θ = (W^c ,b^c ,W ,b,W^s ,b^s ,μ) is the set of all parameters to be trained in the model.

4

Big Data-driven Tourist Behavior Trajectory Preference and Recommendation Performance Analysis

4.1

Analysis of the effect of intelligent recommendation service based on tourists’ behavioral trajectory

In this paper, we use the behavioral trajectory dynamics of tourists at tourist attractions in cities A and B in China from 2020-2024 as a dataset (referred to as A and B for short), which contains tourist attractions and their check-ins in cities A and B collected over 4 years. It contains 250 attractions in city A with 479 check-ins and 550 attractions in city B with 769 check-ins. Each check-in is associated with its timestamp, its GPS coordinates, and its semantics (represented by fine-grained place categories). In order to evaluate the performance of the next point-of-interest recommendation algorithm, this paper employs two evaluation methods widely used in recommender systems, including Recall and MAP. Recall measures the proportion of hits in the recommended top-K list of points of interest to the actual points of interest visited and represents how many points of interest visited by tourists are recommended by the model.

4.1.1

Recommendation Performance Evaluation of Intelligent Service Models

The intelligent recommendation service model based on the behavioral trajectory dynamics of tourists at tourist attractions proposed in this paper incorporates point-of-interest semantic information, spatial contextual information, temporal contextual information, and tourists’ long-term preferences. In order to study how these influencing factors affect the performance of the next point-of-interest recommendation, this paper removes one influencing factor at a time and then designs the corresponding six simplified MALS models to illustrate the effects of the removed influencing factors on the point-of-interest recommendation, respectively. I: Excluding the influence of visitors’ long-term preferences. II: Exclude the influence of tourists’ point-of-interest category shifting preference. III: exclude the effect of absolute moments of the week. IV: exclude the effect of weekend status. V: Exclude the effect of transition distance of point-of-interest-point-of-interest. VI: Exclude the effect of the transition time of point-of-interest-point-of-interest. Experiment with these six models with the MALS model on NYC and TKY datasets and use Recall and MAP as evaluation indexes to judge the effect of each model.

The results of the influence of different contextual factors on the effectiveness of intelligent service recommendations for tourists are shown in Figure 1. It is found that MALS achieves the best performance in datasets A and B when all the influencing factors are included, at which time both Recall and MAP correlation coefficients of the MALS model are the largest, which are 0.2753, 0.3837, 0.3475, and 0.4269, respectively. The performance will be degraded if any of the factors are deleted, which verifies that these factors selected in this paper affect the tourists’ behavioral trajectory to visit the point of interest decision. In addition, the performance drops drastically when visitor preference or category information is excluded (I and II), indicating that visitor preference and category information are critical for the next point of interest recommendation. Among the remaining influencing factors, different moments of the week are the third drastic decrease. It suggests that the demand of visitor behavioral trajectories to visit points of interest is related to different periods. When the transfer distance from the point of interest to the point of interest or the transfer time from the point of interest to the point of interest is excluded, Recall is decreasing sharply, while MAP is decreasing relatively less. One possible reason for this is that when tourists have the opportunity to select similar candidate tourist attractions, tourists tend to visit closer tourist attractions (points of interest), and thus, considering the distance and time interval factors will improve Recall. When tourists do not have to consider these two factors, similar tourist attractions have equal chances of being visited no matter how far away they are, leading to relatively smooth visiting behavior.

4.1.2

Effectiveness of Intelligent Service Model in Recommending Tourist Attractions for Tourists

In the intelligent service recommendation model for tourist behavioral trajectory dynamics, the point-of-interest category vectors represent the information of the previous point-of-interest and the next point-of-interest, and the longer the dimensions of the category vectors indicate that the category of the current point-of-interest is more dependent on the contextual category information. However, the behavioral trajectories of tourists not only have category transfer patterns but also have a strong relationship with the location and time factors of tourists as well as tourists’ preferences. Therefore, in this paper, we set the range of TOP-K for the number of interest points recommended by category dimension as (5, 10, 15, 20, 25, 30), train the models separately, and compare the performance changes of different models using Recall and MAP metrics. The results of intelligent service recommendation for point-of-interest categories are shown in Fig. 2. The results on both datasets show that Recall and MAP have the largest coefficients (0.3574 and 0.2763) when the TOP-K of recommended points of interest is 10 in dataset A. Recall, and MAP in dataset B also have the largest coefficients (0.4325 and 0.3321) when the TOP-K of recommended points of interest dimension is 10. The model performance measured by Recall and MAP metrics is optimal in both datasets. When the dimensionality is just increased, the model performance suddenly improves, indicating that the increase in dimensionality makes the model learn the dynamic category transfer pattern of tourists’ behavioral trajectories. However, as the number of recommended interest points TOP-K is greater than 10, the model performance gradually decreases with less fluctuation.

4.2

Case Study of Tourist Travel Preferences Based on Intelligent Services

This paper takes Xi’an Shaanxi as an example to analyze the behavioral trajectory of tourists and recommend their preferences for intelligent services based on the analysis results. As the ancient capital of the 13th Dynasty, Xi’an, Shaanxi Province has rich cultural and tourism resources that make it one of the most popular tourist cities. During the Spring Festival of 2022, the 18 key tourist attractions belonging to Xi’an City received a total of 2,097,600 tourists, and the average number of days of stay reached 3.99 days. It can be seen that the scale of Xi’an’s tourism market is huge. Moreover, Xi’an is rich in tourism resources, as it was once the ancient capital of the Han and Tang dynasties. Its history and culture are profound, and it is connected to the Qinling Mountains, and its natural landscape resources are varied. Therefore, this paper chooses tourists in Xi’an City, Shaanxi Province, as the research object.

4.2.1

Results of Visitor Profiling and its Analysis

This section analyzes the user portrait from four aspects: basic information about tourists, travel characteristics, cognitive level, and trajectory information. The information related to tourists’ portraits is described as shown in Table 1 below. Among them, the basic information about tourists includes gender, age, and education; the travel characteristics include fellow travelers and travel expenses. The cognitive degree includes the cognitive level and whether they are engaged in tourism-related occupations. The trajectory information includes the tourists’ stay attractions and their stay time. In summary, the four variables of basic information, travel characteristics, cognitive level, and trajectory information are used to portray tourists through second-order clustering.

Table 1.

Information description of tourist portraits

User portrait information	Attribute	Describe
Basic information	Gender	The classification variable, 1= "male, 2=" female"
	Age	Class variable, 1= under 18,2=Between 18 and 25,3=, Between 25 and 35,4=, Between 35 and 60,5= over 60 years old.
	Educational background	Grade variables,1= "junior high school and below;2= "high school and secondary school;3="Bachelor’s degree and junior college;4="Graduate student"
Travel characteristics	Pedestrian	The classification variable, 1= friend, 2= family, 3= parents,4= Couple,5= One person
Travel characteristics	Travel cost	Numerical variables, the average daily travel cost of tourists
Cognitive degree	Cognitive level	The numerical variable, 1= "very low," 2 = "low, 3=" general, "4=" high, "5=" very high"
Cognitive degree	Occupation	The classification variable, 0= "no,"1= "yes"
Trajectory information	Tourist attraction	Nominal variable, tourist attractions
Trajectory information	Playtime	The number of variables, the visitor’s stay in a scenic spot

In this paper, 100 tourists were randomly selected from 2,097,600 people to cluster their trajectories. Tourist clustering results of the analysis of statistical charts shown in Figure 3, tourists are roughly divided into three categories: Category 1: the main characteristics of this group of people for travel spending in general, most of the young adults, not engaged in tourism industry-related occupations, lower education, cognitive level of the middle to lower, and most of the family travel in pairs. This category accounts for 52% of the total number of people. Category 2: The main characteristics of this group are that they spend more on traveling, are mostly young adults, have a higher percentage of people working in tourism-related professions, have higher education and cognitive levels, and mostly travel with friends or alone. The proportion of this group in the total number of people is 21%. Category 3: The main characteristics of this group of people are that they spend less on traveling, are mostly young people between 18-25 years old, mostly travel with friends or couples, have an average level of cognition, and have a bachelor’s degree. This category accounts for 27% of the total number of people.

4.2.2

Results of tourists’ travel preference analysis

In order to better analyze the differences in tourism preferences of different groups of people, this section adopts the multivariate logistic regression model as the research model. The tourist category is selected as the dependent variable, and the eight dimensions of attraction characteristics theme “A humanistic history, B recreational activities, C tour guide service, D natural landscape, E affordability, F food and beverage, G architectural features, H scenic services” are used as independent variables. Among them, the category of tourists is the categorical variable, and the variables related to attraction characteristics are the numerical variables, which were analyzed using SPSS 25.0 statistical software. When the value of coefficient B is greater than 0, it means that the variable has a positive effect on this model. The results of its specific regression analysis are as follows:

1)

Category I tourists group tourism preference difference analysis

The regression results of analyzing the tourism preference of category I tourists are shown in Figure 4. There are differences between category 1 tourist groups and their tourism preferences in E affordable, F catering and food, and H scenic services, and their significance p-values are less than 0.05. Among them, the coefficient B of catering and food is 2.5015, which is greater than 0, indicating that category 1 tourists have an extreme preference for catering and food, but they have little preference for or interest in the other items (B-value is less than 0).

2)

Category II tourists group tourism preference difference analysis

The regression results of analyzing the tourism preference of category II tourists are shown in Figure 5. There is a significant difference between category II tourist groups and their tourism preferences in A humanistic history, D natural landscape, E affordability, G architectural features, and H scenic services (P < 0.05). Among the tourists in category II, except for the coefficient B values of E affordability and G architectural features, which are less than 0 (-4.5831 and -2.4643), the coefficient B values of the remaining six attraction feature subjects of A humanistic history, B recreational activities, C tour guide services, D natural landscapes, F catering and food, and H scenic area services, are respectively 0.6309, 1.4388, 1.5162, 1.5737 1.4486, and 1.084, all of which are greater than 0. By observing the coefficient B value, it is found that compared with category 1, the category 2 group prefers a strong humanistic history, rich recreational activities, diversified catering and food, and intelligent scenic services.

3)

Differences in tourism preferences of category three tourist groups

The regression results of category three tourists’ tourism preference analysis are shown in Figure 6. There is a difference between the preferences of category three groups and their preferences for the attraction characteristics of C tour guide service, D natural landscape, F catering food, and H scenic service, and their significance p-values are 0. In addition, the coefficients B-values of the category three groups for C tour guide service, D natural landscape, and F catering food are greater than 0. This indicates that the category three groups prefer the warm and hospitable tour guide service and picturesque natural landscape to the category one groups. This indicates that the Category III group prefers the hospitable tour guide service and picturesque natural landscape compared to the Category I group. By looking at the value of coefficient B, it is found that group III prefers tour guide service (B=5.5533) and natural landscape (B=7.5528) to group II and has little interest in human history (B=-0.4955) and recreational activities (-2.4981).

4.3

Tourist Attractions Management Intelligent Service Revelation

4.3.1

Emphasize the management thinking of “Internet + Tourism”

Accelerating the development of the tourism industry in the direction of digitization, networking, and intelligence will help break the data silos, enhance the service experience of the scenic spots, innovate the tourism public management service mode, further promote the development of online and off-line integration, and optimize the management mode of the scenic spots. Attention to the “Internet + tourism” management thinking, especially with the new era of the Internet of Things, big data, and other technologies to bring a variety of information data, fully explore the emotional tendencies of tourists for the scenic area, thus helping scenic area managers to understand the tourist experience, accurately grasp the preferences of tourists at the same time, to help scenic area managers clear scenic area management and operation, the existence of advantages and shortcomings, optimize the scenic area management and operation, and optimize the development and operation of the scenic area. Advantages and defects in the management and operation of the scenic area, optimize the scenic area services and create a better, safer, high-quality tourism experience for tourists.

4.3.2

Focus on diversification and refinement of tourism services

Different groups of tourists have different travel preferences. In order to enhance the brand image of the scenic area and improve the scenic area tourism market, it should adhere to the unity of standardization and personalization. Optimize the structure of scenic tourism products, innovative tourism product system for different groups of needs and preferences, the introduction of more customized tourism products, tourism routes, such as for travel budget for tourists customized recreational activities, and a rich, affordable class of tourism products. For tourists with rich travel experience and a high degree of cognitive customization of scenic areas with excellent service, scenic characteristics of significant tourism product packages. For some young college students can be customized for its food and beverage food, rich and diverse, and has significant characteristics of the tourism portfolio products.

5

Conclusion

In this paper, under the big data-driven background, clustering and point-of-interest recommendations of tourists’ behavioral trajectories are carried out from the long and short-term preferences of tourists’ behavioral trajectories, and a case study is carried out in Xi’an City as an example. The preferences of the three categories of tourists are analyzed differently, and finally, insights into the management of intelligent services in tourist attractions are obtained.

1) The recommendation performance of MALS is best when the dynamic dataset of behavioral trajectories of tourists of tourist attractions in cities A and B is not excluded from any influencing factors, at which time the recall and MAP of the MALS model have the largest correlation coefficients. Deleting any factor reduces the recommendation performance of the model, which proves that the factors selected in this paper affect the decision of the tourists’ behavioral trajectories to visit the points of interest. In addition, in both datasets A and B, the model performance measured by both Recall and MAP metrics is optimal when TOP-K is 10. When the TOP-K is greater than 10, the model’s recommendation performance shows a small decrease.

2) In this paper, tourists are profiled based on four variables: basic information, travel characteristics, cognitive level and trajectory information, and tourists are clustered into three categories. 52% of the total number of tourists in category I, whose main characteristics are teenagers with medium spending, lower-middle cognitive level, and lower education, and whose travel process is mostly family-based. The share of category 2 tourists is 21% of the total number of tourists, and this category is dominated by teenagers with a high level of education and awareness. Their main characteristics are that they spend a lot of money on their trips and travel with friends or alone. Category III tourists, the total number of 27%, 18-25 years old undergraduate students are mainly. The travel process is mostly with friends or couples, spending less.

3) Category 1 tourists have an extreme preference for dining and food, but their preference for other items (B-value less than 0) is minimal or of little interest. Category II tourists prefer a strong humanistic history, rich entertainment activities, diverse food and beverage, and intelligent scenic services. The Category III group prefers hospitable tour guide services and picturesque natural landscapes, and it has little interest in humanistic history (B=-0.4955) and recreational activities (-2.4981).

Funding:

This research was supported by the General Project of Statistical Science Research in Shanxi Province in 2023: Comprehensive Performance Evaluation of high-quality Tourism Development - A Case study of Shanxi Province (2023LY018).

Language:: English

Publication timeframe:: 1 times per year
Journal Subjects:: Life Sciences, Life Sciences, other, Mathematics, Applied Mathematics, General Mathematics, Physics, Physics, other

Journal RSS Feed

Big Data-Driven Dynamic Analysis of Tourist Behavioral Trajectories and Intelligent Service Strategies in Tourist Attractions

Guo Hu

Published Online: Mar 17, 2025

Received: Oct 23, 2024

Accepted: Jan 29, 2025

DOI: https://doi.org/10.2478/amns-2025-0291

KeywordsMean filtering technique, LSTM network, Vector embedding, Hierarchical attention mechanism, Tourist behavior trajectory

© 2025 Guo Hu, published by Sciendo

This work is licensed under the Creative Commons Attribution 4.0 International License.

Keywords
Mean filtering technique, LSTM network, Vector embedding, Hierarchical attention mechanism, Tourist behavior trajectory