Investigation of Passengers’ Perceived Transfer Distance in Urban Rail Transit Stations Using XGBoost and SHAP

Abstract

Providing high-quality public transport services and enhancing passenger experiences require efficient urban rail transit connectivity; however, passengers’ perceived transfer distance at urban rail transit stations may differ from the actual transfer distance, resulting in inconvenience and dissatisfaction. To address this issue, this study proposed a novel machine learning framework that measured the perceived transfer distance in urban rail transit stations and analyzed the significance of each influencing factor. The framework introduced the Ratio of Perceived Transfer Distance Deviation (R), which was evaluated using advanced XGBoost and SHAP models. To accurately evaluate R, the proposed framework considered 32 indexes related to passenger personal attributes, transfer facilities, and transfer environment. The study results indicated that the framework based on XGBoost and SHAP models can effectively measure the R of urban rail transit passengers. Key factors that affected R included the Rationality of Signs and Markings, Ratio of Escalators Length, Rationality of Traffic Organization outside The Station, Ratio of Stairs Length, and Degree of Congestion on Passageways. These findings can provide valuable theoretical references for designing transfer facilities and improving transfer service levels in urban rail transit stations.

1. Introduction

The growth of urban rail transit system in China has been remarkable since the turn of the 21st century. The Ministry of Transport of China reported that, as of December 2022, there are 290 urban rail transit lines in operation across mainland China, serving 1.16 billion passengers and covering a total distance of 9584 km. This represents a significant increase of 226.5 km from the previous month, demonstrating the rapid and continuous expansion and development of the urban rail transit system in China [1].

The rapid growth of urban rail transit has given rise to two critical challenges: (1) meeting passengers’ increasing demands for safety, speed, and comfort, and (2) maximizing the commercial potential of urban rail transportation. Addressing these challenges has become a popular research topic among governments and scholars, with many focusing on improving transfer facilities.

Transfer distance at an urban rail transit station is a crucial indicator of service level and passenger satisfaction, typically measured as the actual transfer distance; however, previous research has shown that not all transfer distances are equal. Yao [2] optimized the facility layout in urban rail transit stations to reduce the actual transfer distance. Olszewski [3] introduced the concept of Equivalent Walking Distance, which accounts for obstacles encountered during walking as penalties added to the actual walking distance. The impact of the transfer environment and facilities (such as the Number of Escalators [4], Intersection Density [5], and Facility Congestion [6]) on passenger experience has also been explored in previous research; however, the penalties for poor transfer are often set as fixed values that do not account for differences in the environment or passengers, thus oversimplifying the hindrances in the transfer process.

Perceived transfer distance has been shown to be a more reliable indicator of passengers’ transfer experience and satisfaction with the service compared to actual transfer distance or equivalent walking distance [7]. It refers to the distance between the pick-up and drop-off points of a transfer that passengers estimate based on their perception and experience, which can deviate significantly from the actual distance [7,8]. Furthermore, given individual differences and diverse travel experiences, residents within the same community may have varying perceptions of the actual transfer distance, leading to distinct preferences for urban rail transit travel.

Most people prefer a short transfer distance, and their perceived transfer distance can significantly influence their choice of transfer and even their decision to use urban rail transit [9]; however, a pedestrian-friendly environment can compensate for the inconvenience of a longer transfer distance [10]. For instance, safe intersections, good infrastructure, and attractive buildings can enable passengers to accept longer transfer distances [11,12].

Understanding how passengers estimate transfer distance is crucial for effective passenger flow management and urban rail transit station design and planning. Although research on perceived transfer distance in the field of urban rail transit is limited, findings from psychology, behavior, geography, and other fields on perceived walking distance could be used as references. Relevant literatures suggested that Personal Attributes, Route Characteristics, and Environmental Factors are among the main influencing factors.

Personal Attributes have been identified as influential factors in differentiating perceived distance. For instance, studies [13,14] have discovered that age affects distance estimation, with older travelers more likely to overestimate distance. Gender differences are also found in perceptions of length, with women tending to overestimate distance compared to men [15]. Additionally, travel frequency also have an impact on distance perception, although the effect is less pronounced [16]. Weight-bearing has also been shown to increase distance perception among travelers [17]. Moreover, emotions [18], attention [19], and familiarity with the space during travel [20] could affect judgments of distance.

Route Characteristics also play a crucial role in determining perceived transfer distance. Hess [8] found a positive correlation between the number of route intersections and the perceived distance in urban environments. Sadalla [21] found in a laboratory setting that routes with seven right-angle turns are perceived to be longer than routes with only two turns. Similarly, Kahl [22] suggested that children perceive segmented routes to be longer than continuous routes of the same length. Morgado [23] proposed that participants tend to overestimate distance when two locations are separated by obstacles, while segments without obstacles are relatively underestimated. Tenhundfeld [24] confirmed that perceived distance is longer on slopes, both uphill and downhill, compared to flat ground. The same conclusion is reached for estimating walking distance on stairs [25].

Environmental Factors affect perceived walking distance. For example, Sun [26] experimentally demonstrated that pedestrians are more likely to overestimate walking distances on routes without sheltered walkways. Similarly, Iosa’s study [27] pointed out that there are differences in distance estimation between indoors (single-level) and outdoors environments, with people more likely to underestimate distance in indoor environments. Winter [28] argued that vertical distance increases indoor distance estimation due to the increased complexity of indoor movement in multilevel environments.

The different methods used to assess perceived distance have included the Scaling Method, Regression Analysis [29], t-test, Logistic Regression [30], Graph Theory, and Routing Algorithm. Ankomah [31] used the Scaling Method and found that extroverted men who live in the community for a longer period of time have a more accurate perception of distance. Walmsley [32] demonstrated using Regression Analysis that tourists often overestimate the perceived distance compared to the actual distance. Zhou [33] identified several significant factors that affecte tourists’ perceived distance using independent sample t-test and single-factor analysis of variance, which includes personality, destination, presence of family and friends, purpose of travel, number of trips, and travel experience; however, the aforementioned methods may be less accurate and stable in comprehensively evaluating urban rail transit station transfer distance due to their insensitivity to non-linear relationships between indicators.

To address this issue, recent studies have proposed extreme gradient boosting (XGBoost) [34] and Shapley additive explanations (SHAP) models [35] as potential solutions. XGBoost is a scalable tree boosting system widely used by data scientists, which can provide state-of-the-art results for many problems. It is also a novel sparsity-aware algorithm for handling sparse data and a theoretically justified weighted quantile sketch for approximate learning. XGBoost has been widely used in various areas of transportation, such as traveler perception, traffic volume prediction, transportation safety, and risk analysis, and has achieved better results than traditional algorithms. Asadi-Shekari [36] introduced an algorithm based on XGBoost and analyzed the perceptions of vulnerable road users towards sharing the road with autonomous vehicles. The results showed that XGBoost is highly effective and accurate in predicting the views of cyclists or pedestrians. Liu [37] built a prediction model for bus–metro-transfer ridership based on XGBoost, and the results demonstrated that XGBoost outperforms linear models, random forests, and LightGBM in terms of fitting, accuracy, and stability. Yang [38] explored the relationships between built environment factors and the spatial distribution of freight truck-related crashes using XGBoost. The results showed that XGBoost has a strong ability to capture the nonlinear relationships between certain explanatory variables and the dependent variables. Alshboul [39] developed an ensemble machine learning technique (EMLT) that combines algorithms such as extreme gradient boosting (XGBoost), categorical boosting (CatBoost), k-nearest neighbor (kNN), light gradient boosting machine (LightGBM), artificial neural network (ANN), and decision tree (DT) to predict liquidated damages in highway construction projects. The results showed that, compared to other machine learning-based models, EMLT achieves the highest accuracy of 0.997; however, its application in the field of distance perception is still very limited, and this paper aims to bridge the gap.

Moreover, the use of a tree-based model, such as XGBoost, presents both benefits and drawbacks when compared to more easily interpretable models such as linear regression. Although linear regression models can be easily interpreted, they may not be able to capture the complex and non-linear relationships between predictors. Conversely, the XGBoost model has the ability to capture non-linear relationships and provides higher predictive accuracy, but its interpretability may be limited. In order to address this issue, our study employed SHAP values to explain the model’s predictions and obtained insights into the relative importance of different factors.

To sum up, this paper mainly focused on the perceived transfer distance of passengers, analyzing the influences of 32 different factors related to Passenger Personal Attributes, Transfer Facilities and Transfer Environment. Specifically, we utilized the XGBoost model to assess the perceived transfer distance of passengers, which was then used to evaluate the connection performance of urban rail transit stations. Meanwhile, we adopted the SHAP model to analyze the impact of various factors on passengers’ perceived transfer distance.

2. Index Quantification and Date Collection

2.1. Definition of the Ratio of Perceived Transfer Distance Deviation

In this paper, “transfer” refers to the movement process between the drop-off point of urban rail transit and the pick-up point of another modes of transportation. The Actual Transfer Distance ($D_{\mathrm{Actual}}$) is the actual mileage of the transfer process. The subjective interpretation of this distance by passengers is represented by the variable Perceived Transfer Distance ($D_{\mathrm{Perceive}}$).

To evaluate the transfer service level of urban rail transit stations and compare them regardless of actual transfer distance, this paper introduces the Ratio of Perceived Transfer Distance Deviation ($R$). The equation for calculating R is shown in Equation (1) below,

$$R=\frac{D_{\mathrm{Perceive}}-D_{\mathrm{Actual}}}{D_{\mathrm{Actual}}}\times 100\%$$

(1)

where $R$ is the Ratio Of Perceived Transfer Distance Deviation; $D_{\mathrm{Perceive}}$ is the Perceived Transfer Distance, a scalar measured in meters; and $D_{\mathrm{Actual}}$ is the Actual Transfer Distance, a scalar measured in meters.

A value of $R>0$ means passengers’ perceived transfer distance is longer than the actual distance, indicating low acceptability of the transfer process due to poor actual transfer conditions. The larger the R value, the worse the passenger experience. Conversely, when $R<0$, the $D_{\mathrm{Actual}}$ is shorter than what passengers anticipate based on daily experiences, indicating a high level of acceptability of the transfer process and better passenger experiences. At last, when $R=0$, the $D_{\mathrm{Actual}}$ fully matches what passengers anticipate, indicating that the actual transfer conditions are similar to the average level of passenger cognition.

2.2. Analysis of Influencing Factors

This study examined the perceptual mechanisms that influence transfer distance in the urban rail transit system from three aspects: Passenger Personal Attributes, Transfer Facilities, and Transfer Environment. Figure 1 summarizes all possible influencing factors, and

Table 1. Codes and descriptions of influencing factors.

Categories	Factor	Code	Values and Descriptions
Passenger Personal Attributes	Gender	X1	1 = Male; 2 = Female
	Age	X2	1 = 16–25; 2 = 26–45; 3 = 46–60; 4 = Over 61
	Education	X3	1 = Junior High School or Below; 2 = Technical Secondary School/High School; 3 = College/Undergraduate; 4 = Postgraduate or Above
	Occupation	X4	1 = Civil Servants; 2 = Public Institution Employees; 3 = Enterprise Employee; 4 = Freelancers; 5 = Unemployed; 6 = Retirees; 7 = Students; 8 = Others
	Monthly Income	X5	1 = Less Than 5000 Yuan; 2 = 5000–10,000 Yuan; 3 = 10,000–20,000 Yuan; 4 = More Than 20,000 Yuan
	Health Status	X6	1 = Very Good; 2 = Good; 3 = Poor; 4 = Very Poor
	Fatigue Degree	X7	1 = Very Relaxed; 2 = Relaxed; 3 = A Little Tired; 4 = Very Tired
	Common Travel Modes	X8	1 = Walk; 2 = Bicycle/Electric Bicycle; 3 = Bus; 4 = Rail Transit; 5 = Taxi/Ridesharing; 6 = Private Car
	Daily Walking Steps	X9	1 = Under 2000; 2 = 2000–5000; 3 = 5000–10,000; 4 = 10,000–15,000; 5 = Over 15,000
	Travel Time	X10	1 = Weekday Peak Hours; 2 = Weekday Off-peak Hours; 3 = Holiday Peak Hours; 4 = Holiday Off-peak Hours
	Travel Purpose	X11	1 = On or Off Duty; 2 = Go to School or Leave School; 3 = Visit Relatives and Friends; 4 = Travel; 5 = Entertainment and Shopping; 6 = Official Business; 7 = Others
	Familiarity of Transfer Routes	X12	1 = Very Familiar; 2 = Familiar; 3 = Unfamiliar; 4 = Very Unfamiliar
	Weight-bearing	X13	1 = No Load; 2 = Light; 3 = Heavy; 4 = Very Heavy
	Degree of Urgency	X14	1 = Not Urgent; 2 = Urgent; 3 = Very Urgent
	Companion or Not	X15	1 = With Partner(s); 2 = No Partner
Transfer Facilities	Ratio of Stairs Length	X16	Stair Length/Actual Transfer Distance, Unit: %
	Ratio of Escalators Length	X17	Escalators Length/Actual Transfer Distance, Unit: %
	Ratio of Elevators Operating Height	X18	Elevators Operating Height/Actual Transfer Distance, Unit: %
	Ratio of Passageway Length	X19	Passageway Length/Actual Transfer Distance, Unit: %
	Ratio of Pedestrian Crossing Facilities Length	X20	Length of Pedestrian Crossing Facilities/Actual Transfer Distance, Unit: %
	Number of Turns	X21	Number of Turns in the Transfer Route
	Rationality of Signs and Markings	X22	1 = Very Useful; 2 = Useful; 3 = Useless; 4 = Very Useless
Transfer Environment	Rationality of Traffic Organization inside the Station	X23	1 = Very Good; 2 = Good; 3 = Poor; 4 = Very Poor
	Rationality of Traffic Organization outside the Station	X24	1 = Very Good; 2 = Good; 3 = Poor; 4 = Very Poor
	Comfort of Nighttime Lighting outside the Station	X25	1 = Very Good; 2 = Good; 3 = Poor; 4 = Very Poor
	Lighting Comfort inside the Station	X26	1 = Very Good; 2 = Good; 3 = Poor; 4 = Very Poor
	Air Comfort inside the Station	X27	1 = Very Good; 2 = Good; 3 = Poor; 4 = Very Poor
	Temperature Comfort inside the Station	X28	1 = Very Good; 2 = Good; 3 = Poor; 4 = Very Poor
	Degree of Congestion on Platform	X29	1 = Very Comfortable; 2 = Comfortable; 3 = Congested; 4 = Very Congested
	Degree of Congestion on Stairs	X30	1 = Very Comfortable; 2 = Comfortable; 3 = Congested; 4 = Very Congested
	Degree of Congestion on Passageway	X31	1 = Very Comfortable; 2 = Comfortable; 3 = Congested; 4 = Very Congested
	Degree of Congestion on Entrance	X32	1 = Very Comfortable; 2 = Comfortable; 3 = Congested; 4 = Very Congested

describes the factors in detail.

(1)

Passenger Personal Attributes

Two categories make up Passenger Personal Attributes: individual characteristics and travel characteristics. Individual characteristics include seven factors: Gender, Age, Education, Occupation, Monthly Income, Health Status, and Fatigue Level. Travel characteristics include eight factors: Common Travel Mode, Daily Walking Steps, Travel Time, Travel Purpose, Familiarity of Transfer Routes, Weight-bearing, Degree of Urgency, and Companion or Not.

(2)

Transfer Facilities

Transfer Facilities consist of two categories: facility scale and other features. Facilities scale include six factors: Ratio of Stairs Length, Ratio of Escalators Length, Ratio of Elevators Operating Height, Ratio of Passageway Length, Ratio of Pedestrian, and Crossing Facilities Length. Other features include Number of Turns and Rationality of Signs and Markings.

(3)

Transfer Environment

Transfer Environment is composed of two categories: safety and comfort. Safety includes three factors: Rationality of Traffic Organization outside the Station, Rationality of Traffic Organization inside the Station, and Comfort of Nighttime Lighting outside the Station. Comfort has seven factors: Lighting Comfort inside the Station, Air Comfort inside the Station, Thermal Comfort inside the Station, Degree of Congestion on Platform, Degree of Congestion on Stairs, Degree of Congestion on Passageway, and Degree of Congestion on Entrance.

2.3. Data Collection

2.3.1. Principles of Survey

When conducting a transfer data survey, it is essential to meet the following requirements to ensure accurate and comprehensive data collection:

(1) Site selection should be diverse and representative, with an even distribution of site types covering all the indicators listed in Table 1 for transfer facilities and environment;

(2) Survey time should cover peak and off-peak hours as well as daytime and nighttime on both weekdays and holidays;

(3) The sampling should target passengers who have just completed an urban rail transit transfer and should be drawn randomly to ensure that the sample distribution closely resembles the population distribution;

At last, the sample size of the survey should be determined by Equation (2) below, which specifies the required minimum sample size, denoted as $n$, for a survey with a population of sufficient or infinite size,

$$n⩾{\left(\frac{k}{\alpha }\right)}^{2}P\left(1-P\right)$$

(2)

where $\alpha$ is the confidence level, $k$ is the quantile of a normal distribution, and $P$ is the population proportion. In the behavioral and social sciences, the significance level α is typically set at 0.05, the corresponding value of k is 1.96, and P is conventionally set at 0.5.

Based on Equation (2), the minimum sample size for conducting an efficacious transfer survey should exceed 384.

2.3.2. Survey Methods

Perceptions are difficult to quantify and remember due to their intangible nature. To overcome this challenge, this study adopted the “Experience-Response” based passenger transfer perception survey method [30] to obtain perception data. This method involves conducting a questionnaire survey on passengers who have just completed an urban rail transit transfer. The passengers were not informed of the survey information beforehand, ensuring the timely collection of perception data and reducing the psychological impact of the survey on passengers’ travel experiences.

The survey was conducted at urban rail transit platforms and various transfer points such as bus platforms, non-motorized parking points, and motorized parking points, among others.

2.3.3. Survey Steps

To collect data on perceived and actual transfers of urban rail transit passengers, this study follows six investigation steps:

Step 1: conduct a preliminary investigation of urban rail transit stations to determine their operating hours, peak passenger flow periods, distribution of transfer points, and types of transfer facilities.

Step 2: determine the survey sites, locations, and times based on survey principles.

Step 3: design a questionnaire based on survey indicators, characteristics of the survey object, survey principles, and other requirements. Please refer to Appendix A for a sample questionnaire.

Step 4: survey passengers’ transfer perceptions, who have recently completed an urban rail transit transfer, at survey locations,. The collected data should include passengers’ actual transfer routes, x1–x15, x22–x32, and their perceived transfer distance.

Step 5: take measurements and annotations on the station floor plan to obtain x16-x21 and the actual transfer distance based on the actual transfer routes of passengers. Validate the accuracy of the collected data by performing on-site observations.

Step 6: used Equation (1) to calculate the Ratio of Perceived Transfer Distance Deviation (R) by inserting the collected data of the Perceived Transfer Distance ($D_{\mathrm{Perceive}}$) and the Actual Transfer Distance ($D_{\mathrm{Actual}}$).

3. The Perceived Transfer Distance Model Based on XGBoost-SHAP

3.1. Extracting Key Factors

As multiple factors can influence $R$, factor screening is necessary to reduce the dimensionality and complexity of the data and to improve the accuracy and interpretability of the model. To exclude factors that were not related to R (p < 0.01), the study tested the Pearson correlation between R and the 32 influencing factors listed in Table 1. The m factors that remained after this process were deemed to be strongly correlated with $R$.

Below is the dataset $D$ that consists of n survey samples and m explanatory factors,

$$D=\left\{\left(x_1,y_1\right),\left(x_2,y_2\right),\dots ,\left(x_n,y_n\right)\right\}; x_i=\left(x_i{}^{\left(1\right)},\dots ,x_i{}^{\left(m\right)}\right)$$

where $x_i$ is the i-th sample and $i=1,2,\dots ,n$. $x_i{}^{\left(k\right)}$ is the k-th key factor of i-th sample (1 < $k$ < m), n is the number of samples, and $y_i$ is the R value of the i-th sample.

3.2. The R Prediction Model Based on XGBoost

Due to the non-linear effects among the influencing factors of R, the study utilized the XGBoost regression algorithm, which can effectively handle the heterogeneity between categorical and continuous variables.

XGBoost [34] is a popular machine learning algorithm used for regression and classification problems. It is based on the gradient boosting decision tree algorithm and uses a combination of different decision trees to make predictions. XGBoost has gained popularity due to its ability to handle complex non-linear relationships between variables as well as handle missing data and fast computational speed. It has been widely used in various research and applications, including data mining, natural language processing, and computer vision.

The process for building the $R$ prediction model based on XGBoost is as follows:

Step 1: split the dataset D into a training set and a test set using a ratio of 7:3 due to the small sample size. Use the training set as input for the XGBoost model.

Step 2: initialize the predicted value $f_0\left(x_i\right)$ for each sample; in this study $f_0\left(x_i\right)=0$.

Step 3: establish the objective function Obj and simplify it according to Equations (3) and (4) [34],

$$Obj={\displaystyle \sum }_{i=1}^{n}L\left(y_i,{\widehat{y}}_i\right)+{\displaystyle \sum }_{k=1}^K\Omega \left(f_k\right)=\frac{1}{2n}{\displaystyle \sum }_{i=1}^{\mathrm{n}}{\left({\widehat{y}}_i-y_i\right)}^2+\gamma T+\frac{1}{2}\lambda {\displaystyle \sum }_{j=1}^T{w}_j^2$$

(3)

where ${\widehat{y}}_i$ is the predicted value of $R$ of the i-th sample, $y_i$ is the target value of $R$ of the i-th sample, $L\left(y_i,{\widehat{y}}_i\right)$ is the loss function which measures the difference between ${\widehat{y}}_i$ and $y_i$ and is usually $\frac{1}{2}{\left({\widehat{y}}_i-y_i\right)}^2$, $K$ is the number of decision trees, $\Omega \left(f_k\right)$ is the regularization term that penalizes complex models, $\gamma$ is the complexity coefficient of leaf nodes, $T$ is the number of the leaf node, $\lambda$ is the penalty coefficient, and $w_j$ is the weight of the $j$-th key factor.

To transform the objective function to the Euclidean domain, XGBoost employs Taylor approximation, as shown in Equation (4),

$$Obj=-\frac{1}{2}{\displaystyle \sum }_{j=1}^T\frac{{\left({\sum }_{i\in I_j}{g}_i\right)}^2}{{\sum }_{i\in I_j}{h}_j+\lambda }+\gamma T=-\frac{1}{2}{\displaystyle \sum }_{j=1}^T\frac{{\left[{\sum }_{i\in I_j}\left({\widehat{y}}_i-y_i\right)\right]}^2}{T+\lambda }+\gamma T$$

(4)

where $g_i=-\left({\widehat{y}}_i-y_i\right)$, which represents the gradient of the loss function; $h_i=1$, which represents the Hessian of the loss function; and $I_j$ is the sample set in the leaf node.

Step 4: build decision trees continuously and fit the residuals.

Step 5: accumulate the predicted value of each decision tree, which represents $R$, as shown in Equation (5),

$${\widehat{y}}_i={\displaystyle \sum }_{k=1}^K{f}_k\left(x_i\right)$$

(5)

where $f_k$ is the regression equation corresponding to the $k$-th decision tree.

Step 6: feed the test set into the model and evaluate the model’s performance using various evaluation metrics. This study used two common metrics to measure the accuracy and precision of regression models: RMSE (root mean squared error), and MAE (mean absolute error). Smaller values of these metrics indicate better model performance.

3.3. Analysis of Key Factors Based on SHAP

The XGBoost model has highly accurate in prediction, but it cannot provide an explanation for the impact of factors. This study aims not only to predict $R$ but also to identify the influencing factors that affected it, providing technical support for guiding the planning and design of transfer facilities at urban rail transit stations.

There are various methods to rank the significance of different factors, such as maximum information coefficient, splitting times of tree nodes, Kullback–Leibler divergence, and SHAP [35]. Among these methods, the SHAP framework is preferred because it can consistently present the facilitative or inhibitory effects of each influencing factor on predicted variables, and it can be used for both local and global interpretations [35]. The study analyzed the relationship between R and each influencing factor by calculating the Shapley value (${\varphi }_j$), as shown in Equation (6) [34],

$${\varphi }_j={\displaystyle \sum }_{S\subseteq F\backslash \left\{j\right\}}\frac{\left|S\right|!\left(m-\left|S\right|-1\right)!}{m!}\left[f_x\left(S\cup \left\{j\right\}\right)-f_x\left(S\right)\right]$$

(6)

where ${\varphi }_j$ is the Shapley value of the $j$-th influence factor, which reflects the marginal contribution of the j-th factor to the predicted value ${\widehat{y}}_i$ of $R$; $F$ is the set of $m$ model input factors after extracting key factors; $S$ is the subset of $F$ obtained by removing the $j$-th factor, whose dimension is $\left|S\right|$; $f_x$ is the model of R; $f_x\left(S\cup \left\{j\right\}\right)$ is the predicted value of $R$ with the set $S\cup \left\{j\right\}$ as the input variable; and $f_x\left(S\right)$ is the predicted value of $R$ with the set S as the input variable.

The Shapley value measures the impact of the j-th influence factor on the model’s output, accounting for its interaction with other features in the model. The sign of the Shapley value indicates whether the influence factor has a positive or negative effect on the prediction. A Shapley value > 0 means that the factor increases the predicted value, while a Shapley value < 0 means that the factor decreases the predicted value.

The SHAP framework for calculating the significance of the j-th factor involves the following steps:

Step 1: determine the set $S$.

Step 2: use $S$ and $S\cup \left\{j\right\}$ as input variables in the R prediction model based on XGBoost to obtain the predicted values $f_x\left(S\right)$ and $f_x\left(S\cup \left\{j\right\}\right), \mathrm{respectively}$.

Step 3: substitute $f_x\left(S\right)$ and $f_x\left(S\cup \left\{j\right\}\right)$ into Equation (6) to calculate ${\varphi }_j$ for the j-th factor.

4. Case Study

4.1. Case Overview

This study used the Jin-Yi-Dong Line of the Jinhua Rail Transit System in Zhejiang Province, China, as a case study. This line spans 58.4 km and includes 17 stations.

Following the survey principles outlined in Section 2.3.1, this study conducted an “Experience-Response” questionnaire survey on 39 transfer points in 9 stations along the Jin-Yi-Dong Line between June 6 and 19 June 2022. A total of 1017 respondents completed the survey, and after plausibility checks, 981 responses remained, resulting in an efficiency rate of 96.46%.

After organizing and processing the raw data, this study obtained a total of 33 variables, which were divided into two tables for ease of presentation.

Table 2. Discrete variables descriptive statistics.

Code	Classes: Responses
	1	2	3	4	5	6	7	8
X1	43.83%	56.17%
X2	36.19%	50.76%	11.01%	2.04%
X3	12.64%	39.45%	42.81%	5.10%
X4	14.47%	4.79%	20.90%	27.42%	20.28%	2.45%	2.55%	7.14%
X5	31.50%	50.86%	13.05%	4.59%
X6	55.15%	40.26%	4.28%	0.31%
X7	27.83%	57.49%	12.44%	2.24%
X8	13.05%	31.60%	5.91%	11.42%	22.32%	15.70%
X9	12.64%	35.27%	37.92%	11.42%	2.75%
X10	50.36%	21.20%	21.20%	7.24%
X11	45.16%	4.99%	10.09%	1.84%	12.54%	12.64%	12.74%
X12	32.93%	42.81%	16.21%	8.05%
X13	38.63%	47.20%	11.32%	2.85%
X14	78.18%	19.27%	2.55%
X15	29.05%	70.95%
X22	50.77%	37.82%	9.07%	2.34%
X23	55.66%	34.86%	9.28%	0.20%
X24	58.92%	25.99%	9.28%	5.81%
X25	81.96%	11.62%	4.08%	2.34%
X26	68.81%	27.12%	2.34%	1.73%
X27	81.65%	15.39%	2.14%	0.82%
X28	80.54%	15.49%	2.85%	1.12%
X29	49.24%	35.37%	12.33%	3.06%
X30	43.53%	35.47%	16.41%	4.59%
X31	54.84%	34.66%	9.07%	1.43%
X32	45.26%	39.86%	12.64%	2.24%

contains the discrete variables, while

Table 3. Continuous variable descriptive statistics.

	Unit	Mean	St. Dev.	Min	Median	Max
X16	%	4.26	6.73	0	0	21
X17	%	11.96	7.02	0	13	26
X18	%	0.38	1.16	0	0	6
X19	%	17.08	12.05	0	19	60
X20	%	0.61	1.96	0	0	10
X21	-	6.72	1.62	4	7	11
$R$	%	10.73	21.66	−56.83	11.73	84.74

contains the continuous variables.

4.2. Extracting Key Factors

This study utilized the Pearson correlation coefficient to examine the relationships between $R$ and 32 influencing factors.

Table 4. The result of Pearson correlation coefficient.

Passenger Personal Attributes	r	Connection Facilities	r	Connection Environment	r
X1	0.204 *	X16	0.394 *	X23	0.407 *
X2	0.164 *	X17	−0.482 *	X24	0.372 *
X3	−0.042	X18	0.024	X25	0.133 *
X4	0.057	X19	0.143 *	X26	0.006
X5	0.080	X20	0.093 *	X27	−0.037
X6	0.266 *	X21	0.115 *	X28	−0.014
X7	0.335 *	X22	0.494 *	X29	0.303 *
X8	−0.033			X30	0.318 *
X9	−0.037			X31	0.384 *
X10	0.145 *			X32	0.248 *
X11	0.071
X12	0.254 *
X13	0.248 *
X14	0.080
X15	−0.044

¹ * indicates a significant correlation at the 0.01 level. ² “r” represents the correlation coefficient, which ranges from −1 to 1. The closer the value is to 1 or −1, the stronger the linear relationship between the two variables.

presents the results. Among the 32 factors, 12 factors including Education (X3), Occupation (X4), Monthly Income (X5), Common Travel Mode (X8), Daily Walking Steps (X9), Travel Purpose (X11), Degree of Urgency (X14), Companion or Not (X15), Ratio of Elevators Operating Height (X18), Lighting Comfort inside the Station (X26), Air Comfort inside the Station (X27), and Temperature Comfort inside the Station (X28) were not significantly related to $R$. Therefore, after excluding the 12 factors this study identified the remaining 20 factors as key variables affecting $R$.

4.3. Predicting Effect of XGBoost Model

In order to verify the effectiveness of the XGBoost model in predicting R, this study compared its performance with three other commonly used algorithms: linear regression, random forest, and LightGBM. Moreover, this study selected two different variable input schemes to verify the need for using Pearson correlation coefficient to extract key factors. The first scheme included all factors as model inputs, while the second scheme used only key factors for prediction.

To reduce the impact of randomness during the data splitting, this study evaluated the model’s performance by averaging the results of 10 times 5-fold cross validation. In

Table 5. Evaluation results with all features as model inputs.

Evaluation Metrics		Linear Regression	Random Forest	XGBoost	LightGBM
RMSE	Training set	0.085	0.054	0.034	0.072
	Test set	0.095	0.091	0.058	0.082
MAE	Training set	0.068	0.042	0.027	0.057
	Test set	0.076	0.069	0.045	0.065
Parameter description		/	n_estimators = 545 min_samples_leaf = 3 min_samples_split = 7 max_features = 15 max_depth = 12	n_estimators = 391 eta = 0.03 min_child_weight = 16 colsample_bytree = 0.8 max_depth = 5 subsample = 0.9 reg_lambda = 15	n_estimators = 495 learning_rate = 0.07 min_child_samples = 25 max_depth = 9 num_leaves = 28 subsample = 0.3 colsample_bytree = 0.3 reg_alpha = 2 reg_lambda = 3

, the evaluation results of the model using all factors is presented, while

Table 6. Evaluation results of key features as input to the model.

Evaluation Metrics		Linear Regression	Random Forest	XGBoost	LightGBM
RMSE	Training set	0.089	0.038	0.021	0.051
	Test set	0.095	0.085	0.057	0.066
MAE	Training set	0.071	0.030	0.016	0.041
	Test set	0.076	0.065	0.042	0.053
Parameter description		/	n_estimators = 433 min_samples_leaf = 1 min_samples_split = 4 max_features = 13 max_depth = 13	n_estimators = 1189 eta = 0.02 min_child_weight = 14 colsample_bytree = 0.9 max_depth = 19 subsample = 0.6 reg_lambda = 5	n_estimators = 1344 learning_rate = 0.03 min_child_samples = 20 max_depth = 10 num_leaves = 25 subsample = 0.7 colsample_bytree = 0.8 reg_alpha = 1 reg_lambda = 5

shows the model’s predictive performance solely based on key factors. It is important to note that Figure 2 only displays the performance of each model on the test set.

Upon comparing Table 5 and Table 6, the results showed that after extracting key factors, the random forest, XGBoost, and LightGBM models exhibited a decrease in errors, while the performance of the linear regression model remains relatively poor. This study believes that this is mainly because, after variable selection, tree-based models such as random forest, XGBoost, and LightGBM can better capture nonlinear relationships between variables, reduce model complexity, and minimize noise interference, thereby significantly improving model performance. However, due to the linear assumptions and limited noise factors removal effect of the linear regression model, its performance remains largely unchanged after variable selection.

Moreover, after comparing the performance of all models on the training and test sets, the XGBoost model demonstrated a significant advantage. The study found that the XGBoost model generated lower root mean square error (RMSE) and mean absolute error (MAE) than the other three models. Particularly, the XGBoost model showed a reduction of 40.00%, 32.94%, and 13.64% in RMSE, and 44.74%, 35.38%, and 20.75% in MAE, respectively, when compared with PSO-XGBoost, XGBoost, and DT, the linear regression model, random forest, and LightGBM.

Furthermore, as indicated in Figure 2, the comparison of the actual and predicted R of the four models on the testing set demonstrated that the XGBoost model’s points were positioned in close proximity to the perfect fitting line, indicating a significant alignment between predicted and real values. Conversely, the remaining three models displayed larger errors and greater discreteness, suggesting that the XGBoost model is the most suitable for predicting R and providing superior application results.

4.4. Analysis of the Ranking of Key Factors’ Significance

Figure 3 illustrates the SHAP values of all 20 key factors computed using the trained XGBoost model accompanied by a scatter plot that shows the relationship between the key factors and R. To assess the significance of the key variables on R, this study utilized the degree of significance, which is calculated by averaging the absolute value of SHAP, as presented in Figure 4.

The findings reveal that the Transfer Environment has the most significant impact on R, followed by Transfer Facilities, while the impact of Passenger Personal Attributes is relatively low. Thus, optimizing the Transfer Facilities and Transfer Environment can effectively decrease R for urban rail transit passengers; however, it is also essential to take into account the influence of Passenger Personal Attributes on the perceived transfer distance.

Figure 4 provides valuable insight into the key factors affecting R, with Rationality of Signs and Markings (X22) being identified as the most influential feature. It has the highest average SHAP value, indicating its strongest explanatory power for R overall. Other variables such as Ratio of Escalators Length (X17), Rationality of Traffic Organization outside the Station (X24), Ratio of Stairs Length (X16), and Degree of Congestion on Passageway (X31) also play a crucial role in transfer distance perception, albeit with smaller contributions than X22. Furthermore, Ratio of Passageway Length (X19) is another significant characteristic among the Transfer Facilities variables, which has a critical effect on transfer distance perception.

Additionally, the study identifies several other variables that significantly influence transfer distance perception, including Gender (X1), Rationality of Traffic Organization inside the Station (X23), Degree of Congestion on Stairs (X30), Fatigue Degree (X7), Weight-bearing (X13), Familiarity of Transfer Routes (X12), Ratio of Pedestrian Crossing Facilities Length (X20), Age (X2), and Health Status (X6). In contrast, variables such as Number of Turns (X21), Degree of Congestion on Entrance (X32), Travel Time (X10), Comfort of Nighttime Lighting outside the Station (X25), and Degree of Congestion on Platform (X29) are found to have little effect on transfer distance perception, as their degree of significance is less than 0.01.

This study emphasizes the examination of five key factors impacting R, including Rationality of Signs and Markings (X22), Ratio of Escalators Length (X17), Rationality of Traffic Organization outside the Station (X24), Ratio of Stairs Length (X16), and Degree of Congestion on Passageway (X31).

The Rationality of Signs and Markings (X22) is the most critical factor affecting R with a positive correlation. In other words, improved signs and markings help passengers easily identify directions and locations, reducing their perception of transfer distance. In contrast, poor signage and markings increase R.

In line with everyday experience, the Ratio of Escalators Length (X17) is negatively correlated with passengers’ R. Longer escalators require less physical effort, reducing passengers’ perception of transfer distance.

The Rationality of Traffic Organization outside the Station (X24) plays a crucial role in R. Chaotic organization creates obstacles, such as crowded areas, road crossings, and interweaving passenger and vehicle flows, directly affecting passengers’ perception of transfer distance. Additionally, extreme disorder can pose safety hazards, such as traffic congestion and accidents, further influencing passengers’ perception.

An inverse relationship exists between the Ratio of Stairs Length and R. Longer stairs result in more time spent walking, increased fatigue, and visual illusions, causing passengers to overestimate transfer distance. Therefore, replacing stairs with escalators in urban rail transit stations is recommended to reduce this overestimation.

Lastly, the Degree of Congestion on Passageway (X31) has a significant impact on R. Crowded conditions increase difficulty and travel time, causing anxiety and stress and leading to a perception of a longer distance. Transitioning from spacious areas such as platforms or halls to narrower passageways exposes passengers to uncertain factors, such as crowds and obstacles, requiring frequent changes in steps, reducing comfort during transfers, and intensifying visual information, leading to overestimation of the actual transfer distance. Urban rail transit station operators should closely monitor and guide passageway congestion to ensure passenger safety and travel experience.

Table 7. Discussion on the significance ranking of key factors.

Categories	Factor	Code	Rank	Discussion
Passenger Personal Attributes	Gender	X1	7	Men and women exhibit significant differences in cognition, perception, and physical strength. Women tend to be more influenced by the environment and facilities, while men usually underestimate distances and women overestimate them.
	Age	X2	14	Elder passengers experience declining physical functions with age, but most adapt to these changes, adjusting their psychological perceptions accordingly. As a result, age has minimal impact on perceived distance.
	Health Status	X6	15	Passengers with poor health conditions often experience reduced physical and cognitive abilities, requiring more time and energy for transfers. While those with long-term health issues might gradually adapt to their limitations and adjust their perceptual abilities, short-term health problems, such as temporary injuries, can significantly affect passengers’ ability to perceive distances.
	Fatigue Degree	X7	10	Fatigue can also influence passengers’ walking speed and perceptual accuracy, leading to overestimation of perceived distance.
	Travel Time	X10	18	Passengers have different priorities on weekdays and holidays, which can influence their perceived distances. On weekdays, passengers typically prioritize travel efficiency, whereas on holidays, they place greater emphasis on the overall experience.
	Familiarity of Transfer Routes	X12	12	Unfamiliarity with the transfer route requires more cognitive resources, affecting individual perception and decision-making ability.
	Weight-bearing	X13	11	The degree of load bearing changes the physical and psychological state of passengers. With increased load bearing, passengers are more likely to feel tired, and the perceived distance increases. The greater the load bearing is, the larger the passengers’ perceptual deviation of the transfer distance will be.
Transfer Facilities	Ratio of Stairs Length	X16	4	The length of walking staircases affects passengers’ physical strength and comfort, which increases their perceived distance.
	Ratio of Escalators Length	X17	2	Shorter walking staircases and longer escalators reduce physical effort, resulting in shorter perceived distances.
	Ratio of Passageway Length	X19	6	Enclosed environments such as passageways can create the illusion of elongated time and space, increasing perceived distance.
	Ratio of Pedestrian Crossing Facilities Length	X20	13	Long crossing facilities disrupt the transfer process, leading to increased walking time and fatigue, and reduced passenger satisfaction.
	Number of Turns	X21	16	Although the number of turns affects distance cognition [21], its impact on perceived distance is relatively insignificant compared to other factors.
	Rationality of Signs and Markings	X22	1	Clear and accurate signs and markings help passengers identify directions and locations, shortening perceived distances.
Transfer Environment	Rationality of Traffic Organization inside the Station	X23	8	A reasonable layout of facilities and streamlined settings within stations can reduce conflicts in passenger flow, transfer pressure, and anxiety.
	Rationality of Traffic Organization outside the Station	X24	3	An organized environment outside the station can provide a positive experience and shorten perceived distance, while congestion, passenger flow conflicts, and detours can increase perceived distance.
	Comfort of Nighttime Lighting outside the Station	X25	19	Poor lighting affects passengers’ ability to observe their surroundings, increasing insecurity and anxiety, and impacting their perception of transfer distance.
	Degree of Congestion on Platform	X29	20	Crowded platforms, stairs, and passageways can make passengers uncomfortable, restrict movement, and increase perceived distance, with passageway congestion having the most significant impact.
	Degree of Congestion on Stairs	X30	9
	Degree of Congestion on Passageway	X31	5
	Degree of Congestion on Entrance	X32	17	Congestion at station entrances and exits may increase contact between passengers and cause queues, affecting the comfort of transfer.

presents the discussion on the significance ranking of all 20 key factors.

4.5. Implications and Applications

Based on the research results, this study suggests taking the following measures to improve the perceived status of passenger transfer for the Jin-Yi-Dong Line of the Jinhua Rail Transit System:

Enhancing passenger experience in urban rail transit relies heavily on improving the Rationality of Signs and Markings (X22). One reason for this is that X22 has the most significant impact on passengers’ perceived transfer distance. Fortunately, the cost of improving X22 is relatively low, and optimizing it can be achieved during both station design and management stages. To ensure effective guidance for passengers, signage and markings should be strategically placed with clear and concise information. By doing so, confusion can be minimized, making it easier for passengers to navigate the station and ultimately decreasing their perceived transfer distance.

The optimization of escalators and stairs arrangement, ranked second in importance (X17, X16), is a critical aspect for enhancing the passenger experience in urban rail transit. When planning and designing stations, it is imperative to prioritize the installation of escalators over stairs to minimize physical exertion and reduce the perceived transfer distance, especially for passengers with limited mobility. Additionally, it is crucial to ensure there is an appropriate spacing between escalators and stairs and design them to accommodate effective passenger flow.

In addition, careful attention should be paid to ensure that the traffic organization plan (X23, X24) is adaptable and responsive to changing conditions, such as fluctuating passenger volumes or external events that may impact traffic flow. By continually assessing and adjusting the traffic organization plan as needed, the station can remain efficient and effective in providing a positive passenger experience.

Addressing congestion in passageways (X31) is a critical aspect of optimizing stations. During the station planning and design phase, it is important to estimate passenger flow and establish appropriate passageway sizes. In the operational phase, effective crowd management strategies should be implemented to alleviate congestion, such as deploying staff to guide passengers during peak periods, utilizing real-time information systems to notify passengers of congested areas, and modifying station layouts to promote efficient passenger flow. It is crucial to prioritize passenger safety and comfort by creating clear and unobstructed passageways.

The design and planning of stations should take into account the personal characteristics of the passengers being served, including those with disabilities, the elderly, and those who are unfamiliar with the station. To meet the needs of different passengers, the station design should include seating areas, elevators, and clear wayfinding information. Such features can facilitate passengers’ use of the station and enhance its accessibility and inclusiveness. By considering the diverse needs of passengers, the station can become a more inviting and user-friendly environment for everyone.

By implementing these recommendations, urban rail transit stations can effectively reduce passengers’ perceived transfer distances, improving their overall travel experience and satisfaction.

It is important to note that passenger flow characteristics, transfer facilities, and transfer environment may vary among different stations or even at the same station over different time periods, leading to differing levels of importance for various influencing factors. Therefore, this article provides design and optimization strategies solely based on the Jin-Yi-Dong Line of the Jinhua Rail Transit System at its current stage. These strategies may not be applicable to other stations.

5. Conclusions

This study made a significant contribution to the field of urban rail transit by proposing a novel concept, the Ratio Of Perceived Transfer Distance Deviation (R), which could accurately measure the level of transfer service at urban rail transit stations. Furthermore, this study investigated the potential of machine learning in the study of perceived transfer distance and the use of explainable methods for reviewing model predictions to address the long-standing “black box” problem in computational methods. The comprehensive evaluation system established in this research considered 32 influencing factors relating to Passenger Personal Attributes, Transfer Facilities, and Transfer Environment, providing a comprehensive understanding of the complexities of perceived transfer distance.

The use of advanced XGBoost and SHAP models resulted in a more advanced analysis of transfer distance perception at urban rail transit stations, surpassing other models such as linear regression, random forest, and LightGBM in predicting R, with an RMSE of 0.057 and an MAE of 0.042.

Moreover, this study found that Transfer Facilities, Transfer Environment, and Passenger Personal Attributes were significant influencing factor groups, with the Rationality of Signs and Markings, Ratio of Escalators Length, Rationality of Traffic Organization Outside the Station, Ratio of Stairs Length, and Degree of Congestion on Passageway being the most critical determinants affecting R for urban rail transit passengers.

The research findings provide valuable theoretical guidance for planning, constructing, and optimizing urban rail transit connection facilities, improving the accessibility of urban rail transit stations, and enhancing the level of transfer services provided to passengers. Furthermore, the research can contribute to predicting passenger transfer behavior, offering practical implications for urban rail transit management and planning. Overall, this study presented a novel approach to accurately measuring the perceived transfer distance at urban rail transit stations and provides significant contributions to the field of urban rail transit.