Mobility Behavior Segmentation for Personalized AMoD Service Design: Evidence from Israel

Abstract

For decades, transportation planning has relied on utilitarian principles, which aim to maximize cumulative benefit by meeting the needs of the “average user.” This approach ignores fundamental differences between population groups and produces uniform solutions that fail to address the diverse needs of women, children, the elderly, and other disadvantaged populations. In response, there are growing calls for a transportation justice paradigm that emphasizes individuals’ ability to access meaningful opportunities according to their characteristics, abilities, and life circumstances. Autonomous Mobility on Demand (AMoD) holds the potential to transform future transportation systems. However, without deliberate planning, they risk replicating existing patterns of inequality for populations whose mobility needs differ from those of the average user. This study applies transportation justice principles to examine how AMoD systems can be designed to meet diverse user needs. Using a combination of an Autoencoder for learning reduced representations and an HDBSCAN clustering algorithm, the analysis identifies distinct travel patterns across socioeconomic groups. These findings reveal significant gaps between population segments, particularly among children and older adults, and demonstrate how AMoD systems could expand access to after-school activities, reduce social isolation among elderly women, and reduce various transportation-related social gaps by improving their ability to reach a wider range of opportunities.

1. Introduction

The mainstream approach in transportation planning is grounded in utilitarian reasoning, which prioritizes maximizing overall utility by assuming that serving the needs of the “average user” yields the most significant societal benefit [1,2,3]. This claim is most evident in the widespread use of cost–benefit analysis, which evaluates transportation interventions primarily based on metrics such as travel time savings and willingness-to-pay. While useful for resource allocation, this approach often masks the uneven distribution of benefits and burdens, systematically disadvantaging certain social groups [4].

Specifically, the Utilitarian Approach overlooks the diverse transportation needs of different population groups as it is not concerned with accessibility as a value in itself, but merely as a means to facilitate utility-generating trips [2]. This orientation leads to aggregate measures, while disregarding the distribution of accessibility [4]. By reinforcing a one-size-fits-all approach to infrastructure and services, this logic sidelines the needs of women, children, older adults, and individuals with limited mobility, whose daily patterns and constraints differ significantly from the statistical norm [5,6,7,8]. Flattening differences into averages, this model fails to recognize and respond to the lived realities of many users. As a result, decisions that maximize total welfare may inadvertently reinforce exclusion and inequality, even under the pretense of neutrality.

These theoretical critiques are supported by empirical evidence. Public transit systems are often designed around an “ideal passenger”, typically a financially secure, able-bodied male commuter working standard hour [9]. As a result, those who fall outside this norm, such as women concerned about personal safety, parents traveling with strollers, people with disabilities, and low-income riders, often face numerous barriers, including high costs, limited accessibility, and, in some cases, discriminatory treatment. Consequently, these users may be forced to navigate difficult trade-offs between fulfilling essential family responsibilities and securing reliable access to transportation.

In response to these limitations, alternative ethical frameworks emphasize that fairness in transportation should be measured by an individual’s ability to participate in essential out-of-home activities [10]. Similarly, when applying the “Capability Approach”, developed by Amartya Sen and Martha Nussbaum, mobility should be assessed based on individuals’ fundamental freedoms and opportunities [11,12]. According to this view, access to transportation is not a goal in itself, but a condition for achieving broader capabilities, such as employment, education, and caregiving. Hence, equitable transportation is that which enables individuals to lead lives they have reason to value.

To challenge the utilitarian paradigm, a shift from resource or utility-based evaluations toward a mid-fare approach has been proposed [10]. This justice-theoretical framework focuses on the ability of individuals to translate transportation resources into meaningful opportunities for participation in daily activities. Rather than measuring how many kilometers can be traveled or how much enjoyment a trip provides, the mid-fare perspective highlights whether an individual can realistically access meaningful opportunities within their specific circumstances. Access to transportation determines the extent to which individuals can participate in economic, educational, and social life. Therefore, equity in transportation requires systems tailored to the varied life situations of different population groups, ensuring accessibility aligns with actual abilities [13,14,15].

The rise of emerging technologies such as Autonomous Mobility on Demand (AMoD) systems offers a crucial opportunity to reevaluate the core principles of transportation design and planning [14,15]. Unlike traditional transportation modes, AMoD services integrate vehicle automation with shared mobility and real-time responsiveness, providing dynamic, adaptable, and user-focused transportation solutions [16]. Their ability to tailor mobility services to the specific needs and travel patterns of different users can also enhance system efficiency and accessibility, allowing vehicles to operate more effectively and potentially reducing unnecessary travel and energy consumption [17,18]. As a result, AMoD systems can potentially reduce fleet sizes, lower travel expenses, ease congestion, and mitigate environmental impacts while improving accessibility, especially in regions underserved by traditional transit [17,19]. These environmental benefits may be further strengthened when such fleets operate using zero-emission propulsion technologies, for example battery electric vehicles (BEVs) or hydrogen-based fuel cell vehicles (FCVs), which are increasingly considered viable solutions for high-intensity shared mobility fleets [20,21], with evidence indicating that under appropriately designed deployment and pricing condition such systems can achieve substantial reductions in energy use and air pollutant emissions [19]. In addition, AMoD systems may help address the first/last-mile problem and reduce car dependency, thereby contributing to more sustainable and efficient urban environments [20,22].

However, even with these marked advantages, there is no guarantee that this future service will operate in a way that ensures equal opportunities across different population groups [14,15]. Alongside technological and operational considerations, the deployment of autonomous mobility systems also raises important regulatory and governance challenges. In particular, questions regarding responsibility and liability in the event of road incidents remain an important issue in the literature on autonomous vehicles [23]. These institutional aspects may influence public trust, regulatory acceptance, and ultimately the equitable deployment of AMoD services across different populations.

Without deliberate planning and design focused on fairness, AMoD systems risk replicating the inequalities embedded in existing transport models, continuing to privilege the average user. Bridging this gap requires future AMoD systems to adopt an inclusive approach, enabling them to function as tools for social inclusion rather than exclusion.

Clustering analysis is based on measuring the similarity between individuals in the population using a distance function applied to the selected indicators. A variety of clustering techniques have been employed.

This perspective is further substantiated as transportation behavior cannot be adequately accounted for by income levels alone [24]. An analysis of trip chains in the Greater Toronto and Hamilton Area reveals substantial variation in mobility patterns within the same income bracket, shaped by factors such as gender, car ownership, and the characteristics of the built environment. For instance, women in low-income households without access to a private vehicle frequently undertake complex, caregiving-related journeys via public transportation. In contrast, men in similar households are more likely to make direct car commutes. These findings demonstrate the limitations of average-based planning approaches and highlight the importance of incorporating differentiated mobility needs and the diverse experiences of various user groups into transport policy and design. The Capability Approach offers a particularly valuable perspective for the development of AMoD systems, as it shifts attention from aggregate travel behavior to the structural conditions that enable or constrain individual freedoms.

In empirical applications, operationalizing the Capability Approach in transportation requires identifying relevant capabilities and developing group-specific indicators that reflect the diverse needs and constraints of different population groups [11,12]. While commonly used indicators such as age and gender provide valuable insights, they do not fully capture the complexity of individuals’ mobility needs. More nuanced indicators, such as access to private and public transport, offer a more comprehensive understanding. For example, in Portugal, low-income populations are often compelled to rely on private vehicles due to poor accessibility to public transport, thereby exacerbating their financial burden [25].

Similar accessibility inequalities have also been documented in other national contexts. For example, research conducted in the United States [26] shows that disadvantaged populations, such as older adults, non-drivers, and residents of non-metropolitan areas, often experience lower access to public transport and essential destinations. These disparities highlight the broader challenge of transport disadvantages and its implications for social inclusion and equitable access to opportunities [27]. A systematic review by [28], covering studies from countries such as Canada and China, similarly identified persistent inequalities in public transport accessibility among older adults, with lower levels of accessibility often observed in peripheral and less central areas. These findings suggest that transport disadvantages are a widespread issue and highlight the importance of promoting more socially inclusive and equitable accessibility in transport planning.

Income level is a crucial factor in transportation analysis, as economic opportunities influence mobility behavior and higher income levels are associated with a greater adoption of multimodal transport applications, particularly in large urban areas [29]. This trend enhances users’ travel efficiency by fostering better planning, integrating transport modes, and providing greater flexibility in their daily commutes.

Income-based distinctions are particularly important in transportation analyses, as they help uncover the service potential for socially disadvantaged groups. For example, older women with low incomes demonstrated a high level of awareness regarding electric vehicles, yet their limited financial resources posed a significant barrier to adoption [30]. Each indicator can be considered on its own, but it is often helpful to combine several indicators, particularly through clustering analysis.

Clustering analysis is based on measuring the similarity between individuals in the population using a distance function applied to the selected indicators. A variety of clustering techniques have been employed in transportation behavior research to identify meaningful population segments. Variants of the K-Means algorithm are among the most commonly used, as demonstrated in several studies [25,30,31,32]. Hierarchical clustering has also been applied, for example, by [30,33], the latter primarily to determine the optimal number of clusters. Density-based clustering methods, including hierarchical density-based approaches (HDBSCAN), have also gained traction [34].

The different methods have their advantages and disadvantages, but all require selecting a criterion that defines when two individuals or their actions are considered similar. As not all variables are continuous, defining proximity between individuals is more complex. For example, gender is typically a binary variable, while employment status is a categorical variable. In such cases, one approach is to use distance functions that are not purely geometric, such as Gower distance [35], which can handle a mix of variable types, or Jaccard distance, which is suited for binary data. An alternative approach involves transforming all attributes, including quantitative, binary, and categorical, into continuous representations. This enables the use of methods that map individuals into continuous geometric spaces, such as UMAP and t-SNE [36], which preserve structural relationships in the data. Additionally, deep learning techniques, such as Autoencoders (AEs) [37], can learn compact latent representations that capture underlying behavioral patterns in a data-driven manner.

This paper bridges the gap between behavior-based evidence on disadvantaged groups and their potential mobility behavior under AMoD by combining population segmentation with the analysis of daily activity and travel patterns derived from large-scale travel survey data to understand how emerging technologies may affect such groups.

We thus contribute to the literature in four ways. First, we identify previously overlooked potential AMoD user segments and latent demand, highlighting new market shares that may emerge as automated and personalized mobility becomes available to populations currently constrained by limited transport opportunities. Second, we propose a scalable, data-driven population segmentation framework, presented in Section 2.1, which combines an Autoencoder for representation learning with HDBSCAN clustering to identify population groups based on heterogeneous socioeconomic and transport opportunity variables. Third, we provide evidence on how personalized AMoD could reduce transport-related social gaps, by showing which groups are most constrained today and which mobility opportunities could plausibly expand under tailored on-demand automation. Fourth, we provide large-scale empirical evidence using four Israeli Travel Habits Surveys (2010–2019) covering four metropolitan areas. The scale and diversity of the dataset enable the identification of rare and vulnerable population groups that are often underrepresented in smaller travel surveys.

To explore transportation justice considerations in the design of fair AMoD services, we analyze and compare travel patterns of different population groups using advanced segmentation; the analysis focuses on users with low accessibility and limited economic opportunities. To this end, we examine four large-scale Travel Habits Surveys (THS) conducted in Israel between 2010 and 2019 in the four major metropolitan areas: (1) Jerusalem, (2) Tel Aviv, (3) Haifa, and (4) Be’er Sheva. The consolidated 5 database comprises data from over 129,000 individuals, recorded over 169,000 travel days represented as daily activity logs. The rest of this paper is structured as follows. Section 2 presents the methodology, including the population segmentation and travel pattern analysis. Section 3 describes the results of the clustering and travel behavior analysis. Section 4 discusses the implications of the findings for transportation justice and AMoD service design with key insights.

2. Materials and Methods

The methodology consists of three steps: (1) population segmentation, (2) analysis of travel patterns, and (3) comparative analysis of groups with similar demographics but different opportunities.

2.1. Population Segmentation

Population segmentation is a complex, non-neutral process that involves value-based choices about which characteristics to include in the analysis and which to omit [38]. To address this challenge, the Capability Approach provides a normative framework that focuses on individuals’ real freedoms to achieve the lives they have reason to value, rather than merely accessing resources as emphasized in the Utilitarian Approach.

The Capability Approach emphasizes the need to identify gaps between social groups, such as those based on gender, income, or origin. However, as the variables increases, the number of possible subgroups also increases, and the analysis becomes more complicated [11,12]. Continuous variables, such as age or income, require conversion into categories, a decision that can significantly affect the results of the population segmentation [39]. Each step in this process not only describes reality but also reshapes it.

The survey activity database, which includes over 600,000 daily records, contains information on activity location, start and end times, purpose of the activity, and the mode of transportation used. Each activity is linked to one of 129,000 individuals surveyed, with demographic information including gender, age, employment status, level of education, number of children in the household, household role, place of residence, and availability of a driver’s license. In addition, a third database provides information at the household level, including data on household size, number of children, area of residence, and number of vehicles owned.

Before performing the population segmentation process, we added two variables that were not part of the original dataset. Income data were compiled from the Population and Housing Census [40], which provided average monthly income at the statistical area level. Each household in the survey was assigned an income value corresponding to its residential area, determined through spatial mapping between statistical and survey areas. For public transportation accessibility, the entire country was divided into a 200 m × 200 m grid. Each cell was assigned to a nearby public transportation station, which was classified according to service frequency during peak hours.

The highest public transport accessibility score was assigned to high-frequency services, including trains, light rail, and Bus Rapit Transit, while the lowest score was given to stations with hourly service. Cells without a public transport service were defined as inaccessible. Each surveyed area received an accessibility score between zero and one, reflecting the proportion of accessible grid cells within built-up areas or near paved roads, thereby excluding uninhabited or agricultural land.

The population segmentation framework used in this study consists of three stages: (1) construction of the individual information vector, (2) representation learning using an Autoencoder (AE), and (3) clustering using the HDBSCAN algorithm.

Table 1. Individual Information Vector.

Attributes Group	Vector Attribute	Values Range/Category	Type	Samples	Treatment
Demographic	Age	0–9, 9–16, 16-0, 30–65, 65+	Continuous	128,687	One-Hot Encoding
	Gender	Male	Binary	63,285	One-Hot Encoding
		Female		65,402
	Education level	Less than 12 years	Categorical	59,162	One-Hot Encoding
		High school diploma		42,285
		Undergraduate academic degree		18,620
		Master’s degree/Doctorate		8099
		Other		521
	Employment status	Unemployed	Categorical	77,311	One-Hot Encoding
		Part-time/Other		9841
		Full-time employed		40,641
		Military service		894
	Relationship	Single-person household	Categorical	6958	One-Hot Encoding
		Spouse/Partner		16,405
		Parent		41,560
		Child		60,184
		Grandparent		550
		Other relative		1483
		Housemate/Flatmate		852
		Caregiver/Assistant		338
		Other		357
Transport opportunities	Children under 8	0–7	Continuous	128,687	Min-Max Normalization
	Income	3101–22,948 NIS	Continuous	128,687	Min-Max Normalization
	Number of vehicles	0–7	Continuous	128,687	Min-Max Normalization
	PT accessibility score	0–1	Continuous	128,687	No treatment

presents the structure of the collected data used for population segmentation.

In Table 1, the data attributes are either binary, continuous, or categorical. To address this heterogeneity, we employ a mapping that transforms all attributes into a continuous vector space, enabling the application of standard distance metrics, such as the Euclidean distance. For this purpose, we use an AE neural network. An AE consists of two mapping structures: an encoder (E: x → z), and a decoder (D: z → $\hat{x}$). The network is trained such that the composition of the encoder and decoder reconstructs the input: (E(x)) ≈ x. The AE can also perform dimensionality reduction, facilitate the grouping of similar individuals while effectively capturing nonlinear patterns in the data, particularly in comparison to linear methods such as Principal Component Analysis [41]. In addition, AE also provides a more flexible foundation for learning meaningful representations than non-parametric methods like t-SNE or UMAP, particularly when the goal is clustering or downstream learning [36].

Traditional mixed-distance measures such as Gower distance can also be used to handle heterogeneous datasets containing continuous, categorical, and binary variables. However, such approaches rely on predefined distance formulations and typically treat variables independently when computing similarity. In contrast, the AE framework learns a latent representation of the data through nonlinear transformations, enabling the model to capture interactions between socioeconomic characteristics and transport opportunity variables. In addition, AE provides a compact continuous embedding of the heterogeneous feature space, which improves the stability of density-based clustering methods such as HDBSCAN when applied to high-dimensional survey data. For these reasons, AE was selected as the representation learning step prior to clustering.

The dimension of the individual vector x is the 25 attributes from Table 1. To learn the continuous vector z (the latent space), we use two fully connected layers with 64 and 32 neurons, each with ReLU activation functions. The latent space contains 16 neurons, enabling dimensionality reduction. The reconstructions are symmetrical, such that the decoder has fully connected layers with 32 and 64 neurons, each using a ReLU activation. Finally, in the reconstructed vector, we use the sigmoid function to retrieve the original data, which has been normalized between 0 and 1. The neural network is trained with a learning rate of 0.001 using the mean squared error loss function and the Adam optimizer, achieving a loss of 0.002724 after 100 epochs. The small loss indicates that the AE performs well in reconstructing the data, making z suitable as the continuous vector for the population segmentation algorithm.

The architecture of the Autoencoder was selected following preliminary experimentation aimed at balancing reconstruction accuracy, representation capacity, and computational efficiency. During model development, several alternative configurations were tested, including smaller latent dimensions and different numbers of hidden units. These experiments showed that overly compact representations resulted in higher reconstruction errors and loss of relevant information from the input data. In contrast, architectures achieving similarly low reconstruction errors produced comparable latent representations and led to similar clustering structures when combined with HDBSCAN. The selected layer sizes (64–32–16) also follow a common funnel-shaped architecture with gradually decreasing dimensionality, and powers-of-two layer sizes that are widely used in deep learning implementations due to their computational efficiency in matrix operations [42].

After obtaining a 16-dimensional continuous representation of the data, we apply the HDBSCAN clustering algorithm for population segmentation. The HDBSCAN algorithm was chosen for this task due to its ability to handle large datasets in a reasonable runtime and to identify complex structures in the data, particularly when compared to methods like K-means [43]. The algorithm is well-described in the open-source HDBSCAN library developed by [44]. It has two hyperparameters; the first is the minimum number of samples, which defines how many individuals must be similar to a given person for them to be considered part of a group in the data space. This threshold determines whether a dense region can be identified as a potential cluster. The second hyperparameter is the minimum cluster size. To identify the best combination, we use a random search over four rounds, each evaluating the configuration with the best Density-Based Clustering Validation (DBCV) score. A higher DBCV score indicates clusters that are more compact and better separated from one another. Figure 1 illustrates the search process using the Optuna hyperparameter optimization framework [45], which implements a Bayesian optimization strategy.

The maximum DBCV score achieved, as shown in Figure 1, is 0.37 with a minimum number of samples of 108 and a minimal cluster size of 40 individuals. In heterogeneous behavioral datasets such as travel surveys, clustering quality metrics often yield moderate values due to the inherent variability of human activity patterns. The DBCV metric ranges between −1 and 1 and is primarily intended for the comparative evaluation of density-based clustering results rather than as an absolute quality threshold. Accordingly, in this study, the metric is interpreted relatively across the tested parameter configurations. The reported value (0.37) corresponds to the highest DBCV score obtained among the evaluated HDBSCAN configurations and therefore represents the most appropriate clustering structure within the explored parameter space. A total of 102 different clusters were formed, with 11% of the individuals classified as noise. The smallest cluster includes 43 individuals, while the largest contains 17K individuals. Our next step is to identify the “representative individual” of each cluster, in order to compare different travel patterns.

To identify this information, we calculate the medoid for each cluster. The medoid is the individual observation in the cluster that has the minimal sum of distances from the other individuals in the same cluster, as described in Equation (1)

$$m\mathrm{e}\mathrm{d}\mathrm{o}\mathrm{i}\mathrm{d} \mathrm{r}\mathrm{a}\mathrm{n}\mathrm{k}=\arg \underset{x\in C}{\min }\sum _{y\in C}d\left(x,y\right)$$

(1)

where $C$ is the cluster to which both $x$ and $\mathrm{y}$ belong to, and $d\left(x,y\right)$ is the Euclidean distance function. The medoid is defined as the point with the minimal summed distance to all other points in the cluster. After selecting the 102 medoids in the latent-space clusters, the corresponding data points are reconstructed using the decoder to obtain their original representation. The medoid is used as a representative reference point for interpreting the socio-demographic profile of each cluster. It corresponds to the most central observation within the cluster, i.e., the individual with the smallest total distance to the other members. However, it does not fully capture the internal variability of large clusters and should therefore be interpreted as a typical representative rather than a complete description of the group.

2.2. Travel Patterns

The second part of the methodology involves analyzing travel patterns within each cluster. To systematically represent individuals’ daily mobility behavior, the activity records from the THS survey are transformed into a binary schedule matrix. The daily activity data stored on the THS survey can be represented as a binary matrix, referred to as the schedule matrix ($S_p$, Equation (2)), where each column corresponds to a 5-min interval of the day, and each row represents a specific activity purpose ($a$) or mode ($m$).

$$S_p=\left[s_{i,j}\right]\in \{\mathrm{0,1}{\}}^{\left(m+a\right)\times 288}$$

(2)

$\mathrm{where} j=0,\dots ,287$. If a person $p$ is engaged in activity $j\in \{0,\dots ,a-1\}$ during the time interval $i$, then the matrix cell $s_{i,j}=1$. Likewise, if the person is in transit using a mode $j\in \{a,\dots ,a+m-1\}$ at interval $i$, then $s_{i,j}=1$ as well. All other entries in the column $i$ are set to zero. This representation was previously used as part of the data preparation process for hierarchical clustering of travel patterns [34]. For each population subgroup identified in the segmentation stage, the corresponding schedule matrices are collected and transformed into column vectors (“flattened” representations) that capture the full set of activity and mode indicators across the day. Importantly, the travel activity and mode patterns analyzed in the following stage are computed using all individuals belonging to each cluster, whereas the medoid is used primarily to provide an interpretable socio-demographic label for the group.

2.3. Comparative Analysis

Next, we apply the HDBSCAN clustering algorithm again, conducting two runs for each population subgroup: one on the activity section of the flattened schedule matrix and another on the mode section. Due to the binary nature of the data, we use Jaccard distance [46] as the dissimilarity measure for clustering, following the same procedure used in the population segmentation phase.

This procedure enables the identification of dominant activity and travel mode patterns within each population group. Populations that exhibit clear activity patterns are of great importance in understanding the needs for AMoD services, as they enable more accurate identification of travel habits and optimal adaptation of AMoD services. In contrast, populations that do not exhibit distinct patterns make the process of identifying needs more difficult, which may hinder the ability to plan a tailored and effective service for them.

This study focuses on population groups that exhibit distinct travel patterns, with particular attention to medoids located at demographic extremes, specifically the very young and the elderly, who are particularly vulnerable in terms of mobility. Comparisons are made between groups similar in age or demographic characteristics to assess how automated vehicles may affect these vulnerable groups. This analysis is based on the assumption that, with the availability of AMoD services, the travel behavior of disadvantaged groups could begin to resemble that of more empowered population segments. The comprehensive workflow of the proposed methodology, illustrating the integration of socioeconomic segmentation and the subsequent travel pattern analysis, is summarized in Figure 2.

3. Results

The methodology identified four distinct clusters for children aged 0–8 and four clusters for those aged 9–16. The medoids of these clusters are presented in

Table 2. Cluster Medoids for Children Aged 0–16.

Symbol	Age	Income	Vehicles	Siblings <8	PT Score	Gender	Relation	Education	Employment
A-1	0–8	6689	0	3	0.69	Female	Child	<12 years	Unemployed
A-2	0–8	8885	1	2	0.24	Female	–	–	–
A-3	0–8	6516	0	3	0.68	Male	–	–	–
A-4	0–8	8381	1	2	0.23	Male	–	–	–
B-1	9–16	6621	0	2	0.69	Female	Child	<12 years	Unemployed
B-2	9–16	8872	1	0	0.23	Female	–	–	–
B-3	9–16	7063	0	2	0.73	Male	–	–	–
B-4	9–16	8569	1	1	0.20	Male	–	–	–

. Across all child clusters, we observe that those associated with high family income tend to have low accessibility to public transportation. However, they exhibit high accessibility to private cars, provided that the household owns at least one vehicle. In contrast, the medoids of low-income children are characterized by zero car ownership.

As a validity check, we observe that clusters representing children consistently include the attributes “Unemployed” for employment status and “<12 years” for education, which is self-evident. If such patterns did not appear, it could indicate a flaw in the methodological process. The A-1 and A-3 groups of children are characterized by low income, high accessibility to public transportation, and low accessibility to private vehicles. The main difference between the two groups is the gender of the children.

Figure 3, Figure 4, Figure 5, Figure 6, Figure 7 and Figure 8 in this section visualize daily activity and travel mode patterns at the individual level. In these figures, each row on the vertical axis represents a single individual, while the horizontal axis represents time of day in 5-min intervals. The colors indicate the type of activity or travel mode performed at each time interval. In addition, a horizontal bar displayed above each figure summarizes the temporal intensity of out-of-home activity or travel. In activity figures, darker or more saturated segments indicate time intervals in which a larger share of individuals is engaged in activities outside the home. In travel mode figures, the bar reflects the overall intensity of travel across individuals at each time interval. The legend associated with each figure reports the share of total daily time (in percentage) allocated to each activity type or travel mode within the population group. This provides an aggregate indication of the relative prevalence of different activities or modes across the group.

When examining the travel patterns of children in these groups, it appears that the main destination for travel is educational programs. According to Figure 3, most educational activities begin around 8:00 a.m. and end around 1:00 p.m., with a “tail” on the girls’ group (A-1) who also attend formal care settings between 1:00 p.m. and 4:00 p.m.

In contrast, in groups A-2 and A-4, which are characterized by high income, low accessibility to public transportation, and high accessibility to private vehicles, there is a trend toward a higher proportion of activities between 1:00 p.m. and 4:00 p.m., which is reflected in a more intense yellow coloring at those times in Figure 3. The four groups of children also engage in non-educational activities, mainly in the evenings, but in low-income groups, they are seen less often, as it is more common to stay home after school.

When examining the patterns on A-1 to A-4 groups, it can be seen that among children from low-income households, the rate of staying at home is higher: 77% of boys and 73% of girls stay home after school. In contrast, among children from high-income households, the rate of staying at home is 68–69%.

Among children from high-income families, afternoon activities such as visiting friends and participating in clubs are common. In contrast, boys from families with low socioeconomic status are characterized by a low rate of participation in afternoon activities. When examining the travel patterns of these population groups, it can be seen that children from low-income families tend to travel during three prominent time windows throughout the day: around 8:00 a.m., 1:00 p.m., and to a lesser extent at 4:00 p.m. In low-income families (A-1 and A-3), the use of walking among children as the main mode of transportation is particularly prominent (represented by the red color in Figure 4). Overall, 47% of girls’ travel time and 58% of boys’ travel time are dedicated to walking, mainly during these hours.

In contrast, among children from high-income families (A-2 and A-4), a distinct pattern emerges, characterized by the extensive use of private vehicles (represented by the blue color in Figure 4), which accounts for over 70% of the group’s total travel time. After 4:00 p.m., walking is almost non-existent, and private car use clearly dominates. Walking appears to be limited to the three main time windows identified earlier. In addition, in high-income families, bus use is extremely rare, less than 2.5% of total travel time, compared with rates of 17–19.8% among boys and girls from low-income families. This can be explained by the partial and sometimes problematic availability of the public transportation system, which leads people to use it only when no other option is available.

These results reinforce the activity patterns observed in Figure 3, where children from high-income families show a wider distribution of afternoon activities, likely due to the use of private vehicles, which enables greater mobility compared to walking. Interestingly, when transportation opportunities are available, parents seem to prefer that their children (or themselves as escorts) not walk in the late afternoon. This preference may stem from personal safety considerations.

Figure 5 presents the activity patterns of youth aged 9–16 across the day. In groups B-1 and B-4, representing children aged 9 to 16, similar activity patterns to those of children aged 0 to 8 are observed, with education-related activity remaining dominant (Figure 5). Among this age group, the most notable differences appear along the gender dimension. Boys tend to stay at home at a slightly higher rate (72% in group B2 and 70% in group B4), compared to girls (67% in group B1 and 69% in group B3).

Differences between income groups are not significant, but certain trends can be observed. For example, in the high-income groups (B-1 and B-3), sports activities are not represented at all in the travel patterns. In contrast, in the low-income groups (B2 and B4), sports activities account for 0.2% of the total activity time among boys and 0.5% among girls. Leisure activities and visiting friends are also affected by the level of economic well-being. Among boys, the gap between high- and low-income groups reaches 0.8% of the total activity time for leisure activities, and 1.3% for visiting friends.

Among youths aged 9–16, clear differences emerge in travel mode patterns between groups with different levels of accessibility and opportunities. In the lower income groups (B-1, B-3), the most common transportation modes are walking and bus travel (represented by red and yellow colors, respectively, in Figure 6). Approximately 45% of total travel time is spent walking, while about 30% is spent on public transportation.

Figure 6 shows a clear concentration of trips between 7:00 and 8:00 a.m., alongside additional trips dispersed throughout the afternoon, starting at 1:00 p.m. Even when adolescents from low-income families participate in afternoon or evening activities, they primarily rely on walking or public transportation. On the other hand, in the high-income groups (B2 and B4) characterized by high accessibility to private cars (represented in blue color in Figure 6), a significantly different pattern emerges. In the morning, between 7:00 and 8:00 a.m., the majority of trips are made by private car, accounting for over 50% of the total travel time of these groups.

Between 1:00 and 4:00 p.m., a high intensity of additional trips is recorded in all groups. In the evening hours, trips among children from high-income families are mostly made by private car, usually as passengers. Buses account for 13.6% and 14.7% of trips among the high-income adolescents for girls and boys, respectively.

The clustering process also yielded 23 distinct medoid representations, each reflecting a different profile of individuals aged 65 and over. Among these 23 medoids, only four yielded distinct travel patterns that the HDBSCAN algorithm recognized as valid clusters rather than noise. These four representative profiles are detailed in

Table 3. Cluster Medoids for Elderly Aged 65+.

Symbol	Age	Income	Vehicles	Childerns < 8	PT Score	Gender	Relation	Education	Employment
C-1	65+	8843	1	0	0.42	Female	Spouse/Partner	High school	Unemployed
C-2	65+	9816	1	0	0.32	Female	Spouse/Partner	Bachelor’s	Unemployed
C-3	65+	8349	0	0	0.43	Male	Single	High school	Unemployed
C-4	65+	9789	1	0	0.32	Male	Spouse/Partner	High school	Unemployed

.

The PT accessibility score is relatively low across all these groups. However, the relationship found among children and youth is maintained here as well; higher income is associated with a lower accessibility score, whereas comparatively lower income sometimes coincides with better PT accessibility. In addition, all groups exhibiting clear travel patterns consist of older adults who are unemployed. This may reflect the wide differences in working hours and the nature of activities among people of this age, which make it difficult to identify uniform patterns among older working adults.

Figure 7 shows the activity patterns of groups C-1 andC-4. Across all groups, a very high rate of staying at home is recorded, over 80% of the group’s daily time. At these ages, the time spent at home does not appear to depend on income but is mainly influenced by gender.

Women in this age group tend to spend more time at home than men. In the lower-income groups (C-1, C-3), women spend about 95% of the time at home, compared with 86% among men. There is also a gender gap in the higher income groups (C-2, C-4), with women staying at home about 86% of the day, roughly 3% more than men in this group. Such a high rate of lack of involvement in activities outside the home has previously been identified as a risk factor for reduced mobility among older adults in Israel [47].

Group C-1, characterized by the highest rate of staying at home, exhibits limited patterns of going out. In this group, when women do leave the house, most trips are made for grocery shopping (32.2%), followed by personal errands (28.8%), and finally, medical visits (13.5%). Other activities occur at rates below 10%, including sports, leisure, and visiting friends.

However, this group stands out for its social isolation; only about 9% of the time spent out of the home is dedicated to visiting friends, compared to stable rates of over 23% among the other groups. This finding highlights the vulnerability of this group. In terms of the distribution of activity by time of day, two main columns can be identified in the graph representing the peak times of activity periods: one in the late morning and the other in the afternoon and evening. The first period spans approximately from 9:00 a.m. to 1:00 p.m., while the second shows a high concentration of activity between 4:00 and 8:00 p.m. However, group C-1 deviates from this pattern: it shows only one period of activity in the morning, with no significant presence of activity in the evening.

Figure 8 illustrates the mode choice patterns among elderly groups. The most isolated group, C-1, is characterized by a very high dependence on public transportation, with 60.9% of travel time spent on buses. In contrast, older women with higher incomes (C-2) use the bus for only 8.8% of their travel time and significantly prefer private car transportation, which accounts for about 47% of their travel time. Convenient access to transportation contributes to a higher ability to participate in activities, while maintaining a rate of staying at home similar to that of men from lower-income groups, highlighting the behavioral pattern of the advantaged women group (C-2).

Among men, a clear preference for private cars is evident (represented by green color in Figure 8), almost regardless of socioeconomic status. Men from the lower-income group (C-3) travel by private car at a rate of 57%, while in the higher-income group (C-4), the rate rises to 65%. The pattern observed in Figure 7, which reflects men’s advantage over women in this age group in terms of out-of-home activity participation, may be explained by the high car accessibility shown in Figure 8, as it enables greater independence and mobility.

4. Discussion

The HDBSCAN algorithm was applied after dimensionality reduction with an AE to identify disadvantaged groups in the population that exhibit distinct travel patterns. These groups were compared to advantaged populations by analyzing their travel patterns to examine how personalized autonomous vehicle services tailored to their specific needs could change existing accessibility patterns.

4.1. Implications for Equitable AMoD Design

The findings reveal substantial socioeconomic disparities in children’s travel patterns. Children from low-income families tend to rely heavily on walking, primarily for educational purposes in the morning and, to a lesser extent, for afternoon activities, particularly among boys. These patterns reflect limited access to private vehicles, which, among the Israeli study population, results in reduced ability to engage in after-school activities, particularly during the later hours of the day. The introduction of AMoD-based transportation services, assumed to be less costly and autonomous, could enable children from low-income families to participate in afternoon and evening activities, as is currently observed among children from higher-income families, thereby expanding their opportunities.

Safety considerations likely explain the avoidance of walking with children in the evenings, as indicated by the results. Consequently, this pattern may also manifest in a preference for using autonomous vehicles for evening travel once such services become available. While AMoD adaptation may contribute to reducing social gaps in late-day activities, it could also lead to less desirable outcomes, most notably a decline in walking. A reduction in walking could harm the health and well-being of children, who have traditionally relied on walking as their primary mode of transportation, and also reduce their daily physical activity.

Among teenagers from high-income families, parental involvement in transporting children is high both during and after school hours. This phenomenon reflects the high availability of private vehicles as well as the resources and time that high-income parents can dedicate. Theoretically, autonomous vehicles could free up time for parents’ personal activities and encourage independence among youth. This is particularly relevant since, unlike in the 0–8 age group, an adult’s presence in the vehicle is not always necessary as children mature. On the other hand, populations with lower incomes rely primarily on walking and public transit. Introducing autonomous vehicles to the market might reduce the use of these modes, as already shown in simulation studies [48]. However, autonomous vehicles could enable children and teenagers from disadvantaged populations to participate in evening activities more similarly to their higher-income peers, thereby increasing their opportunities and freedoms.

In analyzing activity patterns among the elderly population, the most vulnerable group in terms of social isolation is middle-income elderly women. This group primarily leaves the house for functional needs, such as shopping and running errands. At the same time, the rate of participation in social activities, including visiting friends or engaging in sports, is significantly lower compared to other groups. Overall, all groups of elderly people examined demonstrated a very high rate of staying at home throughout the day. Furthermore, among higher-income groups at this age, involvement in leisure and sports activities is higher compared to disadvantaged populations of the same gender. These insights indicate the potential of AMoD to expand the range of opportunities and improve accessibility for elderly women with limited mobility. This population, currently characterized by social isolation and limited opportunities, can potentially increase the rate of visiting friends from 9% to ~27%, with the help of a tailored service, similar to the more advantaged groups. Such a change could make a significant contribution to the mental and social health of these women. Older women may therefore represent a key target group that would particularly benefit from autonomous vehicle services. This inference is supported by the observation that, when resources permit, their preferred mode of travel is by private car as a passenger rather than driving themselves or using public transportation.

Overall, the findings indicate significant potential, both in expanding the scope of after-school activities among children and in reducing the rate of staying at home among older women. These findings highlight how personalized AMoD services can broaden accessibility and opportunities, ultimately reduce social disparities and dependency, while also providing decision-makers with insights into existing inequalities. Such insights can guide targeted interventions that extend beyond autonomous vehicle technology, including improvements to PT services in low-income neighborhoods and offering subsidized on-demand services for older adults, particularly older women, and young children. Furthermore, these insights can extend to the design of the vehicles themselves, enabling user-centered interior configurations that accommodate the specific needs, safety considerations, and comfort preferences of diverse population groups.

From a planning perspective, the empirical findings suggest several equity-oriented design principles for future AMoD systems. The results show that children from low-income households participate less in afternoon activities and rely heavily on walking due to limited access to private vehicles. This pattern indicates that pricing mechanisms such as targeted subsidies or reduced fares for disadvantaged households could help reduce economic barriers and enable greater participation in after-school activities. At the same time, the clustering results reveal that populations with lower incomes often experience limited access to private vehicles despite living in areas with relatively good public transport accessibility, which constrains their mobility options. This suggests that AMoD service coverage could be prioritized in neighborhoods characterized by lower car ownership or reduced transport opportunities, ensuring that on-demand services complement existing accessibility gaps. The activity patterns identified in the analysis also highlight clear temporal constraints in mobility, particularly among children during afternoon hours and among older adults during daytime hours. Increasing AMoD service availability during these specific time windows may therefore better support the daily mobility needs of these groups. In addition, the findings regarding social isolation among elderly women suggest that targeted policy instruments, such as mobility credits or partnerships with social services and community programs, could help ensure that AMoD services effectively support populations facing structural mobility barriers.

4.2. Limitations

This work should be considered along with its limitations. Although segmenting the population and analyzing the most common travel patterns within each group offers clear advantages, generalizing a single pattern across all group members presents challenges. This is particularly the case when the probability gaps between the most common pattern and alternative patterns are not significant, as was the case in some of the clusters analyzed. In such cases, the main pattern represents only a partial trend and may not fully capture the internal diversity of the group.

Another critical assumption to the discussion of the results concerns a prevalent claim in contemporary research, according to which AMoD services are expected to reduce consumers’ costs and increase transportation accessibility. Although this is a common assumption, it is essential to remember that it is inherently prospective and uncertain. This study employs this assumption to analyze a possible future scenario, with the aim of understanding the social and spatial implications of deploying personalized AMoD services. Furthermore, this study assumes that groups with fewer opportunities will behave similarly to groups with greater socio-economic power if they are offered an available and accessible transportation solution, such as an affordable AMoD service. However, this assumption is also uncertain. While improving transportation accessibility can remove a significant barrier, disadvantaged populations may still participate less in activities, even when affordable and accessible transportation is available. This may be due to other economic constraints, such as the inability to finance the activities themselves or the lack of free time among working parents. In addition, the actual adoption of AMoD services may depend on further factors such as service pricing, social acceptance, safety perceptions, supervision requirements, and regulatory conditions. Furthermore, the deployment of autonomous mobility systems also raises regulatory and legal challenges, particularly regarding responsibility and liability in the event of road incidents. These issues may influence public trust and the broader acceptance of autonomous mobility services. The feasibility of implementing AMoD services may also depend on the quality of road infrastructure. Autonomous vehicles typically rely on well-maintained road networks, clear lane markings, and reliable infrastructure conditions to operate safely and efficiently. In regions where road infrastructure is limited or poorly maintained, the deployment and scalability of AMoD services may face additional challenges. Future research could further examine how infrastructure conditions influence the practical implementation of autonomous mobility services.