A Federated Personal Mobility Service in Autonomous Transportation Systems

Abstract

A personal mobility service (PMS) is developed to support personalized travel options for users in autonomous transportation systems (ATS), based on a macro-system state and micro-user behavior. However, this functionality necessitates processing and transmitting vast amounts of data, raising concerns about user privacy protection during data processing and transmission within the PMS. Furthermore, the PMS must be maintained and perform well, while preserving privacy. Therefore, we propose a novel federated PMS, denoted as a FPMS. Specifically, the FPMS can serve users’ personal mobility needs by facilitating the collaboration between the physical and information domains. Then, a common framework for FPMS architectures, which captures the features of ATSs, is proposed and discussed from both physical and logical perspectives, which include both the logical architecture and physical architecture; and we present the key algorithms for the FPMS, in conjunction with a artificial neural network (ANN). Additionally, in static estimation scenarios, the FPMS demonstrated a similar accuracy for three different models compared to the traditional PMS, while reducing the computing time by approximately 60% and communication resource consumption by approximately 85%.

1. Introduction

At present, the emergence of ubiquitous GIS and GPS has made it more convenient for travelers to plan and enjoy their trips [1]. Route planning services represent one form of travel information service [2] and mainly provide pertinent guidance information for traveling, contributing to the quality of public transportation. A problem with route planning is related to identifying new routes from existing road networks, with the trip needs of users as the main input [3]. As an important spatial analysis tool, route planning services have been the solution to searching for the shortest path based on travel OD information.

Driven by relevant technologies, such as positioning technologies and communication technologies (e.g., the internet of things and big data [4]), information can be obtained more quickly and accurately, and data can be processed and transmitted more effectively and efficiently; therefore, route planning services have developed and evolved gradually. The current traffic-aware route planning algorithms aim to determine a route that minimizes travel time, while incorporating real-time traffic conditions [3]. Meanwhile, travel demands have been steadily increasing, and travel patterns are diverse and variable. Modern planning and navigation systems usually offer the shortest or fastest route. However, as people have diverse requirements for services, they may not be satisfied with these limited route planning options [5]. Therefore, there is potential for developing next-generation route planning services, which take into account multiple attributes and user preferences.

The progress in intelligent transportation systems (ITS) has been propelled by both increasing demands and emerging technologies, and they have the potential to facilitate the transition toward next-generation autonomous transportation systems (ATS). The realization of a self-organized transportation system requires the implementation of autonomous perception, learning, decision-making, and control [6]. This facilitates active recognition of system status changes, the dynamic learning of potential mobility demands, the rational allocation of relevant resources, and the rapid provision of non-intrusive services. In an ATS, to realize self-awareness, a huge amount of data is involved treatment an ind transmission, which challenges the route planning services when matching user preferences in an ATS.To facilitate personalized services, with consideration of real-time traffic and under the integrated workflow of an ATS, the concept of a personal mobility service (PMS) has been proposed [7]. A PMS is required to work collaboratively with diverse participants within the ATS to comprehensively perceive macroscopic situations. Additionally, a PMS should possess the capability to accurately learn individual user behaviors, to assist decision-making when defining service schedules. Ultimately, a PMS should efficiently and effectively provide personalized solutions that result in the maximum system savings.

The development of PMS services within ATS requires the acquisition and processing of a vast amount of data, to support personalization and optimization. This presents several challenges and problems for a PMS:

• What should the external behaviors and internal structures of a PMS be in an ATS environment?;
• How should a PMS effectively protect user privacy with a large amount of sensitive data?;
• How should a PMS maintain and improve performance, while protecting privacy?

This paper addresses these challenges by proposing a new federated personal mobility service (FPMS), which utilizes a privacy-preserving decentralized mechanism called federated learning (FL) to coordinate service participants in the PMS system [8,9]. First, the external behaviors of the FPMS are explained through an overview. Second, the internal structures of the FPMS are demonstrated using the proposed architecture and key algorithms, which reveal the FPMS privacy-preservation characteristics. Finally, a performance evaluation is performed, to describe the benefits of the FPMS in collaboration with the service cluster consisting of massive smart devices.

In the following sections, we will delve deeper into the topic of federated personal mobility services. First, Section 2 will provide a comprehensive review of the related literature. Section 3 describes the FPMS from two viewpoints (the physical view and the information view). Section 4 proposes two aspects of the FPMS architecture (logical architecture and physical architecture) and key algorithms. Section 5 analyzes the performance differences between the federated PMS and the traditional PMS, through various indicators(training time, resource consumption, training loss, and learning accuracy). Section 6 concludes about the achievements of our research and gives suggestions for future studies.

2. Literature Review

A personal mobility service (PMS) is a type of route planning service that is specific, advanced, and on-demand. In the past few decades, much research has been carried out to develop efficient and effective route planning methods. Previous works related to the topic can be summarized into three main areas: routing data management, location-based route planning, and keyword-aware route planning [3,10]. As for location-based route planning, there are three types of related studies: source–destination-based, multi-location-based, and multi-user-based. A PMS could be regarded as a typical source–destination-based route planning service, which aims to derive the optimal route from a source location to the destination location under user-specific constraints. Previously, many efforts were made to solve the time-dependent shortest problem(TDSP), and Dijkstra was proposed as a shortest path algorithm a few decades ago [11]. Normally, TDSPs are based on static data and focus on variables [1,12,13] such as distance, time, and so on. As relevant technologies continue to evolve, traffic situation information can now be sensed and acquired more promptly and precisely. This enables the integration of the real-time traffic condition status into route planning [5]. Besides static data (e.g., planned timetables), dynamic and real-time data (e.g., timetable deviations, current locations) and dynamic estimation models have been considered in traffic-aware route planning (TA), to find the optimal route [5,14]. Furthermore, studies have also proposed self-aware route planning (SA), which considers planned routes with future traffic conditions [3].

Although TA and even SA systems could optimize route planning, some works [15] showed that, in reality, variables such as fastest and shortest may be unnecessary. Besides time and distance, different travelers have preferences for different features of a journey (e.g., connectivity, fare cost, convenience, etc.), which influence their travel decisions, including route choices [16,17]. For example, in the COVID-19 pandemic, people changed their travel habits [18] and considered the risk of infection more [19], which gave great importance to personalization during route planning. Recommending a route based on user preference can increase the utilization rate of the road network; therefore, personalized route recommendations are meaningful and attractive. Personalization refers to the development of enabling systems that can identify and anticipate the unique requirements of each individual and meet those needs within a familiar context [20]. Personalized information about users can be obtained through direct interaction, by simply asking for their input [21,22]. However, in some cases, indirect acquisition methods that require no interaction with the user are employed for this purpose [23]. The present preference-based routing is learned from historical trajectories [3,16] and travel choice profiles [24,25] and then used to guide routing algorithms to produce personal options.

With the rapid development of technologies and the constant increase in travel demands, it is proposed that transportation systems should be developed and evolved into advanced systems called autonomous transportation system (ATS). An ATS aims to utilize the capabilities of emerging technologies, to enable self-organizing operations and auto-service processes, including an active response to traffic demands, self-driving of vehicles, autonomous management of traffic infrastructure, and active adaptation to the traffic environment. Previous works in the area of preference-based routing have primarily focused on leveraging trajectory data to enhance the accuracy of planned routes [3]. Under the conditions of an ATS, personalized route planning entails higher requirements in data mining, data processing, and knowledge extraction; therefore, a PMS is proposed to provide an active response to personal travel demands and a self-aware route planning service. In particular, personal travel needs and the ATS environment bring a first challenge: what should the external behaviors and internal structures of a PMS be?

Generally speaking, a PMS learns from travel choice profiles, collects historical travel trajectories, and analyzes travel preferences and needs, to recommend optimal route options that match travelers’ habits. PMSs collect large amounts of sensitive data from many travelers and may lead to privacy leakages [26], which raises the second challenge: how to effectively protect data security and user privacy during personalized analysis.

The existing research on route planning privacy protection has primarily focused on anonymization techniques, which can be classified into three categories: clustering-based approaches [27,28], graph encryption approaches [29,30,31], and graph modification-based approaches [32]. However, these approaches are limited by their low graph utility or high search overheads, which can impact the service performance. Hence, novel techniques are required to overcome these limitations and provide effective privacy protection for route planning applications. As a result, the third challenge is how a PMS can, not only protect privacy, but also maintain and improve performance.

3. FPMS Description

The federated personal mobility service (FPMS), as one of the core services of an ATS, comprehensively senses real-time traffic situations and user preference information and then intelligently provides personal travel options, integrating multiple transit modes to assist individual mobility. As shown in Figure 1, FPMS, which plays a role in satisfying users’ personal mobility demands, is enabled by the collaboration between the physical and information domains. In the physical world, the FPMS senses diverse real-world information from different sources for its fundamental input. From the information viewpoint, there are three types of platform: data platform, back-end service, and front-end user interface (UI), working together to satisfy users.

3.1. Physical World

In general, systems or services receive massive amounts of diverse information from the physical world as input. The FPMS mainly collects dynamic and static data from the complex transport infrastructure, multiple transit modes, and individual service users. Dynamic traffic data such as real-time traffic flow, traffic congestion, and traffic incident information, as well as static data such as road network geometry information, network functional characteristics, and historical traffic data, are obtained from the transportation system. As for the multiple transit modes such as busses, the metro, and taxis, they provide real-time vehicle location data and other static data, such as transit and fare schedule information. Data from the transportation system and transportation vehicles provide real-time traffic situation information and assist in dynamic route planning; however, data from individual service users assists in learning the users’ travel habits and travel preferences. Data from users consist of dynamic travel request instructions and static user profile data, such as historical travel choices and historical travel trajectory information, which contains user preferences and personal information.

3.2. Data Platform

A data platform, which is defined as the data storage and data processing section of equipment, devices, systems, etc., is expected to offer high availability, high performance, fast queries, and replication facilities with data redundancy mechanisms. According to functional considerations, a data platform can be divided into three layers: the API, data adapters, and data functions.

Data adapters provide interfaces and libraries for data collection, transformation, and cleaning, as well as storage. Proprietary and standardized data collected from different sources will be adapted and coordinated according to the requirements of the different components [33]. The adaptor should be able to collect data in different formats (e.g., XML, JSON, and text), and transform them for use by other components.

An application programming interface (API) is responsible for the integration of different components through a single interface. External components can integrate their APIs using a set of rules specified by regular models and standards. Internal components allow seamless access to the central repository and facilitate interfaces between components as needed.

Data functions are a set of data processes, such as data collection, data extraction, data fusion, and data analysis. Different components use data for different purposes; for example, roadside monitoring equipment will collect data and personal information devices will analyze data, and data are the basis for the analysis of support services and functions.

To be specific, data collection in a FPMS is normally carried out under two paradigms, as shown in Figure 2. Common data such as traffic data and vehicle data, which have little potential risks, are collected directly for system optimization. As user data are sensitive information with personal privacy concerns, the FPMS processes personal mobility data in a way that preserves privacy, by uploading only desensitized parameters to the service center. This method is an improvement compared to a traditional PMS, which requires personal data to be uploaded. The FPMS enables a novel way of optimizing mobility demands and supply by integrating multi-modal information from both open and private data sources, while maintaining privacy.

3.3. Back-End Service

Traffic data (e.g., traffic conditions, road network geometry, and network functional characteristics), vehicle data (e.g., vehicle location, vehicle types, and vehicle schedules), and user data (e.g., travel requests and travel preferences) are the fundamental inputs for the PMS implementation, which is supported by sub-services: a dynamic route planner (DRP), and personalized option recommendation (POR). As a back-end service, the DRP and POR work together to provide personal, dynamic, and optimal route options to mobility hubs, through various functions characterized as information modules, to complete specific tasks. According to ATS characteristics, all functions can be classified into four categories: autonomous perception, autonomous learning, autonomous decision-making, and autonomous control, as illustrated in Figure 3.

Autonomous perception encompasses the sensing functions that actively acquire information. To ensure user privacy, these functions can operate in two modes: a public mode, and private mode. In the public mode, data such as dynamic traffic data can be obtained with user consent or open accessibility; and in private mode, desensitized aggregation parameters can be exchanged.

Autonomous learning involves learning functions, such as data fusion and option optimization, which also operate in two modes to protect user privacy. In public mode, the cloud can obtain the running status of the system by integrating heterogeneous data from multiple sources. In private mode, the learning functions operate by training the travel choice model locally and uploading the local model to the cloud, creating a shareable global model, while preserving user privacy.

Autonomous decision-making includes rearranging functions to automatically generate personalized travel options. First, global travel options are generated that optimally consider real-time traffic conditions. Second, a personalized travel option matching user preferences is generated.

The autonomous control functions are responsible for executing the selected travel options and are closely associated with the learning functions, to form a feedback loop. These feedback loops enable continuous iteration and facilitate improvement in the service quality and user experience, ultimately measuring the level of PMS service provided to users.

3.4. Front-End User Interface

As shown in Figure 4, the front-end UIs of the FPMS are implemented for users to input travel requests, view route recommendation options, and receive navigation guidance information. In order to better analyze and document the capabilities of the FPMS and dependencies between components, it is considered important to describe the capabilities of FPMS applications. When starting a trip, users can quickly access the personal trip menu via a smartphone app to input the travel destination. According to the travel request information, multiple routes labeled with time, cost, mode, walking distance, and the level of recommendation will be recommended, and concrete route options with real-time traffic conditions shown on maps are displayed by the UI (see Figure 4b,d). In general, the level of recommendation represents the level of personalization and expresses user preferences and travel habits. When choosing each option, users can view the concrete content of the chosen option; for example, the specific metro line information, with the arrival time (see Figure 4c) and the concrete information of the chosen option on a map (see Figure 4e).

4. Proposal Framework of the FPMS

4.1. Architecture

To ensure a consistent design and implementation of the FPMS across ATSs, it is important to develop an FPMS architecture that captures all features. This section proposes and discusses the reference architecture for the FPMS, which includes both physical and logical views. The logical architecture of the FPMS describes the data flow and information interaction among different modules, providing guidelines for the physical architecture. In contrast, the physical architecture maps the function modules in the logical architecture to real-world traffic entities, establishing various information interactions and data flows.

4.1.1. Logical Architecture of the FPMS

The logical architecture of the FPMS mainly the concerns data flow between functions, which follows the ATS operational logic of autonomous perception, autonomous learning, autonomous decision-making, and autonomous control, and defines eight working steps in the logical flow (see Figure 5).

The data collection step involves the acquisition of data from sensors, roadside cameras, and other related sensing devices. This step includes collecting traffic data from roadside monitoring equipment and other sources.

The data extraction step involves extracting key information from various types of data. This includes identifying user preference information through user personal data and extracting other relevant data from different sources.

The data fusion step involves aggregating data in different modes and dispatching the aggregated data to related functions. This step includes aggregating various traffic data and vehicle data for system optimization and integrating data from multiple sources to produce more accurate and insightful results.

The data analysis step involves analyzing the data to train a travel choice model locally. This model is then uploaded and exchanged, to create a shareable global model.

The option generation step utilizes the global travel choice model and global travel optimization model in the cloud, to generate customized travel choices for users.

The option personalization step describes the global choices generated in the previous stage and is customized according to user preferences.

The option execution step involves helping travelers make informed decisions about their travel choices through an interactive user interface.

The option feedback step describes the selection feedback process of collecting user experience information through selection feedback for continuous improvement.

In the logical architecture of the FPMS, the information interaction and data flow between connected functions is commonly realized through exchanging four types of data: open data, consent data, desensitized data, and operation data.

Open data are obtained from open sources, such as road network geometry information, network functional characteristics, and transit vehicle schedule information.

Consent data contains data about user consent, such as travel OD and user selection feedback information.

Desensitized data refers to personal data that have been processed to remove sensitive information, such as user preferences and travel request data.

Operation data include the local and global choice models generated during the training step and other data generated during the operation of the FPMS.

4.1.2. Physical Architecture of the FPMS

The physical architecture of the FPMS refers to the data flow among physical entities, transferring the operation logic into the real world (see Figure 6). The physical architecture of the FPMS is based on physical objects (PO), which refer to the physical traffic entities in the physical view and can be categorized into three groups: entity, module, and system.

The entity is the most basic PO in the ATS and is the primary information source and interacts with the internal functions provided by the system. Entities in the FPMS mainly include users (e.g., commuters, tourists, etc.), vehicles (e.g., private cars, public vehicles, etc.), and roads (e.g., road networks, road environment, etc.). In general, the user consumes the service, the vehicle supports the movement of travelers, and the road may influence the service indirectly.

A module is an intermediate physical object that connects entities and systems and provides basic functions that are utilized at the edges. In the context of the FPMS, the module includes different equipment and devices, such as roadside monitoring equipment (RME), vehicle on-board equipment (OBE), and personal information devices (PID). Each of these modules has a specific function: the RME collects road conditions, the OBE collects transit vehicle information, and the PID serves as a common interface for travelers to interact with the system using personal travel choice models (also called travel utility functions [34]), which are generated through federated learning and can generate personalized travel options for the users.

The system is a compound physical object that implements the complex functions required by the service. In the FPMS, the system includes different centers such as the traffic management center, transit management center, and the personalized recommendation center (PRC). These centers work together to provide efficient and personalized travel options to the users. The traffic management center senses and manages real-time traffic conditions and traffic control information. The transit management center manages multi-mode vehicles on the road and provides various services to users. The PRC maintains the user travel choice model learned from federated learning, optimizes the running of the system, and measures system objectives and user preferences, to maximize the overall savings and enhance the user’s travel experience.

4.2. Algorithm

As outlined above, a suitable travel choice model is the key to achieving a personalized travel recommendation, distinguishing the FPMS from other studies. A high-explainability customized model that integrates FL and discrete choice model (DCM) was previously been developed [35]. To ensure both prediction accuracy and model interpretability, in this section we estimate a multinominal logit (MNL) model as the travel choice model with the choice data from users, where $X=\{x_1,x_2,\dots ,x_K\}$ represent the explanatory variables indicating the observed attributes of the choice alternatives and the individual socio-demographic characteristics, and $\mathcal{Z}=\{{\zeta }_1,{\zeta }_2,\dots ,{\zeta }_k\}$ are the preference parameters associated with the K explanatory variables in the vector $X_{nj}$ that correspond to the explanatory variables of individual n. $ϵ$ is introduced to account for the residual uncertainty in the selection choice context, which fully captures the modeling process. The utility ($U_{nj}$) where individual n associates with alternative j from their choice set $C_n$ is given as Equation (1).

$$U_{nj}=\mathcal{Z}{X}_{nj}+ϵ$$

(1)

We adopted an implementation of an MNL as an artificial neural network (ANN) based on [36], using a simple 2D convolutional neural network (CNN) architecture. The input space X was arranged as a two-dimensional matrix of size $J\times K$, and a convolution filter $\mathcal{Z}$ of shape $(1,K)$ was defined as a set of trainable weights. Subsequently, a CNN with a single convolutional layer was considered. The model slid along one of the K variables of the input X, i.e., choice alternatives j, and calculated the dot product between X and $\mathcal{Z}$, to yield a single scalar value. This mapped X to the output space V of shape $(J,1)$, so that

$$V_j=\sum _{k=1}^K{\zeta }_k{\left(X_k\right)}_j$$

(2)

Therefore, the aforementioned process is equivalent to the linear part in Equation (1). Next, we propagated the output V to the activation layer, which employs the softmax activation function. This function corresponds to the probability of individual n selecting alternative j under standard assumptions in the MNL framework, expressed as

$$P_{nj}=\sigma \left(V_{nj}\right)=\frac{{exp}^{V_{nj}}}{{\sum }_{j^{\prime }\in C_n}{exp}^{V_{n{j}^{\prime }}}}$$

(3)

Finally, cross-entropy is employed as the loss function during the training process to optimize the model’s weights $\mathcal{Z}$ through backpropagation.

Based on the above model, the interpretability and accuracy of the model can be further improved. According to [37], X is a set of continuous explanatory variables and Q is a set of categorical explanatory variables, where $Q=\{Q_1,Q_2,\dots ,Q_M\}$. They represent the observed attributes of the choice alternatives and the individual’s sociodemographic characteristics. Thus, the utility of option j for individual n is given by Equation (4):

$$U_{nj}=\mathcal{Z}{X}_{nj}+{\mathcal{Z}}^{\prime }{E}_j(Q_n;W_j)+ϵ$$

(4)

The function $E(·)$ is defined as an embedding function, which maps each one-hot encoded input $Q_n$ to a latent one-dimensional representation $Q_n^{\prime }$ based on a set of trainable alternative-specific weights $W_j$, where $Q_{nj}^{\prime }=E_j(Q_n;W_j)$. The size of the embedding matrix $\mathbf{W}$ is $S\times J$, where S is the total number of unique categories across all the explanatory categorical variables Q. Therefore, $Q_n$ is projected into the embedding layer that maps each input to a vector of dimensionality $D=J$. Then, the new alternative-specific representations are concatenated with a second convolutional filter and a set of trainable weights ${\mathcal{Z}}^{\prime }$, where the shape of ${\mathcal{Z}}^{\prime }$ is $(1\times M)$.

By associating each embedding dimension with an alternative, the embedding matrix reflects the correlation between different attributes and alternatives. Therefore, the embedding matrix is an interpretable embedding vector, where the embedding value along the j-th dimension represents the correlation with the encoding category of the j-th alternative scheme. In addition, the value of ${\mathcal{Z}}^{\prime }$ is restricted to positive. We can determine the direction of correlation between this attribute and the j-th alternative scheme (positive or negative correlation) according to the notation of the value of the j-th dimension in the embedding matrix.

Ultimately, in [37], to combine the model proposed in [36], the dimension of the above-embedded matrix is increased from J to $J+A$, and the additional dimension A is used as the input to $r_j$. In [36], $r_j$ is the resulting function of a dense neural network (DNN) and is referred to as a representation learning term. For simplicity and convenience, we define the input to the function r as $R_n=\{Q_{J+1,n}^{\prime },\cdots ,Q_{J+A,n}^{\prime }\}.$ Hence, $V_{nj}^{\prime }$ can be expressed as Equation (5).

$$V_{nj}^{\prime }=\mathcal{Z}{X}_{nj}+{\mathcal{Z}}^{\prime }{Q}_{nj}^{\prime }+r_j(R_n;{\mathbf{B}}_j,{\alpha }_j)$$

(5)

where ${\mathbf{B}}_j$ is an alternative-specific trainable parameter, and ${\alpha }_j$ is a bias item. Next, a rectified linear unit (ReLU) is used as the activation function. Therefore, for a single-layer neural network, $r_j$ with H neurons can be written as Equation (6):

$$r_j(R_n;{\mathbf{B}}_j,{\alpha }_j)=\sum _{h=1}^H{\beta }_{jh}(max(0,R_n))+{\alpha }_j$$

(6)

Therefore, the probability that individual n chooses alternative j is naturally expressed as Equation (7). The architecture of the final model is shown in Figure 7, where the model frames of $\mathcal{Z}{X}_{nj}$, ${\mathcal{Z}}^{\prime }{Q}_{nj}^{\prime }$, and $r_j(R_n;{\mathbf{B}}_j,{\alpha }_j)$ correspond to the parts “Utility 1”, “Utility 2”, and “Learning Term”, respectively, in Figure 7. Finally, cross-entropy is implemented as the loss function in the training process.

$$P_{nj}=\frac{{exp}^{V_{nj}^{\prime }}}{{\sum }_{j^{\prime }\in C_n}{exp}^{V_{n{j}^{\prime }}^{\prime }}}$$

(7)

For a FPMS, the objective is to estimate individual-level preference parameters, without compromising the users’ privacy. Hence, we incorporate federated learning into the ANN training process. The federated ANN process is conducted collaboratively by the PID and PRC, as follows:

(1)

PID (in parallel): Training an individual ANN model based on personal information and the global ANN model;

(2)

PID (in parallel): Uploading individual ANN models;

(3)

PRC: Obtaining the global ANN model, which is obtained by aggregating the ANN models of different PIDs;

(4)

PRC: Broadcasting the global ANN model to all PIDs.

After sufficient communication iterations, the individual ANN model can be estimated. The individual ANN model primarily consists of the weights or bias values of different modules in the ANN, without including the original individual information, to protect user data privacy. Meanwhile, during the process of aggregating the global ANN model, we adopt the widely-used federated average algorithm (FedAvg) [8] to calculate the global parameters.

5. Evaluation

This section analyzes the performance differences between the FPMS and the traditional PMS using the same evaluation dataset and scenarios. Figure 8 describes the framework differences, where the traditional PMS transmits data and the FPMS transmits parameters. Meanwhile, the size of the parameters is much smaller than the size of data, which shows the advantages of the algorithm and performance improvements of the FPMS. We compared the differences in the following four indicators: training time, resource consumption, training loss, and learning accuracy.

5.1. Preparation

For fairness, we configured standard settings for the experiments, including the evaluation dataset, models, scenarios, and indicators.

5.1.1. Evaluation Dataset

For this experiment, we used the Swissmetro dataset [38], a publicly available real-world dataset collected in 1998 on trains between St. Gallen and Geneva, Switzerland, as the evaluation dataset. The dataset has sufficient surveys of each respondent to justify the methodology of estimating individual-level preferences. For each respondent, nine hypothetical choice tasks are presented, and each task contains three alternatives: train, Swissmetro (SM), and private car, along with other attributes of the task, including the differential travel cost, travel time, etc.

The dataset contains 10,728 observations from 1192 respondents. To better utilize the data, we had to preprocess the dataset and identify the relatively valuable data. Therefore, we excluded respondents in the following categories:

• Respondents with missing values representing their choices of the three alternatives;
• Respondents who lacked information representing their attributes in the selected category explanatory variable.

After the data preprocessing, the cleaned dataset contained 1004 respondents. Then, for each respondent, we assumed that there was a virtual client, where the first eight choices of the respondent were the training data set, and the ninth choice was the test data set to train and evaluate the models.

5.1.2. Models

To better compare the performance of the traditional PMS and FPMS, in this experiment, we used the following three ANN models, implemented with the PyTorch [39] framework, as travel choice models for both the traditional PMS and FPMS. The three models and their associated hyperparameters were as follows:

• L-MNL (as presented by Sifringer et al. in [36]), with L = 1, H = 100;
• E-MNL (as presented by Arkoudi et al. in [37]), with S = 100, D = J = 3;
• EL-MNL (as presented by Arkoudi et al. in [37]), with S = 100, D = J + A = 3 + 2, H = 15.

For the input feature set of the three models, we followed the input in [37]. X = $\{AS{C}_{Car}$, $AS{C}_{SM}$, TT, TC, HEADWAY} and the remaining 12 variables constitute the feature set Q that are projected to learning term in L-MNL and the embedding layer of E-MNL and ELMNL, where ASC represents the alternative specific constant, TT and TC represent the travel time and cost, and HEADWAY represents the transportation headway, respectively.

5.1.3. Scenarios

To simulate real-world travel selection tasks, namely the work of a PMS system, we virtualized a communication and computing environment with a central server and 1004 clients. For the server, the total network bandwidth was 15 MHz and the CPU cycle frequency was 2.5 GHz. For clients, the transmission power ranged from 0.2 to 1 W, and the CPU cycle frequency ranged from 1.5 to 2 GHz. In addition, we designed a static estimation scenario that was applied to both the traditional PMS and FPMS. In this scenario, it was assumed that all the data had been generated and all the users had been loaded. During the estimation process, default settings of the Adam optimizer [40] were used for all models, with 300 rounds of training for each model. The converged model was ultimately used to estimate the travel choices of the users.

5.1.4. Evaluation Indicators

To comprehensively assess the performance of the traditional PMS and FPMS, we selected four general indicators, namely

• Computation Time (CT): CT is the cumulative time of trained rounds, including both the time spent at the server $T_s$ and all clients $T_c$, as defined in Equation (8). It should be noted that, for the traditional PMS, the calculation time on the client was negligible, so $T_c=0$; For the FPMS, since clients work in parallel, we treated the maximum elapsed time among all clients as $T_c$, so $T_c=max\left(\{T_c^1,T_c^2,\dots ,T_c^n\}\right)$.

  $$CT=\sum _{r\in [1,R]}(T_s+T_c)$$

  (8)
• Resource Consumption (RC): The resource consumption process is mainly communication resource consumption, which can be divided into uplink resource consumption $R{C}_{up}$ and downlink resource consumption $R{C}_{down}$. For the traditional PMS, the main communication resource consumption process is uplink consumption. The communication time in the downlink is ignored, since the time consumed in the downlink is negligible compared to the one of uplink [41], so $R{C}_{down}=0$. For the FPMS, communication resources are consumed when the client uploads individual model parameters and the server distributes global model parameters. Therefore, for the two the PMSs, the communication resource consumption is shown in Equation (9).

  $$RC=\sum _{r\in [1,R]}(R{C}_{down}+R{C}_{up})$$

  (9)
• Training Loss (TL) and Learning Accuracy (LA): TL and LA are the two most commonly used indicators to evaluate the performance of ANN models, as shown in Equations (10) and (11), respectively. In which, cross-entropy is employed as the loss function, the number of clients is N, and $y_i$ and $p_{ij}$ respectively represent the true choice of client i and the probability of alternative j being chosen by client i. Subsequently, we take the choice with the highest probability as the predicted choice $\widehat{y_i}$ for customer i, and if $\widehat{y_i}=y_i$, output 1, otherwise 0.

  $$TL=\frac{1}{N}\sum _{i=1}^{N}L(y_i,p_i)=-\frac{1}{N}\sum _{i=1}^N\sum _{j=1}^J{1}_{\{y_i=j\}}log\left(p_{ij}\right)$$

  (10)

  $$LA=\frac{1}{N}\sum _{i=1}^N{1}_{\{\widehat{y_i}=y_i\}}$$

  (11)

5.2. Results

Based on the aforementioned preparations, the employed model was implemented in a defined static estimation scenario and executed separately under both the traditional PMS and FPMS. The resulting outcomes were then grouped, evaluated, and discussed individually.

First, we compared the TL and LA of the three models under the traditional PMS and FPMS. From Figure 9, it can be seen that when the model was trained to convergence, the performance of the two PMSs in terms of accuracy was not much different.

To reflect the differences in the TL and LA between the two PMS in more detail, we present the values for the TL and LA in the optimal case under the three models. We also selected data from the last 20 rounds, to calculate the mean TL and AL for comparison. For the L-MNL model, the FPMS improved by 1.74% in LA but decreased in E-MNL and EL-MNL. Meanwhile, compared with the traditional PMS, the performance of FPMS for TL was also slightly lower, but the decrease was less than 2%. Details can be obtained from

Table 1. Results for TL and LA under the traditional PMS and FPMS.

Model	Traditional PMS				Federated PMS				Comparison
	TL		LA(%)		TL		LA(%)
	Mean	Min	Mean	Max	Mean	Min	Mean	Max	$\Delta$TL	$\Delta$LA
L-MNL	270.54	270.40	63.63	64.45	274.07	273.89	64.74	65.12	1.30	1.74
E-MNL	268.56	268.54	66.80	67.11	268.91	268.90	65.78	66.28	0.13	−1.52
EL-MNL	264.49	264.19	67.74	68.44	268.14	268.01	67.02	67.44	1.38	−1.05

The best value among the compared methods is bolded, and the second best value among the compared methods is underlined.

, where $\Delta TL$ and $\Delta LA$ are defined as shown in Equation (12).

$$\Delta TL=\frac{(T_{FPMS}-T_{TPMS})}{T_{TPMS}}\Delta LA=\frac{(L{A}_{FPMS}-L{A}_{TPMS})}{L{A}_{TPMS}}$$

(12)

Then, we simulated the working mode of the PMSs to calculate CT and RC, as shown in Figure 10. It can be clearly seen that the FPMS was less time-consuming than the traditional PMS in terms of computation time and resource consumption. This was because the traditional PMS needs to upload gradually accumulated raw data to the server in batches, while the FPMS adopts a parallel mode, so that all clients can run at the same time, reducing the time needed for the accumulation process. In addition, compared to the original data uploaded by the traditional PMS, the FPMS uploads the parameters of the model, which results in a significant difference in the scale of data. Therefore, the traditional PMS consumes more communication resources.

In summary, it can be seen from the evaluation results of the two PMSs that using FL in the FPMS can, not only protect user privacy and improve the utilization rate of edge resources, but also significantly improve model performance in RC and CT, all while ensuring model accuracy. Specifically, the FPMS proposed in this paper has outstanding performance in the following aspects:

• It employs the advantages of FL to better leverage the heterogeneity among users in a collaborative and privacy preserving manner;
• It can approximate the accuracy of the traditional PMS model, meaning it can approximate the effect of the traditional model;
• In static conditions, the FPMS could reduce the CT by about 60% and RC by about 85% for each of the three different models compared to the traditional PMS, greatly increasing the system speed.

5.3. Discussion

We discuss the effects of the three ANN models implemented in the two PMSs and the advantages and disadvantages of the FPMS compared to the traditional PMS from the perspective of three kinds of stakeholder: an individual traveler, system modeler, and service manager.

5.3.1. Perspective of Individual Travelers

In practical applications, travelers can obtain more accurate recommendations through the traditional PMS, but the FPMS can guarantee accuracy and obtain the equivalent services to the traditional PMS more quickly. As can be seen from the above experimental results, although the FPMS was less accurate than the traditional PMS, it could significantly reduce the calculation time and communication resources of the system and ensure that the gap between the accuracy of the FPMS and the traditional PMS was within an acceptable range.

In addition, in terms of data, the traditional PMS needed PID to upload individual data points to the server during operation, to train and aggregate the travel recommendation model. However, the FPMS can upload the desensitized data or the model parameters trained by the PID after local processing according to the local data of the PID and the corresponding model. The server then uses the data it receives to aggregate the global model. Compared with the traditional PMS, the FPMS no longer requires travelers to upload their original data directly, which reduces the risk of data leakage to a certain extent and protects the privacy of individuals. Nevertheless, the transmission of these parameters on a public network is also easy be monitor, which will lead to the leakage of key parameters. These key parameters are still relatively fragile in the existing FPMS algorithm, which could be used to infer the sensitive information of users. Hence, it is necessary to study parameter protection or encryption mechanism for the FPMS in the future, to mitigate the related risks further [42,43].

5.3.2. Perspective of System Modelers

According to the above experiment, it is evident that the different models showed inconsistent performance for travel recommendations. At present, there are many excellent travel recommendation models and algorithms in the traditional PMS, but for the FPMS, continuous research is still needed. In this experiment, FL was combined with three highly interpretable neural networks, and FedAvg was used in the aggregation of the global model. Finally, in the static estimation scenario, the accuracy of the three models was not much different with the traditional PMS and FPMS.

For the FPMS, different PIDs need to train different traveler travel selection models locally. Since the sample size of individual data is small, some methods for processing small sample data can be considered. When the PID trains individual-level parameters with limited individual data, different individual preferences can be reflected in the parameters, thus improving the performance of the FPMS.

In addition to the model algorithm, attention should be paid to the aggregation mechanism of the model. In this experiment, FedAvg [8] was used. After the server receives the model parameters from the different PIDs, only the most straightforward sum of the different parameters is averaged. This approach simplifies the calculation method, but it causes the model to eventually move towards the average of the population level, which eliminates the preference of different individual levels in the model, so that more advanced aggregation mechanisms can be developed (for example, specific weighting functions can be used to aggregate the model participation according to different individuals).

5.3.3. Perspective of Service Managers

For the traditional PMS, with the continuous growth of user volumes, the central cloud may become a performance bottleneck for an immediate response, as it undertakes the majority of functionalities to support the PMS [44]. However, the central cloud of the FPMS only receives the parameters of the model. Compared with the central cloud of the traditional PMS, this can not only reduce the consumption of communication resources between the PID and PRC but also relieve the pressure of receiving, storing, and processing large amounts of data in the central cloud.

Meanwhile, three models with high interpretability were used as PMS algorithms in this experiment. PMS managers can understand the contribution of different attributes to the different user travel choice models using the results of the model operation. In the subsequent data collection process, they can prioritize the data with a high contribution to improving the accuracy of the corresponding data.Thus, the results of the model are more consistent with reality. In addition, for data with a low contribution, we can consider removing it, to reduce the pressure on the central cloud. The FPMS manager can select a different number of attributes in the process of individual model training, according to different PID states (normal mode or power saving mode) for different users, so as to reduce the consumption of resources in the process of PID training, while ensuring the accuracy of the model.

6. Conclusions

To support personalized travel services based on real-time traffic conditions in the ATS environment, a PMS has been proposed. It can comprehensively perceive the macro-system status, accurately learn micro-user behavior, assist in decision-making for designated service plans, and ultimately provide personalized solutions. To realize such a PMS, this article proposed a novel federated PMS that uses the privacy preserving decentralization mechanism of FL to coordinate service participants in the PMS. To illustrate the external behavior and internal structure of the FPMS in the ATS environment, demonstrate the FPMS’s effective protection method for user privacy sensitive data, as well as describe the maintenance and improvement of the FPMS in performance, this paper gave an overview of the FPMS; put forward the FPMS architecture and critical algorithms; and finally, through a performance evaluation, the performance difference between the FPMS and the traditional PMS was analyzed.

An overview of the FPMS can be presented from two perspectives: the physical and the information viewpoints. The physical viewpoint is the basic input to the FPMS, which perceives information from the physical world. The information viewpoint is the transmission and presentation of information through the data platform, backend services, and UI. These three platforms work together to support the functions of the FPMS, to meet user needs. Next, from both physical and logical perspectives, an FPMS architecture that captures ATS characteristics was proposed and discussed, and key algorithms for the FPMS were presented in conjunction with ANN models. The logical architecture mainly represented the information interaction and data flow between functional modules.

In contrast, the physical architecture mapped the functional modules in the logical architecture to the real world, forming various information interactions and data flows between the different traffic entities. The algorithm mines the information input that contributes to travel recommendations for travelers and assists in formulating travel plans. The experimental results showed that, for the three different models, the FPMS and PMS had little difference in the accuracy of their results in the static scenario. However, the FPMS could reduce the calculation time by about 60% and the communication resource consumption by about 85%.

Nonetheless, in terms of model performance, the FPMS is still slightly behind the traditional PMS. Moreover, we only used a synchronous mechanism to interact with the model parameters, which may have led to the issues of overfitting or wasting communication resources, thus impeding the application of the FPMS [45]. In the future, to apply the FPMS proposed in this paper, the following aspects could be improved: (1) In terms of the uploading mechanism, when there are few users in the FPMS, all users can be selected to participate in the training of the model. Federated learning allows multiple participants to collaborate to train an efficient model, without exposing data privacy. However, this distributed machine learning training method is vulnerable to Byzantine clients [46,47], which upload dirty data, such as modified models or false gradients to interfere with the training of the global model. Therefore, we could study the parameter uploading mechanism and select suitable parameters for the training process. This would reduce contamination of the model with dirty data, improve the learning performance, and reduce the consumption of resources; (2) In terms of the interaction mechanism, a synchronous training interaction mechanism was adopted in this paper, and an asynchronous mechanism could be studied in the future to accelerate the updating speed of the model and improve the robustness; (3) In terms of the aggregation mechanism, although FedAvg can guarantee the accuracy of the model, and its simple algorithm can reduce the consumption of computing resources, it is easy to make the model biased to the overall average level. To better show the individual preferences of different travelers, it is necessary to further explore the aggregation mechanism.