Individualised Multimodal Trip Planning for Mobility as a Service

Multimodal mobility services offer alternatives to private car use, yet conventional journey planners typically rank routes using generic criteria and do not reflect individual travel preferences. This study develops and tests an individualised multimodal trip planner integrated into a Mobility as a Service (MaaS) application in Thessaloniki, Greece. The planner combines public buses, walking and shared electric cars, bicycles and e-scooters, including services associated with five shared mobility hubs and supplementary virtual e-bike stations. 

The approach first identifies 23 feasible unimodal and multimodal trip chains and evaluates alternatives according to travel time by mode, overall monetary cost and the number of transfers. Route calculations rely on OpenTripPlanner, while public transport routes and schedules are incorporated using General Transit Feed Specification (GTFS) data. A stated-preference survey of 200 respondents, undertaken in March 2023, informed a binary logistic regression model from which a generalised cost function was estimated. Because bus information was unavailable during the initial survey stage, the public transport travel-time coefficient was subsequently integrated using results from a related Thessaloniki study. 

To personalise recommendations, the researchers introduce dynamically updated mode-specific adaptation factors that reflect users’ recorded histories of trips made with shared cars, bikes and scooters. The approach is evaluated through four constructed test-user profiles with different hypothetical trip histories and route-comparison scenarios. Results demonstrate that the adaptation factors can modify journey rankings, often favouring shared micromobility options when historic choices indicate preferences for these modes despite longer travel times. Nevertheless, route proximity and the availability of shared mobility services continue to influence which itineraries are practical. 

The study illustrates how integrating travel-history information into a generalised cost framework can support adaptive, user-oriented MaaS trip recommendations without requiring explicit preference settings. The evaluation demonstrates changes in algorithmic rankings rather than measured increases in sustainable travel, user satisfaction or public transport ridership. Further field evaluation and richer data integration are needed to assess real-world impacts and expand personalisation to public transport legs.

Supporting evidence

The personalised planner was developed for Thessaloniki’s 2023 MaaS pilot, which included five shared mobility hubs and virtual e-bike stations. It evaluates 23 potential unimodal and multimodal trip combinations and uses OpenTripPlanner to calculate travel times. Its generalised cost function considers mode-specific travel time, monetary cost and transfers. 

A stated-preference questionnaire completed by 200 participants in March 2023 supplied four hypothetical travel-choice scenarios per respondent. A binary logistic regression estimated the initial weights (McFadden’s pseudo-R²=0.182). The estimated transfer coefficient (−3.925) indicates a substantial penalty for switching modes within the survey’s specification. Because bus trips were excluded from the initial choice scenarios, their travel-time parameter was introduced subsequently using an existing local study. 

Four illustrative test-user histories were then constructed: car-oriented, bike-oriented, scooter-oriented and mixed-mode. Their calculated adaptation factors included 0.888 for bike-sharing in the bike-oriented profile and 0.920 for scooter-sharing in the scooter-oriented profile (Table 2). Route examples show that a generic recommendation of bus plus walking could change to bike-sharing plus walking for a personalized profile (Figure 4). In another journey, the leading car-sharing-and-walking route remained unchanged, but secondary options changed (Figure 5). 

These are algorithm demonstrations, not observed modal shifts. The comparison relies on selected origin–destination examples, not a citywide deployment evaluation. 

Key findings

1. Personalisation can change recommended journey rankings. The proposed method adjusts generalized travel-cost calculations using mode-specific factors derived from previously chosen shared-mobility modes. Rather than asking users to declare their preferences, the approach infers them from trip history. Demonstrations show that differently constructed profiles can receive different recommendations for the same origin and destination. 
2. Travel time, monetary cost and transfers all matter. The generalized cost function incorporates time spent in different modes and the number of mode changes. Results from the stated-preference model suggest a pronounced disutility associated with transfers, an important design consideration for MaaS solutions in contexts where travellers are relatively unfamiliar with multimodal journeys. 
3. Shared micromobility is often favoured when prior choices reveal a preference. Profiles that select bikes or scooters despite longer journey times receive stronger personalization toward those services. The test results indicate particular potential to promote bike-sharing. This is a property of the proposed ranking mechanism and does not demonstrate that actual travellers subsequently switched modes. 
4. Feasibility and location still constrain individualisation. In one illustrated journey, personalisation changes the first-ranked itinerary from bus plus walking to bike-sharing plus walking. In another, shared-car access near both endpoints means car-sharing plus walking remains the top-ranked choice across profiles, although lower-ranked alternatives change. Recommendations therefore depend on service availability and the specific origin–destination pair. 
5. Better integration and field validation are priorities. The study identifies real-time data exchange, interoperability, integrated ticketing and potentially personalised incentives as areas for MaaS development. Current adaptation factors apply only to shared cars, bikes and scooters because corresponding histories are available; public transport history is not incorporated. Further research should test actual user responses and investigate broader preference variables and richer mode data. Their operational effectiveness also depends on cooperation among public and private mobility providers.