Travel behavior models are essential for transportation planning and policydevelopment, addressing challenges like traffic congestion, environmentalimpact, and equitable access. By analyzing how individuals make travelchoices, these models support decisions related to infrastructure investmentand resource allocation. These models cover various aspects of travel, including activity planning, route selection and travel time and areconstantly being revised. One of the key ground of improvements are theemergence of novel data sources, significantly advancing the understandingof travel behavior and overall transportation planning. This thesis fitswithin the stream of studies that investigates travel behavior modelsusing novel data sources to guide policies that enhance mobility, supportsustainability, and promote equity in transportation systems, by means of 4distinct papers.

Paper 1 focuses on adapting mobile network data to Scaper, a dynamicdiscrete choice model. The Scaper framework, originally designed foractivity generation and scheduling based on travel survey data, is tailoredto handle the big data source and adapt accordingly. The study developsprobabilistic models by integrating observed and latent states to infertrip attributes from cell tower observations. It employs a backwardinduction method to compute the expected value function, using StochasticExpectation-Maximization for parameter estimation. This paper offersa methodological contribution, demonstrating the potential of how toeffectively adapt an activity based model Scaper to new data sources. Toillustrate the usefulness of this framework, we emphasize its application in Paper 2. This new framework is used for assessing mobility inequality andsegregation before and during COVID-19 in Stockholm. This shows howwe can use these models and data to further investigate mobility patternsduring times of crisis and to envision a more resilient transport system thatpromotes equity.

In line with the thesis’s scope of integrating sustainability into research,we use route choice models and GPS traces to investigate cycling behavior. Paper 3 primarily focuses on cyclists’ route preferences in the Netherlands.Notably, cyclists, including commuters, do not always choose the shortestpath. Instead, various factors influence their decisions, raising the importantquestion: how can we design infrastructure that aligns with cyclists’preferences and encourages more frequent cycling? the detailed GPS tracesallowed for to investigate various aspects of the route beyond distance,for instance, number of junctions, traffic lights, presence of nature, etc.This paper utilizes two approaches to address this question. The firstis theory-driven based on logit models, the Path Size Logit (PSL) andthe Pairwise Combinatorial Logit (PCL), both rooted in random utilitymaximization principles and designed to account for route overlap amongchoices. The second is a data-driven approach using deep learning topredict route choices through a one-dimensional Convolutional NeuralNetwork. We conducted a sensitivity analysis to uncover key patternsin the deep learning model, offering insights into the factors influencingroute preferences. By comparing these two approaches, we emphasize theirstrengths and limitations while showing how GPS data integrates with themto uncover key factors influencing cyclists’ route choices. This paper guidespolicymakers in designing efficient and appealing cycling routes.

Paper 4 expands the scope by incorporating GPS data alongsidesociodemographic information to examine cycling behaviors, particularlyin a cross-border context. Data were collected from three cities, namely,Braga, Istanbul and Tallinn. The focus is on travel time: What are theaverage and range of travel time for cyclists in different cities? How dofactors such as age, and gender influence travel time? Are there differencesbetween different cities? Travel time is a crucial variable for travel demandiimodeling but more so for cyclists, as they do not always prioritize speed. Alonger trip isn’t necessarily worse; it might even be preferred if the shorteralternative is more exhausting. Novel data sources like GPS traces collectedover period of months in three different cities provides the opportunity tounderstand these complex and comparative behavioral contexts. Cyclingunderscores not only the value of time but also the quality of time spentengaging in the activity. It’s within this context that travel time modelingbecomes particularly important to investigate. Using a survival analysisapproach, specifically the Latent Class Accelerated Failure Time (LCAFT)model, Paper 4 reveals how distance, trip purpose and bike type influencethe travel time of cycling and identifies potential latent classes in differentage groups and gender.

Modeller för resebeteende är avgörande för transportplanering ochpolicyutveckling, eftersom de hanterar utmaningar som trafikstockningar,miljöpåverkan och rättvis tillgång till transporter. Genom att analyserahur individer fattar resebeslut stödjer dessa modeller beslut kringinfrastrukturinvesteringar och resursfördelning. Modeller för resebeteendeomfattar olika områden, inklusive aktivitetsplanering, ruttval och resetidenslängd. Dessa modeller förbättras i takt med att nya datakällor blirtillgängliga, vilket leder till en ökad förståelse för resebeteende. Dennaavhandling undersöker resebeteende med hjälp av nya datakällor för attge vägledning för policy som förbättrar mobilitet, stödjer hållbarhet ochtransportjämlikhet.

Artikel 1 fokuserar på att anpassa mobilnätsdata till Scaper, en dynamiskdiskret valmodell. Scaper-ramverket, som ursprungligen utvecklades föraktivitetsgenerering och schemaläggning baserat på resvaneundersökningar,anpassas här till denna typ av storskaliga data. Studien utvecklarsannolikhetsmodeller genom att integrera observerade och latenta tillståndför att härleda reseattribut från mobilmastobservationer. En bakåtrekursivmetod används för att beräkna den förväntade värdefunktionen, ochparametrarna estimeras med hjälp av stokastisk Expectation-Maximization.

Denna artikel bidrar metodologiskt genom att visa hur Scaper effektivtkan anpassas till nya datakällor. För att belysa ramverkets användbarhetlyfter vi fram dess tillämpning i Artikel 2, där den utvecklade modellenanvänds för att analysera mobilitetsjämlikhet och segregation före och underCOVID-19 i Stockholm. Studien visar hur dessa modeller och data kanivbidra till att undersöka mobilitetsmönster i kristider och ge insikter omhur ett mer motståndskraftigt transportsystem kan utformas för att främjajämlik tillgång till transporter.

I enlighet med avhandlingens syfte att integrera hållbarhet i forskningenanvänder vi GPS-spår för att främja aktiva transportmedel, särskilt cykling.Artikel 3 fokuserar främst på cyklisters ruttpreferenser i Nederländerna. Detär värt att notera att cyklister, inklusive pendlare, inte alltid väljer denkortaste vägen. I stället påverkar olika faktorer deras beslut, varvid vi kanställa frågan: hur kan vi utforma infrastruktur som överensstämmer medcyklisters preferenser och uppmuntrar till mer frekvent cykling?

Denna artikel använder två metoder för att besvara denna fråga. Den förstaär en teoridriven metod baserad på logitmodeller: Path Size Logit (PSL)och Pairwise Combinatorial Logit (PCL), båda grundade i principernaför slumpmässig nyttooptimering och utformade för att ta hänsyn tillruttöverlapp mellan valmöjligheter. Den andra är en datadriven metod somanvänder djupinlärning för att förutsäga ruttval genom ett endimensionelltkonvolutiskt neuronnätverk (Conv1D). Vi genomförde en känslighetsanalysför att identifiera viktiga mönster i djupinlärningsmodellen, vilket gerinsikter i de faktorer som påverkar ruttpreferenser.

Genom att jämföra dessa två metoder betonar vi deras styrkor ochbegränsningar samtidigt som vi visar hur GPS-data kan integreras med demför att identifiera nyckelfaktorer som påverkar cyklisters ruttval.

Artikel 4 utvidgar perspektivet genom att inkludera GPS-data tillsammansmed sociodemografisk information för att undersöka cykelbeteenden,särskilt i ett gränsöverskridande sammanhang. Data samlades in från trestäder: Braga, Istanbul och Tallinn. Fokus ligger på restid: Vad är dengenomsnittliga restiden och variationen i restid för cyklister i olika städer?Hur påverkar faktorer som ålder och kön restiden? Finns det skillnadermellan olika städer?

Restid är en avgörande variabel inom modeller för reseefterfrågan, men ännumer så för cyklister, eftersom de inte alltid prioriterar hastighet. En längreresa är inte nödvändigtvis sämre; den kan till och med föredras om detvkortare alternativet är mer ansträngande. Nya datakällor som GPS-spår,insamlade under flera månader i tre olika städer, ger möjligheten att förstådessa komplexa och jämförande beteendemönster.

Cykling understryker inte bara värdet av tid utan också kvaliteten påtiden som spenderas på aktiviteten. Det är i detta sammanhang sommodellering av restid blir särskilt viktig att undersöka. Genom att användaen överlevnadsanalytisk metod, specifikt Latent Class Accelerated FailureTime (LCAFT)-modellen, visar Artikel 4 hur distans, resans syfte ochcykeltyp påverkar cyklisters restid och identifierar potentiella latenta klasserinom olika åldersgrupper och kön.

Over the past years, interest in classifying drivers' behavior from data has surged. Such interest is particularly relevant for car insurance companies who, due to privacy constraints, often only have access to data from Inertial Measurement Units (IMU) or similar. In this paper, we present a semi-supervised learning solution to classify portions of trips according to whether drivers are driving aggressively or normally based on such IMU data. Since the amount of labeled IMU data is limited and costly to generate, we utilize Recurrent Conditional Generative Adversarial Networks (RCGAN) to generate more labeled data. Our results show that, by utilizing RCGAN-generated labeled data, the classification of the drivers is improved in 79% of the cases, compared to when the drivers are classified with no generated data.

Dynamic Discrete Choice Models (DDCMs) allow us to analyze trip attribute choices within an activity-based framework where the sequence of activities matters. These models have been primarily developed for travel survey data. However, the advent of big data presents new opportunities for mobility analysis. This paper bridges traditional approaches with novel data sources by adapting an established activity-based DDCM to handle anonymized mobile network data. The proposed framework advances methods within transportation planning in two ways. First, dynamic models allow for detailed assessments of sequential travel decisions, for instance related to activity, departure time, and modes. Second, automatically collected data are advantageous as compared to travel surveys that are limited by respondents’ ability and willingness to disclose travel behavior. However, this approach poses methodological challenges, particularly because individuals’ exact locations over time are not directly observed. We address these in a latent-based framework in which the positions of the individuals are treated as latent variables and the location of cell towers as observations. We estimate the model using a stochastic expectation-maximization (StEM) algorithm. The performance is highlighted in a case study of daily travel demand in Stockholm, Sweden. Our paper demonstrates the agility of activity-based DDCMs in adapting to new data sources, by incorporating appropriate modifications within a well-established modeling paradigm.

Responding to climate change and sustainability goals, many cities and countries have renewed efforts to upgrade active transportation infrastructure, such as cycling. Good built environment conditions are a necessary prerequisite to promote cycling and so is the understanding of behavioral variations in travel, with travel time being particularly crucial. Because even in comparable built environment, conditions differences in bike usage are visible, which are often linked to socio-demographic factors. A comparative exploration would help shed light on these differences. With that view in mind, this paper examines the impact of trip characteristics and socio-demographic factors on cycling travel time in three European cities: Tallinn, Braga, and Istanbul, using a latent class accelerated failure time (LCAFT) model. The model reveals that socio-demographic factors, such as gender, age, and education, influence travel patterns differently across these cities. In the data from Tallinn and Braga, both age and gender significantly affect travel time. In contrast, in Istanbul, age and education level are the more decisive factors. These findings underscore the importance of considering the specific data contextual factors across different geographic locations. Moreover a comparative analysis further demonstrates that the impact of various factors on cycling travel time varies significantly by city. While distance consistently prolongs travel time, the effects of trip purpose and bike type differ based on socio-demographics and cultural contexts. 

Understanding which factors affect people's choices when traveling helps planners and policymakers build better infrastructure that fosters more sustainable practices. Gaining insights into how cyclists choose their routes is one major key to improving infrastructure and promoting cycling. Particularly commuters and e-bike users have been noted to have preferences unconstrained from shortest path logic. There are other more human centric factors can be of importance, such as closer to nature and away from busy intersections. To that end, this paper aims to uncover cyclists' preferences and the affecting nature-related attributes of routes for commuters to high school in Nijmegen, Netherlands. Based on a Dutch dataset, the study analyzed 1284 e-bike cycling trips, each with four route alternatives, including the chosen one. The primary objective is to identify the most influential parameters affecting e-bike commuters' route choices and understand their contributions. The approach employed both a simple path size Logit (PSL) and a Pairwise Combinatorial Logit (PCL) model, incorporating nature-related and interaction variables. Additionally, the research compared the predictive performance of Logit-based models with deep learning models. The findings shed light on the factors influencing e-bike commuters' route choices and demonstrate the superior predictive capabilities of deep learning techniques, with a validation accuracy of 80.16\%. By adopting a sensitivity analysis approach, we uncovered the key factors that influence our deep learning model in predicting cycling routes, giving a precise interpretation of our results. Our findings show that commuter e-bike cyclists prefer shorter routes with fewer traffic lights and favor routes that have more natural settings. However, their primary concern is efficiency in their commutes.

Dynamic discrete choice models (DDCMs) account for mode choice and departure time in an activity-based framework in which the order of activity is important. The power of such models does not only lie in their precise descriptions of how individuals make decisions, but also in their ability to test different scenarios and thus guide policy assessment. Recently, mobile network data have emerged as new sources of spatio-temporal data for estimation of DDCMs, however knowledge of how to do this effectively for transport policy is still emerging. The aim of this study was to generate individual activity-trips based on anonymized mobile network data to assess the impact of the COVID-19 pandemic on travel behavior among residents in two suburbs with distinct socioeconomic characteristics in Stockholm, Sweden. For both income groups, the results suggested that travel decreased substantially in terms of trip duration and frequency, as well as travel time. Further, the pandemic drastically decreased the number of people present in the central, commercial areas during the day. However, the effects were more considerable in the low-income group. Calculations of the spatial segregation index showed that segregation between the two groups increased during the day, from around 0.6 to around 0.7, i.e., more than 15\%. This suggests that the pandemic disproportionately impacted lower-income groups. This study contributes to the literature on mobility equity and novel, data-driven methods within urban planning and policy.