• Physical (e.g., walking, biking, sports)
• Nature appreciation (e.g., birdwatching, picnicking)
• Environmental (e.g., gardening, conservation)
• Social (e.g., festivals, volunteering)
• Cultural (e.g., performances, exhibitions)
• Mindfulness (e.g., yoga, meditation)

Figure: Visualization of the park scores in London for (a) nature appreciation and (b) physical activities. Explore all parks in 35 cities using the interactive visualization. CC-BY Linus Dietz, map data from OpenStreetMap.

How to Measure Parks’ Support for Well-being

The researchers collected all features found in parks (such as benches, ponds, or sports courts) from OpenStreetMap and linked them to activity types. Because the number of tags is large, they used a validated Large Language Model approach to compile a lexicon of 1432 park features and their corresponding health-promoting activities. To score each park, they built statistical models that measure how well a park supports specific health-related activities. These scores show how many health-promoting elements a park has compared with an average park in the same city. To reflect cultural and climatic differences, each model is separate for each city. The validation of the scores was done by comparing them with a dataset of 10 million geolocated Flickr photos taken in parks. Parks with higher scores had more photos of people engaged in the relevant activities.

The approach is first to systematically assess the qualities of parks for wellbeing on a global scale.

Key Findings

1. Regional differences in park design priorities: European parks prioritize opportunities for nature appreciation, whereas North American parks are focused around physical activities.
2. Spatial gradient: Inner-city parks consistently achieve higher scores across all activity categories compared to peripheral parks.
3. Inequality patterns: The degree of inequality in park activity scores shows no consistent geographic trend across cities, with Copenhagen and Rio de Janeiro exhibiting the greatest disparities and Tokyo, Paris, and Auckland the most equitable offerings. Generally, physical activity infrastructure is fairly evenly distributed but features supporting social interaction tend to be concentrated in a few parks.

Figure: Distribution of park scores by (a) distance quartile to the city center, (b) activity category, and (c) continent. CC-BY Linus Dietz.

Implications

The paper provides actionable insights into the current state of urban parks, enabling the assessment of existing shortcomings and supporting evidence-based decisions for future development. In interviews, park managers and urban designers highlighted that such quantitative evidence can inform decisions about which features to prioritize and can help justify and secure funding for targeted improvements. Furthermore, urban planners can use the park scoring system to identify underserved areas and to prioritize investments that promote more equitable access to high-quality, health-promoting green spaces.

Resources

• Paper: Understanding the Potential of Urban Parks to Promote Well-being. Nature Cities 2025
• Project page
• Interactive Visualization
• Replication package with code and data

Quantifying Urban Greenspace Globally

Recent research has highlighted the role of specific spaces within urban parks in supporting physical and mental wellbeing [1]. Using the Lexicon of Health-Related Park Features [2], a dataset of 23,477 parks across 35 cities was evaluated based on their potential to promote health-related activities.

This project offers students the opportunity to apply and extend their data science skills to a global-scale environmental challenge. Building upon an existing data processing pipeline for urban park analysis [3], the aim is to expand the scope to include all cities worldwide and incorporate a broader range of urban greenspaces—such as community gardens, green corridors, and other forms of urban vegetation. The analysis will be complemented with additional environmental indicators, including urban heat stress [4], air pollution, and vegetation indices such as NDVI, primarily derived from satellite observations.

An important aspect of this work is understanding the urban context of greenspaces; their spatial location, accessibility, and surrounding built environment characteristics such as building density and land-use type. Developing such contextual indices provides valuable insights into how greenspaces are used and helps explain their broader ecological and social roles within cities [5,6].

• [1] Health-promoting Potential of Parks in 35 Cities Worldwide. Nature Cities 2025
• [2] Lexicon of Health-Related Park Features (OpenStreetMap)
• [3] Park Analysis Pipeline
• [4] Parks, S. A., Dillon, G. K., & Miller, C. (2014). A new metric for quantifying burn severity: the relativized burn ratio. Remote Sensing, 6(3), 1827-1844.
• [5] Larson, K. L., Brown, J. A., Lee, K. J., & Pearsall, H. (2022). Park equity: why subjective measures matter. Urban Forestry & Urban Greening, 76, 127733.
• [6] Kaczynski, A. T., Schipperijn, J., Hipp, J. A., Besenyi, G. M., Stanis, S. A. W., Hughey, S. M., & Wilcox, S. (2016). ParkIndex: Development of a standardized metric of park access for research and planning. Preventive medicine, 87, 110-114.

Analyzing and Visualizing the Inequalities of London’s Built Environment and Its Effect on Health

This project aims to develop a comprehensive analytical and visualization framework to quantify and understand inequalities across London’s built environment and their impact on public health. By integrating diverse spatial and socioeconomic datasets, the project will construct multi-dimensional indices of inequality that capture environmental, infrastructural, and health-related disparities across London boroughs. The framework will provide a quantifiable and scalable tool for assessing how variations in the built environment contribute to health inequities, supporting evidence-based urban policy and planning.

The built environment component investigates how the characteristics of places where people live influence their daily wellbeing. This includes examining air quality and pollution levels, noise exposure, access to green spaces and tree coverage, and neighbourhood resilience to extreme weather events such as flooding and heatwaves. Special attention is given to the vulnerabilities of specific population groups and the cumulative effects of environmental stressors. By aggregating and analysing these factors, the project explores how urban form and environmental conditions interact to shape comfort, safety, and overall quality of life across London.

The health equity component examines spatial patterns of health outcomes and access to healthcare. It analyses indicators such as obesity prevalence, GP prescription rates, and accessibility of healthcare services, alongside official measures of health deprivation. This analysis reveals where and why health inequalities persist and how they relate to environmental and infrastructural factors. Together, these insights aim to illuminate the links between the built environment and health outcomes, providing actionable intelligence for local authorities, urban planners, and public health agencies.

• Sanja Šćepanović, Ivica Obadic, Sagar Joglekar, Laura Giustarini, Cristiano Nattero, Daniele Quercia, and Xiao Xiang Zhu. 2024. MEDSAT: a public health dataset for england featuring medical prescriptions and satellite imagery. In Proceedings of the 37th International Conference on Neural Information Processing Systems (NIPS ‘23)
• Schipperijn, J., Madsen, C.D., Toftager, M. et al. The role of playgrounds in promoting children’s health – a scoping review. Int J Behav Nutr Phys Act 21, 72 (2024). https://doi.org/10.1186/s12966-024-01618-2
• Bach, B., Freeman, E., Abdul-Rahman, A., Turkay, C., Khan, S., Fan, Y., & Chen, M. (2022). Dashboard design patterns. IEEE Transactions on Visualization and Computer Graphics, 29(1), 342-352. https://dashboarddesignpatterns.github.io
• https://data.london.gov.uk/
• https://www.gov.uk/government/collections/english-indices-of-deprivation

AI for Data Governance

The exchange of data within and between organizations is governed by company policies and data protection laws. As policies and data flows change over time, maintaining compliance in data exchange poses a complex challenge. Computational data governance refers to automating governance tasks using AI tools that support the organization and its employees to comply with policies and allow data sharing to foster the business needs. Currently, this is a disruptive field in enterprise software with many emerging solutions because all the current and future AI systems need to maintain access to data, which is only possible with highly automated and sophisticated data governance.

The project is about using popular open source software [1, 1b] and open standards [2,3] from the Linux Foundation to develop and evaluate solutions that facilitate data governance within organizations. Based on seminal work [4], further aspects of data governance will be explored, including data quality assurances, metadata tagging, query generation, query validation, AI explanations, and scaling to Big Data problems. An initial dataset of annotated data access requests is provided [5], which is to be adopted to other use cases.

Successful candidates will learn how to employ Large Language Models on a critical problem the IT industry currently faces. Gartner predicts: “By 2027, for example, 60% of organizations will fail to realize the anticipated value of their AI use cases due to incohesive data governance frameworks [6].” This is your chance to prove them wrong and become an expert in a highly-sought after field.

• [1] Data Product MCP https://github.com/entropy-data/dataproduct-mcp
• [1b] Data Contract CLI https://github.com/datacontract/datacontract-cli
• [2] Bitol, Open Data Product Standard (ODPS). Linux Foundation AI & Data, 2025. Version 1.0.0.
• [3] Bitol, Open Data Contract Standard (ODCS). Linux Foundation AI & Data, 2025. Version 3.0.1.
• [4] Dietz, L. W., Wider, A., & Harrer, S. (2025). Automating Data Governance with Generative AI. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(1), 760-771. https://doi.org/10.1609/aies.v8i1.36587
• [5] Supplementary Material for “Automating Data Governance with Generative AI” https://github.com/LinusDietz/Automating-Data-Governance
• [6] https://www.gartner.com/en/data-analytics/topics/data-governance

In the past, data governance relied on checklists, spreadsheets, or domain-specific languages for access control. With advances in natural language processing, data marketplaces within decentralized systems can now help decision-makers manage data sharing more effectively.

Laws such as the EU General Data Protection Regulation, the California Consumer Privacy Act, and the EU AI Act require strict control over personal and sensitive data. Organizations must balance innovation with compliance under these rules.

Data Marketplaces and AI-Assisted Data Governance

In data mesh architectures, domain teams manage their own data products within a federated environment. Each product defines its usage rules and guarantees through a data contract, which states what data can be shared and under what conditions. To coordinate these contracts, an enterprise data marketplace records, reviews, and approves or rejects access requests between data products.

At the center of this marketplace is Governance AI, an LLM-powered tool that checks whether a data access request complies with the provider’s data contract, company policies, and legal requirements such as the GDPR. Governance AI does not make final decisions. Instead, it issues structured warnings and suggestions for correction to guide human experts.

Generating Realistic Testing Datasets

Testing compliance systems is difficult. Real-world datasets contain confidential information, and manually creating synthetic test cases takes time. To evaluate the system, we used LLMs to generate realistic data access requests, building on the work of Herdel et al. (2024). Each test case included metadata about data products, privacy policies, and data contracts to create plausible access scenarios. For example, a marketing team requesting customer purchase data for campaign optimization.

Domain experts from the insurance and e-commerce sectors reviewed the generated requests. Most were judged realistic, and some closely matched real-world use cases; the rest were discarded.

Evaluating Computational Governance

In the evaluation, Governance AI was compared with domain experts who assessed 110 access requests across both sectors.

Governance AI issued 3.6 times more warnings than human experts. It did not miss any case where experts raised a compliance concern. After a secondary review, experts judged 80% of the AI’s warnings to be correct.

This cautious approach may slow some data-sharing workflows, but it supports a key compliance principle: prevent breaches first, optimize later. The AI also offered actionable suggestions, which proved more useful in e-commerce than in highly regulated fields such as insurance.

Findings and Implications

The study highlights several insights for organizations pursuing AI-assisted data governance. Governance AI can handle complex policy reasoning without missing critical cases, showing that automated assistance is both feasible and effective. Synthetic access requests generated by LLMs can realistically simulate real-world governance scenarios, providing a practical way to test systems at scale. Despite the AI’s cautious approach, human oversight remains essential to ensure contextual and legal accuracy. In privacy-sensitive contexts, a stricter approach is often safer, as over-warning carries less risk than under-warning. Differences across sectors also matter: e-commerce scenarios allowed flexible mitigation, such as anonymization, while insurance required precise legal compliance. Beyond compliance, we explored the potential for continuous governance, where AI systems dynamically test and monitor policy adherence as data landscapes evolve.

This is a project in collaboration with Arif Wider from HTW Berlin as well as Simon Harrer, who co-founded Entropy Data and built the data marketplace.

Linus W. Dietz, Arif Wider, & Simon Harrer. Automating Data Governance with Generative AI. AAAI / ACM Conference on Artificial Intelligence, Ethics, and Society 2025.

In a mixed-method study combining interviews with runners and a large-scale online survey, we captured what aspects influence the running experience. Analyzing the answers, we were able to group them into performance and achievement, environmental quality, and mental and social connectedness. These insights led to the creation of the Experience of Running Scale designed to measure how a run feels. Clustering the responses also revealed two major groups of runners’ needs in terms of urban running routes. Sometimes, runners are drawn to scenic paths, which are lush, peaceful, and rich with natural beauty. Others thrive on the energy of urban routes, where the city pulses with life.

Interestingly, these preferences often line up with our personalities. Runners who score higher on neuroticism tend to seek the quiet solitude of nature, while extroverts are more likely to enjoy the buzz of city streets. Using this data, we built a personalized routing engine that accounts for environmental factors like sound, smell, surface type, and traffic to recommend routes that better match what we need.

In a routing engine prototype for London running routes, we mapped the perceivable environment using open data from platforms like Flickr, OpenStreetMap, and local crime reports to be able to recommend these two types of routes. While most runners we asked about their route preferences preferred scenic routes for their calm and beauty, situations like running in the dark flipped this tendency towards the more crowded urban routes as safety concerns trumped serenity.

We hope that by making it easier to plan routes tailored to individual needs and preferences, we can lower the barriers to going out, encouraging people to exercise with greater joy and confidence.

Hänsel, K., Aiello, L. M., Quercia, D., Schifanella, R., Varga, K. Z., Dietz, L. W., & Constantinides, M. (2025). The experience of running: Recommending routes using sensory mapping in urban environments. International Journal of Human-Computer Studies.

Point-of-interest (POI) recommenders have become ubiquitous tools for travelers, guiding them on where to eat, explore, or relax. But while researchers and developers often focus on which algorithm performs best there is little research on understanding why certain models perform better than others across datasets. In this study, we set out to investigate that using data characteristics. Using Foursquare check-in data from New York City, we developed a framework that links the characteristics of a dataset, such as sparsity, popularity bias, and user mobility, to the performance of different recommendation algorithms analyzing accuracy, novelty and item exposure. To uncover how data characteristics shape outcomes we curated 144 distinct data subsets, each reflecting real-world conditions: peak tourist seasons, domestic versus international users, and variations in user activity and mobility. We then trained a mix of classic and POI-specific recommendation models on each subset, analyzing the results using an explanatory framework powered by a multiple-regression model.

What we found confirmed our suspicions: widely studied data characteristics like density and popularity bias do play a role—but so do subtler dynamics, such as how far users tend to travel or how long they’ve been active on the platform. These hidden variables often determine whether a recommendation algorithm succeeds or falls short. Our findings reinforce a key insight: there’s no universal best model. The optimal algorithm depends on the specific characteristics of the data and the goals of the recommendation system, trading-off accuracy, novelty, or equitable item exposure. By quantifying the characteristics of the data set, one can better select or fine-tune models to meet those goals. Ultimately, we hope this work helps pave the way toward more relevant and serendipitous suggestions for users—and more balanced visibility for local businesses. Our framework represents a step toward more informed, intentional choices in recommendation system design to deliver recommendations that are relevant, but also fair, diverse, and genuinely helpful to both travelers and the communities they explore.

Linus W. Dietz, Pablo Sánchez, and Alejandro Bellogín. “Understanding the Influence of Data Characteristics on the Performance of Point-of-Interest Recommendation Algorithms”. In: Information Technology & Tourism 27.1 (Jan. 2025), pp. 75–124. doi: 10.1007/s40558-024-00304-0.

The ongoing urbanization leads to a concentration of population in cities, making it increasingly harder to navigate and exploit their potential. Recommendation technology has emerged as a pivotal tool shaping individuals’ attention and decision-making within these urban landscapes. People heavily rely on recommender systems to overcome the choice overload problem, from selecting accommodations to deciding on dining options and exploring attractions. Thus, recommendation algorithms significantly influence how individuals engage with and navigate cities.

The first Workshop on Urban Recommender Systems (UrbanRec) will bring together industry practitioners and academic scholars dedicated to advancing recommendation systems tailored to the unique dynamics of urban environments. We aim to highlight the advancements in recommender systems technology and also dedicate attention to the aspects of urban health and well-being, which can be positively impacted through the use of recommendation technology. By aligning technological advancements with societal goals, the workshop aims to lay the path towards more inclusive, vibrant, and livable cities where recommendation systems play a constructive role in enhancing the quality of life for all residents and visitors alike.

The workshop aims to identify pressing challenges and seize opportunities at the intersection of recommendation technology and urban sciences through interdisciplinary discussions. By integrating insights from diverse perspectives spanning urban planning, data science, psychology, and public health, participants will explore novel approaches to recommender systems that prioritize user autonomy, community well-being, and environmental sustainability. Ultimately, the workshop aims to foster a deeper understanding of opportunities in the complex interplay between technology and urban dynamics.

Workshop Topics

• Mobile Urban Recommender Systems
• Point-of-Interest Recommendations
• Location-based Services and Route Recommendations
• Health-related Behavioral Change
• Supporting Policymaking and Urban Planning
• Responsible and Ethical Urban Recommender Systems
• Dataset Track

Read the call for papers on the workshop website.

The workshop is co-organized with Sanja Šćepanović, Rossano Schifanella, and Wolfgang Wörndl.

Deciding where to travel is a complex, emotionally involving, and financially relevant decision which people face relatively infrequently. Some aspects of tourist recommendations, such as point-of-interest recommendation, hotel recommendations, or restaurant recommendations, are commercially well established, whereas there are few successful recommender systems for individual travel destinations. In this thesis, we present several contributions in the context of destination recommendation covering traveler mobility analysis, destination characterization, and conversational recommender systems.

Understanding traveler mobility forms the basis for more personalized recommendations. We propose methods to analyze global traveler mobility from location-based social networks to learn which data sources are suitable for analyses in this domain, how people travel around the world, and which types of travelers can be observed. Our cluster analyses of trips and travelers reveal distinct groups, which can serve as an initial preference elicitation step, but we could also show that the common practice of evaluating point-of-interest recommendations without differentiating these groups leads to misleading results. Furthermore, we use the mined trips to construct a specialized map of hierarchical travel regions and to recommend the personalized duration of stay at destinations.

To correctly match traveler preferences with destinations in the content-based recommendation domain, we investigate which data sources are suitable for characterizing destinations. Constructing 18 data models and eliciting the concept of touristic experience in cities using an expert study, we determine that textual data sources, e.g., Wikipedia articles, do a good job of emulating the touristic experience using rank agreement metrics. Additionally, we are able to optimize data sources with explicit features to be competitive by learning the importance of feature weights using black-box learning.

Finally, we present the CityRec conversational destination recommender system. Since users often struggle to verbalize their true preferences and might have unrealistic expectations about destinations, we propose a novel conversational paradigm, “Navigation by Revealing Trade-offs”, to overcome the wishful-thinking problem and inform users of the trade-offs involved in choosing one destination over another. The seamless integration of user interface and algorithms was evaluated using a large-scale user study with 600 participants.

Linus W. Dietz. “Data-Driven Destination Recommender Systems.” Technical University of Munich, 2023.

The involvement of geographic information differentiates point-of-interest recommendation from traditional product recommendation. This geographic influence is usually manifested in the effect of users tending toward visiting nearby locations, but further mobility patterns can be used to model different groups of users. In this study, we characterize the check-in behavior of local and traveling users in a global Foursquare check-in data set. Based on the features that capture the mobility and preferences of the users, we obtain representative groups of travelers and locals through an independent cluster analysis. Interestingly, for locals, the mobility features analyzed in this work seem to aggravate the cluster quality, whereas these signals are fundamental in defining the traveler clusters. To measure the effect of such a cluster analysis when categorizing users, we compare the performance of a set of recommendation algorithms, first on all users and then on each user group separately in terms of ranking accuracy, novelty, and diversity. Our results on the Foursquare data set of 139,270 users in five cities show that locals, despite being the most numerous groups of users, tend to obtain lower values than the travelers in terms of ranking accuracy while they also seem to receive more novel and diverse POI recommendations. For travelers, we observe the advantages of popularity-based recommendation algorithms in terms of ranking accuracy by recommending venues related to transportation and large commercial establishments. However, there are considerable differences in the respective traveler groups, especially between predominantly domestic and international travelers. Due to the considerable influence of mobility on the recommendations, this article underlines the importance of analyzing user groups differently when making and evaluating personalized point-of-interest recommendations.

Pablo Sánchez and Linus W. Dietz. “Travelers vs. Locals: The Effect of Cluster Analysis in Point-of-Interest Recommendation.” In: 30th ACM Conference on User Modeling, Adaptation and Personalization. UMAP’22. New York, NY, USA: ACM, July 2022, pp. 132– 142. DOI: 10.1145/3503252.3531320

Characterizing items for content-based recommender systems is challenging in complex domains such as travel and tourism. In the case of destination recommendation, no feature set can be readily used as a similarity ground truth, which makes it hard to evaluate the quality of destination characterization approaches. Furthermore, the process should scale well for many items, be cost-efficient, and, most importantly, correct. To evaluate which data sources are most suitable, we investigate 18 characterization methods that fall into the following categories: venue data, textual data, and factual data. We make these data models comparable using rank agreement metrics and reveal which data sources capture similar underlying concepts. To support choosing more suitable data models, we capture the desired concept using an expert survey and evaluate our characterization methods toward it. We find that the textual models to characterize cities perform best overall, with data models based on factual and venue data being less competitive. However, we show that data models with explicit features can be optimized by learning weights for their features.

Linus W. Dietz, Mete Sertkan, Saadi Myftija, Sameera Thimbiri Palage, Julia Neidhardt, and Wolfgang Wörndl. “A Comparative Study of Data-driven Models for Travel Destination Characterization.” In: Frontiers in Big Data 5 (Apr. 2022). ISSN: 2624-909X. DOI: 10.3389/fdata.2022.829939

Tourism is a complex domain for recommender systems because of the high cost of recommending an unsuitable item and the absence of ratings to learn user preferences. Conversational recommender systems have been introduced to provide users with an opportunity to give feedback on items in a turn-based dialog until a final recommendation is accepted. In a scenario such as recommending a city to visit, conversational content-based recommendation may be well-suited since users often struggle to specify their preferences without concrete examples. However, critiquing item features comes with challenges. Users might request item characteristics during recommendation that do not exist in reality, for example, demanding very high item quality for a very low price. To tackle this problem, we present a novel conversational user interface that focuses on revealing the trade-offs of choosing one item over another. The recommendations are driven by a utility function that assesses the user’s preference toward item features while learning the importance of the features to the user. This enables the system to guide the recommendation through the search space faster and accurately over prolonged interaction. We evaluated the system in an online study with 600 participants and found that our proposed paradigm leads to improved perceived accuracy and fewer conversational cycles compared to unit critiquing.

Linus W. Dietz, Sameera Thimbiri Palage, and Wolfgang Wörndl. “Navigation by Revealing Trade-offs for Content-based Recommendations.” In: Information and Communication Technologies in Tourism. Ed. by Jason L. Stienmetz, Berta Ferrer-Rosell, and David Mas- simo. Cham: Springer, Jan. 2022, pp. 149–161. ISBN: 978-3-030-94751-4. DOI: 10.1007/978-3-030-94751-4_14