BEAM: Extending the European Assets Map methodology

Suggested by: Johannes Scholz, Daniel Hölbling and Manuel Strohmaier (Joanneum Research Life)

Keywords: Risk Assessment, BEAM Methodology, Geospatial Data Generation

Objective:

1. Extend the Basic European Assets Map (BEAM) methodology, which estimates asset values in €/m², to additional European countries using publicly available data sources such as OpenStreetMap, CORINE Land Cover, Urban Atlas, and Eurostat.
2. Develop a consistent, Europe-wide dataset to support improved and comparable assessments of potential damages from natural hazards, enabling more robust risk analyses at the continental scale.
3. Compare and evaluate the resulting dataset against a dataset developed by JOANNEUM RESEARCH, and explore enhancements through the integration of geographically localized building data.

Short Description:

Natural hazards pose varying levels of risk across Europe, and consistent, spatially resolved asset data are essential for effective risk assessment. The Basic European Assets Map (BEAM) provides monetary asset values per unit area (€/m²) and has already been implemented and validated for Germany (Copernicus Emergency Management 2020). This thesis aims to generalize and apply the BEAM methodology to other European countries using harmonized, openly available data sources. The result will be a comprehensive and standardized dataset for estimating damage potentials across Europe. To ensure reliability and usability, the new dataset will be compared to a reference dataset developed by JOANNEUM RESEARCH and further improved by incorporating detailed, geographically referenced building information. This work contributes to the foundation for large-scale, comparative risk assessments in the context of natural hazards.

Start: ASAP

Copernicus Emergency Management Service Risk and Recovery Mapping 2024: https://mapping.emergency.copernicus.eu/download/EMSN076/EMSN076_Final-Report_v1_20082025.pdf

POI Recommendation Based on "Sense of Place" extracted from LLMs

 Suggested by: Omid Reza Abbasi, Johannes Scholz and Dominik Kowald (KnowCenter)

Keywords: POI Recommendation, Sense of Place, LLMs

Objective:

1. Develop a methodology for extracting features that capture "sense of place" from textual and visual data using LLMs. This involves exploring techniques such as sentiment analysis, topic modeling, style extraction, and image captioning to identify emotional, cultural, and atmospheric attributes of locations. 

2. Evaluate the impact of "sense of place" features on the performance of location-based recommender systems. This includes comparing the performance of recommender systems with and without these features, using metrics such as user satisfaction, engagement, and discovery of unique experiences. 

3. Investigate the applicability of "sense of place" feature engineering across different location types and user demographics. This involves testing the proposed methodology on various datasets, including restaurant reviews, travel blogs, and social media posts related to tourist attractions and local events. 

Short Description: Location-based recommender systems play a crucial role in helping users discover and experience places that align with their interests. However, traditional approaches often focus on objective attributes like distance, ratings, and popularity, neglecting the subjective and emotional dimension of "sense of place." This thesis proposes to investigate how Large Language Models (LLMs) can be utilized to extract features that capture the unique atmosphere, cultural significance, and emotional resonance of locations, thereby enhancing the personalization and relevance of recommendations. The core challenge addressed in this thesis is the inability of existing location-based recommender systems to understand and incorporate the intangible qualities that contribute to a location's "sense of place." User reviews, social media posts, and geotagged photos often contain valuable insights into the emotional and cultural aspects of a location, but these are frequently overlooked by conventional methods. This thesis aims to explore how LLMs can be leveraged to extract these features, enabling a more thorough understanding of user preferences and place characteristics. 

Start: Anytime

Relevant Studies:

Wan, Zhizhong, et al. "LARR: Large Language Model Aided Real-time Scene Recommendation with Semantic Understanding." Proceedings of the 18th ACM Conference on Recommender Systems. 2024. 

Location-Based Recommendation Using Fine-Tuned LLMs

 Suggested by: Omid Reza Abbasi, Johannes Scholz and Dominik Kowald (KnowCenter)

Keywords: Location Recommendation, LLMs, Fine-Tuning

Objective:

1. To investigate the application of LLM reasoning to enhance location-based recommender systems. 

2. To explore the effectiveness of zero-shot CoT prompting for generating relevant location-based recommendations. 

3. To compare the performance of the proposed LLM-based location-based RS approaches against traditional methods. 

Short Description: In this research, the selected LLM will be trained on the preprocessed datasets using appropriate training and validation splits. We will select one or more pre-trained Large Language Models as the base for fine-tuning. The selection will consider factors such as model size, architecture, availability, and prior performance on text-based tasks. Potential candidates include models from the GPT family (e.g., GPT-2, GPT-Neo), open-source alternatives like Llama, or task-specific models. The chosen LLM(s) will be initialized with their pre-trained weights. For models with a fixed input sequence length, we will ensure that the constructed input prompts do not exceed this limit. If necessary, truncation or other techniques to handle long sequences will be explored. The fine-tuning process will involve training the selected LLM(s) on the prepared training data using appropriate training objectives. The specific objective will depend on the recommendation task being addressed (e.g., next POI prediction, rating prediction, recommendation generation).  

Start: Anytime

Relevant Studies:

Lin, Xinyu, et al. "Data-efficient Fine-tuning for LLM-based Recommendation." Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval. 2024. 

Bai, Zhuoxi, et al. "Finetuning Large Language Model for Personalized Ranking." arXiv preprint arXiv:2405.16127 (2024).