Data Quality Challenges in Multimodal Tourism Recommender Systems

Zehui Wang, Wolfram Höpken, Dietmar JannacharXiv 2026 · arXiv

TL;DR

While multimodal data (text, images, social links) can supercharge tourism recommendations, "more data" often means "more noise." This paper identifies critical data quality pitfalls—from LLM hallucinations to the "tourist vs. local" bias—and proposes a roadmap for building more robust, quality-aware recommendation engines.

What Problem Does This Paper Address?

In the world of tourism, a recommendation isn't just about a coordinate on a map; it’s about the "vibe" of a place, captured through reviews, photos, and social connections. To capture this, researchers are increasingly building multimodal recommender systems that fuse diverse data sources.

However, there is a significant gap: most models assume the underlying data is accurate and representative. In reality, user-generated content is messy. Fake reviews, bot activity, and even the tools we use to process this data (like Large Language Models) introduce noise. If a model can’t distinguish between a genuine traveler’s photo and a bot’s "burst" posting, or if it can't reconcile a review saying a cafe is "quiet" with a photo showing a packed crowd, the recommendations will inevitably fail. This paper systematically categorizes these quality challenges to move the field toward more "data-aware" modeling.

Methods

This is a position and challenge paper that draws on the authors' experiences working with real-world datasets. Rather than introducing a single new algorithm, the authors perform a qualitative and forensic analysis of data quality issues across two major sources. They categorize these issues into three main pillars:

  1. Noise: Identifying both human-generated noise (fake accounts) and model-generated noise (LLM hallucinations).
  2. Demographic Bias: Analyzing how the "who" behind the data (e.g., locals vs. tourists) skews the "what" the system recommends.
  3. Semantic Contradictions: Investigating how different modalities (text vs. image) can provide conflicting information about the same location.

Datasets

The authors base their findings on two primary datasets:

  • Multimodal Yelp Dataset: A collection of user-generated reviews and check-in records used to study business representations.
  • Instagram Dataset: A specialized dataset collected from the Lake Constance region (spanning Germany, Switzerland, and Austria). This dataset is particularly useful for studying the behavioral differences between residents and travelers, though it remains private due to platform restrictions.

Key Insights and Ideas

  • The LLM Hallucination Tax: While LLMs are great for summarizing thousands of reviews, they can "hallucinate" features that don't exist. The authors suggest using structured summarization (forcing the LLM to fill specific fields) rather than open-ended generation to keep the model on a leash.
  • The "Tourist vs. Local" Divergence: A local might post about their favorite grocery store, while a tourist wants a landmark. If a model treats all data equally, it risks recommending mundane local spots to travelers.
  • Cross-Modal Conflict: One of the most interesting "aha" moments is the identification of semantic contradictions. A restaurant's text profile might claim it is "clean and quiet," but user-uploaded photos might show a cluttered, noisy environment. Current fusion strategies often ignore these conflicts, leading to unreliable embeddings.
  • Burst Postings as Noise: On platforms like Instagram, users often post multiple photos from the same spot within minutes. These "bursts" can dilute the signal of actual movement trajectories and bias the statistical distribution of activities if not filtered.

Key Results

As a challenge-oriented paper, the results are primarily diagnostic:

  • Detection of Non-Human Users: The authors highlight that suspicious accounts can be flagged by analyzing "implausible movement speeds"—if a user "checks in" at two locations 50 miles apart within five minutes, the data point is likely fraudulent.
  • Geographic Specificity: The Instagram data revealed a heavy bias toward the DACH region (Germany, Austria, Switzerland). This means models trained on this data might reflect specific cultural behaviors that don't generalize to international tourists.
  • Visual Filtering: They found that generating summaries only for images with "clear and interpretable visual cues" significantly reduced noise compared to processing every uploaded photo.

Limitations and Future Work

The authors acknowledge that current recommendation models are too "trusting" of their inputs. They propose several future directions:

  • Quality-Aware Fusion: Developing models that can dynamically "downweight" a specific modality (like a blurry photo or a suspicious review) during the training process.
  • Intent Modeling: Better segmenting users into groups (e.g., locals vs. tourists) to ensure the recommendation matches the user's specific context.
  • New Modalities: Exploring the use of 360° immersive videos and marketing content, which are currently underutilized in the literature.
  • Robustness to Missing Data: Creating fusion mechanisms that don't break when one data stream (like geographic coordinates) is missing or corrupted.

Summary updated

Abstract

Modern recommender systems increasingly consider diverse data modalities to enhance model performance. However, the quality of the underlying multimodal data (e.g. data from social media platforms) has received limited attention so far. Focusing on tourism recommendation, we examine real-world challenges encountered when using datasets from Yelp and Instagram. Issues include noise from fake reviews, hallucinated content from LLMs, modality contradictions, and demographic bias. We outline future directions such as quality-aware fusion, structured and controlled LLM-based content generation to reduce hallucinations, and user group differentiation to improve the robustness and reliability of multimodal recommendation systems.
BibTeX
@misc{wang2026data,
  author = {Zehui Wang and Wolfram Höpken and Dietmar Jannach},
  title = {{Data Quality Challenges in Multimodal Tourism Recommender Systems}},
  year = {2026},
  eprint = {2611.13972},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13972},
}
Figure