Research map / Vision-language and multimodal / Human and embodied modeling
Tell Me What Is Good about This Property: Leveraging Reviews for Segment-Personalized Image Collection Summarization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CrossSummarizer, an unsupervised cross-modal method that personalizes image gallery summaries for Booking.com user segments by matching review-extracted topics to image embeddings after clustering.
Problem
Image collection summarization is generic and ignores user intent; personalizing property image galleries to traveller and trip segments without costly annotations is hard.
Approach
Combine K-Medoids clustering on MuMIC image embeddings with segment-specific review topics (from a Text2Topic BERT cross-encoder) via a tempered-sigmoid image-topic confidence matrix, selecting the best-matching image per cluster to personalize summaries without manual image annotation.
Key results
- CrossSummarizer achieves Diversity 0.963 and Coverage 0.617 on the Small split, outperforming Default Clustering (Cov 0.516) and Clustering with Personalization (Cov 0.578).
- In the user study over 210 samples, CrossSummarizer scored 0.66 ± 0.38 versus 0.34 ± 0.38 for Clustering with Personalization (t(209)=-5.537, p<5e-8).
- Topic-Based Personalization achieves the highest Reviews Coverage on the Big split (0.707) but much lower Representativeness (0.775 vs 0.885 for CrossSummarizer).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add review-topic matching within image clusters | Image collection summarization · Booking.com Small | diversity: 0.947 -> 0.963 | Helped |
| Add review-topic matching within image clusters | Image collection summarization · Booking.com Small | coverage: 0.516 -> 0.617 | Helped |
| Add review-topic matching within image clusters | Image collection summarization · Booking.com Small | reviews coverage: 0.581 -> 0.729 | Helped |
| Add review-topic matching within image clusters | Image collection summarization · Booking.com Small | representativeness: 0.931 -> 0.903 | Hurt |
| Add review-topic matching within image clusters | Image collection summarization · Booking.com Big | diversity: 0.925 -> 0.950 | Helped |
| Add review-topic matching within image clusters | Image collection summarization · Booking.com Big | coverage: 0.430 -> 0.524 | Helped |
| Add review-topic matching within image clusters | Image collection summarization · Booking.com Big | reviews coverage: 0.488 -> 0.677; 0.707 | Mixed |
| Add review-topic matching within image clusters | Image collection summarization · Booking.com Big | representativeness: 0.929 -> 0.885 | Hurt |
| Use CrossSummarizer instead of Clustering with Personalization | Image collection summarization · Booking.com Small | coverage: 0.578 -> 0.617 | Helped |
| Use CrossSummarizer instead of Clustering with Personalization | Image collection summarization · Human paired test across user segments | average preference score: 0.34 ± 0.38 -> 0.66 ± 0.38 | Helped |
| Use only review-topic similarity, without clustering | Topic-Based Personalization · Booking.com Small | representativeness: 0.931 -> 0.733 | Hurt |
| Use only review-topic similarity, without clustering | Topic-Based Personalization · Booking.com Small | coverage: 0.578 -> 0.365 | Hurt |
Surprising findings
- Relying only on review topics (Topic-Based) yields poor visual coverage and representativeness, showing that text alone is insufficient.
- Business traveller type was the only segment where CrossSummarizer did not get higher average user-study scores.
Limitations
- Cold-start properties with few reviews but full galleries are poorly handled.
- Dataset and code are not shared, limiting reproducibility.
- Evaluation is internal to Booking.com with a small human study of 5 participants.
Experimental setup
Datasets: Booking.com property images and reviews (over 6000 properties) Booking.com multi-label image classification dataset
Benchmarks: Booking.com property galleries (Small and Big splits)
Metrics: Diversity Representativeness Coverage Reviews Coverage human paired preference score
Baselines: Default Clustering Clustering with Personalization Topic-Based Personalization
Builds on: CLIP MuMIC Text2Topic K-Medoids
Closest papers by meaning
- Query-Focused Video Summarization: Dataset, Evaluation, and a Memory Network Based Approach (CVPR 2017 · similarity 0.86)
- Beyond Markovian Drifts: Action-Biased Geometric Walks with Memory for Personalized Summarization (ICLR 2026 · similarity 0.85)
- PersonalSum: A User-Subjective Guided Personalized Summarization Dataset for Large Language Models (NeurIPS 2024 · similarity 0.84)
- DIUSum: Dynamic Image Utilization for Multimodal Summarization (AAAI 2024 · similarity 0.83)
- Tailored Visions: Enhancing Text-to-Image Generation with Personalized Prompt Rewriting (CVPR 2024 · similarity 0.82)
- CaptionQA: Is Your Caption as Useful as the Image Itself? (CVPR 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.