CCBR framework lets users steer recommendations via text summaries
Users can now control what they see by editing AI-generated text descriptions of items.
A team of researchers (Fırat Öncel, Jihoon Jeong, Emiliano Penaloza, Mirco Ravanelli, Laurent Charlin, Cem Subakan) has introduced the Controllable and Content-Based Recommendations (CCBR) framework, detailed in a paper submitted to arXiv. CCBR addresses a key limitation of traditional recommendation systems, which rely on dense latent representations that are difficult to interpret and control. Instead, CCBR generates textual user profiles by summarizing item contents—such as images, audio, or video—into natural language bottlenecks. This design allows users to directly edit or provide feedback on the text summaries, steering recommendations toward their preferences. The framework plugs into existing collaborative filtering models and can incorporate multimodal inputs. Across multiple datasets covering images, audio, and video, CCBR achieves competitive performance with standard latent-representation models while offering interpretable control. It also outperforms TEARS, a recent baseline in controllable recommendation systems. The authors demonstrate through systematic interventions that users can effectively steer the model in desired directions.
This work has significant implications for personalization and user agency in AI-driven platforms. By replacing opaque vectors with human-readable text, CCBR makes recommendation logic transparent and modifiable. Users could, for example, tell a music recommender to “focus on acoustic guitar tracks” or a video platform to “show more nature documentaries” simply by altering the generated descriptions. The framework also supports multimodal interventions—users might combine text with a sample image to refine results. While currently under review, CCBR represents a step toward more user-centric recommender systems, balancing performance with interpretability and control. The code and data are expected to be released, potentially accelerating adoption in production environments where transparency and user trust are critical.
- CCBR generates text summaries from item content (images, audio, video) to serve as interpretable bottlenecks in collaborative filtering.
- Users can steer recommendations by editing or intervening on the text summaries, enabling multimodal control.
- Matches standard model performance and outperforms the TEARS baseline across multiple datasets.
Why It Matters
Makes recommendation systems transparent and controllable by users, without sacrificing performance.