TLDR: A new method called “Conformal Risk Control” helps recommender systems like those on YouTube or TikTok reduce unwanted content. Instead of just removing disliked items, it replaces them with “safe” previously viewed content, ensuring users still get a full list of recommendations while provably limiting exposure to unwanted material. The approach is flexible, works with any existing recommender, and was tested on a real-world video platform dataset.
Recommender systems have become an indispensable part of our daily online experience, from shopping to social media and video streaming. While these systems excel at personalizing content, they often face criticism for delivering irrelevant, unwanted, or even harmful recommendations. This can lead to user dissatisfaction, contribute to the spread of misinformation, and erode trust in platforms.
Existing mechanisms, such as ‘Not Interested’ buttons, often fall short. They can be ineffective, slow to adapt to user feedback, and lack transparency, leaving users to guess why they see certain content and how to avoid it.
A New Approach to Mitigate Unwanted Recommendations
A recent research paper, titled “You Don’t Bring Me Flowers: Mitigating Unwanted Recommendations Through Conformal Risk Control”, introduces an innovative method to tackle this challenge. Authored by Giovanni De Toni, Erasmo Purificato, Emilia Gomez, Bruno Lepri, Andrea Passerini, and Cristian Consonni, this paper proposes a model-agnostic and distribution-free approach that uses ‘conformal risk control’ to set a provable limit on unwanted content in personalized recommendations.
Conformal risk control is a technique that allows systems to guarantee that a certain ‘risk’ (in this case, the fraction of unwanted content) stays below a user-defined level. A key challenge with traditional methods is that simply filtering out unwanted items can lead to a smaller set of recommendations, potentially reducing user engagement. To overcome this, the researchers leverage implicit feedback on previously consumed items, using them to expand the recommendation set while still ensuring robust risk mitigation.
Insights from Real-World Data
The researchers conducted an in-depth analysis of data from Kuaishou, a popular online video-sharing platform. Their findings revealed several crucial insights:
- Negative feedback from users (e.g., clicking ‘Do not recommend’) is very sparse, meaning most users report very few videos.
- Surprisingly, user engagement (like watch time) doesn’t always correlate with whether a video is perceived as harmful. Videos with high watch times can still be flagged as unwanted.
- The system sometimes suggests content users have explicitly reported before.
- Very short engagement with a video can be an early indicator of unwanted content.
- Users often rewatch previously seen videos, and some even report a video as unwanted on a second viewing, even if they didn’t on the first.
How the System Works
The proposed method acts as a ‘post-hoc’ pipeline, meaning it can be applied to any existing recommender system without needing to retrain the core recommendation model. Here’s a simplified breakdown:
First, based on user feedback, the system calculates a ‘risk threshold’. Any items initially recommended that fall below this threshold (meaning they are more likely to be unwanted) are identified for removal.
Instead of simply discarding these items and potentially leaving the user with fewer recommendations, the system replaces them with ‘safe’ alternatives. These safe alternatives are typically videos the user has seen before, has not flagged as unwanted, and has engaged with for a significant duration (e.g., watched for a certain percentage of its length). This strategy ensures that the user still receives a full list of recommendations while the overall proportion of unwanted content is provably controlled.
Also Read:
- Strengthening Safety in Diffusion Models Against Fine-Tuning
- Advancing Location Recommendations with Disentangled Graph Learning and POI Weighting
Experimental Validation and Impact
Experiments conducted on the KuaiRand dataset from Kuaishou demonstrated the effectiveness of this approach. The results showed that the method reliably reduces unwanted recommendations to the desired level. Crucially, the ‘Replace’ strategy significantly outperformed a simple ‘Remove’ strategy in maintaining the quality and size of the recommendation list. While removing items drastically degraded recommendation quality, replacing them allowed for better performance, even if some degradation still occurred due to reintroducing previously seen content.
The choice of the underlying scoring function (the model that initially ranks items) also played a role. Interestingly, models designed to incorporate both positive and negative feedback (sign-aware models) sometimes led to more item replacements than those focusing only on positive signals, likely due to the sparsity of negative feedback in the dataset.
The study also highlighted a trade-off: stricter criteria for selecting ‘safe’ replacement items (e.g., only using videos watched for 100% of their duration) further ensures safety but reduces the pool of available replacements, potentially leading to fewer recommendations overall. Furthermore, the global risk control threshold tended to be more conservative for users who rarely report unwanted content, as it was heavily influenced by users who report frequently.
This research marks a significant step towards building more user-centric and trustworthy recommender systems, offering a practical and theoretically sound way to give users more control over the content they see. For more details, you can read the full research paper here.
