You Don't Bring Me Flowers: Mitigating Unwanted Recommendations Through Conformal Risk Control (Extended Abstract)

1 Introduction

The widespread use of recommender systems on online platforms has significantly changed how users engage with content. These systems are integral to modern digital media, facilitating personalized navigation through extensive online content across domains such as e-commerce [Ghanem et al., 2022; Loukili et al., 2023], social media [Gao et al., 2023; S´anchez-Fern´andez and Jim´enez-Castillo, 2021], and content streaming [Deldjoo et al., 2021; Elahi et al., 2022], and tailoring experiences based on historical behaviors and expressed preferences [Fern´andez and Bellog´ın, 2020; Boratto et al., 2021]. By leveraging user-generated data, recommenders aim to enhance user engagement [Xue et al., 2023], satisfaction [Nguyen et al., 2018], and platform profitability [Cai and Zhu, 2019]. However, despite their proven utility and benefits, the basic concepts that make recommendation algorithms effective, such as similarity-based filtering and preference prediction, can inadvertently contribute to the spread of unwanted content [Liu et al., 2024]. Platforms have faced increasing criticism for creating filter bubbles and rabbit holes that reinforce existing beliefs, and for spreading potentially dangerous or harmful content [Tomlein et al., 2021]. Such dynamics are often attributed to algorithms that prioritize engagement, sometimes at the expense of content quality and truthfulness [Tommasel et al., 2021]. Features like the "Not Interested" button on YouTube [Liu et al., 2024] suffer from issues like low user awareness, complexity of use, or limited effectiveness. The lack of transparency and explainability in such strategies further exacerbates these problems, making it difficult for users to understand why they are being shown certain content and how to avoid it. As a result, users develop "folk theories" to explain the recommendations they are getting and try to influence the recommender outcomes [Bernstein et al., 2013; Eslami et al., 2015; Eslami et al., 2016; Liu et al., 2024]. Recent works have studied how to mitigate unwanted content, biases, or ensure fairness in recommendations [Chee et al., 2024; Ahn et al., 2024], but it remains unclear how to grant users precise and transparent control over the systems' risk.

∗ Full version appears in the proceedings of RecSys 2025 [De Toni et al., 2025b], winner of the Best Paper Award.

†Disclaimer: The view expressed in this extended abstract is purely that of the authors and may not, under any circumstances, be regarded as an official position of the European Commission.

Our contributions. We propose a post-hoc method based on conformal prediction [Shafer and Vovk, 2008; Angelopoulos and Bates, 2023] to provably mitigate the exposure to unwanted recommendations in expectation. Conformal prediction is a framework for uncertainty quantification for machine learning models that constructs prediction sets containing the true output with a user-specified confidence level, assuming data exchangeability. It uses non-conformity scores from model outputs to provide distribution-free, finite-sample guarantees. Conformal methods have recently been applied to decision support in classification tasks [Straitouri et al., 2023; Straitouri and Rodriguez, 2024; De Toni et al., 2025a], natural language processing [Campos et al., 2024], and recommendation systems [Angelopoulos et al., 2023b; Xu et al., 2024]. However, in recommender systems, current approaches often overlook key challenges, such as rerecommending unwanted items or balancing personalization with content filtering. To address this, we apply conformal risk control [Angelopoulos et al., 2023a], which generalizes conformal prediction to control general loss functions (or "risk") rather than coverage. Our method builds recommendation sets by filtering potentially unsafe items and replacing them with previously consumed, safe content, while provably respecting a user-defined risk level. Specifically, our contributions include:

1. A comprehensive analysis of real data to understand reporting patterns and watch time behavior related to unwanted content, providing insights into how users interact with and respond to such content.
2. Amethod based on conformal risk control that provably limits unwanted content in recommendations, with theoretical guarantees, by leveraging repeated content.
3. A practical heuristic to choose safe alternative content, balancing the need for personalization and the goal of minimizing unwanted recommendations.
4. Experiments and ablations with a real-world dataset, highlighting our approach's effectiveness in significantly reducing unwanted content.

2 Case Study: AVideo-Sharing Platform

We begin our study by examining data from a popular Chinese short-form video-sharing platform, Kuaishou, which has over 300 million daily users at the time of writing. We use the KuaiRand dataset [Gao et al., 2022], which was collected by sampling 27000 Kuaishou users who collectively watched approximately 32 million videos over one month (from April 8 to May 8, 2022). This dataset contains various user interaction signals, such as watch time for each video and whether users liked or commented on a video. The dataset also includes negative feedback signals, such as when users click the "Do not recommend" or "Report" buttons while watching a video. We focus exclusively on short videos, removing advertisements and videos with zero duration (e.g., dynamic photos). Notably, approximately 29% of users provided negative feedback on at least one video during the data collection period. Consequently, our analysis is restricted to this subset. After further filtering, our dataset consists of 7601 users, 56 million interactions, and 10 million unique videos. On average, a user watched 7, 417 videos during the collection period (median: 4, 623), with a large standard deviation of 9, 223.

Summary of the analysis. From our analysis 1 , we report some key insights that will be useful in designing riskcontrolling recommendation systems:

• The engagement and perceived harmfulness of videos are not fully correlated. Reported videos have the same

1 The full in-depth analysis of KuaiRand is available in Section 3 of the main manuscript [De Toni et al., 2025b].

watch time distribution as non-reported videos. Thus, maximizing engagement does not imply providing less harmful videos to users.

• In Kuaishou, the system suggests undesired content despite explicit user feedback (cf. → / Table 1). Since over 95% of users report less than 1% of videos they watch, recommender systems should maximize the usefulness of this little explicit negative feedback;
• With 21% of interactions resulting in no watch time and a strong peak near zero in view time distribution, very short engagement can serve as an early indicator of harmful or undesired content.

3 Recommender Systems With Risk Control

Given the key insights obtained from the analysis of Kuaishou data, let us now define more formally a risk-controlling recommender system. Consider a finite set of items I ∈ {i1, . . . , iN} and a finite set of users U ∈ {u1, . . . , uM}, each represented by d-dimensional feature vectors, respectively. Our recommender system outputs a set of items S(U) ⊆ I based on some relevance score estimated by a ranker f(U, I) ∈ R predicting how much an item I is relevant for a user U based on the information within U and I (e.g., in terms of watch-time). In two-stage recommender systems, we might perform further post-processing activities to filter certain items, before presenting the final recommendation to the user. Similarly to [Angelopoulos et al., 2023b], let us consider a post-processing step where we employ some threshold λ ∈ R to pick only certain items, based on a score s : U × I → R. For example, consider s(U, I = i) = f(U, I = i), we might want to keep only items with a predicted relevance above a certain base value:

After post-processing, our objective becomes providing the best set Sλ(U) ⊆ Tλ(U) maximizing a target metric computed on the recommendations (e.g., serendipity [Ge et al., 2010]). Generally, we are interested in recommending the most relevant items to each user, under a budget k ≤ N e.g., the number of slots in a YouTube main page. Let us define π(I) = {i1, . . . , iN} as the order induced by sorting the items given the ranker scores f(U, I = ij). If our ranker is reasonably good, then the optimal solution would be simply picking up to the k-th item from π(Tλ(U)):

In classical learning-to-rank tasks, we try to learn the optimal ranker f ∗ = argmaxf∈F E[ℓ(Sλ(U, k))] where ℓ : I → R is our target metric, such as the nDCG. Besides picking the most relevant items for the user, now we want to control the risk of the set Sλ(U, k) induced by the optimal ranker f ∗ . Let us assume we have a risk metric R : I → (0, 1) we would like to provably bound below a certain value α ∈ (0, 1). For example, we might want the fraction of unwanted content in Sλ(U, k) to be below a user-defined value. Since we want to avoid performing costly operations such as re-training the ranker f ∗ , an intuitive way to reach such an objective is by acting on the post-processing step, by finding the optimal λ threshold ensuring our risk is below a user-defined level α ∈ (0, 1):

In our setting, we consider an item unwanted for a user if she expresses distaste for it once seen within the recommendations. Given a set of recommendations Sλ(U, k), we denote its overall risk as the fraction of items the user has flagged once presented with it:

3.1 Conformal Risk Control

By simply plugging the previous risk definition in Eq. (3), we have our optimization objective. However, how do we pick the threshold λ to control the risk defined by Eq. (4)? We can find the optimal threshold via conformal risk control [Angelopoulos et al., 2023a] by employing a held-out calibration set {(u, i, h))j}ni=1 , where for each user-item pair we also record if it was flagged as unwanted (h = 1). We first state the main theorem from [Angelopoulos et al., 2023a], and then we explain its practical implications.

Theorem 1. Assume that RH(Sλ(U, k)) is non-increasing in λ, right-continuous, RH(Sλmax (U, k)) ≤ α and supλ RH(Sλ(U, k)) ≤ 1 < ∞ almost surely. Consider ˆR(λ) = 1n ∑nj=1 RH(Sλ(U = uj , k)), given a held-out calibration set. Given any desired risk upper bound α ∈ [0, 1], if we choose a threshold as:

then, we have E[R(Sˆλ (U, k))] ≤ α.

In practice, as long as our risk function decreases as λ increases, by choosing ˆλ as in Theorem 1, we are guaranteed the risk will be provably bound in expectation. If we consider the ranker output s(U, I) = f(U, I) as a score, as we increase the threshold, fewer items will be contained within Tλ(U). Thus, the likelihood of picking harmful items within Sλ(U, k) decreases monotonically since by removing items, we cannot increase RH(Sλ(U, k)).

4 Risk Control via ItemReplacement

Recent applications of conformal prediction to recommender systems consider only the simple filtering approach as detailed in Eq. (1) [Angelopoulos et al., 2023b; Xu et al.,

Algorithm 1 Risk-controlling Recommender System

Require: U, user, I, items, α, risk level, f : U × I → R, ranker, s : U × I → R+, score function, k set size

• 1: ˆλ ← RISKCONTROL(α) {Theorem 1}
• 2: Tˆλ (U) ← {i ∈ I : s(U, i) ≥ ˆλ}
• 3: T (safe) ← {i′ ∈ IU : H(U, i′ ) = 0 ∧W%(U, i′ ) > β}
• 4: T (replace) ˆλ (U) ← Tˆλ (U) ∪ T (safe) (U)
• 5: return {ij ∈ π(T (replace) ˆλ (U)) : j ≤ k}

2024]. The approach ensures the monotonicity requirements of Theorem 1 but at the price of losing control of the size of the resulting set Sλ(U, k) [De Toni et al., 2025b, Prop. 5.1]. Rather than simply removing items, we take an alternative approach, which consists of replacing potentially unwanted items with safer ones. However, how do we choose these safe items? Our approach exploits a simple fact: users consume previously seen items [Anderson et al., 2014; Dai et al., 2024]. Inspired by the analysis in Section 2, if we consider items the user has seen before, that were not reported as unwanted, we might as well suggest them again in place of potentially unsafe items. As shown in Table 1, there exist rare items (cf. → , 0.11%) that are reported only the second time a user watches them. Therefore, we need a sensible way to select previously seen videos for which we are reasonably sure the user will not report anew. We formalize this desideratum in a simple property that can be tested by employing historical interaction patterns:

Property 1. Denote with H1st ∈ {0, 1} the event that the user has reported an item the first time they saw it. There exists a variable C and a value β ∈ supp(C) such that the probability of an item being flagged a second time is zero P (H = 1 | H1st = 0, C > β) = 0.

In the case of Kuaishou, we can pick as C the watch time of videos. Then, we can simply choose videos on which the users have spent time W%(U, I) above a certain threshold. If Property 1 holds, we can replace items with a score below the risk control threshold with items the user has previously seen and not flagged as harmful, for which we have C > β. Let us denote the pool of safe videos as:

where IU ⊆ I indicates the items previously shown to the user U. Thus, the new pool of items will become:

It is straightforward to show that the new set T (replace) λ (U) satisfies the monotonicity requirements of conformal risk control, and it ensures we can always pick k items to show to the user [De Toni et al., 2025b, Proposition 2]. We consolidate the previous findings into Algorithm 1, a post-hoc algorithm applicable to any pre-trained recommender system.

5 Experiments With Real Data

We evaluate the effects of the risk control strategies (Eqs. (1) and (7)) on a simple two-stage recommender setup for a realistic video recommendation task, towards understanding what the impact of the conformal risk control on standard evaluation metrics (e.g., Recall@k and nDCG@k) is and what the effects of employing different strategies for risk control are (Eq. (1) and Eq. (7)). In the full paper, we further analyze other aspects, such as how many items we have to replace to achieve the desired risk level, or how approximately satisfying Property 1 impacts the desired unwanted content reduction [De Toni et al., 2025b, Sec. 7].

Dataset & Metrics. We use the KuaiRand dataset [Gao et al., 2022], which contains both positive and negative user feedback, along with information about interactions with previously seen videos. For our experiments, we focus on videos viewed at least twice by at least one user, resulting in a dataset that includes both repeated and single interactions. This leaves us with 5657 unique users, 117695 unique items, and more than 3 million interactions. Then, we partition the full interaction set into two datasets based on whether the interaction is repeated or not. We adopt a 10-core setting, randomly splitting the subset with only single interactions into training (70%), validation (15%), and test (15%) sets. The validation set also serves as a calibration set for determining the threshold λ as defined in Theorem 1. To ensure evaluation fidelity, we exclude all test set videos from the repeated interaction set. Thus, the repeated interaction set only considers videos previously seen by users - i.e., those appearing in either training or validation. We compare the effect of the risk control and various strategies on two widely used metrics: Recall@k and nDCG@k. In our experiments, we set k = 20 as is usually done in the literature [Chen et al., 2024]. We also measure the empirical risk, i.e., the fraction of unwanted content in the recommendations, as defined in Eq. (4).

Models. Similar to previous studies [Wang and Joachims, 2023], we evaluate the effectiveness of risk control using a two-stage recommender architecture. We consider two riskcontrol strategies: REMOVE (Algorithm 1 without line 3 and T (safe) = {}) and REPLACE (Algorithm 1). REMOVE is the classical risk-control strategy for recommender systems [Angelopoulos et al., 2023b]. For simplicity, under the REPLACE strategy, repeated items are not re-scored; instead, their original watch time is reused. We experiment with several stateof-the-art methods, including both sign-aware and unsigned graph-based models, depending on whether they incorporate negative user feedback (e.g., reported videos): LightGCL [Cai et al., 2023], GFormer [Li et al., 2023], SiReN [Seo et al., 2022], and SIGFormer [Chen et al., 2024]. For the re-ranking stage, we adopt a simple Neural Collaborative Filtering (NCF) model [He et al., 2017], as our focus is not on maximizing performance on the Kuaishou dataset, but rather on isolating the effects of risk control. Our full implementation is openly available 2 to support further research.

6 Results &Discussion

Fig. 1a shows how the level of harmfulness is tightly controlled across all strategies and ranking models. For any target reduction in unwanted content (α in Algorithm 1), the empirical reduction on the test set is equal to or better (i.e., all points lie on or below the diagonal) than the desired target, confirming that Algorithm 1 reliably enforces risk control. In general, the REMOVE strategy tends to exceed the target reduction, removing more items than strictly necessary. This proves that Algorithm 1 provides a provable link between user actions (e.g., flagging content) and the resulting changes in recommendations. Indeed, by specifying a desired reduction level, users can provably influence their experience with guarantees on the expected reduction in unwanted content. Figs. 1b and 1c show how risk control affects nDCG@20 and Recall@20. Removing items sharply degrades nDCG as the desired reduction level (α in Eq. (3)) increases, due to the shrinking pool of candidate items (Eq. (1)). In contrast, the REPLACE strategy retains a higher nDCG, though some degradation still occurs as previously seen items are reintroduced into the ranking. This performance drop is partly explained by our experimental setup: we do not rescore repeated items but instead reuse their original watch time, which may not reflect user interest upon second exposure. Recall decreases under both strategies, as removing or replacing items may discard potentially relevant, unseen content. Finally, Fig. 1d illustrates that, when full removal of unwanted content is required (α = 0), the sole option is to eliminate or replace all items in the recommendation set. Moreover, for the REPLACE strategy, we might recommend fewer than k items at test time, since we have finite replacement options.

Ultimately, our experimental results demonstrate that our framework can controllably reduce exposure to unwanted recommendations by providing a trade-off between the recommendation set size and utility. These findings point toward developing safer and user-centric recommender systems.

2 https://github.com/geektoni/mitigating-harm-recsys

Acknowledgments

The work of GDT, BL, and AP was partially supported by the following projects: Horizon Europe Programme, grants #101120237-ELIAS and #101120763-TANGO. Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Health and Digital Executive Agency (HaDEA). Neither the European Union nor the granting authority can be held responsible for them.