Many recommenders no longer represent a user as one vector. They keep several "interest vectors", one per facet of taste such as gardening, children's clothing or electronics, and they recommend an item if it matches any one of them strongly. That design makes a particular kind of explanation natural and cheap: "recommended because of your gardening interest." This note argues that such facet-level explanations work as a membership oracle. Each one tells an observer which of a user's facets fired. Over a stream of recommendations, those disclosures add up to the user's set of overlapping communities, meaning the groups of users and items a person belongs to at the same time, including sensitive ones.
We propose to measure this leak and then close most of it with a defense that keeps explanations faithful. Instead of naming the single facet that fired, the system names a small set of facets. The set always includes the true one, every member scores the item nearly as high as the winner, and the set is chosen so that the community profile an observer can assemble over time is shared by at least $k$ users. Put briefly, multi-interest models turn explanations into a way to read off community membership, and choosing explanation sets with an anonymity constraint can block that reading while the explanations stay honest.
Core Claim
In a multi-interest recommender, a user's score for item $i$ is $\max_k \mathbf{u}_k \cdot \mathbf{v}_i$, the best match between the item vector $\mathbf{v}_i$ and any of the user's $K$ interest vectors $\mathbf{u}_k$. This structure is the source of the accuracy gains reported by MIND (Multi-Interest Network with Dynamic Routing, over 60% higher HitRate@50 than the single-vector YouTube DNN on TmallData) and by ComiRec (Controllable Multi-Interest Framework). It also makes the "which facet fired" label a discrete, low-noise signal.
The overlapping-community literature treats a node's overlap as the set of distinct link clusters its edges fall into. One example is the link-community model fit by nonnegative matrix factorization in Link Community Detection Using Generative Model and NMF. The facet labels a user's explanations reveal are close to an edge-level labeling of that user's interactions. Tai et al. (Structural Diversity for Resisting Community Identification) showed that membership across several overlapping groups is a re-identification channel in published graphs. That finding motivates our claim. The mechanism differs: we expect the leak to travel through the explanation interface rather than through graph publication.
We make two predictions. First, an attacker who sees facet explanations will infer a user's overlapping community memberships and sensitive attributes much better than one who sees only the recommended items. Second, much of that excess leak can be removed by k-anonymous ambiguity sets without making explanations unfaithful.
How This Differs From Existing Work
No paper in our search connects multi-vector users to explanation leakage. The closest pieces each lack one ingredient.
Three defense lines tackle the transparency-security tradeoff on tabular models. FaithfulDefense (Models That Are Interpretable But Not Transparent) shrinks rule-set explanations to minimal faithful covers in order to slow model stealing. XRand (XRand) applies local differential privacy to SHAP rankings. Ambiguous Strategic Classification (Ambiguous Strategic Classification) reveals a set of $k$ candidate classifiers instead of one. We borrow the set-valued disclosure idea from the third paper and the cover-style optimization from the first. Our setting differs from all three: the protected quantity is a user's community profile in a recommender, not a classifier's decision rule.
EGSteal (How Explanations Leak the Decision Logic) shows that graph explanations enable model cloning. It works on small molecular graphs with one vector per graph.
Zhao et al. (Can One Embedding Fit All?) show that multi-interest models remove the accuracy penalty suffered by users with diverse tastes. Those diverse users are exactly the ones with the most overlapping memberships, and so the ones our attack should hit hardest. This creates a fairness-privacy tension we measure directly.
Research Program
Ground truth. On each dataset we build the user–item interaction graph and fit overlapping link communities with the NMF link-community method of Baquero et al. Each user's ground-truth profile is the set of link communities their edges belong to. That method has quadratic cost, so we run it on subsampled graphs of at most a few hundred thousand edges and report the subsample sizes.
Models. On the same data we train ComiRec-SA, ComiRec-DR and MIND-style routing. Each model exposes a facet explanation: the index of the winning interest vector, mapped to a human-readable name through the categories of the items closest to that vector.
Attack. The attacker registers shadow accounts with known histories, which lets them learn the map from facet names to communities. They then observe $T$ explanations for a target user and predict that user's community memberships and, where available, sensitive attributes.
Defense. For each recommendation of item $i$ to user $U$, the system discloses a set $S \subseteq \{1,\dots,K\}$ that satisfies two conditions:
$$k^\star \in S, \qquad \mathbf{u}_j \cdot \mathbf{v}_i \ge \max_k \mathbf{u}_k \cdot \mathbf{v}_i - \delta \ \ \text{for all } j \in S,$$
Here $k^\star$ is the facet that actually fired, and $\delta$ is a tolerance on the score gap. Under these conditions every disclosed facet nearly triggered the recommendation, so no explanation is false.
Among valid sets, a greedy cover chooses $S$ so that the cumulative disclosed signature $\sigma(\bigcup_t S_t)$ matches the signatures of at least $k$ other users. The signature $\sigma$ is the set of communities implied by all facets disclosed so far, and $t$ indexes the user's recommendations. When no valid set reaches $k$, the system falls back to the cheapest set that maximizes how many users share the signature.
The two knobs, $\delta$ and $k$, trace a frontier between leakage and explanation precision, measured as $1/|S|$.
Baselines
Recommendation-only reference. The naive reference is a multi-interest model that shows recommendations and no explanation. The attacker trains the same membership classifier on the sequence of recommended items alone. It is the "do nothing" point, and any leak beyond it is attributable to explanations. It runs on identical models and targets, and the attacker has the same shadow-account budget and $T$.
Single-vector reference. This is LightGCN, as used by Zhao et al., with item-level explanations of the form "because you interacted with X." It tests whether multiple vectors are the cause of the leak or whether any explanation leaks this much. It is tuned to match the multi-interest models' recall budget.
Full facet disclosure. This is ComiRec-SA/DR with the single true facet revealed. It is the worst case we aim to defend. For the underlying recommender, Cen et al. report the diversity score rising from 23.2% to 55.1% for about a 5% recall drop on Amazon Books, and a Recall@50 gain of 8.65% on the Alibaba industrial data. There is no reported result on Lacuna for membership-inference leakage through recommender explanations, so the attack numbers will be ours.
Randomized-response facet labels. This is the strongest existing defense. It adapts XRand's idea: report the true facet with probability set by a local-privacy budget $\epsilon$, and otherwise a random facet. XRand reduced explanation-guided backdoor success on EMBER LightGBM from 77.8% to 10.2% at $\epsilon=1.0$. That result is on a different task, so here it serves only to motivate using randomized response as the baseline. We sweep $\epsilon$ and compare at matched leakage.
Beating the baselines. Our defense beats the field if, at equal attacker AUC, it keeps higher explanation fidelity. Fidelity is defined as the true facet always included plus higher precision $1/|S|$, whereas randomized response sometimes outputs a false facet. Recommendation recall is unchanged by construction, since only the explanation changes.
Ablations. Two ablations isolate the key ingredient. In the first, ambiguity sets have the same size distribution but are drawn at random among the $\delta$-valid facets, with no anonymity optimization. In the second, $\delta = \infty$, which drops the near-trigger constraint and shows how much faithfulness costs.
Why not just…?
Why not just withhold explanations? Regulatory pressure and auditing make full opacity costly. XAudit (XAudit) shows that explanations cut audit query cost sharply, for example to a single query for linear models. The recommendation-only baseline also tests whether opacity even protects: if the attacker recovers communities from item lists nearly as well, withholding explanations buys little and our framing collapses. We will report that outcome as is.
Why not just add noise to the facet label? This is the real competitor. Randomized response gives a formal guarantee that our k-anonymity constraint does not. Its costs are that it sometimes tells users a false reason, and that per-explanation noise averages out over a long explanation stream. The second weakness is the same mechanism by which repeated observations erode protection in the analysis of Controlled Privacy Leakage Propagation Throughout Overlapping Grouped Learning. We plot leakage against $T$ for both defenses. If randomized response at matched fidelity leaks no more than our method, we will recommend it instead.
Why not just explain with public item categories instead of learned facets? If coarse category labels leak as much as learned facets, the problem is not specific to multi-vector models. We include category-level explanations as an extra arm. The prediction is that learned facets leak more because they track link communities more closely than a fixed taxonomy does. We test this directly by measuring how well facets and categories each align with the NMF link communities.
Proposed Experiments
Data and attacker targets. We use Amazon Books and Taobao, following the ComiRec setup, and MovieLens-1M, Epinion, Cosmetics and Anime, following Zhao et al. MovieLens-1M's released user age, gender and occupation serve as sensitive attributes. On the other datasets, a designated subset of link communities is treated as sensitive.
Metrics. Membership-inference AUC for sensitive communities and attributes is reported as a function of $T$. We also report Jaccard overlap between the inferred and true community profiles, Recall@50 and diversity, explanation precision $1/|S|$, and the rate of false explanations. Results are broken down by Zhao et al.'s diversity tiers to test whether the most diverse users leak most.
Decisive checks. The project lives or dies on three results. First, full facet disclosure must exceed the recommendation-only AUC by a clear margin; if it does not, the leak claim fails. Second, learned facets must leak more than category explanations; if they do not, the multi-vector angle fails. Third, the anonymity-optimized sets must dominate both random sets and randomized response on the leakage–fidelity frontier at large $T$; if they do not, the defense fails.
References
- Multi-Interest Network with Dynamic Routing for Recommendation at Tmall
- Controllable Multi-Interest Framework for Recommendation
- Link Community Detection Using Generative Model and Nonnegative Matrix Factorization
- Structural Diversity for Resisting Community Identification in Published Social Networks
- Models That Are Interpretable But Not Transparent
- XRand: Differentially Private Defense against Explanation-Guided Attacks
- Ambiguous Strategic Classification
- How Explanations Leak the Decision Logic: Stealing Graph Neural Networks via Explanation Alignment
- Can One Embedding Fit All? A Multi-Interest Learning Paradigm Towards Improving User Interest Diversity Fairness
- XAudit : A Learning-Theoretic Look at Auditing with Explanations
- Controlled privacy leakage propagation throughout overlapping grouped learning