Rec-APC Model: Churn-Aware Recommendation

• Rec-APC is a sequential recommendation model that leverages aggregated user preferences to update belief states from positive feedback while ending sessions on negative responses.
• The model formulates recommendation as a planning problem under partial observability, balancing exploration for improved belief accuracy against the high cost of churn.
• Its branch-and-bound algorithm and finite-horizon approximations yield efficient, near-optimal strategies in environments constrained by privacy and anonymity.

Searching arXiv for the requested topic and closely related papers. The Rec-APC model—short for Recommendation with Aggregated Preferences under Churn—is a sequential recommendation model for privacy-constrained settings in which a recommender system cannot rely on persistent individual-level user data and instead operates from aggregated population-level preference information. It formalizes a session-based decision problem in which an anonymous user is drawn from a known prior over latent user types, the system sequentially recommends content categories, positive feedback updates the posterior belief over user type, and negative feedback immediately terminates the session. In this formulation, recommendation is a planning problem under partial observability with an unusually severe exploration cost: unsatisfactory recommendations do not merely lower reward, they remove all future interaction opportunities. The model and its main theoretical and algorithmic properties are introduced in "Churn-Aware Recommendation Planning under Aggregated Preference Feedback" (Keinan et al., 6 Jul 2025).

1. Definition and problem setting

Rec-APC is designed for recommender-system environments shaped by privacy regulation and platform constraints that limit access to user-level behavioral traces. The motivating setting assumes that a system may know only aggregated preference information, such as user clusters, personas, or look-alike groups, rather than persistent user identities or detailed histories (Keinan et al., 6 Jul 2025). A session therefore begins from a cold-start belief over latent user types, and personalization must occur entirely within-session.

A Rec-APC instance is defined as a tuple

$\mathcal{I} = \langle M, K, \bm q, \bm P \rangle,$

where $M$ is the finite set of latent user types, $K$ is the set of recommendation categories, $\bm q \in \Delta(M)$ is the prior over user types, and $\bm P \in [0,1]^{K \times M}$ is the type-category preference matrix (Keinan et al., 6 Jul 2025). Each entry $\bm P(k,m)$ gives the probability that a user of type $m$ likes category $k$.

The session dynamics are binary and absorbing. At each round, the system recommends some $k \in K$. If the hidden user type is $m$, the recommendation is liked with probability $M$0 and disliked with probability $M$1. A positive response yields reward $M$2 and allows the session to continue; a negative response yields reward $M$3 and immediately ends the session (Keinan et al., 6 Jul 2025). This churn assumption is the defining feature of the model.

A useful clarification is that “aggregated preference feedback” in Rec-APC does not mean that within-session observations are aggregated. During a live session, the system still observes individual binary feedback from the anonymous user. The aggregation lies in the prior knowledge available to the system across sessions: it knows type-level preference statistics rather than individual-level histories (Keinan et al., 6 Jul 2025).

2. Latent types, policies, and belief dynamics

The hidden state of the model is the user’s latent type. The recommender never directly observes that type, but maintains a belief over $M$4. If the current belief is $M$5, the immediate probability that category $M$6 succeeds is

$M$7

This quantity is the expected one-step reward under the current uncertainty (Keinan et al., 6 Jul 2025).

When a recommendation succeeds, the model performs a Bayesian update. The posterior after observing positive feedback to category $M$8 is

$M$9

This is the main state-transition rule in Rec-APC (Keinan et al., 6 Jul 2025). Negative feedback is also informative in principle, but because it terminates the session, that information cannot be exploited operationally in future rounds.

Because only the all-positive history can generate later actions, a policy can be represented as an infinite sequence of recommended categories,

$K$0

with suffix notation $K$1 for the sequence starting at round $K$2 (Keinan et al., 6 Jul 2025). The belief walk induced by a policy $K$3 from an initial belief $K$4 is the deterministic sequence

$K$5

which is defined only along the survival path of repeated successes (Keinan et al., 6 Jul 2025).

This deterministic positive-feedback evolution is a central structural simplification. It distinguishes Rec-APC from more general partially observable recommendation models in which many different observation histories remain viable.

3. Objective and Bellman structure

The objective is to maximize the expected number of liked recommendations before churn. If $K$6 denotes the random number of likes obtained from a type-$K$7 user under policy $K$8, then the value of $K$9 under prior $\bm q \in \Delta(M)$0 is

$\bm q \in \Delta(M)$1

More generally, from belief $\bm q \in \Delta(M)$2,

$\bm q \in \Delta(M)$3

The optimal value is

$\bm q \in \Delta(M)$4

These are the paper’s basic value-function definitions (Keinan et al., 6 Jul 2025).

Rec-APC admits a compact recursive form. For any policy $\bm q \in \Delta(M)$5,

$\bm q \in \Delta(M)$6

The interpretation is immediate: the first recommendation must succeed for any future reward to exist, and conditional on success the system earns one reward and continues from the updated posterior (Keinan et al., 6 Jul 2025). The corresponding optimal Bellman equation is

$\bm q \in \Delta(M)$7

An equivalent closed-form expansion is

$\bm q \in \Delta(M)$8

The first expression views value as a sum over survival probabilities; the second decomposes value by latent user type (Keinan et al., 6 Jul 2025).

There is no explicit discount factor in the model. Instead, future rewards are implicitly discounted by session survival probabilities. This creates the core exploration–exploitation trade-off: exploratory actions may improve posterior concentration if liked, but they also increase the probability of immediate churn if they are poorly aligned with the current belief.

4. Structural properties of optimal policies

A notable result of the Rec-APC analysis is that simple greedy exploitation is not generally optimal. The paper gives examples in which the myopic policy—choosing at each step the category maximizing $\bm q \in \Delta(M)$9—can be arbitrarily worse than the optimal policy (Keinan et al., 6 Jul 2025). This shows that the posterior-improvement term $\bm P \in [0,1]^{K \times M}$0 can dominate immediate reward considerations.

At the same time, the model has strong long-run structure. The paper studies a class of well-separated instances defined via a separator quantity $\bm P \in [0,1]^{K \times M}$1, which is the minimum of three positive margins: cross-type distinguishability of category preferences, the gap between the best and second-best category within each type, and the distance of the maximum preference probability from certainty (Keinan et al., 6 Jul 2025). Formally, the well-separated class is

$\bm P \in [0,1]^{K \times M}$2

Within this class, beliefs can be classified as concentrated or unconcentrated. A belief is $\bm P \in [0,1]^{K \times M}$3-concentrated if $\bm P \in [0,1]^{K \times M}$4, and $\bm P \in [0,1]^{K \times M}$5-unconcentrated if no type has mass at least $\bm P \in [0,1]^{K \times M}$6 (Keinan et al., 6 Jul 2025). The analysis shows that on unconcentrated beliefs, optimal continuation value must increase by a definite amount after each successful recommendation. Because total value is bounded, only finitely many such unconcentrated steps can occur.

A complementary result shows that once the belief is sufficiently concentrated near a type $\bm P \in [0,1]^{K \times M}$7, the first action of the optimal policy is simply the category maximizing $\bm P \in [0,1]^{K \times M}$8. In that regime, pure exploitation becomes optimal because the immediate reward gap dominates any residual future-value differences (Keinan et al., 6 Jul 2025).

These ingredients yield the model’s central structural theorem: for any well-separated instance, there exists a finite time $\bm P \in [0,1]^{K \times M}$9 such that for all $\bm P(k,m)$0,

$\bm P(k,m)$1

Thus optimal policies in Rec-APC converge in finite time to a constant recommendation, i.e. to pure exploitation (Keinan et al., 6 Jul 2025). This does not mean exploration is never useful; rather, it means optimal exploration is always finite. The system eventually settles on repeatedly recommending a single category.

A plausible implication is that Rec-APC transforms the usual recommender-system exploration problem into a front-loaded planning problem: informational actions matter early, but once the posterior enters a sufficiently concentrated region, further experimentation is strictly dominated.

5. Finite-horizon approximations and dynamic programming

Although Rec-APC is defined as an infinite-horizon model, the paper gives finite-horizon approximations. Let

$\bm P(k,m)$2

and assume $\bm P(k,m)$3, excluding degenerate cases in which a category is guaranteed to succeed for some type and repeated forever can yield infinite value (Keinan et al., 6 Jul 2025). Under this assumption, value is bounded by

$\bm P(k,m)$4

The paper defines the finite-horizon value $\bm P(k,m)$5 and shows that the infinite-horizon problem can be approximated to any $\bm P(k,m)$6 using a depth

$\bm P(k,m)$7

This yields an $\bm P(k,m)$8-accurate truncation of the infinite-horizon objective (Keinan et al., 6 Jul 2025).

For small numbers of categories, the authors exploit a commutativity property of positive-feedback Bayesian updates: $\bm P(k,m)$9 This implies that after a fixed number of successful recommendations, the resulting belief depends only on the multiset of categories recommended, not on their order (Keinan et al., 6 Jul 2025). A finite-horizon dynamic program can therefore be built over category-count vectors rather than full action sequences: $m$0 This is useful when $m$1 is small.

For small numbers of user types, the paper also outlines a belief-POMDP approach based on point-based value iteration. These are presented as warm-up methods, not as the main solution strategy (Keinan et al., 6 Jul 2025).

6. Branch-and-bound algorithm

The principal algorithmic contribution of Rec-APC is a branch-and-bound procedure over recommendation prefixes (Keinan et al., 6 Jul 2025). The search space is the infinite tree

$m$2

where each finite prefix $m$3 represents all infinite policies that begin with that recommendation sequence.

For any policy $m$4 and horizon depth $m$5, the value decomposes as

$m$6

This isolates the contribution of a prefix from its continuation value (Keinan et al., 6 Jul 2025).

The algorithm uses simple continuation bounds. An upper bound assumes that after the prefix, the recommender magically learns the user type and can repeatedly recommend that type’s best category: $m$7 A lower bound is given by the best fixed-action policy under the current belief: $m$8 Substituting these into the prefix decomposition yields upper and lower bounds for every subtree rooted at $m$9 (Keinan et al., 6 Jul 2025).

The branch-and-bound procedure maintains the best incumbent lower bound and expands only those children whose optimistic upper bound exceeds the incumbent by more than $k$0. The paper proves finite termination and $k$1-optimality: $k$2 Termination follows because upper-minus-lower bound gaps decay geometrically with prefix depth as a function of $k$3 (Keinan et al., 6 Jul 2025).

This algorithm is specialized to the geometry of Rec-APC. Unlike general POMDP solvers, it exploits the fact that only positive-observation paths generate continuation states and that optimal policies eventually become constant.

7. Relation to POMDPs and recommendation theory

Rec-APC can be expressed as a POMDP in which the hidden state is the user type, actions are recommendation categories, observations are binary feedback, and negative feedback transitions to an absorbing terminal state (Keinan et al., 6 Jul 2025). In belief space, it becomes an MDP over posterior distributions.

What makes Rec-APC distinctive is not merely that it is a POMDP, but that it is a churn-absorbing belief process with deterministic positive-path posterior updates. More general recommendation POMDPs may include nonterminal negative feedback, evolving user states, graded observations, or continued interaction after dissatisfaction. Rec-APC strips the model to a sharply privacy-aware and churn-sensitive core.

Compared with individual-level sequential recommendation models, Rec-APC also omits persistent personalization. Every session begins from the prior $k$4, and all adaptation is intra-session. This suggests that the model is best viewed not as a substitute for full personalized recommendation, but as a normative framework for settings in which regulatory or product-design constraints force recommendation planning to operate at the level of cohorts or personas.

The paper compares its branch-and-bound method against SARSOP, a standard point-based POMDP solver (Keinan et al., 6 Jul 2025). A plausible interpretation is that the comparison highlights where problem structure matters: generic belief-space methods can be competitive when the number of categories is much larger than the number of types, but the specialized Rec-APC solver benefits when latent-type complexity is the main difficulty.

8. Empirical findings

The experiments in the Rec-APC paper use both synthetic instances and MovieLens-derived instances (Keinan et al., 6 Jul 2025). Synthetic preference matrices are generated by sampling latent vectors for types and categories, forming affinities via negated cosine distances, and normalizing to probabilities, while priors over types are sampled from Gaussian logits passed through a softmax. Runtime comparisons average over 500 instances at precision $k$5.

One empirical question is whether optimal policies actually exhibit rapid convergence in practice. The theory guarantees only finite convergence, not a sharp rate. Empirically, however, posterior uncertainty along the optimal belief walk decays rapidly and is well fit by exponentials, with reported curve fits having $k$6 (Keinan et al., 6 Jul 2025). This suggests that optimal exploration tends to be brief.

The main algorithmic comparison shows that the Rec-APC branch-and-bound method outperforms SARSOP in runtime when the number of user types is large or when the numbers of types and categories are comparable (Keinan et al., 6 Jul 2025). When the number of categories grows large while the number of types remains small, SARSOP can be faster. The paper reports that all pairwise comparisons in its synthetic scaling studies are statistically significant under Wilcoxon signed-rank tests with $k$7 (Keinan et al., 6 Jul 2025).

For a more realistic data construction, the authors transform MovieLens 1M into a Rec-APC instance by applying spectral co-clustering to users and movies, using user clusters as latent types and item clusters as recommendation categories (Keinan et al., 6 Jul 2025). The prior $k$8 is estimated from cluster sizes, and the preference matrix $k$9 is estimated from normalized mean ratings between user and item clusters. Across these MovieLens-derived instances, the branch-and-bound method again consistently outperforms SARSOP in runtime, with the advantage increasing as matrix dimension grows (Keinan et al., 6 Jul 2025).

These experiments validate the practical tractability of the model in the regime for which it was designed: anonymous sessions with nontrivial latent-type structure.

9. Applications and interpretation

Rec-APC is most naturally applied to recommendation systems in which users are effectively anonymous at session start and only segment-level preference models are available. Examples include privacy-restricted media platforms, onboarding flows, incognito or cookie-restricted traffic, or systems operating under policies that prohibit persistent cross-session identity tracking. The model is also relevant wherever dissatisfaction has immediate abandonment consequences, such as short-form content feeds, landing-page recommendation, or first-impression personalization.

The latent-type interpretation is intentionally broad. The paper refers to user types as personas, clusters, or look-alike groups (Keinan et al., 6 Jul 2025). This makes the model compatible with many practical segmentation pipelines. At the same time, the action space is category-level rather than item-level, which suggests that Rec-APC is better suited to strategic recommendation planning than to full ranking over extremely large catalogs.

A plausible implication is that Rec-APC can serve as a planning layer above a conventional recommender stack: the upstream system estimates $k \in K$0 and $k \in K$1 from aggregated logs, and the Rec-APC planner selects which coarse content type to expose next under churn risk. The paper itself does not propose such a systems architecture, but the model supports it conceptually (Keinan et al., 6 Jul 2025).

10. Limitations and open questions

The Rec-APC model is deliberately stylized. Its most consequential assumption is that negative feedback immediately ends the session. This makes the mathematics tractable and embeds churn directly into the objective, but it is stronger than the behavior of many real users, who may tolerate multiple poor recommendations. Likewise, feedback is binary, whereas operational recommenders often observe clicks, skips, dwell time, abandonment delay, or multi-level ratings (Keinan et al., 6 Jul 2025).

Another limitation is that the model assumes $k \in K$2 and $k \in K$3 are known. Estimating these from aggregated data is a separate statistical problem not addressed in the paper. The work is therefore a planning contribution rather than a learning contribution. It also assumes that the latent user type remains fixed throughout the session, ruling out contextual drift or fatigue.

At the algorithmic level, the branch-and-bound method is shown to terminate with $k \in K$4-optimality, but the paper does not establish a polynomial-time exact algorithm for the general case, nor a hardness result (Keinan et al., 6 Jul 2025). The authors explicitly identify this as an open direction. Another open direction is extending the convergence analysis to richer models with nonterminal negative feedback or more complex observation processes, where the current deterministic belief-walk arguments would no longer apply.

It would be mistaken to interpret Rec-APC as a general-purpose replacement for full personalized recommendation. The model instead isolates one increasingly important regime: privacy-constrained, anonymous, churn-sensitive sequential recommendation under aggregated preference knowledge.

11. Conceptual significance

Rec-APC is significant because it formalizes a recommendation problem that classical personalized RS frameworks often sidestep: what optimal planning looks like when individual identity is absent, segment-level preference priors are available, and bad actions destroy the chance to learn further. Its main theoretical result—that optimal policies eventually converge to pure exploitation in finite time—shows that even under severe partial observability, the structure of privacy-aware session recommendation can be sharply characterized (Keinan et al., 6 Jul 2025).

Equally important is the methodological message. Although Rec-APC is expressible as a POMDP, its specific probabilistic structure yields stronger insights and faster computation than generic belief-space solvers in the intended regime (Keinan et al., 6 Jul 2025). This suggests that recommendation under regulatory and privacy constraints may benefit from specialized models rather than being treated as a minor variant of classical personalized ranking.

In summary, the Rec-APC model is a hidden-type, binary-feedback, churn-absorbing sequential recommendation framework in which a recommender starts from an aggregated prior over user types, updates beliefs only after positive responses, and seeks to maximize expected successes before session termination. Its theory shows finite-time convergence of optimal policies to constant exploitation on well-separated instances, and its branch-and-bound solver provides an efficient exact-to-$k \in K$5 planning method that performs well empirically on both synthetic and MovieLens-derived tasks (Keinan et al., 6 Jul 2025).