Adaptive Interest Cardinality in Multi Interest Recommendation

42 concepts32 papers1 proposal

Early recommender systems compressed a user’s entire interaction history into a single latent vector. While computationally efficient, this "one-size-fits-all" approach creates a middle-ground trap: it averages distinct preferences into a single, generic profile. If a user likes both heavy metal music and children's lullabies, a single-vector representation often recommends soft rock—a mathematical average that satisfies neither interest.

To solve this, modern architectures transitioned to multi-interest recommendation. As pioneered by Liu et al. in their work on the Multi-Interest Network with Dynamic Routing (MIND), systems began using capsule networks and attention mechanisms to extract multiple distinct "interest vectors" per user.

However, these frameworks introduced a new architectural bottleneck: fixed interest cardinality. They rely on a static hyperparameter, $k$, to pre-define the maximum number of interest slots available for every user or session. Recent research focuses on breaking this constraint, moving toward systems that dynamically scale representation capacity based on the actual complexity of user behavior.

The Problem with Fixed-Capacity Representations

Applying a uniform number of interest clusters across an entire population fails because human behavior varies drastically in breadth and complexity. As highlighted by Xu et al. in Can One Embedding Fit All?, a static allocation forces an unnatural trade-off between under-representing complex users and over-complicating simple ones.

Fixed k=3 Allocation:
User A (Focused):   [ Shopping for Shoes ]  -->  Allocates 3 slots  -->  [Slot 1: Shoes] [Slot 2: Noise]  [Slot 3: Noise]
User B (Diverse):   [ Tech, Books, Cooking, Sports ] --> Allocates 3 slots --> [Slot 1: Tech]  [Slot 2: Books]  [Slot 3: Cook+Sport (Blurred)]

This mismatch manifests in three primary ways:

  1. The Over-Parameterization of Focused Sessions: For a user with a singular, narrow focus (e.g., browsing exclusively for running shoes), a high fixed $k$ forces the model to allocate empty interest slots. This introduces noise, causes data sparsity, and wastes computational resources, as noted by Li et al. in the context of news recommendation (MINER).
  2. The Compression of Multifaceted Histories: For highly active users or long browsing sessions containing diverse categories, a low fixed $k$ forces distinct motivations to merge. Early work by Matsumoto on Query Integration demonstrated that static grouping thresholds either over-cluster distinct interests or under-cluster similar ones, failing to adapt to individual perceptions of similarity.
  3. The Shared-Account Dilemma: On platforms where multiple individuals share a single profile (such as a household Netflix or Amazon account), the number of latent personas shifts constantly. Guo et al. demonstrated in DA-GCN that assuming a fixed number of users per account leads to severe under-modeling or over-modeling, preventing the system from cleanly untangling mixed behavioral streams.

How Dynamic Interest Scaling Works

To accommodate fluid user behavior, recent architectures dynamically infer the optimal number of interest vectors on the fly. Instead of routing items to a fixed set of slots, these models scale their internal representation capacity using several distinct mechanisms.

Session-Length Adaptive Regularization

In session-based recommendations, the number of underlying user intents is highly correlated with time and session length. Short sessions rarely contain multiple distinct goals, whereas long sessions naturally drift.

To exploit this, Lv et al. developed DMI-GNN, which introduces a dynamic regularization term that scales based on session length. This mathematical constraint penalizes the generation of excess interest vectors in short sequences, effectively preventing "interest overload" while allowing long, complex browsing sessions to expand their representation capacity freely.

Graph Coarsening and Message Passing

Rather than forcing items into pre-defined categories, graph-based recommenders learn the interest structure directly from interaction topology.

The Personalized Interest Graph Recommender (PIGR), proposed by Zhou et al., abandons fixed interest slots entirely. Instead, it employs message passing over a learned session graph to aggregate semantically related items into "coarsened nodes." If a session features highly uniform items, the graph collapses into a single coarsened node; if the session explores diverse categories, the network naturally preserves multiple distinct nodes.

Item Interaction Graph         Message Passing & Coarsening       Adaptive Latent Interests
    [Item A] --- [Item B]               (Clustered)                 [Interest Node 1]
          \     /
         [Item C]
            |
    [Item D] --- [Item E]               (Clustered)                 [Interest Node 2]

Reinforcement Learning for Preference Decomposition

Interests are not static; they emerge, interact, and fade over time. Fixed clustering techniques lack the temporal flexibility required to capture this fluidity.

To address this, Shao et al. designed SPLIT, a framework that uses Reinforcement Learning (RL) to decompose a user's single interaction history into an adaptive number of evolving "preference threads." The RL agent decides in real-time whether a new action belongs to an existing interest thread or warrants the creation of a new one, allowing the system's cardinality to expand and contract dynamically.

Multimodal Latent Spaces

In generative sequential modeling, users are traditionally mapped to a single Gaussian distribution within a Variational Autoencoder (VAE). This assumes a unimodal preference structure.

The SemantIc Gaussian Mixture VAE (SIGMA), introduced by Li et al., replaces this unimodal assumption with a flexible Gaussian Mixture Model (GMM) latent space. By allowing the network to dynamically allocate the number of active mixture components (modes) based on the semantic diversity of the user's history, the model naturally scales its representation to match the user's true preference complexity.

The Paradigm Shift

The transition from fixed-capacity modeling to adaptive interest cardinality represents a fundamental shift in how recommender systems conceptualize users. By treating interest capacity as a dynamic variable rather than a static hyperparameter, modern recommendation engines eliminate the need for manual dataset tuning, protect sparse user profiles from noisy over-parameterization, and accurately preserve the rich, multifaceted diversity of highly active human behavior.

Key Foundational Work

Content-Based Music Retrieval Using Query Integration for Users with Diverse Preferences (2007) — contributed 1 concept elements to this direction.
Efficient Model Learning from Joint-Action Demonstrations for Human-Robot Collaborative Tasks (2015) — contributed 1 concept elements to this direction.
Content-Aware Collaborative Music Recommendation Using Pre-trained Neural Networks (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dynamic_interest_scaling_e4c32951,
  title = {dynamic interest scaling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-interest-scaling/rdr_e4c32951f70a48fbbe9a1605e7970f09},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_dynamic_interest_scaling_e4c32951_v1,
  title = {dynamic interest scaling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-interest-scaling/rdr_e4c32951f70a48fbbe9a1605e7970f09/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake