Jointly Modeling Selection and Engagement in Recommendation

28 concepts21 papers1 proposal

Traditional recommendation systems are often built around a single, binary question: Will the user click on this item? While optimizing for this single signal is straightforward, it introduces a fundamental vulnerability known as the "click trap." A click is a highly noisy indicator of true user preference. It captures immediate curiosity or the allure of clickbait, but completely fails to capture post-click satisfaction. Furthermore, relying solely on sparse click or purchase data ignores the rich, multi-faceted spectrum of user behaviors—such as scrolling, favoriting, adding to a cart, or dwelling on content—that occur during a typical session.

To build more robust recommendation engines, a growing body of research shifts the paradigm from single-signal optimization to multi-task engagement modeling. By training models to simultaneously predict both the initial selection (the click) and the subsequent depth of engagement (dwell time, completion rate, or interaction type), these frameworks leverage multi-task learning to regularize sparse data, uncover latent user intent, and align immediate actions with long-term user satisfaction.

The Multi-Task Architecture

The core mechanism behind this direction relies on a shared-bottom architecture. Instead of training isolated models for different user actions, these frameworks feed historical interaction sequences into a unified encoder—such as a Transformer, a Recurrent Neural Network (RNN), or a Multi-Layer Perceptron (MLP)—to learn a global representation of user preference. This shared representation then branches into multiple task-specific heads.

   [Task Head 1: Click?]    [Task Head 2: Dwell Time]    [Task Head 3: Share?]
             \                         |                         /
              \                        |                        /
               +-----------------------+-----------------------+
                                       |
                         [Shared Representation Layer]
                                       |
                         [User & Item Input Sequences]

These heads are typically designed to handle a mix of discrete and continuous objectives:

  • Discrete + Continuous Joint Prediction: In news and video feeds, models pair a binary classification head (predicting click probability) with a continuous regression head (predicting interaction duration). For example, the JUMP framework developed by Tengfei Zhou et al. pairs survival analysis with an RNN to jointly forecast the next click and its corresponding dwell time. Similarly, the HyperNews architecture by Rui Liu et al. utilizes a double-task network to predict binary interaction alongside continuous active reading time.
  • Multi-Behavior Classification: In e-commerce and search, the heads predict distinct types of actions along the conversion funnel. The BMLP framework by Weixin Li et al. expands single-outcome models into a multi-task setup that simultaneously forecasts clicks, favorites, and cart additions. In a similar vein, the IIARec model by Ahmed Rashed et al. uses hierarchical attention to predict both the next item in a sequence and the specific interaction type that will occur.

Why It Works: Regularization and Intent Modeling

Why does adding more tasks make the primary recommendation task perform better? The intuition lies in how auxiliary tasks constrain and enrich the shared latent space.

Preventing Overfitting on Sparse Data

When specific target interactions (like a purchase or a physical visit) are rare, a single-task model easily overfits. Auxiliary tasks provide a denser, more stable learning signal that guides the shared encoder toward robust features. For instance, in Point-of-Interest (POI) recommendation, specific location transitions are highly sparse. Tong Chen et al. address this in their SGRec model by introducing an auxiliary task that predicts the broader semantic category of the next POI alongside the specific venue. This forces the model to learn the underlying intent behind a trajectory, stabilizing representations when data is scarce.

In a different domain, Sajal Halder et al. design a dual-head Transformer that optimizes for discrete item recommendation while simultaneously predicting continuous queuing times, leveraging the structural interdependence between user interest and temporal constraints.

Grounding Representations via Contextual Intent

Auxiliary tasks can also act as situational anchors. In the NEON system developed by Xiaochong Lan et al., the model predicts a user's immediate living needs alongside their preferred fulfillment method (such as delivery versus an in-store visit). Learning the fulfillment modality serves as an auxiliary signal that implicitly informs the model about the user's physical constraints and situational context.

This strategy of modeling high-level intent is mirrored in the travel domain by Jia Xu et al., who use an auxiliary task to predict a user's unconsumed itinerary goals, allowing the system to cleanly distinguish between different behavioral modes, such as a routine business trip versus a vacation.

Aligning Immediate Actions with Long-Term Satisfaction

Beyond acting as a statistical regularizer, multi-task engagement modeling fundamentally alters what the system prioritizes by bridging the gap between immediate curiosity and sustained user value.

Filtering Out Noise with Post-Click Satisfaction

In content recommendation, a user's click preference does not always align with their ultimate reading satisfaction. To decouple these factors, Chuhan Wu et al. propose a multi-task framework that explicitly models both click probability and post-click reading satisfaction. By optimizing for both, the shared encoder learns feature representations that satisfy immediate engagement without sacrificing long-term content quality.

This concept is extended to multi-objective environments in their FeedRec framework (Chuhan Wu et al.), which optimizes for a spectrum of feedback loops—including click probability, content completion rates, and total dwell time—thereby reducing reliance on click-through rates alone and boosting long-term sharing ratios.

Capturing the Intensity of Engagement

When dealing with highly dynamic feeds, such as short videos, the system must distinguish between a user who quickly skips a video and one who watches it to completion. Dasong Li et al. tackle this by training a model to predict whether a user will watch past an initial time threshold alongside the total normalized watch time. This joint optimization forces the network to capture features relevant to both immediate viewer capture and sustained interest.

Similarly, the click-reweighting framework by Ruobing Xie et al. uses a multi-task architecture to predict valid interactions alongside a weighted engagement score, allowing the system to learn general features of user interest while explicitly optimizing for the intensity of the engagement.

Modeling Prerequisite Behaviors

Multi-task modeling also allows systems to capture the structural dependencies of user interfaces. In a scrollable feed, a user cannot click an item unless they choose to scroll past the current screen. The INSCAFER framework by Luo Ji et al. models this relationship by jointly predicting click and scroll probabilities as a Markov Chain. Because the tasks share underlying representations, the model effectively shares knowledge between these tightly coupled user actions, improving the accuracy of both predictions.

Broader Applications

The utility of joint engagement modeling extends across the entire user lifecycle and system architecture:

  • User Retention and Churn: Instead of treating churn prediction as an isolated classification task, Carl Yang et al. incorporate auxiliary tasks—such as identifying a user's behavioral category—into a joint training framework, leveraging latent behavioral patterns as early warning signals for user attrition.
  • Dynamic Environments and Simulators: To train reinforcement learning agents for interactive search, Qianlong Liu et al. build multi-task environment simulators that simultaneously predict user interaction events, session termination, and state transitions, capturing the semantic relationship between user intent and system suggestions.
  • Streaming Representation Learning: In streaming review applications, Xin Zhang 0108 et al. introduce next-item interaction prediction as an auxiliary task alongside sentiment analysis, ensuring that dynamically updating graph embeddings remain anchored to actual behavioral patterns over time.
  • Global Preference Matching: To balance immediate context with long-term alignment, Lingxiao Zhang et al. utilize a dual-task objective that combines bidirectional masked item prediction (local context) with a dedicated user-item matching task (global preference representation).

By moving away from single-metric optimization and embracing the multi-faceted nature of user behavior, multi-task engagement modeling provides a mathematically principled and intuitively sound framework for building recommendation systems that are accurate, resilient to data sparsity, and aligned with genuine user satisfaction.

Key Foundational Work

JUMP: a Jointly Predictor for User Click and Dwell Time (2018) — contributed 2 concept elements to this direction.
I Know You'll Be Back: Interpretable New User Clustering and Churn Prediction on a Mobile Social Application (2018) — contributed 1 concept elements to this direction.
Building Personalized Simulator for Interactive Search (2019) — 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_multi_task_engagement_modeling_0ebbcb8f,
  title = {multi-task engagement modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-task-engagement-modeling/rdr_0ebbcb8f59dd4add84f66c56833c71b5},
  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_multi_task_engagement_modeling_0ebbcb8f_v1,
  title = {multi-task engagement modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-task-engagement-modeling/rdr_0ebbcb8f59dd4add84f66c56833c71b5/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