A rank-aware block decomposition for linear and bilinear operations in recommender models (FM, DCNv2, attention, FC) reduces redundant context feature computation to once per request with identity-equivalent results, plus rDCN variant for deeper layers.
Title resolution pending
7 Pith papers cite this work. Polarity classification is still indexing.
7
Pith papers citing it
fields
cs.IR 7representative citing papers
LENS restores target-specific control in latent-query CTR models via TCQG and TCPB modules plus QueryPos reference, reporting positive gains in all 12 backbone-dataset cells and a density-dependent conditioning rule.
Next Interest Flow models user intent as continuous evolutionary trajectories on a high-dimensional latent interest manifold with kinematic constraints, bidirectional alignment, and temporal causality mechanisms, yielding reported gains on industrial CTR data.
CCN applies contrastive learning on collaborative co-click/co-non-click signals to structure item representations for trigger-induced recommendations, showing 12.3% CTR and 12.7% order lift in an unseen Taobao scenario after training on a year of heterogeneous data.
Introduces versioned late materialization to eliminate data redundancy in ultra-long sequence training for DLRMs by storing histories once and reconstructing via pointers at training time.
AMEN aligns item-scene interactions via homogeneous spaces and a TSP mechanism to let all-domain movelines differentially affect CTR predictions, reporting +11.6% CTCVR lift in A/B tests.
SIREN unifies multi-modal and collaborative features for lifelong user interest modeling via semantic ID retrieval and target-aware transformer interactions, reporting SOTA GAUC and positive GMV gains in production.
citing papers explorer
-
Context Features Are Cheap: Rank-Aware Decomposition for Efficient Feature Interaction in Recommender Systems
A rank-aware block decomposition for linear and bilinear operations in recommender models (FM, DCNv2, attention, FC) reduces redundant context feature computation to once per request with identity-equivalent results, plus rDCN variant for deeper layers.
-
LENS: A Staged Design for Interaction Granularity in Sequential CTR Prediction
LENS restores target-specific control in latent-query CTR models via TCQG and TCPB modules plus QueryPos reference, reporting positive gains in all 12 backbone-dataset cells and a density-dependent conditioning rule.
-
Next Interest Flow: A Generative Pre-training Paradigm for Recommender Systems by Modeling All-domain Movelines
Next Interest Flow models user intent as continuous evolutionary trajectories on a high-dimensional latent interest manifold with kinematic constraints, bidirectional alignment, and temporal causality mechanisms, yielding reported gains on industrial CTR data.
-
Beyond the Trigger: Learning Collaborative Context for Generalizable Trigger-Induced Recommendation
CCN applies contrastive learning on collaborative co-click/co-non-click signals to structure item representations for trigger-induced recommendations, showing 12.3% CTR and 12.7% order lift in an unseen Taobao scenario after training on a year of heterogeneous data.
-
Versioned Late Materialization for Ultra-Long Sequence Training in Recommendation Systems at Scale
Introduces versioned late materialization to eliminate data redundancy in ultra-long sequence training for DLRMs by storing histories once and reconstructing via pointers at training time.
-
All-domain Moveline Evolution Network for Click-Through Rate Prediction
AMEN aligns item-scene interactions via homogeneous spaces and a TSP mechanism to let all-domain movelines differentially affect CTR predictions, reporting +11.6% CTCVR lift in A/B tests.
-
SIREN: Unified Multi-Granularity Semantic Interaction for Multi-Modal Lifelong User Interest Modeling
SIREN unifies multi-modal and collaborative features for lifelong user interest modeling via semantic ID retrieval and target-aware transformer interactions, reporting SOTA GAUC and positive GMV gains in production.