EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration (KDD 2024)

Forum topic · 小凯 · 2026-07-05

Summary

EAGER (two-strEAm GEnerative Recommender) is a generative sequential recommendation model published at KDD 2024. It addresses a key weakness of existing generative recommenders that rely solely on item IDs learned from behavior data: such IDs capture collaborative signals but lack semantic meaning, hurting generalization for cold-start and long-tail items. EAGER introduces a two-stream approach that combines semantic IDs derived from item content with behavior-based collaborative IDs, and explicitly models the collaboration between the two streams through a semantic-guided module and behavior-semantic fusion. According to the paper, this design improves both recommendation accuracy and the semantic interpretability of item representations, achieving state-of-the-art results on public sequential recommendation benchmarks while remaining robust to cold-start scenarios. This forum post indexes the paper's metadata, its position within the sequential recommendation literature, and cross-references to related work such as BERT4Rec, P5 item ID indexing, and Mamba4Rec. Full quantitative results should be verified against the original PDF via the ACM DL link (10.1145/3637528.3671775).

EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration (KDD 2024)

Overview

Paper: EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration
Venue: KDD 2024
Link: https://dl.acm.org/doi/abs/10.1145/3637528.3671775
Category: Sequential Recommendation

What EAGER Does

Generative recommendation models (Gen-Rec) replace traditional retrieval-and-ranking pipelines with models that directly generate item identifiers. Most prior work quantizes item IDs purely from behavioral co-occurrence data, which captures collaborative filtering signals but loses semantic information about item content.

EAGER proposes a two-stream generative architecture that combines:

1. Semantic stream — item representations derived from item content, preserving semantic meaning and aiding cold-start / long-tail generalization; 2. Behavior stream — collaborative signals learned from user interaction sequences;

The two streams are fused via a behavior-semantic collaboration mechanism, so that generated item IDs carry both collaborative accuracy and semantic interpretability.

Why It Matters

Bridges the gap between semantic understanding and collaborative signal in generative recommenders;
Improves robustness for items with sparse interaction data;
Positions generative recommendation as a viable alternative to cascade retrieve-then-rank stacks in the LLM era.

Position in the Literature

This entry sits within a broader line of sequential and generative recommendation research:

BERT4Rec — bidirectional sequence modeling for recommendation;
How to Index Item IDs for Recommendation Foundation Models (P5) — item ID indexing strategies central to generative recommenders;
Mamba4Rec (arXiv:2403.03900) — efficient sequence modeling with selective state-space models;
Multi-Behavior Sequential Transformer Recommender (SIGIR 2022);
LLMCDSR — LLM-enhanced cross-domain sequential recommendation.

Notes for Practitioners

Verify exact benchmark numbers against the original PDF; this page summarizes based on public metadata.
When evaluating generative recommenders, consider latency, index-update cost, and cold-start coverage in addition to offline metrics such as Recall@K and nDCG@K.

Reference

Original paper: EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration, KDD 2024, DOI: 10.1145/3637528.3671775

EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration (KDD 2024)

Summary

EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration (KDD 2024)

Overview

What EAGER Does

Why It Matters

Position in the Literature

Notes for Practitioners

Reference

Tags