When Relevance Meets Novelty: Dual-Stable Periodic Optimization for Serendipitous Recommendation

This paper proposes the Co-Evolutionary Alignment (CoEA) method, which integrates a Dual-Stable Interest Exploration module to model both group identity and individual interests and a Periodic Collaborative Optimization mechanism to establish a dynamic closed-loop feedback system, thereby overcoming the limitations of static optimization and biased interest modeling in LLM-enhanced serendipitous recommendation.

The Gist

Imagine you have a very smart, well-read friend who loves recommending movies, restaurants, and books to you.

The Problem with Current Friends (Traditional Systems):
Right now, most recommendation systems are like that friend who only remembers what you liked last week. If you watched three cooking shows, they will only show you more cooking shows. If you buy running shoes, they only show you more running gear.

• The Trap: You get stuck in a "feedback loop." You see the same things over and over, you get bored (content fatigue), and you never discover that you actually love gardening or jazz music.
• The AI Attempt: Recently, people tried using super-smart AI (Large Language Models) to break this loop. They built a system with two AIs: one that knows what you like (Relevance) and one that tries to find cool, new stuff (Novelty).
• The Flaw: Even these AI duos have two big problems:
  1. They forget who you really are: They only look at what you clicked today. They miss the fact that you are a "tech enthusiast" or a "foodie" in the grand scheme of your life. So, the "New Stuff" AI suggests things that are weirdly random, not actually interesting to you.
  2. They don't learn from mistakes: They set up the system once, and then it just sits there. If you start liking something new, the system doesn't update its strategy. It's like a teacher who grades your test once and never teaches you again.

---

The Solution: CoEA (The "Co-Evolving" Friend)

The authors of this paper propose a new method called CoEA (Co-Evolutionary Alignment). Think of it as upgrading your recommendation friend into a super-observant, lifelong mentor who evolves with you.

Here is how it works, using two main tricks:

1. The "Dual-Stable" Lens (Seeing the Big Picture and the Moment)

Most systems look at your history like a single, blurry photo. CoEA looks at two photos at the same time:

• The Long-Term Portrait (Group Identity): It looks at your entire life history to figure out your "tribe." Are you a "Tech Geek"? A "Parent on the Go"? A "Budget Traveler"? This is your Group ID. It's stable; you don't change your tribe every day.
• The Short-Term Snapshot (Current Mood): It looks at what you clicked just now. Maybe you're a "Tech Geek," but right now you're looking for a "Kitchen Gadget."

The Analogy: Imagine you are a Baker (Long-term identity).

• Old System: You bought flour yesterday. Today, it only shows you more flour.
• CoEA: It knows you are a Baker (Group ID). It sees you are currently looking at a specific type of mixer (Short-term). It suggests, "Hey, since you're a Baker, you might also love this obscure sourdough starter kit you've never seen." It balances who you are with what you need right now.

2. The "Periodic Loop" (The Never-Ending Study Session)

This is the secret sauce. Instead of setting the system and forgetting it, CoEA runs a continuous training loop.

• Step 1 (The Explorer): The "Novelty AI" suggests a bunch of new, weird categories based on your profile.
• Step 2 (The Judge): The "Relevance AI" acts like a strict editor. It checks: "Does this actually fit this user?" It gives a score.
• Step 3 (The Feedback): If the "Novelty AI" suggests something good, the "Relevance AI" says "Good job!" If it suggests something bad, it says "Nope, try again."
• Step 4 (The Update): The "Novelty AI" learns from this feedback and gets smarter. Then, the cycle starts again with new data.

The Analogy: Think of it like a musical band practicing together.

• The "Novelty AI" is the guitarist trying out new, crazy riffs.
• The "Relevance AI" is the drummer keeping the beat and saying, "That riff was cool, but that one was off-key."
• They practice together every day (Periodic Optimization). Over time, the guitarist gets better at playing riffs that fit the band's style perfectly, without losing their creativity.

Why This Matters (The Results)

The researchers tested this on real data from Meituan (a massive Chinese delivery app) and a movie dataset.

• Better Quality: People actually clicked on the recommendations more often because they felt relevant.
• Better Discovery: People found new categories they loved (like finding a new hobby) without getting bored.
• Real World Success: When they put this in the actual Meituan app, it made the company more money (more orders) and showed users more diverse items.

Summary

CoEA is like a recommendation system that finally "gets" you. It knows your deep personality (your group identity) and your current mood. It doesn't just guess once; it constantly learns from your reactions, getting better at suggesting the perfect mix of "familiar comfort" and "exciting new discoveries." It breaks the echo chamber and helps you find the serendipitous moments you didn't know you were looking for.

Technical Summary

Here is a detailed technical summary of the paper "When Relevance Meets Novelty: Dual-Stable Periodic Optimization for Serendipitous Recommendation."

1. Problem Statement

Traditional recommendation systems often suffer from the "feedback loop" trap, where they over-recommend content aligned with a user's historical preferences. This leads to:

• Content Fatigue: Users are trapped in a narrow interest bubble.
• Limited Exploration: The system fails to introduce novel items that users might actually enjoy.
• Bias in Existing LLM Solutions: While Large Language Models (LLMs) show promise for generating diverse content, current dual-model frameworks (using separate "Novelty" and "Relevance" LLMs) have two critical flaws:
  1. Interest Modeling Bias: They rely heavily on short-term behavior, ignoring long-term group identity (e.g., "tech enthusiast"), causing novelty exploration to deviate from the user's stable, deep-seated needs.
  2. Static Optimization: Existing methods perform a one-time alignment between models, failing to leverage incremental user data for continuous, closed-loop optimization.

2. Methodology: Co-Evolutionary Alignment (CoEA)

The authors propose CoEA, a framework designed to balance novelty exploration with preference alignment through two core modules: Dual-Stable Interest Exploration (DSIE) and Periodic Collaborative Optimization (PCO).

A. Dual-Stable Interest Exploration (DSIE)

This module addresses the lack of group identity in interest modeling by jointly processing long-term and short-term behavioral sequences in parallel.

• Long-Term Modeling (Group Identity):
  • Filtering: Noisy long-term sequences are filtered based on click frequency.
  • Embedding: Sparse user attributes and the filtered sequence are processed via Causal Self-Attention (CSA) to generate a long-term interest representation.
  • Group Clustering (RQ-VAE): The representations are mapped to discrete Group Collaborative Semantic IDs (CSID) using a Residual-Quantized Variational AutoEncoder (RQ-VAE). This captures stable group preferences.
  • Profile Generation: Representative users from each cluster are used to generate textual group profiles (e.g., "Tech enthusiasts prefer flagship phones") via a Profile LLM.
• Short-Term Modeling (Individual Timeliness):
  • Recent interactions are mapped directly to fine-grained category sets to capture immediate, dynamic interests without clustering.
• Integration: The DSIE module outputs a user representation that combines the stability of group identity (CSID) with the timeliness of individual short-term interests.

B. Periodic Collaborative Optimization (PCO)

This module resolves the static optimization flaw by establishing a bidirectional closed-loop iteration between the Novelty LLM and the Relevance LLM.

• Novelty LLM (Generator):
  • Input: Combines the Group CSID (profile) with recent short-term categories.
  • Task: Generates candidate novel categories.
  • Training: Fine-tuned via Supervised Fine-Tuning (SFT) to predict future categories and Direct Preference Optimization (DPO) to align with user preferences while avoiding catastrophic forgetting (using KL divergence constraints).
• Relevance LLM (Evaluator):
  • Role: Acts as a reward model.
  • Task: Scores the candidate categories generated by the Novelty LLM against actual user clicks (positive samples) vs. non-clicks (negative samples).
  • Training: Optimized via a discriminative loss function to distinguish between relevant and irrelevant novel items.
• The Loop:
  1. Novelty LLM generates candidates.
  2. Relevance LLM verifies and scores them.
  3. Verified candidates are stored in an offline database.
  4. The Relevance LLM's feedback (scores) is used to incrementally fine-tune the Novelty LLM.
  5. The updated Novelty LLM generates new candidates for re-scoring, creating a sustainable, evolving cycle.

3. Key Contributions

1. Co-Evolutionary Alignment (CoEA): A novel framework that integrates long-term group preference modeling with dynamic dual-model collaboration to solve the Exploration-Exploitation dilemma.
2. Dual-Stable Interest Exploration (DSIE): A mechanism that constructs user representations by parallel processing of long-term (group identity via RQ-VAE/CSID) and short-term (individual timeliness) sequences, preventing novelty exploration from drifting away from user needs.
3. Periodic Collaborative Optimization (PCO): A bidirectional closed-loop mechanism where the Relevance LLM and Novelty LLM continuously co-evolve. It uses incremental data and DPO with KL constraints to prevent catastrophic forgetting while maintaining relevance.
4. Comprehensive Evaluation: Extensive offline and online experiments validating the framework's effectiveness in industrial-scale scenarios.

4. Experimental Results

The method was evaluated on Movielens-1M and a large-scale industrial dataset from Meituan (MTRec).

• Offline Performance:
  • Quality: CoEA outperformed state-of-the-art baselines (including Google-v1/v2 and traditional models like SASRec) in Category Hit Rate (C-H@K) and NDCG (C-N@K).
  • Novelty: It achieved significant gains in Novel Category Proportion (NCP) and Category Long-Tail Proportion (CLTP), with a 38.24% improvement in CLTP@10 on MTRec compared to the strongest baseline.
  • Ablation Studies: Removing the long-term sequence (DSIE) caused a ~5.9% drop in quality metrics, confirming the importance of group identity. Removing the Relevance LLM (PCO) caused quality degradation, proving the necessity of the closed-loop alignment.
• Online A/B Test (Meituan App):
  • Deployed in the Meituan homepage recommendation system.
  • Results: Achieved a +1.203% increase in Gross Transaction Value (GTV) and a +2.364% increase in 7-Day Novel Item Exposure Proportion (7D-NIEP).

5. Significance

This paper addresses a critical bottleneck in modern recommendation systems: how to introduce novelty without sacrificing relevance or user satisfaction.

• Theoretical Advance: It moves beyond simple short-term sequence modeling by introducing Group Collaborative Semantic IDs (CSID) to capture stable, latent user identities.
• Practical Impact: The Periodic Collaborative Optimization mechanism provides a scalable solution for industrial systems to continuously adapt to user behavior changes without manual re-engineering, effectively breaking the "filter bubble" while maintaining high business metrics (GTV).
• Industrial Viability: The successful deployment on Meituan demonstrates that complex LLM-based dual-model frameworks can be operationalized efficiently using offline storage strategies and periodic updates, balancing computational cost with recommendation quality.