Technical Summary: Bridging Passive and Active via Active Expression Modeling

Problem Definition

The paper addresses the limitations of Conversation Starter Recommendation in Large Language Model (LLM)-driven conversational search. Currently, these systems rely on a closed "exposure-click-training" loop, where models are trained exclusively on user interactions with pre-defined, passive starter suggestions. This mechanism creates a systemic feedback loop and an echo chamber:

1. Bias towards Popularity: The system tends to recommend common, generic starters because it only learns from what it has previously exposed, failing to capture dynamic, open-world user intents.
2. Data Sparsity & Distribution Shift: While users can actively type free-text queries (active expressions) to bypass menus, these signals are ignored in traditional starter recommendation. Incorporating them is difficult due to:
   • Distribution Shift: A gap exists between structured, pre-defined starters (closed world) and unstructured, colloquial active queries (open world).
   • Non-ID-able Nature: Active queries are unique text strings, rendering traditional item-ID based popularity statistics (used for debiasing in industrial streaming) ineffective.

Methodology: PA-Bridge Framework

The authors propose PA-Bridge (Passive-Active Bridge), a framework designed to integrate active user expressions into proactive starter recommendation while mitigating distribution shifts and popularity bias. The architecture builds upon a standard Two-Tower model (User Tower and Query Tower) and introduces two core components:

1. Distribution Aligner (Adversarial Learning)

To bridge the semantic gap between active queries ($q_a$) and passive starters ($q_s$), the authors employ an Adversarial Learning mechanism during offline training:

• Generator ($G$): Projects raw textual representations from both sources into a unified latent intent space.
• Discriminator ($D$): Attempts to distinguish whether a representation originates from an active query or a passive starter.
• Objective: The Generator is trained to fool the Discriminator, forcing the encoder to learn domain-invariant representations that strip away source-specific noise, effectively aligning the distributions of the two data sources.

2. Semantic Discretizer (RQ-VAE)

To enable debiasing for the "non-ID-able" active queries, the authors introduce a Semantic Discretizer based on Residual Quantized Variational Autoencoders (RQ-VAE):

• Quantization: The continuous semantic embeddings of queries are quantized into discrete semantic codes (clusters) using a learned Codebook.
• Debiasing Proxy: Instead of relying on unique item IDs, the system uses the streaming frequency of these discrete semantic codes ($P(c_q)$) as a proxy for probability density.
• Loss Function: A semantic-debiased loss is formulated by adjusting logits with $-\log P(c_q)$ and reweighting samples based on the discriminator's classification probability ($\beta_q$). This allows the system to mitigate the bias toward popular active queries and surface long-tail intents without explicit item IDs.

The final training objective combines the debiased loss and the adversarial loss in a multi-task learning framework.

Key Contributions

1. Paradigm Shift: The paper pioneers the integration of active user expressions (manually typed queries) into proactive query recommendation, aiming to shatter the harmful feedback loop and echo chamber effect inherent in closed-world starter systems.
2. PA-Bridge Framework: A novel architecture that utilizes adversarial distribution alignment to unify open-world and closed-world intents, and semantic discretization (RQ-VAE) to enable ID-free debiasing in a unified latent space.
3. Industrial Validation: The method is validated through extensive offline and online evaluations in a real-world production environment, demonstrating robustness against distribution shifts and data sparsity.

Experimental Results

Offline Evaluation

Using a large-scale industrial dataset (7.2 billion samples), the authors compared PA-Bridge against several baselines:

• Semantic Retrieval (Basic): Trained only on starter logs.
• ID-Based Retrieval: Standard industry approach with learnable IDs.
• Naive Augmentation: Directly adding active queries without alignment.
• Ablation Studies: Removing alignment or reweighting modules.

Findings:

• PA-Bridge achieved the highest AUC of 64.65%, a 6.72% relative improvement over the basic semantic model.
• Naive augmentation (Basic + Active) resulted in a performance drop (-2.71%), confirming that distribution alignment is critical.
• Semantic modeling significantly outperformed ID-based retrieval, which suffered from overfitting to historical feedback loops.

Online A/B Testing

Deployed on the Bytedance platform, the model showed significant business value:

• User Active Days: +0.04% increase (noted as significant for a mature, massive-scale application).
• Feature Penetration Rate: +0.54% increase.
• Diversity Metrics:
  • Unique Queries Exposed: +30.65%.
  • Unique Queries Clicked: +56.50%.
    These results indicate that the system successfully exposed a more diverse range of queries and effectively mitigated the feedback loop.

Significance and Claims

The paper claims that by harnessing the "free will" of users through active expressions, it is possible to break the structural flaws of closed-loop recommendation systems. The significance lies in:

1. Breaking the Echo Chamber: Moving beyond reactive, popularity-biased suggestions to capture dynamic, open-world intents.
2. Solving the ID-Free Debiasing Challenge: Providing a viable solution for popularity debiasing in scenarios where items (queries) are unstructured and lack discrete IDs, a common challenge in industrial streaming training for conversational AI.
3. Practical Deployment: Demonstrating that these theoretical advancements translate into measurable improvements in user engagement and system diversity in a production environment.