Himalaya's LLM Self-Iterating Recommendation Parameters: Recall, Ranking, Re-ranking
Himalaya uses LLMs as user preference proxies to automatically optimize recommendation parameters across recall, ranking, and re-ranking stages, leveraging user context and long-term memory to generate personalized parameters via a parameter center, achieving significant metric gains for low-frequency users.
Background
Himalaya, a leading Chinese audio platform, uses a four-stage recommendation funnel: recall → coarse ranking → fine ranking → re-ranking. User feedback (clicks, completion) flows back for model iteration, forming a closed loop. Traditional models rely on user historical behavior as supervision signals, but audio scenarios suffer from three defects: sparse user interactions (lower than text/video), limited interest exploration due to vertical content categories and lack of personalized parameters, and high decision costs with delayed feedback signals (completion requires playback finish).
Overall Design: LLM as User Preference Proxy
The core idea: use LLMs to simulate user reasoning based on user context and long-term memory, generating personalized parameters for recall, ranking, and re-ranking stages. These parameters are deployed via the PBox parameter center, directly affecting online recommendations.
Core Modules
2.1 User Context & Long-Term Memory
User information is split into two layers:
Basic Attributes: Static user info (e.g., demographics) for initial profiling.
Historical Behavior: Sequences of clicks, plays, subscriptions, comments, shares, plus coarse category/tags. This enables LLM to reconstruct interest trajectories (e.g., news in morning, music at night).
2.2 LLM Reasoning Module
Key design choices:
Model Selection: Strong base LLM with good reasoning; adapted via prompt engineering without SFT.
Prompt Structure: Four-part template: role definition, task description, input specification, output constraints (JSON Schema for parseability).
Post-processing: Format validation, range checks (e.g., scores 0-1), anomaly filtering before writing to PBox.
2.3 PBox Parameter Center
PBox stores user-level personalized parameters covering the full funnel:
Recall: Personalized quota per channel, weight allocation, exploration quota.
Coarse Ranking: Paid content preference, quality preference.
Fine Ranking: Multi-objective fusion weights.
Re-ranking: Category diversity, content freshness, timeliness preferences.
Practice 1: LLM-Evaluated Personalized Recall Parameter Self-Iteration
Background
Recall channels include i2i, u2i, u2u, relevance, timeliness. Goal: quantify channel quality via LLM scoring to adjust quotas/weights.
Discriminative Evaluation Framework
Input: Candidate items (album/sound with title/tags) + user context/memory (hundreds of history records, long/short-term category preference scores).
Processing: LLM scores each item 1-5 based on match with user profiles.
Output: Scores aggregated per channel → channel quality profiles → increase quota/weight for good channels, decrease for poor, offline invalid ones. Parameters written to PBox.
Execution Flow
Get user's recall channel info.
Get user context & memory.
Feed both to LLM → output parameters to PBox.
PBox affects online recall.
Recommendation system returns results.
Concrete Example
Before: all users share default quotas (u2u:30, u2i:40, i2i:30) and weights (5.0, 3.0, 2.0). After LLM optimization per user:
User A: u2u match highest → quota 45, weight 6.0.
User B: i2i match highest → quota 55, weight 5.1, priority 3→1.
User C: u2i match highest → priority 2→1, weight 4.0.
Result: traffic shifts toward user-interest-aligned channels.
Results
Identified 10+ high-quality channels, 10+ potential, 20+ downgraded, 10+ taken offline. First quantitative evaluation of recall channels, enabling data-driven iteration.
Practice 2: LLM-Simulated Personalized Ranking Objective Fusion Self-Iteration
Background
Fine ranking uses multi-objective fusion: rank_score = ctr_bias + ctr_weight*ctr_score + dur_bias + dur_weight*dur_score + vipbuy_bias + vipbuy_weight*vipbuy_score (with scaling). Traditional tuning relies on manual A/B tests, slow and not personalized.
Evolution: v1 → v2 → v3
v1 (Generative): LLM directly outputs optimal weights. Failed: high-dimensional space, poor precision/stability.
v2 (Discriminative + User Profile): LLM scores ranked results; user profile added for preference simulation. Reduced complexity but parameter search still manual/grid.
v3 (Discriminative + User Profile + RL Parameter Optimization): LLM scores as reward signal → RL algorithm iterates parameters → closed loop: scoring → reward → parameter update → improvement. Final version deployed.
v3 Optimization Scheme
Input: Ranked candidate list (title, tags, paid info, tokenized title) + user context/memory (identity, search, play/like/subscribe sequences, aggregated interest tags).
Processing: LLM scores ranked list 0-1 (higher = more likely liked). Scores used as reward for RL parameter optimizer to search optimal weight combinations.
Output: Optimal per-user weights (ctr_weight, dur_weight, vipbuy_weight) written to PBox, affecting fine ranking.
Execution Flow
Get user's fine-ranking candidate logs.
RL optimizer proposes a parameter set.
Compute fusion scores, re-rank candidates.
Get user context/memory.
Feed re-ranked list + user info to LLM → score → feedback to RL optimizer.
Repeat until max iterations.
RL optimizer outputs optimal parameters → PBox → online ranking.
Concrete Example
Before: global defaults (ctr_weight 5.0, dur_weight 3.0, vipbuy_weight 2.0). After RL optimization per user:
User A (click-focused): 6.2, 3.4, 2.4.
User B (duration-focused): 4.1, 5.6, 1.2.
User C (paid-conversion-focused): 3.3, 2.2, 4.8.
Results
v3 deployed for low-frequency and veteran users. Low-frequency users: PVCTR +10.83%, avg play duration +12.16%, retention positive trend. Retention gains stabilize over time.
Practice 3: LLM-Optimized Personalized Re-ranking Business Weight Self-Iteration
Background
Re-ranking balances precision and diversity. Global uniform dispersion parameters fail: some users prefer deep vertical content (students → exam courses), others prefer cross-category exploration (professionals → career + emotional + finance). PBox enables personalized dispersion parameters, but manual tuning doesn't scale.
Optimization Scheme
Input: User behavior sequences (hundreds of records), user summary profile, basic info.
Processing: LLM infers preferences, scores three dimensions: category diversity (0-1), sound timeliness (0-1), album timeliness (0-1). Also extracts primary/exploration interest categories with intensities (0.1-1.0). Timeliness scores mapped to high/medium/low tiers: high [0.8,1.0], medium [0.6,0.8), low [0,0.6).
Output: Per-user re-ranking parameters (diversity, sound timeliness, album timeliness scores + interest categories/intensities) → PBox → affects re-ranking.
Execution Flow
Get user context & memory.
LLM infers preferences.
Aggregate & filter preferences → PBox → online re-ranking.
Recommendation system returns results.
Concrete Example
Before: global defaults (diversity 0.1, timeliness 0.9). After LLM:
User A: diversity 0.1→0.3, timeliness 0.9→0.7.
User B: diversity 0.1→0.6, timeliness 0.9→0.4 (strongest exploration).
User C: diversity 0.1→0.2, timeliness 0.9→0.8 (still timeliness-oriented).
Results
10+ LLM-optimized re-ranking parameter sets fully online, covering sounds, albums, cards. All full-traffic experiments show stable positive retention lift. For recall users, timeliness+diversity parameters yield retention improvements at the per-mille level.
Summary & Future Directions
Three practices validate LLM-based user simulation for recommendation parameter optimization, forming a closed loop: simulation → optimization → verification.
Recall: Quantitative channel evaluation (10+ identified each for quality/potential/downgrade/offline), enabling data-driven iteration.
Ranking: v3 (discriminative + profile + RL) deployed; low-frequency users: PVCTR +10.83%, duration +12.16%, retention positive.
Re-ranking: 10+ parameter sets online, stable retention gains, per-mille improvements for recall users.
Future work:
Integrate commercial goals (ad revenue, paid conversion) via multi-objective LLM optimization.
Expand parameter types and granularity across full funnel (recall, coarse, fine, re-ranking) for true per-user personalization.
Auto-research: feed platform-specific knowledge (content taxonomy, user behavior, audio specifics, iteration history) to LLM alongside frontier papers for faster strategy research and tuning.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
