Bilibili Game Center Recommendation System: Architecture, Core Technologies, and Experimental Results

1. Introduction

Bilibili Game Center is a game distribution platform that offers game downloads, reviews, guides, and activities. Users can discover games, reserve tests, download resources, and interact with the community. Figure 1‑1 shows the homepage.

1.2 Background and Significance

Although the Game Center users are a subset of the main site, they form a massive audience. The recommendation system bridges the huge catalog of games with these users, alleviating information overload and driving revenue through game downloads and channel commissions.

1.3 Challenges

System implementation must be flexible to adopt new algorithms and strategies.

Model building faces dynamic user interests, high trial cost for games, sparse and biased interaction data, and the need to balance user satisfaction with platform profit.

1.4 Goals and Evaluation Metrics

The primary goal is to increase B‑side game channel revenue by improving user LTV. Key metrics include CTR, CVR, UV‑CTR, UV‑CVR, revenue per 10k impressions, NAU, and LTV (see Table 1‑1).

2. Overall Architecture

The system consists of a Feature Platform, Deep Learning Platform, Online Service Platform, AB Experiment Platform, and supporting infrastructure for data ingestion, storage, and stability.

3. Core Technologies

3.1 Feature Engineering

User features: demographics, Game Center behavior, in‑game behavior, interests.

Game features: basic attributes, content data, evaluation data, operational metrics.

Context features: time, location, device, etc.

3.1.2 Feature Platform

Provides unified registration, computation, evaluation, selection, and sample generation for all features, enabling reuse across algorithms such as advertising and churn prediction.

3.2 Recommendation Models

The pipeline is split into Recall, Ranking, and Re‑ranking stages (Figure 3‑2).

3.2.1 Recall Models

Popular recall

New‑game recall

ICF (item‑based collaborative filtering)

Tag‑based recall

Dual‑tower u2i and u2i2i models

Multiple channels are weighted and fused to produce the final candidate set (Figure 3‑3).

3.2.2 Ranking Model

A multi‑objective model jointly optimizes click‑through and download probability, leveraging multi‑order feature crossing, attention mechanisms, and sequence modeling (Figure 3‑4).

3.2.3 Re‑ranking Model

Strategy shuffling

MMR diversity shuffling

Activity‑period forced insertion

New‑game support

3.3 Online Services

Recall Service aggregates results from offline inverted indexes and real‑time recall (Figure 3‑5 & 3‑6). Ranking Service scores candidates using cached feature data to reduce latency (Figure 3‑7).

3.4 AB Service

Provides experiment management with group mutual exclusion, layered orthogonal designs, confidence testing, whitelist handling, and traffic reversal (Figure 3‑8).

3.5 Sample Stitching and Model Training

Flink windows are used to balance sample timeliness and label accuracy. A mixed training mode combines high‑integrity low‑latency samples with low‑integrity high‑latency samples for incremental and full‑batch training (Figure 3‑9).

4. Practical Effects

Experiment design compared the current system (80% traffic) with the initial version (20%). Results showed a 78% increase in CTR, 76% increase in CVR, 119% rise in per‑user clicks, and 146% rise in per‑user downloads, confirming the effectiveness of the personalized recommendation system.

5. Summary and Outlook

The system contributed to business growth, robust architecture, and methodology transfer to other algorithmic services. Future directions include multimodal game representations and new recommendation scenarios such as game assets and community content.