STAR: Multi-Domain CTR Predictor

• The paper presents the STAR model which employs a star-shaped neural architecture with domain-specific adapters to enhance multi-domain CTR prediction.
• It introduces innovative Partitioned Normalization and domain indicator embeddings to balance shared learning with domain-specific adaptation.
• The model, deployed in Alibaba’s display advertising, significantly improves AUC, CTR (+8%), and RPM (+6%) while maintaining efficient online inference.

The Star Topology Adaptive Recommender (STAR) is a multi-domain click-through rate (CTR) prediction framework deployed in large-scale industrial recommender systems, notably Alibaba’s display advertising platform. STAR is designed to address the challenges of simultaneously serving multiple “business domains” that exhibit both shared and domain-specific patterns in user-item interactions, which complicate traditional monolithic or fully independent recommendation approaches. By employing a star-shaped neural architecture and adaptive normalization mechanisms, STAR achieves efficient cross-domain parameter sharing while respecting inter-domain distinctions, leading to substantial online and offline gains in CTR and revenue metrics (Sheng et al., 2021).

1. Multi-Domain CTR Prediction: Problem Formulation

Multi-domain CTR prediction requires that, given an incoming sample $(x, y, p)$—where $x$ comprises user, item, and context features; $y \in \{0, 1\}$ is a binary click label; and $p \in \{1, \dots, M\}$ specifies one of $M$ distinct business domains—the recommender accurately estimates a click probability $\hat{y} \in [0, 1]$ at high throughput and low compute cost. The core challenges arise from (i) overlapping user/item sets across domains (justifying parameter sharing) and (ii) significant divergence in domain-specific data distributions (necessitating adaptation). Traditional approaches—using either a fully shared global model or a set of $M$ independent models—suffer from underfitting and resource inefficiency, respectively.

2. STAR Model Architecture

The STAR model’s distinguishing feature is its “star topology,” consisting of a central shared MLP and lightweight domain-specific adapters (spokes). In each hidden layer $\ell$ (input $d_{\ell-1}$, output $d_\ell$):

• Shared parameters: $x$0.
• Domain-specific adapters: $x$1.

Combined domain parameters are defined elementwise: $x$2 ($x$3 is elementwise multiplication). Forward propagation for domain $x$4 proceeds by substituting $x$5 at every layer. This structure preserves global inductive bias while allowing each domain to fine-tune the shared weights.

The input pipeline consists of shared embeddings for user, item, and context raw IDs. These embeddings are concatenated into a fixed-length vector $x$6, then normalized by a domain-aware normalization layer, and entered into the multi-layer star-topology MLP. An auxiliary two-layer MLP consumes a learned domain embedding $x$7 (obtained from the one-hot domain indicator) concatenated with select context features, producing a scalar auxiliary score $x$8. The final prediction

$x$9

combines the main and auxiliary networks.

3. Domain Adaptation Mechanisms

STAR employs two complementary domain adaptation strategies:

1. Domain Indicator Embeddings: Each domain is represented as a learned embedding $y \in \{0, 1\}$0. This embedding is used (i) as input to the auxiliary MLP and (ii) as a source of scale and bias parameters in normalization.
2. Partitioned Normalization (PN): Distinct from typical Batch Normalization, which aggregates moments across all domains, PN maintains separate running means $y \in \{0, 1\}$1 and variances $y \in \{0, 1\}$2 for each domain, plus domain-specific affine parameters $y \in \{0, 1\}$3. The normalization of pre-activation $y \in \{0, 1\}$4 for domain $y \in \{0, 1\}$5 is expressed as

$y \in \{0, 1\}$6

During inference, $y \in \{0, 1\}$7 and $y \in \{0, 1\}$8 are replaced by $y \in \{0, 1\}$9 and $p \in \{1, \dots, M\}$0. This ensures that domain-specific distributional characteristics are respected throughout the network.

A summary of the STAR adaptation components:

Component	Shared	Domain-Specific
Embedding Table	✓
MLP (Hidden Layers)	✓	Adapters ($p \in \{1, \dots, M\}$1)
Partitioned Normalization		PN stats, $p \in \{1, \dots, M\}$2
Auxiliary Network		$p \in \{1, \dots, M\}$3

4. Training Objective and Optimization Scheme

The STAR model is trained end-to-end by minimizing the sum, over all domains, of the standard binary cross-entropy loss (for CTR prediction) plus regularization terms for both shared and domain-specific parameters: $p \in \{1, \dots, M\}$4 Mini-batches of size 2000 are sampled “domain-wise,” i.e., a domain $p \in \{1, \dots, M\}$5 is selected, and $p \in \{1, \dots, M\}$6 datapoints are drawn from $p \in \{1, \dots, M\}$7. The Adam optimizer is employed with a learning rate of 0.001. To handle rapid shifts in domain traffic (“non-stationary traffic splits”), STAR maintains a sliding buffer of recent samples; all batches are drawn i.i.d. from this buffer, mitigating the risk of domain distribution drift and preserving training stability.

5. Empirical Evaluation: Production Data and Baselines

STAR was evaluated offline and online using industrial-scale data from Alibaba display ad serving, comprising 19 business domains (each with a unique click-through rate distribution: 1.27–12.05%) and a training set containing billions of impressions sampled in one-day aggregates.

Baselines included:

• Base: A DIEN-style single model trained on pooled data with standard BatchNorm.
• Shared Bottom: Shared embedding, per-domain MLP.
• MulANN: Domain-adversarial network.
• MMoE: Multi-gate mixture-of-experts architecture.
• Cross-Stitch: Domain-specific MLPs with cross-model parameter transfer.

Key results:

• STAR obtained an overall offline AUC of 0.6506 versus Base (0.6364), Shared Bottom (0.6398), MMoE (0.6403), and Cross-Stitch (0.6415), outperforming all baselines on every domain by 0.005–0.02 AUC.
• Ablation studies revealed that (i) replacing BatchNorm with Partitioned Normalization in the Base model yielded a +0.0121 AUC gain, (ii) using the “star” factorized MLP with BatchNorm provided +0.0091 AUC, and (iii) combining both (full STAR) increased AUC by +0.0142 versus Base.
• Online A/B testing (60+ domains in production) demonstrated +8.0% CTR and +6.0% RPM (Revenue Per Mille) improvement relative to the previous production model (Sheng et al., 2021).

6. Deployment and System Considerations

STAR has been in production for Alibaba’s display advertising system since late 2020. To ensure low-latency prediction and resource efficiency under strict SLAs, each domain’s combined MLP weights $p \in \{1, \dots, M\}$8 are pre-computed and cached, enabling inference-time costs equivalent to a standard Shared-Bottom model. The sliding-buffer mechanism supports continuous online learning and adaptation without major training interruptions.

Resource overhead is minimal: the extra parameters required for domain-specific adapters and PN scales constitute less than 0.1% of the memory footprint compared to the shared embedding tables (which typically dominate model size in industrial recommenders). The architecture is designed for scalable, stable online operation across dozens of business domains.

7. Significance and Related Methodologies

The STAR architecture represents an effective compromise between rigid parameter sharing and full per-domain specialization, leveraging star-topology factorization and domain-partitioned normalization to achieve both model compactness and adaptability. Its empirical superiority over Shared Bottom, multi-expert, domain-adversarial, and cross-stitch baselines demonstrates the benefit of sophisticated domain adaptation over simple model duplication or naive shared models.

By demonstrating robust improvements in both offline metrics (AUC) and online business KPIs (CTR, RPM), as well as operational fit for production-scale, multi-domain recommenders, STAR provides a validated blueprint for future adaptive architectures in real-world multi-distribution settings (Sheng et al., 2021).