💻 computer science

End-to-End Semantic ID Generation for Generative Advertisement Recommendation

This paper proposes UniSID, an end-to-end framework for generative advertisement recommendation that unifies embedding and Semantic ID generation to overcome the limitations of traditional two-stage Residual Quantization, thereby achieving superior performance through multi-granularity contrastive learning and summary-based reconstruction.

Original authors: Jie Jiang, Xinxun Zhang, Enming Zhang, Yuling Xiong, Jun Zhang, Jingwen Wang, Huan Yu, Yuxiang Wang, Hao Wang, Xiao Yan, Jiawei Jiang

Published 2026-05-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jie Jiang, Xinxun Zhang, Enming Zhang, Yuling Xiong, Jun Zhang, Jingwen Wang, Huan Yu, Yuxiang Wang, Hao Wang, Xiao Yan, Jiawei Jiang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The Problem with "Translation"

Imagine you are trying to teach a robot to recommend products to people. To do this, the robot needs to understand what a product is.

The Old Way (The Two-Stage Pipeline):
Think of the old method like a game of "Telephone" played by two different people who don't talk to each other.

Person A (The Encoder): Looks at a product (say, a red leather shoe) and writes a long, complex description on a piece of paper. They then try to summarize that description into a short code (like "Shoe-Red-01").
Person B (The Generator): Takes that short code ("Shoe-Red-01") and tries to guess what the user wants next.

The Problem:
Person A and Person B have different goals. Person A wants to write a perfect summary. Person B wants to predict the next click. Because they work separately, Person A might leave out a tiny detail that Person B really needed. Also, every time Person A summarizes the paper, they lose a little bit of information (like a blurry photocopy). By the time the code reaches Person B, the original "vibe" of the shoe is a bit fuzzy. This is what the paper calls Semantic Degradation and Objective Misalignment.

The New Solution: UniSID (The "All-in-One" Chef)

The authors propose UniSID, which changes the game. Instead of two people playing Telephone, they hire one expert Chef who does everything at once.

How UniSID Works:

The Raw Ingredients: The Chef looks at the actual product data immediately: the photo of the shoe, the text description, the brand name, and the category tags (e.g., "Men's Fashion").
The Simultaneous Cooking: Instead of writing a summary first and then guessing, the Chef cooks the "Code" (the Semantic ID) and the "Flavor Profile" (the Embedding) at the exact same time.
- They look at the raw ingredients and directly decide: "This is a high-quality, stylish, black leather shoe."
- They generate the code and the deep understanding of the shoe simultaneously.
The Result: Because the code is made directly from the raw ingredients, no information is lost in translation. The code perfectly captures the essence of the shoe.

The Secret Ingredients (How They Make It Better)

To make sure this "Chef" doesn't just guess randomly, UniSID uses two special techniques:

1. The "Zoom Lens" Strategy (Multi-Granularity Contrastive Learning)
Imagine looking at a map.

Level 1 (Zoomed out): You see "North America."
Level 2 (Zoomed in): You see "USA."
Level 3 (Close up): You see "New York City."

In the old method, the robot might get confused between "USA" and "Canada" because the levels weren't taught together. UniSID teaches the robot to understand that "New York" is inside "USA," which is inside "North America." It forces the code to be consistent at every level of zoom, ensuring the robot knows exactly how specific or general a product is.

2. The "Hidden Meaning" Detective (Summary-Based Reconstruction)
Sometimes, a product description is literal, but the real meaning is hidden.

Input: "Stainless steel cup, 500ml, for hiking."
Hidden Meaning: "This is for an adventurous, outdoor lifestyle."

UniSID has a special trick. It asks a smart AI to write a short summary of the product's "vibe" (the hidden meaning). Then, it challenges the robot: "Can you look at the code you just made and guess this summary?"
If the robot can't guess the "adventurous lifestyle" from the code, it knows the code is missing something important. This forces the code to capture not just the facts, but the soul of the advertisement.

The Results: Why It Matters

The paper tested this new "Chef" (UniSID) against the old "Telephone players" (RQ-VAE, RQ-KMeans) using real data from Tencent's advertising systems.

Better Codes: The codes (Semantic IDs) UniSID made were much more accurate at grouping similar products together.
Better Recommendations: When used to recommend ads, UniSID got significantly more clicks (up to 4.62% better) than the best existing methods.
No More Blur: By skipping the "two-stage" translation, the system didn't lose the fine details of the products.

In a Nutshell

The paper argues that the current way of turning products into computer codes is like making a photocopy of a photocopy—it gets blurry and loses details. UniSID is a new method that looks at the original document and writes the code directly, using a single, unified process. It also uses a "quiz" system to make sure the code captures the hidden personality of the product, not just its surface features. The result is a smarter, more accurate recommendation system for ads.

Technical Summary: UniSID – End-to-End Semantic ID Generation for Generative Advertisement Recommendation

1. Problem Statement

Generative Recommendation (GR) has emerged as a powerful paradigm by framing recommendation as a next-token prediction task, relying on Semantic IDs (SIDs) to tokenize large-scale items into discrete sequences. However, existing GR approaches predominantly utilize a two-stage cascading compression paradigm based on Residual Quantization (RQ). In this pipeline, items are first encoded into dense embeddings (Stage 1) and subsequently quantized into discrete SIDs (Stage 2).

The authors identify three inherent limitations in this prevailing two-stage paradigm:

Objective Misalignment: The embedding learning stage is optimized for semantic richness, while the SID generation stage is optimized for discrete token prediction. This decoupling prevents end-to-end co-optimization, leading to suboptimal SID representations.
Semantic Degradation: Since SIDs are generated solely from pre-trained embeddings, the process cannot directly utilize raw item features (e.g., multimodal signals or structured attributes). This bottleneck causes the loss of critical semantics and reduces SID fidelity.
Error Accumulation: RQ hierarchically quantizes embeddings by approximating residuals. Quantization noise accumulates across levels, and deeper layers observe only the residual from the previous stage, resulting in progressively sparser information and noisier, less reliable SIDs at deeper levels.

The core research question is whether a unified end-to-end SID generation framework can be designed to break away from this two-stage cascading compression paradigm.

2. Methodology: UniSID

The authors propose UniSID, a unified framework for generative advertisement recommendation that jointly optimizes embeddings and SIDs directly from raw advertising data. The framework consists of three primary components:

2.1 Advertisement-Enhanced Input Schema

To address semantic degradation, UniSID linearizes heterogeneous advertising signals into a unified token sequence processed by a shared Multimodal Large Language Model (MLLM). The input includes:

Ad Instruction Prompts: Explicitly guiding the model to generate SIDs and embeddings.
Unstructured Modalities: Visual content (images) and textual descriptions (titles/descriptions).
Structured Ad Attributes: Industry and hierarchical category information (e.g., Daily Necessities → Tableware → Drinkware → Water Cup) to resolve semantic ambiguity inherent in raw images/text.
Learnable Tokens: A set of learnable SID tokens and an Embedding token appended to the input.

The MLLM encodes these inputs into hidden states. A SID Head projects the representations of the SID tokens into the SID space, while an Embedding Head projects the embedding token representation into the item embedding space. Crucially, the embedding token is conditioned on the preceding SID sequence, allowing it to integrate both raw data and coarse-to-fine semantic information.

2.2 Unified SID and Embedding Generation

Unlike RQ methods that hierarchically quantize residuals, UniSID predicts each layer of the multi-layer SID directly from the same full contextual advertisement. This ensures that all SID layers access complete advertising information, alleviating progressive information sparsification.

The model outputs logits for $L$ semantic layers.
Discrete SID tokens are obtained via an argmax operation on the logits for each layer.
The final SID is the concatenation of these layer-wise tokens.

2.3 Multi-Granularity Contrastive Learning

To enforce semantic consistency across different SID levels, the authors introduce a multi-granularity contrastive learning strategy.

Instead of a fixed positive set, the method adaptively determines positive relationships based on ad relevance at each semantic level $l$ .
Positive sets ( $P_l$ ) consist of items sharing the same semantic category at level $l$ .
The loss function ( $\mathcal{L}_{sid}$ ) is optimized independently at each granularity level, ensuring that fine-grained SIDs do not absorb coarse-level noise and that hierarchical consistency is maintained.
A standard contrastive loss ( $\mathcal{L}_{emb}$ ) is also applied to the item embeddings.

2.4 Summary-Based Ad Reconstruction

To encourage SIDs to capture latent high-level semantics not explicitly present in raw data, a summary-based reconstruction mechanism is proposed.

Summary Generation: A frozen LLM summarizes structured ad attributes into a high-level semantic summary ( $s^{sum}_i$ ) that infers implicit information (e.g., target audience intent).
Reconstruction: The model concatenates the SID embedding and item embedding, projects them through a reconstruction head, and uses an LLM to reconstruct the semantic summary via next-token prediction.
Objective: This process ( $\mathcal{L}_{rec}$ ) acts as an auxiliary supervision signal, compelling SIDs to encode discriminative, high-level semantic features.

2.5 Joint Optimization

The total training objective combines the three components:
$\mathcal{L}_{total} = \mathcal{L}_{sid} + \mathcal{L}_{emb} + \lambda \mathcal{L}_{rec}$
This unified design allows SIDs and embeddings to mutually reinforce each other, avoiding the objective inconsistency of two-stage pipelines.

3. Key Contributions

Identification of Limitations: The paper formally identifies and analyzes the limitations of the prevailing two-stage RQ paradigm: objective misalignment, semantic degradation, and error accumulation.
UniSID Framework: A novel end-to-end framework that jointly optimizes SIDs and embeddings from raw data, utilizing an advertisement-enhanced input schema and full-context multi-layer prediction.
Semantic Preservation Mechanisms: The introduction of multi-granularity contrastive learning to enforce level-specific semantic consistency and a summary-based reconstruction mechanism to capture latent high-level semantics.
Empirical Validation: Comprehensive evaluation across industrial advertising datasets and public benchmarks.

4. Experimental Results

The authors evaluated UniSID on two industrial advertising datasets (Ad-60W, Ad-100W) and a public benchmark (Amazon Beauty).

SID Quality: On the Ad-60W dataset, UniSID achieved a 1.86% to 3.09% improvement in V-Measure (clustering quality) across three semantic layers compared to the strongest baseline (RQ-Kmeans).
Next-Ad Prediction: On the Ad-100W dataset, UniSID improved the Hit Rate (HR@5) by 4.62% over RQ-Kmeans.
Embedding Retrieval: On the Ad-60W dataset, UniSID outperformed state-of-the-art multi-modal embedding methods (e.g., VLM2Vec2) with a massive 45.46% improvement in Recall@1.
Public Benchmark: When integrated into the TIGER framework on the Amazon Beauty dataset, UniSID improved Recall@5 by 11.83% over the best baseline (TIGER-LETTER).

Ablation Studies confirmed that:

Joint Training is superior to single-task training, as the mutual reinforcement between SID and embedding generation enhances both.
Multi-Granularity Contrastive Loss outperforms standard distribution-matching losses (KL, JS) and BCE loss in preserving hierarchical semantic structures.
Summary-Based Reconstruction significantly boosts SID quality by guiding the model to capture implicit, high-level semantic features.

Case Studies demonstrated that UniSID-generated SIDs can successfully reconstruct not only explicit attributes but also implicit high-level semantics (e.g., "comfort," "fashion," "professional image") inferred from the summary mechanism.

5. Significance and Claims

The paper claims that UniSID represents a significant shift from the decoupled, two-stage compression paradigm to a unified end-to-end approach. By directly inducing SIDs from raw advertising data, the method effectively mitigates the information loss and error accumulation inherent in Residual Quantization.

The authors assert that their approach:

Provides a more robust and scalable solution for Generative Recommendation in complex, heterogeneous advertising environments.
Enables SIDs to capture both fine-grained details and latent high-level semantics, improving both the quality of the IDs and the downstream recommendation performance.
Demonstrates strong generalization capabilities, performing well on both large-scale industrial datasets and standard public benchmarks.

The work concludes that UniSID successfully overcomes the inherent limitations of current SID generation methods, offering a unified framework that aligns the objectives of representation learning and generative recommendation.