Meituan's Exploration and Practice in Advertising Algorithm: Information Flow Ad Estimation

This article, compiled from Meituan Tech Salon #81, introduces Meituan's information‑flow advertising business and the current state of click‑through‑rate (CTR) estimation technology, then details the company's practical implementations focusing on decision‑path modeling, ultra‑long/ultra‑wide modeling, and full‑reconstruction modeling.

1. Information‑flow ad business characteristics – strong continuity of user behavior, rich visual and textual information on the card, and diverse textual cues such as merchant name, ratings, and promotions.

1.2 Technical overview – the ad serving pipeline mirrors search advertising (recall → coarse ranking → fine ranking) but is constrained by location‑based services (LBS). Over the past six‑seven years the modeling approach evolved from tree models to DNNs (2017‑2020) and now to sparse large models with ultra‑long sequences (2021‑present).

1.3 Current estimation directions – three main axes: user‑side modeling (timeline, spatial line, combined), link‑side modeling (page and card reconstruction), and NLP/LLM side (knowledge injection). Cross‑feature modeling is de‑prioritized because its incremental benefit is limited.

2. Meituan's practice

• User modeling overview : timeline (multi‑level temporal fusion), spatial line (real and virtual location contexts), and combined behavior patterns.

• Decision‑path modeling : Path Enhance Module (PEM) extracts core paths, Path Augment Module (PAM) expands them with contrastive learning, and Path Matching Module (PMM) performs dual matching of paths and items.

• Ultra‑long/ultra‑wide modeling : sequence length up to 1000, width >10; challenges include SIM/ETA effectiveness and diminishing returns of scaling DIN scores.

• Full‑reconstruction modeling : aims to restore everything the user sees (page + card). Solutions involve side‑model predictions, high‑dimensional KV stores, and context simulation (CSC) with a Context Modeling Transformer (CMT) that encodes/decodes contextual information and applies knowledge distillation from real to simulated pages.

2.3 LLM in CTR – contrasts CTR (ID‑based, engineered networks, strong memory) with NLP (token‑based, large Transformers, reasoning). Three layers of work: knowledge injection via prompts, reasoning injection by leveraging LLM structure, and paradigm iteration using small‑scale token sequences and semantic aggregation to improve effectiveness.

3. Summary and outlook – Effective CTR estimation requires uncovering true user intent through a blend of algorithmic innovation and engineering execution. Full‑reconstruction modeling demonstrates the power of joint algorithm‑engineering effort. While large‑model integration is promising, it remains a long‑term endeavor that demands both software advances and substantial compute resources.