AI‑Powered Automated Advertising Platform: 360 Easy Placement Overview

1. Background and Problem

When advertisers discuss ad placement, they are actually dealing with many variables: budget, creative design, copywriting, bidding price, target audience, media selection, and time slots. For large advertisers these choices are manageable, but small and medium advertisers often encounter two major issues: lack of dedicated creative designers and limited budgets that make manual campaign building and optimization difficult.

Typical manual solutions—outsourcing creative production or hand‑crafting campaign structures—suffer from extra workload, low approval rates, poor material quality, and inefficient optimization cycles.

2. System Architecture

To address these problems, 360 launched an integrated one‑stop platform called 360 Easy Placement , which consists of three core modules:

Easy Creative

Easy Review

Easy Optimization

2.1 Easy Creative – One‑Click Creative Generation

Users only need to select business keywords; the system automatically retrieves and rewrites copy, matches relevant images, and returns a complete set of creative assets. The platform currently holds over 50 million copyrighted images and more than 1 billion generated copy snippets, all freely available to customers.

2.2 Easy Review – Fast Creative Approval

Because creative assets and landing pages are pre‑checked, the review time of Easy Placement is roughly one‑third of the standard review process.

2.3 Easy Optimization – One‑Step Campaign Building

Traditional campaign setup involves up to 15 configuration items (budget, schedule, targeting, bidding, etc.). Easy Placement reduces this to five key items: selected creative, campaign name, budget, schedule, and goal.

The system automatically sets up promotion plans, groups, audience, media, and bidding, leveraging OCPC (Optimized Cost‑Per‑Click) and intelligent scene matching to achieve the desired conversion goals.

3.1 Easy Creative – Creative Generation and Recommendation

The workflow starts with business keywords to retrieve copy templates, perform rewrite and pre‑filtering, and similarly retrieve images, score text‑image relevance, and crop images to produce final creatives.

The copy library is built by collecting raw copy from media, advertiser materials, and manual writing, followed by extensive annotation (business terms, brand tags, pain points, patterns). Automated annotation uses dictionary lookup, NER, FastText for brand detection, and DSSM models for pattern recognition.

To expand the copy pool, the system applies synonym expansion (Word2Vec), near‑sentence generation (DSSM), and iterative human‑in‑the‑loop verification, achieving a ten‑fold increase in annotation speed.

3.1.2 Image Automatic Matching

Image sources include advertiser materials, designer‑created assets, and a 50 million‑image licensed library (with noisy tags). The pipeline retrieves images using keyword search, scores relevance with a dual‑tower model (text and image encoders), and applies a second‑stage recall when needed.

Training data comes from 360 image search logs, limited labeled feedback, and randomly sampled negatives. The recall model uses embedding (triplet) loss, while the ranking model uses a logistic loss with multi‑dimensional supervision.

After retrieval, images pass through a series of binary classifiers (face detection, object detection, OCR, background extension) built on pretrained CNNs to ensure compliance with advertising regulations.

3.2 Easy Optimization – Automatic Effect Optimization

The module consists of three sub‑components: OCPC automatic bidding, intelligent scene recommendation, and personalized display.

3.2.1 OCPC Automatic Bidding

Unlike traditional OCPC, this system does not require a large data warm‑up; it uses a reinforcement‑learning module for exploration, industry‑specific models, real‑time feedback, full‑day conversion correction, and differentiated strategies based on advertiser CPA tolerance.

When sufficient data is collected, CVR is estimated using a multi‑task GBDT model with isotonic regression; CPC bidding equals estimated CVR multiplied by target CPA.

3.2.2 Intelligent Scene Recommendation

Real‑time, asynchronous, and offline modules process ad material, extract features, and match them to user context. Multi‑task learning treats conversion rates of different scenes as separate tasks, enabling balanced bidding across heterogeneous scenarios.

3.2.3 Personalized Display

Based on user demographics and context, the system selects the most suitable creative and presentation style. A lightweight dual‑tower model combines material features with user/media/context features, and a PCTR‑based “creative selection” filter removes bias unrelated to the material itself.

4. Case Study and Conclusions

A case study of a home‑renovation company shows that after adopting Easy Creative and Easy Optimization, daily conversions increased while CPA continuously decreased, demonstrating the platform’s value for SMB advertisers.

Key takeaways:

AI can permeate every stage of advertising, not just ranking or retrieval.

Data quality and utilization are critical.

Clarify problem focus before model selection (e.g., copy generation vs. image‑text relevance).

Decompose complex problems into simpler sub‑tasks.

Practicality of models outweighs sheer complexity.

Author: Zheng Xiaodong, 360 Technical Manager, with six years of experience in RTB advertising algorithms and a background in research (publications at KDD, IJCAI).