Algorithmic Optimization for Information Flow Advertising at Hello Travel

This presentation explains how Hello Travel (哈啰出行) tackles the challenges of information‑flow advertising from the advertiser’s perspective, where limited user‑level exposure data and lack of external new‑user features make modeling difficult.

The talk is organized into four parts:

Current status of information‑flow ad delivery

Hello Travel business background

Advertiser‑side algorithm optimization solutions

Future directions

1. Information‑flow ad delivery status

Information‑flow ads are native, content‑rich ads that appear in social feeds (WeChat Moments, Douyin, Toutiao, etc.). Their market share has grown from a few percent in 2015 to an estimated 40.8% in 2022, making them a primary channel for user acquisition at Hello Travel.

The delivery process is shown from both the platform side (user browsing → ADX request → DSP bidding → ranking & pricing) and the advertiser side (bid decision based on conversion, user value, exposure, and other strategies; followed by account setup, anomaly monitoring, data tracking, and automated delivery).

2. Hello Travel business background

Hello Travel’s external advertising has evolved through four stages:

Exploration stage : basic account opening and market testing.

Initial scaling stage : large‑scale ad spend to capture market share, requiring conversion attribution and monitoring.

Cost‑reduction stage : refined traffic operation using DMP, audience‑management APIs, and marketing APIs.

Mature stage : fully automated, end‑to‑end algorithmic optimization with minimal manual intervention.

The current system architecture integrates three major channels (Juliang Engine, Guangdian Tong, Kuaishou) via a unified API gateway, storage layer (Redis, MySQL, HBase, Elasticsearch), and an application layer that focuses on decision‑making, automation, and attribution.

3. Advertiser‑side algorithm optimization

The optimization is discussed from three dimensions:

Ad‑plan dimension : Modeling the probability that a user who has submitted certification will complete an order within seven days. Challenges include small data volume for new plans and lack of detailed exposure/competition data. The solution reframes the problem as “low‑quality traffic identification” and builds a LightGBM model using user, environment, ad, and time features. Data augmentation converts user‑level samples into order‑level samples to increase training data.

Creative dimension : Predicting whether a new creative will achieve volume. Feature engineering uses ID embeddings (word2vec on creative, plan, and account IDs), cross‑features between numeric and configuration parameters, and interaction of targeting settings. A multi‑class LightGBM model replaces a high‑error regression approach.

Pre‑bid prediction mechanism : A pre‑exposure screening model (uplift model) that assigns a quality score to incoming users, deciding whether to expose them. The model distinguishes natural converters from those needing ad incentives and stratifies users into five buckets for tiered bidding, achieving 10‑20% cost reduction compared with manual bidding.

4. Future directions

Two main goals are outlined:

Enhanced pre‑screening models that incorporate exposure frequency to more precisely block low‑value users.

Fully automated end‑to‑end online delivery, including automatic creative generation, budget allocation across plans, and ROI‑maximizing user targeting.

Additional roadmap items include cross‑channel bidding management, RTB integration, uplift and reinforcement‑learning techniques, and computer‑vision models for creative generation.

Q&A highlights

• Different business lines target distinct user groups, reducing competition for bids. • Similar creatives perform differently across accounts due to varying historical performance and platform weighting (eCPM, CTR, etc.). • The planned “RTA‑based RTB” integrates real‑time bidding functionality into the RTA interface.