Harbor's Passive Growth Algorithms and Growth Engine: Practices and Insights

This article introduces the passive growth algorithms and the growth engine practices used by Haro (哈啰) in its C‑end (consumer) scenarios.

Passive Growth – In‑App Traffic Distribution

Harbor’s passive growth is essentially a traffic‑allocation algorithm. Although it does not involve real‑money bidding like an advertising platform, a bidding mechanism exists between business lines. An attribution tool determines the price of each allocation, and the challenge lies in defining a unified attribution metric that all business lines accept, which requires top‑down coordination.

The system balances short‑term traffic efficiency with long‑term cross‑traffic (LTV) goals. By quantifying long‑term cross‑traffic as an LTV value, it can be compared with short‑term GMV targets in a unified bidding formula. Classic models are used, with a focus on richer feature engineering; Autolnt is employed for feature interaction, and DIN‑style sequence embeddings are applied in certain popup scenarios.

Virtual Mall Ranking

Virtual items differ from physical goods: the item pool is small (tens of items) and user choices are driven by pure price comparison rather than interest. Because items are mutually exclusive and feature dimensions are limited, PointWise ranking is unsuitable. PairWise showed significant improvement, leading to a ListWise framework with a Softmax cross‑entropy loss. A Generator module builds the candidate list, and an Evaluator scores it using a second‑version DCN and a Bi‑LSTM for sequence feature extraction.

Search & Recommendation

In local‑life services (e.g., tours, tickets, rentals, hotels), Haro applies traditional recall embeddings and CTR prediction. A two‑dimensional multi‑task learning approach is used to share a model across homepage and detail‑page recommendation slots, predicting both CTR and CVR. Multi‑domain models with shared and domain‑specific sub‑networks improve CTR across different business lines.

Algorithm Componentization

To accelerate development, each domain’s algorithms are packaged as reusable components consisting of an offline part (standardized tables and task formats) and an online part (standard inference code). This reduces the onboarding time for new scenarios from about a month to one‑to‑two weeks.

Integrated Search Engine

Harbor consolidates search and recommendation into a single engine, with separate intent‑recognition components for query parsing. Service orchestration and configurable algorithm components enable rapid integration of new business scenarios with minimal API changes.

Marketing Engine

The marketing engine connects various promotion tools through a data‑and‑goal bus, preventing excessive pop‑ups and aligning campaigns with global metrics.

L3 Automated User Operation Strategy System

Standardized, abstracted strategies are automated based on triggers (e.g., seasonal promotions during rainy seasons). Many strategies now operate fully automatically.

Rethinking C‑End Algorithms

After a year and a half of practice, the team shifted from point‑optimizations in ranking funnels to a business‑oriented view: algorithms serve the broader goal of improving digital operational efficiency. Passive and active growth both aim at traffic, but from different perspectives. Structural opportunities are sought across the entire business chain rather than isolated tweaks.

Future directions include expanding causal inference for bias reduction, building a causal inference platform, moving toward general models and automated training, and encouraging algorithm engineers to become domain experts.