Design and Implementation of a Recommendation Algorithm PaaS for Scalable Business Scenarios

Currently the recommendation algorithm team supports more than 20 business lines and over 900 recommendation scenarios. By analyzing common requirements across large‑scale promotions and vertical business lines, the team has been consolidating algorithm capabilities and building a PaaS platform to improve efficiency and reuse.

Why PaaS? PaaS provides a flexible, extensible, and reusable framework that reduces repetitive labor, lowers development cost, and enables rapid innovation, similar to other industry players that have adopted PaaS on their business‑center platforms.

How it helps recommendation capabilities The team categorizes business demands into common and custom needs, offering tiered strategies: one‑stop personalized recommendation for generic needs and efficient PaaS tools for custom requirements, thereby shortening delivery cycles.

Solution Design

Business demands are grouped into two major categories: new recommendation‑slot requests and existing recommendation‑strategy optimization requests. For each, the PaaS builds six primary capability dimensions—data, algorithm components, data analysis, operators, scenario templates, and services—comprising 20 secondary capabilities.

Primary Capability

Definition

Secondary Capability

Definition

Data Reuse

Reuse of data across recommendation stages

Direct reuse of recall data

Simple processing reuse of recall data

Reuse of ranking model files

Code (Algorithm Component) Reuse

Non‑model recall (cold‑start, profile, similarity, etc.)

Different recall sources invoke different scripts

KNN recall

Precision ranking model

Data Analysis Reuse

Basic intermediate tables

Project‑level intermediate tables

Operator Reuse

Reuse of operator functions and related data

Filter operator

Weight‑adjustment operator

Top‑placement operator

Deduplication operator

Rendering operator

Diversity operator

Scenario Template Reuse

Full‑site product comprehensive recommendation, detail page, cart, live‑stream, short video, etc.

Feed‑style homepage recommendation based on user behavior and product similarity

Supports multiple material types, KNN recall, configurable model goals, custom features

Main product recommendation for detail pages, cart, and item‑based scenarios

Supports similar‑product recall, free‑shipping bundling, same‑shop recommendation, self‑operated filter, category‑specific add‑to‑cart pop‑ups, LBS recommendation, etc.

Marketing‑driven pool recommendation (e.g., holiday events), tab sorting, store‑SKU recommendation, O2O recommendation

Supports tab sorting, brand recommendation, time‑based recall, filter, special pool creation

Service Reuse

Single‑material service (single product, single shop, short video, etc.)

Filter service

Existing

Precision ranking inference service

Existing

The classification will evolve as the PaaS matures.

Recommendation Algorithm Componentization is the prerequisite for platformization. By abstracting and packaging algorithm capabilities into runnable code bundles, users can plug‑in components without deep code knowledge.

Platformization aims to simplify component usage, offering a visual, configurable, and editable toolset. Two major tracks are addressed: full‑link recommendation capability platformization for new slots, and module‑level platformization for existing slot optimization.

General algorithm strategy configuration builds a shared operator library (recall, ranking, filtering, diversity, etc.) to reduce configuration cost and improve code reuse.

Low‑code development for custom strategies further reduces algorithmic effort, enabling rapid development and deployment of custom operators.

Case 1 – Scenario Template Development describes how a template for product‑aggregation tab recommendation is built, evaluated, and handed over to business users for self‑service.

Case 2 – Automatic Recall Dictionary/Index Creation introduces a one‑click tool that packages recall dictionary/index generation scripts, integrates with DUCC scheduling, and automates updates.

Case 3 – Multi‑Business Ranking Model Support integrates three ranking models (main‑site, discount‑focused, B2B) into templates, allowing business users to select the appropriate model.

Case 4 – "Zhuomu Niao" Low‑Code Filtering Tool automates offline filter/release of products, categories, and sensitive words, replacing manual HDFS text edits with a web‑based configuration interface.

Practical Experience Summary reflects both provider and user perspectives: providers consolidate reusable PaaS tools, while users experience efficiency gains and self‑service capability.

Future Work includes deeper scenario‑template layering, single‑material service expansion, and further algorithm component platform upgrades to achieve fully point‑and‑click operations.