Building Ontology in Daily-Changing Business: JD Retail's UGO Framework Revealed

The prevailing industry consensus holds that ontologies only suit stable domains like banking or manufacturing where processes and entity relationships remain unchanged for years. In fast-changing scenarios — where benefits, channels, creatives, and tactics update weekly — the assumption is that ontology diagrams become obsolete before completion.

JD Retail's User Operations team, led by Zhou Mo, challenged this view. Their user growth domain exemplifies extreme volatility: "always changing" in their own words. Yet they successfully deployed an ontology called UGO (User Group Ontology), with over 95% of growth strategies now solidified in the underlying system. This framework will be publicly presented for the first time at DACon Beijing in October.

01 Fast-Changing Domains Need Ontology Most

Two years ago, the team's problem was not ontology-related. Strategy canvases had swollen to thousands of branches per strategy; business owners couldn't untangle them. Worse, knowledge transfer failed when a colleague left — their configured strategies became untouchable.

They tried knowledge graphs and RAG (feeding Markdown docs to LLMs). Both failed because the knowledge itself expires daily: manual extraction cannot keep pace with business changes. A deeper issue emerged: without a completeness metric, there is no definition of "done." Many ontology projects die mid-way not from drawing difficulty but from not knowing when to stop.

The real blocker is not business volatility but whether knowledge can track business changes. In stable domains, knowledge lasts years; in volatile ones, it spoils in months, and manual maintenance costs skyrocket. Therefore, the faster the business changes, the more ontology shifts from nice-to-have to necessity.

02 Separate the Changing from the Invariant

The team's first correct move: cluster thousands of strategy branches at the canvas level, abstracting marketing's core elements — benefits, touchpoints, channels, audiences, creatives, timing. These became atomic attributes in the ontology.

What changes forever is the combination of these elements; what stays constant are the elements themselves. The ontology stands on this foundation.

The second, more tedious but critical step: pin down fuzzy concept boundaries. Terms like "surprise red packet" vs "mega-promo coupon" were used interchangeably; audience labels like "silver-haired," "student," "Gen Z" had no shared definitions. Every concept was redefined from scratch. This requires domain experts to manually curate each boundary, but once clarified, subsequent knowledge mining and iteration can run autonomously within the ontology.

03 Keep It Alive: Three Capabilities

Drawing the ontology is only the start; sustaining it is the hard part. The team equipped UGO with three capabilities, which form the core framework they will unveil:

Knowledge Completeness Check : Algorithmically map the full knowledge landscape of the domain and measure current coverage. This yardstick tells the team where to iterate next.

Knowledge Tracing : When a critical knowledge gap is detected, identify the true knowledge owner. In large platform teams, the nominal owner may respond weeks later. Their approach fuses multi-source evidence to locate the actual expert, and even uses that person's digital twin to extract knowledge first, followed by expert verification.

Knowledge Freshness : Borrow a "mutation" factor from search/recommendation systems, letting machines actively explore beyond existing cognitive frames and align with latest facts. A sandbox environment runs daily assertions on the most critical metric knowledge.

Only knowledge that can be verified or falsified earns entry into the ontology.

The metric pipeline goes further: the ontology does not directly output answers; it is codified into deterministic workflows and code. Zhou Mo's rationale is blunt: a single metric hallucination destroys leadership trust in the product. Code becomes the final knowledge carrier — run it a hundred times, get the same value .

04 Not a "Perfect Landing" Story

The team openly shares unsolved problems:

OneID matching accuracy improved significantly but still has user association deviations, insufficient for ultra-high-precision attribution and segmentation. Accuracy and timeliness are inherently contradictory: extreme accuracy delays data output; guaranteeing timeliness reduces precision.

On the metric side, only about half of legacy metrics are effective. The tension between standardized governance and operational flexibility remains unresolved: tight control restricts business; loose control muddies definitions.

Investment was smaller than expected: fewer than ten people dedicated to the ontology. Ontology is elite-driven — one person who understands both business and ontology design can cover a domain, unlike traditional knowledge graphs that rely on headcount-heavy entity extraction.

Hard prerequisites exist: historical knowledge standardization, organizational health, and knowledge organization maturity — all three must be present. If your data isn't standardized, build governance first, not ontology.

An interesting observation: organizational boundaries nearly equal ontology boundaries. A "Plus" business maps to a "Plus" team with own OKRs — naturally high cohesion. This emerged organically; teams with unhealthy org structures will likely fail if they force-fit this methodology.

05 UGO Goes Public in October

Timeline: concept formed in May; by 618 (mid-year shopping festival) the metric agent was built and integrated with internal platforms; July–August saw the growth ontology prototype and local strategy generation; recent weeks brought batch rollout.

At DACon Beijing, Zhou Mo will cover: how to decompose ontology for user growth scenarios; solving the three technical hard problems (completeness, tracing, freshness); how the two main tracks — generative strategy and metrics — operate; and the conditions under which not to attempt this. Algorithm implementation details are excluded pending legal review; the talk stops at the framework layer.

Who should attend:

Ontology practitioners stuck at "can't finish, can't sustain" — you'll get a completeness yardstick.

Teams in ultra-fast domains (growth, marketing, operations) who thought ontology irrelevant — this session is for you.

Semantic layer or metric platform builders plagued by inconsistent definitions and LLM hallucinations — the metric-ontology-as-code path is worth hearing.

Decision-makers evaluating "is ontology worth the investment" — Zhou Mo will tell you when not to build, which is more valuable than ten success stories.

Can you build ontology when business changes daily? In October at DACon Beijing, Zhou Mo answers live.