Local Life Comprehensive Demand Knowledge Graph: Design, Algorithms, and Applications

Background

The Local Life Comprehensive Demand Graph (GENE: lifestyle General NEeds net) is a knowledge‑graph‑based approach that starts from user demand to uncover diverse local‑life needs and link them with multi‑industry supplies, aiming to improve platform supply‑demand matching efficiency.

1.1 Business Status

Meituan covers hundreds of industries (food delivery, hotels, travel, medical, education, etc.) serving billions of users. Beyond increasing user and merchant quantity/quality, efficiently matching user demand with merchant supply is a key growth lever.

1.2 Problem Analysis

Current supply organization follows an industry‑category‑merchant‑product hierarchy, which creates several pain points: ambiguous user intents, cross‑category demands, and lack of detailed knowledge in specific categories (e.g., medical‑beauty). These issues stem from a supply‑centric view that neglects the user perspective.

Solution

2.1 Idea

Introduce a user‑centric knowledge graph that captures the first two stages of the decision process—"intent generation" and "consideration"—and connects them to supply via search/recommendation capabilities.

2.2 Concrete Scheme

Adopt a multi‑layer graph structure (industry system, demand object, concrete demand, scene element, scene demand) and model it after Alibaba’s AliCoCo cognitive concept net, extending it to cover the "intent" stage.

Implementation Methods

3.1 Industry System Layer

Construct a pruned and split category tree for the entertainment industry, then associate merchants and products (both real and virtual) with these categories using a multi‑source fusion model (BERT for text, Doc2Vec + Self‑Attention for UGC, One‑Hot for merchant profiles).

3.2 Demand Object Layer

Extract play‑object words via unsupervised Word2Vec expansion and supervised BERT‑CRF sequence labeling, then link objects to categories through frequency‑based matching and hierarchical relation modeling.

3.3 Concrete Demand Layer

Generate candidate phrases around demand objects (using syntax‑aware extraction with ELECTRA + BiAffine) and filter them with a Wide&Deep model that combines statistical and BERT‑derived semantic features, achieving ~92% recall and 85% precision.

3.4 Scene Element Layer

Decompose scene demands into elements (person, time, place, purpose) and mine each element via seed expansion and sequence labeling; then model element‑demand relations as aspect‑based classification using BERT sentence‑pair classification.

3.5 Scene Demand Layer

Assemble scene demands by combining scene elements and concrete demands, then evaluate their plausibility using a Markov Decision Process with reinforcement‑learning‑based reward shaping.

Application Practice

4.1 Parent‑Child

Redesign the parent‑child channel page using graph‑derived icons (e.g., "play with animals", "family spa") and recommendation reasons that surface concrete demands, dramatically improving user decision efficiency.

4.2 Leisure & Entertainment

Introduce scene‑based icons ("spring outing", "nightlife", "team‑building") and fast‑filter lists derived from refined categories, enabling users to quickly find suitable merchants.

Conclusion & Outlook

The graph now contains hundreds of thousands of concrete and scene demand nodes and tens of millions of edges, already deployed in Meituan’s parent‑child, leisure, medical‑beauty, and education services. Future work includes broader industry coverage, multimodal data integration, and deeper graph‑driven recommendation and intent‑recognition.