ZBJ Service Transaction Knowledge Graph

Project Background:

• Developed a service transaction knowledge graph for ZBJ.com, China’s leading crowdsourcing and service marketplace platform.
• Objective: Enhance search accuracy, improve user intent recognition, and optimize category classification and recommendation.

Data Collection & Processing:

• Data source: Millions of user service requests and transaction orders.
• Text preprocessing: Noise removal, tokenization with domain-specific dictionary, stopword filtering.
• Entity recognition: LSTM-CRF for extracting named entities such as service themes, industries, and geographic regions.
• Taxonomy design: Constructed hierarchical labels for Service → Industry → Region, forming the backbone ontology of the knowledge graph.

Modeling & Optimization:

• Knowledge Graph (KG) constructed using Neo4j, linking services, industries, and regions.
• Entity-relationship extraction based on co-occurrence and semantic similarity (Word2Vec + K-Means).
• Service intent classification improved by integrating KG features into Naive Bayes and CNN classifiers.
• Implemented query expansion using KG to capture long-tail user intents.

Workflow:

1. Parse user queries and demand order texts.
2. Extract entities and relations using LSTM-CRF and clustering-based methods.
3. Populate KG with service-industry-region ontology.
4. Enhance search retrieval and category classification with KG-driven features.
5. Recommend related services and providers via knowledge graph traversal.

Project Outcome:

• Increased user intent recognition accuracy by 8%.
• Improved service category prediction performance and search recall.
• Enabled intelligent service matching, increasing transaction success rate and platform revenue.