Heterogeneous Graph Framework for Scalable Company Prediction

The patent introduces a graph framework as an intermediary between traditional machine learning and social network data. This framework uses meta-paths and label propagation to mediate the relationship between companies and their social connections, enabling accurate predictions without requiring extensive training data. The graph structure acts as a mediator that captures complex relationships that traditional classifiers cannot detect.

The patent transitions from traditional feature-based prediction to a graph-based dimensional space. By representing companies, employees, and relationships as nodes and edges in a heterogeneous graph, the system adds structural dimensions to the prediction problem. This dimensional change allows the model to leverage network topology and relationships rather than relying solely on tabular features.

The label propagation algorithm enables self-service prediction by allowing the graph structure to automatically propagate labels from known entities to unknown entities. Instead of requiring retraining on extensive data for each prediction task, the system uses the inherent structure of the social network to self-determine predictions, achieving scalability without proportional increases in training effort.

The graph framework serves as a scalable intermediary that decouples prediction accuracy from training data quantity. By using meta-paths to encode domain knowledge and relationships, the system achieves scalability across millions of businesses without requiring millions of labeled training examples, as the graph structure itself provides the predictive signal.

The patent segments the prediction problem into multiple entity types (companies, employees, industries) represented as separate node types in the heterogeneous graph. Each entity type can have its own features and relationships, allowing the model to capture domain-specific patterns while maintaining overall system coherence. This segmentation enables relationship-aware predictions without requiring complex feature engineering for each relationship type.

The patent creates a composite representation by combining multiple entity types and relationship types into a unified heterogeneous graph structure. Just as composite materials combine different substances to achieve superior properties, the composite graph combines company nodes, employee nodes, industry nodes, and various relationship edges to create a rich predictive model that leverages diverse information sources simultaneously.