A heterogeneous multi-relations-based method for knowledge-intensive service recommendation

Abstract

Recommendation systems (RSs) have been widely used in consumer product trading platforms due to their excellent capability to assist users in finding relevant items, which have also attracted tremendous attention in knowledge-intensive service (KIS) transaction platforms. While numerous RSs have been applied in KIS platforms for task recommendation and worker recommendation, an RS specialized for KIS recommendation to meet employers’ potential needs is still missing. To this end, the heterogeneous multi-relations, including knowledge-service relations, activity-service relations, and worker-service relations, are investigated firstly to bridge the knowledge gap, and a novel methodology for KIS recommendation is proposed based on the heterogeneous multi-relations construction and fusion. This method develops a heterogeneous information network (HIN) to structure high-level interactions among employers, services, activities, and workers, and designs KIS-oriented meta-paths to obtain valuable heterogeneous multi-relations. Furthermore, these valuable relations are aggregated and fused hierarchically through metapath2vec and attention mechanism to learn representations of employers and services and further to predict the targeted employer's potential KIS needs. Finally, the transaction data from ZBJ.com - the largest KIS trading platform in China, is used to validate the effectiveness and high accuracy of the proposed method in KIS recommendation. It is also found that as the service-service knowledge similarity increase, the proposed model has an inverted U-shape performance, in which the composite relation of activity-service and service-service knowledge similarity is the main factor for KIS node representation learning.