API World + AI DevWorld

Recent years, inspired by the huge success of convolutional networks in Computer Vision, Graph Convolutional Networks has emerged and attracted great attention for its potentials in graph-based data world, such as social networks, recommendation, biological graphs and traffic prediction. In this talk, we present our innovation on fraud prediction of using Graph Convolutional Networks (GCNs) in Walmart eCommerce system where it predicts risk scores for entities (e.g. customer, device) in real time on large-scale transaction-based graphs. We demystify how it smartly leverages graph topology via adjacency matrix and learn the weight matrices to nonlinearly propagate information across multi-hop graphs, especially on graphs condensed of fraudulent transactions. We also describe how the semi-supervised classification GCNs is trained with a weighted softmax loss function on partially labeled graph snapshots. By comparing with models that do not rely on graph structures, we show the noticeable performance lift of GCNs in various aspects. To demonstrate how well it works, we evaluate the prediction results through real-world graph examples at WalmartLabs. At the end, we show the significant business/engineering value our entity scores can bring to the company and point out multiple directions in which the methodologies can be generalized and further improved.