Item recommendation for users is a salient feature of many transaction-based systems. Finding the most appropriate items is crucial for both marketing and analytical perspectives. The latest works focus on ranking-based personalized recommenders. However, they recommend the same number of items for everyone and still suffer from the interaction sparsity issue. We propose a complex-graph-oriented supervised learning-based link prediction with a realistic negative sampling application for overcoming these problems. We employ the power-law degree distribution property of the complex graphs to sample the negative instances. The experiments show that our method outperforms ranking-based personalized recommenders with a 20% increase in recommendation success in multiple evaluation metrics. |
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.