Mining High-Value Patents Leveraging Massive Patent Data.

A patent is an inventor’s way of protecting its intellectual property. In recent years, the trend of patent filings has become more prevalent than ever as commercial competition among firms has intensified. While patents have promoted our society forward, only a few patents have made significant contributions, and they are often referred to as high-value patents. However, there is no clear definition of high-value patents, so traditional mining methods often rely on expert reviews in related fields, which is usually labor-intensive and time-consuming. Although existing literatures have resorted to human-designed statistical features to identify high-value patents, they ignore potentially valuable text and images in patents. In this work, we propose a two-phase framework to effectively extract heterogeneous features from the multi-modal text and image to mine high-value patents among the massive patents. In feature extraction phase, features are divided into three categories: statistical features, visual features, and textual features. Among them, statistical features are obtained according to a pretrained graph, textual features are extracted by a BERT-like language model, and a DenseNet-based network is used to extract visual features. In the multi-view learning phase, we use heterogeneous features to train views, then concatenate their features to the final several layers to evaluate the value of a patent. The evaluation result shows that our method outperforms the baseline methods.