多示例嵌入学习的实例关联性挖掘与强化

Multi-instance embedding learning with instance affinity mining and reinforcement

提出了多示例嵌入学习(multi-instance learning,MIL)的实例关联性挖掘与强化算法(multi-instance embedding learning with instance affinity mining and reinforcement,MEMR),包括3个技术。关联性挖掘技术基于自定义的关联性指标,首先在负实例空间中选择初始负代表实例集,然后根据正、负实例间的差异性,选择初始正代表实例集。关联性强化技术分别评估初始正、负代表实例集与整个实例空间的正负关联性,获得整体关联性更强的代表实例集。包嵌入技术通过嵌入函数将包转换为单向量进行学习。实验在4类应用领域和7种对比算法上进行。结果表明,MEMR的准确性总体优于其他对比算法,特别是在图像检索和网页推荐数据集上具有显著优势。

We propose the multi-instance embedding learning with instance affinity mining and reinforcement(MEMR)algorithm,including three techniques.The affinity mining technique is based on a custom affinity metric.First,the initial negative representa-tive instance set(INRI)is selected in the negative instance space.Then,the initial positive representative instance set(IPRI)is chosen according to the difference between positive and negative instances.The affinity reinforcement technique evaluates the posi-tive(negative)affinity between IPRI(INRI)and the entire instance space to obtain a representative instance set with stronger over-all affinity.The bag embedding technique converts bags into single vectors for learning through the designed embedding function.Experiments are carried out across four application domains and seven comparison algorithms.The results show that MEMR generally outperforms other comparison algorithms in accuracy,especially in image retrieval and web recommendation datasets.