结合遗忘特性的多任务多核在线学习算法

对于数据流的处理，多任务多核学习已逐渐成为在线学习算法研究的热点，它在一定程度上可提高数据流预测的准确性。多核方法尽可能使用最少的核函数得到最好的实验效果，当数据量增大、训练模型稳定时，通过阈值限定的方法对核函数进行遗忘，从而减少基本核函数的使用个数，使得计算更加简单；对于算法的优化，通过引入一个遗忘变量，从对偶的角度来进一步优化权重更新过程，这里的权重指多个任务的共有特征权重和每个任务间的特有权重，以提高算法的收敛速度。实验部分对核函数的选取进行了较为详细的分析，通过对UCI数据集和实际的机场客流量数据集进行分析，证明该本算法的合理性和高效性。

Multi-task multi-kernel online learning algorithm combined with forgetting property

For data stream processing, multi-task and multi-kernel learning has gradually become the focus of online learning algorithm. It can improve the accuracy of data flow prediction to a certain extent. The multi-kernel method uses as few kernels as possible to obtain the best experimental results: When the amount of data increases and the training model converges, the kernel functions will be forgotten by introducing a threshold; Thereby, reducing the number of kernel functions can make the calculation simpler. In addition, in order to improve the convergent speed, a forgetting variable is introduced to optimize the weight update process through solving the dual problem to further optimize, in which the weights are defined for common features of multi-tasks and special features of each task. In the experiments, the selection and analysis on kernel functions are carried out in detail. Using the UCI data and the actual airport passenger traffic data, it verifies the best performance and reasonability of the algorithm put forward in this paper.