Abstract: Considering the problems of insufficient information mining and low prediction accuracy in single task time series, a time series prediction method based on Multi Task LS-SVM（MTLS-SVM） is proposed. Multiple time series tasks are simultaneously studied so that task can be pinned down in the training process to induce inductive bias, which improves prediction accuracy. First of all, the several learning tasks are constructed by using the close correlation between adjacent time points. Then the MTLS-SVM model is trained for prediction by the corresponding data sets of each task. This method is applied to several time series data sets. Compared with the single task LS-SVM method, the experimental results show that the proposed method has high prediction accuracy and verify the feasibility and effectiveness.

Key words: time series prediction, multi-task learning, Least Squares Support Vector Machine（LS-SVM）, relativity

摘要： 针对单任务时间序列中存在的信息挖掘不充分、预测精度低等问题，提出了一种基于多任务最小二乘支持向量机（MTLS-SVM）的时间序列预测方法。该方法将多个时间序列任务同时进行学习，使得在训练过程中任务之间能够相互牵制起到归纳偏置作用，最终有效提高模型的预测精度。首先，利用相邻时间点之间的密切相关性，构造多个相邻时间点的学习任务，然后将每个任务对应的数据集同时训练MTLS-SVM模型并将其用于预测。将该方法用于几个时间序列数据集并与单任务LS-SVM方法相比，实验结果表明该方法具有较高的预测精度，验证了方法的可行性和有效性。

关键词: 时间序列预测, 多任务学习, 最小二乘支持向量机, 相关性