Hyperparameter search space table.

The best set of parameters is {number of conv filters, Nc: 256, number of units in LSTM Nh: 64, dropout probability for Dropout Layer: 0, learning rate: 0.001} for read-level prediction on 5-fold cross validation of training data. The window size of convolutional layers, W, is set to 9 and the number of hidden nodes in attention layer, Na, is set to 16 in the default setting. Cole et al. found that the Naïve Bayes classifier performs similarly when using 8-mer and 9-mer as features but slightly worse using smaller k-mers [84] for the taxonomic classification of 16S rRNA reads. Therefore, we similarly use a window size of 9 for convolutional filters. In S11 Appendix, we further show that sample-level classification performance of the model is generally insensitive to the window size. In future work, we will consider other consequences of altering window size, such as computational performance and interpretability. (XLSX)

在训练数据的5折交叉验证中，用于读段级预测的最优参数集为：{卷积滤波器数量（Nc）：256，长短期记忆网络（LSTM）单元数量（Nh）：64，Dropout层的丢弃概率：0，学习率：0.001}。默认设置下，卷积层的窗口大小W设为9，注意力层的隐藏节点数（Na）设为16。在针对16S核糖体RNA（rRNA）读段的分类学分类任务中，Cole等人[84]发现，当以8聚体（8-mer）与9聚体（9-mer）作为特征时，朴素贝叶斯分类器的表现相近，但使用更小的k聚体（k-mer）时表现略逊一筹。因此我们同样将卷积滤波器的窗口大小设为9。在附录S11中，我们进一步证明了该模型的样本级分类性能通常对窗口大小不敏感。在未来的工作中，我们将考虑调整窗口大小带来的其他影响，例如计算性能与可解释性。（XLSX）