Runtime Quality Prediction for Web Services via Multivariate Long Short-Term Memory

Algorithm 2

Online service quality prediction.

	Input: a group of long-term quality time series data and the number of time points for forecasting in the near future.
	Output: a group of quality prediction results .
(1)	Data selection: employ Algorithm 1 to select a subset of time series from , in which any time series is highly similar to each other. In the selection process, align the selected time series according to the optimal time delay.
(2)	Data preparation: divide each long-term time series into multiple time series , each measured during with time points. The moving time window for the division is a single time point. Each time series is related with an observed time series data for training. The data in are measured during , with time points. Thus, the whole dataset is divided into multiple multivariate time series , and each multivariate time series is related with an observed time series for training.
(3)	Model training and evaluation: train the multivariate LSTM network and compute the weight matrices the bias vectors using the input training data set. Evaluate the loss on the evaluation data set to adjust the model parameters.
(4)	Model prediction: during the time slices, take the historical quality time series to predict the value in the near feature. Update the model with the new input prediction results.
(5)	Model update: after time slices elapses, update the historical training set with the new observed result and retrain the prediction model.