<![CDATA[Proactive autoscaling in cloud-native microservices requires anticipatory decisions because reactive controllers often lag under abrupt workload shifts. This study aims to improve autoscaling decision quality through a two-stage machine learning pipeline. The research adopts an experimental design using production-grade microservice traces, with strict time-respecting train-validation-test splits and training-only fitting for preprocessing and oracle-threshold estimation to prevent leakage. In the first stage, multivariate forecasting models predict future CPU and memory utilization from engineered temporal features. In the second stage, the predicted signals are combined with observed features to classify three autoscaling actions: scale down, hold, and scale up. Benchmarking shows recurrent neural models are strong baselines, while an attention-enhanced encoder-decoder performs best. The best Bahdanau-attention model with residual connection reduces test CPU RMSE from 0.030977 to 0.028924 and memory RMSE from 0.010322 to 0.005452 relative to the strongest BiLSTM baseline. For decision learning, the optimized Extreme Gradient Boosting model using prediction-augmented features achieves an accuracy of 0.950602 and an F1 score of 0.951026. Supporting downstream validation also yields lower SLO violation rates than horizontal and vertical baselines while maintaining zero downtime in the evaluated scenarios. These findings indicate that improving forecasting quality and explicitly transferring predictive signals to the decision stage strengthens proactive autoscaling performance.]]>
