Scalable Multi-Class Recognition: Optimizing Deep Convolutional Architecture (VGG-16) for High Dimensional Classification

Waste management requires the effective prediction and classification of waste in order to aid in effective planning and sustainable decision making. As the waste streams can be heterogeneous in nature and can exhibit different kinds of data patterns, there cannot always be an optimal model to predict or classify every kind of waste stream. In this study, we have performed a comparative analysis between various statistical and machine learning models used in waste management. Traditional time-series models like Error-Trend-Seasonality (ETS) and Auto Regressive Integrated Moving Average (ARIMA) have been compared with modern machine learning techniques like Random Forest (RF), XGBoost, LSTM networks, and a CNN VGG16 transfer learning technique used for image classification of wastes. The effectiveness of the different models has been measured using common error measurement metrics like Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Area Under the Curve (AUC). Although in cases where temporal patterns are well defined, ARIMA outperforms ETS, it is observed that machine learning models perform much better due to their ability to capture non-linear relationships in data. Hence, it is clear that the choice of models needs to be made depending on the nature of data itself.