Internet of Things (IoT) systems are driven by numerous Wireless Sensor Networks (WSNs) at the sensing layer supporting various applications. Due to the high volume, complexity, and velocity of IoT data and the limited resource of IoT devices, it is crucial to develop efficient data management to fulfill the required Quality-of-Service (QoS). However, managing IoT traffic based on QoS metrics is insufficient, especially when the required QoS is not attainable even after optimally utilizing all available resources. In this thesis, data analytics is used to develop metrics and methodologies for identifying the quality of IoT data in terms of predictability and contained information. The Quality-of-Information (QoI) of the collected data must be optimized while being subjected to the limited resources to ensure that IoT applications run successfully. Therefore, the QoI is exploited in data management, reduction, and forwarding. Firstly, advanced algorithms are developed for performing Dual-Prediction (DP) to reduce data based on its predictability. The proposed algorithms are based on Deep Neural Networks, which are able to outperform existing techniques in terms of reduction ratio while maintaining the same error levels in recovery. Secondly, information-oriented data reduction and forwarding are developed to maximize the Information-Content (IC) of collected data. The proposed IC metric measures data redundancy and recoverability. The developed schemes aim to improve data delivery by ensuring that all delivered data is important for running applications successfully. Thirdly, the challenges facing the practical implementation of the proposed DP and IC-based data reduction schemes are highlighted. Also, the proposed solutions addressing them are validated through a developed IoT testbed running realistic IoT applications. Finally, a Deep Reinforcement Learning (DRL) approach is proposed for solving problems where unknown-state components in IoT systems need to be managed. The DRL approach has been applied for Edge backhaul selection to fulfill the QoS of data delivery through the backhaul. To improve the convergence time of DRL policies training, Federated Learning (FL) is applied. FL enables collaborative policy training between multiple Edge devices without sharing private training data.