The treatment of complex diseases often involves the use of multiple drugs, a practice known as polypharmacy. However, this approach comes with the risk of drug-drug interactions (DDIs), which can lead to unanticipated adverse effects and even toxicity. Therefore, ensuring the safety of polypharmacy requires identifying DDIs and exploring their underlying mechanisms. Traditional wet lab methods for detecting DDIs are expensive and time-consuming, while computational methods have been developed to predict DDIs. Many of these methods have limitations, and they often struggle to predict potential DDIs between known drugs in the DDI network and drugs from outside that network without any connections to other drugs. Moreover, they may lack the capability to delve into the underlying mechanisms of DDIs and provide meaningful interpretations. In response to these challenges, we introduce a novel machine learning-based method called NGMG, which leverages a knowledge graph to identify feature representations for each drug based on its chemical and topological properties in the DDI network. Then, it combines these features for each drug pair and feeds them into a predictor to obtain a final DDI prediction score. Our experimental results demonstrate the effectiveness of NGMG in two crucial DDI prediction scenarios: identifying potential DDIs among known drugs within the DDI network and predicting interactions between drugs within the DDI network and some drugs from outside the network without any connections to other drugs.