As knowledge recombination increasingly shapes technological advancement, scholarly efforts have focused on how integrating knowledge elements can enhance technological innovation performance. Our study contributes to this literature by adopting a network perspective and assessing the predictive value of various knowledge linkage features. Using granted patents in the pharmaceutical field, we constructed the annual International Patent Classification (IPC) co-occurrence network and extracted the real-time linkage features of pairwise co-occurring IPCs, including tie strength, knowledge distance, node assortativity, and betweenness centrality. Based on explainable machine learning, we found that XGBoost outperformed in predicting both patent impact and patent disruptiveness. More importantly, feature interpretation based on the Shapley value illustrates that patent impact and patent disruptiveness have different determinants, implicating different knowledge combinative strategies. Specifically, we found that betweenness centrality and node assortativity contribute significantly in predicting patent impact, suggesting that localised search combining hotspot and marginal knowledge components are more likely to produce impactful inventions. Conversely, tie strength between combined knowledge components is the most important predictor for patent disruptiveness, indicating that deeper exploitation along existing technology portfolios can effectively enhance patent disruptiveness. These results provide new insights into how knowledge recombination fuels technological innovation and offer practitioners valuable strategies for developing target technologies.