Abstract: To address the challenge that traditional subjective scoring methods struggle to accurately identify anomalies and management risks in high-dimensional, unlabeled multi-source data, this paper proposes an unsupervised machine learning-based approach for anomaly detection and look-through attribution of power grid investment efficiency. Firstly, a multi-disciplinary investment efficiency evaluation index system covering pre-investment decision-making, in-process control, technical benefits, economic benefits and social benefits, is established to achieve quantitative characterization of the entire business process including pre-investment, mid-investment, and post-investment stages. Secondly, an investment efficiency anomaly detection model is constructed based on the isolation forest (iForest) algorithm. The model normalizes and calculates anomaly scores via the expected length of random splitting paths, and combines the algorithm's intrinsic feature weights with the robust deviation of median absolute deviation (MAD) to establish a quantitative derivation formula for the contribution degree of key anomalous indicators. Finally, empirical research and drill-down verification are conducted using real investment evaluation datasets. The results demonstrate that the proposed model can accurately pinpoint anomalous management units and trace underlying business risks without preset thresholds, effectively remedying the defects of traditional evaluation systems and providing solid support for closed-loop life-cycle management of power grid investment.