In recent years, the losses caused by phishing scams on Ethereum have reached a level that cannot be ignored. In such a phishing detection scenario, network embedding is seen as an effective solution. In this paper, we propose an attributed ego-graph embedding framework to distinguish phishing accounts. We first obtain the account labels from an authority site and the transaction records from Ethereum on-chain blocks. Then we extract ego-graphs for each labeled account to represent it. To learn representations for ego-graphs, the graph embedding model graph2vec is applied. Finally, a classifier is adopted to predict phishing accounts. To make graph2vec more suitable for transaction networks, we add a pre-step called attribute-based relabel. Specifically, we take Ethereum transaction attributes into consideration, and propose a novel comprehensive node relabeling strategy including amount-based, number-based, and direction-based relabeling. Experimental results show that our framework achieves effective performance on class imbalanced phishing detection on Ethereum.