Quality prediction for services is a hot research topic for service recommendation and composition. It's a challenge to design an accurate quality prediction approach to meet the user's personalized needs due to the sparsity of QoS historical data. In order to solve the challenging problem, this paper proposes a SOM neural network based service quality prediction approach (SOMQP). First, based on historical QoS data, the new approach clusters on users and services respectively by applying SOM neural network algorithm, and then a new top-k selection mechanism is adopted to obtain similar users and similar services based on the comprehensive consideration of user reputation and service relevance. Finally, a hybrid user-based and item-based strategy is used to predict the missing QoS value. A set of comprehensive experiments are conducted on the real Web service dataset WS-Dream, the results indicate that compared with the classical CF and K-means methods, the presented approach achieves higher QoS prediction accuracy.