We study group correlation structures of financial markets in Japan and U.S. from a network-theoretic point of view. The correlationmatrix of stock price changes, purified by the random matrix theory, is regarded as an adjacency matrix for a network. The weighted links in the networks thus constructed can have negative sign corresponding to anticorrelation between stocks. To identify groups in such a network, we search for the optimum decomposition of nodes which maximizes the total sum of weights of links inside groups. We then find that the network of Tokyo Stock Exchange is decomposed into four groups. The stock prices comove almost perfectly inside the groups and move oppositely across the groups. Also we apply the same analysis to the S&P 500 stocks. The U.S. stock market shows frustrated behavior similar to that embedded in the Japanese market.