Concerning the problem that heterogeneous network representation learning only considers social relations in structure and ignores semantics， combining the social relationship between users and the preference of users for topics， a representation learning algorithm based on topic-attention network was proposed. Firstly， according to the characteristics of the topic-attention network and combining with the idea of the identical-discrepancy-contrary （determination and uncertainty） of set pair analysis theory， the transition probability model was given. Then， a random walk algorithm based on two types of nodes was proposed by using the transition probability model， so as to obtain the relatively high-quality random walk sequence. Finally， the embedding vector space representation of the topic-attention network was obtained by modeling based on two types of nodes in the sequences. Theoretical analysis and experimental results on the Douban dataset show that the random walk algorithm combined with the transition probability model is more comprehensive in analyzing the connection relationship between nodes in the network. The modularity of the proposed algorithm is 0.699 8 when the number of the communities is 13， which is nearly 5% higher than that of metapath2vec algorithm， and can capture more detailed information in the network.