Query-Based Recommendation by HIN Embedding with PRE-LSTM
Published in Yang, X., Wang, C.-D., Islam, Md. S., & Zhang, Z. (Eds.). (2020). Advanced Data Mining and Applications. Lecture Notes in Computer Science. https://doi.org/10.1007/978-3-030-65390-3, 2021
Recommended citation: Liu, ZM., Hui, YM., Huang, L. (2020). Query-Based Recommendation by HIN Embedding with PRE-LSTM. In: Yang, X., Wang, CD., Islam, M.S., Zhang, Z. (eds) Advanced Data Mining and Applications. ADMA 2020. Lecture Notes in Computer Science(), vol 12447. Springer, Cham. https://doi.org/10.1007/978-3-030-65390-3_39
In recent years, as more and more big data platforms have been applied to the government network systems, it’s essential to adopt an effective query-based recommendation algorithm to help officers find out the needed table with a keyword. The key challenge lies in the complex relationship among data from different departments, which cannot be easily solved by the existing database query methods. The Heterogeneous Information Network (HIN) is a specific type of networks developed for modeling complex data relations. However, these existing query-based recommendation algorithms could not make use of HIN. Besides, many query-based recommendation algorithms could not make recommendations with the keyword that is not in the query records. In addition, most of the existing recommendation algorithms do not make full use of the semantic meanings. Therefore, in this paper, by making use of the real dataset provided by the local government, the proposed method is the first to use the pretrained word embeddings and LSTM (PRE-LSTM) to train the network embeddings and to learn the relationship among the tables, departments and the keywords, so that it can make use of the data in HIN and enable the network embeddings to obtain the most precise semantic meaning. Additionally, our algorithm uses word embeddings to represent query keywords so as to let our algorithm make the query-based recommendation for nearly any query. Using the trained embeddings and the PRE-LSTM model, the proposed algorithm is able to show user-specific recommendation results sorted in a reasonable order. Experimental results on the real data application tasks confirm the effectiveness of the proposed method.