Web search involves voluminous data streams that record millions of users' interactions with the search engine. Recently latent topics in web search data have been found to be critical for a wide range of search engine applications such as search personalization and search history warehousing. However, the existing methods usually discover latent topics from web search data in an offline and retrospective fashion. Hence, they are increasingly ineffective in the face of the ever-increasing web search data that accumulate in the format of online streams. In this paper, we propose a novel probabilistic topic model, the Web Search Stream Model (WSSM), which is delicately calibrated for handling two salient features of the web search data: it is in the format of streams and in massive volume. We further propose an efficient parameter inference method, the Stream Parameter Inference (SPI) to efficiently train WSSM with massive web search streams. Based on a large-scale search engine query log, we conduct extensive experiments to verify the effectiveness and efficiency of WSSM and SPI. We observe that WSSM together with SPI discovers latent topics from web search streams faster than the state-of-the-art methods while retaining a comparable topic modeling accuracy.
展开▼
