Mining User’s Location Intention From Mobile Search Log

With the rapid development of mobile internet, the mobile search field has become a hot research field. With the prevalence of smart phones and intelligent personal assistant such as Siri, the market share of mobile search has occupied half of the whole search market. More and more apps try to gain the location information of the user. But it is not enough to have the location information only, we still have to judge whether the location information is helpful for some searches. Thus, how to identify the location intention behind the search becomes the key point.Under this background, we propose a method to classify queries according to its location intention. Then, we can take different actions to different classes of queries. In this paper, we classify queries into three categories according to its location intention, which are "fixed collocation query", "location sensitive query" and "ordinary query"Firstly, fixed collocation query’s characteristic is that each query has a corresponding place name (location name). The corresponding location intention is fixed and won’t change with the user’s location. For example, "Tian’anmen","the Statue of Liberty" and "the Yellow Mountain", their location intentions are "Beijing", "New York" and "Anhui province". These intentions won’t change with the user’s environment. It usually happens when a user searches for famous places of interest, well-known buildings, government office and so on. Each of this kind of queries needs a corresponding place name.Secondly, location sensitive query’s character is that the user needs the information in their location city or nearby. For example, "hotels nearby", "train station", "house price" and "weather forecast". The useful information is limited in the user’s location city.Thirdly, ordinary query is the kind of query which has no location intention. They don’t need location information to improve search results. If we add an arbitrary place name to the query, it may do harm to the search results. For example, "Zhu geliang", "The story of stones", "jokes" and "funny videos"To handle the first class, we devise a select function which contains two factors which are term count and co-occurrence frequency. We choose an appropriate threshold of co-occurrence frequency and vary the term count to get different results of experiments. We get a list of couples of keywords and corresponding place names and finally get an accuracy of 91% in the identification of fixed collocation queries.To classify the second and the third class, we utilize the text classification procedure, which contains word segmentation, stop word elimination, feature selection, and classification modules, to gain a high performance. We devise a new feature selection method (called "FEICHI") and use five feature selection methods as baselines. Experiments show that our new feature selection is valid and superior. In the classification of location sensitive queries, we also use two other methods as baselines to show that our method is correct and get better accuracy and recall. We also do a lot of Auxiliary experiments, including the influence of feature set size, the influence of different dataset and the influence of the split of dataset. All these experiments verify that our method is correct, valid, stable and superior.