|dc.description.abstract||Deep learning techniques have acquired much attention and have been shown to outperform previous state-of-the-art methods in plenty of fields over the last years. This dissertation delves further into the deep time series model and its application to a variety of datasets, including natural language sequence datasets and real estate-related datasets, with a deeper insight using a comprehensive set of analytical methods and algorithms.
The first part is a study of the deep learning techniques for processing natural language data. A spatial translation interface is proposed that focuses on the spatial domain vocabularies and translates the natural language questions to structured queries executable by database management systems (DBMS). Inspired by the deep comprehension model, we propose a natural language interface(NLI) with the spatial comprehension model that is able to recognize the meaning of spatial entities based on the semantics of the context. Our system could support a flexible back-end of multiple database query languages, such as SQL and Prolog, which are all supported based on our effective strategy. A transfer learning strategy is also presented to deal with the challenge of translating spatial language into database queries. A basic model is trained on one sort of database query before being fine-tuned to work with another. The models are verified using the Geoquery dataset, and the performance is demonstrated to outperform conventional approaches.
The second portion covers the application of time series models to real estate forecasting. To complete the real-estate price prediction task, a large-scale real estate-related dataset is constructed, encompassing both static and dynamic features, combining the numerical real estate price history data from Zillow and the survey data from the Census Bureau public dataset. A carefully designed Transformer-based forecasting model which could capture the change of real estate prices and predict hotspot areas for investment in real estate is proposed based on this time series dataset. The model is designed to embed the data with sequential temporal features and combine them with non-temporal features for subsequent prediction tasks. The results of the experiment reveal that our suggested model has a high level of accuracy and surpasses all baseline models.||en_US