Building Location Embeddings from Physical Trajectories and Textual Representations

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 425–434 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 425 Building Location Embeddings from Physical Trajectories and Textual Representations Laura Biester, Carmen Banea, Rada Mihalcea Computer Science & Engineering, University of Michigan, USA {lbiester,carmennb,mihalcea}@umich.edu Abstract Word embedding methods have become the de- facto way to represent words, having been suc- cessfully applied to a wide array of natural lan- guage processing tasks. In this paper, we ex- plore the hypothesis that embedding methods can also be effectively used to represent spa- tial locations. Using a new dataset consisting of the location trajectories of 729 students over a seven month period and text data related to those locations, we implement several strate- gies to create location embeddings, which we then use to create embeddings of the sequences of locations a student has visited. To identify the surface level properties captured in the rep- resentations, we propose a number of probing tasks such as the presence of a specific location in a sequence or the type of activities that take place at a location. We then leverage the rep- resentations we generated and employ them in more complex downstream tasks ranging from predicting a student’s area of study to a stu- dent’s depression level, showing the effective- ness of these location embeddings. 1 Introduction Due to the rising adoption of smartphones over the past decade, the number of services with full or partial information about people’s spatial mobility has skyrocketed. Inspired by the natural language processing (NLP) literature, we investigate vari- ous properties of location embeddings. We explore whether valuable information is encoded in indi- vidual location embeddings, as well as embeddings that encompass a sequence of locations. We be- gin by exploring whether they are able to represent aspects such as location presence or location func- tionality. Ultimately, we test the hypothesis that if enough underlying information is encoded, embed- ding models should aid in predicting user-centered descriptors, such as area of study, academic status, or mental health. Location data can be used by university admin- istrators for applications that improve student life. From the frequency and the type of locations ac- cessed in one’s daily routine, we may be able to identify someone who is depressed or someone who is overworked. Importantly, opt-in frame- works can be established to supplement existing counseling and advising offices, allowing for early intervention in the case of mental health and aca- demic concerns. With proper privacy safeguards in place, such models could readily be applied on most university campuses, as WiFi connection data (from which we infer location) is likely already available. In addition, universities could use this data in an aggregate form to better understand stu- dent life and well-being, and find ways to promote healthy and engaging behaviors on campus. Such aggregate location information can also be used by architectural firms or municipalities to …

Content truncated to 500 words