Predictive Spatial Search

In a typical spatial search problem, (mobile) users search for (stationary or mobile) entities that have spatial attributes. The user’s current location and/or the entities’ locations are considered to assess the relevance of the search result. On one hand, we believe that the user’s future location is more relevant to the search result than the current location. Hence, we study spatial search queries under predictive models of user locations. On the other hand and with the ubiquity of hand held devices, most users do not utilize the full power of spatial search and they do not know what to search for. Hence, we introduce a framework to answer the question: “given a user’s current and predicted locations, what would the user be interested in searching for and seeing as a query result?” More specifically, we propose a predictive spatial search approach that continuously monitors the user’s current location to: (1) predict the user’s future location and integrate this prediction efficiently in the spatial search query processing pipeline, and (2) predict the search keywords that are of relevance to the user, given the user’s location and context. The second type of prediction leverages the knowledge of existing search engines about the behavior of a global set of spatial search users and social media users. Such twophase prediction capability will enable search engines to “pre-search” on behalf of their users and, thereby, leading to gains in user experience, search accuracy, and communication costs. Introduction Commercial search engines recognize the value of tracking the users’ location to provide a better service. The typical straightforward approach is to upload the user’s location every time a search query is issued. However, the vision presented in [1] proposes a novel approach that utilizes the power of data streaming systems to track the users’ locations based on a routeselection preference policy. In this approach, the system accepts (voluntarily) the user’s declared route preferences along with other context-based information. This information helps the system “predict” the user’s location even under the absence of the real-time location feed, survive 2014 Specialist Meeting—Spatial Search Ali, Hendawi, De Cock, Teredesai—2 offline periods, reduce communication cost and predict the user’s future location with acceptable accuracy. The next set of challenges is to identify the issues that can help integrate spatial search capabilities with such predictive data streaming systems. This integration is the theme of our vision. Consider the scenario that a user is driving on the highway and searches for a restaurant. It may be desirable and more relevant to direct the user to his favorite restaurant chain (or similar such chains that are still several minutes ahead from his current location) than to simply identify a nearby less favorite restaurant chain. It is even better to predict that this user is, for example, heading to a football game. Consequently, we direct the user to the parade, events and activities next to the stadium that the user is not aware of. In this scenario, we notice two main characteristics: first, the search engine is geared up to perform a spatial search around a nondeterministic (yet to be predicted) future location ahead of time. Second, the search engine aims at discovering and recommending personalized search topics on a per user basis. Ranking of the search result needs to be performed spatiotemporally taking into consideration the probability of the user’s expected location along the future timeline. Currently, users rely on multiple data sources such as live feeds from social media augmented with organic search to discover related events/facilities at the destination. An advantage of the proposed predictive spatial search frameworks is to help consolidate search results better across various faceted search channels. We believe that future geospatial search engine will inevitably adopt the “we know where you will be and we search before you search” principle. It is the proactiveness of search engines that will shape the future of spatial search. The search engine is the user’s agent that continuously (1) adjusts the search result according to the user’s predicted destination and (2) interacts/socializes with neighboring people/agents to highlight interesting topics that the user is totally unaware of. Predictive Trees: An Index for Predictive Queries on Road Networks Practical experience tells that it is absolutely a myth to assume that commercial search engines know everything about the user’s past locations. Lots of research utilize manufactured databases of historical trajectories and apply machine learning techniques to predict the user’s future location. These trajectory databases are usually collected by researchers or volunteers for research purposes. Yet, from a practical perspective, and due to privacy concerns [2], the user location is revealed on a session basis such that each session is no longer than few minutes. Commercial search engines care about users’ privacy. Consequently, techniques that assume full knowledge of the user’s behavior over extended periods of time are not considered practical in our approach. We propose a new index structure, the predictive tree [3, 4], that enables the evaluation of predictive queries [5] in the absence of the objects’ historical trajectories. Based solely on the connectivity of the road network graph and assuming that the object follows the shortest route to destination, the predictive tree determines the reachable nodes of a moving object within a specified time window T in the future. Moreover, predictive trees utilize every additional piece of information and enhance the probability assignment of the predicted location as more trajectory data becomes available on the user. 2014 Specialist Meeting—Spatial Search Ali, Hendawi, De Cock, Teredesai—3 The predictive tree: (1) provides a generic infrastructure for answering the common types of predictive queries including predictive point, range, KNN, and aggregate queries, (2) updates the probabilistic prediction of the object’s future locations dynamically and incrementally as the object moves around on the road network, and (3) provides an extensible mechanism to customize the probability assignments of the object’s expected future locations, with the help of user defined functions. In our ongoing effort, we leverage predictive trees to support spatial search and integrate this work with predictive data streaming systems. References [1] Mohamed Ali, Badrish Chandramouli, Balan Raman, and Ed Katibah. Spatio-Temporal Stream Processing in Microsoft StreamInsight. IEEE Data Eng. Bull. 33(2): 69–74 (2010). [2] From GPS and Virtual Globes to Spatial Computing—2020: The Next Transformative Technology. A Community Whitepaper resulting from the 2012 CCC Spatial Computing 2020 Workshop. [3] A. M. Hendawi, J. Bao, and M. F. Mokbel. Predictive Tree Source Code and Sample Data. URL:http://www-users.cs.umn.edu/~hendawi/PredictiveTree/, Aug. 2014. [4] A. M. Hendawi and M. F. Mokbel. Panda: A Predictive Spatio-Temporal Query Processor. In ACM SIGSPATIAL GIS, 2012. [5] A. M. Hendawi and M. F. Mokbel. Predictive Spatio-Temporal Queries: A Comprehensive Survey and Future Directions. In MobiGIS, California, USA, Nov. 2012. 2014 Specialist Meeting—Spatial Search Ballatore—4 The Search for Places as Emergent Aggregates ANDREA BALLATORE Postdoctoral Researcher Center for Spatial Studies University of California, Santa Barbara Email: [email protected] earching for places is the most popular geographic online task, in which names and categories are used as the main referents to locate places in the geographic space. By typing “hotels in Santa Barbara” in any popular search engine, the user expects a list of places matching a category (“hotel”) contained within another place called “Santa Barbara.” Answers to such a query can be generated by relying on a gazetteer containing some form of spatial footprint for the symbol “Santa Barbara,” a database of points-of-interest categorized as “hotels” and, indeed, some strategy to compute the relevance of the potential results. This approach satisfies welldefined information needs, but fails to account for more complex, nuanced, fuzzy, and yet cognitively intuitive questions about the town. What places are similar to Santa Barbara with respect to its general atmosphere—but perhaps less expensive? What other towns in Southern California offer a comparable array of amenities? What tourist areas in Italy provide a similar combination of mountain-related and marine activities? Our current computational models of place search do not seem to provide easy answers. Our intimate familiarity with place clashes with the difficulty of dealing with it computationally. Because of its centrality in human cognition and culture, the notion of place is unsurprisingly characterized by high polysemy, strong context-dependence, and innumerable metaphorical uses, carving social meanings from neutral, unbounded spaces (Agnew, 2011). The intellectual prominence of place has waxed and waned over time, being obscured for centuries by more abstract notions of space, and making a reappearance in recent decades (Casey, 1997). Being intensely debated in the social sciences and the humanities, place and its representations have now become an active research frontier in geographic information science (Goodchild, 2011). In this area, a central concern is that place names are often ambiguous, vague, and vernacular. Place categories are culturally-dependent, arbitrary, and inconsistently applied. More strikingly, places are implicitly assumed to have a name and to fit, at least to some degree, known categories. Current efforts focus on more sophisticated place-name interpretation in text documents (Purves, and Jones, 2011), new reference theories tailored to place (Scheider and Janowicz, 2014), and semantically more expressive gazetteers (Keßler et al., 2009). In a complementary approach, I advocate a view of place as an aggregate of objects and processes that interact at a given scale, inter-locked by spatial colocation. This view of place relies on the discovery of implicit relations, and not on some explicit labels assigned by an observer. The approach relies on some assumptions: Places are inescapably multi-faceted (comprising diverse processes), they are socially constructed (emerging as the result of human agency and practices), relational (emerging in a context, not in a vacuum), scale-dependent (different places exist at S 2014 Specialist Meeting—Spatial Search Ballatore—5 different scales), and they are dynamic (emerging, changing, and ultimately disappearing). Following Thrift (1999), I regard places as emergent entities in a complex, non-linear system, and they appear as assemblages of heterogeneous things that meet in space and time. Place can be fruitfully viewed through a holistic lens, emphasizing its contextuality and inherent interconnectedness, rather than as an object in isolation (Ballatore et al., 2012). In practical terms, this approach aims at supporting multi-faceted, context-dependent aggregate search, going beyond the current forms of name-based search for well-defined, individual places. In this sense, places can be searched for on the basis of their emergent distributional characteristics, rather than in an arbitrary, crisp categorization (e.g., city or town). For example, a Japanese tourist in San Francisco might search for places in which architectural landmarks co-occur with museums and galleries, or for places that, as an aggregate, present similar characteristics to Shibuya, the Tokyo shopping district she is familiar with. Text Information Retrieval Place-as-aggregate Search Vector space Set of text documents Geographic space Vector A document as a sequence of words A place as an aggregate of spatially located objects Dimensions Words (high dimensionality) Characteristics of objects (dimensionality defined by application, potentially very high) Index Sparse document-word matrix Sparse object-to-object colocation matrix Search Weighted keyword matching, topic models, similarity Colocation queries, query-by-place Table 1: Place search overview Computationally, this approach to place search can be modeled in analogy to traditional information retrieval in a vector space model, as summarized in Table 1. Given a geographic space, treated as a corpus containing a potentially infinite number of place-as-aggregate, the proposed approach has a number of opportunities and challenges to be tackled. The efficient computation of spatial colocation at a large scale—the identification of categories of objects that co-occur spatially (and temporally) in non-random patterns—is an open problem (Cromley et al., 2014). As many alternative places-as-aggregates can encompass the same entities at different scales, new heuristics are needed to construct optimal aggregates that meet user informational needs at a given scale, based on statistical measures of informational entropy. As I hope I have demonstrated, the uneasy relationship between space and place offers opportunities for unlocking new ways of searching the ocean of geo-information. References Agnew, J. (2011). Space and Place. In J. Agnew & D. Livingstone (Eds.), Handbook of Geographical Knowledge (pp. 316–330). London: Sage Publications. Casey, E. S. (1997). The fate of place: A philosophical history. Berkeley, CA: University of California Press. Cromley, R. G., Hanink, D. M., & Bentley, G. C. (2014). Geographically Weighted Colocation Quotients: Specification and Application. The Professional Geographer 66(1): 138–148. 2014 Specialist Meeting—Spatial Search Ballatore—6 Goodchild, M. F. (2011). Formalizing Place in Geographic Information Systems. In L. M. Burton, S. A. Matthews, M. Leung, S. P. Kemp, & D. T. Takeuchi (Eds.), Communities, Neighborhoods, and Health (pp. 21–33). New York: Springer. Keßler, C., Janowicz, K., & Bishr, M. (2009). An Agenda for the Next Generation Gazetteer: Geographic Information Contribution and Retrieval. In Proceedings of ACM GIS ’09 (pp. 91–100). New York: ACM. Purves, R., & Jones, C. (2011). Geographic information retrieval. SIGSPATIAL Special 3(2): 2–4. Ballatore, A., Wilson, D. C., & Bertolotto, M. (2012). A holistic semantic similarity measure for viewports in interactive maps. In S. Di Martino, A. Peron, & T. Tezuka (Eds.), Web and Wireless Geographical Information Systems (pp. 151–166). Berlin: Springer. Scheider, S., & Janowicz, K. (2014). Place reference systems. Applied Ontology 9: 97–127. Thrift, N. (1999). Steps to an Ecology of Place. In D. Massey, J. Allen, & P. Sarre (Eds.), Human Geography Today (pp. 295–322). Cambridge, UK: Polity Press. 2014 Specialist Meeting—Spatial Search Card et al.—7 The VERP Explorer— A Tool for Applying Recursion Plots to the Eye-Movements of Visual-Cognitive Tasks STUART K. CARD Consulting Professor Department of Computer Science Stanford University Email: [email protected] with ÇAĞATAY DEMIRALP Post Graduate Fellow Department of Computer Science Stanford University Email: [email protected] JESSE CIRIMELE Graduate Student Department of Computer Science Stanford University Email: [email protected] esigns in human-computer interaction (HCI) often involve trading between spatial and textual representations to achieve a nuance of representation that makes a task faster to execute, easier to learn, or less prone to error. Such designs can be very effective, but they can also be subtle, and it can be difficult to understand the mechanisms in play. Even generally successful interfaces can still hide bad combinations of interface, task, and context that could be improved were they identified. One method of approaching this problem is to run chronometric experiments with contrasting conditions. Aside from being expensive for development work, this method is at such an aggregate level that it often does not provide much access or insight into the underlying mechanisms at work. Another method is cognitive simulation (Kieras, 2014). The intent is to specify the likely mechanisms at work and to validate them by their ability to predict chronometric or other data. The validated simulator can then be put to work on inferring other consequences of the design with some claim to knowing why. While this method has advantages, it is even more expensive and is most practical for large projects or projects close to an existing model that can provide a starting point. Figure 1. The VERP Explorer. D 2014 Specialist Meeting—Spatial Search Card et al.—8 A third method is to construct a tool that makes the mechanisms at work visible by when applied to samples of user behavior. In this paper, we propose such a method and tool, The VERP Explorer (VERP stands for Visualization of Eye-Movements based on Recurrence Plots), an interactive visualization based recurrence plots. Eye-movement sequences are taken of users performing visual-cognitive tasks with the subject system. These are mapped into recurrence plot visualizations to highlight patterns of quasi-sequential behavior. In our system, these patterns are then back-mapped into—and overlaid on—the eye-movement scene to help characterize and provide insights into the behavior. Recurrence plots are a type of non-linear analysis that has been used in the study of dynamical systems and other areas (Eichmann et al., 1978; Marwan, 2008). Recently it has been applied to eye-movements (Anderson et al, 2013). Our tool extends and integrates eye-movement and recursion plot analysis into an interactive tool, simplifying exploratory analysis. Eye-movements can be thought of as a sequence of eye gaze positions fi parameterized by time. To obtain the matrix [ ] that is the basis for a recurrence plot, we start with the first eye-position f1 and compare it to all the other eye-positions in the sequence, including itself. If the distance d( fij ) between the two compared eye positions is within some small distance  , then we put a 1 at that position in the matrix, otherwise a 0. rij  1, d( fij )   0, otherwise   