Top-Down AND Bottom-Up CGA Conference: Illuminating Space and Time - - PowerPoint PPT Presentation

Top-Down AND Bottom-Up

CGA Conference: Illuminating Space and Time in Data Science

Krzysztof Janowicz

STKO Lab, University of California, Santa Barbara, USA

April 2018

Top-Down AND Bottom-Up

• K. Janowicz

Semantic Signatures As Analogy To Spectral Signatures

Geospatial (and platial) bands based on geographic location ANND Ripley’s K Bins J Measure 41 geo-stats bands Temporal bands based on geo-social check-ins 24 Hours 7 Days Seasons Thematic bands based on venue tips and reviews 500 LDA topics TF-IDF Make use of data diversity Utilizes data traces actively or passively emitted by humans and their devices

Top-Down AND Bottom-Up

• K. Janowicz

Spatially Explicit Modeling

Invariance test

Spatially explicit models are not invariant under relocation

Representation test

Spatially explicit models include spatial representations in their implementations

Formulation test

spatially explicit models include spatial concepts in their formulations

Outcome test

Spatial structures of inputs and outcomes are different

Top-Down AND Bottom-Up

• K. Janowicz

Place Types and their Similarity

Center POI Active Life Arts & Entertainment Automotive Beauty & Spas Education Event Planning & Services Financial Services Food Health & Medical Home Services Hotels & Travel Local Flavor Local Services Mass Media Nightlife Pets Professional Services Public Services & Government Religious Organizations Restaurants Shopping Distance Bin Street Network

Can we understand place types exclusively via their spatial interaction?

Similarity is a key prerequisite for geographic information retrieval, recommender system,

• ntology alignment, and so forth.

A Word2Vec-based, information-theoretic, distance-lagged approach to model the distributional semantics across 1030 (570) POI type that takes relative Popularity and Frequency (per bin) into account.

βlj

ITDL

• −ωlog2
• 1 −

Ptj 1 + |M|

k1 Ph tk

• − (1 − ω)log2(Fh

tj )

• Top-Down AND Bottom-Up
• K. Janowicz

Experiments

For ω 0.1 − 0.9 and 70 dimensional embeddings Kendall’s W of 0.79 among 25 raters across 10 seven-pair tests Remarkable result, e.g., ρ 0.7, as humans use substantially richer information to reason about similarity, e.g., the meaning of type labels, background knowledge Note that short as well as long-distance bins contribute to these results, e.g., the highest ρ is obtained by a concatenation of bins 4-17-1-5-24 (ω = 0.1)

Top-Down AND Bottom-Up

• K. Janowicz

Places365-CNNs for Scene Classification

Top-Down AND Bottom-Up

• K. Janowicz

Places365-CNNs for Scene Classification

Top-Down AND Bottom-Up

• K. Janowicz

Places365-CNNs for Scene Classification

Top-Down AND Bottom-Up

• K. Janowicz

Places365-CNNs for Scene Classification

Top-Down AND Bottom-Up

• K. Janowicz

Classifying Places Based on Images by Incorporating Spatial Contexts

Tested: Spatial relatedness, spatial co-location, spatial sequence pattern Outperform state-of-the-art image classification systems by over 40% (MRR) and more than double Accuracy@1.

Top-Down AND Bottom-Up

• K. Janowicz

There is no such Thing as RAW Data

Most of you would probably not conceptualize Sky or Wave as place types On top of showing that explicit spatial models matter, we also need to more actively contribute to the used datasets and their schema.

Top-Down AND Bottom-Up

• K. Janowicz

The Role of Space/Place for the Linked Data Knowledge Graphs

The publicly available part of the Linked Data cloud contains approximately 150 billion triples distributed

• ver 10000 diverse datasets

and connected to each other by millions of links. The private part contributes to Google’s new search engine, Apple’s Siri, IBM’s Watson, and so forth.

Top-Down AND Bottom-Up

• K. Janowicz

International Conference on the Internet of Things (IoT 2018)

http://iot-conference.org/

Top-Down AND Bottom-Up

• K. Janowicz