Miroslav Despotovic (p), David Koch, Wolfgang Brunnauer, Eric Stumpe, Simon Thaler, Mathias Zeppelzauer

Ákos jakobi

In order to understand interrelated spatial dependencies in the assesment of real estate location, analysts utilize various methods that examine e.g. neighborhood characteristics, sociodemographics, proximity and accessibility to spatial externalities.
The general problem with this approach is that the location characteristics and parameters of the assessment must be selected and defined by the analyst.
Thus, this approach depends on a considerable number of analyst's subjective and implicit decisions.
One way to address this issue could be to use complementary information from diefferent modalities as the key to more robust information extraction and higher data quality.
We conducted an experimental approach to determine the quality of residential locations based on iterative cognitive voting of image pairs using the Elo rating system of multiple visual modalities like Street View and satellite photos.
Using the obtained rating values, we built a ConvNet-based regression model to predict the Elo scores for a set of independent image representations.
We derive the inference for our experiment by predicting the rental and purchase prices using obtained locational Elo scores along with additional explanatory variables.
From the results of our study, we found that incremental assessment of location images by independent individuals in terms of active learning can serve as a control variable as well as that the utilization of complementary modalities in location quality assessment has potential for further investigation.

Kateryna Zabarina (p)

Miroslav Despotovic

Inside business location studies, one can distinguish between three main approaches - discrete choice models (DCM) (basically solution of optimization problem for each particular location and then choice the best one), count modelling approach (different factors impact on number of a new-born business plants in a given region) and spatial related methods, mainly point pattern analysis (major group of papers is based on application of Ripley's K and its derivatives). All abovementioned approaches are based on aggregated data, which do not seem to be appropriate for such problem - these data do not reflect spatial nature and distribution of data, do not account for spatial factors (such as localization patterns or issue of economic clusters and agglomeration economies) or existence of spatial dependence and heterogeneity.

Among studies operating on point data one should mention study of Bocci and Rocco (2016) which applies inhomogeneous Poisson process to investigate firm location determinants and papers of Sweeney and Gómez‐Antonio (2016; 2018; 2021) exploring determinants of location with Gibbs processes.

Sweeney and Gómez‐Antonio were not the first, who accounted for interaction within the radius – the pioneer attempt was made by Rosenthal and Strange (2003) (although made on aggregated data). However, studies of Sweeney and Gómez‐Antonio opened a wide possibility for further investigation. Innovative element of this paper is assumption that several interaction radii exist, thus instead of simple Gibbs process (considering only one interaction radius) hybrid of several Gibbs processes will be used.

Preliminary results show that consideration of spatial factors and several interaction radii allows to produce a model which passes goodness-of-fit tests. Later, such model may be used to analyse business locations.

Christer Persson (p)

Kateryna Zabarina

A number of entropy measures have been introduced for spatial data. Entropy in Shannon's formulation indicates the degree of heterogeneity in data where one end-point is the uniformly distributed observations, and the other end-point is where observations degenerate to a single value that corresponds to no heterogeneity. Spatial phenomena are characterized by different degrees of heterogeneity, and in this aspect, entropy is a relevant measure for spatial data. However,
entropy by itself says nothing about associations between locations which is a crucial aspect of spatial data. The focus, when constructing spatial entropy measures, is therefore to introduce spatial association in entropy. For the existing measures, this has been solved, for example, by computing the entropy of a spatially weighted distribution. Another way of introducing spatial associations has been to first construct a distribution for the so-called co-occurrence of pairs of observations and then compute this derived distribution's entropy.

This paper presents an alternative approach where the entropy is generalized from Shannon's approach by allowing a different functional form of the surprise part of Shannon's entropy. This form of entropy has been termed generalized entropy. In this application of generalized entropy, a spatial surprise function utilizing the adjacency matrix is used to construct a generalized spatial entropy. This entropy is a LISA (Local Indices of Spatial Association), which can be additively summarized to a global measure. By appropriately setting global parameters, it can be specialized to the ordinary Shannon entropy of the outcome variable.

Ákos jakobi (p)

Christer Persson

This study describes the change of global network characteristics of human movement determined from millions of geolocalized Hungarian mobile cellular network data. Mobile phone cells (or base transceiver towers) were considered as geolocalized nodes of the network and movements between two consecutive cells as network edges. As a result, the network based on locational relationships was appropriate to model mass spatial movements. The aim of the research was to answer the question: to what extent network characteristics change over time and what could we learn from the changes in relation with human dynamics. Alteration of network characteristics were measured by longitudinal change and variation in basic network metrics. Starting from 1st December 2018 until 30th November 2019 it was possible to analyse and compare each day’s network characteristics globally, hence providing 365 individual, but still related networks covering a whole year of data on human mobility. Results confirmed that typical movement periods, such as holidays, have significantly different network characteristics than average workdays. Furthermore, observable difference was measured between intensive movement periods (ie. summer months) and less busy periods of the year (ie. winter seasons). Finally, a reverse possibility also arose when actual mobility processes are presumed from observed network metrics. The study, therefore, aims to introduce this mobile cell based network methodology also as a possible forecasting and measuring tool for spatial human movements.