=Paper=
{{Paper
|id=None
|storemode=property
|title=Making Close to Suitable for Web Searches - a Comparison of two Approaches
|pdfUrl=https://ceur-ws.org/Vol-798/paper9.pdf
|volume=Vol-798
}}
==Making Close to Suitable for Web Searches - a Comparison of two Approaches==
https://ceur-ws.org/Vol-798/paper9.pdf
Making close to Suitable for Web Search
A Comparison of Two Approaches
Iris Helming1 , Abraham Bernstein2 , Rolf Grütter1 , and Stephan Vock3
1
Swiss Federal Research Institute WSL, Birmensdorf, Switzerland {iris.helming,
rolf.gruetter}@wsl.ch
2
University of Zurich, Department of Informatics Zurich, Switzerland
bernstein@ifi.uzh.ch
3
stephan.vock@gmail.com
Abstract. In this paper we compare two approaches to model the vague
german spatial relation in der Nähe von (English: ”close to”) to enable
its usage in (semantic) web searches. A user wants, for example, to find
all relevant documents regarding parks or forestal landscapes close to a
city. The problem is that there are no clear metric distance limits for
possibly matching places because they are only restricted via the vague
natural language expression. And since human perception does not work
only in distances we can’t handle the queries simply with metric dis-
tances. Our first approach models the meaning of these expressions in
description logics using relations of the Region Connection Calculus. A
formalism has been developed to find all instances that are potentially
perceived as close to. The second approach deals with the idea that ev-
erything that can be reached in a reasonable amount of time with a given
means of transport (e.g. car) is potentially perceived as close. This ap-
proach uses route calculations with a route planner. The first approach
has already been evaluated. The second is still under development. But
we can already show a correlation between what people consider as close
to and time needed to get there.
Keywords: Vague Spatial Relation, Local Expression, Region Connec-
tion Calculus (RCC), Route Planning, Reachability.
1 Introduction
Sometimes we want to search for places on the web and restrict the search results
to a specific area. But we don’t have an exact distance restriction in mind, we
just want to look for something that is close or not close to, a bit off and so on.
How can we make this restriction understandable to a search engine? So that
future users could simply apply these expressions as keywords without further
thinking about ”translating” them?
Google4 already delivers results for queries that include near. But these re-
sults show that it’s not really taken care of the meaning of the preposition: if
4
http://www.google.ch
you are looking for a ”hotel in Zurich” for example, it returns also hotels which
are in the area around Zurich. On the other hand, if disliking living in a city cen-
tre but wanting to get there quickly, you could search for ”hotels near Zurich”.
The result will show you hotels near Zurich but also some which are located in
the city centre. Also, these mechanisms can’t scale with regards to the reference
place (i.e. the place to which the first one is supposed to be near). So the area
for searching doesn’t have a bigger size if the reference place is bigger. With our
knowledge representation approach (cf. [2]) such scaling is performed through
the type of the administrative region of the reference place — the granularity of
found nearby places is decreasing if the referent is situated on a more fine-grained
scale, such as a village, or increasing if the referent is on a larger scale, such as
a city. In this paper we will present a novel approach for decoding ”nearness”,
which deals with statistical methods. We compare it to our previous knowledge
engineering approach. With the new approach scaling works through the cho-
sen means of transport — things that are near while driving a car may not be
near while walking. Using this approach one could, in a future implementation,
present the user the best results of nearby places for his traveling context.
Both approaches are meant to map the vague concepts of spatial relations
that occur in natural language onto concrete geographical regions or places.
2 Related Work
Schokaert, De Cock and Kerre [8] suggest augmenting the structured informa-
tion available to a local search service, such as Google Maps, with information
extracted from the web. They show how nearness information in natural lan-
guage and information about the surrounding neighborhood of a place can be
translated into fuzzy restrictions and how such fuzzy restrictions can be used to
estimate the location of a place with an unknown address. The vast amount of
data addressed by the authors, together with the kinds of examples they pro-
vide, suggest that their approach is targeted on mass searches. In our case, the
resources on the web, which could possibly be used to augment the searches,
are sometimes scarce. Our second approach also is a statistical one, but we are
using context information -traveling time- to give the best matching results for
a special purpose.
Yao and Thill [9] also follow a statistical approach to handle vague natural
language expressions of distance. Different to us, they directly transform their
results to distances. We are avoiding this since human perception of vague spatial
expressions doesn’t work on metrical distances.
Also their discussion of general problems when dealing with vague expressions
for distances is of interest for our approach. They are highlighting the importance
of the context when a person has to judge if a place is near another place. Among
others they name transport mode as an influence factor for the perception of
nearness, e.g. Is city A close to city B? Yes for airplane, no for car. With the
statistical approach we make use of this influence factor and show how contextual
information in terms of means of transport can be modelled.
Mata [5] presents an approach to geographic information retrieval integrating
topological, geographical and conceptual matching. For topological matching
topological relations are extracted from overlaying data layers; for geographical
matching constraints are obtained from dictionaries; for conceptual matching a
geographic ontology is used. A constraint defines two geographic objects (points
or polygons) as near provided they are connected by a third object (an arc,
e.g., a road), the length of which is less than a given distance. Different from
the approaches we are comparing, a metric distance measure thus is a necessary
condition for nearness, although not a sufficient. However, the framework seems
general enough to be aligned with that presented in section 3.
3 Knowledge Representation Approach
In this section we briefly summarize the important aspects of the knowledge
representation approach, which we presented in [2]. For modelling nearness with
this first approach we use information of the administrative structure of Switzer-
land, which can be obtained easily. It is freely available as a download for many
countries. An administrative region like a district is responsible for administra-
tive tasks like providing schools, medical supply, organizing elections and so on.
Often borders for such regions are grown where also natural barriers like big
streams exist. It has been shown before that the partitioning of a country into
smaller parts has influence on human perception of space. Maki [3] for example
showed that the affiliation to a state plays an important role in human percep-
tion of locations. In an experiment, subjects had to decide about the location
of two cities regarding their orientation east-west. If the cities in question be-
long to different states, the reaction times were significantly shorter than with
cities which belong to the same state. Human beings are able to judge faster
about entities on a continuum if they can make use of category information.
Among others, Carbon and Leder [1] showed that the membership to different
political systems, structures or hierarchies influences the estimation of distance
between two cities. They used estimation tasks for distances between cities east
and west of the former border inside of Germany. Compared to pairs inside the
same part of the former republic, distances were overestimated if the cities in
question belonged to different parts.
Topologies of regions can be relatively easy obtained via modern geographic
information systems (GIS) or spatial databases. Types of administrative regions
and toplogical relations between them provide us with the possibility to reason on
these regions as polygons. Randell, Cui and Cohn [7] developed a a set of spatial
relations in first-order logic, the Region Connection Calculus (RCC), which we
use for it. How this works is described in the next section. With this approach
we provide a qualitative method for qualitative search queries. As Mark and
Egenhofer already conclude metric is not the most important parameter of the
semantics of most spatial natural-language expressions5.
5
”The topological relations come out as the strongest discriminators approximately
22 times stronger than all metric parameters combined which confirms the under-
3.1 Methods for the Knowledge Representation Approach
Knowledge is organized in a sample OWL DL Knowledge Base KB, consisting of
a TBox T and an ABox A. Partitions of regions are represented in T, partially
ordered in a way that each element of a partition is a subset of an element of
the next upper level of partitions. Each partition is typed and the concepts for
typing are mutually disjoint, so that an individuum can only be of one type. So
assume you have the partitions C(xi )i∈I and D(yk )k∈K of the same region, their
types are C and D. C(xi )i∈I is understood as more fine-grained than D(yk )k∈K
if each element of C(xi )i∈I is a subset of an element of D(yk )k∈K . For instance,
District(yk )i∈I is partitioned by elements of Community(xi )i∈I and both are
partitions of a canton. PartOf relations are kept functional, which means that
regions are only asserted as part of the next upper region but not as part of
the region above the next upper hierarchical step. In the ABox A one finds
assertions like partOf(Dietlikon, District-Buelach), stating that the individual
Dietlikon, which is of the type community, belongs to Bülach, which is a district
of type. Further all individuals are asserted as different from each other. The
Region Connection Calculus can be used to represent spatial relations in first
order logic. There are different sets available (e.g. RCC-3, RCC-5, RCC-8). For
our purpose, we are using RCC-8, which means, we are using a RCC set that
differentiates 8 relations. You can see the 8 relations in Figure 1.
Fig. 1: The Region Connection Calculus with 8 relations
Altogether these relations form a jointly exhaustive and pairwise disjoint set.
RCC relations can be interpreted temporal and spatial. Within the spatial inter-
pretion, regions are considered as sets of points. According to that two regions
which are connected to each other have at least one point in common. Rules
are formulated in a subset of the Semantic Web Rule Language (SWRL)[6]. Our
Rule Base is relatively small, in it, existing relations of the Region Connection
Calculus (RCC) are used as basis to form Composition Rules. For example there
are rules for the additional relation close to (cf. [2]).
lying assumption that topology is more critical for the semantics of spatial relations
than metric” (Mark and Egenhofer 1994, p. 227 [4]).
From RCC-8, we are using the relations part of P (TPP, NTPP, TPPi,
NTPPi), partially overlaps PO and externally connected EC to form the ba-
sic relation close to. Disjunctions of RCC relations in the bodies of composition
rules are represented by auxiliary roles, such as {P, PO} subsuming the roles
partOf and partiallyOverlaps. This allows composition rules that are expressed
as (non-disjunctive) Horn rules (see equation 1).
3.2 The added Relation CLOSETO
To the set of RCC-8 relations we added a composed relation CLOSETO. A
location x is close to y, stated asCL(x,y). The following equation shows the
actual CLOSETO rule:
∀x∀y∀z[CLap (y, x) ∧ z{P, P O}y → CL(z, x)] (1)
It is read as region z is close to region x if region y is a priori close to x
and z is part of or partially overlaps y. This rule makes up the basic building
block of this approach. In addition to the basic rule the knowledge representation
approach also includes the notion of ”a priori”-closeness, which is derived by a
second rule (cf. [2]). This second rule enables us to include functional micro
regions additionally to the administrative regions into reasoning. These micro
regions, consisting of mountane and space planning regions, were introduced to
analyze the behavior of commuters. Since these micro regions are also related
to traveling their addition seems to be useful for the comparison of the two
approaches. For more details please refer to [2].
3.3 Results of the Knowledge Representation Approach
In previous papers it has already been shown that this approach works for the
part of Switzerland that is covered by the sample ontology. Also an evaluation
has been performed using the search engine ”GoForIt”6 , which provides general
search and directory search as shown in [2]. For 170 communities two different
kinds of queries were performed. Firstly plain queries, such as ”communities close
to Dietlikon” and afterwards a query which contained a concatenation of all the
communities which have been inferred as ”close to”, such as ”Nürensdorf OR
Dübendorf OR Rümlang OR Wallisellen OR Kloten OR Wangen-Brüttisellen
OR Bassersdorf OR Opfikon” (all communities inferred as close to Dietlikon).
Finally, the results showed that recall was significantly higher for the rewritten
query.
4 Statistical Approach
We will represent now the second approach which is based on the idea that
people speak of places as close to if they can reach them quickly. Imagine you
6
http://www.goforit.com/
plan a picnic in a forest nearby and because you have lots of food to take with
you you want to go there by car. Then you will speak of a forest you can reach
in a reasonable amount of time as close to. In terms of metric distance this
place could be farer away from your location than another one. But the other
place doesn’t appear as being close to you because it would take longer to go
there by car. Sometimes the occurrences of close to will differ from one means
of transport to another. If you are using a car you can reach things in greater
distance easier than while you are walking. On the other hand sometimes you
find paths through woods which you can take while you are walking but not if
you are driving a car.
4.1 Methods for the Statistical Approach
To gain language data in an appropriate amount the german newspaper text
corpora of the Institut für Deutsche Sprache (IDS, Mannheim7 ) were used. Al-
together it has a size of 5.3 million tokens. German is used as object language
because the application of the approach should start in the German speaking
part of Switzerland8 . The keyword string in der Nähe von (i.e. ”close to”) was
looked up in the corpus. Because of the great ambiguity of toponyms and since
there are not yet good filters for toponyms available items were annotated man-
ually for the two close places. Then all identified place names were looked up in
a gazetteer to obtain the coordinates. Geonames9 and Swissnames10 were used
for this. The additional inclusion of Swissnames aimed at getting rich data of
Switzerland so that we could guarantee the comparability to the first approach
which is only implemented for a part of Switzerland yet. Then the coordinates
of the identified places were fed into a route planner. The routing API of cloud-
made11 , which is based on OpenStreetMap12 , served best for this purposes. If
the place name was ambigous, the nearest match was chosen. The routes were
calculated for trips by car, bike and walking. To get counter examples also hits
for ”nicht in der Nähe von” (not close to), ”weitab von” (further away from)
and other expressions for counterparts of close to were annotated. With these
there were some difficulties since they are - except the not close to - not direct
antonyms to close to and often they were used as a subjective statement of how
far something is away with regard to some topic. Part of the instances, for ex-
ample weit entfernt von (”far away from”), were often used to neutrally express
distance in combination with a metric distance measure. An example would be:
7
http://www.ids-mannheim.de/
8
In other countries other amounts of traveling time maybe felt as near. For example,
this amount of traveling time for Switzerland may be around 12 minutes. But for
larger countries with only few cities this may even be 2 hours. In the future, one
could calculate country specific traveling values.
9
http://www.geonames.org/
10
http://www.swisstopo.admin.ch/internet/swisstopo/en/home/products/
landscape/toponymy.html
11
http://cloudmade.com/
12
http://www.openstreetmap.org/
10km weit entfernt von (”10km far away from”). And mostly, if places are not
near each other, this fact is not explicitly mentioned.
Further there was a technical problem. Sentences like Frankfurt, weitab von
Asien (”Frankfurt, far away from Asia”) occured. One can imagine, that it would
not be difficult to calculate a route from Frankfurt to anywhere, but how to man-
ifest the endpoint for that route in a whole continent like Asia? Therefore such
routes could not be calculated. Nevertheless, there are some negative examples
and calculated trips by car, bike and walking for them as well. Analogous to the
”near-matches”, when a name was ambigous in geonames, we picked the match
with farest distance between the two places.
5 Preliminary Evaluation
The new approach is intended to model closeness via reachability with different
means of transport (cf. section 4). Reachability is lowered by path obstacles.
5.1 Qualitative Illustration
See Figure 2 for an example where a mountain ridge would force you to make a
detour of 3 times the length of the direct path when traveling by car. Because
of this detour you would not say that Arosa and Davos, the places in question,
are close to each other. The statistical approach can take care of path obstacles
like mountains and lakes and missing direct connections between neighboured
suburbs and so on. (For the knowledge representation approach these two places
would be close to each other when applying the basic rule, because the commu-
nity of Arosa borders the district of Davos. They would not be close to each other
when applying the rule which has been extended for micro regions. Here we have
an example where the inclusion of micro regions underlines the reachability.)
5.2 Quantitative Evaluation
For all the 345 pairs of places which occurred in the corpus connected with in
der Nähe von (English: ”close to”), we calculated the route for going by car,
going by bike and walking. We collected traveling time and traveling distance.
The same holds for the 30 pairs of places which are connected by nicht in der
Nähe von (”not close to”) and the above mentioned (cf. 4.1) synonyms. Then
the true distance was also calculated for every pair.
Some data points are quite far away from the rest. A manual check of the 10
sentences in the corpus to the most far away data points showed that there have
been mismatches to geonames, because for one of the place names there was
only another item in geonames which did not match the place actually meant.
Therefore SPSS13 was used to draw a histogram of traveling time by car only
13
The SPSS software can be obtained under
http://www-01.ibm.com/software/analytics/spss/
Fig. 2: Arosa and Davosa are departed via a mountain ridge. Since
there is no direct street over the mountain, a car has to take all the
way around the mountains along Chur, Landquart, Küblis and Klosters.
Source of Map: Kantonale Verwaltung Graubünden, GIS-Kompetenzzentrum
(http://mapserver1.gr.ch/kantonalesstrassennetz/
kantonalesstrassennetz.phtml)
Fig. 3: histogram with 20 equal-distance classes for traveling time up to 10,000
seconds
up to 10,000 seconds. You can see it in Figure 3. It has 20 equal distance classes
which have a size of 500 seconds each. This histogram still includes 240 hits. It
shows a right-skewed distribution: in the first class we only have 28 occurrences.
With 71 occurrences, most points are found in the second class 500 to 1,000
seconds. Afterwards the amount of datapoints declines for the next two classes
like normal distributed data. That there are relatively few occurrences in the
first class let us conclude that things which are connected to each other are not
mentioned as being close to each other.
Fig. 4: Overview of Results in Numbers
A table with the results of the comparison can be seen in Figure 4. Like in
the histogram, we chose to display the results for the 240 cases up to 10,000
seconds traveling time as well. For example the maximum of time needed for a
route to a place would be over 8 days (706788 seconds). As already mentioned
this is because of mismatched place names with the gazetteer items. For the
same reason average (8.67 hours; 31,222 seconds) and standard deviation (28.63
hours; 103,053 seconds) show high values. The shortest trip to a closeby place
takes 2,15 minutes (129 seconds). For the 240 cases up to 10,000 seconds of
traveling time we have better results. The maximum value is 2,74 hours (9,859
seconds). Average is 34,53 minutes (2,072 seconds). The standard deviation 36
minutes (2,160 seconds) is still high. A reason for this could be that also other
means of transport have an influence here. So for example, one feels as close to
a place where one has a direct flight connection to. This could explain that the
histogram declines in waves, the second peak could for example make up the
places with direct flight connections. And since the second peak is much smaller,
we could conclude that this is because it is much more ususal to go by car than to
go by plane. The minimum traveling time of the not close to matches is slightly
below 10,000 seconds. This may back up that our decision to have a closer look
at occurrences below 10,000 seconds. With the not close to matches it is quite
natural that standard deviation is high, since we also used differnt synonyms for
not close to. Also the range for things which are felt as not close to may be very
widespread.
Results for pairs of places in Switzerland are listed again to compare them
to the knowledge representation approach which is only applicable to a part of
Switzerland right now. Since the knowledge representation approach deals with
administrative regions but not with villages we mapped every occurring village
to its community for the comparison. For Switzerland we only have 2 negative
matches, too few to say much about them. What can be seen by the positive
matches is the smaller scale: the range is between 4.42 minutes (265 seconds)
and 39.7 minutes (2,382 seconds) and the average for traveling time is only 12.13
minutes (728 seconds). We can conclude from that, that when both places of a
close to relation are situated in a relatively small country like Switzerland, also
the distances between closeby places are small.
Fig. 5: Sensitivity and Specificity of time (car), distance (car), time-bike,
distance-bike and true distance. It can be seen that time (car) and time-bike
are slightly better predictors than true distance.
A logistic regression calculation using SPSS with nearness as dependent vari-
able (1 for near pairs; 0 for not-near pairs) showed that distance, traveling dis-
tance and time are all correlating to nearness. But still we have too few values
for pairs which are not close to to do valuable prediction. Also with SPSS sensi-
tivity and specificity were calculated under the assumption that traveling time
predicts what is felt as close to. Results are shown in the roc-curve in Figure 5.
As you can see, the correlation of time was a bit stronger but with the data base
available now it shows not significance.
5.3 Comparison with the Knowledge Representation Approach
The comparison with the knowledge representation approach shows, that places
that occurred in the corpus as close to each other are also found to a great
amount by the knowledge representation approach. The best fit was gained with
application of the basic rule: 28 of 33 matches were found, which makes 84.85 %.
For the micro region extension there have been 21 of 33 matches, which makes
up 63.64 %. So the extension with the micro regions is more restrictive and
the basic rule found many of the closeby places. Maybe in the future one could
use additional regions, like the micro regions, to limit results of the knowledge
representation approach for a typing structure of regions that is related to the
context of searching.
The evaluation is not finished yet, we still have to gain more corpus data for
places which are not near. But we already have established the process to get
final results for our approach.
6 Discussion
Fig. 6: Overview of Advantages and Disadvantages of both Approaches
The table in Figure 6 shows an overview of the advantages and disadvantages
discussed in this section. The knowledge representation approach will be more
precise wherever there is only little data about transport connections like streets
available. But information about hierarchical structures of regions, information
about topology and the type of a region is needed. It is good for modelling all
kinds of landscapes (e.g. swamps, mountains) as polygons and reason on them.
The statistical approach needs much data for setting up critical times for things
which are closeby. Once after the approach started working with good predictions
it will be applicable wherever you have route information, but not if the route
cannot be calculated.
close to-calculations with the knowledge representation approach can be done
with all types of places that are specified in the ontology and only these. If one
wants to calculate closeby cinemas or bakeries, the ontology has to be extended
with such types of places and entities which are of these types. An example for
this is used in the discussion of the table in Figure 4: some places are villages and
only the types of community, district and canton are available. So a mapping of
the village to the community to calculate the close to-factor for the community
in which it was situated was performed.
The statistical approach is applicable whenever there is a known path be-
tween places. It works with point data, not polygons. For cities, states and so on
there is always used a point which lies within. Depending on where you are in
the state/ city the appraoch is more or less accurate. For that reason we have not
been able to calculate a route from Frankfurt to Asia (cf. the Frankfurt weitab
von Asien (”Frankfurt far away from Asia”) example from the not close to part
of the corpus, section 4.1). Nevertheless, under normal circumstances it is not
very likely that such a route is needed.
While the Knowledge Representation uses administrative regions, the statis-
tical approach uses the context of traveling. Traveling is important for perception
of things closeby but also the hierarchical administrative structure of a country
has some influence (cf. section 2). The Knowledge Representation method will
always lead to clearcut judgements close to or not close to. But since we are
dealing with natural language data which is often quite vague and has many
influence factors, in one context a place may be seen as close to the reference
place whereas in another it is not. The statistical approach can also say some-
thing about the ”shaded” areas, it can give the degree to which something is
near.
7 Conclusion and Outlook
We have shown that it is possible to model human language concepts of spatial
relations via description logical expressions using administrative regions as back-
ground knowledge or via reachability by different means of transport. They both
have advantages and disadvantages. As already mentioned we have to extend the
evaluation for the statistical approach. When this is done with satisfying results,
we want to embed more possibilities to model humans perception of vague nat-
ural language expressions for spatial relations, for example bei (English: ”next
to”), zwischen (English: ”in between”), etc. We will do this for the two presented
approaches and maybe for others which we will develop in the future. Ontologies
are providing good background knowledge for such additional models. There are
many influence factors for the perception of spatial relation, one could build up
a user-friendly system which first evaluates the most important models for the
users needs and then computes the best-matching results.
Bibliography
[1] Carbon, C.C., Leder, H.: The wall inside the brain: Overestimation of dis-
tances crossing the former iron curtain. In: Psychonomic Bulletin & Review,
p. 746750. No. 12 (4) (2005)
[2] Grütter, R., Helming, I., Bernstein, A., Speich, S.: Rewriting queries for
web searches that use local expressions. In: Bassiliades, N., Governatori, G.,
Paschke, A. (eds.) Rule-Based Reasoning, Programming and Applications;
Proceedings of 5th International Symposium, RuleML 2011-Europe. pp. 345–
359. Springer, Heidelberg (2011)
[3] Maki, R.: Categorization and distance effects with spatial linear orders. Jour-
nal of Experimental Psychology: Human Learning and Memory 7 (1), 1532
(1981)
[4] Mark, D., Egenhofer, M.: Modeling spatial relations between lines and re-
gions: Combining formal mathematical models and human subjects testing.
Tech. Rep. 94-1, National Center for Geographic Information and Analysis,
University of California, Santa Barbara, CA (1994)
[5] Mata, F.: Geographic information retrieval by topological, geographical, and
conceptual matching. In: F., I.F., A., R.M., S., L. (eds.) Proceedings of the
Second International Conference on GeoSpatial Semantics (GeoS 2007). p.
98113. Springer, Lecture Notes in Computer Science No 4853, Berlin (2007)
[6] Motik, B., Horrocks, I., Rosati, R., Sattler, U.: Can owl and logic program-
ming live together happily ever after? In: Cruz, I.e.a. (ed.) Proceedings of
the 5th International Semantic Web Conference (ISWC 2006). LNCS, vol.
4273, pp. 501–514. Springer, Heidelberg (2006)
[7] Randell, D.A., Cui, Z., Cohn, A.G.: A spatial logic based on regions and
connections. In: Nebel, B., Rich, C., Swartout, W. (eds.) Proceedings of the
Third International Conference on Principles of Knowledge Representation
and Reasoning (KR92). pp. 165–176. Kaufmann, San Mateo, CA Morgan
(1992)
[8] Schockaert, S., De Cock, M., Kerre, E.: Location approximation for local
search services using natural language hints. International Journal of Geo-
graphic Information Science 22 (3), 315–336 (2008)
[9] Yaoh, X., Thill, J.C.: How far is too far? - a statistical approach to context-
contingent proximity modeling. Transactions in GIS 9 (2), 157–178 (2005)