Computers, Society, and Nature

I appreciate the title of JMonterey’s blog! Spatial data mining, as described in the article by Shekhar et al., seems exactly like extracting precious resources out of the underground or a kind of ‘homogeneous’ set of data. The article gives great examples of the type of ‘gold’ that we can get from the data mining processes and it’s applications (e.g.: crime, safety, floods,…). The article points out again to me that the techniques and methods of data mining depend on what is the application and what is the type of information that the researcher is looking for. Remember the example of data mining in the article starts from the question related to bird-nesting habits. This implies choosing the right set of data for our question and establishing where we are going to find the information we are looking for.

I’m left with the question of time… The ‘gold’ or outliers are spatial objects with non-spatial characteristics that differ from their neighbors’ characteristics. But what is a neighbor? I’m wondering where the notion of time fits in the data mining models, because two spatial neighbors could have a very distant relationship if we consider time, processes, change and interactions (recall the notion of absolute space/relative space in Marceau’s article on the problem of scale).

S_Ram

I found myself more interested in what you guys had to say about scale than the texts! or at least more inspired…
One point that is really interesting is that we have fancy techniques to choose a right scale to study a problem but in the end, the scale problem persists. Here again I think that this is a good scientific problem because the researcher has to ask questions and analyze the tools he is using in defining his questions and his methods. This is why I disagree with Point McPolygon’s point at the end of his blog saying that we should develop the technology more rather than attempting to define what is the right scale of study. I think that revealing scale issues is part of the science, and that it is important to study this instead of merely relying on the technology, which anyway is subject to human development and thus human’s understanding of the scale issue…
Wyatt ask is we can be accountable for issues of scale. Maybe by questioning the appropriate scale to use and the appropriate way to transfer information across scale is revealing a problem that we might not completely solve but being transparent about the process is a step towards not lying with the map and accountability.
The example of  adapting to climate change brought up by Victor Manuel is really interesting on my point of view. Different strategies are implemented at different scale of governance (country, region, municipality, communities), using different types of information or different ‘granularity of information’. On top of it dynamics occur between the scale of governance and thus between information. Defining the appropriate scale of information would depend on the scale of governance that your studying. Although most importantly in this case would probably be the problem of transferring information from one scale to the other. This is a very difficult task because of the issues related to MAUP,… mentioned in the texts, but also because of the meaning that the different scale of governance gives to geospatial areas, entities, concepts, processes,…

S_Ram

Guo and Mennis outline the emerging field of spatial data mining for the introduction to a special journal issue.  Work in this field has been prompted by the ever-increasing availability of finer and finer grained data, from an exponentially increasing number of sensors ranging from satellites to cell phones to surveillance cameras.  There has been a great interest in tapping into and making sense of these streams of “big data”, but in order to do so we must develop new ways of exploring, processing and analyzing them.  This is essentially what was alluded to in relation to geostatistics last week: our data and technology have surged way ahead of our available methodological toolbox.

One of the biggest issues with these new data sources is that they are largely unstructured: for example, they may just be a string of text such as a tweet, with some locational metadata.  In order to analyze a large number of unstructured data points, it is necessary to impose a structure via classification.  This is no easy task!  Although computer programs such as qualitative coding software packages exist and can group phrases by theme, most classification algorithms that exist necessitate a training process where the user tweaks the parameters of the classifier manually on a subset of the data.  The development of foolproof unsupervised classifiers that can not only sort unstructured data effectively, but also do so in a way that the output is of use to researchers, is a major challenge in this domain.

A key related idea to the advent of big data is the long-standing trade-offs between resolution and extent in spatial scale.  Though big data presents us with both extent and resolution of unprecedented magnitude, there still remain the limits imposed by humans’ own cognitive abilities.  Computer programs developed to make sense of big data must classify and generalize the raw input data in a way that allows geographers to effectively navigate this sea of data, rather than simply leaving us lost.

-FischbobGeo

The article “Spatial Data Mining, by Shashi Shekhar, explains what data mining is and how it has made great strides across various categories such as location prediction, spatial outlier detection, co-location mining, and clustering.  Data mining is finding meaningful patterns or information in data from a large data set that would otherwise have been imperceptible. This can be done in many ways, using either statistical tools or modeling, or a combination of both. The modeling usually takes a training set of data, and applies it to a testing set of data in order to build the model. One of the classic challenges of data mining is to take into account the spatial autocorrelation and spatial heterogeneity during this process.

The main unsolved problems of spatial data mining lie in the intersection of geospatial data and information technology. As GIScientists, this relates directly to many of the other subjects that we study. As is often the case in more modern applications of this science, we are limited more in methodology than by the technology itself. The advantage to a concept such as this is that patterns may emerge that nobody had previously considered, as opposed to doing statistical tests on hypothesized meanings of datasets. This technology opens a whole new world of possible advances in knowledge, both relating to GIScience and otherwise.

Pointy McPolygon

There is no doubt that scale plays a large role in the way in which data is interpreted. The article by Atkinson and Tate provide a good overview of scale of measurement, scales of spatial variation, and the issues inherent to spatial data. However, if we draw from Kathryn’s seminar about spatial statistics, and realize that large scale spatial processes impact smaller scale processes and their patterns, I’m still unclear as to how rescaling of data top down, bottom up or applying geostatistical techniques  can quantify the effects of large scale processes on smaller spatial processes.

An interesting suggestion that the authors cited from Milne (1991) was that to understand heterogeneity, conduct analysis across a wide range of measurement scales and extract the parameters that remained consistent to changes in scale. Though an expensive and time consuming task, if this could be done then it seems like great way to extract these features and focus on parameters that are scale sensitive to determine the appropriate scale necessary for analysis. Also, can those parameters that are robust to change tell the researching something about the study at hand? With the intensification of geospatial data and larger datasets, developing the necessary tools to better integrate multi-scale datasets for a more comprehensive evaluation is a mountainous task.  A tough GIScience topic with no easy answer, but it is crucial that we recognize that the scale of measurement we choose and the changes to data variability once we rescale data can greatly affect the final results of the analysis.

-tranv

Spatial data mining relies on the geographic attributes of the data to uncover spatial relationships within the dataset leading to knowledge discovery.  No doubt that if implemented successfully, then it has contributed to developing spatial theories, and contributing to geographic research. It’s a science!

The authors contend that with the increasing contribution of volunteered geographic information, GPS and LBS technology provides new research directions for the field. Though there is an abundance spatial data, and borrowing from Beth’s theme of critical GIS – we have to be mindful that those that can/do contribute is not representative an entire population. Bias is inherent in obtaining data from these sources because certain groups of people have been known to contribute more, certain locales can be headlined more often, and there is a large group of individuals who are disenfranchised because they do not have access to such technologies and completely sidelined. The authors concern themselves with developing the right questions and methodologies to solicit the answers, but is the data appropriate to answer the questions being asked?

Though there are several techniques to begin the spatial data mining process, how does the user decide which technique is appropriate for their analysis? Given that the end user may not be well versed in spatial data mining and nuances that exist within different types of techniques, different results will be generated by different rule sets, classifications. Since spatial data mining is a multidisciplinary field, who should be ultimately responsible for teaching the theories and methodologies?

-tranv

After a good half-century of quantitative developments in geography and geographic developments in statistics, Nelson takes stock of the field of spatial statistics by asking eminent spatial statisticians (statistical geopgraphers?) for their take on how the field has developed, current and future challenges and seminal works that should be on any quantitative geographer’s bookshelf.  She synthesizes the researchers’ responses to get at the broader trends characterizing spatial statistics.

A key shift discussed by Nelson is the state of advances in methodology vis-à-vis data availability and size. As spatial statistics grew in tandem with the Quantitative Revolution in the mid-20^th century, geostatistical methods were in many ways ahead of the available data and technology: computers and automated data management technologies were still nascent, limiting the quantity of data that could be analyzed to what could be managed by hand or by using punchcard systems.  Meanwhile, data collection and organization was onerous and typically manual.  Today, we have the opposite problem: we have TONS of processing power to perform complex calculations, and programming languages to implement new methods easily, so much so that our technology is now ahead of most conventional methods of spatial analysis.  Of particular importance is the new problem of how to work with big data, which may provide more comprehensive samples (even data for the entire population!), in a finer temporal resolution and a richer detail than ever before.

Rising up to the challenges presented by big data and stagnant methods will be paramount for the continued relevance of spatial statistics into the future.  However, today’s cohort of geography students may be falling behind the curve in their technical ability to respond to these challenges.  Mathematics and computer science, absolutely crucial to working in advanced spatial statistics, are receiving less and less of a focus in our Geography departments (though this is starting to change with computer science, specifically in relation to GIS curricula).  Indeed, the very core of quantitative methodology in geography has been shaken by the Cultural Turn.  While qualitative approaches are doubtlessly important to a rich understanding of geographical processes, there is a risk that geographers will lose their quantitative toolkit in a policy context where, increasingly, ‘numbers talk’.  Spatial statistics itself may also suffer if geographers are unable to bring their nuanced views of spatial considerations to the table.

When thinking about these issues, I come back to Nelson’s Figure 1, the Haggett view of progress in geography and helical time. It is still an open question whether we are on the cusp of a second Quantitative Revolution in geography, or whether spatial statistics and geographic thought will continue to drift away from each other, with potentially dire consequences.

-FischbobGeo

“It is ironic that ontology is proposed as a mechanism for resolving common semantic frameworks, but a complete understanding and a shared meaning for ontology itself are yet to be achieved”
–Agarwal

Agarwal, a computer scientist, takes on the challenging task of applying the concept of ontology to Geographic Information Science.  After introducing the concept, the author discusses some applications of ontologies in GIScience, stressing that ontologies underpin much of the ‘scientific’ potential of the field.  She then outlines and discusses the considerations and challenges associated with developing and implementing a common ontology for GIScience.

It is clear from the article that the concept of ontology is a complex one, grounded in classical philosophical thought.  It is also apparent that ontologies are of the utmost importance to applications such as artificial intelligence, but why exactly should GIScientists drop what they’re doing and consider ontology?  Agarwal gives some reasons.  First, ontologies require us to come to a consensus on terms in the discipline that may currently be fuzzily or ambiguously defined.  Additionally, ontologies make the underlying assumptions and relationships of a GIS data and/or analysis model more explicit and transparent.  These advantages lend themselves well to the overall goal of data interoperability, the achievement of which will make everyone’s lives easier in the long run.

More fundamentally, Agarwal asserts that “the lack of an adequate underlying theoretical paradigm means that GIScience fails to qualify as a complete scientific discipline.”  Fighting words!  Is the legitimacy of GIScience, or even geography in general, rendered null and void by its lack of a unifying ontology?  Agarwal contends that the conceptual fuzziness and ambiguities of many terms and processes in GIScience and geography, as well as the difficulties associated with the concepts of scale and spatial cognition, not only make ontologies very difficult to develop, but should be considered problematic to the entire geographic discipline.  Problems of scale and other disciplinary ambiguities are typically handled by geographers arbitrarily based on the researcher’s individual research question: do we need a more strictly defined framework?  I would argue that it is not possible to get rid of geography’s messiness; that it is the complex subjectivities of geography that are part of what distinguishes it as a discipline and it is not really possible to separate these subjectivities from GIScience.  If anything, the challenge of applying ontology to GIScience should be a humble reminder that making sweeping and universal conclusions is problematic, but it shouldn’t stop us in our tracks.

-FischbobGeo

I disagree with the statement that Anselin and Getis’s article is completely irrelevant today. I think that the point they are making about being careful with the toolbox is even more crucial today. Yes, tremendous steps have been made in terms of tools and methods in spatial stats but the questions behind the computerized processes still need to be well asked. We cannot let ourselves be led by the GIS tools. The authors give great examples of wrong maps being made because of misinterpretation of data. We can do statistics blindly with any data, and we will have results at the end no matter what. However, it’s not guarantee that the results will mean anything. The analyses have to be well supported to give meaning and interpretation to the results. I will end by quoting the authors: ”the technology should be led by theoretical and methodological developments in the field itself.”

S_Ram