Spatio-temporal data correlation

giobbigio

Hi everybody,

I'm new here and I'll explain what my problem is:

I've a database composed of a collection of positions and timestamps per object (the idea is that at irregural time intervals, one of the many objects moving, sends timestamp, latitude and longitude). I've searched through various topics here but I've found nothig about correlation of geospatial data.

 

My goal is to find a way to correlate data of this type and, if it is possible, I would undestand how can I realize a non-supervised anomaly detection system (just a point where to start will be appreciated)

 

Regards

SGolbert

Hi giobbigio,

 

this looks like a active research topic, I would find possible approaches in the scientific literature, for example in Google Scholar. I think that a problem could be tackled with appropiate feature engineering and a time series approach, but with only a brief description there is not much we can do.

 

If you search the literature and don't find a satisfying answer, we may be able to help further if you post some of the data and explain the problem in detail.

 

Regards,

Sebastian

giobbigio

Thanks for the reply.

I'm not a complete newbie to this kind of problems but i'm new to Rapidminer. The theory behind is quite clear.

Il' explain how my data are organized:

. My database is composed of objects moving in a geographical area. For this reason it is a collection of ids latitudes, longitudes and timestamps. The database is organized in this way: multiple examples are different latitude,longitude,timestamp of an object (the id repeats for each timestamp of the object):

 

id              lat           long          timestamp

obj.1      xx.xxxx   yy.yyyy       year/month/day h:mm

obj.2      xx.xxxx   yy.yyyy       year/month/day h:mm

....(until obj.n)

obj.1      xx.xxxz   yy.yyyz       year/month/day h:mm+x min

obj.2      xx.xxxz   yy.yyyz       year/month/day h:mm+x min

....(until obj.n)

....until N

 

(n = number of objects, N = (number of positions for each object)*n)

 

For example i'd like to find:

. correlation between position and timestamps of each single object (for example check if the position is similar to the  position of an hour ago)

. anomalies between objects (if they are near multiple times during the day or if an object that moves always on a same patter deviates)

 

Hoping I've been clear,

Regards

SGolbert

Hi,

 

I think that a good starting point would be to install the time series extension (you can find it under Extensions -> Marketplace). You would have to play around with it a while and read about time series if necessary.

 

When you get to the point in which you can analyze windows, it could be a good idea to analyze the differences of first derivatives of the time series: if the direction changes too fast, it could be an anomalous behaviour. Furthermore, other aggregations could be taken in order to generate features for a clustering approach.

 

As far as I know the implementation in RapidMiner is not more complicated than other use cases, but to learn the necessary machine learning concepts and process and generate features for it can be challenging.

 

Regards,

Sebastian