Now we study another interesting issue called Mining Periodic Movement Patterns. Usually the sensor may register movement data, but in many cases you may have very sparse data. Just giving you a very interesting example. For example, animal scientists or bird scientists, they would like to study animal movements and bird movements. They put censors on the body of those animals. But remember the sensor usually is pretty small, on the other hand it will last, it will need to be last for a long time. Like, for a study of bird migration, a sensor on the body of the bird is expected to last for a year. So, how could a sensor be lasting for a year without charging the battery? The only thing you can think is we will get very sparse data, like 24 hours we may only register one GPS location although 24 hours the birds may fly very far away. Then if we want to find the periodicity, the interesting thing could be it's very hard to find the periodicity if there's not overlap, just because the data is so sparse. Then just to look at how to, thinking about bird flying patterns, we just think the bird, suppose this place is the nest of the bird, okay. Then, if you just look at the bird, how they exactly fly with a very sparse sensor data, you'll not be able to find any meaningful periodic patterns of bird flying. However, you may find every night, every evening the birds were flying back to the nest. [COUGH] In the morning it will fly out from the nest. So, from the nest point of view or from the cluster center point of view it does have some periodicity. Then we may take such location as our reference spots. That means the reference spots can be detected using density-based method because they come back quite frequently. Then with this, if we do just based on the nest, we say when they're in the nest, when they're out of the nest, so we will be able to find a nice and interesting periodic patterns. That means the periodicity is more obvious when we look at the binary sequence like in and out of the nest. So that's interesting [INAUDIBLE] can see if we only think of a nest the periodicity will be able to be detected for each reference spot using Fourier Transform and auto-correlation. Now we look at an interesting example on mining periodic patterns with sparse data. This is a real data sets about birds flying in North America. So this actually accumulate three-year bird migration data, they are very, very sparse. However, if you based on the dense points, you take those dense regions as reference spots, you will be able to find a periodicity because in the winter they're going to fly down to the South. In the summer they may fly back to the North. In the middle they may fly some places to have some other activities, so you will find a few interesting reference spots. Once we detect periodicity based on those reference points, we will be able to summarize such patterns in a periodic way because we find a period and we find the nice periodic patterns. So this is an interesting study, based on the very sparse data we will be able to mine periodic behaviors for moving objects. Then the previous study we will assume the periodicity can be detected because the period is sort of knowing, okay. Then, in many cases, the time related data can be scattered and sparse. For example, phone calls at a location, they could be very sparse. And then for such sparse data, if we look at the time, we will be able to map the calling time in the sequence based on the time you're say 5, 13, 26, 29. If you map this into the time sequence, it's very hard to find a periodicity. Then what about if we can guess a right period, like t. If t is 20, if we guess it right, you probably can easily see they form a very dense cluster. However, if you got a wrong guess on the period, likely this mapping will be scattered along the whole period. So that means the projection on the true period, it will show highly skewed (clustered) distribution. Based on this observation, some in-depth study and algorithm have been developed to detect the true pattern in a mining event periodicity from incomplete observations. So in this whole session, we discussed a few interesting issues. We studied Mining Spatial Associations, Mining Spatial Colocation Patterns, Mining and Aggregating Patterns over Multiple Trajectories, Mining Semantics-Rich Movement Patterns and Mining Periodic Movement Patterns. In this session we actually used the following research papers. If you would like to know more, we strongly encourage you to read those research papers. Thank you. [MUSIC]
