(no subject)
Today I met with Raffay and Amos, then plus Aaron.
Activity recognition at the loading docks of the GT Barnes and Noble. They have defined a grammer of 61 primitive events which are things like Truck Arrives Right Dock, Person Opens Door, Box Goes in Truck, etc. To capture the local statistical information, they use 3-grams of events. Then, as a similarity metric between activities, they take the difference of histogram of 3-grams. Using a dominant sets clustering algorithm, they learn different types of activities. And using ROC curves, Amos can predict error rates if new examples are added without having to rerun the dominant sets clustering.
I'll be building a subsystem that explains why the anomalies are anomalous, based on a model called the Absent-Present model. In the first step of this process, you select the highest importance elements (event ngrams) of the typical cluster member, using the relative frequency of an ngram in the cluster overall as a measure of importance. In the corresponding anomalous event you can mark the sequences which are most notably absent.
Next, you find ngrams in the anomoly that have a high relative frequency against the typical event in the cluster. You have identified important features in the typical that are Absent in the anomaly, and features Present in the anomaly that are not present in the typical.
The final output should be something like this:
Sequence 567 was anomalous to subclass Delivering Books.
A typical Loading Books sequence contains Open Door, Open Door, Open Door which was absent from the Sequence 567.
Sequence 567 contains Truck Pulls Up, Open Door, Box goes in Truck which was absent from a typical Loading Books sequence.
Raffay has an explanation here. Amos has an explanation here.
Activity recognition at the loading docks of the GT Barnes and Noble. They have defined a grammer of 61 primitive events which are things like Truck Arrives Right Dock, Person Opens Door, Box Goes in Truck, etc. To capture the local statistical information, they use 3-grams of events. Then, as a similarity metric between activities, they take the difference of histogram of 3-grams. Using a dominant sets clustering algorithm, they learn different types of activities. And using ROC curves, Amos can predict error rates if new examples are added without having to rerun the dominant sets clustering.
I'll be building a subsystem that explains why the anomalies are anomalous, based on a model called the Absent-Present model. In the first step of this process, you select the highest importance elements (event ngrams) of the typical cluster member, using the relative frequency of an ngram in the cluster overall as a measure of importance. In the corresponding anomalous event you can mark the sequences which are most notably absent.
Next, you find ngrams in the anomoly that have a high relative frequency against the typical event in the cluster. You have identified important features in the typical that are Absent in the anomaly, and features Present in the anomaly that are not present in the typical.
The final output should be something like this:
Sequence 567 was anomalous to subclass Delivering Books.
A typical Loading Books sequence contains Open Door, Open Door, Open Door which was absent from the Sequence 567.
Sequence 567 contains Truck Pulls Up, Open Door, Box goes in Truck which was absent from a typical Loading Books sequence.
Raffay has an explanation here. Amos has an explanation here.