Numenta 2014 Spring NuPIC Hackathon Demos, 8 May 2014

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Moksha: 2014 Spring NuPIC Hackathon Demo

Description

Sergey & Mohamed

A

All right, so, I think we're the last teams I'm going to try to be really really quick. um The premise of our work is that, for some reasons that people who like to be called evolutionary psychologists would be eager to explain to you. People have friends and.

A

And we decided to uh run an anomaly detection on the friends check-ins into different places on facebook.

A

So uh we got the data from facebook, which is a relatively difficult process, considering they changed their api four days ago and nobody knows how it works. um But basically we um go ahead.

A

Yeah, so now it's downloading and uh analyzing the data to figure out which of the latest friends check-ins are anomalous. The reason why we think that's interesting is that most of our friends kind of do the same things every day, so we don't necessarily want to talk to them about it, but if they do something interesting, something outside of their range will likely to be to ask about it. The interesting part about this project is the encoders. You could.

B

Use I have no idea what the data is: you're, actually looking yeah.

A

So the data for a check-in is uh location of a place they checked in with whom they checked in um how many people like the place and categories, there's a lot of categorical data. How facebook categorizes different types of venues and.

B

You can get this for all your friends.

A

B

Okay, all right got it. Thank you.

A

Some friend, which is the default, uh so, uh as you can see, we we tried to uh make the things that are very well very um predictable kind of transparent. So you see that people who go to bars and assembly shops kind of transparent, but whenever it's an event that doesn't really occur. All that often it's a bit more opaque. The interesting part about this is creating the encoder because we're not predicting any scalar value, but rather the similarity between check-ins. So you can't do any swarming.

A

You have to manually, create the encoder and we assigned uh half of the sdr to location, that's half of what we consider to be anomaly and approximately a little bit more than half of the rest, two categories uh and the remainders to things like how popular is the place, how frequently the people check in to it and uh how many people have people checked in with.

B

How much data do you have how many.

A

So these are the people who have the most data um on average, a little bit more than 100 uh check-ins for per person.

B

A

Time period uh through throughout their history, so on facebook, yeah.

C

So all my friends basically and all their check-ins and all his friends and all their check-ins and then we analyze those and we surface the one that have the most data kind of. So we have a better dimension, are.

B

You building a separate model for each friend. Is that the yes.

C

B

Have a date 100 data points for each friend? Yes, so that's not a lot of training data.

A

Exactly and so even the transparent ones, the anomaly scores are really high, they're, like at point uh 3.4, but compared to the real anomalies which are 0.8.9. So the difference is noticeable.

B

You're saying it actually worked, yeah! Oh okay! I wasn't. I wasn't sure if this you're saying what the results you're going to show us- or this is actually the results. Yeah.

C

It's it's. The opacity.

B

D

C

Actually boring, so you could look at it as what kind of sorry for my all, my friends, but.

B

I mean, I guess, you'd, have to know your friends and know whether this is unusual or not.

C

Yes, but I mean we based it like on if they do something different than what they do. Usually, then it's something interesting but like our concepts and furthermore, is really like imagine, you're actually going to your social media and you have all these people that say, I'm brushing my teeth. I.

B

Don't want to see that I got, I just don't know if mama mama bear goes to microsoft. A lot I mean you know mama.

A

Bear conference, that's the name of the event. Oh I see so since it's not transparent. That's unusual.

B

Yes, I guess I have to trust you that that's the right score. I mean yeah.

C

Yeah it's the name of the event where they check. I.

B

Got it, I guess I wouldn't know you know, maybe the person checking in the fifth street distillery is a mormon, never drinks at all. That would be pretty unusual. You know right.

C

Now we we find that detection anomaly on like context that are news or checks or like even is really interesting, because you could actually um on the fly, look at a news for example and know or like teach it while you're doing it. This is a content that I like, and since we have an sdr on that content, the page could be like filtered on the fly, while I'm looking at new york time- and it's actually showing me only the thing that interests me.

D

C

It's it's really the the kind of bigger server, that's cool. I thought it was cool.

B

Yeah, I was just thinking.

D

While you were just saying that it might be interesting to if you took the text associated with the event or the place that the people had checked into, we thought about it. We.

A

Didn't have the time for it and.

D

Combine that with the set got a sept sdr for that and then use that to further refine. What the other thing worked.

B

In grack we have this, we take all these server metrics and server data, and we we we show them in order ranked by how knowledge they are. But this is a really cool thing. You just you're using the sort of opacity to to to sort of make it pop out at you right much cooler than what we did cool really nice.