Example uses

[Peter Nguyen]

Hi,

Skydb seems to be an awesome technology, but I wonder if there are examples about how to use the data to answer questions like:

- What content is a particular user most likely want to read next?

- What products is a particular user most likely is interested in?

- When will an actor most likely trigger a certain event (over time)?

Besides the use cases, what about its performance and scaling, is there any guide on how to scale to multiple servers?

[Ben Johnson]

hey Peter-

Sky doesn't currently do machine learning techniques like collaborative filtering or predictive analytics yet. We're in the process of moving the query engine backend to use Julia (http://julialang.org/) which supports the statistical functions required for machine learning. I'll post more about that as it gets further along.

Scaling across multiple nodes is something else that's being added. It should be up and running in the next month or so. Once that's ready then I'll release the next version (v0.4.0) and do a major update to the docs. 

-- 
Ben

[Peter Nguyen]

Hi Ben,

which type of collaborative filtering algorithm do you think would be suitable for Skydb? It seems that the Slope One (http://en.wikipedia.org/wiki/Slope_One) is effective and fit well for real-time predictions. I also found implementation of it in different languages which should be no problem to convert to Go. Although some implementation use traditional relational databases such as Mysql, I was wonder if Skydb would be better for such applications?

[Ben Johnson]

Peter-

I haven't looked much at collaborative filtering algorithms yet. Slope One looks good from what it says on the Wikipedia page. I've personally been interested in figuring out how to predict next actions based on past behavior and also trying to cluster behavior.

The query engine is currently being moved over to LLVM to get better performance and to get around some LuaJIT limitations. Whichever algorithms end up being used will need to be implemented in C since LLVM can interop with that easily.

I'm currently holding off on implementing predictive algorithms until LLVM integration and distributed processing are done. Those are two major changes that need to be implemented before v0.4.0 is released.

-- 
Ben

[Peter Nguyen]

Hi Ben,

after further research for different collaborative filtering algorithms, it seems that Slope One is suitable for item based filtering which is only effective for rating systems. What I'm trying to find, and which is much more interesting from a ecommerce point of view, is an algorithm similar to the Amazon Recommender. It's probably the same thing as predictive algorithms which interest you. I've been working with a product recommendation software for some time now and although I don't know exactly the algorithm that it implements, I can probably give some pointers if it is of any help to you.