Well, IMO big data tensor math is Mahout’s strongest point and GPUs on immediately on the roadmap.
On May 21, 2016, at 7:30 AM, Steven NASa <[email protected]> wrote: Hi Pat, Thank you for your reply, I fully understand that core algorithms and data are 2 different part of the system, this is why we have 2 major idea: "Big data" and "Machine Learning". My requirements of Recommenders are just like what Amazon does: Item-based, but the number of items and users is very big, so there comes to a very huge matrix. So I am still learning using Mahout to make the matrix computing on a distributed system. After I am familiar with Mahout, I think I can have some works on GPU acceleration for Matrix computing and some other mathematical optimization. About the data prep, I think we can define an abstraction of conventions in data prep, data ingestion, and serving components. Users can following some conventions to feed data to Mahout. Steven NASa 2016/05/21 2016-05-21 22:06 GMT+08:00 Pat Ferrel <[email protected]>: > Hi Stephen, > > We have implemented SVD, ALS, and CCO for recommender, but these are only > core algorithms, not really recommenders as Mahout has done in the past. > The reason for this is that there are data prep, data ingestion, and > serving components that, in a modern system, must be supplied also. So far > Mahout has stayed aways from actually including servers, either for input > of output. > > That said there is plenty of room for algorithm development in Mahout. I > worked on the CCO algorithm, which uses PredictionIO (proposed for the > Apache Incubator) to supply the serving components. > > Someone with your experience in real-life use of recommenders is certainly > welcome. > > What type of project did you have in mind? > > > On May 20, 2016, at 10:00 AM, Suneel Marthi <[email protected]> wrote: > > Welcome to the project Steven!! > > On Fri, May 20, 2016 at 10:07 AM, Steven NASa <[email protected]> wrote: > >> Hi Folk & Masters, >> >> My name is *NASa*. I am now working for an e-commerce B2C company in > China, >> dealing with Transaction Process development in C++ & Java on Linux >> environment. >> >> As you know, *Recommender System* is quite valuable and important to an >> e-commerce online shopping website like Amazon. I was told and required > to >> design and implement a Recommender System which can bring some value to > my >> Company. Our System is based on C++ codes. So I was searching for an > robust >> Machine Learning framework in C++ which can help me to easily implement a >> Recommender System. I did not find any one which can satisfy my >> requirements, but only some C++ math libraries. >> >> Our system is based on an internal distributed frameworks like RPC and DB >> access on Linux environment based on C++ programming language. But I find >> it is really inconvenient to implement a Recommender System in C++ from >> zero without distributed computing library supporting, like >> implementing *Collaborative >> Filtering* with SVD in a distributed computing way. So I am trying to > find >> a framework/library with is designed based on Distributed-System. There I >> come to *Mahout*. >> >> I wish I can build a library that can help people easily and quickly > build >> up a Recommender System based on Distributed System and also use the >> Machine Learning Algorithms in distributed way. Apache has many amazing >> projects which can help people to build up robust distributed system >> easily. So I am moving to using “Java” environment. >> >> I am new to *Mahout* and *Hadoop*, *Spark*, *Scala* and I learned Andrew >> Ng’s “Machine Learning” from Coursera >> <https://www.coursera.org/learn/machine-learning/home/welcome>. So I > have >> the basic knowledge of Machine Learning, and now I am keeping forward to >> *Deep >> Learning* and *Convex Optimization*, some other Mathematical Optimization >> implementation. I am now still learning and getting famiIiar with > Mahout. I >> hope I can contribute some codes to Mahout in the early future with >> learning by coding and coding by learning. >> NASa 2016/05/20 >> >> > >
