implicit_als.jl

# A Simple Julia Version of the algorithm described in the paper Collaborative Filtering for Implicit Feedback Datasets 
# http://yifanhu.net/PUB/cf.pdf

# This was mainly an experiment for me in learning Julia, unfortunately the code below ended up
# being about the same speed as my pure python version here https://github.com/benfred/implicit
# and much much slower than the Cython version I included there.

# (fwiw The slowness is probably because I don't understand the language as well as I do Python)

function alternatingLeastSquares(input, factors, regularization)
    rows, cols = size(input)

    # initialize factors randomly
    colFactors = randn(factors, cols)
    rowFactors = randn(factors, rows)

    # transpose once for fast access
    inputT = input'

    for iteration = 1: 20
        println(iteration)
        leastSquares(input, colFactors, rowFactors, regularization)
        leastSquares(inputT, rowFactors, colFactors, regularization)
    end

    rowFactors, colFactors
end


function leastSquares(Cui, X, Y, regularization)
    factors, users = size(X)
    rows = rowvals(Cui)
    vals = nonzeros(Cui)

    YtY = Y * Y'

    for u = 1: users
        A = YtY + regularization * eye(factors)
        b = zeros(factors)

        for i in nzrange(Cui, u)
            factor = Y[:, rows[i]]
            b += (vals[i] + 1) * factor
            A += factor * vals[i] * factor'
        end
        X[:, u] = A\b
    end
end

using MatrixMarket
m = MatrixMarket.mmread("small.mtx")
@time alternatingLeastSquares(m, 50, 1)