s****h 发帖数: 3979 | 1 two questions:
1.
For recommendation engine based on collaborative filtering, the result of
ALSWR in Mahout would be very similar to result of SVD in MLlib of spark,
right?
As the SVD with spark + MLlib performance is very good, can we forget about
ALSWR in Mahout?
2.
How to evaluate SVD?
My understanding: for a known user/item matrix M, we remove some of the
known user/item pair and get new matrix M1, then do the SVD for M1 and get
the reconstructed matrix M2. Comparing removed user/item pairs between M and
M2, we can evaluate SVD.
This would make SVD evaluation very slow, as you might want to generate lots
of M1 matrix and do the SVD to get M2. And we know SVD is computationally
intensive here.
How people deal with the evaluation of ALSWR with mahout, the evaluate
process would be much longer, right?
thanks. | l******n 发帖数: 9344 | 2 Depend on your data, I do not think you can make a general conclusion.
about
【在 s****h 的大作中提到】 : two questions: : 1. : For recommendation engine based on collaborative filtering, the result of : ALSWR in Mahout would be very similar to result of SVD in MLlib of spark, : right? : As the SVD with spark + MLlib performance is very good, can we forget about : ALSWR in Mahout? : 2. : How to evaluate SVD? : My understanding: for a known user/item matrix M, we remove some of the
| s****h 发帖数: 3979 | 3 SVD对binary 数据,sparse数据不友好,所以用ALS.
其实我真正想知道的是:
不考虑计算复杂度,计算时间,并行性等等,仅在accuracy上相比较,SVD是否优于
ALSWR,毕竟SVD有个singular value,而且从空间转换的角度,更有说服力。
那么如果算法A比SVD好,是否就肯定比ALSWR好? |
|