Sometimes, the quality of the output is simply assessed by eyeballing it. Looking at the examples provided below, it’s clear that the expected user ratings are 2 and 3 (please refer to the empty data cells).
The examples are simplified – I’ve used much larger datasets and still harvested the same results.

Table 1

Users Movie 1 Movie 2 Movie 3
1 – 3 4
2 2 3 4
3 2 3 4

Table 2

Users Movie 1 Movie 2 Movie 3
1 – 3 3
2 3 3 3
3 3 3 3

The ALS model delivers exactly these ratings, but the SVD model fails miserably. Does anyone have any idea why that is? Why would we get results other than 2 and 3?