First, about interpreting these confusion matrices: the sum of every row is 1, which implies that every value is a conditional probability p( predicted label | true label ), i.e. the probability of a given true label to be a particular predicted label. Example: the top left cell in both matrices is 0.01, which means that when the true label is 5 the probability that the system predicts label 1 is 1%.
The two confusion matrices show the results of the predictions by two different systems. This means that in general there is not one which is correct and the other which makes errors, just two different ways to predict labels which might lead to different types of errors.
- The diagonal shows the True Positive cases, i.e. cases where the predicted label is the same as the true label (this is important since these are the "correct" cases). The probabilities show very little difference between the two systems on the diagonal, the top one being slightly better for labels 3 and 4 and the bottom one slightly better for label 5.
- In general it's important not to look only at the diagonal especially for ordinal values, because if the predicted value is far from the true value it's a more serious error than if it's close to it. Example: if the true value is 4 it's better for a system to predict 3 than 1 (both are errors but the latter is worse). However here again there is very little difference between the two systems, they appear to have a very similar behaviour.
In order to quantitatively summarize and compare the performance of the two systems, confusion matrices are too complex. Typically one would use an appropriate evaluation measure, for instance micro or macro f-score (classification evaluation) or Mean Absolute Error (regression evaluation).
