In the present study, we compared the effects of temporal compression (averaging across multiple scans) and space selection (i.e. selection of "regions of interest" from the whole brain) on single-subject and multi-subject classification of fMRI data using the support vector machine (SVM). Our aim was to investigate various data transformations that could be applied before training the SVM to retain task discriminatory variance while suppressing irrelevant components of variance. The data were acquired during a blocked experiment design: viewing unpleasant (Class 1), neutral (Class 2) and pleasant pictures (Class 3). In the multi-subject level analysis, we used a "leave-one-subject-out" approach, i.e. in each iteration, we trained the SVM using data from all but one subject and tested its performance in predicting the class label of the this last subject's data. In the single-subject level analysis, we used a "leave-one-block-out" approach, i.e. for each subject, we selected randomly one block per condition to be the test block and trained the SVM using data from the remaining blocks. Our results showed that in a single-subject level both temporal compression and space selection improved the SVM accuracy. However, in a multi-subject level, the temporal compression improved the performance of the SVM, but the space selection had no effect on the classification accuracy.