In transcriptome (RNA-Seq) libraries, unmapped reads are reads that fail to map to known exons. More often than not, they represent genomic DNA. RSeQC has tools to help identify genomic (also sometimes called intergenic) sequences.
For the most part, the percentage of unmapped reads just provides you with QC information, like how much gDNA you purified along with your mRNA. It's unlikely to be very useful beyond that, but it really comes down to what you're looking for in those sequences.
If you're doing DNA-Seq, then unmapped reads generally represent reads that failed to map unambiguously to known sequences. This will depend on a threshold (usually provided by the user) for alignment stringency.