Heterogeneous　information　networks　(HINs)　composed　of　multiple　types　of　nodes　and　links,　play　increasingly　important　roles　in　real　life　applications.　Classification　of　the　related　data　is　an　essential　work　in　network　analysis.　Existing　methods　can　effectively　solve　these　classification　tasks　when　they　are　applied　to　homogeneous　information　networks　and　simple　data,　but　not　for　the　noisy　and　sparse　data.　To　address　the　problem,　we　propose　Stacked　Denoising　Auto　Encoder　(SDAE)　with　sparse　factors　to　learn　features　of　nodes　in　heterogeneous　networks.　In　particular,　sparse　factors　are　added　in　each　hidden　layer　of　the　proposed　stacked　denoising　auto-encoder　to　efficiently　extract　features　from　noisy　and　sparse　data.　Moreover,　a　relax　strategy　is　employed　to　construct　class　hierarchy　with　high-quality　based.　Finally,　nodes　of　the　heterogeneous　information　network　can　be　classified.　Our　proposed　framework　Relax　strategy　on　Stacked　Denoising　Auto　Encoder　with　sparse　factors　(RSDAEf)　comparison　with　several　existing　methods　clearly　indicates　RSDAEf　outperforms　the　existing　methods　and　achieves　a　classification　precision　of　88.3%　on　DBLP　dataset.