In　the　context　of　Cyber　Physical　Social　Intelligence　(CPSI),　efficiently　training　and　inferring　from　samples　with　limited　labels　poses　critical　challenges　due　to　the　scarcity　and　high　cost　of　label　acquisition　for　big　data.　The　aim　is　to　attain　high　accuracy　at　minimal　cost,　thereby　enhancing　adaptation　to　the　CPSI　scenario.　To　tackle　the　challenges　in　CPSI,　we　present　a　multi-level　feature　learning　framework　for　semi-supervised　classification　tasks.　Initially,　we　employ　a　mapping　operation　for　each　view,　extracting　view-specific　features　with　a　feature-level　reconstruction　loss.　These　features　are　fused　to　obtain　a　shared　feature.　Simultaneously,　a　learnable　graph　neural　network　captures　global　topology　using　a　graph　structure-level　reconstruction　loss.　Subsequently,　a　scalable　graph　convolution　fusion　module　combines　these　features.　Our　evaluations　on　eight　benchmark　datasets　show　promising　results　and　empirically　prove　the　effectiveness　of　our　approach,　surpassing　eight　state-of-the-art　methods　in　multi-view　semi-supervised　classification　tasks.　©　2018　Tsinghua　University　Press.