Semi-supervised　learning　has　always　been　a　hot　topic　in　machine　learning.　It　uses　a　large　number　of　unlabeled　data　to　improve　the　performance　of　the　model.　This　paper　combines　the　co　training　strategy　and　random　forest　to　propose　a　novel　semi-supervised　regression　algorithm:　semi　supervised　random　forest　regression　model　based　on　co-training　and　grouping　with　information　entropy　(E-CoGRF),　and　applies　it　to　the　evaluation　of　depression　symptoms　severity.　The　algorithm　inherits　the　ensemble　characteristics　of　random　forest,　and　combines　well　with　co-training.　In　order　to　balance　the　accuracy　and　diversity　of　co-training　random　forests,　the　algorithm　proposes　a　grouping　strategy　to　decision　trees.　Moreover,　the　information　entropy　is　used　to　measure　the　confidence,　which　avoids　unnecessary　repeated　training　and　improves　the　efficiency　of　the　model.　In　the　practical　application　of　evaluation　of　depression　symptoms　severity,　we　collect　cognitive　behavioral　data　of　emotional　conflict　based　on　the　depressive　affective　disorder.　And　on　this　basis,　feature　construction　and　normalization　preprocessing　are　carried　out.　Finally,　the　test　is　conducted　on　35　labeled　and　80　unlabeled　depression　patients.　The　result　shows　that　the　proposed　algorithm　obtains　MAE　(Mean　Absolute　Error)　=　3.63　and　RMSE　(Root　Mean　Squared　Error)　=　4.50,　which　is　better　than　other　semi-supervised　regression　algorithms.　The　proposed　method　effectively　solves　the　modeling　difficulties　caused　by　insufficient　labeled　samples,　and　has　important　reference　value　for　the　diagnosis　of　depression　symptoms　severity.