Given　that　citations　are　not　equally　important,　various　techniques　have　been　presented　to　identify　important　citations　on　the　basis　of　supervised　machine　learning　models.　However,　only　a　small　volume　of　instances　have　been　annotated　manually　with　the　labels.　To　make　full　use　of　unlabeled　instances　and　promote　the　identification　performance,　the　semi-supervised　self-training　technique　is　utilized　here　to　identify　important　citations　in　this　work.　After　six　groups　of　features　are　engineered,　the　SVM　and　RF　models　are　chosen　as　the　base　classifiers　for　self-training　strategy.　Then　two　experiments　based　on　two　different　types　of　datasets　are　conducted.　The　experiment　on　the　expert-labeled　dataset　from　one　single　discipline　shows　that　the　semi-supervised　versions　of　SVM　and　RF　models　significantly　improve　the　performance　of　the　conventional　supervised　versions　when　unannotated　samples　under　75%　and　95%　confidence　level　are　rejoined　to　the　training　set,　respectively.　The　AUC-PR　and　AUC-ROC　of　SVM　model　are　0.8102　and　0.9622,　and　those　of　RF　model　reach　0.9248　and　0.9841,　which　outperform　their　counterparts　and　the　benchmark　methods　in　the　literature.　This　demonstrates　the　effectiveness　of　our　semi-supervised　self-training　strategy　for　important　citation　identification.　Another　experiment　on　the　author-labeled　dataset　from　multiple　disciplines,　semi-supervised　learning　models　can　perform　better　than　their　supervised　learning　counterparts　in　term　of　AUC-PR　when　the　ratio　of　labeled　instances　is　less　than　20%.　Compared　to　our　first　experiment,　insufficient　amount　of　instances　from　each　discipline　in　our　second　experiment　enables　the　performance　of　the　models　to　be　unsatisfactory.