The　presence　of　outliers　in　the　training　data　affects　the　accuracy　of　the　constructed　model.　To　cope　with　the　outlier　interference　in　the　model　construction　process,　some　robust　methods　have　been　proposed　on　the　basis　of　the　nonparametric　method,　Gaussian　process　regression　(GPR),　without　eliminating　the　outliers　previously.　However,　the　high　complexity　of　these　robust　GPR　methods　makes　them　unable　to　cope　with　situations　where　the　amount　of　data　is　too　large.　In　this　article,　we　analyze　the　impact　of　outliers　on　model　construction　in　the　setting　of　big　data　and　propose　a　robust　version　based　on　the　sparse　GPR.　Empirical　evaluations　conducted　on　two　publicly　available　datasets,　as　well　as　a　nitrogen　oxides　soft　sensor　designed　for　a　physical　diesel　engine　whose　data　exist　outliers　that　are　difficult　to　distinguish　from　normal　data,　provide　compelling　evidence　to　support　the　notion　that　the　proposed　method　leads　to　significant　enhancements　in　performance.