In this work, we propose a temporal aggregation scheme for sentiments expressed in social networks. The proposed method discounts for the bias caused in aggregation due to classification errors while providing confidence intervals. A computationally efficient prediction and interpolation scheme of temporal progression is discussed that accounts for the heteroscedastic nature of noise. To this end, we use a heteroscedastic gaussian process model. To test the efficacy of our proposed method, we use tweets about Donald Trump obtained for a period of twelve hours. The results are generalized using six state of art classification schemes for predicting sentiments. Our method shows improvement in R2 statistics with better coverage under proposed uncertainty for all the six classification schemes. Finally, the results of variational heteroscedastic gaussian process (VHGP) regression are discussed and the normalized mean square error with negative log-probabilty density of the prediction are reported. It is further shown that the volatility of opinion tracking in social network data streams is better captured with a heteroscedastic noise model.