Transfer learning for just-in-time design smells prediction using temporal convolutional networks

This paper investigates whether the adoption of a transfer learning approach that can be effective for just-in-time design smells prediction. The approach uses a variant of Temporal Convolutional Networks to predict design smells and a carefully selected fine-grained process and product metrics. The validation is performed on a dataset composed of three open-source systems and includes a comparison between transfer and direct learning. The hypothesis, which we want to verify, is that the proposed transfer learning approach is feasible to transfer the knowledge gained on mature systems to the system of interest to make reliable predictions even at the beginning of development when the available historical data is limited. The obtained results show that, when the class imbalance is high, the transfer learning provides F1-scores very close to the ones obtained by direct learning.