Recently, Twitter search has drawn much attention of researchers in social networks. Although rich features of Twitter can be incorporated into rank learning, the retrieval effectiveness can be hurt by the lack of training data. Transductive learning, as a common semi-supervised learning method, has been playing an import role in dealing with the lacking of training data. Due to the fact that noise is generated during the iterative process of transductive learning, a clustering-based transductive method is proposed. There exist two important parameters in the clustering-based transductive approach, namely the threshold of clustering and the number of the documents that will be clustered. This paper extends the method by utilizing a different clustering algorithm. As shown by extensive experiments on the standard TREC Tweets11 collection, both of the two parameters have an effect on the retrieval effectiveness. Furthermore, the robustness of the clustering-based transduction approach on different query sets is also studied. Finally, the paper proposes an adaptive clustering-based approach by introducing a so called cluster coherence as quality controller. The experimental results show that the robustness of the proposed method is better.