For multi-homed networks, inter-domain traffic engineering (TE) consists in selecting the best path via available transit providers, so that the transmission quality is improved in front of network events, such as congestion and fail-over. In practice, this choice bases on end-to-end (e2e) measurements toward destination networks. These measurements, especially Round-Trip Time (RTT), are expected to offer an faithful view on inter-domain path properties. Hosts in destination networks with open ports are deliberately discovered for active measurement. RTT traces so obtained can be influenced by host-local factors that are not relevant to inter-domain routing and eventually mislead route decisions. We data-mined the RTT time-series between two ASes with unsupervised learning method - clustering, on a set of statistic features. Achieved results showed that our method was capable of improving data quality, by excluding less reliable traces. Moreover, we considered traceroute measurements. Early results suggested that most variations of e2e delay actually occured in access networks. We thus believe that the proposed scheme can improve the accuracy and stability of the route selection for multi-homed networks.