Advanced Clustering Methods for Mining

Heroin and cocaine gas chromatography data are analyzed using several clustering techniques. A database with clusters confirmed by police investigation is used to assess the potential of the analysis of the chemical signature of these drugs in the investigation process. Results are compared to standard methods in the field of chemical drug profiling and show that conventional approaches miss the inherent structure in the data, which is highlighted by methods such as spectral clustering and its variants. Also, an approach based on genetic programming is presented in order to tune the affinity matrix of the spectral clustering algorithm. Results indicate that all algorithms show a quite different behavior on the two datasets, but in both cases, the data exhibits a level of clustering, since there is at least one type of clustering algorithm that performs significantly better than chance. This confirms the relevancy of using chemical drugs databases in the process of understanding the illicit drugs market, as information regarding drug trafficking networks can likely be extracted from the chemical composition of drugs.