We present the novel, semi-automated clustering tool ASPECT for analysing voluminous archives of spectra. The heart of the program is a neural network in form of Kohonen's self-organizing map. The resulting map is designed as an icon map suitable for the inspection by eye. The visual analysis is supported by the option to blend in individual object properties such as redshift, apparent magnitude, or signal-to-noise ratio. In addition, the package provides several tools for the selection of special spectral types, e.g. local difference maps which reflect the deviations of all spectra from one given input spectrum (real or artificial). ASPECT is able to produce a two-dimensional topological map of a huge number of spectra. The software package enables the user to browse and navigate through a huge data pool and helps him to gain an insight into underlying relationships between the spectra and other physical properties and to get the big picture of the entire data set. We demonstrate the capability of ASPECT by clustering the entire data pool of 0.6 million spectra from the Data Release 4 of the Sloan Digital Sky Survey (SDSS). To illustrate the results regarding quality and completeness we track objects from existing catalogues of quasars and carbon stars, respectively, and connect the SDSS spectra with morphological information from the GalaxyZoo project.

We exploit the spectral archive of the Sloan Digital Sky Survey (SDSS) Data Release 7 to select unusual quasar spectra. The selection method is based on a combination of the power of self-organising maps and the visual inspection of a huge number of spectra. Self-organising maps were applied to nearly 10^5 spectra classified as quasars by the SDSS pipeline. Particular attention was paid to minimise possible contamination by rare peculiar stellar spectral types. We present a catalogue of 1005 quasars with unusual spectra. This large sample provides a useful resource for both studying properties and relations of/between different types of unusual quasars and selecting particularly interesting objects. The spectra are grouped into six types. All these types turn out to be on average more luminous than comparison samples of normal quasars after a statistical correction is made for intrinsic reddening. Both the unusual broad absorption line (BAL) quasars and the strong iron emitters have significantly lower radio luminosities than normal quasars. We also confirm that strong BALs avoid the most radio-luminous quasars. Finally, we create a sample of quasars similar to the two "mysterious" objects discovered by Hall et al. (2002) and briefly discuss the quasar properties and possible explanations of their highly peculiar spectra. (Abstract modified to match the arXiv format)