Current subject studies and data-driven approaches in lighting research often use manually selected light spectra, which usually exhibit a large bias due to the applied selection criteria. This paper, therefore, presents a novel approach to minimize this bias by using a data-driven framework for selecting the most diverse candidates from a given larger set of possible light spectra. The spectral information per wavelength is first reduced by applying a convolutional autoencoder. The relevant features are then selected based on Laplacian Scores and transformed to a two-dimensional embedded space for subsequent clustering. The low dimensional embedding, from which the required diversity follows, is done with respect to the locality of the features. In a second step, photometric parameters are considered and a second clustering is performed. As a result of this algorithmic pipeline, the most diverse selection of light spectra complying with a given set of relevant photometric parameters can be extracted and used for further experiments or applications.