In computational homogenization approaches, data-driven methods entail advantages due to their ability to capture complex behavior without assuming a specific material model. Within this domain, constitutive model-based and model-free data-driven methods are distinguished. The former employ artificial neural networks as models to approximate a constitutive relation, whereas the latter directly incorporate stress–strain data in the analysis. Neural network-based constitutive descriptions are one of the most widely used data-driven approaches in computational mechanics. In contrast, distance-minimizing data-driven computational mechanics enables substituting the material modeling step entirely by iteratively obtaining a physically consistent solution close to the material behavior represented by the data. The maximum entropy data-driven solver is a generalization of this method, providing increased robustness concerning outliers in the underlying data set. Additionally, a tensor voting enhancement based on incorporating locally linear tangent spaces enables interpolating in regions of sparse sampling. In this contribution, a comparison of neural network-based constitutive models and data-driven computational mechanics is made. General differences between machine learning, distance minimizing, and entropy maximizing data-driven methods are explored. These include the pre-processing of data and the required computational effort for optimization as well as evaluation. Numerical examples with synthetically generated datasets obtained by numerical material tests are employed to demonstrate the capabilities of the investigated methods. An anisotropic nonlinear elastic constitutive law is chosen for the investigation. The resulting constitutive representations are then applied in structural simulations. Thereby, differences in the solution procedure as well as use-case accuracy of the methods are investigated.