Many powerful methods for intelligent data analysis have become available in the fields of machine learning and data mining. However, almost all of these methods are based on the assumption that the objects under consideration are represented in terms of feature vectors, or collections of attribute values. In the present paper we argue that symbolic representations, such as strings, trees or graphs, have a representational power that is significantly higher than the representational power of feature vectors. On the other hand, operations on these data structure that are typically needed in data mining and machine learning are more involved than their counterparts on feature vectors. However, recent progress in graph matching and related areas has led to many new practical methods that seem to be very promising for a wide range of applications.