In this work, the cycles in reversible non-uniform elementary cellular automata (CAs) are used as a tool for data classification. For the non-uniform lattice, based on the individual information propagation of the rule for each cell, a graph theoretic network flow approach is applied to find the overall information propagation in the lattice. Considering this overall information propagation, a greedy selection strategy is devised that drastically reduces the possible number of effective rules suitable for data classification. To apply these CAs for classification, configurations within the cycles of the appropriate cellular automaton (CA) are assigned class labels to group the cycles into distinct classes. Finally, through experimentation and analysis, the performance of cellular automaton in classifying data based on accuracy and execution time is evaluated and shown to be comparable to the traditional classification algorithms.