One prominent task in graphs is property prediction, where a property of some of the graph's nodes is known and used to make predictions for those individuals for which this property is unknown. In this paper, we look at topic prediction for papers organized in a so-called co-authorship graph (CAG) where the individuals are scientific papers with links between them if they share some author. A CAG tends to have a large number of cliques, each formed by all the papers published by the same author. Thus, topic prediction in a CAG tends to be computationally expensive. We investigate in how far we can reduce this complexity without sacrificing the prediction quality by reducing the number of links in the CAG based on the papers' publication dates. We apply an inexpensive iterative neighbourhood's majority vote based algorithm to predict unknown topics based on the papers with known topics and the CAG's link structure. For three data sets, we evaluate our algorithm in terms of classification accuracy and computational time on both the full graph G and subgraphs of it. On substantially smaller subgraphs of G, our algorithm obtains classification accuracies that are similar to the results obtained on G, while achieving a reduction in execution time of up to one order of magnitude.