Contrasting offline and online results when evaluating recommendation algorithms