The stronger the impact, the rarer the situation tends to be (the Maxwell-Boltzmann distribution formalised this). This rarity makes the situation less predictable because we cannot learn trustworthy causes of the situation. This problem has been disturbing predictive data mining. You can predict the next event by following a frequent episode (partial sequence) found in the eventsequence of the past data, as long as no unusual change occurs (Mannila, 1995). Association rules learned from the data of baskets ± each containing a data set of items ± can also be used to predict items that will occur in a new basket if they were somehow frequently seen in the past baskets (Agrawal, 1993). For example, if you frequently buy blue cheese in combination with wine, then the prediction that you will want blue cheese if you buy wine will help the owner of the liquor store recommend blue cheese to you, unless you find something else to go with the wine. However, if there is an unusual (new) food product in the liquor store one day, this can cause a significant change in your choice if the