This paper proposed Apriori algorithm for frequent itemset mining. As data accumulate in database, it would be valuable to make use of the data, or discovering regular patterns from them. For categorical data, if P(item B|item A) is high, there might exist a rule: A => B. To draw such conclusion, we have to observe the occurrence of A and B at least min_sup times/percentage. Therefore itemset counting becomes a first step of association rule mining. The Apriori algorithm is essentially a clever method to count frequent itemsets. The main intuition is that ""for an itemset to be frequent, all its subsets must be frequent"". This ""downward closure"" property can be exploited to prune the candidate subsets, quickly shrinking the search space. To reduce the passes of database scan (or disk access), AprioriTid and AprioriHybrid were proposed. Association rule mining is very useful in exploratory data analysis, because it is unsupervised, does not require any modeling or labeling. The only assumption is that ""highly frequent phenomena are potentially interesting"". However it has several limitations: 1) The definition of ""interestingness"" needs to be refined in specific applications, for example in text data ""it"" and ""is"" co-ocur a lot not is not interesting. 2) Some associations are mutual exclusive ""when A appears, B is likely to disappear"", or mutual information I(A;B) < 0. Such rules call for a fundamentally different formulation than ""frequency thresholding"", and may consequently render Apriori algorithms useless. 3) The data have to be categorical, not numerical. One has to put the numerical data into bins to make them categorical. An interesting observation: at the end of the paper, the author mentioned that IBM was planning to integrate the algorithm into their database products. The paper was written when data mining was still in its infancy. As the scope and functionality of data mining grows, nowadays it is more of a separate service provided by a few big database vendors. As the need of data mining grows, it is possible that functionality such as frequent pattern mining, clustering, regression/classification will become commonly used, standard part of database/filesystems.