Generalized Sequential Patterns (AI Studio Core)

Description

This operator searches sequential patterns in a set of transactions. The ExampleSet must contain one attribute for the time and one attribute for the customer. Moreover, each transaction must be encoded as a single example. The time and customer attributes are specified through the time attribute and customer id parameters respectively. This pair of attributes is used for generating one sequence per customer containing every transaction ordered by the time of each transaction. The algorithm then searches sequential patterns in the form of: If a customer bought item 'a' and item 'c' in one transaction, he bought item 'b' in the next. This pattern is represented in this form: <a, c> then <b>. The minimal support describes how many customer must support such a pattern for regarding it as frequent. Infrequent patterns will be dropped. A customer supports such a pattern, if there are some parts of his sequence that includes that pattern. The above pattern would be supported by a customer, for example, with transactions: <s, g> then <a, s, c> then <b> then <f, h>. The minimum support criteria is specified through the min support parameter.

The min gap, max gap and window size parameters determine how transactions are handled. For example, if the above customer forgot to buy item 'c', and had to return 5 minutes later to buy it, then his transactions would look like: <s, g> then <a, s> then <c> then <b> then <f, h>. This would not support the pattern <a, c> then <b>. To avoid this problem, the window size determines, how long a subsequent transaction is treated as the same transaction. If the window size is larger than 5 minutes then <c> would be treated as being part of the second transaction and hence this customer would support the above pattern. The max gap parameter causes a customers sequence not to support a pattern, if the transactions containing this pattern are too widely separated in time. The min gap parameter does the same if they are too near.

This technique overcomes some crucial drawbacks of existing mining methods, for example:

• absence of time constraints: This drawback is overcome by the min gap and max gap parameters.
• rigid definition of a transaction: This drawback is overcome by the sliding time window.

Please note that all attributes (except customer and time attributes) of the given ExampleSet should be binominal, i.e. nominal attributes with only two possible values. If your ExampleSet does not satisfy this condition, you may use appropriate preprocessing operators to transform it into the required form. The discretization operators can be used for changing the value of numerical attributes to nominal attributes. Then the Nominal to Binominal operator can be used for transforming nominal attributes into binominal attributes.

Please note that the sequential patterns are mined for the positive entries in your ExampleSet, i.e. for those nominal values which are defined as positive in your ExampleSet. If your data does not specify the positive entries correctly, you may set them using the positive value parameter. This only works if all your attributes contain this value.