TW201933142A - Frequent time-gap sequential pattern mining system and method - Google Patents

Abstract

A frequent time-gap sequential pattern mining system and method are disclosed. The system includes a time-gap sequence database, a pseudo projection database construction module, a candidate pattern generation module and a candidate pattern pruning module. The time-gap sequence database stores time-gap sequences. The pseudo projection database construction module constructs a pseudo projection database of candidate patterns according to the time-gap sequence database and calculates support of the candidate patterns. The candidate pattern generation module searches all L-length candidate patterns from the pseudo projection database to calculate all L+1-length candidate patterns, and updates support of the L+1 length candidate patterns to gather into a set. The candidate pattern pruning module applies a pruning algorithm to prune or eliminate the L+1-length candidate patterns in the set, thereby the mining effectiveness would be improved.

Description

Frequent time interval sequence style exploration system and method

The present invention relates to a technique for time interval sequence pattern exploration, and more particularly to a frequent time interval sequence pattern exploration system and method.

There are many materials in life that have chronological characteristics, so analyzing sequential data is an important issue, and it is widely used both scientifically and commercially. Projects in the sequence represent different content according to different applications or situations, and are usually of interest to people. Whenever an item occurs, it is accompanied by an occurrence time point, and the records of multiple events are formed into a specific sequence according to their chronological order.

For example, in the e-commerce website, each purchase behavior of the customer is regarded as a project, and it can be known that the customer first purchases the product A and the product B, and separately purchases the sequence of the product C and the product D on different days, and the sequence can be {( A, B), C, D} are indicated. If the interval between occurrences of each event is also recorded in the sequence, this is called a time interval sequence.

Therefore, following the above example, the time sequence information of the sequence {(A, B), C, D} is restored to obtain a new sequence {A, 0 B, 10 C, 30 D}, and the new sequence represents the same purchase of the customer. Product A and Product B, purchase product C after ten days, Product D is purchased only one month later, where the number can represent any time unit.

Therefore, if many customers have a time interval sequence of such experiences, the time correlation of products A, B and C is displayed, which can be used as a reference for when to properly recommend product C to customers who have purchased products A and B. Obviously, exploring the frequent time interval sequence pattern from the time interval sequence database will help various application areas explore the temporal correlation between projects and provide more detailed meanings to assist users in making effective decisions.

However, the exploration of all frequent patterns may result in low efficiency in exploration efficiency or interpretation of commercial meanings. Both the maximum frequent pattern and the closed frequent pattern are often used in the data exploration field to avoid excessive candidate patterns or frequent patterns. If a style does not have any supersets called the maximum frequent style, and if a style does not have any super styles with the same support, it is called closed frequent style. However, the novel nature of the time interval information in the time interval sequence pattern, that is, the sequence pattern with a shorter time interval provides a more complete and detailed commercial meaning than the sequence pattern with a longer time interval, and the traditional style trimming is applied. The method is not appropriate.

In summary, most of the conventional techniques only consider the sequence order phenomenon, and apply it to different fields of data to generate different analysis methods, but ignore the time interval information and characteristics. Moreover, the memory required by the prior art technology is relatively large, the exploration efficiency is low, and important and detailed information may be lost.

Therefore, how to solve the above problems of the prior art has become one of the major problems of those skilled in the art.

The present invention provides a frequent time interval sequence pattern mining system and method that can be used to solve one or more of the above-discussed problems.

The frequent time interval sequence style exploration system of the present invention comprises: a time interval sequence database, a virtual projection database creation module, a candidate pattern generation module and a candidate style clipping module. The time interval sequence database is used to store time interval sequences. The virtual projection database building module establishes a virtual projection database of the candidate style according to the time interval sequence database and calculates the support degree of the candidate pattern. The candidate pattern generation module searches for all L length candidate patterns from the virtual projection database, and calculates all L+1 length candidate patterns according to the L length candidate patterns, and then updates the L+1 length candidates. The support of the style, the candidate patterns of the L+1 lengths are aggregated into a set, where L is a positive integer equal to or greater than 1. The candidate style pruning module trims or deletes the candidate patterns of the L+1 lengths in the set by a pruning algorithm.

The frequent time interval sequence pattern exploration method of the present invention comprises: searching for a frequent pattern of all L lengths from a time interval sequence database to be aggregated into a set, wherein L is a positive integer equal to or greater than 1; all frequent patterns in the set are Establishing a virtual projection database; searching for all L+1 length candidate patterns from the virtual projection database to calculate the support of the L+1 length candidate patterns; updating the L+1 length candidate patterns Support for updating the set; and trimming or deleting the candidate patterns of the L+1 lengths within the set with a pruning algorithm.

The above described features and advantages of the invention will be apparent from the description and appended claims. In the following description The additional features and advantages of the present invention are set forth in part in the description of the invention. The features and advantages of the present invention are realized and attained by the <RTIgt; It is to be understood that both the foregoing general description

1‧‧‧Frequency Interval Sequence Pattern Exploration System

10‧‧‧Time interval sequence database

11‧‧‧time interval sequence

20‧‧‧Virtual projection database building module

21‧‧‧Virtual Projection Database

30‧‧‧Candidate style generation module

40‧‧‧Candidate style trimming module

41‧‧‧First trimming algorithm

42‧‧‧Second trimming algorithm

50‧‧‧ Frequent sequence style

S1 to S6‧‧‧ steps

1 is a schematic structural diagram of a frequent time interval sequence pattern exploration system in the present invention; FIG. 2 is a schematic flow chart of a frequent time interval sequence pattern exploration method in the present invention; FIG. 3A is a time interval of the present invention; Schematic diagram of the sequence database; FIG. 3B illustrates the support of the items in the U _L=1 set in the present invention; FIG. 3C illustrates the virtual projection database prefixed with the pattern <A> and its support in the present invention; The figure shows a 3-pattern virtual projection database prefixed with the pattern <A> and its support degree in the present invention; FIG. 3E shows a virtual projection database with frequent 2-patterns in the present invention and its support degree; FIG. 3F shows The virtual projection database of 3-pattern is frequently used in the present invention and its support; the 3G diagram shows the virtual projection database of the frequent 4-pattern in the present invention and its support degree; FIG. 4A shows the step S3 in the second diagram of the present invention. The virtual code; and FIG. 4B illustrates the virtual code of step S4 in FIG. 2 of the present invention.

The embodiments of the present invention are described in the following specific embodiments, and those skilled in the art can easily understand other advantages and functions of the present invention by the disclosure of the present disclosure, and can also be implemented by other different embodiments. Or application.

Different from the maximum frequent pattern or closed frequent pattern of the prior art, if a pattern does not have any pattern having the same sequence and any time interval is smaller, it is called a minimum frequent time interval sequence pattern, and the present invention proposes A groundbreaking minimum frequent interval sequence style exploration algorithm to exploit this style. Moreover, the present invention will present a generalized representation to represent a time interval sequence, which can simultaneously record the order of the items and the length of the time interval, and proposes an efficient exploration algorithm to search all the frequent items from the time interval sequence database. Time interval sequence style.

The following time interval sequences are simply referred to as patterns, and the basic definitions are explained first. E={e1,e2,...,ez} is a collection of project types, z .

Definition 1: Sequence Database

A time interval sequence database TSDB={s ₁ ,s ₂ ,...,s _N }, a time interval sequence , t _{i-1, i} is the time interval between the project e _i-1 and the project e _i , 1 I-1<i j,i,j , e _i E. For example, s={A, 2 B, 3 C} represents B after two time units after A occurs, and C occurs after three time units after B occurs.

Definition 2: Style length

If there are j items in the interval sequence s, the length of s is L=j or j-pattern. Taking definition 1 as an example, the length of the sequence s is 3 or 3-pattern.

Definition 3: Support and suffix

given versus And ,1 m 1< m 2 j Mf g , then s ₁ supports s ₂ , and the suffix If mf = g , ; s ₁ is the prefix of s ₂ . For example, the prefix s ₁ =<A,5 C> and s ₂ =<A,2 B,2 C,1 D,0 E>, then s ₁ supports s ₂ , .

Definition 4: suffix interval accumulation pattern and space

Following the above definition 3, given the suffix interval cumulative sequence pattern , or <t _mf,mf+1 e _mf+1 ,at _mf,mf+2 e _mf+2 ,...,at _mf,g e _g 〉 indicates that its space is a paired set of all time intervals and items {at _mf+1,mf+2 e _mf+2 },...,{at _g-1,g e _g }. Continue with the above example, <1 D,1 E>, .

Definition 5: Virtual Projection Database

A virtual projection database of style s pdb(s)={{SID ₁ , Index ₁ },...,{SID _M ,Index _M }}, a set of paired SIDs supported by the style s in the virtual projection database The serial number, and Index is the last item number in the SID that is supported by the style s. The first item in a sequence is numbered 1, the second item is numbered 2, and so on. For example, s=<A,1 C>, then the virtual projection database ps(s)={{2,2},{4,4}} of s, see Figure 3D.

Definition 6: Style support

The support for a style s is the number of sequences in the database that are supported by the style s, represented by sup(s).

Definition 7: Frequent styles

The support of a frequent pattern s should not be lower than the minimum support threshold (min_sup).

1 is a schematic architectural diagram of a frequent time interval sequence pattern exploration system 1 in the present invention. As shown, the frequent time interval sequence style mapping system 1 includes a time interval sequence database 10, a virtual projection database creation module 20, a candidate pattern generation module 30, and a candidate pattern clipping module 40.

The time interval sequence database 10 is used to store the time interval sequence 11. The virtual projection database creation module 20 creates a virtual projection database 21 of the candidate style according to the time interval sequence database 10 and calculates the support degree of the candidate pattern. The candidate pattern generation module 30 searches for all L length candidate patterns from the virtual projection database 21 to calculate all L+1 length candidate patterns according to the L length candidate patterns, and then updates the L+1 lengths. The support of the candidate pattern, the candidate patterns of the L+1 lengths are aggregated into a set, where L is a positive integer equal to or greater than 1. The candidate style pruning module 40 trims or deletes the L+1 length candidate patterns within the set, for example, by the first pruning algorithm 41 or the second pruning algorithm 42. The aforementioned virtual projection database 21 can be used to store the location index of the candidate pattern.

For example, the virtual projection database creation module 20 can search all the length patterns of the L length (such as L=1 length, or 1-pattern) from the time interval sequence database 10 to be aggregated into a set (or U). _L=1 ), and the frequent patterns of the L=1 lengths in the U _L=1 set are all established into the virtual projection database 21. The candidate pattern generation module 30 may search for all L=2 length (or 2-pattern) candidate patterns from the virtual projection database 21 to calculate the support degree of the L=2 length candidate patterns, and then update the candidate pattern. For the support of the candidate styles of L=2 length, the candidate patterns of L=2 lengths are aggregated into a set (such as U _L=2 sets). The candidate style pruning module 40 trims or deletes the L=2 length candidate patterns in the U _L=2 set by the first pruning algorithm 41 or the second pruning algorithm 42.

The first pruning algorithm 41 may be, for example, if there is a first pattern P and a second pattern P ' in the set , the virtual projection database pdb(P) having the first pattern is equal to the virtual projection having the second pattern. The database pdb(P ' ), and the first style P supports the second style P', then the first style P is deleted. Furthermore, the second pruning algorithm 42 may be, for example, if the first memory P and the second pattern P ′ are present in the set, the first pattern P supports the second pattern P ′ and has the first style. The style support degree sup(P) minus the style support degree sup(P ' ) having the second style plus the maximum support degree in the paired set SP _{P '} is lower than the minimum support degree threshold (min_sup), the first style is deleted P.

The candidate pattern trimming module 40 can delete the non-minimum frequent time interval sequence pattern according to the characteristics of the time interval, and the minimum frequent time interval sequence pattern can generally refer to all the sequences with the same item and sequence, between the two sequences of the single sequence. The time interval is smaller than the time interval of the other two items.

The candidate style pruning module 40 can determine whether there is a frequent pattern of the L+1 length in the set. If yes, all the frequent patterns in the set are built into the virtual projection database 21 in a recursive manner, and then the candidate pattern is generated. Group 30 updates the support for the candidate pattern of the L+1 length to update the set.

For example, the U _{L=2 in-} set support is not less than the minimum support threshold, and the L=2 length candidate pattern is a frequent pattern of L=2 length. When the U _L=2 set has a frequent pattern of L=2 length, the candidate style pruning module 40 recursively sets the frequent patterns of the L+1 lengths in the U _L+1 set to the virtual projection database. 21. Simultaneously output a candidate sequence pattern in which all the support values of the U _L+1 set are not less than the minimum support threshold value is the frequent sequence pattern 50.

FIG. 2 is a schematic flow chart showing a method for exploring a frequent time interval sequence pattern in the present invention. Please refer to FIG. 1 above for a description.

As shown in FIG. 2, in step S1, all the L (such as L = 1) length frequent patterns are searched from the time interval sequence database 10 to be aggregated into a set (eg, U _{L = 1} set), where L Is a positive integer equal to or greater than 1 (ie, L≧1).

In step S2, all the frequent patterns in the set are built into the virtual projection database 21, and the virtual projection database 21 can be used to store the position index of the candidate pattern.

In step S3, all L+1 length candidate patterns are searched from the virtual projection database 21 to calculate the support degrees of the L+1 length candidate patterns.

In step S4, the support of the candidate patterns of the L+1 lengths is updated to update the set (such as the U _L+1 set).

In step S5, the candidate patterns of the L+1 lengths in the set (e.g., U _L+1 sets) are clipped or deleted by the first pruning algorithm 41 or the second pruning algorithm 42.

For example, the first pruning algorithm 41 is: if both the style P and the style P ' exist in the set, the virtual projection database pdb(P) is equal to the virtual projection database pdb(P ' ), and the style P supports the style P', Then delete the style P. Moreover, the second pruning algorithm 42 is: if both the style P and the style P ' exist in the set, the style P supports the style P ' , and the style support degree sup(P) minus the style support degree sup(P ' ) plus If the maximum support in the paired set SP _{P '} is lower than the minimum support threshold, the style P is deleted.

In step S6, it is judged whether there are frequent patterns of the L+1 lengths in the set (such as the U _L+1 set). If it exists, it is returned to step S2 in a recursive manner. If it does not exist, it ends. In other words, if the set (eg, U _L+1 set) is a non-empty set, steps S2 through S6 are repeatedly performed until the set (eg, U _L+1 set) is an empty set. The foregoing recursive manner may include establishing all of the frequent patterns in the set to the virtual projection database 21, and then updating the support of the candidate pattern of the L+1 length to update the set.

In addition, the frequent time interval sequence pattern exploration method of the present invention can delete the non-minimum frequent time interval sequence pattern according to the characteristics of the time interval, and the minimum frequent time interval sequence pattern can generally refer to all sequences having the same item and sequence, a single sequence. The time interval between the two items is smaller than the time interval between the two items in the other sequences.

The frequent time interval sequence style exploration method of the present invention can remove candidate patterns in the set with support below the minimum support threshold and continue All frequent patterns of L+1 are generated from a collection of frequent patterns of L lengths until no frequent patterns of longer length are produced.

3A is a schematic diagram showing a time interval sequence database in the present invention, FIG. 3B is a diagram showing the support degree of the items in the U _L=1 set in the present invention, and FIG. 3C is a diagram showing the prefix style in the present invention. The virtual projection database of A> and its support degree, the 3D figure shows the 3-pattern virtual projection database prefixed with the pattern <A> in the present invention and its support degree, and the 3E figure shows the frequent 2- in the present invention. The virtual projection database of pattern and its support degree, the 3F figure shows the virtual projection database with frequent 3-pattern in the invention and its support degree, and the 3G figure shows the virtual projection database of frequent 4-pattern in the invention. With its support. 4A is a virtual code of step S3 in FIG. 2 of the present invention, and FIG. 4B is a virtual code of step S4 in FIG. 2 of the present invention.

In detail, in step S1 of FIG. 2, all 1-patterns in the time interval sequence database are listed first, and the support degree of the item is calculated according to the above definition 3, that is, the number of sequences in which the item occurs. For each 1-pattern support exceeds the minimum support threshold and join the U _L=1 set, given the time interval sequence database of Figure 3A. Assuming that the minimum support threshold is 3 (ie, the support is greater than 3), there are items A, B, C, and E in the U _L=1 set, as shown in Figure 3B.

In step S2 of FIG. 2, each 1-pattern in the U _L=1 set establishes its virtual projection database according to definition 5, taking the <A> prefix as an example, as shown in the second column of FIG. 3C. .

In step S3 of FIG. 2, the suffix interval accumulation space is searched by using the obtained virtual projection database, as shown in the third column of FIG. 3C (the detailed steps are as shown in the virtual code of FIG. 4A), the prefix Matching the suffixes in the interval accumulation space to form candidate patterns of multiple L+1 lengths and establishing the virtual projection database to which they belong and join The collection is shown in columns 3 and 4 of Figure 3C.

The virtual projection database obtained in step S3 of FIG. 2 only considers the case where the suffix appears in the database (ie, the time interval in the suffix is the same as the item), but the time interval in the pattern represents the meaning of less than or equal to Therefore, in the case where the virtual projection database update is considered to be smaller in step S4 of FIG. 2, reference is made to the virtual code of FIG. 4B.

For example, as shown in the fourth column of FIG. 3C, the virtual projection database of <A, 5 B> is {{1, 5}}, that is, only the sequence in the sequence 1 of the virtual projection database occurs in the same order and its time interval. However, <A, 5 B> means that A occurs within five time units, and B contains the pattern of B within three time units, so pdb(<A, 3 B>) should be added to pdb (<A , 5 B>).

Therefore, step S4 of FIG. 2 accumulates each S in the space for the suffix interval, and if there is S' of the same item in the space, and the time interval of S' is smaller than the time interval of S, the virtual projection data of S' The library should be merged into the virtual projection database of S. For example, pdb(A 3 C) should contain pdb(<A,2 C>) and pdb(<A,1 C>), while pdb(<A,2 C>) ) should contain pdb(<A,1 C>).

It is stated that for each pair {SID', Index'} in each pdb (prefix + S'), if the sequence number SID' does not appear in pdb (prefix + S), then pair it {SID', Index'} is added to pdb (prefix + S). If the sequence number SID' has appeared in pdb (prefix + S), and the index ' of the SID' is smaller than pdb (previous When the index of the same SID in +S) is added, the Index is updated to index'. For example, as shown in column 5 of Figure 3C, all paired SIDs in pdb(<A,1 C>) have not appeared in pdb(<A,2 C>), then the entire pdb(<A,1 C) 〉) added to pdb(<A, 2 C>); while pdb(<A, 2 C>) has no pairing of SID=2, 3, 4 in pdb(<A, 3 C>), so All are added to pdb (<A, 3 C>), where index=2 of SID=1 in pdb(<A, 2 C>) is small, so SID=1 in pdb(<A, 3 C>) will be updated. The index is 2.

In step S5 of FIG. 2, the support degree of the L+1 length candidate pattern is calculated according to the updated virtual projection database, and the L+1 length candidate pattern whose support exceeds the minimum threshold support is regarded as L+1 length. The frequent pattern is added to the U _L+1 set, and the support is deleted below the minimum threshold support, as indicated by the third column in the 3C chart.

For example, assuming a minimum threshold support of 2, the frequent patterns in the 3C chart are only <A, 3 B>, <A, 5 B>, <A, 1 C>, <A, 2 C>, and <A, 3 C>, the rest are deleted. However, since the frequent styles are still numerous, the present invention proposes a pruning algorithm to further delete the combination of frequent patterns. If there are frequent patterns in the set to be deleted by the pruning algorithm, the process returns to step S2 and restarts until there are no frequent patterns in the set. So far, the entire loop is executed using the recursive method. Taking the prefix <A, 2 C> of the 3D figure as an example, it is found that no candidate pattern support exceeds the minimum threshold support, and the exploration of the prefix is ended.

In step S5 of FIG. 2, the use of the pruning algorithm is mentioned, mainly by deleting the styles in the set to reduce the number of exploration styles, thereby improving the efficiency of the exploration. The principle of this pruning algorithm is to assume that the style P and the style P', and the style P supports the style P', if it can be determined by the style P prefix If there are frequent patterns, they will be the same as the style P', you can delete them. For example, if it is determined that all frequent patterns of <A, 3 C> and <A, 2 C> are the same for the prefix, the deletion time is larger because the time-consuming style does not bring more information.

According to the above definition and observation, the present invention can obtain the following lemma.

Lemma 1: There are both style P and style P', if pdb(P)=pdb(P ' ) .

Lemma 2: There is both style P and style P'. If style P supports style P ' , then sup(P) Sup(P ' ).

Lemma 3: There are both style P and style P', and style P supports style P ' , if Sup(P)-sup(P ' )+sup(P ' +{at,e})<min_sup, then g SP _P g SP _{P '} stsup(g) Min_sup.

According to the above three lemmas, the present invention proposes the following two trimming algorithms.

The first trimming algorithm 41 (see Fig. 1): If there are two styles P and a pattern P ' , pdb(P) = pdb(P ' ) and the pattern P supports the style P', the pattern P is deleted. Based on Lemma 1, it can be seen that the styles grown with the prefix of style P appear in the same prefix as the style P ' . Therefore, in Fig. 3C, under the cumulative space of the prefix <A>, <A, 5 B> supports <A, 3 B> and <A, 3 C> supports <A, 2 C>, and <A, 3 C> is the same as pdb of <A, 2 C>, so delete <A, 3 C>, as shown in Fig. 3D, <A, 2 C> and <A, 3 C> indicate that the growth pattern is exactly the same. If it is deleted by this first trimming algorithm, it is represented by △ .

The second trimming algorithm 42 (see Fig. 1): If there is a pattern P and a pattern P ' , the pattern P supports the style P ' , the maximum support degree in the sup(P)-sup(P ' )+SP _{P '} <min_sup, Then delete the style P. Based on Lemma 2, you can know that sup(P)-sup(P ' ) 0. Based on Lemma 3, it can be known that the frequent pattern obtained by prefixing the style P will be the same as the frequent pattern prefixed by the style P ' , so there is no need to be longer.

For example, suppose min_sup=2, <A, 5 B> in Fig. 3C supports <A, 3 B>, and argmax(sup(<A,5 B>)-sup(<A,3 B>)+sup( SP _{<A,3 B>} ))=sup(<A,5 B>)-sup(<A,3 B〉+argmax(sup(SP _{〈A,3 B>} ))))=3-2+1< 3, therefore, delete the larger time <A, 5 B>, and argmax (sup(SP _{<A, 3 B>} )) is 1 as shown in the 1st line of Fig. 4. If it is deleted by this second trimming algorithm, it is represented by ◇ .

In summary, the present invention is described by way of example in the example database of FIG. 3A. In step S1 of Fig. 2, all frequent patterns of length L = 1 are searched from the time interval sequence database. Assuming that min_sup=2, the frequent patterns of all L=1 lengths are <A>, <B>, <C>, and <E>, and the result is as shown in FIG. 3B.

Then, all frequent patterns of L=1 length will be sequentially performed in step by step from step S2 to step S5 of FIG. It is assumed that the execution order is sequentially <A>, <B>, <C>, and <E>. First, when <A> is the start of the prefix, step S2 to step S5 are performed until the stop condition is satisfied, and then <B> is changed, and until The system and method of the present invention are stopped when all of the frequent patterns of length L = 1 are completed.

In detail, <A> lists the frequent patterns of all L=1 lengths as a prefix, and obtains a virtual projection database via step S2 of FIG. 2 (eg, column 3E, row 1, column 2), step S3 of FIG. Then, by using the virtual projection database, the suffix interval accumulation space (such as the first column and the third column of the 3E figure) and its virtual projection database (such as the third column and the fourth column of the 3E chart) are found, and then the second figure. Step S4 Update its virtual projection database and calculate the true support, then delete the <A, 3 C> using the first trimming algorithm of step S5 of Figure 2, and use the second trimming algorithm to <A, 5 B >delete.

Then, prefixed with <B>, there is no frequent pattern of L=2 length, so stop and change to prefix <C>. The candidate pattern of L=2 prefixed by <C> is only <C, 2 B>. Since the support exceeds the minimum support threshold, the candidate of L=3 is immediately prefixed with <C, 2 B>. The style, as shown in the 1st line of the 3F chart, is stopped and replaced with the prefix <E> because there is no frequent pattern of L=4 length. The frequent patterns of L=2 length under the prefix of <E> are <E, 2 A> and <E, 4 C>, so the candidate pattern of L=3 length is prefixed by <E, 2 A>, such as As shown in the second line of the 3F chart, it is found that the support of <E, 2 A, 2 C> exceeds the minimum support threshold, so continue to find L=4 with the prefix of <E, 2 A, 2 C>. Candidate style for length. As shown in Fig. 3G, the entire exploration system is stopped because there is no frequent pattern of L=4 length. Therefore, the minimum frequent interval sequence patterns are <A>, <B>, <C>, <E>, <A, 3 B>, <A, 1 C>, <A, 2 C>, <A, 2 C, 2 B>, <C, 2 B>, <E, 2 A>, <E, 4 C> and <E, 2 A2 C>.

In summary, in the frequent time interval sequence pattern exploration system and method of the present invention, the minimum frequent time interval sequence pattern can be explored, and the style exploration method has a groundbreaking contribution. Furthermore, the present invention can utilize the storage method of the virtual projection database to greatly reduce the required memory resources. At the same time, the present invention proposes two pruning algorithms according to the characteristics of the time interval, which can reduce the search space of candidate patterns and increase the efficiency of exploration.

Further, the advantages or technical effects of the present invention include at least: (1) using a data structure of a virtual projection database to avoid the need for conventional techniques. The performance bottleneck caused by the steps of copying and updating the quantity data; (2) proposing a unique pruning algorithm based on the characteristics of the time interval observation and recursive structure, greatly reducing the pattern growth space to improve the exploration efficiency; (3) presenting more detailed Time interval information.

The above-described embodiments are merely illustrative of the principles, features, and effects of the present invention, and are not intended to limit the scope of the present invention. Any person skilled in the art can recite the above without departing from the spirit and scope of the present invention. The embodiment is modified and changed. Any equivalent changes and modifications made by the disclosure of the present invention should still be covered by the scope of the patent application. Therefore, the scope of protection of the present invention should be as set forth in the scope of the patent application.

Claims (15)

A frequent time interval sequence style exploration system, comprising: a time interval sequence database for storing a time interval sequence; a virtual projection database building module, wherein the virtual projection data of the candidate pattern is established according to the time interval sequence database The library and the support degree of the candidate pattern are calculated; a candidate pattern generation module searches for all L length candidate patterns from the virtual projection database to calculate all L+1 according to the L length candidate patterns. a candidate pattern of the length, and further updating the support of the candidate patterns of the L+1 lengths, the candidate patterns of the L+1 lengths are aggregated into a set, wherein L is a positive integer equal to or greater than 1; and a candidate A style trimming module that trims or deletes the candidate patterns of the L+1 lengths within the set using a pruning algorithm. The system of claim 1, wherein the virtual projection database is used to store a location index of the candidate pattern. The system of claim 1, wherein the pruning algorithm is that if the first and second styles exist in the set, the virtual projection database having the first style is equal to the first The second style virtual projection database, and the first style supports the second style, the first style is deleted. The system of claim 1, wherein the pruning algorithm is such that if the first and second styles exist in the set, The first style supports the second style, and the style support with the first style minus the style support with the second style plus the maximum support in the paired set is lower than the minimum support threshold, then deleting The first style. The system of claim 1, wherein the candidate pattern trimming module deletes the non-minimum frequent time interval sequence pattern according to the characteristics of the time interval. The system of claim 5, wherein the minimum frequent interval sequence pattern refers to all sequences having the same item and sequence, and the time interval between two items of the single sequence is the same as the other sequences. The time interval between items is smaller. The system of claim 1, wherein the candidate pattern trimming module further determines whether the frequent pattern of the L+1 length exists in the set, and if present, returns all of the set in the recursive manner. The virtual projection database is established in the frequent style, and the candidate pattern generation module updates the support degree of the candidate pattern of the L+1 length to update the set. A frequent time interval sequence style exploration method includes: searching for a frequent pattern of all L lengths from a time interval database to be aggregated into a set, wherein L is a positive integer equal to or greater than 1; all frequent patterns in the set are Establishing a virtual projection database; searching for all L+1 length candidate patterns from the virtual projection database to calculate the support degree of the L+1 length candidate patterns; Updating the support of the L+1 length candidate styles to update the set; and trimming or deleting the L+1 length candidate styles within the set by a pruning algorithm. The method of claim 8, wherein the virtual projection database is used to store a location index of the candidate pattern. The system of claim 8, wherein the pruning algorithm is such that if the first and second styles exist in the set, the virtual projection database having the first style is equal to the first The second style virtual projection database, and the first style supports the second style, the first style is deleted. The method of claim 8, wherein the trimming algorithm is that if the first and second styles exist in the set, the first style supports the second style and has the first The style of the style is subtracted from the style support with the second style plus the maximum support in the paired set is lower than the minimum support threshold, and the first style is deleted. The method of claim 8, further comprising deleting the non-minimum frequent time interval sequence pattern according to the characteristics of the time interval. The method of claim 12, wherein the minimum frequent interval sequence pattern refers to all sequences having the same item and sequence, and the time interval between two items of the single sequence is the same as the other sequences. The time interval between items is smaller. For example, the method described in claim 8 of the patent scope includes determining the set. Whether there is a frequent pattern of the L+1 length in the combination, if any, all the frequent patterns in the set are established in the recursive manner to establish the virtual projection database, thereby updating the support degree of the L+1 length candidate pattern to Update the collection. The method of claim 8, further comprising removing candidate patterns in the set having a support below a minimum support threshold, and continuously generating all L+1 from the set of frequent patterns of the L length. Frequent patterns until no frequent patterns of longer length are produced.