A commodity real-time recommendation method and system based on efficient repeated negative sequence pattern mining

Abstract

The application discloses a commodity real-time recommendation method and system based on high-efficiency repeated negative sequence pattern mining, and relates to the technical field of data mining.The application designs a prefix tree structure for storing and maintaining high-efficiency repeated positive sequence patterns, and realizes efficient and dynamic management of data; and the application further proposes a new high-efficiency repeated negative sequence candidate pattern utility value calculation mode, specifically: for any high-efficiency repeated negative sequence candidate pattern, the utility values of the high-efficiency repeated positive sequence patterns matched with the high-efficiency repeated negative sequence candidate pattern are added to obtain the utility value of the high-efficiency repeated negative sequence candidate pattern.The application can effectively improve the accuracy of high-efficiency repeated negative sequence pattern mining, and the commodities recommended based on the more accurate high-efficiency repeated negative sequence patterns mined by the application layer can be more in line with the needs of users.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data mining, and particularly relates to a commodity real-time recommendation method and system based on efficient repeated negative sequence pattern mining. BACKGROUND

[0002] With the rapid development of e-commerce shopping platforms, commodities on the online platform are constantly updated in real time, and the number of commodities grows exponentially. The existing commodity recommendation system is usually difficult to realize real-time analysis on continuously generated and dynamically changed shopping data streams, resulting in a lag in the recommendation results and the inability to timely capture the dynamic changes in user interest. Therefore, the present application proposes a commodity real-time recommendation method and system capable of efficiently mining repeated negative sequence patterns in real time on shopping data streams. SUMMARY

[0003] In order to solve the above-mentioned deficiencies in the prior art, the present application proposes a commodity real-time recommendation method and system based on efficient repeated negative sequence pattern mining.

[0004] A commodity real-time recommendation method based on efficient repeated negative sequence pattern mining comprises the following steps:

[0005] S1, collecting shopping data streams and sequentially performing data cleaning and classification to obtain user shopping sequences, commodity categories ITEM and their utility values; storing the user shopping sequences and their corresponding user IDs into a user shopping sequence data table in chronological order; then, dividing the shopping sequences in the user shopping sequence data table into different data batches in chronological order, each data batch containing the same number of shopping sequences; the present application uses a sliding window to real-time receive and manage the latest data batches;

[0006] S2, sequentially processing the encoded shopping sequences contained in the data batches in the sliding window in chronological order to obtain efficient repeated negative sequence patterns; then, moving the sliding window to delete the encoded shopping sequences of the old data batches and incorporate the encoded shopping sequences of the new data batches; then, processing the encoded shopping sequences contained in the new data batches to obtain efficient repeated negative sequence patterns; wherein the efficient repeated negative sequence patterns are obtained based on efficient repeated negative sequence candidate patterns, and the utility value of the efficient repeated negative sequence candidate pattern with a length greater than or equal to 2 is obtained by matching the efficient repeated negative sequence candidate pattern with a corresponding efficient repeated positive sequence pattern and accumulating the utility value of the efficient repeated positive sequence pattern;

[0007] S3, adjusting the commodities displayed on the platform interface based on the efficient repeated negative sequence patterns, their utility values and the corresponding user IDs.

[0008] Preferably, step S2 specifically comprises the following steps:

[0009] S2-1, performing an encoding operation on the shopping sequences in the user shopping sequence data table and the commodity elements in the element data table to obtain an encoded element data table and an encoded user shopping sequence data table;

[0010] S2-2, performing efficient utility repeated positive sequence pattern mining on the encoded shopping sequences contained in each data batch in the sliding window in chronological order, and storing the efficient utility repeated positive sequence patterns and the utility list for storing information related to the efficient utility repeated positive sequence patterns into the prefix tree respectively; performing negative transformation on the efficient utility repeated positive sequence patterns to obtain efficient utility repeated negative sequence candidate patterns; then, calculating the utility values of the efficient utility repeated negative sequence candidate patterns, and comparing the utility values of the efficient utility repeated negative sequence candidate patterns with the minimum negative utility threshold ξ', and taking the efficient utility repeated negative sequence candidate patterns with utility values greater than or equal to the minimum negative utility threshold ξ' as the efficient utility repeated negative sequence patterns;

[0011] S2-3, moving the sliding window, deleting the encoded shopping sequences of a number of old data batches, and incorporating the encoded shopping sequences of a number of new data batches; processing the encoded shopping sequences contained in the new data batches by using the algorithm layer according to step S2-2.

[0012] Preferably, step S2-2 specifically comprises the following steps:

[0013] S2-2-1, performing efficient utility repeated positive sequence pattern mining on the shopping sequences contained in each data batch in the sliding window in chronological order by using the algorithm layer;

[0014] S2-2-2, obtaining efficient utility repeated negative sequence candidate patterns based on the efficient utility repeated positive sequence patterns;

[0015] S2-2-3, traversing all efficient utility repeated negative sequence candidate patterns with a length greater than or equal to 2 in the efficient utility repeated negative sequence pattern list, matching the efficient utility repeated negative sequence candidate patterns to the corresponding efficient utility repeated positive sequence patterns, and all the matched efficient utility repeated positive sequence patterns constituting an efficient utility repeated positive sequence pattern set; calculating the utility values of the efficient utility repeated negative sequence candidate patterns with a length greater than or equal to 2 based on the utility values of the efficient utility repeated positive sequence patterns matched with the efficient utility repeated negative sequence candidate patterns; taking the efficient utility repeated negative sequence candidate patterns with utility values greater than or equal to the minimum negative utility threshold ξ' as the efficient utility repeated negative sequence patterns; and storing all the efficient utility repeated negative sequence patterns into the efficient utility repeated negative sequence pattern list.

[0016] Preferably, step S2-2-3 specifically comprises the following steps:

[0017] (1) obtaining a user ID set containing high-utility repetitive negative sequence candidate patterns;

[0018] (2) based on the user ID set, querying high-utility repetitive positive sequence patterns matched with the high-utility repetitive negative sequence candidate patterns according to the hash table in the prefix tree; the matched high-utility repetitive positive sequence patterns constitute a high-utility repetitive positive sequence pattern set;

[0019] (3) accumulating the utility values corresponding to the high-utility repetitive positive sequence patterns satisfying the matching condition to obtain the utility value of the high-utility repetitive negative sequence candidate pattern.

[0020] Preferably, step S2-2-2 specifically comprises the following steps:

[0021] S2-2-2-1, obtaining a high-utility repetitive negative sequence candidate pattern with a length of 1 by using a high-utility repetitive positive sequence pattern with a length of 1, including the following steps: performing negative transformation on the high-utility repetitive positive sequence pattern with a length of 1 to obtain a high-utility repetitive negative sequence candidate pattern with a length of 1; calculating the utility value of the high-utility repetitive negative sequence candidate pattern with a length of 1, and taking the high-utility repetitive negative sequence candidate pattern with an utility value greater than or equal to the minimum negative utility threshold ξ' as the high-utility repetitive negative sequence pattern with a length of 1;

[0022] S2-2-2-2, connecting the high-utility repetitive positive sequence pattern with a length of 1 with the high-utility repetitive negative sequence candidate pattern with a length of 1 to obtain a high-utility repetitive negative sequence candidate pattern with a length of 2;

[0023] S2-2-2-3, obtaining a high-utility repetitive negative sequence candidate pattern with a length of n based on a high-utility repetitive positive sequence pattern with a length of (n-1), wherein n is equal to the number of commodity elements in the shopping sequence, and n≥3.

[0024] Preferably, in S2-2-2-3, the way of obtaining a high-utility repetitive negative sequence candidate pattern with a length of n based on a high-utility repetitive positive sequence pattern with a length of (n-1) includes way one and way two; wherein,

[0025] way one: adding a high-utility repetitive negative sequence pattern with a length of 1 at the end of the high-utility repetitive positive sequence pattern with a length of (n-1) to obtain a high-utility repetitive negative sequence candidate pattern with a length of n, n≥3;

[0026] Mode two, at the end of the high-efficiency repeated negative sequence candidate mode with a length of (n-1), add the corresponding mode according to the negative constraint condition, wherein the negative constraint condition refers to that the high-efficiency repeated negative sequence candidate mode does not allow to exist two negative commodity elements in succession; if the last element of the high-efficiency repeated negative sequence candidate mode is a negative commodity element, add a high-efficiency repeated positive sequence mode with a length of 1 at the end; if the last element of the high-efficiency repeated negative sequence candidate mode is a positive commodity element, add a high-efficiency repeated negative sequence mode with a length of 1 or a high-efficiency repeated positive sequence mode with a length of 1 at the end, to obtain a high-efficiency repeated negative sequence candidate mode with a length of n, n≥3;

[0027] Repeat mode one and mode two until no new high-efficiency repeated negative sequence candidate mode is generated; or until the length of the high-efficiency repeated negative sequence candidate mode is greater than m+1, wherein m is the maximum length of the high-efficiency repeated positive sequence mode.

[0028] A high-efficiency repeated negative sequence mode mining commodity real-time recommendation system, comprising a data storage layer, an algorithm layer and an application layer, wherein,

[0029] The data storage layer is used for collecting shopping data streams, data cleaning and classification, obtaining user shopping sequences, commodity categories ITEM and utility values thereof; storing the user shopping sequences and corresponding user IDs in the user shopping sequence data table in chronological order; dividing the shopping sequences in the user shopping sequence data table into different data batches in chronological order, each data batch containing the same number of shopping sequences; using a sliding window to continuously slide on the data batches with a preset window size, and real-time receiving and managing the latest data batches;

[0030] The algorithm layer is used for processing the encoded shopping sequences contained in the data batches in the sliding window in chronological order to obtain high-efficiency repeated negative sequence modes; then, moving the sliding window, deleting the encoded shopping sequences of the old data batches, and incorporating the encoded shopping sequences of the new data batches; and processing the encoded shopping sequences contained in the new data batches to obtain high-efficiency repeated negative sequence modes;

[0031] The application layer adjusts the commodities displayed on the platform interface based on the high-efficiency repeated negative sequence modes and utility values thereof, and combines the user IDs; at the same time, saves the adjustment records, generates log files and transmits them to the data storage layer for archival storage.

[0032] Compared with the prior art, the beneficial technical effects of the present application are:

[0033] The prefix tree structure is designed for storing and maintaining high-efficiency repeated positive sequence patterns, and realizing efficient and dynamic management of data. The application also proposes a new high-efficiency repeated negative sequence candidate pattern utility value calculation method. Specifically, for any high-efficiency repeated negative sequence candidate pattern, the utility value of the high-efficiency repeated positive sequence pattern matched with the high-efficiency repeated positive sequence pattern is accumulated to obtain the utility value of the high-efficiency repeated negative sequence candidate pattern. This method of obtaining the utility value of the high-efficiency repeated negative sequence candidate pattern does not directly discard the influence of negative commodity elements, but indirectly quantifies the utility of the repeated negative sequence candidate pattern through the logic of "excluding positive sequences containing negative commodity elements", which is equivalent to starting from the "real preference of the user when selecting positive commodities", and can effectively improve the accuracy of high-efficiency repeated negative sequence pattern mining. The recommended commodities based on the more accurate high-efficiency repeated negative sequence pattern mined by the application layer can better meet the needs of users. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 The method and HUNSPM The method in the application has a minimum utility threshold ξ of 2x10 4 , the data batch is 1000, and the window size is set to 5, 6, 7, 8, and 9 respectively, the test results of testing the DS1 data set under the condition that the minimum utility threshold ξ is 2x10

[0035] Figure 2 The method and HUNSPM The method in the application has a minimum utility threshold ξ of 3x10 4 , the data batch is 1000, and the window size is set to 4, 5, 6, 7, and 8 respectively, the test results of testing the DS2 data set under the condition that the minimum utility threshold ξ is 3x10

[0036] Figure 3 The method and HUNSPM The method in the application has a minimum utility threshold ξ of 2x10 4 , the data batch is 1000, and the window size is set to 5, 6, 7, 8, and 9 respectively, the test results of testing the DS3 data set under the condition that the minimum utility threshold ξ is 2x10

[0037] Figure 4 The method and HUNSPM The method in the application has a minimum utility threshold ξ of 4x10 4 , the data batch is 1000, and the window size is set to 5, 6, 7, 8, and 9 respectively, the test results of testing the DS4 data set under the condition that the minimum utility threshold ξ is 4x10

[0038] Figure 5 The method and HUNSPM The method has a minimum utility threshold ξ of 2 × 10 4 Test results of the DS1 dataset under the conditions of a window size of 8 and data batches of 800, 850, 900, 950, and 1000 respectively;

[0039] Figure 6 The method described in this application and HUNSPM The method has a minimum utility threshold ξ of 3 × 10 4 Test results of the DS2 dataset under the conditions of a window size of 6 and data batches of 800, 850, 900, 950, and 1000 respectively;

[0040] Figure 7 The method described in this application and HUNSPM The method has a minimum utility threshold ξ of 2 × 10 4 Test results of the DS3 dataset under the conditions of a window size of 8 and data batches of 800, 850, 900, 950, and 1000 respectively;

[0041] Figure 8 The method described in this application and HUNSPM The method has a minimum utility threshold ξ of 4 × 10 4 Test results of the DS4 dataset under the conditions of a window size of 8 and data batches set to 800, 850, 900, 950, and 1000 respectively;

[0042] Figure 9 The method described in this application and HUNSPM The method uses a data batch size of 1000, a window size of 6, and a minimum utility threshold ξ set to 2×10. 4 2.5×10 4 3×10 4 3.5×10 4 4×10 4 The test results of testing the DS1 dataset under the given conditions;

[0043] Figure 10 The method described in this application and HUNSPM The method uses a data batch size of 1200, a window size of 5, and a minimum utility threshold ξ set to 2×10. 4 2.5×10 4 3×10 4 3.5×10 4 4×10 4 The test results of testing the DS2 dataset under the given conditions;

[0044] Figure 11 The method and HUNSPM The test results of testing the DS3 data set under the condition that the data batch is 1000, the window size is 5, and the minimum utility threshold ξ is set to 2x10 4 , 2.5x10 4 , 3x10 4 , 3.5x10 4 , 4x10 4 , respectively;

[0045] Figure 12 The method and HUNSPM The test results of testing the DS4 data set under the condition that the data batch is 1000, the window size is 5, and the minimum utility threshold ξ is set to 4x10 4 , 4.2x10 4 , 4.4x10 4 , 4.6x10 4 , 4.8x10 4 , respectively. DETAILED DESCRIPTION

[0046] A commodity recommendation system based on high-utility repeated negative sequence pattern mining, comprising a data storage layer, an algorithm layer and an application layer, a commodity real-time recommendation method based on the commodity recommendation system, comprising the following steps:

[0047] S1, using the data storage layer to collect the user shopping data stream on the shopping platform in chronological order, and sequentially performing data cleaning and data classification on the collected user shopping data stream to obtain the user shopping sequence and all commodity categories ITEM and their utility values, wherein all commodity categories ITEM and their utility values are counted into an element data table; then, the user shopping sequence and the corresponding user ID are stored in the user shopping sequence data table in chronological order; then, the shopping sequence in the user shopping sequence data table is divided into different data batches in chronological order, each data batch containing the same number of shopping sequences; then, a sliding window is used to continuously slide on the data batch with a preset window size, to real-time receive and manage the latest data batch, thereby ensuring timely processing and updating of the latest data, in the embodiment, the preset window size is 2, the sliding window slides each time, only one data batch is removed, and one data batch is added, the removed data batch is the oldest data batch, and the added data batch is the latest data batch, the element data table and the data batch received and managed by the sliding window are transmitted to the algorithm layer by the data storage layer; specifically:

[0048] In the present application, data cleaning refers to cleaning the data irrelevant to the commodity recommendation and the abnormal data in the collected user shopping data stream; wherein the data irrelevant to the commodity recommendation refers to the user's shopping time, the user's shopping time has no direct influence on judging the user's preference for the commodity, so the user's shopping time data is cleaned; the abnormal data refers to the case that the same user shopping record is written multiple times, such data will cause data redundancy and interfere with the subsequent accurate analysis of the user's shopping preference, so it will also be deleted in the data cleaning process.

[0049] The data classification processing is performed on the user shopping data stream after data cleaning, and the user's shopping sequence and all commodity types ITEM and utility values are obtained; specifically including the following steps:

[0050] 1) For the shopping data stream after data cleaning, a shopping sequence set is created for each user, and a unique user ID is set for each user;

[0051] 2) The types of each commodity ITEM in the shopping data stream and the purchase quantity of each commodity type ITEM are counted in the user set;

[0052] 3) All commodity types ITEM and utility values in the shopping data stream are counted into the element data table; in this embodiment, the element data table is shown in Table 1:

[0053] Table 1 Element data table

[0054]

[0055] In Table 1, the commodity types ITEM include down jackets, cotton shoes and cotton pants, and the corresponding utility values are 1000 yuan, 500 yuan and 200 yuan respectively. These utility values represent the value of each type of commodity.

[0056] 4) All purchased commodity types ITEM and corresponding purchase quantities of the same user in the shopping data stream after data cleaning constitute a shopping sequence; the user's shopping sequence and the corresponding user ID are stored in the user shopping sequence data table in chronological order; in this embodiment, the user shopping sequence data table is shown in Table 2:

[0057] Table 2 User shopping sequence data table

[0058]

[0059] The user shopping sequence data table shown in Table 2 in this embodiment contains six shopping sequences, which are divided into three data batches in chronological order, each data batch contains two shopping sequences, specifically,

[0060] Data batch 1 contains the shopping sequences of user 1 and user 2, wherein the shopping sequence of user 1 is <(down jacket, 2) (cotton shoes, 2)>, which means that user 1 bought two down jackets and two pairs of cotton shoes in the same time period 1; the shopping sequence of user 2 is <(down jacket, 1) (down jacket, 2) (cotton shoes, 1) (cotton shoes, 2) (cotton pants, 3)>, which means that user 2 bought one down jacket, two down jackets, one pair of cotton shoes, two pairs of cotton shoes and three cotton pants in the same time period 1.

[0061] Data batch 2 contains the shopping sequences of user 3 and user 4, wherein the shopping sequence of user 3 is <(down jacket, 2) (cotton shoes, 1) (cotton pants, 3)>, which means that user 3 bought two down jackets, one pair of cotton shoes and three cotton pants in the same time period 2; the shopping sequence of user 4 is <(down jacket, 3) (cotton pants, 2) (cotton shoes, 1)>, which means that user 4 bought three down jackets, two cotton pants and one cotton shoe in the same time period 2.

[0062] Data batch 3 contains the shopping sequences of user 5 and user 6, wherein the shopping sequence of user 5 is <(down jacket, 2) (cotton shoes, 2) (cotton pants, 1)>, which means that user 5 bought two down jackets, two pairs of cotton shoes and one cotton pants in the same time period 3; the shopping sequence of user 6 is <(cotton shoes, 2) (cotton pants, 3)>, which means that user 6 bought two pairs of cotton shoes and three cotton pants in the same time period 3.

[0063] In this embodiment, two data batches are contained in the sliding window, and the two data batches included in the sliding window in the initial state are data batch 1 and data batch 2; after the encoded shopping sequences contained in data batch 1 and data batch 2 are processed by the algorithm layer, the sliding window slides, deletes the encoded shopping sequences in data batch 1, and includes the encoded shopping sequences in data batch 3.

[0064] S2, the encoded shopping sequences contained in the data batches in the algorithm layer sliding window are processed in time sequence to obtain high-efficiency repeated negative sequence patterns, and in this embodiment, the encoded shopping sequences contained in data batch 1 and data batch 2 in the sliding window are processed by the algorithm layer to obtain high-efficiency repeated negative sequence patterns; then, the sliding window moves, deletes the encoded shopping sequences of the oldest data batch, and includes the encoded shopping sequences of the latest data batch, wherein the oldest data batch refers to the earliest data batch in time in the sliding window; then, the encoded shopping sequences contained in the latest data batch are processed by the algorithm layer to obtain high-efficiency repeated negative sequence patterns; specifically, the following steps are included:

[0065] S2-1, the algorithm layer is used for encoding operation on the shopping sequence in the user shopping sequence data table and the commodity elements in the element data table; wherein the shopping sequence in the user shopping sequence data table is the encoded shopping sequence contained in each data batch; in the embodiment, different commodity elements are encoded into different symbols, wherein the down jacket is encoded as a, the cotton shoes are encoded as b, and the cotton trousers are encoded as c; in the user shopping sequence list, each commodity element is replaced by the corresponding symbol, and the purchase quantity is retained; wherein the encoded element data table is shown in Table 3; and the encoded user shopping sequence data table is shown in Table 4;

[0066] Table 3: Encoded element data table

[0067]

[0068] Table 4: Encoded user shopping sequence data table

[0069]

[0070] S2-2, the algorithm layer is used for processing the encoded shopping sequence contained in the data batch 1 and the data batch 2 in the sliding window in sequence, obtaining the high-efficiency repeated negative sequence mode, specifically: the algorithm layer is used for mining the high-efficiency repeated positive sequence mode from the encoded shopping sequence contained in the data batch 1 and the data batch 2 in the sliding window in sequence, obtaining the high-efficiency repeated positive sequence mode, storing the high-efficiency repeated positive sequence mode and the utility list for storing the high-efficiency repeated positive sequence mode related information into the prefix tree respectively; the high-efficiency repeated positive sequence mode is negatively transformed to obtain the high-efficiency repeated negative sequence candidate mode; then, the utility value of the high-efficiency repeated negative sequence candidate mode is calculated, and the utility value of the high-efficiency repeated negative sequence candidate mode is compared with the minimum negative utility threshold ξ'; the high-efficiency repeated negative sequence candidate mode with the utility value greater than or equal to the minimum negative utility threshold ξ' is taken as the high-efficiency repeated negative sequence mode; step S2-2 specifically includes the following steps:

[0071] S2-2-1, the encoded shopping sequence contained in each data batch in the sliding window is mined according to the time sequence, and in the embodiment, the encoded shopping sequence contained in the data batch 1 and the data batch 2 in the sliding window in the initial state is mined according to the high-efficiency repeated positive sequence mode by the algorithm layer in sequence, specifically including the following steps:

[0072] S2-2-1-1, using the algorithm layer to sequentially mine the encoded shopping sequences contained in the data batch 1 and the data batch 2 in the sliding window for length-1 high-efficiency repetitive positive sequence candidate patterns, to obtain length-1 high-efficiency repetitive positive sequence candidate patterns of each type of commodity element in each encoded shopping sequence, and store the high-efficiency repetitive positive sequence candidate patterns in a high-efficiency repetitive positive sequence candidate pattern list; for example, mining the encoded shopping sequences of user 2 in Table 4 for length-1 high-efficiency repetitive positive sequence candidate patterns can obtain length-1 high-efficiency repetitive positive sequence candidate patterns 、 and .

[0073] S2-2-1-2, calculate the utility values of all length-1 high-efficiency repetitive positive sequence candidate patterns obtained in step S2-2-1-1, and compare the utility values with the minimum utility threshold ξ, and take the high-efficiency repetitive positive sequence candidate patterns with utility values greater than or equal to the minimum utility threshold ξ as length-1 high-efficiency repetitive positive sequence patterns; store all length-1 high-efficiency repetitive positive sequence patterns in a high-efficiency repetitive positive sequence pattern list, and store all length-1 high-efficiency repetitive positive sequence patterns and their utility values in a prefix tree.

[0074] In step S2-2-1-2 of the present application, the utility value of the length-1 high-efficiency repetitive positive sequence candidate pattern is calculated as shown in formula (1):

[0075] (1)

[0076] In formula (1), represents the utility value corresponding to the high-efficiency repetitive positive sequence candidate pattern when a commodity element type appears for the th time in a certain encoded shopping sequence in the th data batch B; represents the external utility value (i.e. unit price) of a commodity element type represents the purchase quantity corresponding to the th appearance of a commodity element type in a certain encoded shopping sequence S of a user ID d.

[0077] ​​For example, the encoded shopping sequence of user 2 in Table 4 is <(a, 1) (a, 2) (b, 1) (b, 2) (c, 3)>, which reflects the order of the user's 2 purchase behavior and the specific purchase quantity in the same time period. For example, in the encoded shopping sequence of user 2 in Table 4, the same kind of goods appears multiple times. In order to accurately distinguish the different purchase behaviors of the same kind of goods, it is necessary to mark the purchase order of the same kind of goods separately; for example, the product element category a appears twice, the first occurrence (a, 1) is marked as (a1, 1), indicating the first purchase behavior of down jacket in the sequence, and the second occurrence (a, 2) is marked as (a2, 2), indicating the second purchase behavior of down jacket in the sequence; therefore, the encoded shopping sequence of user 2 is marked as <(a1, 1) (a2, 2) (b1, 1) (b2, 2) (c, 3)>. represents the external utility value of product category a in the encoded shopping sequence, represents the purchase quantity of user 2 when purchasing product category a for the first time.

[0078] To realize the dynamic storage of high-utility repetitive sequential patterns, the present application designs a prefix tree structure, the specific structure of which is defined as follows:

[0079] The high-utility repetitive sequential pattern with a length of 1 and its utility value are added to the first layer of the prefix tree; in the first layer, each node of the prefix tree stores a high-utility repetitive sequential pattern and its utility list with information related to the high-utility repetitive sequential pattern.

[0080] In the present application, the information related to the high-utility repetitive sequential pattern in the utility list includes data batch, utility array and hash table structure; wherein,

[0081] 1) Utility array; the length of the utility array is consistent with the window size, and the utility array is a set of record collections, each record storing the utility value of the high-utility repetitive sequential pattern corresponding to each data batch contained in the current window. When the window slides, only the utility value corresponding to the high-utility repetitive sequential pattern of the oldest data batch in the utility array needs to be deleted, and the utility value corresponding to the high-utility repetitive sequential pattern of the latest data batch needs to be added in the utility array; in the present application, the cumulative result of each record in the utility array is taken as the total utility value of the high-utility repetitive sequential pattern in the current window, which directly reflects the preference heat of the product category in the user group of the last N data batches, and the recommendation system can improve the display priority of the product with the largest utility value on the homepage of the recommendation system according to the cumulative result of each record in the utility array.

[0082] 2) a hash table structure for storing the user IDs associated with the high-utility repetitive positive sequential patterns and the utility values corresponding to each user ID; since a key step in the subsequent mining of high-utility repetitive negative sequential patterns is to calculate the utility values of the high-utility repetitive negative sequential candidate patterns, and the calculation of the utility values of the high-utility repetitive negative sequential candidate patterns depends on the utility values and position information of the high-utility repetitive positive sequential patterns;

[0083] The present application can be quickly located by a prefix tree, and the hash table structure is used to realize accurate query of the high-utility repetitive positive sequential patterns, so that the query complexity is reduced from O (N) to O (1).

[0084] S2-2-1-3, the high-utility repetitive positive sequential candidate patterns of different lengths of 1 of the same time and the same user are connected, and recursive cycles are performed to obtain high-utility repetitive positive sequential candidate patterns of lengths of 2 to n, respectively, n being the same as the number of types of commodities in the encoded shopping sequence; wherein, before the mode connection, the high-utility repetitive positive sequential candidate patterns of length 1 with the weighted utility upper bound or the remaining utility upper bound being lower than the minimum utility threshold ξ are deleted; in the present application, the high-utility repetitive positive sequential candidate patterns of lengths of 2 or more are stored in a high-utility repetitive positive sequential candidate pattern list;

[0085] For example, the high-utility repetitive positive sequential candidate patterns of length 1 obtained based on the encoded shopping sequence of user 2 in table 4 are connected, and recursive cycles are performed to obtain high-utility repetitive positive sequential candidate patterns of lengths of 2 and 3, respectively, and the high-utility repetitive positive sequential candidate patterns of lengths of 2 and 3 are stored in the high-utility repetitive positive sequential candidate pattern list.

[0086] In the mining process of high-utility repetitive sequential patterns, in order to reduce the computational complexity and avoid the problem of reduced computational efficiency caused by the large size of high-utility repetitive sequential candidate patterns with high utility, the application deletes the high-utility repetitive sequential candidate patterns with a length of 1 and a weighted utility upper bound or a remaining utility upper bound lower than the minimum utility threshold ξ before performing the pattern connection operation, and then performs the pattern connection operation to generate high-utility repetitive sequential candidate patterns with a longer length. Specifically, the application adopts two pruning strategies of weighted utility upper bound and remaining utility upper bound, calculates the weighted utility upper bound and the remaining utility upper bound of the high-utility repetitive sequential candidate pattern, compares the two utility upper bounds with the minimum utility threshold ξ, and determines that the high-utility repetitive sequential candidate pattern does not have the potential to become a high-utility repetitive sequential pattern if the weighted utility upper bound or the remaining utility upper bound of the high-utility repetitive sequential candidate pattern is less than the minimum utility threshold ξ. The high-utility repetitive sequential candidate pattern is deleted from the high-utility repetitive sequential candidate pattern list, and only the high-utility repetitive sequential candidate patterns with a weighted utility upper bound and a remaining utility upper bound greater than or equal to the minimum utility threshold ξ are reserved for the pattern connection operation. The two utility upper bounds and the mode connection method using the two utility upper bounds are consistent with the contents of sections 4.3 and 4.4 disclosed in the paper USpan: An Efficient Algorithm for Mining High Utility Sequential Patterns.

[0087] S2-2-1-4, all high-utility repetitive sequential candidate patterns with a length greater than or equal to 2 stored in the high-utility repetitive sequential candidate pattern list are constrained in a non-overlapping manner to obtain non-overlapping sequential candidate patterns, the length of the non-overlapping sequential candidate pattern is greater than or equal to 2, the utility value of all non-overlapping sequential candidate patterns is calculated, and the non-overlapping sequential candidate pattern with a utility value greater than or equal to the minimum utility threshold ξ is taken as a high-utility repetitive sequential pattern with a length greater than or equal to 2; and all non-overlapping sequential candidate patterns are stored in the high-utility repetitive sequential candidate pattern list, and all high-utility repetitive sequential patterns with a length greater than or equal to 2 and their utility values are stored in the second layer of the prefix tree; in the second layer, each node of the prefix tree stores a high-utility repetitive sequential pattern and its utility list with information related to the high-utility repetitive sequential pattern. In the application, the utility array corresponding to all high-utility repetitive sequential patterns with a length greater than or equal to 2 is accumulated, that is, the utility value of the high-utility repetitive sequential pattern with a length greater than or equal to 2 in the sliding window.

[0088] In the application, the way of constraining the high-utility repetitive sequential candidate patterns with a length greater than or equal to 2 in a non-overlapping manner is consistent with the content disclosed in S2-1-5 of the invention patent CN120146968B.

[0089] For example, the encoded shopping sequence of user 2 in Table 4 is <(a, 1)(a, 2)(b, 1)(b, 2)(c, 3)>, and the high utility repetitive sequential pattern candidates of length greater than or equal to 2 that can be mined are <ab> 、 <ac> 、 <bc> 、 <abc>; when a high utility repetitive sequence candidate pattern is mined <ab>The positions of the encoded shopping sequence of the user 2 are (1, 3), (1, 4), (2, 3), (2, 4) respectively. According to the non-overlapping constraint, if the same product element in the same position repeats in the multiple matching positions of the encoded shopping sequence of the high utility repetitive positive sequence candidate mode with a length greater than or equal to 2, only the first matching position is retained, and the subsequent repeated positions are deleted. For example, in (1, 3) and (1, 4), the product category a repeats in position 1, so (1, 4) is deleted; in (1, 3), (2, 3) and (2, 4), the product category b repeats in position 3, so (2, 3) is deleted. Finally, the high utility repetitive positive sequence candidate mode is mined <ab>When, (1, 3), (2, 4) are two times of non-overlapping appearance of the high efficient repeated positive sequence candidate mode.

[0090] The non-overlapping matching position obtained after deduplication is used to extract the utility value of the corresponding commodity element at each position and perform accumulation operation, and the utility value of the non-overlapping positive sequence candidate mode is obtained.

[0091] The utility value of all non-overlapping positive sequence candidate modes in the present application is calculated according to formula (2).

[0092] (2)

[0093] In formula (2), represents the utility value of the non-overlapping positive sequence candidate mode in the first data batch B; represents the utility value of the non-overlapping positive sequence candidate mode in the encoded shopping sequence S of the user ID d.

[0094] The utility value of the non-overlapping positive sequence candidate mode in the encoded shopping sequence S of the user ID d is calculated according to formula (3).

[0095] (3)

[0096] In formula (3), represents an encoded shopping sequence, specifically, an encoded shopping sequence S of the user ID d, represents the non-overlapping positive sequence candidate mode; represents the non-overlapping positive sequence candidate mode in the encoded shopping sequence ; represents the number of all non-overlapping appearances of the non-overlapping positive sequence candidate mode t in the encoded shopping sequence ; ω represents the number of appearances of the non-overlapping positive sequence candidate mode in the encoded shopping sequence that meets the non-overlapping condition, represents the utility value of the commodity element in the encoded shopping sequence .

[0097] As shown in formula (3), when the non-overlapping positive sequence candidate mode does not appear in the encoded shopping sequence , 0; when the non-overlapping positive sequence candidate mode in a certain coded shopping sequence appears only once, then the utility value corresponding to the item element at the position of the occurrence is equal to when there is no overlapping positive sequence candidate pattern in a certain coded shopping sequence appears twice without overlapping, then the utility value corresponding to the item element at the position of the occurrence is equal to the maximum value of the utility values corresponding to the item element at the position of the occurrence of all the overlapping positive sequence candidate patterns in a certain coded shopping sequence appears more than or equal to twice without overlapping, then the sum of the utility values corresponding to the item element at the position of the occurrence of all the overlapping positive sequence candidate patterns t is equal to .

[0098] For example, the coded shopping sequence of user 2 in Table 4 is <(a, 1)(a, 2)(b, 1)(b, 2)(c, 3)>, and the high-utility repeating positive sequence candidate patterns with length more than or equal to 2 that can be mined are <ab> 、 <ac> 、 <bc> 、 <abc>When mining high utility repetitive sequential candidate patterns <ab>In this case, (1, 3), (2, 4) are high-efficiency repeated normal sequence candidate modes <ab>In the user- encoded shopping sequence 2, the two non-overlapping occurrences of "milk" and "bread" are then U r ( <ab>S2) is the maximum of the utility values of the two matching positions.

[0099] S2-2-2, obtaining high-utility repetitive negative sequence candidate patterns based on high-utility repetitive positive sequence patterns; wherein all the high-utility repetitive positive sequence patterns include the high-utility repetitive positive sequence patterns with length 1 in S2-2-1-2 and the high-utility repetitive positive sequence patterns with length greater than or equal to 2 in S2-2-1-4; and the high-utility repetitive positive sequence patterns include the following steps:

[0100] S2-2-2-1, obtaining high-utility repetitive negative sequence candidate patterns with length 1 by using high-utility repetitive positive sequence patterns with length 1, and the steps include the following steps: performing negative transformation on the high-utility repetitive positive sequence patterns with length 1 to obtain high-utility repetitive negative sequence candidate patterns with length 1; calculating the utility values of the high-utility repetitive negative sequence candidate patterns with length 1, and taking the high-utility repetitive negative sequence candidate patterns with utility values greater than or equal to the minimum negative utility threshold ξ' as the high-utility repetitive negative sequence patterns with length 1; wherein the calculation method of the utility value u(ns) of the high-utility repetitive negative sequence candidate patterns with length 1 is shown in formula (4):

[0101] (4)

[0102] In formula (4), represents the high-utility repetitive negative sequence candidate pattern with length 1; and DB represents the encoded shopping sequence data table. represents the total utility value of all the commodity elements in the encoded shopping sequence data table; represents the sequence pattern formed by converting the negative commodity elements in the high-utility repetitive negative sequence candidate pattern with length 1 into positive commodity elements;

[0103] S2-2-2-2, obtaining high-utility repetitive negative sequence candidate patterns with length 2 by using the high-utility repetitive positive sequence patterns with length 1 and the high-utility repetitive negative sequence candidate patterns with length 1, and the steps include the following steps: connecting the high-utility repetitive positive sequence patterns with length 1 and the high-utility repetitive negative sequence candidate patterns with length 1 to obtain high-utility repetitive negative sequence candidate patterns with length 2;

[0104] S2-2-2-3, obtaining high-utility repetitive negative sequence candidate patterns with length n based on high-utility repetitive positive sequence patterns with length (n-1), wherein n is equal to the number of commodity elements in the shopping sequence, and n≥3; in this application, the way of obtaining high-utility repetitive negative sequence candidate patterns with length n based on high-utility repetitive positive sequence patterns with length (n-1) includes way one and way two; wherein the application generates high-utility repetitive negative sequence candidate patterns by using way one and way two;

[0105] Mode one: add a high-efficiency utility repetition negative sequence pattern with length of 1 at the end of a high-efficiency utility repetition positive sequence pattern with length of (n-1), to obtain a high-efficiency utility repetition negative sequence candidate pattern with length of n, n≥3;

[0106] Mode two: add a corresponding pattern according to a negative constraint condition at the end of a high-efficiency utility repetition negative sequence candidate pattern with length of (n-1), wherein the negative constraint condition refers to that two negative commodity elements are not allowed to exist continuously in the high-efficiency utility repetition negative sequence candidate pattern; if the last element of the high-efficiency utility repetition negative sequence candidate pattern is a negative commodity element, add a high-efficiency utility repetition positive sequence pattern with length of 1 at the end; if the last element of the high-efficiency utility repetition negative sequence candidate pattern is a positive commodity element, add a high-efficiency utility repetition negative sequence pattern with length of 1 or a high-efficiency utility repetition positive sequence pattern with length of 1 at the end, to obtain a high-efficiency utility repetition negative sequence candidate pattern with length of n, n≥3;

[0107] Repeat mode one and mode two until no new high-efficiency utility repetition negative sequence candidate pattern is generated; or until the length of the high-efficiency utility repetition negative sequence candidate pattern is greater than m+1, wherein m is the maximum length of the high-efficiency utility repetition positive sequence pattern;

[0108] Then, add all the high-efficiency utility repetition negative sequence candidate patterns to the high-efficiency utility repetition negative sequence candidate pattern list; and add the high-efficiency utility repetition negative sequence pattern with length of 1 to the high-efficiency utility repetition negative sequence pattern list.

[0109] S2-2-3, traverse all the high-efficiency utility repetition negative sequence candidate patterns with length greater than or equal to 2 in the high-efficiency utility repetition negative sequence pattern list, match the high-efficiency utility repetition negative sequence candidate pattern to the corresponding high-efficiency utility repetition positive sequence pattern, and all the matched high-efficiency utility repetition positive sequence patterns constitute a high-efficiency utility repetition positive sequence pattern set; calculate the utility value of the high-efficiency utility repetition negative sequence candidate pattern with length greater than or equal to 2 based on the utility value of the high-efficiency utility repetition positive sequence pattern matched with the high-efficiency utility repetition negative sequence candidate pattern; and take the high-efficiency utility repetition negative sequence candidate pattern with utility value greater than or equal to the minimum negative utility threshold ξ' as the high-efficiency utility repetition negative sequence pattern; and store all the high-efficiency utility repetition negative sequence patterns to the high-efficiency utility repetition negative sequence pattern list; wherein: the application, the utility value calculation method of the high-efficiency utility repetition negative sequence candidate pattern with length greater than or equal to 2, specifically includes the following steps:

[0110] (1) obtain a user ID set containing the high-efficiency utility repetition negative sequence candidate pattern;

[0111] When the high-efficiency utility repetition negative sequence candidate pattern contains only one negative commodity element, the obtaining mode of the user ID set in which the high-efficiency utility repetition negative sequence candidate pattern with length greater than or equal to 2 is located is as shown in formula (5):

[0112] (5)

[0113] In formula (5), MPS(ns) represents an ordered list of all positive commodity elements in a high-efficiency repeated negative sequence candidate mode with a length greater than or equal to 2 arranged in a sequence; {MPS(ns)} represents a user ID set containing all positive commodity elements in a high-efficiency repeated negative sequence candidate mode with a length greater than or equal to 2; and ns represents a high-efficiency repeated negative sequence candidate mode with a length greater than or equal to 2. represents a user ID set in which a high-efficiency repeated negative sequence candidate mode with a length greater than or equal to 2 is located; represents a sequence mode formed by converting negative commodity elements in a high-efficiency repeated negative sequence candidate mode with a length greater than or equal to 2 into positive commodity elements; satisfies the sequence mode , that is, a set containing all converted positive commodity elements.

[0114] When the number of negative commodity elements contained in a high-efficiency repeated negative sequence candidate mode is greater than or equal to 2, the user ID set in which a high-efficiency repeated negative sequence candidate mode with a length greater than or equal to 2 is located is obtained in the manner shown in formula (6):

[0115] (6)

[0116] In formula (6), represents that n sequence modes generated by a high-efficiency repeated negative sequence candidate mode with a length greater than or equal to 2 are combined with corresponding user ID sets, and after removing duplicate user IDs, a user ID set is obtained, in which the sequence mode represents a sequence mode formed by converting all negative commodity elements in into positive commodity elements, and m represents the number of negative commodity elements in a high-efficiency repeated negative sequence candidate mode with a length greater than or equal to 2; represents a subsequence formed by all positive commodity elements in a certain high-efficiency repeated negative sequence candidate mode with a length greater than or equal to 2 and any one negative commodity element contained in the high-efficiency repeated negative sequence candidate mode; MPS ( ns ) represents an ordered list of all positive commodity elements in a high-efficiency repeated negative sequence candidate mode with a length greater than or equal to 2 arranged in a sequence; { MPS (ns)} represents a user ID set containing all positive commodity elements in a high-efficiency repeated negative sequence candidate mode with a length greater than or equal to 2;

[0117] (2) Based on the user ID set, the hash table in the prefix tree is used to query the high-efficiency repeated positive sequence mode matched with the high-efficiency repeated negative sequence candidate mode, and the matched high-efficiency repeated positive sequence mode constitutes a high-efficiency repeated positive sequence mode set; specifically comprising the following steps:

[0118] According to the negative containing definition, that is, given a sequence and a negative sequence , then and only then:

[0119] and , , sequence contains negative sequence , .

[0120] For example, the high-efficiency repeated negative sequence candidate mode = <a¬bc>, = <ac>, the user 2's encoded shopping sequence <(a, 1)(a, 2)(b, 1)(b, 2)(c, 3)> is arranged in time order to form the sequence <aac> ， <ac>is <aac>a subsequence of the sequence satisfying the condition = <abc> ， <abc>not <acc>subsequence of the sequence of the form , under the condition that; thus, <a¬bc> <acc>.

[0121] Based on the definition of negative inclusion above, it can be seen that: Efficiently using candidate patterns of repeating negative sequences... Each occurrence corresponds to an ordered list of all positive item elements in a candidate pattern of repeated negative sequences of length greater than or equal to 2. Therefore, in this application, the utility value of a high-efficiency repeating negative sequence candidate pattern is calculated by summing the utility values ​​of high-efficiency repeating positive sequence patterns that match the high-efficiency repeating negative sequence candidate pattern, wherein the high-efficiency repeating negative sequence candidate pattern is matched with the corresponding high-efficiency repeating positive sequence pattern to filter out the high-efficiency repeating positive sequence pattern that meets the matching condition from the set of high-efficiency repeating positive sequence patterns;

[0122] In this application, the matching steps for matching the efficient repeating negative sequence candidate pattern to the corresponding efficient repeating positive sequence pattern are as follows: 1) The intersection of the user ID set containing the efficient repeating positive sequence pattern and the user ID set containing the efficient repeating negative sequence candidate pattern is not empty; 2) The efficient repeating positive sequence pattern contains an ordered list composed of all positive product elements in the efficient repeating negative sequence candidate pattern arranged in chronological order; then the efficient repeating positive sequence pattern is the efficient repeating positive sequence pattern that matches the efficient repeating negative sequence candidate pattern.

[0123] The efficient method for obtaining the set of repeating positive sequence patterns is shown in equation (7):

[0124] (7)

[0125] In equation (7), HURPSP' represents the efficient repeating positive sequence pattern that matches the efficient repeating negative sequence candidate pattern, {HURPSP'} represents the set of efficient repeating positive sequence patterns, and r represents any efficient repeating positive sequence pattern among all efficient repeating positive sequence patterns in the efficient repeating positive sequence pattern list; NS represents the set of all user IDs containing the efficient repeating positive sequence pattern r; NS represents the set of user IDs containing the efficient repeating negative sequence candidate pattern. MPS ( ns () represents an ordered list composed of all positive product elements in the candidate pattern of repeated negative sequences arranged in chronological order.

[0126] (3) The utility values ​​corresponding to the efficient repetitive positive sequence patterns that meet the matching conditions are accumulated to obtain the utility value of the efficient repetitive negative sequence candidate patterns, as shown in equation (8):

[0127] (8)

[0128] In equation (8), represents the utility value of the high-utility repetitive negative sequence candidate pattern whose length is greater than or equal to 2; represents the high-utility repetitive positive sequence pattern matched with the high-utility repetitive negative sequence candidate pattern represents the high-utility repetitive positive sequence pattern matched with the high-utility repetitive negative sequence candidate pattern represents the high-utility repetitive positive sequence pattern matched with the high-utility repetitive negative sequence candidate pattern represents the high-utility repetitive positive sequence pattern matched with the high-utility repetitive negative sequence candidate pattern

[0129] S2-3, the sliding window is moved, the encoded shopping sequence of the oldest data batch is deleted, the encoded shopping sequence of the newest data batch is included, and the encoded shopping sequence contained in the newest data batch is processed by the algorithm layer according to step S2-2; taking the first sliding of the sliding window as an example, after the sliding model is slid for the first time, the oldest data batch in the sliding window is deleted, and the newest data batch is included, that is, the data batch 1 is deleted, and the data batch 3 is included; the encoded shopping sequence contained in the data batch 3 is processed by the algorithm layer to mine the high-utility repetitive positive sequence pattern, and the high-utility repetitive positive sequence pattern obtained based on the newest data batch (that is, the data batch 3) and the utility list for storing the information related to the high-utility repetitive positive sequence pattern are stored into the prefix tree respectively; the high-utility repetitive positive sequence pattern is negatively transformed to obtain the high-utility repetitive negative sequence candidate pattern; then, the utility value of the high-utility repetitive negative sequence candidate pattern is calculated, and the utility value of the high-utility repetitive negative sequence candidate pattern is compared with the minimum negative utility threshold ξ', the high-utility repetitive negative sequence candidate pattern whose utility value is greater than or equal to the minimum negative utility threshold ξ' is taken as the high-utility repetitive negative sequence pattern; in step S2-3, if the high-utility repetitive positive sequence pattern mined based on the newest data batch (that is, the data batch 3) in the sliding window does not exist in the prefix tree, a new node is created in the prefix tree; if it already exists, the utility array is shifted one bit to the left, and the utility value of the high-utility repetitive positive sequence pattern in the newest data batch is added at the end of the utility array. In this embodiment, the window is dynamically slid to update, one newest data batch is automatically included, one oldest data batch is excluded, and the state of managing the newest data batch is always maintained.

[0130] S3, the application layer receives the high utility repetitive negative sequence pattern and its utility value output from the algorithm layer, adjusts the goods displayed on the platform interface in combination with the user ID, saves the adjustment record, generates a log file and transmits it to the data storage layer for archiving and saving. The application layer of the present application presents the complete information of the recommended goods in the goods display system, including the goods name, goods picture, goods introduction, goods price and other contents, to ensure that the user can fully understand the details of the goods and provide sufficient reference for the purchase decision. With the continuously generated user shopping data, the application layer continuously receives the new recommended results output by the algorithm layer, and dynamically adjusts the priority order of the recommended goods in the platform interface.

[0131] The existing high utility sequence pattern mining method is difficult to dynamically manage the continuously generated user shopping data stream in real time, resulting in that the mining result is prone to lag behind the latest user shopping behavior; in view of this problem, the present application designs a prefix tree structure for storing and maintaining high utility repetitive positive sequence patterns, realizing efficient and dynamic management of data; in addition, the existing high utility sequence pattern mining method is not reasonable in calculating the utility value of the high utility repetitive negative sequence candidate pattern, for example, the existing high utility sequence pattern mining method often simply sets the utility value of the negative goods element as 0, and only uses the utility value of the positive goods element in the high utility repetitive negative sequence candidate pattern to replace the utility value of the entire high utility repetitive negative sequence candidate pattern, while essentially ignoring the influence of the negative goods element on the actual utility of the high utility repetitive negative sequence candidate pattern, which leads to the inability to mine more accurate high utility repetitive negative sequence patterns, and further leads to the inaccuracy of the application layer in adjusting the priority order of the recommended goods in the platform interface based on the high utility repetitive negative sequence pattern, its utility value and its user ID, so that the recommended result cannot well meet the potential shopping demand of the user; therefore, the present application proposes a new high utility repetitive negative sequence candidate pattern utility value calculation method, specifically: for any high utility repetitive negative sequence candidate pattern, the utility value of the high utility repetitive positive sequence pattern matched therewith is accumulated to obtain the utility value of the high utility repetitive negative sequence candidate pattern, this way of obtaining the utility value of the high utility repetitive negative sequence candidate pattern does not directly discard the influence of the negative goods element, but indirectly quantifies the utility of the repetitive negative sequence candidate pattern through the logic of "excluding positive sequences containing negative goods elements", which is equivalent to starting from "the real preference of the user when selecting positive goods", and can effectively improve the accuracy of high utility repetitive negative sequence pattern mining. For example, for the negative candidate sequence < down jacket, ¬ The greater the utility value, the more likely the user will not purchase cotton shoes when purchasing down jackets in the current window, and the user is more inclined to match other goods. Therefore, the application layer can increase the display frequency of down jackets and other goods elements, and reduce the display frequency of down jackets and negative goods elements. In summary, the present application can realize real-time mining of high-utility repetitive negative sequence patterns based on continuously changing shopping data streams, so that the application layer can continuously reflect the latest user interest changes in the recommendation list, put truly attractive goods in the front, and make the recommendation results well meet the potential shopping needs of users.

[0132] Test:

[0133] The present application uses IBM data generator to generate test set for testing, and the acquisition website of IBM data generator is: https: / / www.philippe-fournier.viger.com / s pmf / index.php?link=datasets.php; the present application tests the mining efficiency of the method described in the present application by adjusting the parameters to control the data size of the test set, the complexity of the shopping sequence and the pattern distribution, wherein the adjusted parameters are C: the number of users; T: the average number of transactions of each user; S: the average number of goods per transaction; I: the length of the maximum pattern (i.e. the maximum size of the pattern obtained by pattern connection); DB: the number of sequence patterns; N: the number of different goods.

[0134] The present application generates the following four data sets through IBM data generator to test the proposed goods recommendation method, and the four data sets specifically include: DS1 data set is obtained under the condition that the parameters are set as C=6, T=1, S=4, I=1, DB=10K (10K=10000) and N=50; DS2 data set is obtained under the condition that the parameters are set as C=6, T=1, S=6, I=1, DB=10K (10K=10000) and N=50; DS3 data set is obtained under the condition that the parameters are set as C=8, T=1, S=4, I=1, DB=10K (10K=10000) and N=50; DS4 data set is obtained under the condition that the parameters are set as C=8, T=1, S=6, I=1, DB=10K (10K=10000) and N=50.

[0135] The existing high-utility negative sequence pattern mining method (from HUNSPM: An Efficient Algorithm for Mining High Utility Negative Sequential Patterns) is only applicable to static data sets. In order to make it have the ability to mine high-utility repetitive negative sequence patterns in real time, and make it comparable to the mining effect of the method described in the present application, the existing high-utility negative sequence pattern mining method is improved to obtain an improved high-utility negative sequence pattern mining method, referred to as HUNSPM Method, and then the HUNSPM Method and the method described in the present application are tested and compared on four data sets.

[0136] The improvement of the existing high-utility negative sequence pattern mining method is that the sliding window described in step S1 of the present application, the prefix tree structure in step S2 of the present application, and the technology disclosed in the specification of the existing patent CN120146968 B in paragraphs

[0074] -

[0028] are added to the step of obtaining high-utility positive sequence patterns disclosed in the existing high-utility negative sequence pattern mining method (from HUNSPM: An Efficient Algorithm for Mining High Utility Negative Sequential Patterns). This improvement enables the HUNSPM Method to have the ability to mine high-utility repetitive negative sequence candidate patterns and high-utility repetitive negative sequence patterns in real time.

[0137] The method described in the present application and the HUNSPM Method have the ability to mine high-utility repetitive negative sequence candidate patterns and high-utility repetitive negative sequence patterns in real time. 4 , the data batch is 1000, and the window size is set to 5, 6, 7, 8, and 9, respectively. The test results are shown in Figure 1 Figure 1 DS1 represents the DS1 data set.

[0138] The method described in the present application and the HUNSPM Method have the ability to mine high-utility repetitive negative sequence candidate patterns and high-utility repetitive negative sequence patterns in real time. 4 , the data batch is 1000, and the window size is set to 4, 5, 6, 7, and 8, respectively. The test results are shown in Figure 2 Figure 2 ​​DS2 represents the DS2 dataset; wherein the sliding window size is not set to 9 when testing the DS2 dataset, because the size of the DS2 dataset is only 7518 sequences, which can cause the window to not extract enough valid sequences, and the experimental results lose statistical significance;

[0139] The method and the HUNSPM The method in the present application has a minimum utility threshold value ξ of 2 x 10 4 , a data batch of 1000, and a window size of 5, 6, 7, 8, and 9, respectively, and the test results are as shown in Figure 3 , Figure 3 DS3 represents the DS3 dataset;

[0140] The method and the HUNSPM The method in the present application has a minimum utility threshold value ξ of 4 x 10 4 , a data batch of 1000, and a window size of 5, 6, 7, 8, and 9, respectively, and the test results are as shown in Figure 4 , Figure 4 DS4 represents the DS4 dataset;

[0141] The method and the HUNSPM The method in the present application has a minimum utility threshold value ξ of 2 x 10 4 , a window size of 8, and a data batch of 800, 850, 900, 950, and 1000, respectively, and the test results are as shown in Figure 5 , Figure 5 DS1 represents the DS1 dataset;

[0142] The method and the HUNSPM The method in the present application has a minimum utility threshold value ξ of 3 x 10 4 , a window size of 6, and a data batch of 800, 850, 900, 950, and 1000, respectively, and the test results are as shown in Figure 6 , Figure 6 DS2 represents the DS2 dataset;

[0143] The method and the HUNSPM The method in the present application has a minimum utility threshold value ξ of 2 x 10 4 , a window size of 8, and a data batch of 800, 850, 900, 950, and 1000, respectively, and the test results are as shown in Figure 7 , Figure 7 DS3 represents the DS3 dataset;

[0144] The method described in this application and HUNSPM The method has a minimum utility threshold ξ of 4 × 10 4 With a window size of 8 and data batch sizes of 800, 850, 900, 950, and 1000 respectively, the DS4 dataset was tested, and the results are as follows. Figure 8 As shown, Figure 8 In this context, DS4 represents the DS4 dataset;

[0145] The method described in this application and HUNSPM The method uses a data batch size of 1000, a window size of 6, and a minimum utility threshold ξ set to 2×10. 4 2.5×10 4 3×10 4 3.5×10 4 4×10 4 Under the given conditions, the DS1 dataset was tested, and the test results are as follows: Figure 9 As shown, Figure 9 In this context, DS1 represents the DS1 dataset;

[0146] The method described in this application and HUNSPM The method uses a data batch size of 1200, a window size of 5, and a minimum utility threshold ξ set to 2×10. 4 2.5×10 4 3×10 4 3.5×10 4 4×10 4 Under the given conditions, the DS2 dataset was tested, and the test results are as follows: Figure 10 As shown, Figure 10 In this context, DS2 represents the DS2 dataset;

[0147] The method described in this application and HUNSPM The method uses a data batch size of 1000, a window size of 5, and a minimum utility threshold ξ set to 2×10. 4 2.5×10 4 3×10 4 3.5×10 4 4×10 4 Under the given conditions, the DS3 dataset was tested, and the test results are as follows: Figure 11 As shown, Figure 11 In this context, DS3 represents the DS3 dataset;

[0148] The method described in this application and HUNSPM The method uses a data batch size of 1000, a window size of 5, and a minimum utility threshold ξ set to 4 × 10⁻⁶. 4 4.2×10 4 4.4×10 4 4.6×10 4 4.8×10 4 Under the given conditions, the DS4 dataset was tested, and the test results are as follows: Figure 12 As shown, Figure 12 In this application, DS4 represents the DS4 dataset. The minimum utility threshold ξ used when testing the DS4 dataset is different from that used for the DS1 dataset because the sequence lengths and sizes of the DS4 and DS1 datasets differ.

[0149] Figures 1 to 12 In this context, Our represents the method described in this application, HUNSPM HUNSPM Methods, for example, Figures 1 to 12 The term "efficient use of repeating negative sequence patterns" (HUNSPM) in this application refers to the efficient use of repeating negative sequence patterns mined by the method described in this application. ) indicates HUNSPM The method mines efficient repeating negative sequence patterns; efficient repeating negative sequence candidate patterns (0ur) represent efficient repeating negative sequence candidate patterns mined by the method described in this application; efficient repeating negative sequence candidate patterns (HUNSPM) ) indicates HUNSPM The method discovers efficient candidate patterns for repetitive negative sequences; Figures 9 to 12 In this context, the average mining time represents the runtime for generating an efficient negative sequence pattern. The average mining time reflects the performance of the method described in this application and HUNSPM. The mining efficiency of the method.

[0150] from Figures 1 to 4 It can be seen from this:

[0151] As the window size increases, the number of efficient candidate patterns for repeated negative sequences and the number of efficient patterns for repeated negative sequences mined by the method described in this application based on four datasets both show a significant upward trend.

[0152] from Figure 1 It can be seen that when the method described in this application mines based on the DS1 dataset, as the window size increases from 5 to 9, the number of efficient repeating negative sequence candidate patterns mined by the method in this application increases from 2669 to 8300, and the number of efficient repeating negative sequence patterns mined increases from 1849 to 5827; while HUNSPM When the window size is increased from 5 to 9, the number of high-utility repetitive negative subsequence candidate patterns and the number of high-utility repetitive negative subsequence patterns mined by the method based on the DS1 dataset remain at tens, regardless of the mining;

[0153] From Figure 2 It can be seen that when the method based on the DS2 dataset is mined, the number of high-utility repetitive negative subsequence candidate patterns mined by the method increases from 308 to 5197, and the number of high-utility repetitive negative subsequence patterns increases from 186 to 3663, as the window size increases from 4 to 8; while the HUNSPM The number of high-utility repetitive negative subsequence candidate patterns and the number of high-utility repetitive negative subsequence patterns mined by the method under the same test conditions remain between 24 and 130;

[0154] From Figure 3 It can be seen that when the method based on the DS3 dataset is mined, the number of high-utility repetitive negative subsequence candidate patterns mined by the method increases from 3772 to 8747, and the number of high-utility repetitive negative subsequence patterns increases from 2777 to 6769, as the window size increases from 5 to 9; while the HUNSPM The number of high-utility repetitive negative subsequence candidate patterns and the number of high-utility repetitive negative subsequence patterns mined by the method under the same test conditions remain between 30 and 70, still at a very low level.

[0155] From Figure 4 It can be seen that when the method based on the DS4 dataset is mined, the number of high-utility repetitive negative subsequence candidate patterns mined by the method increases from 889 to 6105, and the number of high-utility repetitive negative subsequence patterns increases from 592 to 4215, as the window size increases from 5 to 9; while the HUNSPM The number of high-utility repetitive negative subsequence candidate patterns and the number of high-utility repetitive negative subsequence patterns mined by the method under the same test conditions remain between 20 and 100, still at a very low level.

[0156] As can be seen from the above, the number of high-utility repetitive negative subsequence candidate patterns and the number of high-utility repetitive negative subsequence patterns mined by the method increase exponentially as the window size increases, while the HUNSPM The number of high-utility repetitive negative subsequence candidate patterns and the number of high-utility repetitive negative subsequence patterns mined by the method under the same conditions grow slowly, and the number of high-utility repetitive negative subsequence candidate patterns and the number of high-utility repetitive negative subsequence patterns mined remain at tens.

[0157] From Figures 5 to 8 It can be seen that:

[0158] As the data batch size gradually increases, the number of high utility repetitive negative subsequence candidate patterns and the number of high utility repetitive negative subsequence patterns mined by the method described in the present application based on the four data sets all show a stable and obvious upward trend.

[0159] From Figure 5 It can be seen that: when the method described in the present application is based on the DS1 data set mining, as the data batch size increases from 800 to 1000, the number of high utility repetitive negative subsequence candidate patterns mined by the method described in the present application increases from 5058 to 7235, and the number of high utility repetitive negative subsequence patterns increases from 3423 to 4914; while the HUNSPM The method based on the DS1 data set mining, when the data batch size increases from 800 to 1000, the number of high utility repetitive negative subsequence candidate patterns and high utility repetitive negative subsequence patterns mined is basically maintained between 60 to 100;

[0160] From Figure 6 It can be seen that: when the method described in the present application is based on the DS2 data set mining, as the data batch size increases from 800 to 1000, the number of high utility repetitive negative subsequence candidate patterns mined by the method described in the present application increases from 683 to 2060, and the number of high utility repetitive negative subsequence patterns increases from 470 to 1396; while the HUNSPM The method based on the DS2 data set mining, when the data batch size increases from 800 to 1000, the number of high utility repetitive negative subsequence candidate patterns and high utility repetitive negative subsequence patterns mined is basically maintained between 38 to 96;

[0161] From Figure 7 It can be seen that: when the method described in the present application is based on the DS3 data set mining, as the data batch size increases from 800 to 1000, the number of high utility repetitive negative subsequence candidate patterns mined by the method described in the present application increases from 5399 to 7530, and the number of high utility repetitive negative subsequence patterns increases from 3965 to 5583; while the HUNSPM The method based on the DS3 data set mining, when the data batch size increases from 800 to 1000, the number of high utility repetitive negative subsequence candidate patterns and high utility repetitive negative subsequence patterns mined is basically maintained between 40 to 120, still at a very low level.

[0162] From Figure 8 It can be seen that: when the method described in the present application is based on the DS4 data set mining, as the data batch size increases from 800 to 1000, the number of high utility repetitive negative subsequence candidate patterns mined by the method described in the present application increases from 2462 to 4572, and the number of high utility repetitive negative subsequence patterns increases from 1658 to 3163; while the HUNSPM The method can mine 20-100 high-utility repetitive negative sequence candidate patterns and high-utility repetitive negative sequence patterns under the same test conditions.

[0163] Therefore, the method can effectively utilize the sequence information in the data batch to comprehensively mine the high-utility repetitive negative sequence patterns.

[0164] From Figure 9 It can be seen that, when the method is based on the DS1 data set and the minimum utility threshold ξ increases from 2×10 4 to 4×10 4 , the number of high-utility repetitive negative sequence patterns mined by the method decreases from 3229 to 284; and the number of high-utility repetitive negative sequence patterns mined by the HUNSPM method decreases from 62 to 2. Obviously, in the range of the minimum utility threshold ξ from 2×10 4 to 4×10 4 , the number of high-utility repetitive negative sequence patterns mined by the method is still significantly higher than that of the HUNSPM method; in addition, when the method is based on the DS1 data set and the minimum utility threshold ξ increases from 2×10 4 to 4×10 4 , the average mining time of the method and the HUNSPM method shows an overall upward trend, but the average mining time of the method is significantly lower than that of the HUNSPM method, which indicates that the mining efficiency of the method is higher.

[0165] From Figure 10 it can be seen that, when the method is based on the DS2 data set and the minimum utility threshold ξ increases from 2×10 4 to 4×10 4 , the number of high-utility repetitive negative sequence patterns mined by the method decreases from 3144 to 204; and the number of high-utility repetitive negative sequence patterns mined by the HUNSPM method decreases from 62 to 4. Obviously, in the range of the minimum utility threshold ξ from 2×10 4 to 4×10 4 , the number of high-utility repetitive negative sequence patterns mined by the method is still significantly higher than that of the HUNSPM method; in addition, when the method is based on the DS2 data set and the minimum utility threshold ξ increases from 2×10 4 to 4×10 4 The method and HUNSPM The average mining time of the method shows an overall upward trend, wherein the minimum utility threshold ξ is increased from 2 x 10 4 to 3.5 x 10 4 The method and HUNSPM The average mining time of the method is gradually approaching, and after the minimum utility threshold ξ is greater than 3.5 x 10 4 The average mining time of the method is significantly lower than that of the HUNSPM The method, which shows that the mining efficiency of the method is higher;

[0166] From Figure 11 It can be seen that: when the method based on the DS3 data set is mined, as the minimum utility threshold ξ is increased from 2 x 10 4 to 4 x 10 4 , the number of high-utility repetitive negative sequential patterns mined by the method is reduced from 2777 to 225; while the number of high-utility repetitive negative sequential patterns mined by the HUNSPM method is reduced from 48 to 4, obviously, in the range of the minimum utility threshold ξ from 2 x 10 4 to 4 x 10 4 , the number of high-utility repetitive negative sequential patterns mined by the method is still significantly higher than that of the HUNSPM method; in addition, when the method based on the DS3 data set is mined, as the minimum utility threshold ξ is increased from 2 x 10 4 to 4 x 10 4 , the average mining time of the method and the HUNSPM method shows an overall upward trend, and the average mining time of the method is significantly lower than that of the HUNSPM method, which shows that the mining efficiency of the method is higher;

[0167] From Figure 12 It can be seen that: when the method based on the DS4 data set is mined, as the minimum utility threshold ξ is increased from 4 x 10 4 to 4.8 x 10 4 , the number of high-utility repetitive negative sequential patterns mined by the method is reduced from 592 to 199; while the number of high-utility repetitive negative sequential patterns mined by the HUNSPM method is reduced from 14 to 2, obviously, in the range of the minimum utility threshold ξ from 4 x 10 4 to 4.8 x 10 4 The number of high-utility repetitive negative sequence patterns mined by the method of the present application is still significantly higher than that of HUNSPM The average mining time of the method of the present application is in an overall upward trend, and the average mining time of the method of the present application is significantly lower than that of HUNSPM 4 when the method of the present application is based on the DS4 dataset 4 The average mining time of the method of the present application is in an overall upward trend, and the average mining time of the method of the present application is significantly lower than that of HUNSPM The average mining time of the method of the present application is in an overall upward trend, and the average mining time of the method of the present application is significantly lower than that of HUNSPM The average mining time of the method of the present application is in an overall upward trend, and the average mining time of the method of the present application is significantly lower than that of HUNSPM The advantage of the average mining time of the method of the present application is more significant, which indicates that the mining efficiency of the method of the present application is higher.

[0168] Obviously, the method of the present application has higher mining efficiency while ensuring the mining effect, and can generate high-utility repetitive negative sequence patterns more quickly, and can better meet the needs of quickly analyzing user purchase behavior and timely recommending goods in scenarios such as user shopping which have higher real-time requirements.< / acc> < / acc> < / abc> < / abc> < / aac> < / ac> < / aac> < / ac> < / ab> < / ab> < / ab> < / abc> < / bc> < / ac> < / ab> < / ab> < / ab> < / abc> < / bc> < / ac> < / ab>

Claims

1. A high utility repetitive negative sequence pattern mining based commercial real-time recommendation method, characterized in that: The method comprises the following steps: S1, collecting shopping data flow, data cleaning and classification, storing the obtained shopping sequence of the user and the corresponding user ID in the user shopping sequence data table in chronological order; Divide the shopping sequence in the data table into data batches according to time; adopt a sliding window to continuously slide on the data batches with a preset window size; S2, sequentially process the encoded shopping sequence contained in the data batches in the sliding window according to time sequence, move the sliding window, delete a number of old data batches, and incorporate a number of new data batches, process the encoded shopping sequence contained in the new data batches, and obtain high-efficiency repeated negative sequence patterns; wherein the high-efficiency repeated negative sequence pattern is obtained based on a high-efficiency repeated negative sequence candidate pattern, and the utility value of the high-efficiency repeated negative sequence candidate pattern with a length greater than or equal to 2 is obtained in the following manner: matching the high-efficiency repeated negative sequence candidate pattern with a corresponding high-efficiency repeated positive sequence pattern, and accumulating the utility value of the high-efficiency repeated positive sequence pattern; The matching step of matching the high-efficiency repeated negative sequence candidate pattern to the corresponding high-efficiency repeated positive sequence pattern is as follows: 1) the intersection of the user ID set where the high-efficiency repeated positive sequence pattern is located and the user ID set where the high-efficiency repeated negative sequence candidate pattern is located is not empty; 2) the high-efficiency repeated positive sequence pattern contains an ordered list composed of all positive commodity elements in the high-efficiency repeated negative sequence candidate pattern in the order of sequence; then the high-efficiency repeated positive sequence pattern is the high-efficiency repeated positive sequence pattern matched with the high-efficiency repeated negative sequence candidate pattern; S3, adjusting the platform interface display of the commodity based on the high-efficiency repeated negative sequence pattern, the utility value and the corresponding user ID. 2.The high-efficiency commodity real-time recommendation method with repeated negative sequence pattern mining of utility according to claim 1, characterized in that: Step S2 specifically comprises the following steps: S2-1, performing an encoding operation on the shopping sequence in the user shopping sequence data table and the commodity element in the element data table to obtain an encoded element data table and an encoded user shopping sequence data table; S2-2, sequentially mining the high-efficiency repeated positive sequence pattern from the encoded shopping sequence contained in each data batch in the sliding window according to time sequence, and storing the high-efficiency repeated positive sequence pattern and the utility list for storing the related information of the high-efficiency repeated positive sequence pattern in the prefix tree; performing negative transformation on the high-efficiency repeated positive sequence pattern to obtain a high-efficiency repeated negative sequence candidate pattern; calculating the utility value of the high-efficiency repeated negative sequence candidate pattern, and taking the high-efficiency repeated negative sequence candidate pattern with an utility value greater than or equal to the minimum negative utility threshold as the high-efficiency repeated negative sequence pattern; S2-3, moving the sliding window, deleting the encoded shopping sequence of a number of old data batches, and incorporating a number of new data batches; using the algorithm layer to process the encoded shopping sequence contained in the new data batches according to step S2-2. 3.The method of claim 2, wherein: Step S2-2 specifically comprises the following steps: S2-2-1, using the algorithm layer to sequentially mine the high-efficiency repeated positive sequence pattern from the shopping sequence contained in each data batch in the sliding window according to time sequence; S2-2-2, obtaining high utility repetitive negative sequence candidate patterns based on high utility repetitive positive sequence patterns; S2-2-3, traversing all high utility repetitive negative sequence candidate patterns with length greater than or equal to 2 in the high utility repetitive negative sequence pattern list, matching the high utility repetitive negative sequence candidate patterns to corresponding high utility repetitive positive sequence patterns, constructing a high utility repetitive positive sequence pattern set with all the matched high utility repetitive positive sequence patterns, calculating the utility value of the high utility repetitive negative sequence candidate patterns with length greater than or equal to 2 based on the utility value of the high utility repetitive positive sequence patterns matched with the high utility repetitive negative sequence candidate patterns, and taking the high utility repetitive negative sequence candidate patterns with utility value greater than or equal to the minimum negative utility threshold as high utility repetitive negative sequence patterns. 4.The method of claim 3, wherein: Step S2-2-3 specifically includes the following steps: (1) obtaining a user ID set containing the high utility repetitive negative sequence candidate patterns; (2) based on the user ID set, querying the high utility repetitive positive sequence patterns matched with the high utility repetitive negative sequence candidate patterns according to the hash table in the prefix tree; the matched high utility repetitive positive sequence patterns constitute a high utility repetitive positive sequence pattern set; (3) accumulating the utility values of the high utility repetitive positive sequence patterns satisfying the matching conditions to obtain the utility value of the high utility repetitive negative sequence candidate patterns.

5. The method of claim 3, wherein the method further comprises: Step S2-2-2 specifically includes the following steps: S2-2-2-1, obtaining high utility repetitive negative sequence candidate patterns with length 1 using high utility repetitive positive sequence patterns with length 1, including the following steps: performing negative transformation on the high utility repetitive positive sequence patterns with length 1 to obtain high utility repetitive negative sequence candidate patterns with length 1; calculating the utility value of the high utility repetitive negative sequence candidate patterns with length 1, and taking the high utility repetitive negative sequence candidate patterns with utility value greater than or equal to the minimum negative utility threshold as high utility repetitive negative sequence patterns with length 1; S2-2-2-2, connecting the high utility repetitive positive sequence patterns with length 1 with the high utility repetitive negative sequence candidate patterns with length 1 to obtain high utility repetitive negative sequence candidate patterns with length 2; S2-2-2-3, obtaining high utility repetitive negative sequence candidate patterns with length n based on high utility repetitive positive sequence patterns with length (n-1), wherein n is equal to the number of commodity elements in the shopping sequence, and n≥3.

6. The method of claim 5, wherein the method further comprises: In S2-2-2-3, the way of obtaining high utility repetitive negative sequence candidate patterns with length n based on high utility repetitive positive sequence patterns with length (n-1) includes way one and way two; Way one, adding high utility repetitive negative sequence patterns with length 1 at the end of high utility repetitive positive sequence patterns with length (n-1); Mode two, at the end of the high-efficiency negative sequence candidate mode with a length of (n-1), add the corresponding mode according to the negative constraint condition, wherein the negative constraint condition refers to the high-efficiency negative sequence candidate mode does not allow the existence of two consecutive negative commodity elements; if the last element of the high-efficiency negative sequence candidate mode is a negative commodity element, a high-efficiency positive sequence mode with a length of 1 is added at the end; if the last element of the high-efficiency negative sequence candidate mode is a positive commodity element, a high-efficiency negative sequence mode with a length of 1 or a high-efficiency positive sequence mode with a length of 1 is added at the end; Repeat mode one and mode two until no new high-efficiency negative sequence candidate mode is generated; or until the length of the high-efficiency negative sequence candidate mode is greater than m+1, wherein m is the maximum length of the high-efficiency positive sequence mode.

7. A high utility repetitive negative sequence pattern mining based commercial real-time recommendation system characterized by: The high-efficiency negative sequence pattern mining real-time commodity recommendation system is used to realize the steps of the high-efficiency negative sequence pattern mining real-time commodity recommendation method of any one of claims 1 to 6, and the high-efficiency negative sequence pattern mining real-time commodity recommendation system comprises a data storage layer, an algorithm layer and an application layer, wherein The data storage layer is used to collect shopping data streams, data cleaning and classification, obtain user shopping sequences, commodity categories ITEM and their utility values; store the user shopping sequences and their corresponding user IDs in the user shopping sequence data table in chronological order; divide the shopping sequences in the user shopping sequence data table into different data batches in chronological order, and each data batch contains the same number of shopping sequences; use a sliding window to continuously slide on the data batches with a preset window size; The algorithm layer is used to process the encoded shopping sequences contained in the data batches of the sliding window in chronological order to obtain high-efficiency negative sequence patterns; move the sliding window, delete some old data batches, include some new data batches, and process the encoded shopping sequences contained in the new data batches to obtain high-efficiency negative sequence candidate patterns and their utility values; the high-efficiency negative sequence candidate pattern utility value calculation method is as follows: for any high-efficiency negative sequence candidate pattern, the utility value of the high-efficiency positive sequence pattern matched therewith is added to obtain the utility value of the high-efficiency negative sequence candidate pattern; The application layer adjusts the commodities displayed on the platform interface based on the high-efficiency negative sequence patterns and their utility values, and combines the user IDs; at the same time, the adjustment record is saved, a log file is generated and transmitted to the data storage layer for archival storage.

Images