CN112100483B - Association rule recommendation method fusing user interest weights - Google Patents

Abstract

The invention discloses an association rule recommendation method for fusing user interest weights, which is implemented according to the following steps: step 1, constructing a scoring matrix of a user and a custom item; step 2, calculating the interest weight of the user on the customized content; step 3, constructing a UIFP-tree according to the user customized content interest weight; and 4, carrying out frequent pattern mining on the UIFP-tree to obtain frequent item combinations, and recommending the frequent item combinations to a user. According to the method and the device for obtaining the customized content rules, the user interest weight is fused, the obtained customized content rules retain the user preference more, a large number of nonsensical rules are filtered, the data mining efficiency is improved, and the finally obtained strong association rules have practical application values.

Description

Association rule recommendation method fusing user interest weights

Technical Field

The invention belongs to the technical field of customized content association rule recommendation, and relates to an association rule recommendation method integrating user interest weights.

Background

In recent years, association rule mining algorithms are widely applied to the recommendation field, and by analyzing data, association between the data is found, and recommendation is realized by using association rules. The custom-made clothing content recommendation is based on clothing styles, and custom-made content combination recommendation with specific rules is mined and recommended to users according to relevant order information of certain clothing styles. The conventional FP-growth association rule algorithm only uses the minimum support degree as a filtering and screening standard of the item set in the transaction database, and in the recommendation process, the support degree and the confidence degree are two thresholds which are frequently applied in the conventional association rule algorithm, but sometimes the strong association rule with high support degree and high confidence degree has no practical application value, particularly in the e-commerce field, the preference rule obtained according to the subjective score of the user is more valuable, and the preference is easily filtered out by using the conventional FP-growth association rule algorithm.

Disclosure of Invention

The invention aims to provide an association rule recommendation method integrating user interest weights, which solves the problems of low association rule recommendation efficiency and inaccurate recommendation result in the prior art.

The technical scheme adopted by the invention is that the association rule recommendation method for fusing the user interest weight is implemented according to the following steps:

step 1, constructing a scoring matrix of a user and a custom item;

step 2, calculating the interest weight of the user on the customized content;

step 3, constructing a UIFP-tree according to the user customized content interest weight;

and 4, carrying out frequent pattern mining on the UIFP-tree to obtain frequent item combinations, and recommending the frequent item combinations to a user.

The invention is also characterized in that:

the step 1 is specifically implemented according to the following steps:

step 1.1, setting K items in the preparation items, wherein the customization content is expressed as content, and the customization combination is expressed as item= { content ₁ ,content ₂ ,......,content _k }；

Step 1.2, n user custom combinations and m users are provided, wherein the user custom combinations are expressed as I= { item ₁ ,item ₂ ,......,item _n User is denoted as u= { user } ₁ ,user ₂ ,......,user _m }；

And 1.3, constructing a scoring matrix of the user and the custom item according to the historical scoring of the custom item by the user.

Step 1.3 is specifically that the common 5-star user scores are preprocessed, all 0-3 scores are set to 0, the current custom combination is disliked, all 4-5 scores are set to 1, the current custom combination is liked, and the scoring matrix of the user and the custom items is expressed as follows:

wherein ,R_i,j (1.ltoreq.i.ltoreq.m, 1.ltoreq.j.ltoreq.n) represents a user _i Given a custom combined item _j Is a score of (2).

In step 2, specifically, the content of the set system includes n choices, content= { option ₁ ,option ₂ ,......,option _n If user _i Combining items to custom content _j Is 1, item _j The option corresponding to each custom content score is 1, and the custom content interest weight is:

wherein ,the number of users with a score value of 1 for the customized content is represented, and |u| represents the total number of users.

The step 3 is specifically implemented according to the following steps:

step 3.1, scanning a transaction database D, and solving a frequent item set F meeting the minimum support count min_sup in the D;

step 3.2, all the user-customized content interest degree weights W meeting the user-customized content interest degree weights are obtained in the set F _i Greater than or equal to min_u _option Is set W;

step 3.3, sorting the set W in descending order according to the support degree, marking the result as a table L, creating UIFP-tree root nodes, marking the root nodes as roots, and marking the root nodes as NULL;

and 3.4, inserting nodes under the root node to obtain the UIFP-tree.

Step 3.4 is specifically implemented according to the following steps:

step 3.4.1, min_u is satisfied in each transaction _option Frequent items are ordered in the order of L, and the ordered list is [ p|P ]]Where P is the first element and P is a table of the remaining elements;

step 3.4.2, call insert_tree ([ p|P ], T).

Step 4 is specifically implemented according to the following steps:

step 4.1, obtaining a condition mode base, namely obtaining a prefix path for each node from the bottom to the top from the leaf node of the UIFP-tree, and replacing the support count with the current node support count;

step 4.2, constructing a conditional FP tree by using a conditional mode base, recursively calling a tree structure, deleting nodes smaller than the minimum support, and if the tree structure of a single path is finally presented, directly listing all combinations, namely frequent item combinations; if the tree structure of the non-single path is finally presented, continuing to call the tree structure until the tree structure of the single path is formed;

step 4.3, repeating steps 4.1 and 4.2 until only one element is contained in the tree, i.e. the frequent item combinations that are eventually recommended to the user.

The beneficial effects of the invention are as follows: compared with the traditional FP-growth algorithm, the method has the advantages that the user preference is reserved more in the obtained custom-made content rule by fusing the user interest weight, a large number of nonsensical rules are filtered, the data mining efficiency is improved, and the finally obtained strong association rule has practical application value

Drawings

FIG. 1 is a frequent pattern tree diagram filtered by minimum support and user interest weight in an association rule recommendation method integrating user interest weight;

FIG. 2 is a graph of experimental results of comparison of algorithm running time of the UIFP-growth algorithm adopted by the invention and the conventional FP-growth algorithm under different minimum supporters;

FIG. 3 is a graph of comparing the number of excavation results under different minimum supporters of the UIFP-growth algorithm adopted by the invention with the conventional FP-growth algorithm;

FIG. 4 is a graph showing the comparison of the running time of the UIFP-growth algorithm and the running time of the conventional FP-growth algorithm under different numbers of transactions.

Detailed Description

The invention will be described in detail below with reference to the drawings and the detailed description.

The invention discloses an association rule recommendation method for fusing user interest weights, which is implemented according to the following steps:

step 1, constructing a scoring matrix of a user and a custom item; the method is implemented according to the following steps:

step 1.1, all customizable contents of a certain clothing style have k items, and the customized contents comprise: content such as fabric, color, pattern, and the like, and customized content is expressed as content, and customized combination is expressed as item= { content ₁ ,content ₂ ,......,content _k }；

Step 1.2, assuming that a certain garment style has n custom-made combinations of users and m users, the custom-made combinations are expressed as i= { item ₁ ,item ₂ ,......,item _n User is denoted as u= { user } ₁ ,user ₂ ,......,user _m }；

Step 1.3, constructing a scoring matrix of the user and the custom item according to the historical scoring of the custom item by the user; preprocessing the commonly used 5-star user scores, setting 0-3 score to 0 to represent dislike of the current customization combination, setting 4-5 score to 1 to represent like of the current customization combination, and expressing the scoring matrix of the user and the customization items as follows:

wherein ,R_i,j (1.ltoreq.i.ltoreq.m, 1.ltoreq.j.ltoreq.n) represents a user _i Given a custom combined item _j Is a score of (2);

step 2, calculatingInterest weight of the user for the customized content; let the custom content contain n choices, content= { option ₁ ,option ₂ ,......,option _n If user _i Combining items to custom content _j Is 1, item _j The option corresponding to each custom content score is 1, and the user-custom content interest scores are shown in table 1;

TABLE 1 user-customized content interest scoring

The custom content option weights are set to find items that can better describe the user's interests, the custom content interest weights are:

wherein ,user number with score value 1 of the customized content is represented, and U represents total number of users; the higher W indicates that the user is interested in the content, the lower W indicates that the user is less interested in the content, especially when +.>When the content is selected by all users, the content has no difference in interest level of all users;

step 3, constructing a UIFP-tree according to the user customized content interest weight; the method is implemented according to the following steps:

step 3.1, scanning a transaction database D, and solving a frequent item set F meeting the minimum support count min_sup in the D;

step 3.2, all the user-customized content interest degree weights W meeting the user-customized content interest degree weights are obtained in the set F _i Greater than or equal to min_u _option Is set W;

step 3.3, sorting the set W in descending order according to the support degree, marking the result as a table L, creating UIFP-tree root nodes, marking the root nodes as roots, and marking the root nodes as NULL;

step 3.4, inserting nodes under the root node to obtain a UIFP-tree; the method is implemented according to the following steps:

step 3.4.1, min_u is satisfied in each transaction _option Frequent items are ordered in the order of L, and the ordered list is [ p|P ]]Where P is the first element and P is a table of the remaining elements;

step 3.4.2, calling an insert_tree ([ p|P ], T);

the specific steps of the insert_tree are as follows:

1) If T has children N such that n. term = p. term, the count of N is incremented by 1;

2) Otherwise, creating a new node N, setting the count of the new node N to be 1, and linking the new node N to the father node T;

3) Linking the node chain of N to the nodes with the same item names through a node chain structure;

4) Recursively invoking an insert_tree ([ q|Q ], N) if P is not null, where Q is one element in P and Q is a table of the remaining elements;

step 4, frequent pattern mining is carried out on the UIFP-tree, frequent item combinations are obtained, and the frequent item combinations are recommended to users; the method is implemented according to the following steps:

step 4.1, obtaining a condition mode base, namely obtaining a prefix path for each node from the bottom to the top from the leaf node of the UIFP-tree, and replacing the support count with the current node support count;

step 4.2, constructing a conditional FP tree by using a conditional mode base, recursively calling a tree structure, deleting nodes smaller than the minimum support, and if the tree structure of a single path is finally presented, directly listing all combinations, namely frequent item combinations; if the tree structure of the non-single path is finally presented, continuing to call the tree structure until the tree structure of the single path is formed;

step 4.3, repeating steps 4.1 and 4.2 until only one element is contained in the tree, i.e. the frequent item combinations that are eventually recommended to the user.

Examples

Taking suit customization as an example, the application of the method in the recommendation of the suit customized content is analyzed in detail, and according to the research on the customized content of a suit of a certain style, the names and parameters of the customized content are shown in table 2:

table 2 customized content for western style clothes

According to the study on the customized content of a suit of a certain style, part of experimental data is extracted for specific analysis, and the data set is shown in table 3:

table 3 dataset

Let min_sup=3, min_u be the minimum support count _option After parameterizing the data in the dataset, and scanning the database once, the results are shown in table 4:

TABLE 4 transaction database D and filtered ordered entries

The process of mining frequent items for the transactional database D is as follows:

(1) Deriving a frequent item set;

scanning the database for the 1 st time to derive a frequent 1-term set f= [ L1:5, n1:5, e1:3, F1:3, d1:3];

(2) Solving all item sets meeting the minimum custom content interest degree weight;

for a collectionFind out that all custom content interests weight is greater than min_u _option And ordering the items according to the support degree in a descending order to obtain an ordered set L. The interestingness weights of the items in F are F= [ L1:3/7, N1:4/7, E1:3/7, F1:3/7, D1:2/7]Since the interestingness weight of D1 is lower than the threshold value min_u _option Thus, D1 is cut off, then l= [ L1:5, n1:5, e1:3, f1:3]The set of frequent items in each transaction that meet the minimum interestingness weight is shown in table 5:

table 5 sets of frequent items meeting minimum interestingness weights

(1) Constructing a UIFP-tree;

creating a root node of the tree, denoted as root, with a value of "NULL", scanning the transaction database D again, creating a branch for each transaction, for the first transaction "T1: L1, N1, E1, F1" in the transaction database, creating a first branch < (L1: 1), (N1: 1), (E1: 1), (F1: 1) >, for the second transaction, counting each node by 1 because it is identical to the first transaction, counting the prefix path < L1> for the third transaction, and if there is a new node, creating only new node links at the back, and so on, obtaining a frequent pattern tree as shown in fig. 2, (2) UIFP-growth excavation;

according to the FP-growth algorithm, the following mining is performed on the UIFP-tree shown in FIG. 2: firstly, searching a head list according to a bottom-up sequence, finding the last item F1 in the head list, and finding two branches where the item F1 is located according to a node chain of the F1: < L1:3, N1:2, E1:2, F1:2> and < L1:3, E1:1, F1:1>, it can be seen that the prefix paths of F1 are < L1, E1, N1:2> and < L1, E1:1>, which constitute the conditional pattern base of F1, for such a sub-database, a conditional frequent pattern tree is built, the conditional frequent pattern tree of F1 contains only nodes < L1, E1>, and no nodes < N1>, and the nodes N1 are filtered out because the support count of the nodes N1 is less than the minimum support count of 3, the resulting frequent pattern set is { (F1:3), (L1, F1:3), (E1, F1:3), (L1, E1, F1:3) }, and so on, and the mining results for the frequent pattern tree are shown in Table 6.

TABLE 6 UIFP-growth mining results

The experimental data of the invention come from a background database of a nogara clothing customization mall, all user customization combinations under a certain business suit style are derived from the background database, and 2000 pieces of customization information are taken as a test data set; the hardware environment of the experiment is as follows: windows10, memory 8G, CPUi7-8550, software environment is: python version 3.6.3, pyCharm integration environment; in order to verify the effectiveness of the UIFP-growth algorithm, the present invention compares with the FP-growth algorithm in three ways:

1) The results of the algorithm running time comparison experiment under different minimum supporters are shown in fig. 2;

as can be seen from FIG. 2, under different minimum supporters, the UIFP-growth algorithm is shorter than the running time of the FP-growth algorithm, the mining efficiency is higher, the running time of the FP-growth algorithm gradually decreases with the increase of the minimum supporters, and the running time of the UIFP-growth algorithm is customized with the minimum supporters and given user-customized content interest weight threshold value min_u _option Gradually decreasing as the minimum support and min_u _option Above 0.3, there is no significant change in the run time of the algorithm, therefore, when the minimum support and min_u _option At 0.3, the mining result is more efficient and the result is more meaningful;

2) The comparison experiment of the number of the digging results under different minimum supporters is shown in fig. 3;

as can be seen from FIG. 3, the number of mining results of the UIFP-growth algorithm under the same minimum support degree is smaller than that of the FP-growth algorithm, and the number of mining results of the FP-growth algorithm is gradually reduced along with the increase of the minimum support degree, and the UIFP-growth is supported along with the given minimum support degreeDegree and user customized content interest weight min_u _option The mining results are gradually reduced, and the mining results are less than the FP-growth algorithm under different minimum supporters, because the UIFP-growth algorithm filters the interestingness weight on the basis of the minimum supporters in the mining process, and certain meaningless rules are reduced, so that the mining results are more meaningful;

3) Under the condition that the minimum support degree and the custom content interest weight are both 0.3, the UIFP-growth algorithm and the FP-growth algorithm run time pairs under different numbers of transactions, such as shown in FIG. 4;

as can be seen from FIG. 4, the running time of the UIFP-growth algorithm is lower than that of the FP-growth algorithm when the mining is performed in the same transaction number, and the running time of the FP-growth algorithm and the running time of the UIFP-growth algorithm are in an ascending trend along with the increase of the database transaction number, but the growth speed of the UIFP-growth algorithm is slower, so that the mining time of the UIFP-growth algorithm is shorter and the performance is better compared with that of the FP-growth algorithm.

Claims (3)

1. The association rule recommendation method integrating the user interest weight is characterized by comprising the following steps of:

step 1, constructing a scoring matrix of a user and a custom item; the method is implemented according to the following steps:

step 1.1, setting K items in the preparation items, wherein the customization content is expressed as content, and the customization combination is expressed as item= { content ₁ ,content ₂ ,......,content _k }；

Step 1.2, n user custom combinations and m users are provided, wherein the user custom combinations are expressed as I= { item ₁ ,item ₂ ,......,item _n User is denoted as u= { user } ₁ ,user ₂ ,......,user _m }；

Step 1.3, constructing a scoring matrix of the user and the custom item according to the historical scoring of the custom item by the user; specifically, the common 5-star user score is preprocessed, all 0-3 scores are set to 0, the current custom combination is disliked, all 4-5 scores are set to 1, the current custom combination is liked, and the scoring matrix of the user and the custom item is expressed as:

wherein ,R_i,j I is more than or equal to 1 and less than or equal to m, j is more than or equal to 1 and less than or equal to n, and represents a user _i Given a custom combined item _j Is a score of (2);

step 2, calculating the interest weight of the user on the customized content; specifically, the set content includes n choices, content= { option ₁ ,option ₂ ,......,option _n If user _i Combining items to custom content _j Is 1, item _j The option corresponding to each custom content score is 1, and the custom content interest weight is:

wherein ,the number of users with the score value of 1 representing the customized content, and U represents the total number of users;

step 3, constructing a UIFP-tree according to the user customized content interest weight; the method is implemented according to the following steps:

step 3.1, scanning a transaction database D, and solving a frequent item set F meeting the minimum support count min_sup in the D;

step 3.2, all the user-customized content interest degree weights W meeting the user-customized content interest degree weights are obtained in the set F _i Greater than or equal to a threshold value min_u _option Is set W;

step 3.3, sorting the set W in descending order according to the support degree, marking the result as a table L, creating an FP-tree root node, marking the root node as a root, and marking the root node as NULL;

step 3.4, inserting nodes under the root node to obtain a UIFP-tree;

and 4, carrying out frequent pattern mining on the UIFP-tree to obtain frequent item combinations, and recommending the frequent item combinations to a user.

2. The association rule recommendation method for fusing user interest weights as claimed in claim 1, wherein the step 3.4 is specifically implemented as follows:

step 3.4.1, min_u is satisfied in each transaction _option Frequent items are ordered in the order of L, and the ordered list is [ p|P ]]Where P is the first element and P is a table of the remaining elements;

step 3.4.2, call insert_tree ([ p|P ], T).

3. The association rule recommendation method for fusing user interest weights as claimed in claim 1, wherein the step 4 is specifically implemented according to the following steps:

step 4.1, obtaining a condition mode base, namely obtaining a prefix path for each node from the bottom to the top from the leaf node of the UIFP-tree, and replacing the support count with the current node support count;

step 4.2, constructing a conditional FP tree by using a conditional mode base, recursively calling a tree structure, deleting nodes smaller than the minimum support, and if the tree structure of a single path is finally presented, directly listing all combinations, namely frequent item combinations; if the tree structure of the non-single path is finally presented, continuing to call the tree structure until the tree structure of the single path is formed;

step 4.3, repeating steps 4.1 and 4.2 until only one element is contained in the tree, i.e. the frequent item combinations that are eventually recommended to the user.