CN104809626A - Customized commodity recommending method based on user credit assessment - Google Patents

Abstract

The invention discloses a customized commodity recommending method based on user credit assessment. The method comprises the steps that 1) a matrix M is constructed according to the collected user-commodity order dealing number of times and order cancelling number of times; 2) total commodity order dealing sum of money and total order cancelling sum of money of each user are calculated according to the matrix M so that the initial credit value of the users is obtained; 3) total commodity dealing number of times and total canceling number of times of the users are calculated according to the matrix M so that the credit index of the users is obtained; 4) the credit value of the users is calculated according to the initial credit value and the credit index; 5) dealing commodities of each user are classified according to the matrix M, and the type of the main purchasing body contained by each user is confirmed; and 6) the users of the same credit value section are selected as for any one user A, and if the characteristic of the commodity purchased by the user B belongs to the characteristic corresponding to the type of the main purchasing body in the user A, the commodity is recommended to the user A. More accurate and customized commodity recommending service is realized by the customized commodity recommending method based on user credit assessment.

Description

A kind of individual commodity recommendation method based on user credit assessment

Technical field

The present invention relates to a kind of individual commodity recommendation method based on user credit assessment, belong to e-commerce field.

Background technology

In recent years, the change of people's life pattern at any time and improving constantly of quality of life, the type of merchandize that e-commerce initiative and the sharp increase of Transaction Information amount, businessman provide and quantity also grow with each passing day.The merchandise news of magnanimity in e-commerce system, consumer is difficult to the commodity fast and effeciently picked out required for him.How choose consumer and must continue as them after commodity and recommend relevantly to have it valuable and meet the commodity of its preference, how to promote the loyalty of consumer to businessman, how meeting the problems such as the more personalized demand of user becomes the key issue that e-commerce development must solve.

At present a lot of e-commerce ventures and many researchists propose many service recommendation methods, mainly comprise content-based, Knowledge based engineering, rule-based, based on effectiveness and based on the recommend method of correlation rule.In recent years, collaborative filtering method achieves good achievement in business application, such as: Amazon, CDNow, the technology that MovieFinder etc. have employed collaborative filtering is improved service quality, but the shortcoming of collaborative filtering has: (1) user is very sparse to the evaluation of commodity, the similarity between the user obtained based on the evaluation of user like this may inaccurate (i.e. openness problem); (2) increasing along with user and commodity, the performance of system can more and more lower (i.e. scalability problem); (3) if never user is evaluated a certain commodity, then these commodity just impossible recommended (i.e. initial evaluation problem).Therefore, present ecommerce is recommended all to adopt multiple technologies to combine the recommendation of the service of realizing or commodity.

But, current most recommend method depends on the comment/scoring of user to commodity, or the scoring of commodity is predicted according to product features, a lot of individual commodity recommendation system or method are mainly just analyzed the feature of user to the preference of commodity or analysis commodity and are found similar user or article, the result of TOP-N is also showed consumer by last calculated recommendation, such as patent No. ZL 201210050057.3 " a kind of e-commerce website Method of Commodity Recommendation based on key word ", ZL 201010538437.2 " a kind of online Products Show system of selection, Apparatus and system ", and the patent of ZL 201110306941.4 " a kind of personalized recommendation method in conjunction with score data and label data ", and application number 201410020517 " a kind of item property clustering method based on user comment ", the patented claims such as application number 201310332537 " the product features recommend method based on Ontology ", application number 201310586279 " a kind of product information recommend method based on trust evaluation and system ", application number 201410196044 " Technologies of Recommendation System in E-Commerce and method thereof based on product similarity ".Although, people also proposed some new methods, and they combine the geographical location information of user, user's interests change in time, the Expression analysis of user, the answer-mode of user, the intention of user are excavated, the behavior trend prediction of user and the method such as user clustering by bought commodity.But in these methods, still there is following two problems: first problem is, they all do not consider the credit how passing through assessment user, and carry out commercial product recommending according to the credit similarity of different user.In fact, the transaction record number of times produced in user transaction process and dealing money have good reference value for the preference of an assessment user.Second Problem is, they all do not consider the so a kind of phenomenon in real trade process, the demand of multiple individuality may have been contained in the trading activity of an i.e. user, as: in certain family, shopping online usually be same user name (account name of namely concluding the business), but in fact the trading activity produced relates to multiple members in family.Therefore, the present invention will for this two problems, sets up a kind of individual commodity recommendation method based on user credit assessment, achieves more precisely, more personalized commercial product recommending service.

Summary of the invention

In order to solve the problem, the invention provides a kind of individual commodity recommendation method based on user credit assessment.The present invention mainly improves the precision of commercial product recommending based on the credit evaluation value of user's historical trading behavior, consider the demand that may contain multiple individuality in the trading activity of a user simultaneously, as: in certain family, shopping online usually be same user name, but the trading activity produced can relate to the multiple members in family, so the commodity each user being produced to transaction are carried out cluster analysis by commending system, accurately distinguish the individual demand of the multiple members in each user, and by introducing the credit value of user, improve the existing collaborative filtering based on commodity, realize more accurate, personalized commercial product recommending service.

Technical scheme of the present invention is:

Based on an individual commodity recommendation method for user credit assessment, the steps include:

1) cancel number of times according to the user-goods orders conclusion of the business number of times collected from e-commerce website and user-goods orders, build the matrix of user-commodity transaction number of times;

2) according to the matrix of user-commodity transaction number of times, and the transaction value of each commodity, the goods orders transaction value and the goods orders that calculate each user cancel total charge, calculate the initial credit value of each user.

3) according to the matrix of user-commodity transaction number of times, total degree cancelled by the commodity conclusion of the business total degree and the commodity that calculate each user, calculates the credit index of each user;

4) according to 2) 3) the initial credit value of each user that calculates and credit index, calculate the credit value of each user, wherein also introduce one when there are these transaction in user system or user may there is the probable value of misoperation/make mistakes.

5) according to the matrix of user-commodity transaction number of times, to the commodity that each user strikes a bargain, be suitable for the sex of user according to commodity, the feature such as age and occupation classifies, produce multiple classification, namely determine the purchase principal classes of material implication in user.

6) according to 5) classification results, buy main body for each class of containing in each user, adopt the collaborative filtering based on commodity respectively, introduce the credit value of each user simultaneously, finally form the Recommendations to each user.

Further, when building the matrix M of user-commodity transaction number of times, the row of matrix is user name, and matrix column is trade name, matrix entries m _ijrepresent user u _ito commodity item _jproduce the number of times of transaction, if strike a bargain 1 time, this entry value cumulative+1, if cancel 1 time, this entry value cumulative-1.

Further, when calculating the initial credit value of each user, by commodity item _jtransaction value count p (item _j), then user u _igoods orders transaction value be all commodity knockdown prices of this user and the summation of conclusion of the business number of times product, count it is all commodity order cancellation prices of this user and the summation cancelling number of times product that goods orders cancels total charge, counts user u _ithe computing method of initial credit value be:

${\mathrm{Cred}}_0\left(u_i\right)=1-\frac{C\left(u_i\right)}{C\left(u_i\right)+S\left(u_i\right)},$ Wherein Cred ₀(u _i) ∈ [0,1].

Further, when calculating the credit index of each user, user u _icommodity conclusion of the business total degree be the summation of all number of times of bargain of this user, count user u _iorder cancellation commodity total degree be the summation of all commodity of order cancellation number of times of this user, count user u _ithe computing method of credit index be: $\mathrm{\λ}\left(u_i\right)=1-\frac{\left|f_c\left(u_j\right)\right|}{f_s\left(u_i\right)+\left|f_c\left(u_j\right)\right|},$ Wherein λ (u _i) ∈ [0,1].

Further, when calculating the credit of each user, wherein introduce one when there are these transaction in user system or user may there is the probable value ε ∈ [0,1] of misoperation/make mistakes, the credit value computing method of each user are:

$\mathrm{Cred}\left(u_i\right)=(1-\mathrm{\ϵ})\×{\mathrm{Cred}}_0{\left(u_i\right)}^{\mathrm{\λ}\left(u_i\right)},$ Wherein Cred (u _i) ∈ [0,1].

Further, when determining the purchase principal classes of material implication in each user, extracting the feature of user purchased item, comprising commodity and being suitable for the features such as the sex of user, age and occupation, classifying.Therefore, by a user u _iin the purchase principal classes that comprises count:

t is that total category set closes, and m is classification set sizes, t _ii-th classification

Further, when forming the Recommendations to each user, adopt user u _icredit value Cred (u _i) and user u _iin the individual segregation class (u that contains _i), form a kind of new collaborative filtering method, the main thought of the method filters out and targeted customer u exactly _iat the interval Cred (u of same credit value _i) ± μ (μ ∈ [0,1] real number is practical experience value) and belong to class (u _i) in each classification TOP-N position user, targeted customer u is recommended in the bargain list of these users _i.

Compared with prior art, the beneficial effect that brings of technical solution of the present invention

The beneficial effect main manifestations that brings of method that the present invention proposes is in the following areas:

(1) classic method is mainly considered to assess the credit rating of commodity, the present invention mainly carries out credit evaluation by the behavior of user's historical trading to user, come for its Recommendations according to the credit evaluation value of targeted customer, this can improve the precision of commercial product recommending, embodies the common preference with similar credit rating user group.

(2) the present invention considers the demand that may contain multiple individuality in the trading activity of a user simultaneously, as: in certain family, shopping online usually be same user name, but the trading activity produced can relate to the multiple members in family, so the present invention carries out cluster analysis to the commodity that each user produced transaction, accurately distinguish the individual demand of the multiple members in each user.

(3) individual demand containing member in the credit of user and user combines by the present invention, improves existing collaborative filtering, achieves more accurate, personalized commercial product recommending service.

(4) advantage of the present invention is also embodied in: for different user, and recommendation results has diversity, and both different user can obtain different recommendation results; Moreover for same user, recommendation results also can have diversity.

Embodiment

Below the present invention is explained in further detail.

According to step, collect e-commerce website (as sky cat) and commodity (as user in sky cat bought dotey's list) information has been bought to each user, counting user is to the number of times of the conclusion of the business of every part commodity or cancellation (as sky cat is shown as closedown), build user-commodity transaction degree matrix, if after wherein certain goods orders conclusion of the business record occurs in same goods orders cancellation record, then this transaction count adds up for 1 time by conclusion of the business, otherwise, then add up for 1 time by cancellation.As shown in table 1, example gives three users A, B, C transaction count matrix to commodity 1,2,3,4,5, and wherein the transaction count of user to commodity 1 counts 3, represents that user A successfully have purchased 3 commodity 1; The transaction count of user A to commodity 3 counts-1, shows that user A finally eliminates 1 time to the order of commodity 3, does not namely produce actual purchase behavior.

Table 1 is the matrix example of user-commodity transaction number of times

	Commodity 1	Commodity 2	Commodity 3	Commodity 4	Commodity 5
						User A	3	0	-1	1	2
User B	-1	0	1	1	1
						User C	2	1	-1	2	0

Wherein, commodity 1-200 unit, commodity 2-1000 unit, commodity 3-500 unit, commodity 4-200 unit, commodity 5-100 unit.

According to step 2, the price of all commodity listed in matrix constructed by step one is collected at e-commerce website, here can suppose that same commodity have identical price in repeatedly purchasing process, if price variance is larger, then can list these part commodity in matrix difference statistics number (such as: for the commodity 1 that price variance is large by the commodity that two pieces is different, commodity 1-A and commodity 1-B can be classified as, this is rational in eCommerce transaction process, because in rational time range, commodity have daily sale price and promotion price two kinds usually, generally there will not be more price change), then, according to the price of commodity and user to the conclusion of the business number of times of commodity, calculate the transaction value of each user, and cancel the total charge bought.As shown in table 1, the commodity transaction value result of calculation of user A is: S (A)=3 × 200+1 × 200+2 × 100=1000 unit, cancel buy total charge result of calculation be: C (A)=1 × 500=500 unit.Accordingly, the initial credit value that we calculate user A is similar, the initial credit value calculating user B is Cred ₀(B)=0.8, that user C is Cred ₀(C)=0.7826.

According to step 3, except the initial credit value calculating each user, also will according to user-commodity transaction degree matrix, total degree cancelled by the commodity conclusion of the business total degree and the commodity that calculate each user, thus calculates the credit index of each user.As shown in table 1, the commodity of user A strike a bargain 6 times, and cancel 1 time, then credit index result of calculation is: similar, the credit index calculating user B be λ (B)=0.75, user C be λ (C)=0.8333.

According to step 4, except the initial credit value that calculates each user according to step 2, step 3 and credit index, consider that user is in the process of exchange of commodity, system may break down or accident temporarily, and user may occur misoperation because of carelessness or make mistakes, thus cause the conclusion of the business in customer transaction record or cancel record producing deviation.So when adopting the initial credit value of user and credit index to calculate the current credit value of this user, we also introduce the probable value ε that misoperation/make mistakes may appear in a system or user, suppose that its value is 0.001, the credit value that we can calculate user A is: Cred (A)=(1-0.001) × Cred ₀(A) ^{λ (A)}=0.999 × 0.6667 ^0.8571=0.7058, similar, the credit value calculating user B be Cred (B)=0.8458, user C be Cred (C)=0.8151.

According to step 5, in order to provide commercial product recommending service more personalized, consider that each account logging on e-commerce website (i.e. user) in fact may buy commodity for multidigit member, such as, one family is usually only set up a public account Alice for the multiple members (as father, mother and child) in family and is bought commodity.For this reason, here need the commodity to customer transaction (comprising the commodity striking a bargain and cancel) to be suitable for the features such as the sex of user, age and occupation according to commodity to classify, thus determine the main body of multiple actual purchases of containing in the middle account (i.e. user), by the classification to purchased commodity, define Alice user and contained father, mother and child three purchase main body.

According to step 6, the credit value of each user is drawn in evaluates calculation, and after the type of subject that each user is contained segmented, collaborative filtering will according to targeted customer B or targeted customer C at interval Cred (the B) ± μ (supposing μ=0.05) of same credit value, namely [0.7958, the preference of the user 0.8958] carries out commercial product recommending, in recommendation process, also whether will belong to the demand of the actual purchase main body that targeted customer B or targeted customer C contains according to product features.It is as shown in table 2,

1) credit value of user C meets above-mentioned collaborative filtering condition, and plan the commodity 2 that C bought and recommend targeted customer B, before recommendation, detect the some purchase main bodys (as father) finding that the feature of commodity 2 belongs to main body B and contains, then recommend user B.The feature of the commodity 2 that the user B such as selected buys belongs to the purchase principal classes character pair in this user A.

2) in the collaborative filtering process recommending targeted customer C, the credit value of user B also meets based on above-mentioned filtercondition, and plan the commodity 5 that B bought and recommend user C, but before recommendation, detect and find that commodity 5 feature does not belong to any one purchase main body in main body C, therefore do not recommend user C.

Table 2 is the collaborative filtering result example based on user credit and individual segregation

	Commodity 1	Commodity 2	Commodity 3	Commodity 4	Commodity 5
						User A	3	0	-1	1	2
User B	-1	√	1	1	1
						User C	2	1	-1	2	Ⅹ

Wherein, " √ " represents recommended commodity, and " Ⅹ " represents not recommended commodity.

Claims (9)

1., based on an individual commodity recommendation method for user credit assessment, the steps include:

1) cancel number of times according to the user-goods orders conclusion of the business number of times collected, user-goods orders, build user-commodity transaction degree matrix M;

2) according to the transaction value of user-commodity transaction degree matrix M and each commodity, the goods orders transaction value and the goods orders that calculate each user cancel total charge, then cancel according to goods orders transaction value and goods orders the initial credit value that total charge calculates respective user;

3) according to user-commodity transaction degree matrix M, total degree cancelled by the commodity conclusion of the business total degree and the commodity that calculate each user, then cancels according to commodity conclusion of the business total degree and commodity the credit index that total degree calculates respective user;

4) according to 2), 3) the initial credit value that calculates and credit index, calculate the credit value of each user;

5) according to user-commodity transaction degree matrix M, the commodity that each user strikes a bargain are classified, determine the purchase principal classes that each user is contained;

6) for arbitrary user A, determine that a credit value is interval, then select the user being positioned at this credit value interval, if the feature of the commodity b of the user B purchase selected belongs to the purchase principal classes character pair in this user A, then these commodity b is recommended this user A.

2. the method for claim 1, is characterized in that, the described initial credit value of user is directly proportional to the turnover accounting of this user; The described credit index of user is directly proportional to the conclusion of the business number of times accounting of this user.

3. method as claimed in claim 1 or 2, it is characterized in that, the computing method of described initial credit value are: by commodity item _jtransaction value count p (item _j), first calculate user u _igoods orders transaction value $S\left(u_i\right)=\underset{m_{\mathrm{ij}}\≥1}{\mathrm{\Σ}}{m}_{\mathrm{ij}}\×p\left({\mathrm{item}}_j\right),$ Goods orders cancels total charge $C\left(u_i\right)=\underset{m_{\mathrm{ij}}\≤-2}{\mathrm{\Σ}}\left|m_{\mathrm{ij}}\right|\×p\left({\mathrm{item}}_j\right);$ Then user u is calculated _iinitial credit value: wherein Cred ₀(u _i) ∈ [0,1], m _ijrepresent user u _ito commodity item _jproduce the number of times of transaction.

4. method as claimed in claim 3, is characterized in that, each user u _ithe computing method of described credit value be: wherein Cred (u _i) ∈ [0,1], probable value ε ∈ [0,1] represent is that system or user maloperation or unexpected probability of makeing mistakes may occur, λ (u _i) be user u _icredit index.

5. method as claimed in claim 1 or 2, it is characterized in that, the computing method of described credit index are: by commodity item _jtransaction value count p (item _j), first calculate user u _icommodity conclusion of the business total degree f _s order cancellation commodity total degree f _c $\left(u_i\right)=\underset{m_{\mathrm{ij}}\≤-1}{\mathrm{\Σ}}{m}_{\mathrm{ij}};$ Then user u is calculated _icredit index: $\mathrm{\λ}\left(u_i\right)=1-\frac{\left|f_c\left(u_j\right)\right|}{f_s\left(u_i\right)+\left|f_c\left(u_j\right)\right|},$ Wherein λ (u _i) ∈ [0,1], m _ijrepresent user u _ito commodity item _jproduce the number of times of transaction.

6. method as claimed in claim 1 or 2, it is characterized in that, the construction method of described user-commodity transaction degree matrix M is: using the row of user name as matrix, trade name, as matrix column, uses matrix entries m _ijrepresent user u _ito commodity item _jproduce the number of times of transaction, if strike a bargain 1 time, then this entry value cumulative+1, if cancel 1 time, this entry value cumulative-1.

7. method as claimed in claim 1 or 2, is characterized in that, describedly determines that the method for the purchase principal classes that each user is contained is: first extract user u _ithe feature of purchased item, then calculates this user u _iin the purchase principal classes that comprises: t is that total category set closes, and m is classification set sizes, t _ii-th classification.

8. method as claimed in claim 7, is characterized in that, described feature comprises commodity and is suitable for the sex of user, age and occupation.

9. method as claimed in claim 1 or 2, it is characterized in that, buy principal classes for each in user A, adopt the collaborative filtering method based on commodity to select the TOP-N position user of each classification respectively, this user A is recommended in the bargain list of these users.