CN101206752A - Electric commerce website related products recommendation system and method - Google Patents

Abstract

The invention discloses a corresponding commodity on electronic business websites recommendation system which mines characteristic words from the historical purchasing data and purchasing contents of customers, builds similar relations of different eigenvectors by mathematical methods, acquires the similar relations of commodities and thus finds the suitable commodities for customers. The invention also discloses a recommendation method of corresponding commodities which comprises the following steps of: extracting data of a certain time span from the database, choosing and calculating the useful data information segment to build a 'commodity-customer' calculating model, calculating the similarity degree of commodities according to the commodity-customer-other commodities model, and recommending the results to the customers. The invention has the advantages of grasping the mentality of customers, encouraging the customers to purchase corresponding commodities and thus promoting the sales volume.

Description

Electric commerce website related products recommendation system and method thereof

Technical field

The present invention relates to the related products recommendation field of e-commerce website, particularly relate to e-commerce website based on the commercial product recommending technology of shopping behavior Item and content-based commercial product recommending technology.

Background technology

Current era is the epoch of infotech, and the internet has occupied quite great proportion in people's life, and shopping online is exactly one of them.For this reason, e-commerce website is a target to obtain maximum benefit all, with the method for technology and non-technology, increases registered customers, increases order volume, quality services are provided.Under these prerequisites, personalized commercial product recommending technology is arisen at the historic moment.

At present, some domestic recommended technologies adopt more original method mostly.

The most original method is " data base querying ", finds in the database and client browses, collects or the commodity of buying have some other commodity of same authors, same category, same subject etc. by the sql statement, recommends client." questionnaire feedback " mode by puing question to, allows client answer some problems, is directly acquainted with client's hobby, recommends suitable commodity.The commercial product recommending method that also has " correlation rule " or the like form in addition.More than these methods, not ideal enough at aspects such as the accuracy of recommending, real-times, the automaticity and the degrees of permanency of recommendation are low, lacking individuality.Summary is got up, and domestic goods recommend to exist following shortcoming at present:

The recommendation of lacking individuality.A lot of recommendation results is at all users, or at most of user, not recommendation at the personalization of certain user's current behavior.In these recommendation results, may a lot of recommendations not conform to certain user's interest.This is the disadvantage that Electronic Commerce in China is recommended.

The automaticity of recommending is low.Most recommendation function all needs the user through mutual with computing machine after a while, imports the interest information of oneself, just can obtain result's (" questionnaire feedback " mode) then.

The degrees of permanency of recommending is low.Domestic most recommended technology all is based upon on active user's session basis at present, can not utilize user's session information in the past, and the degrees of permanency of recommendation is very low, and this also is a major defect of domestic recommended technology.

Recommend method is single.It is exactly classified browse and content-based retrieval basically that the great majority that used are recommended strategy, lacks the multiple mixing use of recommending strategy, especially lacks the use that mixes of recommendation strategy personalized and non-personalization.

Real-time is poor, can not onlinely recommend.The recommendation strategy that has can not be accomplished online recommendation, recommends as the letter formula, and recommendation results can not in time feed back to the user.

The limitation of recommending is big.Major part can only be recommended those commodity that certain sales volume, pass fluence are arranged, and for those commodity that newly advance of just having put on the shelf, at a loss as to what to do---cold beginning problem.

Summary of the invention

In order to overcome the above problems, give the customer recommendation commodity that they buy possibly, can recommend the commodity of newly putting on the shelf simultaneously again, the invention provides a kind of easy method of using at e-commerce website:

At first use based on the historical recommended technology that excavates of shopping, excavate from client's actual purchase data, the associated recommendation commodity that obtain meet objectively actual shopping trend of client and interest custom.

Then, use the natural language processing correlation technique of content-based excavation, excavate the feature speech of merchandise news, commodity list is shown as relevant information characteristics vector, obtain similarity relation between commodity according to the similarity relation between the different characteristic vector, thereby be the customer recommendation dependent merchandise.

In sum, system is by constituting based on historical excavation part of shopping and content-based excavation part.First excavates part based on shopping history:

As shown in Figure 1, the employed functional module of system of the present invention comprises: source data preparation module, data field extraction module, computation model module, dependent merchandise computing module.Each part wherein all is that the present invention is achieved the also basis of successful Application, has constituted a technological system.

1. source data preparation module: from database, extract order data, travel log, search daily record of certain hour span or the like, can reflect the data set that " commodity-client " concerns.

2. data field extraction module: extract the useful information in the every order, as: buy data such as date, purchaser, purchase commodity.

3. computation model module:, set up " commodity-client " computation model by analyzing these data.

4. dependent merchandise computing module: the relation of these other commodity of purchase of customer of---buying the client of these commodity---according to commodity, at each commodity in the model, from these commodity itself, find the client who bought it, again from client, find the commodity that have relation with it, use the core formula to calculate similarity between these two commodity then.

At last, obtain the inverted file of the recommendation results of each commodity after calculating finishes.

Specifically may further comprise the steps:

1) raw data is prepared: can be order data, or travel log, search daily record or the like.

2) data field extracts: obtain computing information useful in the data source, as time, people, commodity etc.

3) set up computation model:, set up " commodity-client " computation model (represent which purchase of customer which commodity, which commodity which purchase of customer relation such as to be crossed) by by the data of analyze extracting.

4) dependent merchandise computation model: the dependent merchandise set of calculating each commodity with the core calculations formula.

Wherein, the core calculations formula of above-mentioned steps 4 has a variety of selections, such as:

(1) L1-Norm algorithm:

$L1(B,R)=\frac{|B\∩R|}{\left|B\right|\·\left|R\right|}$

(2) L2-Norm algorithm:

$L2(B,R)=\frac{|B\∩R|}{\sqrt{\left|B\right|\·\left|R\right|}}$

(3) MI1 algorithm:

$\mathrm{MI}1(B,R)=-|B\∩R|\·\lg \frac{|B\∩R|}{\left|B\right|\·\left|R\right|}$

(4) COS algorithm: $\cos (\stackrel{\→}{A},\stackrel{\→}{B})=\frac{\stackrel{\→}{A}\·\stackrel{\→}{B}}{\left|\right|\stackrel{\→}{A}\left|\right|\·\left|\right|\stackrel{\→}{B}\left|\right|}$

More than exemplified 4 formula relatively more commonly used, the emphasis and the effect of calculating have nothing in common with each other, and other also have some formula, and which kind of selects for use, by the real needs decision, can use separately, also can be used in combination.

The content-based excavation part of second portion:

As shown in Figure 5, the employed functional module of system of the present invention comprises: merchandise news initialization module, word-dividing mode, feature phrase module, feature speech module, Vector Groups compound module and relatedness computation module.Each part wherein all is that the present invention is achieved and the basis of successful Application and constituted a technological system.

Merchandise news initialization module: read the relevant information of every commodity, set up the correspondence collection of commodity ID and descriptor.

Word-dividing mode: descriptor is carried out Chinese word segmentation, the result behind the acquisition participle.

Feature phrase module: the feature phrase behind the calculating descriptive labelling information participle.

Feature speech module: the feature speech behind the calculating descriptive labelling information participle.

Vector Groups compound module: feature phrase and feature vocabulary are shown as the set of commodity proper vector.

Relatedness computation module:, calculate the similarity between commodity, as the alternative set of dependent merchandise by the set of proper vector.

May further comprise the steps:

1) prepares the information of dependent merchandise, comprise that commodity title, content of good, synopsis etc. describe the information of commodity.

2) merchandise news is carried out lexical analysis, obtain the alternative set of calculated characteristics phrase and feature speech.

3) calculated characteristics phrase can use several different methods that alternative word is made up as the feature phrase, extracts as the portmanteau word based on n-gram.

4) calculated characteristics speech to the alternative word ordering, obtains the weight of alternative word, as calculating the TFIDF value of speech.

5) vector set that formation is represented the commodity feature united in feature phrase and feature speech.

6) calculate the dependent merchandise of each commodity, and the result shown recommend client.

The present invention can be applicable to based on the commercial product recommending of shopping behavior Item, content-based commercial product recommending and the recommendation etc. that do not have the commodity of newly putting on the shelf of purchase of customer record.Can complement one another between several persons, solve the commercial product recommending problem from different perspectives,, thereby reach the sales volume that increases business web site for user's shopping provides favorable experience.

Description of drawings

Fig. 1 is of the present invention based on the historical system flowchart that excavates part of shopping.

Fig. 2 is of the present invention based on the historical partial data processing module process flow diagram that excavates of shopping.

Fig. 3 is of the present invention based on historical computation model partly row and the index structure figure of excavating of shopping.

Fig. 4 is of the present invention based on the historical calculating dependent merchandise process flow diagram that excavates part of shopping.

Fig. 5 is the content-based commercial product recommending process flow diagram that excavates part of the present invention.

Fig. 6 is the content-based extracting keywords group of part and the process flow diagram of keyword of excavating of the present invention.

Fig. 7 is the process flow diagram that the content-based dependent merchandise that excavates part of the present invention calculates.

Embodiment

Example 1: the processing of three months orders of certain shopping website

Step 1: raw data is prepared

As extracting 3 months order data OrderInfo.txt, form is (a time client ID commodity ID trade name)

For example: the data presentation of preceding 20 row of OrderInfo.txt

20070715 C1 P1 Name1

20070717 C1 P2 Name2

20070721 C3 P1 Name1

20070722 C5 P2 Name2

20070725 C1 P4 Name4

20070803 C4 P7 Name7

20070805 C1 P3 Name3

20070812 C3 P2 Name2

20070825 C2 P2 Name2

20070827 C5 P5 Name5

20070827 C5 P3 Name3

20070828 C4 P9 Name9

20070828 C3 P4 Name4

20070828 C3 P8 Name8

20070829 C1 P5 Name5

20070829 C5 P8 Name8

20070829 C4 P4 Name4

20070830 C1 P8 Nane8

20070830 C2 P5 Name5

20070903 C2 P4 Name4

20070903 C4 P8 Name8

20070903 C2 P9 Name9

20070903 C3 P9 Name9

20070903 C3 P6 Name6

…… …… …… ……

Step 2: data field extracts, as shown in Figure 2

Press the CustID ordering:
			20070715 20070717 20070721 20070722 20070725 20070803 20070805 20070812 20070825 20070827 20070827 20070828 20070828 20070828 20070829 20070829 20070829 20070830 20070830 20070903 20070903 20070903 20070903 20070903 ……	C1 C1 C3 C5 C1 C4 C1 C3 C2 C5 C5 C4 C3 C3 C1 C5 C4 C1 C2 C2 C4 C2 C3 C3 ……	P1 P2 P1 P2 P4 P7 P3 P2 P2 P5 P3 P9 P4 P8 P5 P8 P4 P8 P5 P4 P8 P9 P9 P6 ……

Press the ProID ordering:
			20070715 20070717 20070721 20070722 20070725 20070803 20070805 20070812 20070825 20070827 20070827 20070828 20070828 20070828 20070829 20070829 20070829 20070830 20070830 20070903 20070903 20070903 20070903 20070903 ……	C1 C1 C3 C5 C1 C4 C1 C3 C2 C5 C5 C4 C3 C3 C1 C5 C4 C1 C2 C2 C4 C2 C3 C3 ……	P1 P2 P1 P2 P4 P7 P3 P2 P2 P5 P3 P9 P4 P8 P5 P8 P4 P8 P5 P4 P8 P9 P9 P6 ……

Extract contiguous items (useful information), data item is sorted.In this embodiment, when extracting contiguous items, extract OrderDate (time buying), CustID (client ID), ProID (commodity ID) in the order data, be stored in the structure that comprises these three integer fields, write in the binary file, and sort two intermediate result files below generating respectively according to CustID in each cellular construction and ProID:

Step 3: generate computation model

Two destination files by previous step generates generate " commodity-client " computation model.

In this example, computation model mainly represent which purchase of customer which commodity, which commodity which purchase of customer relation such as to be crossed by, may be summarized to be client do shopping model and the purchased model of commodity.

Which commodity client each client of model representation that does shopping bought, and was an inverted list structure (C-P inverted list).Each commodity of the purchased model representation of commodity by which client were enough bought, and also were an inverted list structure (P-C inverted lists).

By two intermediate result files that read step 2 produces, two inverted lists below generating, and two search indexes (C-P inverted index and P-C inverted index), as shown in Figure 3.

C-P inverted list C1:P1 P2 P3 P4 P5 P8 C2:P2 P4 P5 P9 C3:P1 P2 P4 P6 P8 P9 C4:P4 P7 P8 P9 C5:P2 P3 P5 P8

P-C inverted list P1:C1 C3 P2:C1 C2 C3 C5 P3:C1 C5 P4:C1 C2 C3 C4 P5:C1 C2 C5 P6:C3 P8:C1 C3 C4 C5 P9:C2 C3 C4

Step 4: dependent merchandise calculates

(1), core calculations formula:

Here be example with " COS algorithm ", it is to adopt a kind of proposed algorithm based on the collaborative filtering between the clauses and subclauses, i.e. cosine formula.It matches to the commodity with same or similar sales histories, and adds up their similarity degree, and vectorization is carried out in marketing activity, then vector is carried out formula and calculates, and forms a similar entry matrix at last.

$\mathrm{similarity}(\stackrel{\→}{A},\stackrel{\→}{B})=\cos (\stackrel{\→}{A},\stackrel{\→}{B})=\frac{\stackrel{\→}{A}\·\stackrel{\→}{B}}{\left|\right|\stackrel{\→}{A}\left|\right|\·\left|\right|\stackrel{\→}{B}\left|\right|}$

Wherein, in instantiation:

Vector A: the sale person-time amount of expression commodity A, vectorial B: the sale person-time amount of expression commodity B.

Molecule implication: the identical commodity amount of buying at commodity A and commodity B.

The denominator implication: expression commodity A sells the evolution of person-time amount and the product that commodity B sells the evolution of person-time amount.

(2), flow process as shown in Figure 4.

A) carry out memory-mapped and read two inverted lists on the disk,

B) for the commodity Pi in the P-C inverted list, inquire about its sales figure, obtain client and gather CustList,

C) each client among the traversal CustList, inquiry inverted list C-P obtains client's purchase and gathers ProList,

D) traversal ProList matches in twos with former commodity Pi (thinking that here two commodity being crossed by same purchase of customer have correlativity), create-rule (Proi-Proj-Count),

E) by computing formula and rule, calculate the similarity of each commodity among commodity Pi and the ProList, ordering back record result,

F) return b, up to having traveled through all Pi.

When for example, calculating the dependent merchandise of P1:

A) search the P-C inverted list, obtain sales figure C1, the C3 of P1

B) search the C-P inverted list, obtain C1 and also bought (P2, P3, P4, P5, P8), C3 has also bought (P2, P4, P6, P8, P9), thinks that there is correlativity in other purchase commodity of P1 and C1 and C3

C) statistical rules: P1-P2-2, P1-P3-1, P1-P4-2, P1-P5-1, P1-P6-1, P1-P8-2, P1-P9-1

D) similarity between calculating P1 and these dependent merchandises

P1-P2：

P1-P3：

P1-P4： P1-P5：

P1-P6：

P1-P8：

P1-P9：

Therefore, obtain the recommendation results (commodity relevant with P1 are by sequencing of similarity) of P1:

P2(0.707)、P4(0.707)、P6(0.707)、P8(0.707)、P9(0.707)、P3(0.5)、P5(0.4087)

Example 2: client's recommendation dependent merchandise of buying books

Step 1

The ready message data mainly are commodity textual description information.It comprises commodity title, content of good, synopsis etc., as:

C program design (third edition)---new century Basis of Computer Engineering education book series author: Tan Haoqiang work

The design of JavaScript advanced procedures---scheme clever program design book series author: Zha Kasi work, Cao Li etc. translate

…….

Step 2

Above data message is carried out participle.At different user demands, can adopt different cutting methods.The word algorithm of cutting as based on the forward maximum match can guarantee certain cutting speed, and shortened the working time of system.

Step 3

The analytical characteristic phrase.In order to obtain the feature phrase, can adopt several different methods: based on the n-gram of word with based on n-gram of speech etc.On the basis of participle, consider the result of participle fragment, better based on the n-gram effect of speech.N-gram based on speech: consider the continuous combination of speech, individual character, can constitute the phrase of complete meaning.Here select to use the 2-5 metagrammar, calculated the combined probability of all 2-5 unit speech, word, become the pattern of speech.The speech of the speech that the frequency is lower also can be included into like this.

Step 4

The analytical characteristic speech extracts.To the most significant word of difference document should be that those frequencies of occurrences in document are enough high, but at the enough few word of other document medium frequencys of entire document collection.So term weighing is frequently relevant with its frequency and literary composition.Here use the TF-IDF formula to calculate

${\mathrm{\ω}}_i\left(d\right)=\frac{{\mathrm{tf}}_i\left(d\right)\log (\frac{N}{n_i}+0.1)}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\left({\mathrm{tf}}_i\left(d\right)\right)}^2\×{\log }^2(\frac{N}{n_i}+0.1)}}---\left(1\right)$

According to the TF-IDF formula, the document that comprises a certain entry in the document sets is many more, illustrates that the ability of its differentiation document category attribute is low more, and its weights are more little; On the other hand, the frequency that a certain entry occurs in a certain document is high more, illustrates that the ability of its differentiation document content attribute is strong more, and its weights are big more.

Step 5

The pattern of feature phrase and feature phrase composite vector,, next step lays the first stone for calculating the degree of correlation.

Step 6

Relatedness computation.The similarity of calculating two vectors has several different methods: based on the similarity calculating method of vector space model; Similarity calculating method based on aggregation model.Here consider speed and efficient, adopt similarity calculating method based on aggregation model.The decision function of text similarity: make the fingerprint collection of F (A) expression document A, the fingerprint collection of F (B) expression document B, (then decision function is S for A, the B) similarity of expression document A and B

$S(A,B)=\frac{|F\left(A\right)\∩F\left(B\right)|}{|F\left(A\right)\∪F\left(B\right)|}---\left(2\right)$

When document vector dimension changes when little, can be reduced to

S(A，B)＝|F(A)∩F(B)| (3)

Formula (3) can expression (2) trend, application of formula (3) can be simplified calculation process.At first, write down the proper vector of each merchandise news; Then write down the corresponding commodity of each vector; When calculating the commodity similarity, only need to calculate the corresponding commodity of each vector.In program, the vectorial corresponding goods ID sequence table according to commodity data is set up the row of falling when calculating similarity, can draw the big commodity of the degree of correlation as long as the vector traversal of commodity correspondence is searched the back counting.

Example as a result

The new commodity that c program design (third edition) is recommended is:

The C programmer design

The C++ program design is resolved

C++ program design study course (the 2nd edition)

The design of JavaScript advanced procedures---scheme clever program design book series, the new commodity of recommendation is:

The practical study course of ASP.NET

Claims (8)

1. electric commerce website related products recommendation system, excavate the feature speech from client's shopping historical data and shopping content, utilize mathematical method to set up similarity relation between the different characteristic vector, obtain the similarity relation between commodity, thereby be the customer recommendation dependent merchandise, wherein client's shopping history portion comprises:

Source data preparation module: the data set that can reflect " commodity-client " relation that from database, extracts the certain hour span;

Data field extraction module: extract the useful information in the every order, set up " commodity-client " computation model;

The relation of these other commodity of purchase of customer that dependent merchandise computing module: according to commodity---is bought the client of these commodity---, utilization core formula calculates the similarity between them;

At last, obtain the inverted file of the recommendation results of each commodity;

The content part of wherein doing shopping comprises:

Merchandise news initialization module: read the relevant information of every commodity, set up the correspondence collection of commodity ID and descriptor;

Word-dividing mode: descriptor is carried out Chinese word segmentation, the result behind the acquisition participle;

Feature phrase module: the feature phrase behind the calculating descriptive labelling information participle;

Feature speech module: the feature speech behind the calculating descriptive labelling information participle;

Vector Groups compound module: feature phrase and feature vocabulary are shown as the set of commodity proper vector;

Relatedness computation module:, calculate the similarity between commodity, as the alternative set of dependent merchandise by the set of proper vector.

2. a kind of electric commerce website related products recommendation system as claimed in claim 1 is characterized in that the source data preparation module is meant that execution extracts the order data of certain hour span, travel log, search daily record from database.

3. a kind of electric commerce website related products recommendation system as claimed in claim 1 is characterized in that the data field extraction module is meant purchase date, purchaser, the purchase commodity data of carrying out in every order of extraction.

4. a kind of electric commerce website related products recommendation system as claimed in claim 1 is characterized in that the dependent merchandise computation model is meant the one or more combination that adopts in L1-Norm algorithm, L2-Norm algorithm, MI1 algorithm, the COS algorithm.

5. an electric commerce website related products recommendation method comprises the steps:

(1) data division in the processing shopping historical data base:

1. from database, prepare raw data,

2. extract computing information data field useful in the data source,

3. by analyzing the data of extracting, set up " commodity-client " computation model,

4. calculate the dependent merchandise set of each commodity with the core calculations formula;

(2) data division of processing shopping content of good:

1. prepare the information of dependent merchandise,

2. merchandise news is carried out lexical analysis, obtains the alternative set of calculated characteristics phrase and feature speech,

3. the calculated characteristics phrase makes up alternative word as the feature phrase,

4. the calculated characteristics speech to the alternative word ordering, obtains the weight of alternative word,

5. the vector set that formation is represented the commodity feature united in feature phrase and feature speech,

6. calculate the dependent merchandise of each commodity, and show that the result who recommends gives client.

6. a kind of electric commerce website related products recommendation method as claimed in claim 5 is characterized in that the core calculations formula is meant the one or more combination that adopts in L1-Norm algorithm, L2-Norm algorithm, M11 algorithm, the COS algorithm.

7. a kind of electric commerce website related products recommendation method as claimed in claim 5 is characterized in that using portmanteau word based on n-gram to extract and calculates the calculated characteristics phrase.

8. a kind of electric commerce website related products recommendation method as claimed in claim 5 is characterized in that making the TFIDF value of word to calculate the weight of speech.

Images