CN109670914A - A kind of Products Show method based on time dynamic characteristic - Google Patents
A kind of Products Show method based on time dynamic characteristic Download PDFInfo
- Publication number
- CN109670914A CN109670914A CN201811544768.XA CN201811544768A CN109670914A CN 109670914 A CN109670914 A CN 109670914A CN 201811544768 A CN201811544768 A CN 201811544768A CN 109670914 A CN109670914 A CN 109670914A
- Authority
- CN
- China
- Prior art keywords
- product
- user
- attribute
- similarity
- level
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
- G06Q30/0203—Market surveys; Market polls
Landscapes
- Business, Economics & Management (AREA)
- Engineering & Computer Science (AREA)
- Strategic Management (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Finance (AREA)
- Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- Marketing (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Game Theory and Decision Science (AREA)
- Data Mining & Analysis (AREA)
- Human Resources & Organizations (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of Products Show methods based on time dynamic characteristic, it include: purchaser record, review record and the product description information for collecting user in target time section, target time section is divided into time slice sequence, and generates user to the rating matrix of product；The attribute relative level for obtaining each product, for indicating class of each attribute of product in the affiliated timeslice of product；According to collected information and the attribute relative level of product, the product similarity of the user's similarity and any two product of any two user under same user's attention rate is obtained；Scoring according to user's similarity and product similarity prediction user to product, and majorized function is determined according to prediction result；Rating matrix is decomposed, and matrix decomposition result is adjusted according to majorized function, so that majorized function value is minimum, to generate the recommendation list of each user, completes the Products Show to user.The present invention can effectively improve the accuracy of Products Show.
Description
Technical field
The invention belongs to the Products Show technical fields based on machine learning, dynamic based on the time more particularly, to one kind
The Products Show method of step response.
Background technique
The explosive growth of e-commerce results in the rapid development of recommender system, and recommender system has become data mining
(DM) and the research hotspot in the field artificial intelligence (AI) it, has been widely used for realizing E-business service at present, based on position
Service and the various services such as news push among.Recommender system is effectively alleviated by generating the recommendation list of property one by one
Information overload caused by as product substantial increase, and the purchase suggestion of customization is provided by active to excite the shopping of user
Desire.For many hyundai electronics business web sites, such as Amazon, Taobao, Ebay, Products Show is being effectively improved user's body
It tests, sale etc. is promoted to play a crucial role, cause extensive concern in business and academia.
Products Show method and collaborative filtering model based on content are two kinds of Products Show methods being widely used at present,
Both Products Show methods can both consider each product attribute of product, (such as can also browse for the historical behavior of user
History, grading, comment) it is modeled, user can be accurately predicted to the interest of different product, to preferably meet user
Demands of individuals.But based on the method for content just for the feature occurred in user's purchasing history, and have ignored product category
Property real time characteristic, it is difficult to directly processing the faster product of renewal frequency (such as electronic product)；Collaborative filtering model is by user
Preference and product summary file are considered as static information, with user preference and product configuration file may over time and
The fact that dynamic change, is not inconsistent, and accuracy is recommended to cannot be guaranteed.
In view of in real life, the competitiveness of product is not constant, it will usually over time and by
Gradually decline, many recommended methods based on context are suggested, these methods can be improved recommendation by introducing time factor
Accuracy.Such method often relies on one of people cognition, i.e., nearest information (for example, the product finally bought) than with
It is preceding more useful.Therefore, the example closer with current time point can obtain higher weight.However, due to Buying Cycle electronics
Product is usually more than 1 year (for example, computer, mobile phone etc.), and the speed that electronic product itself updates is also very fast, therefore most
Close purchaser record is likely to lose timeliness, in other words, in the recommendation of the fields specific products such as electronic product, on
It states the method for product based on context and is not suitable for.
Summary of the invention
In view of the drawbacks of the prior art and Improvement requirement, the present invention provides a kind of products based on time dynamic characteristic to push away
Method is recommended, it is intended that improving the exactness of Products Show in the case where the competitiveness dynamic change at any time of product.
To achieve the above object, the present invention provides a kind of Products Show methods based on time dynamic characteristic, comprising:
(1) purchaser record, review record and the product description information for collecting user in target time section, by target time section
It is divided into time slice sequence, and user is generated to the rating matrix of product according to review record；
(2) according to the attribute relative level of each product of product description information acquisition, for indicating each attribute of product in product
Class in affiliated timeslice；
(3) according to collected information and the attribute relative level of product, the user for obtaining any two user is similar
The product similarity of degree and any two product under same user's attention rate；
(4) scoring according to user's similarity and product similarity prediction user to product, and determined according to prediction result
Majorized function；
(5) rating matrix is decomposed, and matrix decomposition result is adjusted according to majorized function, to obtain so that
The smallest target user's recessive character matrix of majorized function value and target product recessive character matrix, to generate each user's
Recommendation list completes the Products Show to user；
Wherein, the affiliated timeslice of product is the timeslice where the time point of production, and attribute relative level includes belonging to
The horizontal lower bound of property and the attribute level upper bound.
Products Show method provided by the present invention based on time dynamic characteristic, it is discrete by that will be divided into continuous time
Time slice sequence, the visit of user can be studied according to a specific module (for example, with week, being divided the moon)
Behavior is asked, so as to preferably compare the product of different times.
Further, step (2) includes:
For any one product pj, according to product description information acquisition product pjAffiliated timeslice TqInterior production is owned
Product, to obtain product set
Obtain product setIn be weaker than product p in the configuration of any k-th of product attributejProduct quantity Ninf
(pj), to calculate product pjAttribute level lower bound on k-th of product attribute are as follows:
Obtain product setIn be better than product p in the configuration of k-th of product attributejProduct quantity Nsup(pj), with
Calculate product pjThe attribute level upper bound on k-th of product attribute are as follows:
According to attribute level lower boundWith the attribute level upper boundObtain product pjCategory on k-th of product attribute
Property relative level；
Wherein,Indicate product setIn product quantity.
For any one product, it is weaker than ratio shared by the product of the product by obtaining attribute configuration in same timeslice
Example obtains the attribute level lower bound of product, and is better than shared by the product of the product by obtaining attribute configuration in same timeslice
Ratio, obtain the attribute level upper bound of product, can clearly react in the class where each attribute of the product of a certain period,
To the characteristic of the competitiveness dynamic change at any time of more enough preferably reactor products.
Further, step (3) further include:
It obtains the number that each product attribute occurs in the review record of each user respectively according to review record, and thus obtains
Each user is obtained to the degree of concern of different product attribute, calculation formula is as follows:
Wherein, i indicates that Customs Assigned Number, k indicate product attribute number, ci,kIndicate k-th of product attribute in i-th of user
Review record in the number that occurs, Xi,kI-th of user is indicated to the degree of concern of k-th of product attribute, N indicates scoring area
Between number.
Each user is obtained to the degree of concern of different product attribute, can more precisely obtain specific user for certain
The Sentiment orientation of product attribute, to improve recommendation accuracy.
Further, in step (3), any two user u is obtainedi1And ui2The method of user's similarity include:
Obtain user u respectively according to purchaser recordi1And ui2The product bought, and the product bought according to each user
Attribute relative level obtain the level of consumption of each user；The level of consumption of each user includes: user ui1Level of consumption lower boundWith the level of consumption upper boundAnd user ui2Level of consumption lower boundWith the level of consumption upper bound
According to level of consumption lower boundWithCalculate user ui1And ui2Level of consumption lower bound similarity are as follows:
According to the level of consumption upper boundWithCalculate user ui1And ui2The level of consumption upper bound similarity are as follows:
According to similarity KLL (ui1,ui2) and similarity KLU (ui1,ui2) calculate user ui1With user ui2User it is similar
Degree are as follows:
Wherein, ε3For preset adjustment factor, k indicates product attribute number, and K indicates product attribute sum.
Class as where the attribute relative level of product has measured product in each attribute of a certain period, is purchased by user
The attribute relative level of the product bought calculates the level of consumption of user, and according to customer consumption level calculation user's similarity,
Influence of the time factor to user's similarity can be fully considered, when calculating user's similarity, also by user to different product
The degree of concern of attribute is taken into account, so that user's similarity calculated, enables to user's similarity calculated more
The similarity degree between user is accurately portrayed, to improve the accuracy of Products Show.
Further, for any one user ui, according to user uiThe attribute relative level for the product bought obtains
User uiThe level of consumption, comprising:
It obtains by user uiThe product set that the product bought is constituted
According to product setIn attribute level lower bound of each product on k-th of product attribute calculate user uiDisappear
The flat lower bound of water wasting are as follows:
According to product setIn the attribute level upper bound of the product on k-th of product attribute calculate user uiConsumption
The horizontal upper bound are as follows:
According to level of consumption lower boundWith the level of consumption upper boundObtain user uiThe level of consumption；
Wherein,WithRespectively indicate product setIn product pjAttribute level on k-th of product attribute
Lower bound and the attribute level upper bound,Indicate product setIn product quantity.
Further, it in step (3), obtains in any one user uiAttention rate under, any two product pj1With
pj2The method of product similarity include:
Sentiment analysis is carried out to review record to obtain public users for the emotion of each product attribute and score；
According to product pj1And pj2Attribute level lower bound on k-th of product attributeWithCalculate user uiIt is right
Product pj1And pj2Attribute level lower bound preference similarity are as follows:
According to product pj1And pj2Attribute level lower bound on k-th of product attributeWithCalculate user uiIt is right
Product pj1And pj2The attribute level upper bound preference similarity are as follows:
According to preference similarity KIL (pj1,pj2,ui) and preference similarity KIU (pj1,pj2,ui) calculate product
pj1And pj2Relative to user uiProduct similarity are as follows:
Wherein, Yj1,kAnd Yj2,kRespectively public users are to product pj1And pj2The emotion scoring of k-th of product attribute, Tq1With
Tq2Respectively product pj1And pj2The timeslice of appearance,WithRespectively timeslice Tq1And Tq2Middle time point,For time pointWithBetween time interval, ε1And ε2It is preset adjustment factor.
A large amount of subjectivity speeches with personal emotion, therefore the sight to implying in review record are often implied in review record
Point and emotion are analyzed in real time, are gone out public users with efficient detection and are taken in time in the past section to the emotional attitude of the product
To effectively assess the traction etc. of the product；When calculating product similarity, by user to different product attribute
Degree of concern is taken into account, and the same user of reflection for enabling to product similarity calculated more accurate is for different product
Preference similarity degree；In addition, the time interval between timeslice described in product is examined when calculating product similarity
Including worry, influence of the time factor to product similarity can be effectively embodied, to improve the accuracy of Products Show.
Further, in step (4), any one user u is predictediTo any one product pjScoringMethod packet
It includes:
It obtains relative to user ui, with product pjThe maximum N of product similarity1A product, to obtain product set N
(pj)；By product set N (pj) in each product and product pjRelative to user uiProduct similarity be added, obtain the first canonical
Change coefficient are as follows:The gain of product similarity is calculated according to the first regularization coefficient
Are as follows:
It obtains and user uiThe maximum N of user's similarity2A user, to obtain user set N (ui)；By product collection
Close N (ui) in each user and user uiUser's similarity be added, obtain the second regularization coefficient are as follows:The gain of user's similarity is calculated according to the second regularization coefficient are as follows:
According to the gain of product similarity and user similarity prediction of gain user uiTo product pjScoring are as follows:
Wherein, α and γ is balance parameters, and V and Q respectively indicate the user's recessive character matrix and production that matrix decomposition obtains
Product recessive character matrix, ViAnd VhRespectively indicate the i-th row and h row in user's recessive character matrix V, QcAnd QjIt respectively indicates
C row and jth row in product recessive character matrix Q.
Further, in step (4), the majorized function according to determined by prediction result are as follows:
Wherein, i and j respectively indicates Customs Assigned Number and product number, RijIndicate the element of the i-th row jth column in rating matrix,
λ is regularization parameter, ‖ ‖FRepresenting matrix F- norm.
The scoring to product according to product similarity and user's similarity prediction user, can according to the competitiveness of product with
The characteristic of time dynamic improves the accuracy of prediction scoring.
Further, in step (4), majorized function is determined according to prediction result, comprising:
The attribute value upper bound of any k-th of product attribute is obtained respectivelyWith attribute value lower boundTo calculate k-th
The weight of product attribute are as follows:
Obtain any two product pj1And pj2In the attribute value p of k-th of product attributej1,kAnd pj2,k, to calculate product pj1
And pj2Dominant character similarity are as follows:With dominant character similarity Sj1,j2For
The element of 1 row jth 2 of matrix jth column, obtains dominant character similarity matrix S, for recording the dominant character of any two product
Similarity；
Product p is calculated according to product recessive character matrix Qj1And pj2Recessive character similarity are as follows:With recessive character similarity Mj1,j2For the element of 1 row jth 2 of matrix jth column, recessive character phase is obtained
Like degree matrix M, for recording the recessive character similarity of any two product；
Recessive character similarity matrix M and dominant character similarity matrix S are subtracted each other, feature difference matrix Δ is obtained, is used
Difference between the recessive character similarity and dominant character phase similarity of record any two product；
Bound term is constructed according to feature difference matrix Δ, according to user's recessive character matrix V and product recessive character matrix V
Regularization term is constructed, iteration item is constructed according to prediction result, and optimization letter is determined according to bound term, regularization term and iteration item
Number；
Wherein, bound term is for promoting recommendation effect, and regularization term is for preventing over-fitting, and iteration item is for describing reality
Difference between scoring and prediction scoring.
Majorized function is constructed according to the consistency between the dominant character similarity and recessive character similarity of product, it can
More precisely preference of the prediction user to different product.
Further, in step (4), identified majorized function are as follows:
Wherein, i and j respectively indicates Customs Assigned Number and product number, RijIndicate the element of the i-th row jth column in rating matrix,
Q indicates timeslice number, | T | indicate timeslice sum, λ is regularization parameter, and β is balance parameters, EqFor companion matrix, it is used for
All products occurred in q-th of timeslice, ‖ ‖ are extracted from feature difference matrix ΔFRepresenting matrix F- norm.
In general, contemplated above technical scheme through the invention, can obtain it is following the utility model has the advantages that
(1) the Products Show method provided by the present invention based on time dynamic characteristic, by the division of timeslice, and is pressed
The attribute relative level of product is obtained according to timeslice, and product similarity and user are calculated according to the attribute relative level of product
Similarity can fully consider the characteristic of product competitiveness dynamic change at any time in Products Show, push away to improve product
The accuracy recommended.
(2) the Products Show method provided by the present invention based on time dynamic characteristic, in Products Show by each user
The degree of concern of different product attribute is taken into account, can more precisely obtain specific user for certain product attributes
Sentiment orientation, to improve the accuracy of Products Show.
(3) the Products Show method provided by the present invention based on time dynamic characteristic, according to the dominant character phase of product
Majorized function is constructed like the consistency between degree and recessive character similarity, can more precisely predict user to different product
Preference, to improve the accuracy of Products Show.
Generally speaking, the Products Show method provided by the present invention based on time dynamic characteristic, can be in the competing of product
Strive the accuracy that power effectively improves Products Show in the case where dynamic change at any time.
Detailed description of the invention
Fig. 1 is the Products Show method flow diagram provided in an embodiment of the present invention based on time dynamic characteristic.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below
Not constituting a conflict with each other can be combined with each other.
The present invention provides a kind of Products Show method based on time dynamic characteristic, Integral Thought is: will be continuous
Period be divided into discrete timeslice, pass through attribute relative level measure product each attribute in affiliated timeslice shelves
It is secondary, preferably to compare the product of different times, and any two user is further calculated according to the attribute relative level of product
Between product similarity relative to same user of user's similarity and any two product, thus abundant in Products Show
The characteristic for considering the competitiveness dynamic change at any time of product, effectively improves the accuracy of Products Show；In addition, according to product
Consistency between dominant character similarity and recessive character similarity constructs majorized function, more precisely to predict user couple
The preference of different product, to improve the accuracy of Products Show.
Products Show method provided by the invention based on time dynamic characteristic, as shown in Figure 1, comprising:
(1) purchaser record, review record and the product description information for collecting user in target time section, by target time section
It is divided into time slice sequence, and user is generated to the rating matrix of product according to review record；
The purchaser record of user includes which product user bought；The review record of user include user to product and
The comment information of product attribute；Product description information include product occur time, attribute value of the product on each product attribute,
And the maximum value and minimum value of each product attribute value；
The partition strategy of timeslice, can according to product competitiveness at any time dynamic change characteristic with week, the moon or other when
Between unit divide, it is preferably more different so as to study the access behavior of user according to a specific module
The product in period；
(2) according to the attribute relative level of each product of product description information acquisition, for indicating each attribute of product affiliated
Class in timeslice；
Attribute relative level includes attribute level lower bound and the attribute level upper bound；The affiliated timeslice of product is production
Timeslice where time point；
In an optional embodiment, step (2) is specifically included:
For any one product pj, according to product description information acquisition product pjAffiliated timeslice TqInterior production is owned
Product, to obtain product setProduct setContain product pjAnd with product pjAll products of the same period；
Obtain product setIn be weaker than product p in the configuration of any k-th of product attributejProduct quantity Ninf
(pj), to calculate product pjAttribute level lower bound on k-th of product attribute are as follows:
Obtain product setIn be better than product p in the configuration of k-th of product attributejProduct quantity Nsup(pj), with
Calculate product pjThe attribute level upper bound on k-th of product attribute are as follows:
Attribute level lower boundIllustrate that all products of the same period are weaker than production in the configuration on k-th of product attribute
Product pjProbability, the attribute level upper boundIllustrate that all products of the same period are better than production in the configuration of k-th of product attribute
Product pjProbability；
According to attribute level lower boundWith the attribute level upper boundObtain product pjCategory on k-th of product attribute
Property relative level；
Wherein,Indicate product setIn product quantity；
For any one product, it is weaker than by attribute configuration in the same timeslice of acquisition general shared by the product of the product
Rate obtains the attribute level lower bound of product, and is better than shared by the product of the product by obtaining attribute configuration in same timeslice
Probability, obtain the attribute level upper bound of product, can clearly react in the class where each attribute of the product of a certain period,
To the characteristic of the competitiveness dynamic change at any time of more enough preferably reactor products；
(3) according to collected information and the attribute relative level of product, the user for obtaining any two user is similar
The product similarity of degree and any two product under same user's attention rate；
In order to more accurately obtain specific user for the Sentiment orientation of certain product attributes, pushed away with further increasing product
The accuracy recommended, step (3) further include:
It obtains the number that each product attribute occurs in the review record of each user respectively according to review record, and thus obtains
Each user is obtained to the degree of concern of different product attribute, calculation formula is as follows:
Wherein, i indicates that Customs Assigned Number, k indicate product attribute number, ci,kIndicate k-th of product attribute in i-th of user
Review record in the number that occurs, Xi,kI-th of user is indicated to the degree of concern of k-th of product attribute, N indicates scoring area
Between number；Under most of points-scoring systems, the value of N is 5, such as Taobao and Amazon are all 5 grades of evaluation systems, share 5 and comment
By stages；
In an optional embodiment, in step (3), any two user u is obtainedi1And ui2User's similarity
Method include:
Obtain user u respectively according to purchaser recordi1And ui2The product bought, and the product bought according to each user
Attribute relative level obtain the level of consumption of each user；The level of consumption of each user includes: user ui1Level of consumption lower boundWith the level of consumption upper boundAnd user ui2Level of consumption lower boundWith the level of consumption upper bound
According to level of consumption lower boundWithCalculate user ui1And ui2Level of consumption lower bound similarity are as follows:
According to the level of consumption upper boundWithCalculate user ui1And ui2The level of consumption upper bound similarity are as follows:
According to similarity KLL (ui1,ui2) and similarity KLU (ui1,ui2) calculate user ui1With user ui2User it is similar
Degree are as follows:
Wherein, ε3For preset adjustment factor, K indicates product attribute sum；
For any one user ui, according to user uiThe attribute relative level for the product bought obtains user uiDisappear
Water wasting is flat, comprising:
It obtains by user uiThe product set that the product bought is constituted
According to product setIn attribute level lower bound of each product on k-th of product attribute calculate user uiDisappear
The flat lower bound of water wasting are as follows:
According to product setIn the attribute level upper bound of the product on k-th of product attribute calculate user uiConsumption
The horizontal upper bound are as follows:
According to level of consumption lower boundWith the level of consumption upper boundObtain user uiThe level of consumption；
Wherein,WithRespectively indicate product setIn product pjAttribute level on k-th of product attribute
Lower bound and the attribute level upper bound,Indicate product setIn product quantity；
Class as where the attribute relative level of product has measured product in each attribute of a certain period, is purchased by user
The attribute relative level of the product bought calculates the level of consumption of user, and according to customer consumption level calculation user's similarity,
Influence of the time factor to user's similarity can be fully considered, when calculating user's similarity, also by user to different product
The degree of concern of attribute is taken into account, so that user's similarity calculated, enables to user's similarity calculated more
The similarity degree between user is accurately portrayed, to improve the accuracy of Products Show；
In an optional embodiment, in step (3), any one user u is obtainediAttention rate under, any two
A product pj1And pj2The method of product similarity include:
Sentiment analysis is carried out to review record to obtain public users for the emotion of each product attribute and score；
A large amount of subjectivity speeches with personal emotion, therefore the sight to implying in review record are often implied in review record
Point and emotion are analyzed in real time, are gone out public users with efficient detection and are taken in time in the past section to the emotional attitude of the product
To effectively assess the traction etc. of the product；In the present embodiment, used sentiment analysis method is word-based
The sentiment analysis method of allusion quotation, this method fully consider context relation, in conjunction with negative and degree adverb, sentence pattern clause to emotion
Influence, carry out Sentiment orientation category division using supervised learning, can accurate judgement user to the emotion state of product different attribute
Degree；
According to product pj1And pj2Attribute level lower bound on k-th of product attributeWithCalculate user uiIt is right
Product pj1And pj2Attribute level lower bound preference similarity are as follows:
According to product pj1And pj2Attribute level lower bound on k-th of product attributeWithCalculate user uiIt is right
Product pj1And pj2The attribute level upper bound preference similarity are as follows:
According to preference similarity KIL (pj1,pj2,ui) and preference similarity KIU (pj1,pj2,ui) calculate product
pj1And pj2Relative to user uiProduct similarity are as follows:
Wherein, Yj1,kAnd Yj2,kRespectively public users are to product pj1And pj2The emotion scoring of k-th of product attribute, Tq1With
Tq2Respectively product pj1And pj2The timeslice of appearance,WithRespectively timeslice Tq1And Tq2Middle time point,For time pointWithBetween time interval, ε1And ε2It is preset adjustment factor.
When calculating product similarity, degree of concern of the user to different product attribute is taken into account, institute is enabled to
Similarity degree of the more accurate same user of reflection of the product similarity of calculating for the preference of different product；In addition,
When calculating product similarity, the time interval between timeslice described in product is taken into account, can effectively embody the time because
Influence of the element to product similarity, to improve the accuracy of Products Show；
(4) scoring according to user's similarity and product similarity prediction user to product, and determined according to prediction result
Majorized function；
In an optional embodiment, in step (4), any one user u is predictediTo any one product pj's
ScoringMethod include:
It obtains relative to user ui, with product pjThe maximum N of product similarity1A product, to obtain product set N
(pj)；By product set N (pj) in each product and product pjRelative to user uiProduct similarity be added, obtain the first canonical
Change coefficient are as follows:The gain of product similarity is calculated according to the first regularization coefficient
Are as follows:
It obtains and user uiThe maximum N of user's similarity2A user, to obtain user set N (ui)；By product collection
Close N (ui) in each user and user uiUser's similarity be added, obtain the second regularization coefficient are as follows:The gain of user's similarity is calculated according to the second regularization coefficient are as follows:
According to the gain of product similarity and user similarity prediction of gain user uiTo product pjScoring are as follows:
Wherein, α and γ is balance parameters, for adjusting influence of the various pieces to final fitting result；V and Q difference
The user's recessive character matrix and product recessive character matrix that representing matrix decomposes, ViAnd VhIt is recessive special to respectively indicate user
Levy the i-th row and h row in matrix V, QcAnd QjRespectively indicate the c row and jth row in product recessive character matrix Q；
In the present embodiment, in step (4), the majorized function according to determined by prediction result are as follows:
Wherein, i and j respectively indicates Customs Assigned Number and product number, RijIndicate the element of the i-th row jth column in rating matrix,
λ is regularization parameter, ‖ ‖FRepresenting matrix F- norm；
The scoring to product according to product similarity and user's similarity prediction user, can according to the competitiveness of product with
The characteristic of time dynamic improves the accuracy of prediction scoring；
(5) rating matrix is decomposed, and matrix decomposition result is adjusted according to majorized function, to obtain so that
The smallest target user's recessive character matrix of majorized function value and target product recessive character matrix, to generate each user's
Recommendation list completes the Products Show to user；
In the present embodiment, matrix decomposition result is adjusted according to majorized function, used method is random
Gradient descent method；After obtaining target user's recessive character matrix and target product recessive character matrix, can according to this two
Scoring of each user of a Matrix prediction for each product, to choose one or more produce according to actual application demand
Product recommend user.
In another real-time example of the invention, the Products Show method provided by the present invention based on time dynamic characteristic
It is similar to the aforementioned embodiment, the difference is that, in the present embodiment, in step (4), optimization letter is determined according to prediction result
Number, comprising:
The attribute value upper bound of any k-th of product attribute is obtained respectivelyWith attribute value lower boundTo calculate k-th
The weight of product attribute are as follows:
Obtain any two product pj1And pj2In the attribute value p of k-th of product attributej1,kAnd pj2,k, to calculate product pj1
And pj2Dominant character similarity are as follows:With dominant character similarity
Sj1,j2For the element of 1 row jth 2 of matrix jth column, dominant character similarity matrix S is obtained, for recording the aobvious of any two product
Property characteristic similarity；
Product p is calculated according to product recessive character matrix Qj1And pj2Recessive character similarity are as follows:With recessive character similarity Mj1,j2For the element of 1 row jth 2 of matrix jth column, recessive character phase is obtained
Like degree matrix M, for recording the recessive character similarity of any two product；
Recessive character similarity matrix M and dominant character similarity matrix S are subtracted each other, feature difference matrix Δ is obtained, is used
Difference between the recessive character similarity and dominant character phase similarity of record any two product；
Bound term is constructed according to feature difference matrix Δ, for promoting recommendation effect；According to user's recessive character matrix V and
Product recessive character matrix V constructs regularization term, for preventing over-fitting；Iteration item is constructed according to prediction result, for describing
Difference between practical scoring and prediction scoring；And majorized function is determined according to bound term, regularization term and iteration item；
In the present embodiment, in step (4), identified majorized function are as follows:
Wherein, i and j respectively indicates Customs Assigned Number and product number, RijIndicate the element of the i-th row jth column in rating matrix,
Q indicates timeslice number, | T | indicate timeslice sum, λ is regularization parameter, and β is balance parameters, EqFor companion matrix, it is used for
All products occurred in q-th of timeslice are extracted from feature difference matrix Δ, | | | |FRepresenting matrix F- norm；
Due to having divided timeslice, recessive character similarity matrix M, dominant character similarity matrix S and feature difference
Row and column difference in matrix Δ is orderly, and the product of contemporaneity is located in the same partitioning of matrix, therefore, passes through construction one
Simple companion matrix Eq, all products occurred in q-th of timeslice can be extracted from feature difference matrix Δ；Example
Such as, there are four product p altogether1~p4, corresponding feature difference matrix are as follows:
Wherein, product p1And p2It is all the product for belonging to first timeslice, in order to by product p of the same period1And p2Phase
It closes information to extract, the companion matrix of construction are as follows:
The transposition of corresponding companion matrix are as follows:
It can be obtained by matrix multiplication:
It has obtained in product of the same period, between the recessive character similarity and dominant character phase similarity of two products
Consistency；When extracting the information of product in different time piece, companion matrix can refer to above method construction, not do herein superfluous
It states；
The above method constructs optimization according to the consistency between the dominant character similarity and recessive character similarity of product
Function can more precisely predict user to the preference of different product.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all include
Within protection scope of the present invention.
Claims (10)
1. a kind of Products Show method based on time dynamic characteristic characterized by comprising
(1) purchaser record, review record and the product description information for collecting user in target time section, by the target time section
It is divided into time slice sequence, and user is generated to the rating matrix of product according to the review record；
(2) according to the attribute relative level of each product of product description information acquisition, for indicating each attribute of product in product
Class in affiliated timeslice；
(3) according to collected information and the attribute relative level of product, user's similarity of any two user is obtained, with
And product similarity of any two product under same user's attention rate；
(4) scoring according to user's similarity and product similarity prediction user to product, and according to prediction result
Determine majorized function；
(5) rating matrix is decomposed, and matrix decomposition result is adjusted according to the majorized function, to obtain
So that the smallest target user's recessive character matrix of the majorized function value and target product recessive character matrix, to generate
The recommendation list of each user completes the Products Show to user；
Wherein, the affiliated timeslice of product is the timeslice where the time point of production, and the attribute relative level includes belonging to
The horizontal lower bound of property and the attribute level upper bound.
2. as described in claim 1 based on the Products Show method of time dynamic characteristic, which is characterized in that the step (2)
Include:
For any one product pj, according to product p described in the product description information acquisitionjTimeslice TqInterior production is owned
Product, to obtain product set
Obtain the product setIn be weaker than the product p in the configuration of any k-th of product attributejProduct quantity
Ninf(pj), to calculate the product pjAttribute level lower bound on k-th of product attribute are as follows:
Obtain the product setIn be better than the product p in the configuration of k-th of product attributejProduct quantity Nsup
(pj), to calculate the product pjThe attribute level upper bound on k-th of product attribute are as follows:
According to the attribute level lower boundWith the attribute level upper boundObtain the product pjIn k-th of product attribute
On attribute relative level；
Wherein,Indicate the product setIn product quantity.
3. as described in claim 1 based on the Products Show method of time dynamic characteristic, which is characterized in that the step (3)
Further include:
The number that each product attribute occurs in the review record of each user is obtained respectively according to the review record, and is thus obtained
Each user is obtained to the degree of concern of different product attribute, calculation formula is as follows:
Wherein, i indicates that Customs Assigned Number, k indicate product attribute number, ci,kIndicate k-th of product attribute commenting in i-th user
By the number occurred in record, Xi,kI-th of user is indicated to the degree of concern of k-th of product attribute, N indicates scoring interval number.
4. as claimed in claim 3 based on the Products Show method of time dynamic characteristic, which is characterized in that the step (3)
In, obtain any two user ui1And ui2The method of user's similarity include:
The user u is obtained respectively according to the purchaser recordi1And ui2The product bought, and the production bought according to each user
The attribute relative level of product obtains the level of consumption of each user；The level of consumption of each user includes: the user ui1Disappear
The flat lower bound of water wastingWith the level of consumption upper boundAnd the user ui2Level of consumption lower boundIn the level of consumption
Boundary
According to the level of consumption lower boundWithCalculate the user ui1And ui2Level of consumption lower bound similarity are as follows:
According to the level of consumption upper boundWithCalculate the user ui1And ui2The level of consumption upper bound similarity are as follows:
According to the similarity KLL (ui1,ui2) and the similarity KLU (ui1,ui2) calculate the user ui1With the user
ui2User's similarity are as follows:
Wherein, ε3For preset adjustment factor, k indicates product attribute number, and K indicates product attribute sum.
5. as claimed in claim 4 based on the Products Show method of time dynamic characteristic, which is characterized in that for any one
User ui, according to the user uiThe attribute relative level for the product bought obtains the user uiThe level of consumption, comprising:
It obtains by the user uiThe product set that the product bought is constituted
According to the product setIn attribute level lower bound of each product on k-th of product attribute calculate the user ui
Level of consumption lower bound are as follows:
According to the product setIn the attribute level upper bound of the product on k-th of product attribute calculate the user ui's
The level of consumption upper bound are as follows:
According to the level of consumption lower boundWith the level of consumption upper boundObtain the user uiThe level of consumption；
Wherein,WithRespectively indicate the product setIn product pjAttribute level on k-th of product attribute
Lower bound and the attribute level upper bound,Indicate the product setIn product quantity.
6. as claimed in claim 3 based on the Products Show method of time dynamic characteristic, which is characterized in that the step (3)
In, it obtains in any one user uiAttention rate under, any two product pj1And pj2The method of product similarity include:
Sentiment analysis is carried out to the review record to obtain public users for the emotion of each product attribute and score；
According to the product pj1And pj2Attribute level lower bound on k-th of product attributeWithCalculate the user
uiTo the product pj1And pj2Attribute level lower bound preference similarity are as follows:
According to the product pj1And pj2Attribute level lower bound on k-th of product attributeWithCalculate the user
uiTo the product pj1And pj2The attribute level upper bound preference similarity are as follows:
According to the preference similarity KIL (pj1,pj2,ui) and the preference similarity KIU (pj1,pj2,ui) calculate
The product pj1And pj2Relative to the user uiProduct similarity are as follows:
Wherein, Yj1,kAnd Yj2,kRespectively public users are to the product pj1And pj2The emotion scoring of k-th of product attribute, Tq1With
Tq2The respectively described product pj1And pj2The timeslice of appearance,WithThe respectively described timeslice Tq1And Tq2Interlude
Point,For the time pointWithBetween time interval, ε1And ε2It is preset adjustment factor.
7. as described in claim 1 based on the Products Show method of time dynamic characteristic, which is characterized in that the step (4)
In, predict any one user uiTo any one product pjScoringMethod include:
It obtains relative to the user ui, with the product pjThe maximum N of product similarity1A product, to obtain product collection
Close N (pj)；By the product set N (pj) in each product and the product pjRelative to the user uiProduct similarity phase
Add, obtain the first regularization coefficient are as follows:According to first regularization coefficient
Calculate the gain of product similarity are as follows:
It obtains and the user uiThe maximum N of user's similarity2A user, to obtain user set N (ui)；By the production
Product set N (ui) in each user and the user uiUser's similarity be added, obtain the second regularization coefficient are as follows:The gain of user's similarity is calculated according to second regularization coefficient are as follows:
According to user u described in the product similarity gain and user's similarity prediction of gainiTo the product pjScoring
Are as follows:
Wherein, α and γ is balance parameters, and V and Q respectively indicate user's recessive character matrix that matrix decomposition obtains and product is hidden
Property eigenmatrix, ViAnd VhRespectively indicate the i-th row and h row in user's recessive character matrix V, QcAnd QjIt respectively indicates
C row and jth row in the product recessive character matrix Q.
8. as claimed in claim 7 based on the Products Show method of time dynamic characteristic, which is characterized in that the step (4)
In, the majorized function according to determined by prediction result are as follows:
Wherein, i and j respectively indicates Customs Assigned Number and product number, RijIndicate the element that the i-th row jth arranges in the rating matrix,
λ is regularization parameter, | | | |FRepresenting matrix F- norm.
9. as claimed in claim 7 based on the Products Show method of time dynamic characteristic, which is characterized in that the step (4)
In, majorized function is determined according to prediction result, comprising:
The attribute value upper bound of any k-th of product attribute is obtained respectivelyWith attribute value lower boundTo calculate k-th of product
The weight of attribute are as follows:
Obtain any two product pj1And pj2In the attribute value p of k-th of product attributej1,kAnd pj2,k, to calculate the product pj1
And pj2Dominant character similarity are as follows:With the dominant character similarity
Sj1,j2For the element of 1 row jth 2 of matrix jth column, dominant character similarity matrix S is obtained, for recording the aobvious of any two product
Property characteristic similarity；
The product p is calculated according to the product recessive character matrix Qj1And pj2Recessive character similarity are as follows:With the recessive character similarity Mj1,j2For the element of 1 row jth 2 of matrix jth column, recessive spy is obtained
Similarity matrix M is levied, for recording the recessive character similarity of any two product；
The recessive character similarity matrix M and the dominant character similarity matrix S are subtracted each other, feature difference matrix is obtained
Δ, for recording the difference between the recessive character similarity of any two product and dominant character phase similarity；
Bound term is constructed according to the feature difference matrix Δ, according to user's recessive character matrix V and product recessive character
Matrix V construct regularization term, according to the prediction result construct iteration item, and according to the bound term, the regularization term and
The iteration item determines the majorized function；
Wherein, the bound term is for promoting recommendation effect, and for preventing over-fitting, the iteration item is used for the regularization term
Difference between the practical scoring of description and prediction scoring.
10. as claimed in claim 9 based on the Products Show method of time dynamic characteristic, which is characterized in that the step (4)
In, identified majorized function are as follows:
Wherein, i and j respectively indicates Customs Assigned Number and product number, RijIndicate the element that the i-th row jth arranges in the rating matrix,
Q indicates timeslice number, | T | indicate timeslice sum, λ is regularization parameter, and β is balance parameters, EqFor companion matrix, it is used for
All products occurred in q-th of timeslice are extracted from the feature difference matrix Δ, | | | |FRepresenting matrix F- norm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811544768.XA CN109670914B (en) | 2018-12-17 | 2018-12-17 | Product recommendation method based on time dynamic characteristics |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811544768.XA CN109670914B (en) | 2018-12-17 | 2018-12-17 | Product recommendation method based on time dynamic characteristics |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109670914A true CN109670914A (en) | 2019-04-23 |
| CN109670914B CN109670914B (en) | 2020-11-17 |
Family
ID=66143936
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811544768.XA Active CN109670914B (en) | 2018-12-17 | 2018-12-17 | Product recommendation method based on time dynamic characteristics |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109670914B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110852846A (en) * | 2019-11-11 | 2020-02-28 | 京东数字科技控股有限公司 | Processing method and device for recommended object, electronic equipment and storage medium |
| CN111339439A (en) * | 2020-02-18 | 2020-06-26 | 广东工业大学 | Collaborative filtering recommendation method and device fusing comment text and time sequence effect |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8214264B2 (en) * | 2005-05-02 | 2012-07-03 | Cbs Interactive, Inc. | System and method for an electronic product advisor |
| US8781916B1 (en) * | 2012-05-18 | 2014-07-15 | Google Inc. | Providing nuanced product recommendations based on similarity channels |
| CN105956089A (en) * | 2016-05-03 | 2016-09-21 | 桂林电子科技大学 | Recommendation method capable of aiming at classification information with items |
| CN106294859A (en) * | 2016-08-22 | 2017-01-04 | 南京邮电大学盐城大数据研究院有限公司 | A kind of item recommendation method decomposed based on attribute coupling matrix |
| CN108090794A (en) * | 2016-11-22 | 2018-05-29 | 阿里巴巴集团控股有限公司 | A kind of merchandise items sort method and device based on dynamic sliding time window |
| CN108681739A (en) * | 2018-03-26 | 2018-10-19 | 安徽师范大学 | One kind recommending method based on user feeling and time dynamic tourist famous-city |
-
2018
- 2018-12-17 CN CN201811544768.XA patent/CN109670914B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8214264B2 (en) * | 2005-05-02 | 2012-07-03 | Cbs Interactive, Inc. | System and method for an electronic product advisor |
| US8781916B1 (en) * | 2012-05-18 | 2014-07-15 | Google Inc. | Providing nuanced product recommendations based on similarity channels |
| CN105956089A (en) * | 2016-05-03 | 2016-09-21 | 桂林电子科技大学 | Recommendation method capable of aiming at classification information with items |
| CN106294859A (en) * | 2016-08-22 | 2017-01-04 | 南京邮电大学盐城大数据研究院有限公司 | A kind of item recommendation method decomposed based on attribute coupling matrix |
| CN108090794A (en) * | 2016-11-22 | 2018-05-29 | 阿里巴巴集团控股有限公司 | A kind of merchandise items sort method and device based on dynamic sliding time window |
| CN108681739A (en) * | 2018-03-26 | 2018-10-19 | 安徽师范大学 | One kind recommending method based on user feeling and time dynamic tourist famous-city |
Non-Patent Citations (4)
| Title |
|---|
| 冯小翼: "在线评论的产品属性提取与情感分析研究", 《中国优秀硕士学位论文全文数据库（电子期刊） 经济与管理科学辑》 * |
| 朱思丞等: "推荐算法时间动态特性研究进展", 《工业控制计算机》 * |
| 杨静等: "顾客偏好的动态挖掘算法", 《信息与控制》 * |
| 胡新明: "基于商品属性的电子商务推荐系统研究", 《中国博士学位论文全文数据库（电子期刊） 经济与管理科学辑》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110852846A (en) * | 2019-11-11 | 2020-02-28 | 京东数字科技控股有限公司 | Processing method and device for recommended object, electronic equipment and storage medium |
| CN111339439A (en) * | 2020-02-18 | 2020-06-26 | 广东工业大学 | Collaborative filtering recommendation method and device fusing comment text and time sequence effect |
| CN111339439B (en) * | 2020-02-18 | 2023-05-09 | 广东工业大学 | Collaborative filtering recommendation method and device fusing comment text and time sequence effect |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109670914B (en) | 2020-11-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102902691B (en) | Recommend method and system | |
| Amrhein | Searching for the elusive aggregation effect: evidence from statistical simulations | |
| Zibarzani et al. | Customer satisfaction with Restaurants Service Quality during COVID-19 outbreak: A two-stage methodology | |
| Al-Gasawneh et al. | The relationship between perceived destination image, social media interaction and travel intentions relating to Neom city | |
| CN107016026B (en) | User tag determination method, information push method, user tag determination device, information push device | |
| CN103927675B (en) | Judge the method and device of age of user section | |
| CN105630946B (en) | A kind of field intersection recommended method and device based on big data | |
| EP2474945A1 (en) | Analyzing transactional data | |
| Scholz et al. | Using PageRank for non-personalized default rankings in dynamic markets | |
| CN103700005A (en) | Association-rule recommending method based on self-adaptive multiple minimum supports | |
| CN111724235A (en) | Online commodity recommendation method based on user novelty | |
| CN109493151A (en) | Method for Sales Forecast method and system | |
| CN110046228A (en) | Short text subject identifying method and system | |
| Chan et al. | IoT data acquisition in fashion retail application: Fuzzy logic approach | |
| Greenstein-Messica et al. | Machine learning and operation research based method for promotion optimization of products with no price elasticity history | |
| CN109670914A (en) | A kind of Products Show method based on time dynamic characteristic | |
| Al-Gunaid et al. | A survey of fuzzy cognitive maps forecasting methods | |
| CN113487117A (en) | Method and system for simulating e-commerce user behavior data based on multi-dimensional user portrait | |
| Ayuningtyas et al. | The pattern of online purchase behavior by millennial family | |
| Lipianina-Honcharenko et al. | Intelligent method for forming the consumer basket | |
| Halkin | Assessing the utility of retailer based on generalized costs of end-consumers | |
| Kamil et al. | THE INFLUENCE OF MATERIALISM SELF CONTROL AND FINANCIAL ATTITUDE ON FINANCIAL MANAGEMENT BEHAVIOR IN ONLINE MARKETPLACE WITH IMPULSIVE BUYING AS AN INTERVENING VARIABLE | |
| Xiaona | Personalized recommendation model for mobile E-commerce users | |
| KR20150079353A (en) | Apparatus and method for measuring brand personality | |
| Al-Basha | Forecasting Retail Sales Using Google Trends and Machine Learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |