CN106296242A - A kind of generation method of commercial product recommending list in ecommerce and the system of generation - Google Patents
A kind of generation method of commercial product recommending list in ecommerce and the system of generation Download PDFInfo
- Publication number
- CN106296242A CN106296242A CN201510268039.6A CN201510268039A CN106296242A CN 106296242 A CN106296242 A CN 106296242A CN 201510268039 A CN201510268039 A CN 201510268039A CN 106296242 A CN106296242 A CN 106296242A
- Authority
- CN
- China
- Prior art keywords
- commodity
- purchase probability
- model
- data
- behavior
- 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.)
- Pending
Links
Abstract
The invention discloses a kind of generation method of commercial product recommending list in ecommerce, the method comprises the following steps: S11 gathers the characteristic of user, and is merged by each terminal data, the real-time estimate characteristic vector after being merged;S12 calculates the purchase probability of behavior commodity;The purchase probability of the behavior commodity that S12 is obtained by S13 is modified, the purchase probability of the behavior commodity after being merged;S14, according to the purchase probability of revised behavior commodity, calculates the purchase probability of similar dependent merchandise, and sorts according to buying rate size, generate commercial product recommending list.Also disclose a kind of generation system of commercial product recommending list in ecommerce simultaneously.This production method and production system can improve the precision of prediction of commercial product recommending list.
Description
Technical field
The present invention relates to e-commerce field, it particularly relates to a kind of commercial product recommending row in ecommerce
The generation method of table and the system of generation.
Background technology
At present, the standard buying prediction algorithm of internet electronic business is based on multi-dimensional feature data source, uses
Logistic regression model, model learning training method is substantially maximum likelihood algorithm, or gradient
Descent algorithm.The data source that the many employings of forecast model are unified, unified model, the algorithm predicts model of unit/terminal.
In current commodity purchasing prediction, the use of single model is the most universal, but also exposes it simultaneously
It is difficult to describe the problem that whole relations of complicated business, precision of prediction and robustness are not fully up to expectations comprehensively.Equally,
Unit/terminal and unified data source, be also unfavorable for improving the precision of prediction of commodity purchasing.
Summary of the invention
Technical problem: the embodiment of the present invention is to be solved be technical problem is that: provide one in ecommerce
The generation method of commercial product recommending list and the system of generation, it is possible to increase the precision of prediction of commercial product recommending list.
Technical scheme: for solving above-mentioned technical problem, on the one hand, the present embodiment provides one for ecommerce
The generation method of middle commercial product recommending list, the method comprises the following steps:
S11 gathers the characteristic of user, and by each terminal data Feature Fusion, after being merged the most in advance
Survey characteristic vector;
S12 calculates the purchase probability of behavior commodity;
The purchase probability of the behavior commodity that S12 is obtained by S13 is modified, and obtains revised behavior commodity
Purchase probability;
S14, according to the purchase probability of revised behavior commodity, calculates the purchase probability of similar dependent merchandise, and
Sort according to buying rate size, generate commercial product recommending list.
As a kind of embodiment, in described step S11, described characteristic derives from PC end, WAP
Terminal data under data and third party's line in data, and third party's line in end and/or APP end station;Described
Characteristic includes history offline feature data and real-time characteristic data.
As a kind of embodiment, described real-time characteristic data acquisition resolves access log, click logs, exposure
The method of daily record, event log and/or order daily record obtains.
As a kind of embodiment, described S11 farther includes: set up the mapping relations of terminal, gets through or seeks
The data cube computation passage of each terminal in road, by each characteristic by mapping relations association merged after feature to
Amount, by each terminal feature Vector Fusion, the real-time estimate characteristic vector after being merged.
As a kind of embodiment, described S12 farther includes: first classify user, and sets up each
The sub-model that class user is corresponding;Then the real-time estimate characteristic vector obtained by S11 brings point mould of relative users into
In type, calculate the purchase probability of behavior commodity.
As a kind of embodiment, described user carried out classification include: according to terminal type, user type and
Three kinds of dimensions of visitor's type, classify to user.
As a kind of embodiment, in described S13, the purchase probability of the behavior commodity obtaining S12 is repaiied
Just: use using the training result first of each sub-model and new corrected parameter as the skew of Model Fusion merge because of
Son, the purchase probability of the behavior commodity obtaining S12 is modified, and obtains revised commodity purchasing probability.
As a kind of embodiment, in described S14, the purchase probability of similar dependent merchandise is surveyed according to formula (1)
Calculate:
Score_i=Master_SPU_Pos*SKU_Score_i/max (SKU_Score_i) formula (1)
Wherein, Score_i represents the purchase probability of similar dependent merchandise, and Master_SPU_Pos represents behavior
The purchase probability of commodity, max (SKU_Score_i) represents degree of association peak in similar dependent merchandise list,
SKU_Score_i represents the degree of association of similar dependent merchandise SKU_i and behavior commodity.
On the other hand, the present embodiment provides a kind of generation system of commercial product recommending list in ecommerce, should
System includes:
Acquisition module: for gathering the characteristic of user, and each terminal data is merged, after being merged
Real-time estimate characteristic vector;
Computing module: for calculating the purchase probability of behavior commodity;
Fusion Module: for the purchase probability of behavior commodity is modified, the behavior commodity after being merged
Purchase probability;
Generation module: for calculating the purchase probability of similar dependent merchandise, and sort according to buying rate size, raw
Become commercial product recommending list.
As a kind of embodiment, in described acquisition module, characteristic derives from PC end, WAP end, APP
Terminal data under data and line in data in end station, and third party's line;Described characteristic include history from
Line characteristic and real-time characteristic data.
As a kind of embodiment, described acquisition module specifically for: set up the mapping relations of terminal, get through each
The data cube computation passage of terminal, the characteristic vector after each characteristic is merged by mapping relations association,
By each terminal feature Vector Fusion, the real-time estimate characteristic vector after being merged.
As a kind of embodiment, described computing module includes:
Classification submodule: for classifying user, forms catergories of user;
Modeling submodule: for all types of user is set up corresponding sub-model;
Calculating sub module: the real-time estimate characteristic vector obtained for acquisition module brings the sub-model of relative users into
In, calculate the purchase probability of behavior commodity.
As a kind of embodiment, in described Fusion Module, the purchase probability of behavior commodity is modified be,
Use the training result first of each sub-model and new corrected parameter as the skew fusion factor of Model Fusion, right
The purchase probability of behavior commodity is modified, the commodity purchasing probability after being merged.
As a kind of embodiment, in described generation module, the purchase probability of similar dependent merchandise is according to formula (1)
Measuring and calculating:
Score_i=Master_SPU_Pos*SKU_Score_i/max (SKU_Score_i) formula (1)
Wherein, Score_i represents the purchase probability of similar dependent merchandise, and Master_SPU_Pos represents behavior
The purchase probability of commodity, max (SKU_Score_i) represents degree of association peak in similar dependent merchandise list,
SKU_Score_i represents the degree of association of similar dependent merchandise SKU_i and behavior commodity.
Beneficial effect: compared with prior art, the commercial product recommending in ecommerce that the embodiment of the present invention provides
The generation method and system of list, gather the characteristic of multiple terminals, multi-data fusion, can solve data
Integrity, improves model accuracy.Meanwhile, set up multi-model according to user's classification, utilize multi-model to calculate respectively
Purchase probability, multi-model can improve the accuracy of prediction probability and the accuracy of recommendation.Multi-model merges permissible
The result normalization of multiple models so that the precision of the commercial product recommending list of generation is higher, more meets user's
Demand.It addition, multiple terminals data fusion can solve stability and the accuracy of the behavior recommendation of different terminals,
The user behavior data of the data cover of multiple terminals than single terminal end is more comprehensive, abundant, so that robust
Property is more preferable.
Accompanying drawing explanation
Fig. 1 is the FB(flow block) of one embodiment of the invention;
Fig. 2 is the FB(flow block) of model training in one embodiment of the invention;
Fig. 3 is the scattergram of model in one embodiment of the invention;
Fig. 4 is the structured flowchart of another embodiment of the present invention.
Detailed description of the invention
Below in conjunction with the accompanying drawings, the technical scheme of the embodiment of the present invention is described in detail.
As described in Figure 1, the present embodiment provides a kind of generation method of commercial product recommending list in ecommerce,
Comprise the following steps:
S11 gathers the characteristic of user, and is merged by each terminal data, and the real-time estimate after being merged is special
Levy vector;
S12 calculates the purchase probability of behavior commodity;
The purchase probability of the behavior commodity that S12 is obtained by S13 is modified, the behavior commodity after being merged
Purchase probability;
S14 calculates the purchase probability of similar dependent merchandise, and sorts according to buying rate size, generates commercial product recommending
List.
In above-described embodiment, the user characteristic data of collection derives from data in PC end, WAP end, APP end station,
And terminal data under data and line on third party's line.So, for data acquisition angle, these characteristic numbers
According to covering multiple terminal.With single terminal, such as WAP end compares, the Data Source model that the present embodiment gathers
Enclose wide.Multiple terminals Data Source, provides basic data more accurately for follow-up purchase probability prediction.As
A kind of preferred version, characteristic includes history offline feature data and real-time characteristic data.History offline feature
Data can select: member's attribute or label, member's loyalty, membership buying power, member's preference, Hui Yuanli
History accesses index etc..Member's history accesses index and adds shopping cart behavior number of times, mistake in the N of may is that over days
Go N days to add most PV numbers in collection behavior number of times, N days past PV number, the session of past N days,
Past N days website visiting duration, the maximum duration of a session of past N days, past N days browse commodity details
Most access details number of pages, past N days number of sessions, past N in number of pages, the session of past N days
Quantity on order etc. everyday.Real-time characteristic data acquisition is with resolving access log, click logs, exposure daily record, thing
Part daily record, the method for order daily record obtain.Such as, real-time characteristic data include that level Four page commodity details district clicks on
Number of times (including evaluating), commodity level Four page collection number of clicks, list page recommend number of clicks, list page to receive
Hide number of clicks, list page commodity number of clicks, user browses pv, search page collects number of clicks, search
Page recommends number of clicks, new dynamic promotional page commodity number of clicks, user static state commercial promotions page pv etc..
In S11, the determination method of real-time estimate characteristic vector is: set up the mapping relations of terminal, gets through each
The data cube computation passage of terminal, the characteristic vector after each characteristic is merged by mapping relations association,
By each terminal feature Vector Fusion, the real-time estimate characteristic vector after being merged.For non-member, Ke Yitong
Cross the characteristic index after the correspondence mappings relationship such as cookie, mobile phone string number is merged.For member,
Can be by association real time data and historical data, the characteristic indexs after being merged such as member's codings.By meeting
Member's relation, each terminal recognition code carry out data cube computation and association, obtain fused data.
After real-time estimate characteristic vector after being merged by S11, the purchase needing calculating behavior commodity is general
Rate.The measuring method of the purchase probability of the behavior commodity that the present embodiment provides is different from traditional method.The present embodiment
Use the Forecasting Methodology of multi-model, specifically: first, user is classified, and set up all types of user pair
The sub-model answered;Then, the real-time estimate characteristic vector obtained by S11 is brought in the sub-model of relative users,
The purchase probability of calculating behavior commodity.
The mode classifying user has a lot, and as a kind of preferred version, the present embodiment selects according to terminal
Type, user type, three kinds of dimensions of visitor's type, classify to user, and every class user set up one point
Model.User as one example, is divided into 8 classes, is specifically classified as follows by example:
PC new visitor model: for without historical behavior feature and the PC visitor that accesses website today and set up
Model.
PC old visitor model: for the model having the PC visitor of historical behavior feature and set up.
PC recruit's model, for without historical behavior feature and the PC member that is registered as member today and build
Vertical model.
PC old member model, builds for PC member that is that have historical behavior feature and that be registered as member before
Vertical model.
WAP visitor's model, the model set up for the WAP visitor of non-registered member.
WAP member's model, the model set up for the WAP member of registered members.
APP visitor's model, the model set up for the APP visitor of non-registered member.
APP member's model, the model set up for the APP member of registered members.
The most in addition, other user classification method is the most feasible, as long as setting up the model of corresponding classification respectively,
And be trained.
Above-mentioned multi-model is used to calculate the purchase probability of the behavior commodity obtained also for multiple.The present embodiment is to multiple
Model merges, and obtains the purchase probability of unified behavior commodity, and the purchase of the behavior commodity after i.e. merging is general
Rate.The purchase probability of the behavior commodity obtaining S12 is modified, and detailed process is: use each sub-model
Training result first and new corrected parameter as the skew fusion factor of Model Fusion, the behavior that S12 is obtained
The purchase probability of commodity carries out probability correction, obtains the probability revised, the commodity purchasing probability after i.e. being merged.
The method revised is that multi-model probability fusion model calculates.
The new corrected parameter of each sub-model can freely set, and such as, new corrected parameter is conversion ratio, clicking rate
Or category of model.
Multi-model merges the machine-learning process that can be considered again.It is provided with 3 disaggregated models: sub-model M1,
Sub-model M2 and sub-model M3, utilizes test sample data to carry out model measurement, the result of test data output
It is respectively F1, F2 and F3.Predict the outcome as f1 by training sample data by the training data that model exports,
F2 and f3.By f1, the conversion ratio of f2, f3 and each sub-model is merged in training sample data, to each point of mould
Type is trained again, obtains the model revised.Again by the model training test sample data revised, obtain result a,
B and c.
In the present embodiment, behavior commodity refer to that user has the commodity of operation behavior on e-commerce website.Operation
Behavior is such as: browses, click on, the behaviors such as folder of puting into collection.Similar dependent merchandise refers to similar to behavior commodity
Or relevant commodity.Similar dependent merchandise by behavior commodity according to the correlating method meter such as correlation rule, collaborative filtering
The items list drawn, then chooses according to support and degree of belief and meets the threshold value that user gives respectively, press
Commodity are obtained according to threshold filtering.The association scene wherein chosen includes but not limited to browse finally to be bought, has seen also
See, seen final purchase, accessory collocation scene etc..The present embodiment calculates similar dependent merchandise with formula (1)
Purchase probability:
Score_i=Master_SPU_Pos*SKU_Score_i/max (SKU_Score_i) formula (1)
Wherein, Score_i represents the purchase probability of similar dependent merchandise, and Master_SPU_Pos represents behavior
The purchase probability of commodity, max (SKU_Score_i) represents degree of association peak in similar dependent merchandise list,
SKU_Score_i represents the degree of association of similar dependent merchandise and behavior commodity.
The present embodiment uses multi-data source, the data of multiple terminals type, is trained calculating in real time of multi-model and purchases
Buy conversion prediction probability.Model uses linear burst and merges the training sample data got through simultaneously, and to dividing
The multi-model of sheet uses the side-play amount factor to carry out Model Fusion normalization, finally gives a precision high, the most eventually
The model of the real-time purchase probability prediction that end merges.The present embodiment is according to different terminals, the real-time visit of different crowd
Ask behavior, it is provided that real-time, a personalized Method of Commodity Recommendation based on user's purchase probability.The present embodiment
With real-time data collection, and can calculate, to user's real-time recommendation commodity.
The present embodiment, according to different data terminals and fused data, is respectively trained model according to crowd, then presses
According to the offset correction of section of hiving off, merge each model, obtain unified model.Method used by fusion has a lot,
Most people are directly by all kinds of methods in machine learning, or with statistical regression etc..The present embodiment selects each point
Model preliminary forecasting result adds the method for new corrected parameter re-training, obtains Fusion Model.Other machines learns
Method precision of prediction improve limited, model is more complicated.In the present embodiment, the data that step S11 gathers are come
Come from multiple terminal.In step S13, crowd is classified, and every class crowd is established a point mould respectively
Type;Then each sub-model is trained, using inclined as Model Fusion of training result first and new corrected parameter
Move fusion factor, carry out each sub-model merging normalization, obtain revised unified model, after i.e. merging
Unified model;The unified model after this fusion is finally utilized to carry out probability measuring and calculating.The other machines that compares learns
Method, the present embodiment in terms of Data Source and model set up two, especially merge after unified model, improve
Precision of prediction.Compared with setting up a total model in prior art, the present embodiment uses and divides each classification crowd
Not setting up sub-model, model is simple.Different classes of crowd, the factor affecting purchase probability is different.Set up one
Individual total model needs to consider the influence factor of various people.And each classification crowd is set up sub-model, it is thus only necessary to
Consider the influence factor of such crowd, it is not necessary to consider the influence factor of other classes crowd.Therefore, the present embodiment
Model simple.
The method and the method utilizing single model to be predicted that use the present embodiment compare.According to single
Model prediction, grader correctly judges that the value of positive sample is higher than the probability of negative sample (English full name: Area
Under the ROC Curve, is called for short in literary composition: be AUC) 0.70.Use the present embodiment method, AUC=0.85.
The precision of prediction of the method for the present embodiment is higher than the precision of prediction of single model.
Shown in Fig. 2, the off-line of the model for relating in the present embodiment trains flow process automatically, specific as follows:
S21 chooses real-time behavior characteristics data and history feature data according to different terminals.
These real-time behavior characteristics data and history feature data include data in the stations such as PC end, WAP end, APP end,
Also terminal data under data and line is included on third party's line.Such as, real-time behavior characteristics data include: details are visited
Ask that feature, search characteristics, list page access feature, sales promotion accesses feature, page click feature, collection
Number, shopping cart number of times etc..History feature data include: member's attribute or label, member's loyalty, member purchase
Buy power, member's preference, member's history access index etc..For example, member's history access index includes in the past
Within N days, add shopping cart behavior number of times, past N days added collection behavior number of times, past N days PV number, mistake
Go in session in N days at most PV numbers, past N days website visiting duration, the session of past N days
Big duration, N days past browse commodity details number of pages, the session of past N days in most access details numbers of pages
Amount, past N days number of sessions, past N days quantity on order.
Real-time behavior characteristics data that S22 gathers according to S21 and history feature data, respectively extraction unit divided data
As training sample data and off-line test sample data, obtain characteristic vector.
This feature vector includes real-time behavior characteristics and history feature index.Extraction unit divided data can be 1-30
They data produced, it is also possible to be the interior data produced of other natural law.
The data that S23 obtains according to S22, training pattern.
Described in above-described embodiment, according to terminal type, user type and visitor's type dimension, user is entered
Row classification, and every class user is set up a sub-model.Use each point of logistic regression model training
Model.
For example, each sub-model of training includes:
PC new visitor model, the sample data of training includes the PC visitor (non-member) without historical behavior feature;
PC old visitor model, the sample data of training includes PC visitor's (non-meeting of history historical behavior feature
Member);
PC recruit's model, the sample data of training includes the PC member without historical behavior feature;
PC old member model, the sample data of training includes the PC member of history historical behavior feature;
WAP visitor's model, the sample data of training includes WAP visitor (non-member);
WAP member's model, the sample data of training includes WAP member;
APP visitor's model, the sample data of training includes APP visitor (non-member);
APP member's model, the sample data of training includes APP member.
Logistic regression function (model of user's purchase probability) is represented, as shown in formula (2) with p (y=1 | x):
Wherein, and p (y=1 | x) represent transition probability, f (x) represents the linear function of characteristic vector.
Owing to Feature Selection is more, and model is more complicated, can produce the problem such as over-fitting, feature synteny,
The method selecting LASSO to return carries out variable selection and the regularization of model, and the RSS form that lasso returns is such as
Under:
Wherein, y represents predictor variable, β0Represent constant, βjRepresenting variable parameter, λ represents that lasso punishes
Penalty factor, i represents that sample size, j represent model variable quantity, xijRepresent variable.
Model carries out burst process according to terminal, user type, visitor's type dimension, the such as formula of the model after burst
(4):
Wherein, p (y=1 | x) represents transition probability, i.e. purchase probability;π represents sub-model, use of respectively classifying
The sub-model that family is corresponding, m represents model quantity, and x represents that variable, ω represent that weight, i represent sample size.
S24, the training pattern as shown in formula (4) obtained according to S23, add training result and Xin Xiu first
Positive parameter merges, the unified model after being merged.
As shown in Figure 4, for another embodiment.This embodiment provides one commercial product recommending row in ecommerce
The generation system of table, including:
Acquisition module: for gathering the characteristic of user, and each terminal data is merged, after being merged
Real-time estimate characteristic vector;
Computing module: for calculating the purchase probability of behavior commodity;
Fusion Module: for the purchase probability of behavior commodity is modified, the behavior commodity after being merged
Purchase probability;
Generation module: for calculating the purchase probability of similar dependent merchandise, and sort according to buying rate size, raw
Become commercial product recommending list.
In this embodiment, the user characteristic data of acquisition module collection derives from PC end, WAP end, APP end station
Terminal data under data and line in interior data, and third party's line.So, for data acquisition angle, this
A little characteristics cover multiple terminal.With single terminal, such as WAP end compares, the number that the present embodiment gathers
Wide according to carrying out source range.Multiple terminals Data Source, provides basis number more accurately for follow-up purchase probability prediction
According to.As a kind of preferred version, characteristic includes history offline feature data and real-time characteristic data.History
Offline feature data can select: member's attribute or label, member's loyalty, membership buying power, member's preference,
Member's history accesses index etc..Member's history accesses index and adds shopping cart behavior time in the N of may is that over days
In number, past N days interpolation collection behavior number of times, past N days PV number, session of past N days at most
PV number etc..Real-time characteristic data acquisition with resolve access log, click logs, exposure daily record, event log,
The method of order daily record obtains.Such as, real-time characteristic data include level Four page commodity details district number of clicks (bag
Include evaluation), commodity level Four page collection number of clicks, list page recommend number of clicks, list page collection click on time
Number, list page commodity number of clicks etc..
In acquisition module, the method for real-time estimate characteristic vector is: set up the mapping relations of terminal, gets through each
The data cube computation passage of terminal, the characteristic vector after each characteristic is merged by mapping relations association,
By each terminal feature Vector Fusion, the real-time estimate characteristic vector after being merged.For non-member, Ke Yitong
Cross the characteristic index after the correspondence mappings relationship such as cookie, mobile phone string number is merged.For member,
Can be by association real time data and historical data, the characteristic indexs after being merged such as member's codings.By meeting
Member's relation, each terminal recognition code carry out data cube computation and association, obtain fused data.
After real-time estimate characteristic vector after being merged by acquisition module, need the purchase of calculating behavior commodity
Probability.Computing module in the present embodiment includes:
Classification submodule: for classifying user, forms catergories of user;
Modeling submodule: for all types of user is set up corresponding sub-model;
Calculating sub module: the real-time estimate characteristic vector obtained for acquisition module brings the sub-model of relative users into
In, calculate the purchase probability of behavior commodity.
This computing module is different from other computing modules existing.This computing module by classification submodule to user
Classify, form catergories of user;Then by modeling submodule, all types of user is set up corresponding sub-model;
Finally utilize the real-time estimate characteristic vector that acquisition module is obtained by calculating sub module, bring point mould of relative users into
In type, calculate the purchase probability of behavior commodity.The mode classifying user has a lot, preferred as one
Scheme, the present embodiment selects according to terminal type, user type, three kinds of dimensions of visitor's type, carries out user
Classification, and every class user is set up a sub-model.This computing module is owing to have employed multiple sub-model, and it is calculated
The purchase probability of the behavior commodity obtained is multiple.Multiple models are merged by the present embodiment, obtain unification
The purchase probability of behavior commodity, the purchase probability of the behavior commodity after i.e. merging.The behavior that computing module is obtained
The purchase probability of commodity is modified, and detailed process is: use the training result first of each sub-model and Xin Xiu
Positive parameter is as the skew fusion factor of Model Fusion, and the purchase probability of the behavior commodity obtaining computing module enters
Row probability correction, obtains the probability revised.The method revised is that multi-model probability fusion model calculates.
The new corrected parameter of each sub-model can freely set, and such as, new corrected parameter is conversion ratio, clicking rate
Or category of model.
Multi-model merges the machine-learning process that can be considered again.It is provided with 3 disaggregated models: sub-model M1,
Sub-model M2 and sub-model M3, utilizes test sample data to carry out model measurement, the result of test data output
It is respectively F1, F2 and F3.Predict the outcome as f1 by training sample data by the training data that model exports,
F2 and f3.By f1, the conversion ratio of f2, f3 and each sub-model is merged in training sample data, to each point of mould
Type is trained again, obtains the model revised.Again by the model training test sample data revised, obtain result a,
B and c.
In the present embodiment, behavior commodity are the commodity that user has the behaviors such as access, and similar dependent merchandise refers to and uses
The commodity that family behavior commodity are similar or relevant.As a kind of preferred version, calculate similar relevant according to formula (1)
The purchase probability of commodity.
Score_i=Master_SPU_Pos*SKU_Score_i/max (SKU_Score_i) formula (1)
Wherein, Score_i represents the purchase probability of similar dependent merchandise, and Master_SPU_Pos represents behavior
The purchase probability of commodity, max (SKU_Score_i) represents degree of association peak in similar dependent merchandise list,
SKU_Score_i represents the degree of association of similar dependent merchandise SKU_i and behavior commodity.
Those skilled in the art should know, it is achieved the method for above-described embodiment or system, can be by calculating
Machine programmed instruction realizes.This computer program instructions is loaded in programmable data processing device, such as, calculate
Machine, thus in programmable data processing device, perform corresponding instruction, for the method realizing above-described embodiment
Or the function that system realizes.
Those skilled in the art, according to above-described embodiment, can carry out the technological improvement of non-creativeness to the application,
Spirit without deviating from the present invention.These improve still should be regarded as the application scope of the claims it
In.
Claims (14)
1. the generation method of commercial product recommending list in ecommerce, it is characterised in that the method includes
Following steps:
S11 gathers the characteristic of user, and by each terminal data Feature Fusion, after being merged the most in advance
Survey characteristic vector;
S12 calculates the purchase probability of behavior commodity;
The purchase probability of the behavior commodity that S12 is obtained by S13 is modified, and obtains revised behavior commodity
Purchase probability;
S14, according to the purchase probability of revised behavior commodity, calculates the purchase probability of similar dependent merchandise, and
Sort according to buying rate size, generate commercial product recommending list.
The most in accordance with the method for claim 1, it is characterised in that in described step S11, described spy
Levy Data Source data and the 3rd in data, and third party's line in PC end, WAP end and/or APP end station
Terminal data under side's line;Described characteristic includes history offline feature data and real-time characteristic data.
The most in accordance with the method for claim 2, it is characterised in that described real-time characteristic data acquisition resolves
The method of access log, click logs, exposure daily record, event log and/or order daily record obtains.
The most in accordance with the method for claim 3, it is characterised in that described S11 farther includes: set up
The mapping relations of terminal, are got through or the data cube computation passage of each terminal of pathfinding, each characteristic are closed by mapping
System associates the characteristic vector after being merged, by each terminal feature Vector Fusion, the real-time estimate after being merged
Characteristic vector.
The most in accordance with the method for claim 1, it is characterised in that described S12 farther includes: first
User is classified, and sets up the sub-model that all types of user is corresponding;Then the real-time estimate obtained by S11 is special
Levy in the sub-model that vector brings relative users into, calculate the purchase probability of behavior commodity.
The most in accordance with the method for claim 5, it is characterised in that described user is carried out classification include:
According to terminal type, user type and three kinds of dimensions of visitor's type, user is classified.
The most in accordance with the method for claim 5, it is characterised in that in described S13, S12 is obtained
The purchase probability of behavior commodity is modified: use the training result first of each sub-model and new corrected parameter
As the skew fusion factor of Model Fusion, the purchase probability of the behavior commodity obtaining S12 is modified,
To revised commodity purchasing probability.
The most in accordance with the method for claim 1, it is characterised in that in described S14, similar dependent merchandise
Purchase probability according to formula (1) calculate:
Score_i=Master_SPU_Pos*SKU_Score_i/max (SKU_Score_i) formula (1)
Wherein, Score_i represents the purchase probability of similar dependent merchandise, and Master_SPU_Pos represents behavior
The purchase probability of commodity, max (SKU_Score_i) represents degree of association peak in similar dependent merchandise list,
SKU_Score_i represents the degree of association of similar dependent merchandise SKU_i and behavior commodity.
9. the generation system of commercial product recommending list in ecommerce, it is characterised in that this system includes:
Acquisition module: for gathering the characteristic of user, and each terminal data is merged, after being merged
Real-time estimate characteristic vector;
Computing module: for calculating the purchase probability of behavior commodity;
Fusion Module: for the purchase probability of behavior commodity is modified, the behavior commodity after being merged
Purchase probability;
Generation module: for calculating the purchase probability of similar dependent merchandise, and sort according to buying rate size, raw
Become commercial product recommending list.
10. according to the system described in claim 9, it is characterised in that in described acquisition module, characteristic number
According to deriving from PC end, WAP end, APP end station in data, and third party's line terminal data under data and line;
Described characteristic includes history offline feature data and real-time characteristic data.
11. according to the system described in claim 9, it is characterised in that described acquisition module specifically for:
Set up the mapping relations of terminal, get through the data cube computation passage of each terminal, each characteristic is passed through mapping relations
Associating the characteristic vector after being merged, by each terminal feature Vector Fusion, the real-time estimate after being merged is special
Levy vector.
12. according to the system described in claim 9, it is characterised in that described computing module includes:
Classification submodule: for classifying user, forms catergories of user;
Modeling submodule: for all types of user is set up corresponding sub-model;
Calculating sub module: the real-time estimate characteristic vector obtained for acquisition module brings the sub-model of relative users into
In, calculate the purchase probability of behavior commodity.
13. according to the system described in claim 9, it is characterised in that in described Fusion Module, to behavior
The purchase probability of commodity is modified, use using the training result first of each sub-model and new corrected parameter as
The skew fusion factor of Model Fusion, is modified the purchase probability of behavior commodity, the commodity after being merged
Purchase probability.
14. according to the system described in claim 9, it is characterised in that in described generation module, phase patibhaga-nimitta
The purchase probability of underlying commodity is calculated according to formula (1):
Score_i=Master_SPU_Pos*SKU_Score_i/max (SKU_Score_i) formula (1)
Wherein, Score_i represents the purchase probability of similar dependent merchandise, and Master_SPU_Pos represents behavior
The purchase probability of commodity, max (SKU_Score_i) represents degree of association peak in similar dependent merchandise list,
SKU_Score_i represents the degree of association of similar dependent merchandise SKU_i and behavior commodity.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510268039.6A CN106296242A (en) | 2015-05-22 | 2015-05-22 | A kind of generation method of commercial product recommending list in ecommerce and the system of generation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510268039.6A CN106296242A (en) | 2015-05-22 | 2015-05-22 | A kind of generation method of commercial product recommending list in ecommerce and the system of generation |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106296242A true CN106296242A (en) | 2017-01-04 |
Family
ID=57632985
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510268039.6A Pending CN106296242A (en) | 2015-05-22 | 2015-05-22 | A kind of generation method of commercial product recommending list in ecommerce and the system of generation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106296242A (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107992530A (en) * | 2017-11-14 | 2018-05-04 | 北京三快在线科技有限公司 | Information recommendation method and electronic equipment |
CN108549674A (en) * | 2018-03-30 | 2018-09-18 | 咪咕文化科技有限公司 | A kind of recommendation method, apparatus and storage medium |
CN108805594A (en) * | 2017-04-27 | 2018-11-13 | 北京京东尚科信息技术有限公司 | Information-pushing method and device |
CN108830689A (en) * | 2018-06-13 | 2018-11-16 | 深圳乐信软件技术有限公司 | Item recommendation method, device, server and storage medium |
CN109034973A (en) * | 2018-07-25 | 2018-12-18 | 北京京东尚科信息技术有限公司 | Method of Commodity Recommendation, device, system and computer readable storage medium |
WO2019061990A1 (en) * | 2017-09-30 | 2019-04-04 | 平安科技(深圳)有限公司 | User intention prediction method, electronic device, and computer readable storage medium |
CN109903063A (en) * | 2017-12-07 | 2019-06-18 | 北京京东尚科信息技术有限公司 | Data processing method and its system, computer system and readable storage medium storing program for executing |
CN110163722A (en) * | 2019-05-13 | 2019-08-23 | 南京邮电大学 | The big data analysis system and analysis method precisely sold for agricultural product |
CN110533515A (en) * | 2019-09-04 | 2019-12-03 | 深圳创新奇智科技有限公司 | A kind of the electric business personalized recommendation method and device of high low latency of handling up |
CN110827104A (en) * | 2018-08-07 | 2020-02-21 | 北京京东尚科信息技术有限公司 | Method and device for recommending commodities to user |
CN112132660A (en) * | 2020-09-25 | 2020-12-25 | 尚娱软件(深圳)有限公司 | Commodity recommendation method, system, device and storage medium |
CN113868542A (en) * | 2021-11-25 | 2021-12-31 | 平安科技(深圳)有限公司 | Attention model-based push data acquisition method, device, equipment and medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102902691A (en) * | 2011-07-28 | 2013-01-30 | 上海拉手信息技术有限公司 | Recommending method and recommending system |
CN103679494A (en) * | 2012-09-17 | 2014-03-26 | 阿里巴巴集团控股有限公司 | Commodity information recommendation method and device |
CN103744917A (en) * | 2013-12-27 | 2014-04-23 | 东软集团股份有限公司 | Mixed recommendation method and system |
CN104050272A (en) * | 2014-06-25 | 2014-09-17 | 上海艾瑞市场咨询有限公司 | Cross-screen user behavior data statistics and analysis method and device |
-
2015
- 2015-05-22 CN CN201510268039.6A patent/CN106296242A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102902691A (en) * | 2011-07-28 | 2013-01-30 | 上海拉手信息技术有限公司 | Recommending method and recommending system |
CN103679494A (en) * | 2012-09-17 | 2014-03-26 | 阿里巴巴集团控股有限公司 | Commodity information recommendation method and device |
CN103744917A (en) * | 2013-12-27 | 2014-04-23 | 东软集团股份有限公司 | Mixed recommendation method and system |
CN104050272A (en) * | 2014-06-25 | 2014-09-17 | 上海艾瑞市场咨询有限公司 | Cross-screen user behavior data statistics and analysis method and device |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108805594A (en) * | 2017-04-27 | 2018-11-13 | 北京京东尚科信息技术有限公司 | Information-pushing method and device |
WO2019061990A1 (en) * | 2017-09-30 | 2019-04-04 | 平安科技(深圳)有限公司 | User intention prediction method, electronic device, and computer readable storage medium |
CN107992530A (en) * | 2017-11-14 | 2018-05-04 | 北京三快在线科技有限公司 | Information recommendation method and electronic equipment |
CN109903063A (en) * | 2017-12-07 | 2019-06-18 | 北京京东尚科信息技术有限公司 | Data processing method and its system, computer system and readable storage medium storing program for executing |
CN108549674B (en) * | 2018-03-30 | 2021-11-09 | 咪咕文化科技有限公司 | Recommendation method, recommendation device and storage medium |
CN108549674A (en) * | 2018-03-30 | 2018-09-18 | 咪咕文化科技有限公司 | A kind of recommendation method, apparatus and storage medium |
CN108830689A (en) * | 2018-06-13 | 2018-11-16 | 深圳乐信软件技术有限公司 | Item recommendation method, device, server and storage medium |
CN109034973A (en) * | 2018-07-25 | 2018-12-18 | 北京京东尚科信息技术有限公司 | Method of Commodity Recommendation, device, system and computer readable storage medium |
WO2020020137A1 (en) * | 2018-07-25 | 2020-01-30 | 北京京东尚科信息技术有限公司 | Commodity recommendation method, apparatus and system, and computer readable storage medium |
CN110827104A (en) * | 2018-08-07 | 2020-02-21 | 北京京东尚科信息技术有限公司 | Method and device for recommending commodities to user |
CN110163722A (en) * | 2019-05-13 | 2019-08-23 | 南京邮电大学 | The big data analysis system and analysis method precisely sold for agricultural product |
CN110163722B (en) * | 2019-05-13 | 2022-08-23 | 南京邮电大学 | Big data analysis system and analysis method for accurate sale of agricultural products |
CN110533515A (en) * | 2019-09-04 | 2019-12-03 | 深圳创新奇智科技有限公司 | A kind of the electric business personalized recommendation method and device of high low latency of handling up |
CN112132660A (en) * | 2020-09-25 | 2020-12-25 | 尚娱软件(深圳)有限公司 | Commodity recommendation method, system, device and storage medium |
CN112132660B (en) * | 2020-09-25 | 2023-12-26 | 尚娱软件(深圳)有限公司 | Commodity recommendation method, system, equipment and storage medium |
CN113868542A (en) * | 2021-11-25 | 2021-12-31 | 平安科技(深圳)有限公司 | Attention model-based push data acquisition method, device, equipment and medium |
CN113868542B (en) * | 2021-11-25 | 2022-03-11 | 平安科技(深圳)有限公司 | Attention model-based push data acquisition method, device, equipment and medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106296242A (en) | A kind of generation method of commercial product recommending list in ecommerce and the system of generation | |
Fan et al. | Product sales forecasting using online reviews and historical sales data: A method combining the Bass model and sentiment analysis | |
CN106844787B (en) | Recommendation method for searching target users and matching target products for automobile industry | |
CN111222332B (en) | Commodity recommendation method combining attention network and user emotion | |
CN106296257A (en) | A kind of fixation of advertisement position put-on method based on user behavior analysis and system | |
CN104391849B (en) | Incorporate the collaborative filtering recommending method of time contextual information | |
CN105868847A (en) | Shopping behavior prediction method and device | |
CN106251174A (en) | Information recommendation method and device | |
Bauer et al. | Recommender systems based on quantitative implicit customer feedback | |
CN111784455A (en) | Article recommendation method and recommendation equipment | |
CN103824192A (en) | Hybrid recommendation system | |
CN105469263A (en) | Commodity recommendation method and device | |
CN106485562A (en) | A kind of commodity information recommendation method based on user's history behavior and system | |
CN104239338A (en) | Information recommendation method and information recommendation device | |
CN105528374A (en) | A commodity recommendation method in electronic commerce and a system using the same | |
CN101206751A (en) | Customer recommendation system based on data digging and method thereof | |
KR101963817B1 (en) | Apparatus and method for generating prediction information based on a keyword search volume | |
CN103514255A (en) | Method for collaborative filtering recommendation based on item level types | |
CN108509573A (en) | Book recommendation method based on matrix decomposition collaborative filtering and system | |
CN108205775A (en) | The recommendation method, apparatus and client of a kind of business object | |
CN107292713A (en) | A kind of rule-based individual character merged with level recommends method | |
CN112149352B (en) | Prediction method for marketing activity clicking by combining GBDT automatic characteristic engineering | |
CN111695024A (en) | Object evaluation value prediction method and system, and recommendation method and system | |
CN114861050A (en) | Feature fusion recommendation method and system based on neural network | |
CN111400613A (en) | Article recommendation method, device, medium and computer equipment |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170104 |