CN103744917A - Mixed recommendation method and system - Google Patents
Mixed recommendation method and system Download PDFInfo
- Publication number
- CN103744917A CN103744917A CN201310737345.0A CN201310737345A CN103744917A CN 103744917 A CN103744917 A CN 103744917A CN 201310737345 A CN201310737345 A CN 201310737345A CN 103744917 A CN103744917 A CN 103744917A
- Authority
- CN
- China
- Prior art keywords
- user
- degree
- interest
- algorithm
- various dimensions
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/958—Organisation or management of web site content, e.g. publishing, maintaining pages or automatic linking
Abstract
The invention provides a mixed recommendation method and system. The method comprises dividing user data in a database into single-dimensional user characteristic values and multi-dimensional user characteristic values; calculating the single-dimensional user characteristic values through a Bayesian average algorithm to obtain user recommendation algorithm interestingness; calculating the multi-dimensional user characteristic values through a naive Bayes model to obtain user recommendation algorithm interestingness; determining recommendation lists according to the recommendation algorithm interestingness of the single-dimensional users and the recommendation algorithm interestingness of the multi-dimensional user characteristic values, and recommending the recommendation lists to database users. According to the method and the system, the problems of low recommendation algorithm interestingness prediction accuracy, calculating and evaluation of recommendation algorithm interestingness of different users and time bottlenecks for training user interestingness models of users can be solved.
Description
Technical field
The present invention relates to recommended technology field, more specifically, relate to a kind of mixing recommend method and system.
Background technology
Recommended technology is developed so far the history of existing more than ten years, and numerous proposed algorithms is suggested and in the industry cycle application therebetween.Through a large amount of practices, it is found that and seem to go for all scenes without any a kind of method, because every kind of proposed algorithm has the limitation that self is difficult to go beyond.
In the face of the limitation of various proposed algorithms, released in the industry mixing recommended technology, mixing at present recommended technology has a lot, as: weighted blend recommended technology, the mediation that intersects mix recommended technology etc.Fig. 1 shows the flow process of traditional mixing recommended technology, and as shown in Figure 1, the flow process of traditional mixing recommended technology comprises:
S110: start;
S120: user's browsing objective business web site;
S130: website sends request to commending system;
S140: commending system external interface;
S150: proposed algorithm; Wherein, in proposed algorithm, can call A and B, A is information of goods information data storehouse, party B-subscriber's behavior and User Information Database;
S160: whether use to mix and recommend strategy, if so, carry out S180; If not, carry out S170;
S170: mix and recommend policy components;
S180: recommendation list;
S190: finish.
Above-mentioned is the whole flow process of traditional mixing recommended technology, and at present, the mixing of main flow recommends strategy to comprise: weighting type mixing recommended technology, classification-type mixing recommended technology, chiasma type mixing recommended technology, recommended models hybrid technology etc.
But above-mentioned these mix the problem that recommended technology all exists some general character.
Because above-mentioned commending system is identical to all users' proposed algorithm mixed strategy, and this strategy may and not be suitable for all users.Such as: certain system provides multiple proposed algorithm system, adopts stepped mixing strategy, preferentially represents the result of the proposed algorithm generation that accuracy rate is higher, the follow-up result that adds other algorithms that recall rate is higher.For the recommendation list of final recommendation results (Top-N), the user that interest is single can please oneself to recommendation results; And interest widely user always preferentially see his an interested aspect, As time goes on, can inevitably cause user's low, the novel degree of algorithm interest-degree predictablity rate inadequate, and the time that commending system need to complete recommendation increases, thereby reduce the satisfaction of user to recommendation results.
Therefore, need now a kind of new mixing recommended technology to solve the problem of above-mentioned appearance.
Summary of the invention
In view of the above problems, the object of this invention is to provide a kind of mixing recommend method and system, to solve the problem of user's the proposed algorithm interest-degree that proposed algorithm interest-degree predictablity rate is low, different user is assessed in calculating and the time bottleneck of training user interest model.
On the one hand, the invention provides a kind of mixing recommend method, comprise the user data in database is divided into one-dimensional degree user characteristics value and various dimensions user characteristics value;
Adopt Bayes's average algorithm to calculate one-dimensional degree user characteristics value, obtain one-dimensional degree user's proposed algorithm interest-degree; And,
Adopt model-naive Bayesian to calculate various dimensions user characteristics value, set up various dimensions user interest model, and various dimensions user interest model is trained, obtain various dimensions user's proposed algorithm interest-degree;
According to one-dimensional degree user's proposed algorithm interest-degree and various dimensions user's proposed algorithm interest-degree, determine recommendation list, and recommendation list is recommended to the user of database.
In addition, preferred scheme is, according to one-dimensional degree user's proposed algorithm interest-degree and various dimensions user's proposed algorithm interest-degree, determining in the process of recommendation list, by mixing, recommend policy components to merge one-dimensional degree user's proposed algorithm interest-degree and various dimensions user's proposed algorithm interest-degree, the interest-degree of the user of acquisition database to each proposed algorithm, according to the interest-degree of each proposed algorithm, obtain the ratio of each proposed algorithm in all proposed algorithms, determine recommendation list.
In addition, preferred scheme is, in the process that adopts Bayes's average algorithm to calculate one-dimensional degree user characteristics value, to adopt parallel calculating method to calculate one-dimensional degree user's proposed algorithm interest-degree;
In the process that adopts model-naive Bayesian to calculate various dimensions user characteristics value, adopt parallel calculating method to calculate various dimensions user's proposed algorithm interest-degree.
In addition, preferably scheme is, in the process that adopts Bayes's average algorithm to calculate one-dimensional degree user characteristics value,
Bayes's average algorithm formula is:
Wherein, C is a constant of setting voluntarily;
M is the mean value of project in database;
N is the data bulk of described project;
X is the value of each data;
In above-mentioned formula, m is prior probability, and the value of each x is an adjustment factor, along with the increase of n, and the value of project
its correct value will constantly be tended to.
In addition, preferably scheme is, in the process that adopts model-naive Bayesian to calculate various dimensions user characteristics value,
Model-naive Bayesian formula is:
Wherein, C is object event; X is eigenwert;
P (C|X) is the probability that under X condition, C occurs;
P (X|C) is the probability that under C condition, X occurs;
The probability that P (C) is C;
The probability that P (X) is X.
On the other hand, the present invention also provides a kind of mixing commending system, comprising:
Database community subdivision, for being divided into the user data of database one-dimensional degree user characteristics value and various dimensions user characteristics value;
Bayes's average algorithm computing unit, for adopting Bayes's average algorithm to calculate one-dimensional degree user characteristics value, obtains one-dimensional degree user's proposed algorithm interest-degree;
Model-naive Bayesian computing unit, for adopting model-naive Bayesian to calculate various dimensions user characteristics value, sets up various dimensions user interest model, and various dimensions user interest model is trained, and obtains various dimensions user's proposed algorithm interest-degree;
Algorithm recommendation unit, for according to one-dimensional degree user's proposed algorithm interest-degree and various dimensions user's proposed algorithm interest-degree, determines recommendation list, and recommendation list is recommended to the user of database.
From technical scheme above, mixing recommend method provided by the invention and system, can calculate the proposed algorithm interest-degree of assessment different user, by adopting Bayes's average algorithm to obtain one-dimensional degree user's proposed algorithm interest-degree, can improve the accuracy rate of one-dimensional degree user's algorithm interest prediction; By adopting model-naive Bayesian to obtain various dimensions user's proposed algorithm interest-degree, adopt the algorithm interest that naive Bayesian study can digging user group recessiveness, improve user's the pre-measuring precision of algorithm interest; By parallel computation, can shorten the training time of user interest model and the time of interest forecast assessment.
In order to realize above-mentioned and relevant object, one or more aspects of the present invention comprise below by the feature that describes in detail and particularly point out in the claims.Explanation below and accompanying drawing describe some illustrative aspects of the present invention in detail.Yet, the indication of these aspects be only some modes that can use in the variety of way of principle of the present invention.In addition, the present invention is intended to comprise all these aspects and their equivalent.
Accompanying drawing explanation
By reference to the content below in conjunction with the description of the drawings and claims, and along with understanding more comprehensively of the present invention, other object of the present invention and result will be understood and easy to understand more.In the accompanying drawings:
Fig. 1 is the schematic flow sheet of traditional mixing recommended technology;
Fig. 2 is according to the schematic flow sheet of the mixing recommend method of the embodiment of the present invention;
Fig. 3 is according to the logical organization block diagram of the mixing commending system of the embodiment of the present invention.
In institute's drawings attached, identical label is indicated similar or corresponding feature or function.
Embodiment
In the following description, for purposes of illustration, for the complete understanding to one or more embodiment is provided, many details have been set forth.Yet, clearly, also can in the situation that there is no these details, realize these embodiment.Below with reference to accompanying drawing, specific embodiments of the invention are described in detail.
In the mixing recommended technology of aforementioned conventional, commending system external interface is received after request, calls proposed algorithm, and result of calculation (Top-N) is sent to and mixes recommendation policy components.If proposed algorithm is single, directly result is illustrated in recommendation list.If adopt multiple proposed algorithm, to mix and recommend policy components each algorithm generation result to be merged according to fixed policy, recommendation list is shown the recommendation results after final fusion.The problems such as the algorithm recommendation rate that exists for the mixing recommended technology of aforementioned conventional is inaccurate, commending system is flexible, the present invention proposes a kind of personalized recommended technology that mixes, this technology can be calculated the level of interest of user to algorithm according to user data, thereby adjust automatically the ratio that each algorithm generates the shared recommendation list of result, for user provides personalized mixing, recommend.
In the present invention, during user's browsing objective business web site, website is to commending system request recommendation results, and commending system external interface calls proposed algorithm after receiving request.Personalized mixing of the present invention recommends policy components according to user's self interest, to be inclined to the recommendation list of determining that user is final, and training data source comes from user behavior and User Information Database.
Below with reference to accompanying drawing, specific embodiments of the invention are described in detail.Fig. 2 shows according to the flow process of the mixing recommend method of the embodiment of the present invention; As shown in Figure 2, the invention provides a kind of mixing recommend method, concrete recommend method comprises:
S210: the user data in database is divided into one-dimensional degree user characteristics value and various dimensions user characteristics value.
Particularly, in the present invention, according to commending system service environment, choose the eigenwert of calculated recommendation algorithm interest-degree, this eigenwert can be one-dimensional degree user data, can be also various dimensions user data.Database comprises user behavior data and user profile data.
In the present invention, user's data are divided into two large classes: a class is one-dimensional degree user characteristics value, and if user is to the click of content recommendation, scoring, collection etc., the feature of these type of data is to determine intuitively user's algorithm interest preference; Another kind of is various dimensions user characteristics value, as do not determined by intuitive judgment the data of user's algorithm hobby in user's age, sex, address etc.
That is to say, one-dimensional degree user characteristics value can the user profile of intuitive judgment user to proposed algorithm interest-degree, is generally user's behavioral data or the information of user feedback; That various dimensions user characteristics value is multidimensional, stealthy, conventionally can not the user profile of intuitive judgment user to proposed algorithm interest-degree.
S220: adopt Bayes's average algorithm to calculate one-dimensional degree user characteristics value, obtain one-dimensional degree user's proposed algorithm interest-degree.
Particularly, the present invention adopts the thought of Bayes's average algorithm, and one-dimensional degree user characteristics value is optimized and is revised.Its core concept is by cold start-up one-dimensional degree user's the interest popular interest that furthers, and constantly revises its point of interest by improving of one-dimensional degree user data, makes result of calculation more and more near the real interest of user.
It should be noted that, user's cold start-up refers to when new user arrives there is no this user's behavioral data or information data, thus cannot predict its interest according to this user's historical behavior or user profile, thus cannot do personalized recommendation to whereby this user.
Suppose to provide three kinds of proposed algorithm A, B, C in recommendation process.Represent is to mix the recommendation list that three kinds of algorithms generate to final user.Do you how obtain user most interested to the result of which kind of algorithm recommendation? there is several different methods to obtain, wherein, a kind of simple method: calculate user's clicking rate, concrete grammar is as follows:
In the situation that every kind of algorithm exposure recommendation results number is identical, the maximum algorithm of number of times of being clicked by targeted customer, is liked by this user more.If user is C to the clicks of algorithm A
a, in like manner obtain the clicks C to B and C
b, C
c; Like this, pass through C
a, C
band C
ccapable of dynamic adjustment algorithm exposes the number that generates result; Algorithm A recommends number to account for the number percent P of the total number of recommendation list
afor:
Yet these class data of clicking rate are affected by environment larger, can not completely objectively react user preferences.Than the better solution of clicking rate, be to recommending the scoring judgement user preferences of article by user.Suppose that user U (gives a mark interval at 1-5) as follows to the result scoring of recommending:
?
in like manner can calculate W
band W
c.The benefit of using scoring is that scoring is determined by user self, the impact of the environment of avoiding on user's operation.
Mean value W can not objectively react the hobby of user to algorithm.Reason is, is similar to n Bernoulli trials, and along with the increase of experiment number, final probability can tend to its expectation P.The increase of the scoring of the project algorithm A being produced along with user, W convergence is truly liked with it, and the number of times of marking crosses that I haven't seen you for ages causes the value of W unreliable.
The present invention uses for reference the problem that Bayes's average algorithm thought solves mean value W underconfidence under cold start.The average more rank problem that is applicable to of Bayes.It has used for reference the thought of Bayesian inference (Bay esian Inference): the in the situation that of uncertain result, first estimate a value, then, by the continuous correction of fresh information, make it more and more approach correct value.
Bayes's average algorithm general type:
Wherein, C is a constant of setting voluntarily;
M is the mean value of project in database;
N is the data bulk of described project;
X is the value of each data.
Above-mentioned formula 3. in, m is prior probability, the value of each x is one and adjusts the factor, along with the increase of n, the value of this project
its correct value will constantly be tended to.
Use for reference formula thought 3., the W to formula in 2.
arevise optimization, optimum results is as follows:
Wherein,
user's average score number of times for algorithm A;
T
afor the average score number of times of targeted customer to algorithm A;
W
afor the average score of targeted customer to algorithm A.
By formula, 4. can calculate targeted customer's prior probability interest tendency W
a', W
b' and W
c', and along with the increase of user behavior data is constantly revised, make the true interest hobby of its continuous convergence.
In the present invention, adopt Bayes's average algorithm to calculate user's proposed algorithm interest-degree.The method be applicable to can intuitive judgment user algorithm the data of one-dimensional degree user behavior of hobby, if user is to the scoring of algorithm recommendation results etc.This algorithm can solve the not high problem of degree of confidence that predicts the outcome that cold start-up user brings when behavioral data is less, to improve the accuracy rate of user's algorithm interest prediction.
The calculating of said process is to be based upon on the hypothesis basis of " user's scoring is higher, and user is higher to the level of interest of algorithm ", and obviously, this hypothesis is set up.Yet some information can not be used the method to calculate.As: in a garment marketing website, its user's hobby has very large associated with its age and sex.And can not determine intuitively that by us which kind proposed algorithm the user of certain age bracket sex is applicable to, they are higher to the interest level of that proposed algorithm.And often, these information can play critical effect.Therefore, just adopt in the following method and recommend:
S230: adopt model-naive Bayesian to calculate various dimensions user characteristics value, set up various dimensions user interest model, and various dimensions user interest model is trained, obtain various dimensions user's proposed algorithm interest-degree.
Particularly, the present invention uses the model-naive Bayesian in machine learning, by various dimensions user's historical data, trains user interest model, and calculates user's proposed algorithm interest-degree.
Bayesian probability model general type:
Above-mentioned formula 5. in, C is object event; X is eigenwert;
P (C|X) is the probability that under X condition, C occurs;
P (X|C) is the probability that under C condition, X occurs;
The probability that P (C) is C; The probability that P (X) is X.
In a specific embodiment of the present invention, the age of user of take is example as characteristic variable.The age of hypothetical target user U is a, and two kinds of proposed algorithm A, B are provided in recommendation process, and user U is P (A|a) to the interested probability of algorithm A, and in like manner obtaining the probability interested of algorithm B is P (B|a).By formula, 5. obtaining user U is P (A|a) to the interested probability of algorithm A:
It should be noted that, 6. above-mentioned formula is only concerned about the molecular moiety in formula, because denominator does not rely on the value of C and feature a, is given, so denominator can be thought a constant.
Prior probability P (A) is not a distribution function, is only a value, and it has expressed the ratio of the shared quantity of sample of each class in sample space.According to law of great number, in training set, sample size is abundant and while coming from the choosing at random of sample space, can estimate with the shared ratio of Different categories of samples in training set the value of P (A).Therefore, can make P (A) equal the mean value of recommendation results number of the exposure of algorithm A here, formula is as follows:
P (a|A) is the probability density function with certain (as normal distribution) formal distribution, need to from training set, the distribution situation of sample characteristics estimate; Method of estimation can be divided into parameter estimation and non-parametric estmation.Common method is that the maximum likelihood of probability is estimated.
Here targeted customer's age Normal Distribution; Therefore the formula that, P (a|A) represents is as follows:
Wherein, μ is the expectation to the interested user's of algorithm A age;
σ
2for the variance to the interested user's of algorithm A age.
The expectation at age and variance can obtain in the historical data training by website, by formula 7. with formula 8., can compute age the user U that is a to the interest level of algorithm A, be P (a|A) * P (A).In like manner, obtain the interest level of U to algorithm B.
Above-mentioned is the unique computation process of characteristic variable, situation more generally, and user profile is multidimensional characteristic variable.According to model-naive Bayesian, obtain:
P (the F of above-mentioned formula in 9.
i| C| computing method are with formula 8..
The above-mentioned formula general type that 9. first half section is naive Bayesian, F1 ..., Fn is various dimensions user characteristics value; Second half section is not consider that the formula in denominator situation represents.
As can be seen from the above, this model-naive Bayesian is in the situation that user profile is more, and comprehensively each dimensional information is made prediction to user's proposed algorithm tendency.In the present invention, adopt the model-naive Bayesian training study hidden customer information of machine learning, this type of information is various dimensions often.Proposed algorithm interest-degree that can digging user group recessiveness by this model, improves the pre-measuring precision of interest.The result of calculating by this model can merge with Bayes's average algorithm result or supplement.Model-naive Bayesian and Bayes's average algorithm have solid Fundamentals of Mathematics and theoretical model; And Bayes's average algorithm and model-naive Bayesian are widely used in fields such as machine learning, probability statistics, data minings.
In the present invention, the data in database are divided into one-dimensional degree user characteristics value and various dimensions user characteristics value, and choosing by concrete business scenario of eigenwert determines; Then according to the eigenwert of choosing, determine again and adopt which kind of method to learn eigenwert.
It should be noted that: the applicable eigenwert quantity of NB Algorithm is not limited only to various dimensions information.One-dimensional degree eigenwert can be used the interest-degree of model-naive Bayesian predictive user to proposed algorithm equally.The one-dimensional degree eigenwert here refers to the eigenwert of user profile, and more definite saying is: in concrete business scenario, only provide the eigenwert of single dimension, but this user itself exists a lot of other implicit informations that will consider; Therefore in actual applications, adopt NB Algorithm to calculate it.In actual application, the quantity of eigenwert is generally all greater than 1.Above-mentioned one-dimensional degree user characteristics value adopts a specific embodiment of NB Algorithm as follows:
During prediction rent, can be using floor space as single features value prediction rent price, floor space is larger, and rent is higher.And in actual applications, in order to improve accuracy rate, tend to the information such as room number in the facility in the geographic position in house, house, house to calculate as eigenwert.That is to say, the room number in the geographic position in house, the facility in house, house etc. is the necessary implicit information of the prediction rent except floor space.Wherein, the method for eigenvalue is not limited to the choosing method described in the embodiment of the present invention, can also be according to concrete business scenario flexible selected characteristic value.
Bayes's average algorithm is with adopting the model-naive Bayesian difference of single dimensional characteristics value: the eigenwert of Bayes's average algorithm is generally user's behavioral data or user's feedback information.The feature of these class data is, there is cold start-up and constantly increasing, the in the situation that of data volume deficiency, predict the outcome insincere, therefore the present invention adopts Bayes's average algorithm to revise data, and along with the continuous increase of user behavior makes the true interest hobby of its continuous convergence.The eigenwert of model-naive Bayesian is generally user's attribute information; As user's age, sex etc.For this type of information, the present invention adopts the naive Bayesian in machine learning to set up monitor model-user interest model, and according to the model after training, the interest-degree of predictive user to algorithm.Comprehensive above-mentioned steps S220 and step S230, the present invention, in the process that adopts Bayes's average algorithm to calculate one-dimensional degree user characteristics value, adopts parallel calculating method to calculate one-dimensional degree user's proposed algorithm interest-degree; And, in the process that adopts model-naive Bayesian to calculate various dimensions user characteristics value, adopt parallel calculating method to calculate various dimensions user's proposed algorithm interest-degree.
It should be noted that, in order to improve personalized performance of mixing recommended technology, the present invention proposes the algorithm interest tendency that adopts parallel computation framework to calculate user.It is advantageous that and can process the system information of calculating magnanimity.For most of on-line systems, user's quantity and user's behavioural information are ever-increasing.User's personalized interest modeling is closely-related with user's the order of magnitude.If number of users is too huge, the time exponentially level that can make to calculate user's interest model increases, and can not to user's algorithm interest model, upgrade in time, directly affects commending system performance index.
Parameter estimation and prior probability wherein need to be added up the user data in system.The acquisition of these parameters can be used the framework of distributed parallel to calculate.
The parameter estimation of take in above-mentioned NB Algorithm can realize by parallel computation framework the expectation of the interest-degree of algorithm A and variance as example (formula 8.), user; Its formula is as follows:
Every group of user's of parallel computation interest-degree and, merging phase calculates the expectation of all users' interest-degree, in like manner can obtain the Distributed Calculation process of standard deviation, formula is as follows:
Poor average of each user's of parallel computation interest and expectation and, then result is done and is merged.Adopt distributed parallel to calculate the renewal frequency that interest model can guarantee the user interest model of high PV website, thus sharp perception user's interests change.
In the present invention, by parallel computation, can solve the time bottleneck of model training and interest forecast assessment, promote personalized performance of mixing recommended technology; Can also be verified by parallel computation training pattern the commercial viability of the modeling method of Bayes's average algorithm and model-naive Bayesian proposition.
S240: according to one-dimensional degree user's proposed algorithm interest-degree and various dimensions user's proposed algorithm interest-degree, determine recommendation list, and recommendation list is recommended to the user of described database.
Particularly, can recommend policy components to merge one-dimensional degree user's proposed algorithm interest-degree and various dimensions user's proposed algorithm interest-degree by mixing, the interest-degree of the user of acquisition database to each proposed algorithm, according to the interest-degree of each proposed algorithm, obtain the ratio of each proposed algorithm in all proposed algorithms, determine recommendation list, recommendation list is recommended to the user of database.
Wherein, personalized major function of mixing recommendation policy components is that the result of calculation of two kinds of interest models is merged, obtain the interest-degree of user to each proposed algorithm, and calculate the shared recommendation list ratio of each proposed algorithm by formula below, its formula is as follows:
Wherein, W
ifor merging the interest-degree after Bayes's average algorithm result of calculation and model-naive Bayesian result of calculation.
It should be noted that, calculate W
imethod have a lot, except simple linear model, also can adopt regression model, RBM etc.Finally, personalized mixing recommends policy components to pass through P
idetermine recommendation results in recommendation list, represent to final user, complete recommendation.
In addition, it should be noted that, personalized mixing recommends policy components that two kinds of result of calculations are merged, and how to merge and had suitable proven technique, and those skilled in the art can understand, and do not do detailed introduction here.
In the present invention, the proposed algorithm interest-degree that obtains user by Bayes's average algorithm and model-naive Bayesian can be recommended strategy for user provides personalized mixing, strengthen the personalized degree of recommending, improve the degree of accuracy of recommendation results, and take into account good algorithm recall rate.Meanwhile, when recommending, can catch user interest and change, adjust and recommend strategy automatically, and without any need for human intervention.Can avoid experiencing because of the user that on line, A/B test causes the situation of decline.On the contrary, enriching user's algorithm interest model and can more and more understand user along with user data; Make recommendation results more and more suit user's point of interest, improve the user satisfaction of recommending.By the method, can add up very easily the actual effect of proposed algorithm, thereby commending system developer is revised timely and optimize proposed algorithm, form benign cycle.
Corresponding with said method, the present invention also provides a kind of mixing commending system, and Fig. 3 shows the mixing commending system logical organization according to the embodiment of the present invention.
As shown in Figure 3, mixing commending system 300 provided by the invention comprises database community subdivision 310, Bayes's average algorithm computing unit 320, model-naive Bayesian computing unit 330 and algorithm recommendation unit 340.
Wherein, database community subdivision 310 is for being divided into the user data of database one-dimensional degree user characteristics value and various dimensions user characteristics value.
Bayes's average algorithm computing unit 320, for adopting Bayes's average algorithm to calculate one-dimensional degree user characteristics value, obtains one-dimensional degree user's proposed algorithm interest-degree.
Model-naive Bayesian computing unit 330 is for adopting model-naive Bayesian to calculate various dimensions user characteristics value, set up various dimensions user interest model, and various dimensions user interest model is trained, obtain various dimensions user's proposed algorithm interest-degree.
Algorithm recommendation unit 340, for according to one-dimensional degree user's proposed algorithm interest-degree and various dimensions user's proposed algorithm interest-degree, is determined recommendation list, and recommendation list is recommended to the user of database.
Algorithm recommendation unit 340 further comprises interest-degree acquiring unit 341 and recommendation list determining unit 342.
Wherein, interest-degree acquiring unit 341, for recommend policy components to merge described one-dimensional degree user's proposed algorithm interest-degree and described various dimensions user's proposed algorithm interest-degree by mixing, obtains the interest-degree of the user of described database to each proposed algorithm.
Recommendation list determining unit 342, for obtaining each proposed algorithm in the ratio of all proposed algorithms according to the interest-degree of each proposed algorithm, is determined recommendation list.
Wherein, Bayes's average algorithm computing unit 320, in the process that one-dimensional degree user characteristics value is calculated, adopts parallel calculating method to calculate one-dimensional degree user's proposed algorithm interest-degree; Model-naive Bayesian computing unit, in the process that various dimensions user characteristics value is calculated, adopts parallel calculating method to calculate various dimensions user's proposed algorithm interest-degree.
Wherein, in the process that Bayes's average algorithm computing unit 330 calculates one-dimensional degree user characteristics value at employing Bayes average algorithm,
Bayes's average algorithm formula is:
Wherein, C is a constant of setting voluntarily;
M is the mean value of project in database;
N is the data bulk of described project;
X is the value of each data;
In above-mentioned formula, m is prior probability, and the value of each x is an adjustment factor, along with the increase of n, and the value of described project
its correct value will constantly be tended to.
Wherein, in the process that model-naive Bayesian computing unit calculates various dimensions user characteristics value at employing model-naive Bayesian,
Model-naive Bayesian formula is:
Wherein, C is object event; X is eigenwert;
P (C|X) is the probability that under X condition, C occurs;
P (X|C) is the probability that under C condition, X occurs;
The probability that P (C) is C;
The probability that P (X) is X.
By above-mentioned embodiment, can find out, mixing recommend method provided by the invention and system, can calculate the proposed algorithm interest-degree of assessment different user, by adopting Bayes's average algorithm and model-naive Bayesian to obtain user's proposed algorithm interest-degree, improve user's the pre-measuring precision of algorithm interest; By parallel computation, can shorten the training time of user interest model and the time of interest forecast assessment.
Above with reference to accompanying drawing, in the mode of example, the mixing recommend method and the system that according to the present invention, propose have been described.But, it will be appreciated by those skilled in the art that the mixing recommend method and the system that for the invention described above, propose, can also on the basis that does not depart from content of the present invention, make various improvement.Therefore, protection scope of the present invention should be determined by the content of appending claims.
Claims (10)
1. mix a recommend method, comprising:
User data in database is divided into one-dimensional degree user characteristics value and various dimensions user characteristics value;
Adopt Bayes's average algorithm to calculate described one-dimensional degree user characteristics value, obtain described one-dimensional degree user's proposed algorithm interest-degree; And,
Adopt model-naive Bayesian to calculate described various dimensions user characteristics value, set up various dimensions user interest model, and described various dimensions user interest model is trained, obtain described various dimensions user's proposed algorithm interest-degree;
According to described one-dimensional degree user's proposed algorithm interest-degree and described various dimensions user's proposed algorithm interest-degree, determine recommendation list, and described recommendation list is recommended to the user of described database.
2. mixing recommend method as claimed in claim 1, wherein, is determining in the process of recommendation list according to described one-dimensional degree user's proposed algorithm interest-degree and described various dimensions user's proposed algorithm interest-degree,
By mixing, recommend policy components to merge described one-dimensional degree user's proposed algorithm interest-degree and described various dimensions user's proposed algorithm interest-degree, obtain the interest-degree of the user of described database to each proposed algorithm;
According to the interest-degree of described each proposed algorithm, obtain the ratio of each proposed algorithm in all proposed algorithms, determine recommendation list.
3. mixing recommend method as claimed in claim 1, wherein,
In the process that adopts Bayes's average algorithm to calculate described one-dimensional degree user characteristics value, adopt parallel calculating method to calculate described one-dimensional degree user's proposed algorithm interest-degree;
In the process that adopts model-naive Bayesian to calculate described various dimensions user characteristics value, adopt parallel calculating method to calculate described various dimensions user's proposed algorithm interest-degree.
4. mixing recommend method as claimed in claim 1, in the process that adopts Bayes's average algorithm to calculate described one-dimensional degree user characteristics value,
Described Bayes's average algorithm formula is:
Wherein, C is a constant of setting voluntarily;
M is the mean value of project in database;
N is the data bulk of described project;
X is the value of each data;
In above-mentioned formula, m is prior probability, and the value of each x is an adjustment factor, along with the increase of n, and the value of described project
its correct value will constantly be tended to.
5. mixing recommend method as claimed in claim 1, in the process that adopts model-naive Bayesian to calculate described various dimensions user characteristics value,
Described model-naive Bayesian formula is:
Wherein, C is object event; X is eigenwert;
P (C|X) is the probability that under X condition, C occurs;
P (X|C) is the probability that under C condition, X occurs;
The probability that P (C) is C;
The probability that P (X) is X.
6. mix a commending system, comprising:
Database community subdivision, for being divided into the user data of database one-dimensional degree user characteristics value and various dimensions user characteristics value;
Bayes's average algorithm computing unit, for adopting Bayes's average algorithm to calculate described one-dimensional degree user characteristics value, obtains described one-dimensional degree user's proposed algorithm interest-degree;
Model-naive Bayesian computing unit, be used for adopting model-naive Bayesian to calculate described various dimensions user characteristics value, set up various dimensions user interest model, and described various dimensions user interest model is trained, obtain described various dimensions user's proposed algorithm interest-degree;
Algorithm recommendation unit, for according to described one-dimensional degree user's proposed algorithm interest-degree and described various dimensions user's proposed algorithm interest-degree, determines recommendation list, and described recommendation list is recommended to the user of described database.
7. method, system is recommended in mixing as claimed in claim 6, and wherein, described algorithm recommendation unit further comprises:
Interest-degree acquiring unit, for recommend policy components to merge described one-dimensional degree user's proposed algorithm interest-degree and described various dimensions user's proposed algorithm interest-degree by mixing, obtains the interest-degree of the user of described database to each proposed algorithm;
Recommendation list determining unit, for obtaining each proposed algorithm in the ratio of all proposed algorithms according to the interest-degree of described each proposed algorithm, determines recommendation list.
8. method, system is recommended in mixing as claimed in claim 6, wherein,
Described Bayes's average algorithm computing unit, in the process that described one-dimensional degree user characteristics value is calculated, adopts parallel calculating method to calculate described one-dimensional degree user's proposed algorithm interest-degree;
Described model-naive Bayesian computing unit, in the process that described various dimensions user characteristics value is calculated, adopts parallel calculating method to calculate described various dimensions user's proposed algorithm interest-degree.
9. mixing commending system as claimed in claim 6,
In the process of described one-dimensional degree user characteristics value being calculated at described Bayes's average algorithm computing unit employing Bayes average algorithm,
Described Bayes's average algorithm formula is:
Wherein, C is a constant of setting voluntarily;
M is the mean value of project in database;
N is the data bulk of described project;
X is the value of each data;
In above-mentioned formula, m is prior probability, and the value of each x is an adjustment factor, along with the increase of n, and the value of described project
its correct value will constantly be tended to.
10. mixing commending system as claimed in claim 6,
In the process that described model-naive Bayesian computing unit calculates described various dimensions user characteristics value at employing model-naive Bayesian,
Described model-naive Bayesian formula is:
Wherein, wherein, C is object event; X is eigenwert;
P (C|X) is the probability that under X condition, C occurs;
P (X|C) is the probability that under C condition, X occurs;
The probability that P (C) is C;
The probability that P (X) is X.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310737345.0A CN103744917B (en) | 2013-12-27 | 2013-12-27 | Method and system are recommended in mixing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310737345.0A CN103744917B (en) | 2013-12-27 | 2013-12-27 | Method and system are recommended in mixing |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103744917A true CN103744917A (en) | 2014-04-23 |
| CN103744917B CN103744917B (en) | 2018-04-06 |
Family
ID=50501935
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310737345.0A Active CN103744917B (en) | 2013-12-27 | 2013-12-27 | Method and system are recommended in mixing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103744917B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104205100A (en) * | 2014-06-05 | 2014-12-10 | 深圳市推想大数据信息技术有限公司 | Method and device for estimating recessive character distribution of users |
| CN105913286A (en) * | 2016-05-16 | 2016-08-31 | 达而观信息科技（上海）有限公司 | Method for automatically integrating various kinds of personalized recommendation models |
| CN106296242A (en) * | 2015-05-22 | 2017-01-04 | 苏宁云商集团股份有限公司 | A kind of generation method of commercial product recommending list in ecommerce and the system of generation |
| CN106980629A (en) * | 2016-01-19 | 2017-07-25 | 华为技术有限公司 | A kind of network resource recommended method and computer equipment |
| CN108629608A (en) * | 2017-03-22 | 2018-10-09 | 腾讯科技（深圳）有限公司 | User data processing method and processing device |
| CN108876537A (en) * | 2018-06-15 | 2018-11-23 | 重庆知遨科技有限公司 | A kind of mixed recommendation method for on-line mall system |
| CN110070391A (en) * | 2019-04-17 | 2019-07-30 | 同盾控股有限公司 | Data processing method, device, computer-readable medium and electronic equipment |
| CN111209486A (en) * | 2019-12-19 | 2020-05-29 | 杭州安恒信息技术股份有限公司 | Management platform data recommendation method based on mixed recommendation rule |
| CN111861605A (en) * | 2019-04-28 | 2020-10-30 | 阿里巴巴集团控股有限公司 | Business object recommendation method |
| CN109936498B (en) * | 2019-02-26 | 2021-11-19 | 企知道网络技术有限公司 | Full ordering method for prompt of instant communication message |
| CN114254193A (en) * | 2021-12-20 | 2022-03-29 | 北京达佳互联信息技术有限公司 | Content recommendation method and device, electronic equipment and storage medium |
| WO2022213662A1 (en) * | 2021-04-06 | 2022-10-13 | 深圳先进技术研究院 | Application recommendation method and system, terminal, and storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101271558A (en) * | 2008-05-16 | 2008-09-24 | 华东师范大学 | Multi-policy commercial product recommending system based on context information |
| CN103309866A (en) * | 2012-03-09 | 2013-09-18 | 华为技术有限公司 | Method and device for generating recommendation result |
-
2013
- 2013-12-27 CN CN201310737345.0A patent/CN103744917B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101271558A (en) * | 2008-05-16 | 2008-09-24 | 华东师范大学 | Multi-policy commercial product recommending system based on context information |
| CN103309866A (en) * | 2012-03-09 | 2013-09-18 | 华为技术有限公司 | Method and device for generating recommendation result |
Non-Patent Citations (2)
| Title |
|---|
| 向志明: "面向海量用户行为数据的并行个性化推荐算法的研究与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 阮一峰: "基于用户投票的排名算法（六）：贝叶斯平均", 《SONGSHUHUI.NET/ARCHIVES/67443》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2015184619A1 (en) * | 2014-06-05 | 2015-12-10 | 深圳市推想大数据信息技术有限公司 | Method and apparatus for estimating recessive character distribution of users |
| CN104205100B (en) * | 2014-06-05 | 2018-02-02 | 北京推想科技有限公司 | A kind of method and device for the recessive character distribution for estimating user |
| CN104205100A (en) * | 2014-06-05 | 2014-12-10 | 深圳市推想大数据信息技术有限公司 | Method and device for estimating recessive character distribution of users |
| CN106296242A (en) * | 2015-05-22 | 2017-01-04 | 苏宁云商集团股份有限公司 | A kind of generation method of commercial product recommending list in ecommerce and the system of generation |
| US10839032B2 (en) | 2016-01-19 | 2020-11-17 | Huawei Technologies Co., Ltd. | Network resource recommendation method and computer device |
| CN106980629A (en) * | 2016-01-19 | 2017-07-25 | 华为技术有限公司 | A kind of network resource recommended method and computer equipment |
| CN105913286A (en) * | 2016-05-16 | 2016-08-31 | 达而观信息科技（上海）有限公司 | Method for automatically integrating various kinds of personalized recommendation models |
| CN105913286B (en) * | 2016-05-16 | 2020-02-11 | 达而观信息科技（上海）有限公司 | Method for automatically fusing multiple personalized recommendation models |
| CN108629608A (en) * | 2017-03-22 | 2018-10-09 | 腾讯科技（深圳）有限公司 | User data processing method and processing device |
| CN108629608B (en) * | 2017-03-22 | 2023-02-24 | 腾讯科技（深圳）有限公司 | User data processing method and device |
| CN108876537A (en) * | 2018-06-15 | 2018-11-23 | 重庆知遨科技有限公司 | A kind of mixed recommendation method for on-line mall system |
| CN108876537B (en) * | 2018-06-15 | 2021-08-17 | 上海百秋电子商务有限公司 | Mixed recommendation method for online marketplace system |
| CN109936498B (en) * | 2019-02-26 | 2021-11-19 | 企知道网络技术有限公司 | Full ordering method for prompt of instant communication message |
| CN110070391B (en) * | 2019-04-17 | 2020-06-19 | 同盾控股有限公司 | Data processing method and device, computer readable medium and electronic equipment |
| CN110070391A (en) * | 2019-04-17 | 2019-07-30 | 同盾控股有限公司 | Data processing method, device, computer-readable medium and electronic equipment |
| CN111861605A (en) * | 2019-04-28 | 2020-10-30 | 阿里巴巴集团控股有限公司 | Business object recommendation method |
| CN111209486A (en) * | 2019-12-19 | 2020-05-29 | 杭州安恒信息技术股份有限公司 | Management platform data recommendation method based on mixed recommendation rule |
| CN111209486B (en) * | 2019-12-19 | 2023-04-11 | 杭州安恒信息技术股份有限公司 | Management platform data recommendation method based on mixed recommendation rule |
| WO2022213662A1 (en) * | 2021-04-06 | 2022-10-13 | 深圳先进技术研究院 | Application recommendation method and system, terminal, and storage medium |
| CN114254193A (en) * | 2021-12-20 | 2022-03-29 | 北京达佳互联信息技术有限公司 | Content recommendation method and device, electronic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103744917B (en) | 2018-04-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103744917A (en) | Mixed recommendation method and system | |
| US8572011B1 (en) | Outcome estimation models trained using regression and ranking techniques | |
| US8296253B2 (en) | Managing online content based on its predicted popularity | |
| US8744989B1 (en) | Ranking and vote scheduling using statistical confidence intervals | |
| US20090037410A1 (en) | System and method for predicting clickthrough rates and relevance | |
| CN101354714B (en) | Method for recommending problem based on probability latent semantic analysis | |
| CN105824912A (en) | Personalized recommending method and device based on user portrait | |
| CN104391849A (en) | Collaborative filtering recommendation method for integrating time contextual information | |
| CN103678431A (en) | Recommendation method based on standard labels and item grades | |
| Fu et al. | User preference learning with multiple information fusion for restaurant recommendation | |
| EP4181026A1 (en) | Recommendation model training method and apparatus, recommendation method and apparatus, and computer-readable medium | |
| CN109543840B (en) | Dynamic recommendation system design method based on multidimensional classification reinforcement learning | |
| CN109635206B (en) | Personalized recommendation method and system integrating implicit feedback and user social status | |
| Goodwin et al. | The challenges of pre-launch forecasting of adoption time series for new durable products | |
| Wang et al. | Interactive social recommendation | |
| US20140181102A1 (en) | Method and system for robust social choices and vote elicitation | |
| CN103309894A (en) | User attribute-based search realization method and system | |
| CN103761266A (en) | Click rate predicting method and system based on multistage logistic regression | |
| Fung et al. | A fuzzy expected value-based goal programing model for product planning using quality function deployment | |
| CN105761154A (en) | Socialized recommendation method and device | |
| CN110520886B (en) | System and method for eliminating bias in media mix modeling | |
| CN109657962A (en) | A kind of appraisal procedure and system of the volume assets of brand | |
| CN106651427A (en) | Data association method based on user behavior | |
| Cheng et al. | Understanding the impact of individual users’ rating characteristics on the predictive accuracy of recommender systems | |
| CN109636184B (en) | Method and system for evaluating account assets of brands |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |