CN105653683A - Personalized recommendation method and device - Google Patents
Personalized recommendation method and device Download PDFInfo
- Publication number
- CN105653683A CN105653683A CN201511020219.9A CN201511020219A CN105653683A CN 105653683 A CN105653683 A CN 105653683A CN 201511020219 A CN201511020219 A CN 201511020219A CN 105653683 A CN105653683 A CN 105653683A
- Authority
- CN
- China
- Prior art keywords
- predictive model
- sample
- sampled data
- score value
- loss function
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a personalized recommendation method and device. The method comprises the following steps: obtaining current operation information of an operation action executed on a to-be-tested article by a to-be-tested user; obtaining a prediction model, wherein the prediction model is used for expressing corresponding relationship between the operation information and a score value and is obtained through training sample data, and the sample data comprises a sample score value from a sample user to a sample article and sample operation information of an operation action executed on the sample article by the sample user; obtaining a predicted score value corresponding to the current operation information by utilizing the prediction model; and carrying out personalized recommendation on the basis of the predicted score value. In such way, the personalized recommendation can be realized on the basis of recessive behaviors carried out on the articles by users.
Description
Technical field
The present invention relates to data processing field, specifically, it relates to a kind of personalized recommendation method and device.
Background technology
Along with the development of information technology, in order to better serve user, personalized recommendation technology is arisen at the historic moment, for providing, for user, the personalized recommendation content meeting its demand. Usually, personalized recommendation technology based on user behavior data, then in conjunction with certain data analysing method, for user generates personalized recommendation content.
Specifically, user's behavior can be divided into dominant behavior and recessive behavior. Dominant behavior can be the fancy grade to article that user intuitively shows, such as user for article are marked. Recessive behavior can be the operational motion that article are performed by user, as operational motion can praise for browsing, buy, evaluate, collect, forward, put, download etc.
When carrying out personalized recommendation, as in the personalized recommendation scheme that realizes based on collaborative filtering, can first utilize user that the score value of article is calculated the similarity between user, determine the neighbor user with target user with same or similar interest preference, according to neighbor user, the score value of article is carried out target of prediction user again and treat the score value pushing away article, and then using article to be pushed away higher for prediction score value as content recommendation, it is sent to target user. Wherein, predict that higher can be understood as of score value has exceeded preset value.
It can thus be seen that the score value of article is played a part most important by user in personalized recommendation scheme. For dominant behavior, it is possible to directly utilize the user grabbed to the score value of article, carry out personalized recommendation. But, for recessive behavior, user usually can only be captured to the operation information of operational motion performed by article, e.g., operational motion is for browsing, then the information that operates can be presented as when browsing long; Operational motion for buying, then operates information and can be presented as purchase number of times, etc., how to utilize the operation information grabbed, carry out personalized recommendation, become current problem demanding prompt solution.
Summary of the invention
It is an object of the invention to provide a kind of personalized recommendation method and device, it is possible to based on user to the recessive behavior of article, it is achieved personalized recommendation.
Embodiments providing a kind of personalized recommendation method, described method comprises:
Obtain user to be measured to the current operation information of operational motion performed by article to be measured;
Obtain predictive model, the corresponding relation of described predictive model for representing between operation information and score value, obtaining described predictive model by sampled data carries out training, described sampled data comprises sample of users to the sample score value of sample article and sample of users to the sample operations information of the operational motion that sample article perform;
Utilize described predictive model, obtain the prediction score value that described current operation information is corresponding;
Based on described prediction score value, carry out personalized recommendation.
Can selection of land, described predictive model is: y=��1x1+��2x2+��+��ixi+��+��nxn+��0;
Wherein, y represents score value; ��iRepresent the weighted value of i-th operational motion; xiRepresent the operation information of i-th operational motion; 1��i��n, n is the number of types of operational motion; ��0For constant.
Can selection of land, described sampled data is carried out training obtain described predictive model, comprising:
Obtain the described sampled data of many groups;
Set up original predictive model, and utilize described original predictive model, obtain and often organize prior estimate corresponding to the sample operations information in sampled data;
Set up loss function, the deviation that described loss function is often organized between the prior estimate corresponding with the sample operations information in this group sampled data of the sample score value in sampled data for representing;
Adjust described original predictive model, and when described loss function reaches minimum value, obtain described predictive model.
Can selection of land, described loss function is:
Wherein, J (��) represents loss function; yjRepresent the prior estimate that the sample operations information in jth group sampled data is corresponding; zjRepresenting the sample score value in jth group sampled data, 1��j��m, m is the number of sampled data.
Can selection of land, described original predictive model is:
Y=��11x1+��12x2+��+��1ixi+��+��1nxn+��10, wherein, ��1iRepresent the initial weight value of i-th operational motion; ��10Represent initial constant; Then
The described original predictive model of described adjustment, and when described loss function reaches minimum value, obtain described predictive model, comprising:
The �� in described original predictive model is adjusted by following formula1i: ��2i=��1i+��(zj-yj)xji; Wherein, ��2iRepresent after described original predictive model is carried out once adjustment, weighted value after the adjustment of i-th operational motion; �� represents learning rate; (zj-yj)xjiBy described loss function J (��) to ��1iAsk and partially lead acquisition; xjiRepresenting the sample operations information of i-th operational motion in jth group sampled data, 1��j��m, m is the number of sampled data;
Judge described ��2iDescribed loss function whether is made to reach minimum value, if it does, then by described ��2iIt is defined as described ��i, obtain described predictive model; If not, then at described ��2iBasis on, utilize described �� and described (zj-yj)xji, adjustment obtains described predictive model.
The embodiment of the present invention additionally provides a kind of personalized recommendation device, and described device comprises:
Operation information acquisition unit, for obtaining user to be measured to the current operation information of operational motion performed by article to be measured;
Predictive model acquiring unit, for obtaining predictive model, the corresponding relation of described predictive model for representing between operation information and score value, obtaining described predictive model by sampled data carries out training, described sampled data comprises sample of users to the sample score value of sample article and sample of users to the sample operations information of the operational motion that sample article perform;
Score value obtaining unit, for utilizing described predictive model, obtains the prediction score value that described current operation information is corresponding;
Personalized recommendation unit, for based on described prediction score value, carrying out personalized recommendation.
Can selection of land, described predictive model acquiring unit, specifically for obtaining following predictive model:
Y=��1x1+��2x2+��+��ixi+��+��nxn+��0;
Wherein, y represents score value; ��iRepresent the weighted value of i-th operational motion; xiRepresent the operation information of i-th operational motion; 1��i��n, n is the number of types of operational motion; ��0For constant.
Can selection of land, described device also comprises:
Sampled data acquiring unit, for obtaining the described sampled data of many groups;
Unit set up by model, for setting up original predictive model, and utilizes described original predictive model, obtains and often organizes prior estimate corresponding to the sample operations information in sampled data;
Loss function sets up unit, for setting up loss function, and the deviation that described loss function is often organized between the prior estimate corresponding with the sample operations information in this group sampled data of the sample score value in sampled data for representing;
Model adjustment unit, for adjusting described original predictive model, and when described loss function reaches minimum value, obtains described predictive model.
Can selection of land, described loss function sets up unit, loses function as follows specifically for setting up:
Wherein, J (��) represents loss function; yjRepresent the prior estimate that the sample operations information in jth group sampled data is corresponding; zjRepresenting the sample score value in jth group sampled data, 1��j��m, m is the number of sampled data.
Can selection of land, unit set up by described model, specifically for setting up following original predictive model:
Y=��11x1+��12x2+��+��1ixi+��+��1nxn+��10, wherein, ��1iRepresent the initial weight value of i-th operational motion; ��10Represent initial constant; Then
Described model adjustment unit, specifically for adjusting the �� in described original predictive model by following formula1i: ��2i=��1i+��(zj-yj)xji; Wherein, ��2iRepresent after described original predictive model is carried out once adjustment, weighted value after the adjustment of i-th operational motion; �� represents learning rate; (zj-yj)xjiBy described loss function J (��) to ��1iAsk and partially lead acquisition; xjiRepresenting the sample operations information of i-th operational motion in jth group sampled data, 1��j��m, m is the number of sampled data; Judge described ��2iDescribed loss function whether is made to reach minimum value, if it does, then by described ��2iIt is defined as described ��i, obtain described predictive model; If not, then at described ��2iBasis on, utilize described �� and described (zj-yj)xji, adjustment obtains described predictive model.
In technical solution of the present invention, it is possible to by representing the predictive model of corresponding relation between recessive behavior and score value, recessive for the user grabbed behavior is converted to user to the prediction score value of article, so, just can realize personalized recommendation based on prediction score value. In addition, obtained the mode of predictive model by learning sample data, also contribute to improving the accuracy of score value of the present invention prediction.
Other features and advantages of the present invention are described in detail in embodiment part subsequently.
Accompanying drawing explanation
Accompanying drawing is used to provide a further understanding of the present invention, and forms a part for specification sheets, is used from explanation the present invention with embodiment one below, but is not construed as limiting the invention. In the accompanying drawings:
Fig. 1 is the schema of personalized recommendation method of the present invention;
Fig. 2 is the schema of the method obtaining predictive model in the present invention;
Fig. 3 is the schematic diagram of consumer articles hobby matrix in the present invention;
Fig. 4 is the structural representation of personalized recommendation device of the present invention.
Embodiment
Below in conjunction with accompanying drawing, the specific embodiment of the present invention is described in detail.Should be understood that, embodiment described herein, only for instruction and explanation of the present invention, is not limited to the present invention.
See Fig. 1, show the schema of a kind of personalized recommendation method of the present invention, can comprise:
S101, obtains user to be measured to the current operation information of operational motion performed by article to be measured.
S102, obtain predictive model, the corresponding relation of described predictive model for representing between operation information and score value, obtaining described predictive model by sampled data carries out training, described sampled data comprises sample of users to the sample score value of sample article and sample of users to the sample operations information of the operational motion that sample article perform.
In the present invention program, it is possible to use train the predictive model obtained, estimate user corresponding to recessive behavior to the prediction score value of article, and then utilize prediction score value to carry out personalized recommendation.
First, the present invention program is when carrying out score value and predict, it is possible to obtain following two aspect information:
On the one hand, user to be measured is to the current operation information of operational motion performed by article to be measured.
, it is possible to use grab current operation information in real time, for example online score value prediction is carried out; Or, it is possible to when needs carry out personalized recommendation, recycling the current operation information previously grabbed, carry out off-line score value prediction, this can be not specifically limited by the embodiment of the present invention.
Such as, for user A watch video A grab following information: user A watch video A time long, user A has collected video A, it can be seen that operational motion 1 is for browsing, long when current operation information 1 is for browsing; Operational motion 2 is collection, and current operation information 2 is collection number of times is 1.
On the other hand, predictive model, it is possible to the corresponding relation between expression operation information and score value. In the present invention program, predictive model can be: y=��1x1+��2x2+��+��ixi+��+��nxn+��0. Wherein, y represents score value; ��iRepresent the weighted value of i-th operational motion; xiRepresent the operation information of i-th operational motion; 1��i��n, n is the number of types of operational motion; ��0For constant.
In application process, it is possible to choose the type of operational motion in predictive model in conjunction with actual demand. For example, it is possible to determined by statistical study, for most of article, which operational motion is relatively big on the impact of prediction score value, and the impact as bought and browse is relatively big, then at least can comprise in predictive model and buy and browse this two operational motions. Or, it is possible in conjunction with the characteristic of some special article, it is determined that go out for this special article, which operational motion is bigger on the impact of prediction score value, as for special article A, the impact of evaluation is relatively big, then predictive model at least can comprise and evaluate this operational motion. Or, it is possible to comprehensively choosing more operational motion type, this can be not specifically limited by the embodiment of the present invention as far as possible.
As a kind of example, the present invention also provides a kind of scheme obtaining predictive model, specifically can introduce see hereafter Fig. 2 place, wouldn't describe in detail herein.
S103, utilizes described predictive model, obtains the prediction score value that described current operation information is corresponding.
After getting above-mentioned two aspect information, predictive model just can be utilized to obtain the user to be measured prediction score value that the recessive behavior of article to be measured is corresponding.
As in above-mentioned example, if y=is ��1x1+��2x2+��0, wherein, ��1The weighted value of length when score value is predicted during for browsing;X1During for browsing long; ��2For being collected in weighted value when score value is predicted; x2For collection number of times, then according to this predictive model and current operation information, it is possible to prediction obtains prediction score value corresponding to recessive behavior when user A watches video A.
As in above-mentioned example, if y=is ��1x1+��2x2+��3x3+��0, that is, predictive model comprises the operation information of the operational motion not grabbed, such as ��3For buying the weighted value when score value is predicted; x3For buying number of times. Corresponding to this, it is possible to the operation information of the operational motion not grabbed by this kind is considered as 0, as in above-mentioned example, it is possible to understand that be x3=0, so, according to this predictive model and current operation information, the prediction score value that recessive behavior when obtaining user A viewing video A is corresponding can be predicted equally.
S104, based on described prediction score value, carries out personalized recommendation.
For example, user to be measured is U1, article to be measured are T1, utilize the present invention program to obtain U1To T1The prediction score value X that performed operational motion is corresponding11After, it is possible to by X11Fill in the hobby matrix of consumer articles shown in Fig. 3, during for personalized recommendation. Matrix shown in Fig. 3 can represent that g user is to the score value of k article.
Such as, when carrying out personalized recommendation based on collaborative filtering, it is possible to use X11Calculate U1And the similarity between other users, it is determined that go out U1Neighbor user, and then according to neighbor user to the score value of article, be U1Determine content recommendation, it is achieved for U1Personalized recommendation. Or, when other users in matrix are carried out personalized recommendation, as to U2Carry out personalized recommendation, it is possible to based on X11Judge U1Whether it is U2Neighbor user, if U1It is U2Neighbor user, also can further according to U1To the score value of article, it is U2Determine content recommendation, it is achieved for U2Personalized recommendation. That is, based on the prediction score value in the present invention program, it is possible not only to user to be measured is carried out personalized recommendation, it is also possible to other users are carried out personalized recommendation.
To sum up, utilize the present invention program, just recessive for the user grabbed behavior can be converted to user to the prediction score value of article, it is achieved the present invention carries out the object of personalized recommendation based on recessive behavior. Furthermore, it may be desirable to utilize train the predictive model obtained, estimate the prediction score value that recessive behavior is corresponding, also contribute to improving the accuracy of score value of the present invention prediction.
See Fig. 2, show the schema that the present invention obtains the method for predictive model, it is possible to comprising:
S201, obtains the described sampled data of many groups.
In the present invention program, by capturing the history behavioral data of sample of users, obtain described sampled data, for example, sampled data can be presented as: the recessive behavior (can be specially operational motion, operation information two aspect content) of sample of users sample article sample score value, wherein, recessive behavior can be Dan Weidu, namely only includes a kind of operational motion; Or recessive behavior can be various dimensions, namely comprises at least two kinds of operational motions.
Understandably, sample article can for having general representative article, or, sample article can for having the article of higher degree of relation with article to be measured, and this can be not specifically limited by the embodiment of the present invention.
In addition, about the number of the sampled data participating in training, usually, the number of sampled data is many, and the accuracy of the predictive model obtained can be higher, therefore, it is possible to suitably choose the sampled data of more group, train;Or, it is also possible to according to the accuracy of training gained predictive model, the number of the selected sampled data of the adjustment of adaptability.
For example, it is possible to first choose the sampled data that quantity is N, and train based on this N group sampled data, obtain predictive model; And then selected part test sample book data, and utilize the accuracy of described test sample book Data Detection predictive model, if the accuracy of predictive model is low, then can on the basis of above-mentioned N group sampled data again, choose M group sampled data more more, and utilize (N+M) to organize sampled data and train, obtain the predictive model after upgrading. That is, the present invention can by suitably increasing the mode of sampled data capacity, it is to increase the accuracy of predictive model. Wherein, the accuracy of predictive model is low it is understood that the accuracy of predictive model is lower than preset value, and preset value can not be limited by the embodiment of the present invention, it is possible to determines in conjunction with practical application request.
S202, sets up original predictive model, and utilizes described original predictive model, obtains and often organizes prior estimate corresponding to the sample operations information in sampled data.
As a kind of example, the original predictive model in the present invention program can be presented as following formula:
Y=��11x1+��12x2+��+��1ixi+��+��1nxn+��10
Wherein, ��1iRepresent the initial weight value of i-th operational motion; ��10Represent initial constant. Specifically, it is possible to be �� at random1iAnd ��10Assignment, this can be not specifically limited by the embodiment of the present invention.
S203, sets up loss function, the deviation that described loss function is often organized between the prior estimate corresponding with the sample operations information in this group sampled data of the sample score value in sampled data for representing.
Except original predictive model, the present invention program also can set up loss function (LossFunction), prior estimate corresponding to sample operations information that this function can represent sample score value, obtain based on predictive model, gap therebetween. Usually, loss function is when reaching minimum value, when namely gap between prior estimate and sample score value is minimum, it is possible to think that the predictive model now used is the most accurate. So, just loss function can being made to reach minimum value as condition, adjustment original predictive model, obtains the predictive model in the present invention program.
Specifically, it is possible to set up loss function in the following manner.
Assume ��j=yj-zjRepresent the prior estimate y of jth group sampled datajWith sample score value zjBetween deviation. Usually, produce ��jReason may be relevant with the operational motion type chosen in predictive model, or in sampled data exist some random noises. Assume ��jFor independent same distribution, and Normal Distribution, i.e. ��j��N (0, ��2), then have:
Thus can derive:
P(yj|xj& ��) represent given xjWith the y after ��jDistribution probability.
Specifically, P (yj|xj& ��) it is the probability obtained for this group sampled data of jth group, the present invention should be the most punctual in the prediction of all sampled datas, seeks the value of ��, the product of probability of all sampled datas namely should be made maximum. As a kind of example, it is possible to solving maximum probability by method of maximum likelihood and amass, specifically, likelihood function L (��) can be presented as following formula:
Wherein, m is the number of sampled data.
L (��) is taken the logarithm:
Make logL(��)Maximum, also it is exactly makeMinimum, therefore, loss function J (��) in the present invention program can be presented as following formula:
S204, adjusts described original predictive model, and when described loss function reaches minimum value, obtains described predictive model.
By analyzing above it will be seen that lose function when reaching minimum value, prior estimate is closest to sample score value, therefore the present invention program obtains the process of predictive model, it is possible to be converted into the process solving loss function minimum. Specifically, it is possible to by modes such as stochastic gradient descent method, method of least squares, solving loss functional minimum value, this can be not specifically limited by the embodiment of the present invention.
Below to obtain predictive model by stochastic gradient descent method, the process of adjustment original predictive model is explained.
First, it is possible to adjust the �� in original predictive model by following formula1i:
Wherein, ��2iRepresent after original predictive model is carried out once adjustment, weighted value after the adjustment of i-th operational motion; �� is learning rate, can be used for representing the adjustment step-length of iteration adjustment; (zj-yj)xjiCan by loss function J (��) to ��1iAsk and partially lead acquisition, can be used for representing the adjustment direction of iteration adjustment; xjiRepresent the sample operations information of i-th operational motion in jth group sampled data.
To ��1iAsk and partially lead and can be presented as:
The present invention program adjusts original predictive model, it is possible to understand that be the parameter �� in adjustment model1iAnd ��10Value, specifically can relate to following two aspect information:
On the one hand, can be used for representing the �� of adjustment step-length. Specifically, it is possible to arrange �� according to experience, usually, it is less that �� is arranged, and the speed of convergence of loss function can be slow, but can obviously find the convergence point losing function, i.e. minimum value; If �� arrange relatively big, the speed of convergence of loss function than comparatively fast, but may can not easily find convergence point. As a kind of example, it is possible to arrange several �� more, by repeatedly attempting, choose suitable ��. The value of �� can be not specifically limited by the embodiment of the present invention.
On the other hand, can be used for representing (the z of adjustment directionj-yj)xji. Specifically, it is possible to by loss function J (��) to ��1iAsking and partially lead acquisition, usually, adjustment direction is the opposite direction of required partial derivative. In addition, it is necessary to explanation, if the predictive model in the present invention program is for the recessive behavior of Dan Weidu, namely original predictive model is y=��11x1+��10, then adjustment direction can by loss function J (��) to ��11Ask and lead acquisition.
In addition, it is necessary to explanation, with adjustment ��1iMutually similar, it is possible to adjust the �� in original predictive model according to lower formula10:
Secondly, judge the �� after adjustment2iLoss function whether is made to reach minimum value.
Whether, as a kind of example, it can be determined that the weighted value of operational motion, constantly hover at convergence point place, namely in nearest several iteration adjustment processes, whether weighted value is substantially constant. If it does, then can think that loss function reaches minimum value.
Such as, it is possible to use �� and (zj-yj)xji, obtain ��2i�� after adjustment3i, and ��3i�� after adjustment4iDeng, if ��2i����3iAnd ��4iSubstantially remain unchanged, then can judge that loss function has reached minimum value, it is possible to by ��2iThe �� being defined as in predictive modeli, obtain the predictive model in the present invention program. Specifically,
If through ��1iTo ��2iAn iteration adjustment, loss function do not reach minimum value, then can at ��2iBasis on, utilize �� and (zj-yj)xji, proceed iteration adjustment, until when determining that loss function reaches minimum value, obtaining predictive model. Detailed process see introducing above, no longer can be illustrated herein.
Corresponding with method shown in Fig. 1, the embodiment of the present invention also provides a kind of personalized recommendation device 300, schematic diagram shown in Figure 4, and described device can comprise:
Operation information acquisition unit 301, for obtaining user to be measured to the current operation information of operational motion performed by article to be measured;
Predictive model acquiring unit 302, for obtaining predictive model, the corresponding relation of described predictive model for representing between operation information and score value, obtaining described predictive model by sampled data carries out training, described sampled data comprises sample of users to the sample score value of sample article and sample of users to the sample operations information of the operational motion that sample article perform;
Score value obtaining unit 303, for utilizing described predictive model, obtains the prediction score value that described current operation information is corresponding;
Personalized recommendation unit 304, for based on described prediction score value, carrying out personalized recommendation.
Can selection of land, described predictive model acquiring unit, specifically for obtaining following predictive model:
Y=��1x1+��2x2+��+��ixi+��+��nxn+��0;
Wherein, y represents score value; ��iRepresent the weighted value of i-th operational motion; xiRepresent the operation information of i-th operational motion; 1��i��n, n is the number of types of operational motion; ��0For constant.
Can selection of land, described device also comprises:
Sampled data acquiring unit, for obtaining the described sampled data of many groups;
Unit set up by model, for setting up original predictive model, and utilizes described original predictive model, obtains and often organizes prior estimate corresponding to the sample operations information in sampled data;
Loss function sets up unit, for setting up loss function, and the deviation that described loss function is often organized between the prior estimate corresponding with the sample operations information in this group sampled data of the sample score value in sampled data for representing;
Model adjustment unit, for adjusting described original predictive model, and when described loss function reaches minimum value, obtains described predictive model.
Can selection of land, described loss function sets up unit, loses function as follows specifically for setting up:
Wherein, J (��) represents loss function; yjRepresent the prior estimate that the sample operations information in jth group sampled data is corresponding; zjRepresenting the sample score value in jth group sampled data, 1��j��m, m is the number of sampled data.
Can selection of land, unit set up by described model, specifically for setting up following original predictive model:
Y=��11x1+��12x2+��+��1ixi+��+��1nxn+��10, wherein, ��1iRepresent the initial weight value of i-th operational motion; ��10Represent initial constant; Then
Described model adjustment unit, specifically for adjusting the �� in described original predictive model by following formula1i:Wherein, ��2iRepresent after described original predictive model is carried out once adjustment, weighted value after the adjustment of i-th operational motion; �� represents learning rate; (zj-yj)xjiBy described loss function J (��) to ��1iAsk and partially lead acquisition; xjiRepresenting the sample operations information of i-th operational motion in jth group sampled data, 1��j��m, m is the number of sampled data; Judge described ��2iDescribed loss function whether is made to reach minimum value, if it does, then by described ��2iIt is defined as described ��i, obtain described predictive model; If not, then at described ��2iBasis on, utilize described �� and described (zj-yj)xji, adjustment obtains described predictive model.
Below the preferred embodiment of the present invention is described by reference to the accompanying drawings in detail; but; the detail that the present invention is not limited in above-mentioned enforcement mode; within the scope of the technical conceive of the present invention; the technical scheme of the present invention can being carried out multiple simple variant, these simple variant all belong to protection scope of the present invention.
It should be noted that in addition, each concrete technology feature described in above-mentioned embodiment, when not contradiction, it is possible to combined by any suitable mode, in order to avoid unnecessary repetition, various possible array mode is illustrated by the present invention no longer separately.
In addition, can also carrying out arbitrary combination between the various different enforcement mode of the present invention, as long as it does not run counter to the thought of the present invention, it should be considered as content disclosed in this invention equally.
Claims (10)
1. a personalized recommendation method, it is characterised in that, described method comprises:
Obtain user to be measured to the current operation information of operational motion performed by article to be measured;
Obtain predictive model, the corresponding relation of described predictive model for representing between operation information and score value, obtaining described predictive model by sampled data carries out training, described sampled data comprises sample of users to the sample score value of sample article and sample of users to the sample operations information of the operational motion that sample article perform;
Utilize described predictive model, obtain the prediction score value that described current operation information is corresponding;
Based on described prediction score value, carry out personalized recommendation.
2. method according to claim 1, it is characterised in that, described predictive model is:
Y=��1x1+��2x2+��+��ixi+��+��nxn+��0;
Wherein, y represents score value; ��iRepresent the weighted value of i-th operational motion; xiRepresent the operation information of i-th operational motion; 1��i��n, n is the number of types of operational motion; ��0For constant.
3. method according to claim 1 and 2, it is characterised in that, described sampled data being carried out trains the described predictive model of acquisition, comprising:
Obtain the described sampled data of many groups;
Set up original predictive model, and utilize described original predictive model, obtain and often organize prior estimate corresponding to the sample operations information in sampled data;
Set up loss function, the deviation that described loss function is often organized between the prior estimate corresponding with the sample operations information in this group sampled data of the sample score value in sampled data for representing;
Adjust described original predictive model, and when described loss function reaches minimum value, obtain described predictive model.
4. method according to claim 3, it is characterised in that, described loss function is:
Wherein, J (��) represents loss function; yjRepresent the prior estimate that the sample operations information in jth group sampled data is corresponding; zjRepresenting the sample score value in jth group sampled data, 1��j��m, m is the number of sampled data.
5. method according to claim 3, it is characterised in that,
Described original predictive model is: y=��11x1+��12x2+��+��1ixi+��+��1nxn+��10, wherein, ��1iRepresent the initial weight value of i-th operational motion; ��10Represent initial constant; Then
The described original predictive model of described adjustment, and when described loss function reaches minimum value, obtain described predictive model, comprising:
The �� in described original predictive model is adjusted by following formula1i: Wherein, ��2iRepresent after described original predictive model is carried out once adjustment, weighted value after the adjustment of i-th operational motion; �� represents learning rate; (zj-yj)xjiBy described loss function J (��) to ��1iAsk and partially lead acquisition; xjiRepresenting the sample operations information of i-th operational motion in jth group sampled data, 1��j��m, m is the number of sampled data;
Judge described ��2iDescribed loss function whether is made to reach minimum value, if it does, then by described ��2iIt is defined as described ��i, obtain described predictive model; If not, then at described ��2iBasis on, utilize described �� and described (zj-yj)xji, adjustment obtains described predictive model.
6. a personalized recommendation device, it is characterised in that, described device comprises:
Operation information acquisition unit, for obtaining user to be measured to the current operation information of operational motion performed by article to be measured;
Predictive model acquiring unit, for obtaining predictive model, the corresponding relation of described predictive model for representing between operation information and score value, obtaining described predictive model by sampled data carries out training, described sampled data comprises sample of users to the sample score value of sample article and sample of users to the sample operations information of the operational motion that sample article perform;
Score value obtaining unit, for utilizing described predictive model, obtains the prediction score value that described current operation information is corresponding;
Personalized recommendation unit, for based on described prediction score value, carrying out personalized recommendation.
7. device according to claim 6, it is characterised in that, described predictive model acquiring unit, specifically for obtaining following predictive model:
Y=��1x1+��2x2+��+��ixi+��+��nxn+��0;
Wherein, y represents score value; ��iRepresent the weighted value of i-th operational motion; xiRepresent the operation information of i-th operational motion; 1��i��n, n is the number of types of operational motion; ��0For constant.
8. device according to claim 6 or 7, it is characterised in that, described device also comprises:
Sampled data acquiring unit, for obtaining the described sampled data of many groups;
Unit set up by model, for setting up original predictive model, and utilizes described original predictive model, obtains and often organizes prior estimate corresponding to the sample operations information in sampled data;
Loss function sets up unit, for setting up loss function, and the deviation that described loss function is often organized between the prior estimate corresponding with the sample operations information in this group sampled data of the sample score value in sampled data for representing;
Model adjustment unit, for adjusting described original predictive model, and when described loss function reaches minimum value, obtains described predictive model.
9. device according to claim 8, it is characterised in that, described loss function sets up unit, loses function as follows specifically for setting up:
Wherein, J (��) represents loss function; yjRepresent the prior estimate that the sample operations information in jth group sampled data is corresponding; zjRepresenting the sample score value in jth group sampled data, 1��j��m, m is the number of sampled data.
10. device according to claim 8, it is characterised in that,
Unit set up by described model, specifically for setting up following original predictive model:
Y=��11x1+��12x2+��+��1ixi+��+��1nxn+��10, wherein, ��1iRepresent the initial weight value of i-th operational motion; ��10Represent initial constant; Then
Described model adjustment unit, specifically for adjusting the �� in described original predictive model by following formula1i:Wherein, ��2iRepresent after described original predictive model is carried out once adjustment, weighted value after the adjustment of i-th operational motion; �� represents learning rate; (zj-yj)xjiBy described loss function J (��) to ��1iAsk and partially lead acquisition; xjiRepresenting the sample operations information of i-th operational motion in jth group sampled data, 1��j��m, m is the number of sampled data; Judge described ��2iDescribed loss function whether is made to reach minimum value, if it does, then by described ��2iIt is defined as described ��i, obtain described predictive model; If not, then at described ��2iBasis on, utilize described �� and described (zj-yj)xji, adjustment obtains described predictive model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201511020219.9A CN105653683B (en) | 2015-12-30 | 2015-12-30 | Personalized recommendation method and device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201511020219.9A CN105653683B (en) | 2015-12-30 | 2015-12-30 | Personalized recommendation method and device |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105653683A true CN105653683A (en) | 2016-06-08 |
CN105653683B CN105653683B (en) | 2020-10-16 |
Family
ID=56478136
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201511020219.9A Active CN105653683B (en) | 2015-12-30 | 2015-12-30 | Personalized recommendation method and device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105653683B (en) |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106600372A (en) * | 2016-12-12 | 2017-04-26 | 武汉烽火信息集成技术有限公司 | Commodity recommending method and system based on user behaviors |
CN106779204A (en) * | 2016-12-08 | 2017-05-31 | 电子科技大学 | The method and device that a kind of personalized path is recommended |
CN106846191A (en) * | 2016-11-25 | 2017-06-13 | 北京粉笔蓝天科技有限公司 | A kind of method of combination of curriculums table, system and server |
CN107577736A (en) * | 2017-08-25 | 2018-01-12 | 上海斐讯数据通信技术有限公司 | A kind of file recommendation method and system based on BP neural network |
CN108334729A (en) * | 2017-08-28 | 2018-07-27 | 江西博瑞彤芸科技有限公司 | Health information management method and management system |
WO2018214503A1 (en) * | 2017-05-23 | 2018-11-29 | 北京三快在线科技有限公司 | Method and device for setting sample weight, and electronic apparatus |
CN109408731A (en) * | 2018-12-27 | 2019-03-01 | 网易(杭州)网络有限公司 | A kind of multiple target recommended method, multiple target recommended models generation method and device |
CN110008404A (en) * | 2019-03-22 | 2019-07-12 | 成都理工大学 | Enigmatic language justice model optimization method based on the optimization of NAG momentum |
CN110032498A (en) * | 2018-11-23 | 2019-07-19 | 浙江每日互动网络科技股份有限公司 | A kind of prediction technique of user APP behavior |
CN110049079A (en) * | 2018-01-16 | 2019-07-23 | 阿里巴巴集团控股有限公司 | Information push and model training method, device, equipment and storage medium |
CN110110205A (en) * | 2018-01-16 | 2019-08-09 | 北京京东金融科技控股有限公司 | Recommendation information generation method and device |
CN110309417A (en) * | 2018-04-13 | 2019-10-08 | 腾讯科技(深圳)有限公司 | The Weight Determination and device of evaluation points |
CN111159555A (en) * | 2019-12-30 | 2020-05-15 | 北京每日优鲜电子商务有限公司 | Commodity recommendation method, commodity recommendation device, server and storage medium |
WO2021081962A1 (en) * | 2019-10-31 | 2021-05-06 | 华为技术有限公司 | Recommendation model training method, recommendation method, device, and computer-readable medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120179638A1 (en) * | 2011-01-11 | 2012-07-12 | National Tsing Hua University | Relative variable selection system and selection method thereof |
CN103377296A (en) * | 2012-04-19 | 2013-10-30 | 中国科学院声学研究所 | Data mining method for multi-index evaluation information |
CN104156472A (en) * | 2014-08-25 | 2014-11-19 | 四达时代通讯网络技术有限公司 | Video recommendation method and system |
CN104537114A (en) * | 2015-01-21 | 2015-04-22 | 清华大学 | Individual recommendation method |
CN105069122A (en) * | 2015-08-12 | 2015-11-18 | 天津大学 | Personalized recommendation method and recommendation apparatus based on user behaviors |
-
2015
- 2015-12-30 CN CN201511020219.9A patent/CN105653683B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120179638A1 (en) * | 2011-01-11 | 2012-07-12 | National Tsing Hua University | Relative variable selection system and selection method thereof |
US8706670B2 (en) * | 2011-01-11 | 2014-04-22 | National Tsing Hua University | Relative variable selection system and selection method thereof |
CN103377296A (en) * | 2012-04-19 | 2013-10-30 | 中国科学院声学研究所 | Data mining method for multi-index evaluation information |
CN104156472A (en) * | 2014-08-25 | 2014-11-19 | 四达时代通讯网络技术有限公司 | Video recommendation method and system |
CN104537114A (en) * | 2015-01-21 | 2015-04-22 | 清华大学 | Individual recommendation method |
CN105069122A (en) * | 2015-08-12 | 2015-11-18 | 天津大学 | Personalized recommendation method and recommendation apparatus based on user behaviors |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106846191A (en) * | 2016-11-25 | 2017-06-13 | 北京粉笔蓝天科技有限公司 | A kind of method of combination of curriculums table, system and server |
CN106779204A (en) * | 2016-12-08 | 2017-05-31 | 电子科技大学 | The method and device that a kind of personalized path is recommended |
CN106779204B (en) * | 2016-12-08 | 2021-02-09 | 电子科技大学 | Personalized path recommendation method and device |
CN106600372A (en) * | 2016-12-12 | 2017-04-26 | 武汉烽火信息集成技术有限公司 | Commodity recommending method and system based on user behaviors |
WO2018214503A1 (en) * | 2017-05-23 | 2018-11-29 | 北京三快在线科技有限公司 | Method and device for setting sample weight, and electronic apparatus |
CN107577736A (en) * | 2017-08-25 | 2018-01-12 | 上海斐讯数据通信技术有限公司 | A kind of file recommendation method and system based on BP neural network |
CN107577736B (en) * | 2017-08-25 | 2021-12-17 | 武汉数字智能信息科技有限公司 | File recommendation method and system based on BP neural network |
CN108334729A (en) * | 2017-08-28 | 2018-07-27 | 江西博瑞彤芸科技有限公司 | Health information management method and management system |
CN110049079A (en) * | 2018-01-16 | 2019-07-23 | 阿里巴巴集团控股有限公司 | Information push and model training method, device, equipment and storage medium |
CN110110205A (en) * | 2018-01-16 | 2019-08-09 | 北京京东金融科技控股有限公司 | Recommendation information generation method and device |
CN110309417A (en) * | 2018-04-13 | 2019-10-08 | 腾讯科技(深圳)有限公司 | The Weight Determination and device of evaluation points |
CN110032498A (en) * | 2018-11-23 | 2019-07-19 | 浙江每日互动网络科技股份有限公司 | A kind of prediction technique of user APP behavior |
CN109408731A (en) * | 2018-12-27 | 2019-03-01 | 网易(杭州)网络有限公司 | A kind of multiple target recommended method, multiple target recommended models generation method and device |
CN110008404A (en) * | 2019-03-22 | 2019-07-12 | 成都理工大学 | Enigmatic language justice model optimization method based on the optimization of NAG momentum |
CN110008404B (en) * | 2019-03-22 | 2022-08-23 | 成都理工大学 | Latent semantic model optimization method based on NAG momentum optimization |
WO2021081962A1 (en) * | 2019-10-31 | 2021-05-06 | 华为技术有限公司 | Recommendation model training method, recommendation method, device, and computer-readable medium |
CN111159555A (en) * | 2019-12-30 | 2020-05-15 | 北京每日优鲜电子商务有限公司 | Commodity recommendation method, commodity recommendation device, server and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN105653683B (en) | 2020-10-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105653683A (en) | Personalized recommendation method and device | |
US20190026639A1 (en) | Neural architecture search for convolutional neural networks | |
CN104090912B (en) | Information-pushing method and device | |
CN101556553B (en) | Defect prediction method and system based on requirement change | |
CN103853786B (en) | The optimization method and system of database parameter | |
JP6964689B2 (en) | Sample weight setting method and device, electronic device | |
CN105260390B (en) | A kind of item recommendation method based on joint probability matrix decomposition towards group | |
EP4083857A1 (en) | Information prediction model training method and apparatus, information prediction method and apparatus, storage medium, and device | |
CN103455522A (en) | Recommendation method and system of application extension tools | |
CN109739768B (en) | Search engine evaluation method, device, equipment and readable storage medium | |
CN107862022A (en) | Cultural resource commending system | |
US20160019651A1 (en) | Method and apparatus of determining relative driving characteristics using vehicular participative sensing systems | |
WO2024067387A1 (en) | User portrait generation method based on characteristic variable scoring, device, vehicle, and storage medium | |
CN110472363B (en) | Surrounding rock deformation grade prediction method and system suitable for high-speed railway tunnel | |
CN105260171A (en) | Virtual item generation method and apparatus | |
CN116165988A (en) | Production quality control method and system for automobile center console | |
CN103177086A (en) | Method and device for ranking order through using searching result correspondingly | |
CN106373057A (en) | Network education-orientated poor learner identification method | |
CN107577736A (en) | A kind of file recommendation method and system based on BP neural network | |
CN104636407A (en) | Parameter choice training and search request processing method and device | |
CN109558544A (en) | Sort method and device, server and storage medium | |
CN105631535A (en) | Method and device for predicting scoring data | |
CN105740354A (en) | Adaptive potential Dirichlet model selection method and apparatus | |
CN105630880A (en) | Score data prediction method and apparatus | |
CN110516164A (en) | A kind of information recommendation method, device, equipment and storage medium |
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 |