Information recommendation method and device
Technical field
The present invention relates to areas of information technology, more particularly to a kind of information recommendation method and device.
Background technology
Today's society, the network information is vast as the open sea, and search engine is the weight that user searches the information oneself liked
One of means are wanted, but the premise of search engine is that user knows that he/her needs anything, if family oneself is used in conjunction all
Not knowing oneself can obtain what or want what when obtained, and search engine is helpless.Therefore
Intelligent recommendation algorithm and system are introduced, helps user to find for oneself needing/liking by the method for recommendation
Property content.Current proposed algorithm has content-based recommendation algorithm, Collaborative Filtering Recommendation Algorithm and mixed
Close proposed algorithm.
Content-based recommendation system main thought is according to the historical information (text such as evaluate, share, collected
Shelves) structuring user's content-preference document, recommended project and the similarity of user preference document are then calculated, will
Most like project recommendation is to user.The problem of this recommendation method, is that the information of recommendation is substantially ratio
It is more popular, it can only recommend to be interested in similar resource with user, it is impossible to find for user more new emerging
Interest.
Collaborative Filtering Recommendation Algorithm is broadly divided into two classes:
1:Collaborative filtering based on user:Scoring of the user to article is more similar, then they are to other
The scoring of article is also more similar, so as to find the arest neighbors with similar interests, is formed and recommended.
2:The similitude that collaborative filtering based on article scores different articles according to user estimates the use
Scoring of the family to some article, is recommended with this.
Above two collaborative filtering recommending method has problems with:
One is sparse sex chromosome mosaicism, when data volume is very big in commending system and the explicit score data of user is seldom
When, it is difficult to similitude is calculated, and can not be recommended.Two be cold start-up problem, when new article (i.e. fresh information)
When having just enter into system, it is evaluated without user, causes collaborative filtering can not recommend the resource.Three be to expand
Malleability problem, user and resource can quickly increase with the time, and the complexity of collaborative filtering and data
Amount is linear to be increased, and causes to calculate performance and efficiency under big data quantity, scalability is poor.
The content of the invention
In view of this, the embodiment of the present invention is expected to provide a kind of information recommendation method and device, at least partly solves
Certainly in the prior art information recommendation the problem of.
To reach above-mentioned purpose, the technical proposal of the invention is realized in this way:
First aspect of the embodiment of the present invention provides a kind of information recommendation method, and methods described includes:
Record the historical operation behavior of user;
According to the historical operation behavior, content-preference label is generated;
According to the content-preference label, it is determined that meeting the neighbour user of default similarity condition with targeted customer;
Obtain the historical operation behavioral data of the neighbour user;
According to the historical operation behavioral data of the proximal subscribers, information recommendation is carried out to the targeted customer.
It is described according to the historical operation behavior based on such scheme, content-preference label is generated, including:
According to the historical operation behavior, it is determined that the information that user accessed in the time of specifying;
Determine the weight for the information that user accessed;
According to the weight, preference vector is generated.
It is described according to the content-preference label based on such scheme, it is determined that meeting default phase with targeted customer
Like the neighbour user for the condition of spending, including:
Calculate the cosine value of the first preference vector of targeted customer and the second preference vector of neighbour's alternative user;
According to the cosine value, the neighbour that the default similarity condition is met with the targeted customer is determined
User.
Based on such scheme, methods described also includes:
According to the historical operation behavior, the liveness of user is determined;
Based on the liveness, neighbour's alternative user that Similarity Measure is carried out with the targeted customer is filtered out;
It is described according to the content-preference label, it is determined that meeting the neighbour of default similarity condition with targeted customer
User, including:
According to content-preference label, the similarity of alternative neighbour user and the targeted customer are calculated;
The neighbour user that default similarity condition is met with the targeted customer is determined according to the similarity.
It is described according to the historical operation behavior based on such scheme, the liveness of user is determined, including:
The liveness of each user in the counting statistics cycle in user's set;
The corresponding user of rejecting abnormalities liveness, and preset function relation is utilized, calculate the liveness of each user
Label;
Neighbour's alternative user is determined according to liveness label.
Second aspect of the embodiment of the present invention provides a kind of information recommending apparatus, and described device includes:
Recording unit, the historical operation behavior for recording user;
Generation unit, for according to the historical operation behavior, generating content-preference label;
First determining unit, for according to the content-preference label, it is determined that meeting default phase with targeted customer
Like the neighbour user for the condition of spending;
Acquiring unit, the historical operation behavioral data for obtaining the neighbour user;
Recommendation unit, for the historical operation behavioral data according to the proximal subscribers, to the targeted customer
Carry out information recommendation.
Based on such scheme, the generation unit, specifically for according to the historical operation behavior, it is determined that referring to
The information that the interior user that fixes time accessed;Determine the weight for the information that user accessed;According to the weight,
Generate preference vector.
Based on such scheme, first determining unit, for calculate targeted customer the first preference vector and
The cosine value of second preference vector of neighbour's alternative user;According to the cosine value, determine and the target
User meets the neighbour user of the default similarity condition.
Based on such scheme, described device also includes:
Second determining unit, for according to the historical operation behavior, determining the liveness of user;
Screening unit, Similarity Measure is carried out for based on the liveness, filtering out with the targeted customer
Neighbour's alternative user;
First determining unit, specifically for according to content-preference label, calculating alternative neighbour user and institute
State the similarity of targeted customer;Determined according to the similarity and meet default similarity with the targeted customer
The neighbour user of condition.
Based on such scheme, second determining unit gathers interior specifically for user in the counting statistics cycle
Each user liveness;The corresponding user of rejecting abnormalities liveness, and preset function relation is utilized, calculate
The liveness label of each user;Neighbour's alternative user is determined according to liveness label.
The embodiment of the present invention provides information recommendation method and device, will determine meet pre- with targeted customer first
If the neighbour user of similarity condition, historical operation behavioral data based on neighbour user enters for targeted customer
Row information is recommended, and can so excavate and new point of interest is excavated for targeted customer, while can realize some
The recommendation just released news, realizes the cold start-up of information recommendation, can lift information recommendation Efficiency and accuracy.
Brief description of the drawings
Fig. 1 is the schematic flow sheet of the first information recommendation method provided in an embodiment of the present invention;
Fig. 2 is the schematic flow sheet provided in an embodiment of the present invention for determining content-preference label;
Fig. 3 is the schematic flow sheet of second of information recommendation method provided in an embodiment of the present invention;
Fig. 4 is a kind of structural representation of information recommending apparatus provided in an embodiment of the present invention.
Embodiment
Technical scheme is done below in conjunction with Figure of description and specific embodiment and further explained in detail
State.
Embodiment one:
As shown in figure 1, the present embodiment provides a kind of information recommendation method, methods described includes:
Step S110:Record the historical operation behavior of user;
Step S120:According to the historical operation behavior, content-preference label is generated;
Step S130:According to the content-preference label, it is determined that meeting default similarity condition with targeted customer
Neighbour user;
Step S140:Obtain the historical operation behavioral data of the neighbour user;
Step S150:According to the historical operation behavioral data of the proximal subscribers, carried out to the targeted customer
Information recommendation.
Information recommendation method described in the present embodiment can be in the server applied to various information recommendations.For example,
Commercial product recommending information, the application downloaded (Application, App), the video searched for, picture, under
The network information of the storage such as the e-book of load or the goods links collected in a network.
The historical operation behavior of the user in the present embodiment can non-be login behavior and positive act.Here
Login behavior may include log-on webpage, login community, log in application etc. various register behaviors.Here
Positive act be user's active conduct, it may include using APP carry out social, shopping, collection, comment,
The various operations carried out based on user's active consciousness such as search, download or upload.The historical operation of these users
Behavior can reflect preference and the detest of user.Therefore in the present embodiment, in the step s 120 will according to
The historical operation behavior at family, generates content-preference label.
It will determine according to content-preference label and meet default similarity bar with targeted customer in step s 130
The neighbour user of part.The targeted customer is the user of information to be recommended in the present embodiment.Here neighbour
User is the user beyond the targeted customer, is the user larger with targeted customer's similarity.For example, with
Family A is targeted customer, and user B is targeted customer A neighbour user, represents that user B and user A expires
The foot default similarity condition.When carrying out the operation such as information browse, some information of user B concerns can
The information that energy and non-user A are currently also paid close attention to;But it is due to user A and user B content-preference label
Similarity is very big, illustrates to have closely similar preference and detest, some information of this when of user B concerns,
It is also likely to be some information of user A concerns and hobby.Therefore will be obtained in step S140 in the present embodiment
Take the historical operation behavioral data of the neighbour user.It will be grasped in step S150 according to the history of neighbour user
Make behavioral data to targeted customer's recommendation information.For example, some product informations that user B is paid close attention to, advertisement
The information recommendation that the user such as information or news A is not paid close attention to currently gives user A.Used in this way to target
Family carries out information recommendation, and the information that on the one hand can largely ensure recommended user A is that user A is willing to
The information that meaning receives and understood, while active that can be intelligent is to excavate more points of interest with targeted customer,
Expand the visual field and the information attention width of targeted customer.After a new information enters information recommendation system,
User B the operation such as comments on even without progress, can also be similar to targeted customer (user A) because of it
Property, user A is given by the new information recommendation, so as to realize that the information of some new entrance information systems can be by
Suitable recommended user is quickly found, the cold start-up of information is realized, information recommendation efficiency is improved.
Further, as shown in Fig. 2 the step S120 includes:
Step S121:According to the historical operation behavior, it is determined that the information that user accessed in the time of specifying;
Step S122:Determine the weight for the information that user accessed;
Step S123:According to the weight, preference vector is generated.
The specified time described in the present embodiment can be the current statistic time before a period of time in, such as one
The information that user accessed before week, one day or several hours.Here the information accessed may include to open
The information read, clicks on the information downloaded, the information for the product bought in shopping website and collected
The information such as document.The weight for the information that user accessed will be determined in step S122.Here weight is being carried out
When determining, weight calculation and assignment will be carried out according to preset algorithm.For example, according to user access information
The frequency, duration, user's action type and operation frequency etc. carry out weight assignment.Usual institute in the present embodiment
The span of weight is stated between 0 to 1, weight is bigger, represent that user is interested in the information.
For example, user accesses shopping website, the information to product A performs browse operation, to product B's
Information carries out browse operation and collection is operated, and the information to products C has carried out browsing and have purchased products C.
It is determined that above three information weight when, can according to action type, action type here includes browsing,
Collection and buy, browse, collect and buy in the present embodiment the positive influence value of weight is increased successively.
It is determined that during weight, while the information such as number of visits, purchase number of packages are further accounted for, to further determine that weight.
The weight can be determined according to the historical operation behavior to information in the present embodiment in a word.
Preference vector will be generated according to weight in the present embodiment.For example, the information accessed is sorted successively,
By the corresponding weight of the information accessed, vector is built according to the sequence of the information accessed, is thus constituted
The preference vector.For example, information A weight is a, information B weight is b, information C power
Weight is c;Information A, information B and information C are the information that user operated;If the sequence of these information
For information A, information B and information C, then the preference vector built is (a, b, c).Obvious user accesses
The weight for the information crossed as the different dimensions of preference vector value.
For example, it is assumed that the frequency of the information operated with user determines the weight for the information that user accessed,
This when, the weight can be:Wherein, sjThe letter accessed for j-th of user represented
The frequency of breath;The SiThe frequency of the information accessed for i-th of user;The I is user when specified
The sum of the interior information accessed or the I accessed information.Here j and i value is not less than 1
And no more than described I.What is be so not only simple has carried out weights determination, and the weights determined using the example
Span one be scheduled on 0 to 1.
Certainly when implementing, weighted value directly can also be equal to visitation frequency or the operation frequency,
When building the preference vector, it is being normalized, is utilizingAs preference vector each
The exploitation of dimension.
As further improvement of this embodiment, the step S130 may include:
Calculate the cosine value of the first preference vector of targeted customer and the second preference vector of neighbour's alternative user;
According to the cosine value, the neighbour that the default similarity condition is met with the targeted customer is determined
User.
In the present embodiment the neighbour user is determined using cosine.The first preference vector of targeted customer and
Second preference vector of neighbour's alternative user, if angle of two vectors in vector space is smaller, cosine
Value is bigger, therefore can symbolize the similarity between two users using cosine value in the present embodiment, in this reality
Apply that cosine value in example is bigger, then illustrate that the similarity of two users is bigger.In the present embodiment can be according to cosine
Value selects neighbour alternative user of the cosine value more than predetermined threshold value as formal neighbour user;Can also be right
The cosine value calculated is ranked up, and selects forward N positions of sorting as the neighbour user.Here
N is the integer not less than 1.
In specific calculating process, first preference vector and the second preference vector identical dimensional are corresponding
For same information or same type information.For example, the first preference vector and the second preference vector are including M
The M dimensional vectors of individual element.1st element of vector is value of the preference vector in the 1st dimension.At this
The first preference vector and the second preference vector are to same information or same type letter in m dimensions in embodiment
Breath carries out the weight of operation formation.When targeted customer or neighbour's alternative user to a certain information can a certain class do not have
When operation behavior is recorded, corresponding weights desirable 0.Here m is not less than 1 and no more than described M
Integer.The M is the integer not less than 2.
The calculating of the structure and cosine value of easy utilization preference vector is just capable of determining that institute in the present embodiment
Neighbour user is stated, the characteristics of with realizing easy.
When carrying out information recommendation according to the historical operation behavioral data of neighbour user in step S150, it may include
Filter out, the recommendation alternate information that targeted customer within a specified time has not visited and neighbour user accessed;
The historical operation behaviors such as the frequency, action type are operated to carry out weight tax according to user these recommendation alternate informations
Value, the recommendation information for recommending targeted customer is selected according to the weight of assignment.In the present embodiment here
The method of weight assignment may be referred in previous embodiment it is determined that method used when preference vector.Herein
Place, the weighted value of assignment is bigger, then it represents that neighbour user is bigger to the interest of the information, then recommends target
User, the probability welcome by targeted customer is higher, and pouplarity also can be higher.Therefore in the present embodiment
It can be ranked up based on the weight of assignment, the forward recommendation alternate information of selected and sorted is used as formal recommendation
Information, can also select recommendation alternate information of the weights value more than preset value as formal recommendation information.
Provide a kind of simple and efficient method for determining the neighbour user in the present embodiment in a word, have
The characteristics of realizing simple and accurate high.
As further improvement of this embodiment, in order to reduce the operand for determining neighbour user, in this implementation
In example, as shown in figure 3, methods described also includes:
Step S101:According to the historical operation behavior, the liveness of user is determined;
Step S102:Based on the liveness, filter out and carry out the near of Similarity Measure with the targeted customer
Adjacent alternative user;
The step S130 may include:According to content-preference label, alternative neighbour user and the target are calculated
The similarity of user;Determined according to the similarity and meet default similarity condition with the targeted customer
Neighbour user.
Neighbour's alternative user will be filtered out according to the liveness of user first in the present embodiment.For example, a certain
The online frequency of individual neighbour's alternative user recently is very low, and the information of access is also seldom, and it is used being defined as target
The effect of the recommendation information at family is smaller.For example, neighbour's alternative user once week in have accessed 1 website,
A piece of article is seen;Than targeted customer access information quantity it is few, thereby increases and it is possible to its information accessed all by
Targeted customer accessed, if being defined as neighbour user, it is possible to have no idea to recommend to targeted customer
More information;And occur calculating excessive phenomenon when being determined neighbour user.In the present embodiment
Liveness can be generated according to user's history operation behavior.Here liveness is higher, shows what user accessed
The quantity of information is read, it is possible to which the referential of the recommendation information provided for targeted customer is bigger.
Therefore it is alternative as the neighbour compared with large user that liveness will be selected in the step S102 of the present embodiment
User, on the one hand reduces amount of calculation, while providing information recommendation offer more ginsengs to be expected for targeted customer
Examine foundation.
Specifically, the step S102 may include:
The liveness of each user in the counting statistics cycle in user's set;
The corresponding user of rejecting abnormalities liveness, and preset function relation is utilized, calculate the liveness of each user
Label;
Neighbour's alternative user is determined according to liveness label.
Here proposition Showed Very Brisk degree may include to propose the maximum user of liveness and liveness in user's set
For 0 user, such as the liveness average and variance of all users in user's set, can be being calculated, will
Liveness is removed more than the user that the user of average and the exhausted angle value of P times of variance is considered as Showed Very Brisk degree.This
In P be positive number as many as 1, for example, value can be the values such as 2 or 3 or 4.
Further, liveness label is calculated;The liveness label can be a kind of to rejecting abnormalities liveness
The value that the normalization of the liveness of later remaining users is obtained.It is for instance possible to use
Wherein, feThe liveness accessed for e-th of remaining users represented;The frFor r-th of remaining users
Liveness;The R is the sum of the remaining users of rejecting abnormalities liveness.Here e and r value
All it is not less than 1 and no more than described R.Certainly it is full of a kind of method for determining liveness label here, below
The computational methods of another liveness label are also provided.
Liveness label=100* (minimum liveness in liveness-remaining users of r-th of remaining users)/
(liveness minimum in maximum liveness-remaining users in remaining users).When the minimum of remaining users is enlivened
Degree equal to remaining users maximum liveness when, the liveness label can value be 100.
The liveness label is bigger in the present embodiment, represents liveness of the correspondence user in measurement period
It is higher.Here measurement period can be time span set in advance, generally can be one before current time
In the section time.When progress neighbour's alternative user is determined, liveness label can be selected according to liveness label
The forward user of sequence also may be selected liveness label and be more than label threshold value as neighbour's alternative user
User is used as neighbour's alternative user.In a word, by the calculating of liveness, calculating can greatly be reduced
Amount, while guaranteeing to provide the reference frame of recommendation information for targeted customer.
Embodiment two:
As shown in figure 4, the present embodiment provides a kind of information recommending apparatus, described device includes:
Recording unit 110, the historical operation behavior for recording user;
Generation unit 120, for according to the historical operation behavior, generating content-preference label;
First determining unit 130, for according to the content-preference label, it is determined that meeting pre- with targeted customer
If the neighbour user of similarity condition;
Acquiring unit 140, the historical operation behavioral data for obtaining the neighbour user;
Recommendation unit 150, for the historical operation behavioral data according to the proximal subscribers, to the target
User carries out information recommendation.
Information recommending apparatus described in the present embodiment can be applied to various information recommendation servers or information recommendation
Device in platform.The recording unit 110 may correspond to storage medium or database, be able to record that user
Historical operation behavior.The generation unit 120, the first determining unit 130, acquiring unit 140 and recommendation
Unit 150 may both correspond to the processor or process circuit of server or service platform.The processor can be wrapped
Include the processing structures such as application processor, central processing unit, digital signal processor or programmable array.It is described
Process circuit may include application specific integrated circuit.These processing structures can be realized above-mentioned by the execution of appointment codes
Unit, can also realize aforesaid operations by the operation such as logical operation of circuit.Pass through neighbour in the present embodiment
The determination of user, the historical operation behavioral data based on neighbour user, can be fine to carry out information recommendation
Solve the problem of in the prior art can not be for new usage mining new point of interest.
In embodiment one and embodiment, the targeted customer and the neighbour user can be not know each other completely
User, the two possible users are probably different provinces, in addition country variant user, but be all net
When network carries out message reference, there is very big similitude to the operation behavior of the network information, therefore each other can be near each other
Adjacent user.
As further improvement of this embodiment, the generation unit 120, specifically for according to the history
Operation behavior, it is determined that the information that user accessed in the time of specifying;Determine the weight for the information that user accessed;
According to the weight, preference vector is generated.The mode and dress of the similarity of preference between two users of calculating
Putting structure has many kinds, is determined in the present embodiment by generating preference vector, simple special with calculating
The characteristics of point and apparatus structure are simple.Further, first determining unit 130, for calculating target
The cosine value of the first preference vector of user and the second preference vector of neighbour's alternative user;According to the cosine
Value, determines the neighbour user that the default similarity condition is met with the targeted customer.The meter of cosine value
A kind of calculating operational ton is few at last and calculates simple calculation, is calculated in the present embodiment using cosine value
The first determining unit 130 carry out neighbour user determination, with calculate it is simple and simple in construction the characteristics of.
In addition, described device also includes:Second determining unit, for according to the historical operation behavior, really
Determine the liveness of user;Screening unit, for based on the liveness, filtering out and entering with the targeted customer
Neighbour's alternative user of row Similarity Measure;First determining unit 130, specifically for inclined according to content
Good label, calculates the similarity of alternative neighbour user and the targeted customer;Determined according to the similarity
The neighbour user of default similarity condition is met with the targeted customer.
Described information recommendation apparatus also includes the second determining unit and screening unit in the present embodiment, the two
The corresponding hardware configuration of functional unit, hardware corresponding with the first determining unit, acquiring unit and recommendation unit
Structure is similar, and different is that the second determining unit is used for the liveness for determining user, and screening unit is used for root
Neighbour's alternative user is determined according to the liveness of user.Described information recommendation apparatus passes through neighbour in the present embodiment
The screening of alternative user, can reduce the amount of calculation determined in neighbour's user procedures, lift computational efficiency, and
More reference frames can be provided for recommendation information.Specifically such as, second determining unit is specific to use
In the liveness of each user in the counting statistics cycle in user's set;The corresponding user of rejecting abnormalities liveness,
And preset function relation is utilized, calculate the liveness label of each user;According to being determined liveness label
Neighbour's alternative user.In the present embodiment in order to ensure living caused by the abnormal access behaviors such as some corpses access
Jerk is abnormal, and second determining unit can also propose Showed Very Brisk degree in the present embodiment, to determine screening
Go out really to be able to carry out neighbour's alternative user that information recommendation provides reference frame for targeted customer, to carry
Rise the accuracy of information recommendation.
In a word, the calculating and neighbour that second determining unit can be by liveness in the present embodiment are alternative
The screening of user, on the one hand can reduce amount of calculation, on the other hand can be provided more to carry out information recommendation
Accurately reference frame.
Below in conjunction with any one above-mentioned embodiment, there is provided a specific example:
This example information recommends method to include:
The first step:All users are calculated with the liveness label of user.Liveness label is used for representing that user unites
Active degree in the meter cycle on website, the numerical value span of liveness label can be 0-100 integer,
Value is bigger to represent that user's active degree is higher.
Liveness tag computation step is as follows:
(1) login behavior and positive act, wherein positive act bag are divided into behavior of the user on website
Include use to product, collection, comment, search and the behavior such as download.
User in the website behavior of measurement period (initial value is 1 month) the interior all users in website is always logged in secondary
Number, the total positive act number of times of user are counted.
(2) in the counting statistics cycle (initial value be 1 month) all users liveness.
The total positive act number of times of the total login times+2* users of liveness=user.
(3) rejecting processing is carried out to liveness abnormal behaviour user.Processing method is as follows;
Reject the use that liveness is more than liveness threshold value upper (upper can be set, initial value upper=10000)
Family, calculates the average avg and standard deviation std of remaining all user activities, rejects liveness and is more than
(avg+3*std) user, user activity label is calculated just for remaining user.
(4) user activity label is calculated to all users after above-mentioned (3) are handled.
User activity label=100* (minimum liveness in liveness-all users)/(all users
Minimum liveness in middle maximum liveness-all users).
Liveness label rounds up, and span is 0-100 integer.When minimum liveness and maximum are lived
When jerk is equal, the liveness label value of user is 100.
Second step:Content-preference label is calculated to all users.Content-preference label is used for describing user to production
The preference of content representated by each all label of product.For the product of website, each product has one
The Product labelling list being pre-designed.Record is accessed by product of the user in measurement period on website, can
To obtain preference of the user to product, and then obtain preference of the user to the label of all products.In this example
The middle measurement period for calculating content-preference label can be with the foregoing measurement period length phase for calculating liveness label
Together, and original position is identical.Corresponding in previous embodiment it is believed that the specified time and measurement period pair
The duration answered is identical and initial time is identical.
Content-preference label takes the form of user content label preference weight vector, and user content label is inclined
Good vector is represented by each label weight composition of vector.Line label is entered to the label of product all in website,
Label label is corresponding with the component label of user content label preference vector.
Each representation in components user in user content label preference vector to the preference weight of corresponding label,
Each component takes decimal, and span is [0,1], and component of a vector value is bigger to represent user to corresponding mark
The preference tendency of label is bigger.The algorithm of content-preference label is as follows:
Pre-prepd historical operation behavioral data.
In measurement period (initial value of the measurement period can be 1 month), user accessed or used
Product list, product library, Product labelling table.
Using Do statement, count each user and accessed or used list of websites, to travel through each
The website that user accessed;Specifically such as, for user's in website user's lists.
Using Do statement, consumer products list under each website is counted, is accessed with each convenient user
The product crossed under website, specifically such as, " user uses product list to for products in.
Product labelling table is taken from product library.
Using Do statement, the label in each Product labelling list is compiled, specifically such as:
For labels in " Product labelling table "
The label counting=the tag computation+1
The user of user-label frequency vector;Here label frequency vector, is one kind of foregoing preference vector.
User-label frequency vector of user is normalized according to below equation
3rd step:The personalized recommendation based on customer relationship.
Application scenarios are when users log on, according to the product history access record of user, to calculate the interior of user
Hold the corresponding preference vector of preference label, by the similarity calculating method of preference vector, obtain login user
Most like user list, behavior is accessed by product of the nearest-neighbors user in measurement period, generation is worked as
The Products Show list of preceding login user.Recommendation based on customer relationship has following algorithm and step:
(1) list of targeted subscribers delimited.
The user list logined in website daily, and the recommendation list generation Time of Day of the user and work as
The difference on preceding date, which is more than, recommends expired threshold value, and threshold value initial value is set to 3.
(2) the related any active ues of targeted customer are obtained.The related any active ues generating algorithm of targeted customer is such as
Shown in lower:
Prepare the historical operation behavioral data of each user in advance.
(initial value of measurement period here can be 1 month), list of targeted subscribers are determined in measurement period.
Preparation product library, Product labelling table, product and the user list using the product.
Using Do statement neighbour's alternative user is determined for each targeted customer in list of targeted subscribers.Example
Such as, using for Do statements, for user in " list of targeted subscribers ".
Obtain product access list of the user in measurement period beyond targeted customer to product library.
Initialize related any active ues list USER_ACTIVE_ ID={ } of neighbour's alternative user.
Initialize casual user list USER_TEMP, using for Do statements, traversal targeted customer with
The product access list of outer other users;Specifically such as, for products in " product access list ", and obtaining
The user list for using product is taken, USER_PRO is denoted as.Wherein, USER_TEMP=
USER_ACTIVE+USER_PRO。
Reject the user repeated in USER_TEMP.
Using Do statement, such as Do statements travel through USER_TEMP user, specifically such as, for
User in USER_TEMP.
If the liveness label of USER_TEMP user is more than 50, corresponding user is inserted
Into USER_ACTIVE.After finally having circulated, the larger user of liveness is selected alternative as neighbour
User.
4th step:Determine the k nearest neighbor user of targeted customer.
Using the cosine similarity of targeted customer and the content-preference label of neighbour's alternative user as two users it
Between similarity measurement.Cosine similarity formula is as follows:
If the preference vector U of close-target user1=(x1,x2,…,xn), the preference vector of neighbour user
U2=(y1,y2,…,yn), then
It is determined that during neighbour user, specifically may include following steps:
1. prepare the historical operation behavioral data of targeted customer and neighbour's alternative user in advance.
2. obtain list of targeted subscribers, neighbour's alternative user list USER_ACTIVE ID.
3. Do statement is utilized, for example for Do statements, each mesh in traversal list of targeted subscribers
Mark user, specifically such as, for targeted customers u1In " list of targeted subscribers ".
4. initialize K=500.
5. initialized target user u1K nearest neighbor user list KNN_USER_ user u2ID={ }.
6. targeted customer u1Content-preference label U1=(x1,x2,…,xn)。
7. Do statement is utilized, for example for Do statements, in the neighbour's user list for traveling through targeted customer
The liveness of each neighbour user, for example, for neighbour users u2In " USER_ACTIVE use
Family u2ID”。
8. neighbour user u2Content-preference label U2=(y1,y2,…,yn)。
9. user u is calculated by cosine similarity formula1With user u2Similarity
10. " USER_ACTIVE_ user u2User presses and targeted customer u in ID "1Similarity descending is arranged
Row.
11.KNN_USER_ user u1ID=USER_ACTIVE_ user u2With first K use in ID
Family.
5th step:Result of calculation more than the Products Show list generating algorithm integrated use of targeted customer,
The Products Show list of targeted customer is calculated by the k nearest neighbor user list of targeted customer.Specifically
Step can be as follows:
1. pre-prepd neighbour user and the historical operation behavioral data of targeted customer.
2. alternative target user list, the k nearest neighbor user table of targeted customer and consumer products list.
3. initialising subscriber product scoring threshold value LIMIT=100.
4. the Products Show list products quantity k=100 of initialising subscriber.
5. traveling through each targeted customer, for example, targeted customer is traveled through using for Do statements, specifically such as,
For targeted customers u1In list of targeted subscribers.
6. initialized target user u1Products Show list REC_PROC_ user u1ID={ }.
7. travel through the neighbour user of each targeted customer, specifically such as, for user u2In targeted customers
u1K nearest neighbor user list.
8. if fruit product a is not in targeted customer u1Product list in, then according to functional relation calculate product
A scoring, the functional relation can be:Product a scoring=product a scoring+neighbour user u2
Access product a frequency.
9.REC_PROC_ user u1ID=scoring highests and the k product scored more than LIMIT,
This corresponding information of k product is to recommend the information of correspondence targeted customer.
When triggering recommendation action, the information recommendation list of the user, one are found into list of targeted subscribers
As for obtain information recommendation list in preceding 100 information as recommendation information, if in information recommendation list
Including information number less than 100, then regard all information in recommendation information list as recommendation information
Recommend corresponding targeted customer.
In several embodiments provided herein, it should be understood that disclosed apparatus and method,
It can realize by another way.Apparatus embodiments described above are only schematical, for example,
The division of the unit, only a kind of division of logic function, can there is other division when actually realizing
Mode, such as:Multiple units or component can be combined, or be desirably integrated into another system, or some spies
Levying to ignore, or does not perform.In addition, the coupling each other of shown or discussed each part,
Or direct-coupling or communication connection can be the INDIRECT COUPLINGs or logical of equipment or unit by some interfaces
Letter connection, can be electrical, machinery or other forms.
The above-mentioned unit illustrated as separating component can be or may not be it is physically separate, make
It can be for the part that unit is shown or may not be physical location, you can with positioned at a place,
It can also be distributed on multiple NEs;It can select therein part or all of according to the actual needs
Unit realizes the purpose of this embodiment scheme.
In addition, each functional unit in various embodiments of the present invention can be fully integrated into a processing module
In or each unit individually as a unit, can also two or more unit collection
Into in a unit;Above-mentioned integrated unit can both be realized in the form of hardware, it would however also be possible to employ
Hardware adds the form of SFU software functional unit to realize.
One of ordinary skill in the art will appreciate that:Realize all or part of step of above method embodiment
It can be completed by the related hardware of programmed instruction, it is computer-readable that foregoing program can be stored in one
Take in storage medium, the program upon execution, performs the step of including above method embodiment;And it is foregoing
Storage medium include:Movable storage device, read-only storage (ROM, Read-Only Memory),
Random access memory (RAM, Random Access Memory), magnetic disc or CD etc. are various
Can be with the medium of store program codes.
The foregoing is only a specific embodiment of the invention, but protection scope of the present invention is not limited to
In this, any one skilled in the art the invention discloses technical scope in, can be easily
Expect change or replacement, should all be included within the scope of the present invention.Therefore, protection of the invention
Scope should be based on the protection scope of the described claims.