CN107038169A - Object recommendation method and object recommendation equipment - Google Patents
Object recommendation method and object recommendation equipment Download PDFInfo
- Publication number
- CN107038169A CN107038169A CN201610078101.XA CN201610078101A CN107038169A CN 107038169 A CN107038169 A CN 107038169A CN 201610078101 A CN201610078101 A CN 201610078101A CN 107038169 A CN107038169 A CN 107038169A
- Authority
- CN
- China
- Prior art keywords
- user
- vector
- photo
- historical data
- sight spot
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
- G06F16/43—Querying
- G06F16/435—Filtering based on additional data, e.g. user or group profiles
Abstract
The application is related to a kind of object recommendation method and object recommendation equipment, and the object recommendation method includes：Characteristic vector extraction step, the characteristics of objects vector that the user characteristics vector of at least one corresponding user characteristics of expression represents corresponding objects feature is extracted with the historical data of object according to user, each in the case where historical data updates in the user characteristics vector updates independently of one another and each in the characteristics of objects vector updates independently of one another；Weight matrix generation step, usage history data represent weight matrix of the user characteristics relative to the mapping of characteristics of objects to generate；And, recommender score determines step, and the fraction of multiple candidate targets for being selected user is determined based on user characteristics vector, characteristics of objects vector weight matrix.
Description
Technical field
The application is generally related to based on historical data come to the object recommendation method of user's recommended
With object recommendation equipment.
Background technology
On Internet provide for user many object recommendation services, such as book, music,
Film and news are recommended.Recommendation is the historical data based on collection to learn the feature of user and right
The feature of elephant is carried out.Generally, the set of historical data can be expressed as user object matrix
Form (for example, Fig. 1).The matrix has a dimension and list object for user list
A dimension, each element relatively recorded with user and object.The content of the record
The scoring (such as 0-5 points) that can be user to object, or it is digitized access record,
Such as, 1：Like；0：Do not like；Or, 1：Bought；0：Do not bought, etc..With
Family feature can from user object matrix a line learning corresponding with the user to.Similarly,
Characteristics of objects can from a user object matrix row learning corresponding with the object to.
It is Deta sparseness to recommending the related typical problem of application.As explained above, root
Learn user characteristics and characteristics of objects according to user object matrix.In actual recommendation, Yong Huhe
The quantity of object becomes increasing.But, it is very difficult to collect each element of the matrix
Historical data.Therefore, increasing element is changed into " unknown " in matrix, such as institute in Fig. 1
Show, this causes learnt user characteristics and characteristics of objects to become unstable and unreliable.
Especially in recommending scenery spot application is shot, historical data (access of such as user to sight spot
Record etc.) it is related to the position height at user and sight spot, this frequently results in more serious data
It is openness.
Disclosed in U.S. Patent Publication US 2010/0030764A1 " using pass through it is neighbouring and
The commending system of explicit and implicit feedback the combined filtration of the combination of latency model
(Recommender System Utilizing Collaborative Filtering Combing
Explicit and Implicit Feedback With Both Neighborhood and Latent
Factor Models)”.The document includes following recommendation method：The user of expression historical data
Object matrix decomposition is user's factor matrix and object factor matrix.In user's factor matrix,
Each user is respectively by a line factor vector representation.In object factor matrix, each object point
Not by a row factor vector representation.Pass through the point of correspondence user's factor vector object factor vector
The long-pending recommender score to calculate the object for specific user.Calculating for specific user's
After the recommender score of each object, to be recommended at least to determine according to the arrangement of these fractions
One object.
The document can solve the Deta sparseness in object recommendation application.However, in the document
In, user's factor matrix and object factor matrix are according to the history number as training data
According to off-line learning.It means that user object matrix is once have updated, user's factor matrix
All needed with least one in object factor matrix by re -training.If for example, we are new
Know that user have accessed some object, then we need the whole user's factor matrix of re -training or
Whole object factor matrix.Re -training process is the time-consuming operation of height, therefore prior art
The method proposed in document is infeasible for the recommendation application of frequent updating user object matrix
's.Moreover, recommendation results are unaccountable, it is difficult to illustrate user's factor square in this method
The implication of each element of battle array and object factor matrix.
For example, in recommending scenery spot application is shot, shooting recommending scenery spot and user and the position at sight spot
Height correlation is put, this often results in Deta sparseness and user sight spot matrix is in actual applications
Frequent updating.Using the recommendation method of the document be accomplished by frequently re -training and therefore with
Such dynamic data source can not be compatible.Finally, recommendation results are also to be difficult to what is explained, this meaning
Taste, which, to be difficult to the reason for explanation sight spot is recommended to user and is analyzed and improved.
The content of the invention
The first purpose of the application avoids the whole user's factor of re -training when historical data updates
Matrix or whole object factor matrix, so that not only accurate but also efficiently carry out object recommendation.In addition,
It is expected that object recommendation result can be explained.
The application's relates in one aspect to a kind of object recommendation method, and the object recommendation method includes：
Characteristic vector extraction step, according to user's and object multiple historical datas extract at least one
Individual at least one expression corresponding objects feature of the user characteristics vector for representing correspondence user characteristics
Characteristics of objects vector, it is each in the user characteristics vector in the case that historical data updates
It is individual all independently of one another update and the characteristics of objects vector in each independently of one another
Update；Weight matrix generation step, usage history data come generate represent user characteristics relative to
The weight matrix of the mapping of characteristics of objects；And, recommender score determines step, special based on user
Vector, characteristics of objects vector weight matrix is levied to determine multiple candidates for being selected user
The fraction of object.
The another aspect of the application is related to a kind of object recommendation equipment, including：Characteristic vector is extracted
Device, is configured as according to user's and object multiple historical datas extract at least one table
Show the object of at least one expression corresponding objects feature of the user characteristics vector of correspondence user characteristics
Characteristic vector, each in the case where historical data updates in the user characteristics vector
Update independently of one another and in the characteristics of objects vector each independently of one another more
Newly；Weight matrix generating means, are configured with historical data to generate expression user characteristics
Relative to the weight matrix of the mapping of characteristics of objects；And, recommender score determining device, by with
It is set to based on user characteristics vector, characteristics of objects vector weight matrix to determine to be used for be selected
The fraction of multiple candidate targets of user.
Therefore, according to each side of the application, enable to when updating historical data, it is to avoid
The whole user's factor matrix of re -training or whole object factor matrix, so that not only accurate but also efficient
Ground carries out object recommendation.In addition, object recommendation result is also to explain.
Brief description of the drawings
With reference to specific embodiment, and referring to the drawings, to above and other mesh of the application
And advantage be further described.In the accompanying drawings, identical or corresponding technical characteristic or portion
Part will be represented using identical or corresponding reference.
Fig. 1 is the user object matrix for being shown as historical data；
Fig. 2 shows the flow chart of the recommending scenery spot method according to one embodiment of the application；
Fig. 3 shows that the characteristic vector in the recommending scenery spot method according to embodiments herein is carried
Take the flow chart of process；
Extracted user characteristics vector sight spot characteristic vector is shown respectively in Fig. 4 A and Fig. 4 B
Example；
The weight matrix generated according to the not be the same as Example of the application is shown respectively in Fig. 5 A and 5B；
The determination recommending scenery spot of the not be the same as Example according to the application is shown respectively in Fig. 6 A and 6B
The schematic diagram of fraction；
Fig. 7 is the exemplary configuration for showing the recommending scenery spot equipment according to embodiments herein
Block diagram；
Fig. 8 shows the recommendation for being used to implement recommending scenery spot method according to embodiments herein
The schematic diagram of system；And
Fig. 9 illustrates the ability to implement the hardware configuration of the computer system of embodiments herein
Block diagram.
Embodiment
The one exemplary embodiment of the application is described hereinafter in connection with accompanying drawing.In order to clear
For the sake of Chu and simplicity, all features of embodiment are not described in the description.However, should
Understand, must make many specific to embodiment during implementing to embodiment
Set, to realize the objectives of developer, for example, meeting related to equipment and business
Those restrictive conditions, and these restrictive conditions may with embodiment difference and have
Changed.In addition, it also should be appreciated that, although development is likely to be extremely complex and time-consuming
, but for the those skilled in the art for having benefited from present disclosure, this development is only
Only it is routine task.
Herein, it should be noted that in order to avoid having obscured the application because of unnecessary details,
Illustrate only in the accompanying drawings with according at least to the closely related process step of the scheme of the application and
/ or device structure, and eliminate the other details little with the application relation.
It will be illustrated below exemplified by shooting recommending scenery spot application, because as mentioned before
User sight spot matrix frequent updating.It is to be understood that the example and non-limiting, the application can also be fitted
For in the application of other object recommendations, such as books to recommend application, music to recommend application, film
Recommend application etc., this will be illustrated later.Next, being described in the following order.
1. the flow chart of recommending scenery spot method
2. characteristic vector extraction process
2-1 user characteristics vector extraction process
2-2 sight spots characteristic vector extraction process
3. weight matrix generating process
4. recommender score determination process
5. the block diagram of recommending scenery spot equipment
6. the advantage of recommending scenery spot method and apparatus
7. realize the commending system of object recommendation method
8. other embodiments
9. the computer equipment to implement the present processes and equipment
1. the flow chart of recommending scenery spot method
First, reference picture 2 is described according to the recommending scenery spot method of one embodiment of the application
Flow chart.In the method, it is following three using the user sight spot matrix decomposition as historical data
Individual matrix：User characteristics matrix, weight matrix, and sight spot eigenmatrix.Then, it is based on
These three matrixes calculate the recommender score at sight spot, that is, the unknown with user sight spot matrix
Plain related fraction.
Step S101 is user characteristics vector extraction step, wherein according to user's and sight spot is more
Individual historical data come extract at least one represent correspondence user characteristics user characteristics vector, going through
History data update in the case of in the user characteristics vector each independently of one another more
Newly.Step S101 is triggered by the user's recommendation request being described later on.
The exemplary forms of the historical data of user and the historical data at sight spot can be as shown in Figure 1
Such user sight spot matrix, wherein have recorded access situation of the user to sight spot, each row is visual
To represent the historical data of user characteristics and respectively arranging to can be considered the history number for representing sight spot feature
According to.Historical data is not limited to the exemplary forms, such as another form can be that different user is clapped
The collection of photographs at the different sight spots taken the photograph, or be different user on the internet to different sight spots
Evaluation score etc..
Therefore, the historical data of user can come from for example including at least one in following item：
A) sight spot that user accesses；
B) sight spot of user's scoring, and evaluation score；
C) the sight spot photo that user shoots；
D) the sight spot photo that user shoots, photo has photo tag；And
E) the sight spot photo of user's scoring, and evaluation score.
On the other hand, the historical data at sight spot can come from for example including at least one in following item
It is individual：
A) user at sight spot is accessed；
B) user scored sight spot, and evaluation score；
The photo shot when c) shooting sight spot；And
The photo shot when d) shooting sight spot, photo has photo tag.
Photo tag is, for example, time, place and photographer's name of photograph taking etc..
In one example, the historical data for shooting sight spot can be by by using photo files
Middle storage can be that the positional information of GPS information is clustered to obtain to the photo of collection.
Herein, user characteristics makes each user be distinguished from each other, and can use going through for correspondence user
History data obtain representing the user characteristics vector of user characteristics based on various features extracting method.
As previously described, in recommending scenery spot application, historical data such as user sight spot matrix is frequent more
Newly.The application cause in the user characteristics vector extracted in the case each can
Update independently of one another, this avoids the whole user's factor matrix of renewal and therefore will greatly improved
The efficiency of recommending scenery spot.The process that characteristic vector is extracted will be described in detail later.
Step S102 is sight spot characteristic vector extraction step, wherein according to the multiple of user and sight spot
Historical data come extract at least one represent correspondence sight spot feature sight spot characteristic vector, in history
Each in the case that data update in the sight spot characteristic vector updates independently of one another.
The step is similar with step S101, it is, for example, possible to use the historical data base at correspondence sight spot
Obtain representing the sight spot characteristic vector of sight spot feature in various features extracting method.Similarly,
The application also cause in extracted sight spot characteristic vector each can independently of one another more
Newly.This avoids the effect for updating whole sight spot factor matrix and therefore greatly improving recommending scenery spot
Rate.
Step S103 is weight matrix generation step, wherein using user's history data and sight spot
Historical data represents weight of the user characteristics vector relative to the mapping of sight spot characteristic vector to generate
Matrix.
As it was previously stated, the user's history data or sight spot history number in any source can be based respectively on
According to extracting user characteristics vector or sight spot characteristic vector using appropriate feature extracting method.For
Make it possible to know the relation between the characteristic vector of user characteristics vector sight spot so as to more accurate
Really determine recommender score of the sight spot for user, it would be desirable to which generation represents that user characteristics is relative
In the weight matrix of the mapping of sight spot feature.For example, representing user characteristics and with row with list
Show sight spot feature, then the element (weight) of the weight matrix can be regarded as with this feature
User likes the degree of this feature in sight spot.Therefore the recommendation results of the method proposed are can
To explain the reason for sight spot is recommended to user.The process for generating weight matrix will in detail below
Description.
Step S104 is that recommender score determines step, wherein special based on user characteristics vector, sight spot
Vector weight matrix is levied to determine the fraction of multiple candidate sights for selecting user.
After being extracted user characteristics vector sight spot characteristic vector and generating weight matrix,
Recommender score of the sight spot to selected user can be just calculated, this essentially defines user sight spot
The content of the unknown in matrix.Because the forward sight spot of arrangement generally more preferably matches user, institute
Recommended with that should select to arrange forward sight spot.Recommend the quantity at sight spot can be according to multiple
Factor determines, total threshold value of such as recommender score, personal recommender score or artificially defined
Recommended amount.
Therefore, it can be avoided in historical data according to the recommending scenery spot method of embodiments herein
The whole user's factor matrix of re -training or whole sight spot factor matrix during renewal, so as to improve
The efficiency of recommending scenery spot.In addition, recommending scenery spot result is also to explain.By under reading
These advantages may be better understood in the detailed description in face.
2. characteristic vector extraction process
Fig. 3 shows the flow chart of the characteristic vector extraction process according to embodiments herein.
In Fig. 3, step S201 is information receiving step, wherein receive recommendation request from user and
Specifying information.Step S202 is data acquisition step, wherein obtaining newest user and sight spot is gone through
History data.Step S203 is feature definition step, wherein, pre-defining will be according to historical data
The feature of extraction.Step S204 is that characteristic vector updates step, wherein being carried based on historical data
Take and update user characteristics vector or sight spot characteristic vector.This will be further elaborated with later.
2-1 user characteristics vector extraction process
The process for extracting user characteristics vector is specifically described referring next to Fig. 3 flow chart.
First, the shooting recommending scenery spot request sent from user is received.
The request such as communication module from user equipment.The request preferably includes at least two
Category information：ID, is general index in the database of storage user characteristics vector；With with
The position that family is specified, user just seeks to shoot sight spot at the position.
The mode of specified location in user can be explicit or implicit.In explicit way, use
Family can by, for example, input position title, manually interaction map on drawing area or to
Its equipment sends voice command etc. and carrys out specified location.In implicit, position can be by user
Equipment is automatically determined.For example, the current location of user can be by the GPS module in user equipment
There is provided.In addition, position radius can expect the region for carrying out recommending sight spot wherein with instruction user.
Position radius can be specified by user, or be automatically determined according to the contextual information of user.
Next, obtaining the latest history data at user and sight spot.These historical datas can be come
The information of the record at sight spot, or the various letters that user provides are accessed on user from internet
Breath source, such as one group photo.The various sources of historical data are as it was noted above, no longer weigh herein
It is multiple.
Then, the user characteristics to be extracted according to historical data is pre-defined.For example in history number
According to the record for being one group of sight spot photo sight spot that either user accessed or commenting undue sight spot
In the case of, user characteristics can be the scene information in these photos, such as " street ", " dynamic
Thing ", " plant ", " sea " etc..Thus user characteristics vector can include following item
At least one of in：
A) quantity of the photo shot by relative users of predetermined scene is belonged to；
B) percentage of the photo shot by relative users of predetermined scene is belonged to；
C) quantity of the photo scored by relative users of predetermined scene is belonged to；
D) percentage of the photo scored by relative users of predetermined scene is belonged to；
E) quantity at the sight spot accessed by relative users of predetermined scene is belonged to；
F) percentage at the sight spot accessed by relative users of predetermined scene is belonged to；
G) quantity at the sight spot scored by relative users of predetermined scene is belonged to；And
H) percentage at the sight spot scored by relative users of predetermined scene is belonged to.
Or, user characteristics can be the statistical information that sight spot is accessed by relative users, the system
Meter information can be type, access frequency and the duration and user characteristics arrow that user accesses sight spot
Amount is made up of its corresponding statistical value.
Finally, extracted and more based on user profile, historical data and defined user characteristics
New user characteristics vector.
In one example, user characteristics vector is scene tag histogram herein.Extracting method
It is as follows：Using the N number of photo (N for the user's upload for proposing recommendation request>0)；Shine N number of
The photo of each in piece is iterated, so as to be known based on corresponding picture material using such as object
Other method recognizes photo scene information, and the photo scene information recognized of each photo is pre-
Define one or more of scene tag set (feature) scene tag；Whole N number of photos
The summation of the scene tag information recognized be calculated to form calculating vector；Then with N pairs
Calculate vector to be normalized as histogram, histogrammic each element is correspondence scene tag
Percentage.Then scene tag histogram is stored as the user characteristics in customer data base
Vector, as shown in Figure 4 A.When user uploads one group of new photo, although user's history number
According to changing, but independently can only update the scene tag histogram of correspondence user without
Influence the characteristic vector of other users.
In other examples, as the replacement of the one group of photo uploaded to user, historical data can
Since the scoring of free relative users sight spot.Scene information is being defined as the situation of user characteristics
Under, user characteristics vector can similarly be extracted with aforementioned exemplary.
In another example, each scape for showing that the user accesses can be counted based on historical data
Type, access frequency and duration etc. of point as the user user characteristics vector.
2-2 sight spots characteristic vector extraction process
The process is similar with user characteristics vector extraction process.Difference is only described below.
First, if specifying position and radius in user profile, according to position and radius come
Choose candidate sight.If herein without specified location and radius, step S201 can be ignored.
Therefore, the specified location in user with known latitude and longitude, specified location in user are given
It can be computed with the distance between each sight spot.Calculating and comparing from user's specific bit
After the distance for putting each sight spot, it is only located at the sight spot within radius and is considered candidate sight.
Then, it is the sight spot or comment that one group of sight spot photo or user accessed in historical data
In the case of the record at undue sight spot, sight spot feature can also be the scene letter in these photos
Breath, " street ", " animal ", " plant ", " sea " etc..But sight spot feature
Vector and user characteristics vector are different and including at least one in following item：
A) quantity of the photo for belonging to predetermined scene shot to corresponding sight spot；
B) percentage of the photo for belonging to predetermined scene shot to corresponding sight spot；
C) quantity with the photo for specifying mean opinion score shot to corresponding sight spot；And
D) percentage with the photo for specifying mean opinion score shot to corresponding sight spot.
Or, sight spot feature can be the statistical information for each user for accessing corresponding sight spot, described
Statistical information can be age, sex and education degree and the sight spot characteristic vector for accessing user
It is made up of its corresponding statistical value.
Finally, extracted and more based on user profile, historical data and defined sight spot feature
New sight spot characteristic vector.
In one example, sight spot characteristic vector is scene tag histogram.Extracting method is as follows：
Using M photo (M in the range of targeted candidate sight>0)；To M photo
In each photo be iterated so that recognized based on corresponding picture material photo scene believe
Breath.The photo scene information recognized of each photo is predefined scene tag set (feature)
One or more of scene tag；Calculate the scene tag information recognized of M photo
Summation is to form calculating vector；Then it is normalized with M to calculating vector as histogram,
Histogrammic each element is the percentage of correspondence scene tag.Then scene tag histogram quilt
Storage is as the sight spot characteristic vector in scene data storehouse, as shown in Figure 4 B.When the sight spot
In the range of when uploading one group of new photo, can be independent although sight spot historical data changes
Ground updates the scene tag histogram at correspondence sight spot without influenceing the characteristic vector at other sight spots.
In other examples, as the replacement to one group of photo in the range of candidate sight, go through
History data can be the user and evaluation score scored sight spot.It is defined as evaluation score
In the case of the feature of sight spot, sight spot characteristic vector can similarly be extracted with aforementioned exemplary.
In another example, it can be counted based on historical data and draw each use for accessing the sight spot
Age, sex and the education degree at family etc. as the candidate sight sight spot characteristic vector.
3. weight matrix generating process
Weight matrix is, for example, square formation, wherein each element is real number.The element of weight matrix is
User characteristics is transferred to the weight of sight spot feature.Approximately, weight is bigger, special with certain class
The user levied and the matching degree at the sight spot with certain category feature are bigger.
In one example, weight matrix is trained and generated by usage history off-line data.
If stem algorithm can apply in training process, such as stochastic gradient descent SGD algorithms, at this
In algorithm：
1) input learning rate factor gamma and normalization factor λ；
2) random initializtion weight matrix S；
3) loss function is defined：
Wherein,
ru,iIt is true value of the user u for sight spot i；
It is recommender scores of the user u for sight spot i
puIt is user u characteristic vector；
qiIt is sight spot i characteristic vector；
4) for user u and sight spot i, gradient is calculated：
5) corresponding S is updated：
6) for each user u in training data and each sight spot i repeat steps 4) and 5),
Until convergence；
7) training completes and exports the weight matrix S of study.
The weight matrix S now obtained can be together with the characteristic vector of user characteristics vector sight spot
The data of the unknown at least corresponded in the matrix of user sight spot are calculated by multiplying.
Fig. 5 A show the weight matrix generated according to the example.Each element representation tool in the matrix
The user for having certain category feature likes the degree at the sight spot with certain category feature.It is used as historical data
Training data is more sufficient, then the weight matrix generated will update user's with subjects history data
When keep more stablize.Therefore this is by as the key factor for adapting to dynamic data source.
In another example, as the replacement of off-line training method, history number is also based on
Weight matrix is generated according to using statistical method.Certain of each element representation user in the matrix
Each joint probability sum in the joint probability of certain category feature at category feature and sight spot, matrix should be waited
In 1.Fig. 5 B show the weight matrix generated according to the example, wherein such as 0.11 represents
User with " (liking shooting) plant " feature shot the scape containing " street " feature
The probability of point.In statistics, in the case of the collection of photographs that given each user shoots, calculate
Go out the joint probability distribution of each user characteristics and each sight spot feature.Statistical method is not limited to herein
The joint probability enumerated, other manner, such as conditional probability can also be used as needed.
4. recommender score determination process
User characteristics vector, sight spot characteristic vector and after generating weight matrix are being extracted,
The recommender score of the candidate sight for the user is assured that, and is then recommended as needed
One or more sight spots.
In one example, for extracting the history number of user characteristics vector sight spot characteristic vector
According to source be used for generate historical data used in weight matrix source it is identical.It is specific and
Speech, the photograph collection that user provides is used as the user's scape that can therefrom obtain as shown in Figure 1
The historical data of dot matrix.User characteristics vector sight spot characteristic vector be based on scene information from
What photograph collection was extracted, as previously described.In addition, weight matrix be based on the user sight spot matrix,
User characteristics vector and sight spot characteristic vector come what off-line training was obtained, as previously described.
Therefore, the recommendation of each candidate sight is calculated respectively for same user u with equation below
Fraction：
That is the characteristic vector p based on user uu, weight matrix S and sight spot i characteristic vector
qiTo calculate recommender scores of the sight spot i to user uAs shown in Figure 6A.
Recommender score indicates the matching journey of the characteristic vector of correspondence user and the characteristic vector at sight spot
Degree.For specific user, if sight spot has the recommender score bigger than another sight spot, the scape
Point has the bigger possibility being easily accepted by a user and is consequently adapted to be recommended.In this example,
Because the source of historical data is identical, the correlation of data and the stability of recommendation results are more
It is good.
In a modification, above-mentioned weight matrix can not be off-line training generation.As
Substitute, it can use statistical method to generate based on the collection of photographs as historical data,
As previously described.Using such statistics generation weight matrix as shown in Figure 6B that
Recommender score is similarly determined in sample
In another example, and unlike preceding example, for extracting user characteristics vector
With the source of the historical data of sight spot characteristic vector and for generating history used in weight matrix
The source of data is different.Specifically, this kind of data in sight spot that user accesses are used as history number
According to user sight spot matrix as shown in Figure 1 can be directly obtained from such historical data.
On the other hand, user characteristics vector sight spot characteristic vector is that the photograph collection provided from user is based on
What scene information was extracted, as previously described.In addition, weight matrix is to be based on the user sight spot square
Battle array, user characteristics vector and sight spot characteristic vector come what off-line training was obtained, as previously described.
Analogously it is possible to determine recommender score as shown in figure 6 a
In the modification similar with aforementioned variant of the example, weight matrix can not be by offline
Training generation.Alternatively, the weight matrix can be based on the photograph as historical data
What piece set was generated using statistical method, wherein certain category feature of each element representation user with
The joint probability of certain category feature at sight spot.Analogously it is possible to determine to push away as shown in fig. 6b
Recommend fraction
Therefore, need not be in historical data more according to the recommending scenery spot method of embodiments herein
The whole user's factor matrix of re -training or whole sight spot factor matrix when new.It is additionally, since power
Element in weight matrix has specific implication, therefore is easy to explain recommendation results.
5. the block diagram of recommending scenery spot equipment
The recommending scenery spot equipment of one embodiment according to the application is described referring next to Fig. 7
The block diagram of 700 exemplary configuration.The recommending scenery spot equipment 700 includes：Characteristic vector is extracted
Device 701, is configured as according to user's and sight spot multiple historical datas extract at least one
At least one expression correspondence sight spot feature of the individual user characteristics vector for representing correspondence user characteristics
Sight spot characteristic vector, it is each in the user characteristics vector in the case that historical data updates
It is individual all independently of one another update and the sight spot characteristic vector in each independently of one another
Update；Weight matrix generating means 702, are configured with historical data and represent to use to generate
Weight matrix of the family feature relative to the mapping of sight spot feature；And recommender score determining device
703, it is configured as determining to use based on user characteristics vector, sight spot characteristic vector and weight matrix
In the fraction of multiple candidate sights of selected user.
Apparatus above 701-703 can be configured to perform foregoing recommending scenery spot method
Step S101-S104.
Preferably, characteristic vector extraction element 701 can also include：Information receiver 704,
It is configured as receiving recommendation request and specifying information from user；Data acquisition facility 705,
It is configured as obtaining newest user's history data and sight spot historical data；Characterizing definition device
706, it is configured as the pre-defined feature to be extracted according to historical data；And vector more new clothes
707 are put, is configured as that characteristic vector is extracted and updated based on historical data.
Arrangement described above is the example for implementing recommending scenery spot method described in this application
Property and/or preferred device.These devices can be hardware cell (such as central processing unit (figure
CPU 901 in 9), field programmable gate array, digital signal processor, application specific integrated circuit
Or computer etc.) and/or software service (such as computer-readable program).It is not detailed above
Ground describes the device for implementing each step.As long as performing the step of some is handled however, having,
Just can be with useful in the corresponding device for implementing same processing (by hardware and/or software implementation).
The technology limited by all combinations of described step and device corresponding with these steps
Scheme is all included in present disclosure, as long as these technical schemes that they are constituted
It is complete and applicable.
In addition, the said equipment being made up of various devices can be incorporated into as functional module it is all
In such as hardware unit of computer etc.In addition to these functional modules, computer certainly may be used
With with other hardware or software part.
6. the advantage of recommending scenery spot method and apparatus
According to the recommending scenery spot method and apparatus of embodiments herein user sight spot matrix decomposition
For three below matrix：The user characteristics matrix being made up of user characteristics vector, weight matrix with
And the sight spot eigenmatrix being made up of sight spot characteristic vector.Then, corresponding user characteristics is utilized
Vector, sight spot characteristic vector and weight matrix calculate the recommended hour at the sight spot for relative users
Number.
First, user characteristics and sight spot are characterized in be described using such as photo scene histogram.
Because the photo sight spot histogram is a class low dimensional feature, it is possible to reduce position correlation
Sparse sex chromosome mosaicism in recommendation.Secondly, user characteristics matrix and sight spot eigenmatrix difference
Formed by user picture scene histogram and sight spot photo scene histogram, be independently of user's scape
Dot matrix.Weight matrix is according to historical data off-line training or statistical learning, as going through
The training data of history data is more sufficient, then the weight matrix acquired will be in user and sight spot history number
More stablize according to holding when updating.When updating the historical data relevant with user or sight spot, only
Update corresponding user characteristics vector or sight spot characteristic vector, other user characteristics vectors or sight spot
Characteristic vector will not change.Dependent on such local re -training mechanism, the method proposed
Realize the compatibility with dynamic data source.Finally, because weight matrix can be considered mapping
Matrix, each element therein is that special characteristic and the special characteristic at sight spot of user matches journey
Degree, so the reason for sight spot is recommended to user can be explained in the method proposed.
7. realize the commending system of recommending scenery spot method
Fig. 8 shows the recommendation system for being used to implement recommending scenery spot method according to embodiments herein
The schematic diagram of system.Commending system include customer data base, scene data storehouse, image analysis module,
Off-line training module, user's extractor, sight spot filter and recommended engine.Customer data base
User characteristics vector (as shown in Figure 4 A) and sight spot Characteristic Vectors are stored respectively with scene data storehouse
(as shown in Figure 4 B), these vectors are that (characteristic vector extracts dress by image analysis module to amount
Put) generated using the view data of user's upload.According to images above data, off-line training
Weight matrix (such as Fig. 5 A that module (weight matrix generating means) study is used by recommended engine
It is shown).When commending system receives the request for shooting recommending scenery spot from specific user, use
Family extractor is called to select corresponding user characteristics vector, and scape from customer data base
Coordinate filter is called to select multiple sight spot features with the position according to sight spot from scene data storehouse
Vector simultaneously forms candidate sight set.Recommended engine (recommender score determining device) uses user
Characteristic vector (such as pu), sight spot characteristic vector (such as qi) and weight matrix (such as S) count
Calculate the recommender score of each candidate sight (such as).Higher recommender score refers to user and right
Answer the preferable matching between sight spot.Candidate sight is sorted in lists according to recommender score, and
And several candidate sights in the top are output as recommendation results.Finally, recommendation results quilt
Return to user equipment and presented to user.
8. other embodiments
Although embodiment above concentrates on recommending scenery spot application, the application is clearly not limited to this
Application is planted, but various object recommendation fields common on such as network can be widely used for.
For example, recommend field in books, and it is similar with recommending scenery spot field, by the way that user is accessed
The types (or scoring to books) of books user characteristics vector is extracted as feature simultaneously
And the education degree of the users of access books is extracted book feature vector as feature, and so
Carry out off-line training using user characteristics vector book feature vector and user's books matrix afterwards to weigh
Weight matrix, it may be determined that recommender score of some books for selected user.
In another example, it is similar with recommending scenery spot field in video recommendations field, by the way that user is visited
(or scorings to video) such as the types and viewing duration of the video asked is extracted as feature
User characteristics vector and it assign age, sex and the education degree of the user of access video as spy
Levy to extract video features vector, and then using user characteristics vector video features vector with
And user video matrix carrys out off-line training weight matrix, it is possible to determine some videos for selected
The recommender score of user.
9. the computer equipment to implement the present processes and equipment
Fig. 9 is the hardware configuration for illustrating the ability to implement the computer system of embodiments herein
Block diagram.
As shown in Figure 9, computer system includes the processing list connected via system bus 904
Member 901, read-only storage 902, random access memory 903 and input/output interface 905,
And connected via input/output interface 905 input block 906, output unit 907, deposit
Storage unit 908, communication unit 909 and driver 910.Program can be previously recorded in as meter
The ROM (read-only storage) 902 or memory cell 908 of built-in recording medium in calculation machine
In.Or, program can store (record) in removable media 911.Herein,
Removable media 911 include such as floppy disk, CD-ROM (compact disk read-only storage),
MO (magneto-optic) disk, DVD (digital versatile disc), disk, semiconductor memory etc..
Input block 906 be used for input user request, be configured with keyboard, mouse, touch-screen,
Microphone etc..In addition, output unit 907 is configured with LCD (liquid crystal display), loudspeaker
Deng.
Communication unit 909 may, for example, be wireless communication unit, including at least one transceiver mould
Block and locating module.Transceiver module is used to send recommendation request and from long-range to remote server
Server receives object recommendation result.Locating module is, for example, GPS module 912, for obtaining
The position of user.
Memory cell 908 or the storage elemental users of ROM 902 information, historical data, interest
Etc..RAM 903 can store provisional status information and results of intermediate calculations.
In addition, except by driver 910 from above-mentioned removable media 911 program
It is installed to outside the configuration of computer, can be by communication network or radio network download program
To computer with built-in storage unit 908.In other words, can be for example with wireless parties
Formula is led to by the satellite for digital satellite broadcasting from download point to computer or in a wired fashion
The network of LAN (LAN) or internet etc. is crossed to computer transmission procedure.
If by user's manipulation to input block 906 etc., via input/output interface 905
Order is have input to computer system, then CPU 901 is performed in ROM 902 according to order
The program of storage.Or, CPU 901 is carried in the program stored in memory cell 908
With configuration processor on RAM 903.
Therefore, CPU 901 is performed according to some processing of above-mentioned flow chart or passed through
The processing that the configuration of above-mentioned block diagram is performed.Next, if it is necessary, then CPU 901
The result of processing is allowed for example to export, pass through from output unit 907 by input/output interface 905
Transmitted, recorded in memory cell 908 by communication unit 909 etc..
In addition, program can be performed by a computer (processor).In addition, program can be with
Handled in a distributed fashion by multiple computers.Furthermore it is possible to program transportation to long-range meter
Calculation machine is performed.
Computer system shown in Fig. 9 be merely illustrative and be never intended to the application,
It is applied or purposes carries out any limitation.Computer system shown in Fig. 9, which can be incorporated in, appoints
What embodiment, can be as the processing system in equipment as stand-alone computer, or also, can
To remove one or more components from computer system as needed, one can also be added to
Individual or more component is used as additional component.
The present processes and system can be implemented in many ways.For example, can pass through
Software, hardware, firmware or its any combinations implement the present processes and system.It is above-mentioned
The order of method and step be merely illustrative, it is specific that the present processes step is not limited to the above
The order of description, unless otherwise clearly stated.In addition, in certain embodiments, this
Application can also be implemented as recording program in the recording medium, and it includes being used to realize basis
The machine readable instructions of the present processes.Thus, the application, which also covers storage, to be used to realize root
According to the recording medium of the program of the present processes.
Although some embodiments of the application are described in detail by example, this
Art personnel should be appreciated that the model that above-mentioned example is merely illustrative without limiting the application
Enclose.It should be appreciated by those skilled in the art that above-described embodiment can be changed without departing from this Shen
Scope and spirit please.Scope of the present application is limited by appended claim.
Claims (14)
1. a kind of object recommendation method, it is characterised in that including：
Characteristic vector extraction step, according to user's and object multiple historical datas come extract to
At least one expression corresponding objects of the user characteristics vector of few expression correspondence user characteristics are special
The characteristics of objects vector levied, in the user characteristics vector in the case that historical data updates
Each update independently of one another and in the characteristics of objects vector each is only each other
On the spot update；
Weight matrix generation step, usage history data represent that user characteristics vector is relative to generate
In the weight matrix of the mapping of characteristics of objects vector；And
Recommender score determines step, based on user characteristics vector, characteristics of objects vector weight square
Battle array determines the fractions of multiple candidate targets for being selected user.
2. object recommendation method according to claim 1, wherein, for extracting user
The source of the historical data of characteristic vector and characteristics of objects vector is different from being used to generate weight matrix
The source of used historical data.
3. object recommendation method according to claim 1, wherein, it is special for extracting user
The source of historical data of vector characteristics of objects vector is levied with being used for generating weight matrix
Historical data source it is identical.
4. the object recommendation method according to one of claim 1-3, wherein, use use
Family and object historical data statistically generates weight matrix.
5. the object recommendation method according to one of claim 1-3, wherein, use use
Family and object historical data carrys out off-line training generation weight matrix.
6. object recommendation method according to claim 5, wherein, off-line training generation weight
The step of matrix, includes：
Using user's and object historical data and the use that is obtained based on the historical data
Family characteristic vector and characteristics of objects vector carry out off-line training weight matrix.
7. the object recommendation method according to one of claim 1-3, wherein, user's
The historical data comes from least one in following item：
A) object that user accesses；
B) object of user's scoring, and evaluation score；
C) photo that user shoots；
D) photo that user shoots, photo has photo tag；And
E) photo of user's scoring, and evaluation score.
8. the object recommendation method according to one of claim 1-3, wherein, object
The historical data comes from least one in following item：
A) user of object is accessed；
B) to the user of object score, and evaluation score；
C) photo shot during reference object；And
D) photo shot during reference object, photo has photo tag.
9. object recommendation method according to claim 7, wherein, the user characteristics
Vector includes at least one in following item：
A) quantity of the photo shot by relative users of predetermined scene is belonged to；
B) percentage of the photo shot by relative users of predetermined scene is belonged to；
C) quantity of the photo scored by relative users of predetermined scene is belonged to；
D) percentage of the photo scored by relative users of predetermined scene is belonged to；
E) quantity of the object accessed by relative users of predetermined scene is belonged to；
F) percentage of the object accessed by relative users of predetermined scene is belonged to；
G) quantity of the object scored by relative users of predetermined scene is belonged to；
H) percentage of the object scored by relative users of predetermined scene is belonged to；And
I) statistical information of object is accessed by relative users, the statistical information can be user
Access type, access frequency and the duration of object.
10. object recommendation method according to claim 8, the characteristics of objects vector bag
Include at least one in following item：
A) quantity of the photo for belonging to predetermined scene shot to corresponding object；
B) percentage of the photo for belonging to predetermined scene shot to corresponding object；
C) quantity with the photo for specifying mean opinion score shot to corresponding object；
D) percentage with the photo for specifying mean opinion score shot to corresponding object；With
And
E) statistical information of the user of corresponding object is accessed, the statistical information can access to use
Age, sex and the education degree at family.
11. the object recommendation method according to one of claim 1-3, wherein, Duo Gehou
Selecting object is determined according to the position of selected user.
12. the object recommendation method according to one of claim 1-3, wherein, it is described right
As if shoot sight spot.
13. object recommendation method according to claim 12, wherein, shoot sight spot
The historical data can by by using stored in photo files can be GPS information position
Confidence ceases to be clustered to obtain to the photo of collection.
14. a kind of object recommendation equipment, it is characterised in that including：
Feature deriving means, are configured as according to user's and object multiple historical datas are carried
At least one is taken to represent that the user characteristics vector of correspondence user characteristics represents corresponding objects feature
Characteristics of objects vector, it is each in the user characteristics vector in the case that historical data updates
It is individual all independently of one another update and the characteristics of objects vector in each independently of one another
Update；
Weight matrix generating means, are configured with historical data to generate expression user characteristics
Relative to the weight matrix of the mapping of characteristics of objects；And
Recommender score determining device, is configured as being based on user characteristics vector, characteristics of objects vector
The fraction of multiple candidate targets for being selected user is determined with weight matrix.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610078101.XA CN107038169B (en) | 2016-02-04 | 2016-02-04 | Object recommendation method and object recommendation device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610078101.XA CN107038169B (en) | 2016-02-04 | 2016-02-04 | Object recommendation method and object recommendation device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107038169A true CN107038169A (en) | 2017-08-11 |
| CN107038169B CN107038169B (en) | 2021-03-12 |
Family
ID=59532145
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610078101.XA Active CN107038169B (en) | 2016-02-04 | 2016-02-04 | Object recommendation method and object recommendation device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107038169B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108121211A (en) * | 2017-12-12 | 2018-06-05 | 美的智慧家居科技有限公司 | Control method, server and the computer readable storage medium of home appliance |
| CN110688581A (en) * | 2019-10-30 | 2020-01-14 | 南京领行科技股份有限公司 | Information real-time pushing method and device, computing equipment and medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101076032A (en) * | 2007-02-15 | 2007-11-21 | 腾讯科技（深圳）有限公司 | Method and system for transmitting information |
| CN101383942A (en) * | 2008-08-01 | 2009-03-11 | 深圳市天威视讯股份有限公司 | Hidden customer characteristic extracting method and television program recommendation method and system |
| US20100030764A1 (en) * | 2008-07-30 | 2010-02-04 | At&T Corp. | Recommender System Utilizing Collaborative Filtering Combining Explicit and Implicit Feedback with both Neighborhood and Latent Factor Models |
| CN102103634A (en) * | 2009-12-22 | 2011-06-22 | 索尼公司 | Information processing apparatus and method |
| CN102651011A (en) * | 2011-02-27 | 2012-08-29 | 祁勇 | Method and system for determining document characteristic and user characteristic |
| CN102999540A (en) * | 2011-09-10 | 2013-03-27 | 祁勇 | Method and system for determining user features on Internet |
| CN104572707A (en) * | 2013-10-18 | 2015-04-29 | 北京卓易讯畅科技有限公司 | Preferable object information providing method and device |
-
2016
- 2016-02-04 CN CN201610078101.XA patent/CN107038169B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101076032A (en) * | 2007-02-15 | 2007-11-21 | 腾讯科技（深圳）有限公司 | Method and system for transmitting information |
| US20100030764A1 (en) * | 2008-07-30 | 2010-02-04 | At&T Corp. | Recommender System Utilizing Collaborative Filtering Combining Explicit and Implicit Feedback with both Neighborhood and Latent Factor Models |
| CN101383942A (en) * | 2008-08-01 | 2009-03-11 | 深圳市天威视讯股份有限公司 | Hidden customer characteristic extracting method and television program recommendation method and system |
| CN102103634A (en) * | 2009-12-22 | 2011-06-22 | 索尼公司 | Information processing apparatus and method |
| CN102651011A (en) * | 2011-02-27 | 2012-08-29 | 祁勇 | Method and system for determining document characteristic and user characteristic |
| CN102999540A (en) * | 2011-09-10 | 2013-03-27 | 祁勇 | Method and system for determining user features on Internet |
| CN104572707A (en) * | 2013-10-18 | 2015-04-29 | 北京卓易讯畅科技有限公司 | Preferable object information providing method and device |
Non-Patent Citations (2)
| Title |
|---|
| QIAN WANG ET AL.: ""Collaborative filtering recommendation algorithm based on hybrid user model"", 《INTERNATIONAL CONFERENCE ON FUZZY SYSTEMS AND KNOWLEDGE DISCOVERY》 * |
| 谭志国 等: ""基于面积比矩阵的加权特征向量点匹配算法"", 《国防科技大学学报》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108121211A (en) * | 2017-12-12 | 2018-06-05 | 美的智慧家居科技有限公司 | Control method, server and the computer readable storage medium of home appliance |
| CN108121211B (en) * | 2017-12-12 | 2021-09-14 | 美的智慧家居科技有限公司 | Control method of household appliance, server and computer readable storage medium |
| CN110688581A (en) * | 2019-10-30 | 2020-01-14 | 南京领行科技股份有限公司 | Information real-time pushing method and device, computing equipment and medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107038169B (en) | 2021-03-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11087413B2 (en) | Identity mapping between commerce customers and social media users | |
| Leonelli et al. | Data and society: A critical introduction | |
| US10726050B2 (en) | Assigning social networking system users to households | |
| US8935263B1 (en) | Generating rankings of reputation scores in reputation systems | |
| US20150161529A1 (en) | Identifying Related Events for Event Ticket Network Systems | |
| CN107924401A (en) | Video recommendations based on video title | |
| CN102591868B (en) | System and method for automatic generation of photograph guide | |
| EP2518679A1 (en) | Method and system fo recommending geo-tagged items | |
| US20150262081A1 (en) | Estimating reputation scores in reputation systems | |
| CN108664658B (en) | Collaborative filtering video recommendation method considering dynamic change of user preference | |
| AU2015336942A1 (en) | Learning with transformed data | |
| JP6261547B2 (en) | Determination device, determination method, and determination program | |
| CN109635206B (en) | Personalized recommendation method and system integrating implicit feedback and user social status | |
| US9081825B1 (en) | Querying of reputation scores in reputation systems | |
| KR102203253B1 (en) | Rating augmentation and item recommendation method and system based on generative adversarial networks | |
| CN108897750B (en) | Personalized place recommendation method and device integrating multiple contextual information | |
| KR20210082106A (en) | Method for Providing Real Estate Estimated Real Transaction Price Calculation Service Using Decision Tree Based Time Series Trend Prediction Learning Model | |
| CN112199609A (en) | POI recommendation system and method for sensing space-time semantic interval under self-attention | |
| CN110647683A (en) | Information recommendation method and device | |
| Du et al. | Pcard: Personalized restaurants recommendation from card payment transaction records | |
| CN116204714A (en) | Recommendation method, recommendation device, electronic equipment and storage medium | |
| CN113656699B (en) | User feature vector determining method, related equipment and medium | |
| CN107038169A (en) | Object recommendation method and object recommendation equipment | |
| KR101743169B1 (en) | System and Method for Searching Missing Family Using Facial Information and Storage Medium of Executing The Program | |
| CN112699311A (en) | Information pushing method, storage medium and electronic equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |