CN107203558A - Object recommendation method and apparatus, recommendation information treating method and apparatus - Google Patents
Object recommendation method and apparatus, recommendation information treating method and apparatus Download PDFInfo
- Publication number
- CN107203558A CN107203558A CN201610154736.3A CN201610154736A CN107203558A CN 107203558 A CN107203558 A CN 107203558A CN 201610154736 A CN201610154736 A CN 201610154736A CN 107203558 A CN107203558 A CN 107203558A
- Authority
- CN
- China
- Prior art keywords
- score
- probability
- accumulated value
- target
- value
- 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
- G06F16/951—Indexing; Web crawling techniques
Abstract
The present invention relates to a kind of object recommendation method and apparatus, recommendation information treating method and apparatus, the object recommendation method includes:Obtain the raw score of each object in candidate target set;Obtain existing object and recommend default target score probability cumulative distribution under scene;According to the target score probability cumulative distribution, the raw score isotonic regression is mapped as target score;Corresponding object is selected from the candidate target set according to the target score;Recommend the object picked out.Object recommendation method and apparatus that the present invention is provided, recommendation information treating method and apparatus so that recommendation results are accurate.
Description
Technical field
The present invention relates to field of computer technology, at more particularly to a kind of object recommendation method and apparatus, recommendation information
Manage method and apparatus.
Background technology
Recommended is a kind of important way to user's transmission information, the object such as application program, Yong Huhuo of recommendation
Person's commodity etc., the recommendation information for recommending these objects can be sent to user terminal, recommendation information is rung by user terminal
Should.Current object recommendation method, usually calculates each object in candidate target set using specific scoring algorithm
Score value, a part of object is selected from candidate target set as recommendation results further according to score value.
However, current object recommendation method, the scoring algorithm used when calculating score value is different, each calculated is right
The score value of elephant is also different, and the distribution situation of the various score values calculated is also different, object recommendation under so different scoring algorithms
Probability has different, causes recommendation results not precisely, it is necessary to improve.
The content of the invention
Based on this, it is necessary to for current object recommendation method when calculating score value the scoring algorithm that uses it is different lead
Causing the problem of recommendation results are not accurate, there is provided a kind of object recommendation method and apparatus, recommendation information treating method and apparatus.
A kind of object recommendation method, methods described includes:
Obtain the raw score of each object in candidate target set;
Obtain existing object and recommend default target score probability cumulative distribution under scene;
According to the target score probability cumulative distribution, the raw score isotonic regression is mapped as target score;
Corresponding object is selected from the candidate target set according to the target score;
Recommend the object picked out.
A kind of object recommendation device, described device includes:
Raw score acquisition module, the raw score for obtaining each object in candidate target set;
Target score probability cumulative distribution acquisition module, default target score under scene is recommended for obtaining existing object
Probability cumulative distribution;
Mapping block, for according to the target score probability cumulative distribution, the raw score isotonic regression to be mapped
For target score;
Choosing module, for selecting corresponding object from the candidate target set according to the target score;
Recommending module, for the object for recommending to pick out.
A kind of recommendation information processing method, methods described includes:
Receive the recommendation information of recommended;The recommendation information is picked out according to the target score of the object,
The target score is after obtaining raw score, to be recommended scoring to the object and according to existing object to preset under scene
Target score probability cumulative distribution the raw score isotonic regression mapped;
The recommendation information is ranked up according to the target score of the object;
The recommendation information is shown according to clooating sequence.
A kind of recommendation information processing unit, described device includes:
Recommendation information receiving module, the recommendation information for receiving recommended;The recommendation information is according to described right
What the target score of elephant was picked out, the target score be the object is scored and after obtaining raw score, according to
Default target score probability cumulative distribution maps the raw score isotonic regression under existing object recommendation scene;
Order module, is ranked up for the target score according to the object to the recommendation information;
Recommendation information display module, for showing the recommendation information according to clooating sequence.
Above-mentioned object recommendation method and apparatus, recommendation information treating method and apparatus, carry out scoring by object and obtain corresponding
Original score value, according still further to existing object recommend scene under default target score probability cumulative distribution original score value is protected
Sequence Hui-Hui calendar is to target score, so as to recommend corresponding object according to target score.Even if using different scoring algorithms
To obtain raw score, can order preserving map recommend to the target score for meeting target score probability cumulative distribution, so pair
As the probability distribution of target score used is accurate, the recommendation results further according to target score recommended are also accurate
's.
Brief description of the drawings
Fig. 1 is the applied environment figure of commending system in one embodiment;
Fig. 2 is the internal structure schematic diagram of terminal in one embodiment;
Fig. 3 is the internal structure schematic diagram of server in one embodiment;
Fig. 4 is the schematic flow sheet of object recommendation method in one embodiment;
Fig. 5 be one embodiment according to target score probability cumulative distribution, raw score isotonic regression is mapped as mesh
The schematic flow sheet for the step of marking score value;
Fig. 6 be one embodiment according to target score probability cumulative distribution, raw score probability accumulated value is mapped as
The schematic flow sheet of the step of target score probability accumulated value matched with raw score probability accumulated value;
Fig. 7 is the raw score probability cumulative distribution curve synoptic diagram that is obtained by GBDT algorithms in one embodiment;
Fig. 8 be one embodiment in map acquisition raw score probability cumulative distribution curve synoptic diagram;
The schematic flow sheet of Fig. 9 is generates target score probability cumulative distribution in one embodiment the step of;
The schematic flow sheet of Figure 10 is generates target score probability cumulative distribution in another embodiment the step of;
Figure 11 is the target score probability after the target score probability density curve before one embodiment alignment and calibration
Density curve contrast schematic diagram;
Figure 12 is the schematic flow sheet of recommendation information processing method in one embodiment;
Figure 13 is the structured flowchart of object recommendation device in one embodiment;
Figure 14 is the structured flowchart of mapping block in one embodiment;
Figure 15 is the structured flowchart of object recommendation device in another embodiment;
Figure 16 is the structured flowchart of object recommendation device in further embodiment;
Figure 17 is the structured flowchart of recommendation information processing unit in one embodiment.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.
As shown in figure 1, in one embodiment there is provided a kind of commending system, including terminal 110 and server 120, its
Middle terminal 110 includes mobile terminal, mobile unit and personal computer etc., and mobile terminal includes mobile phone, tablet personal computer, intelligent hand
At least one of table or personal digital assistant (PDA) etc..Server 120 can be that independent physical server can also be
Physical server cluster.
As shown in Fig. 2 there is provided a kind of terminal 110, including the processing connected by system bus in one embodiment
Device, non-volatile memory medium, built-in storage, network interface, display screen and input equipment.Wherein processor has computing function
The function of being worked with control terminal 110, the processor is configured as performing a kind of recommendation information processing method.Non-volatile memories
Medium includes at least one of magnetic storage medium, optical storage media and flash memory type storage medium.Non-volatile memory medium is deposited
Operating system is contained, be also stored with recommendation information processing unit.The recommendation information processing unit is used to realize a kind of recommendation information
Processing method.Network interface is used for by network connection to server 120.Input equipment can be physical button or with display
The overlapping touch control layer of screen, touch control layer and display screen constitute touch screen.
As shown in figure 3, there is provided a kind of server 120, including the place connected by system bus in one embodiment
Manage device, non-volatile memory medium, built-in storage and network interface.Wherein processor has computing function and control server
The function of 120 work, the processor is configured as performing a kind of object recommendation method.Non-volatile memory medium includes magnetic storage
At least one of medium, optical storage media and flash memory type storage medium.Non-volatile memory medium is stored with operating system, also
Be stored with object recommendation device.The object recommendation device is used to realize a kind of object recommendation method.Network interface is used to pass through net
Network is connected to terminal 110.
As shown in figure 4, in one embodiment there is provided a kind of object recommendation method, the present embodiment is applied in this way
Server 120 in above-mentioned Fig. 3 is illustrated.This method specifically includes following steps:
Step 402, the raw score of each object in candidate target set is obtained.
Wherein, candidate target set is the set that some recommendable objects are constituted, and the object in candidate target set can
To be represented with corresponding object identity, object identity is then the character string with uniqueness.Object in candidate target set
Can be specifically user, commodity or application program etc., commodity such as virtual goodses or physical commodity etc..
Service implement body can obtain the object properties of multiple dimensions of each object in candidate target set, so that according to multiple
The object properties of dimension are simultaneously scored by the first scoring algorithm, obtain raw score.Such as server can obtain commodity
The item property of multiple dimensions such as price, knock-down price, number of clicks, clicking rate, conclusion of the business number of times and conversion ratio, and to each dimension
The item property of degree carries out scoring rear weight summation and obtains raw score.Raw score is relative to following target scores
Speech, is that mapping obtains the score value before target score.
In one embodiment, service implement body can extract the feature of the sample in training set, using machine learning algorithm
Training obtains forecast model, so as to extract in candidate target set the feature of each object and be input to forecast model, so as to predict
Go out the raw score of each object in candidate target set.Wherein forecast model is that the Feature Mapping of object that will need to predict score value is
The function of the score value of prediction.The machine learning algorithm of use such as GBDT (Gradient Boosting Decision Tree,
Gradient lifted decision tree) algorithm, CART (Classification and Regression Trees, Taxonomy and distribution) calculate
Method or algorithm of support vector machine etc..
Step 404, obtain existing object and recommend default target score probability cumulative distribution under scene.
Specifically, server is previously stored with the corresponding target score probability cumulative distribution of various object recommendation scenes, from
And after raw score is obtained, obtain the target score probability cumulative distribution under existing object recommendation scene.Wherein object recommendation
Scene refers to the scene for recommending certain types of object, such as the scene of recommended user, the scene of Recommendations or recommendation should
Scene with program etc..
Target score is to use the machine learning algorithms different from the current machine learning algorithm for calculating raw score to learn
The accurate score value obtained, target score probability cumulative distribution is the data for representing target score probability accumulated value distribution situation.
One target score one target score probability accumulated value of correspondence, the target score probability accumulated value is represented less than equal to corresponding
All target scores of target score account for the ratio of all target scores in target score set.Target score be it is discrete, generally
Rate cumulative distribution can be represented with discrete functional relationship, can also be represented with mode is enumerated.
Illustrate, it is assumed that the target score in target score set is followed successively by 3,5,9 and 10, corresponding mesh according to ascending order
Mark score value is followed successively by 0.2,0.4,0.2 and 0.2 relative to the probability of target score set, then corresponding target score probability accumulation
Value is followed successively by 0.2,0.6,0.8 and 1.
Step 406, according to target score probability cumulative distribution, raw score isotonic regression is mapped as target score.
Specifically, order-preserving refers to that raw score is mapped as after target score, and corresponding target score retains raw score
Magnitude relationship.Such as raw score a and b are each mapped to target score A and B, and a < b, then A < B.Recurrence refers to original
Beginning score value passes through the probability distribution of the target score obtained after certain mapping, target corresponding with target score probability cumulative distribution
Score value probability distribution is basically identical.
In one embodiment, service implement body can be by raw score according to ascending order or descending sort, and then according to row
Sequence order is sequentially adjusted in each raw score for corresponding target score so that target score meets target score probability iterated integral
Cloth.
Step 408, corresponding object is selected from candidate target set according to target score.
Specifically, server can set suitable screening conditions as needed, according to the screening conditions from candidate target collection
Object is selected in conjunction, the screening conditions enter row constraint according to target score.Screening conditions are such as selected with more than preset value
The object of target score, either the corresponding object of target score of the maximum predetermined number of selection target score value or selection with
Corresponding object of target score of score value matching generated at random etc..
Step 410, the object picked out is recommended.
Specifically, server can generate the recommendation information for the object for recommending to pick out, and recommendation information is pushed into terminal.
Recommendation information includes object identity, can also include object factory text and/or description picture, can also include each picking out
The corresponding target score of object.
Above-mentioned object recommendation method, carries out scoring by object and obtains corresponding original score value, pushed away according still further to existing object
Recommend default target score probability cumulative distribution under scene and original score value isotonic regression is mapped to target score, so that according to
Target score recommends corresponding object., can order preserving map even if obtaining raw score using different scoring algorithms
To the target score for meeting target score probability cumulative distribution, the probability distribution of the target score used in such recommended is accurate
True, the recommendation results further according to target score recommended are also accurately.
As shown in figure 5, in one embodiment, step 406 specifically includes following steps:
Step 502, by raw score according to ascending sort, each raw score is obtained according to ranking results corresponding original point
It is worth probability accumulated value.
Wherein, raw score is referred to according to ascending sort arrange raw score according to the order of raw score from small to large
Sequence.Specifically, server is after by raw score according to ascending sort, and the raw score in traversal ranking results obtains current
The corresponding probable value of raw score of traversal, thus by the ranking results since the first to the raw score currently traveled through
All probable values are added, and obtain the corresponding raw score probability accumulated value of raw score currently traveled through.
For example, being shown in Table one:
The raw score of ascending sort | Raw score probability accumulated value |
0.001 | 0.0013 |
0.002 | 0.0018 |
0.004 | 0.0019 |
0.006 | 0.0020 |
... | ... |
0.32 | 1 |
First row is the raw score according to ascending sort in above-mentioned table one, and secondary series is then raw score pair in first row
The raw score probability accumulated value answered, with raw score probability accumulated value be per the raw score in a line it is one-to-one, the
The raw score that the raw score probability accumulated value of two row represents less than raw score corresponding equal in first row accounts for all original
The accounting of score value.
Step 504, according to target score probability cumulative distribution, raw score probability accumulated value is mapped as and raw score
The target score probability accumulated value of probability accumulated value matching.
Specifically, server can travel through raw score probability accumulated value according to target score probability cumulative distribution, for traversal
Raw score probability accumulated value find equal or close target score probability accumulated value, and then the raw score of traversal is reflected
The target score probability accumulated value for penetrating to search out.Wherein target score probability accumulated value is matched with raw score probability accumulated value
Refer to that target score probability accumulated value is met close to condition with raw score probability accumulated value.Server can also be by interpolation
The target score probability accumulated value of the raw score probability accumulated value distribution matching of traversal.
Step 506, raw score is mapped as target score corresponding with target score probability accumulated value.
Specifically, target score probability accumulated value after ascending sort by target score according to generating, and target score is general
Rate accumulated value has one-to-one relationship with target score, and server just may be used after mapping obtains target score probability accumulated value
So that raw score to be mapped as to the target score corresponding to the target score probability accumulated value.
In the present embodiment, raw score is after ascending sort, and the raw score probability accumulated value of acquisition can reflect
The raw score for going out to be less than or equal to current raw score accounts for the ratio of all raw scores, and target score probability accumulated value is also anti-
The target score mirrored less than or equal to corresponding target score accounts for the ratio of all target scores.So raw score probability is tired out
After product value and the matching of target score probability accumulated value, corresponding raw score is mapped as corresponding target score, just to reflect
Target score after penetrating can both retain the magnitude relationship of raw score, and default target score probability distribution can be met again,
And amount of calculation is small.
As shown in fig. 6, in one embodiment, step 504 specifically includes following steps:
Step 602, according to target score probability cumulative distribution, the target point equal with raw score probability accumulated value is searched
It is worth probability accumulated value;If finding, step 604 is performed;If not finding, step 606 is performed.
Target score probability cumulative distribution represents the probability that various target scores occur, and service implement body can be by target score
Various target score probability accumulated values represented by probability cumulative distribution are sorted in ascending order, and are searched in the ranking results and are equal to original
The target score probability accumulated value of beginning score value probability accumulated value.
Step 604, raw score probability accumulated value is mapped as to the target score probability accumulated value found.
Specifically, if the target score probability accumulated value equal with raw score probability accumulated value can be found, this is illustrated
Position of the corresponding target score of target score probability accumulated value found in whole target score set, with corresponding original
Position of the corresponding raw score of beginning score value probability accumulated value in whole raw score set is identical, by the raw score
Probability accumulated value is mapped directly into the target score probability accumulated value found, and isotonic regression can be achieved.
For example, being shown in Table two:
Raw score | Raw score probability accumulated value | Target score probability accumulated value | Target score |
0.001 | 0.0013 | 0.0012 | 0.00001 |
0.002 | 0.0018 | 0.0015 | 0.00002 |
0.004 | 0.0019 | 0.0015 | 0.00002 |
0.006 | 0.0020 | 0.0020 | 0.00004 |
... | ... | ... | ... |
0.32 | 1 | 1 | 0.05432 |
Tertial target score probability accumulated value is obtained by lookup table mode in above-mentioned table two, is specifically according to target point
It is worth probability cumulative distribution and searches the maximum target score value probability accumulated value for being less than or equal to corresponding raw score probability accumulated value.The
Target score probability accumulated value during target score in four row is arranged with the 3rd is one-to-one relationship.
For example, in above-mentioned table two, when raw score probability accumulated value is 0.0020, can find equal to 0.0020
Target score probability accumulated value, then corresponding raw score 0.006 is mapped as corresponding target score 0.00004.
Step 606, raw score probability accumulated value is mapped as meeting close to condition with raw score probability accumulated value
Target score probability accumulated value.
Specifically, server can not find the target score probability accumulated value equal with raw score probability accumulated value
When, can be according to target score probability cumulative distribution, then search tired with the close target score probability of raw score probability accumulated value
Product value as mapping target.
In one embodiment, step 606 includes:Raw score probability accumulated value is mapped as with raw score probability to tire out
The immediate target score probability accumulated value of product value.Specifically, server can search the gap with raw score probability accumulated value
Minimum target score probability accumulated value, and raw score probability accumulated value is mapped as the target score probability accumulated value.
In one embodiment, step 606 includes:Raw score probability accumulated value is mapped as to be less than raw score probability
The maximum target score value probability accumulated value of accumulated value.
For example, in above-mentioned table two, when raw score probability accumulated value is 0.0018, in the absence of the mesh equal to 0.0018
Score value probability accumulated value is marked, but the maximum target score value probability accumulated value that can be found less than 0.0018 is 0.0015, then will
Corresponding raw score 0.002 is mapped as corresponding target score 0.00002.Also such as raw score probability accumulated value is
When 0.0019, in the absence of the target score probability accumulated value equal to 0.0019, but can be with difference to the maximum mesh less than 0.0019
It is 0.0015 to mark score value probability accumulated value, then corresponding raw score 0.004 is mapped as into corresponding target score 0.00002.
As can be seen that raw score 0.0018 and 0.0019 is mapped as 0.00002, but the equal situation of this part have no effect on it is whole
The order of the raw score retained on body, is still isotonic regression.
If raw score is obtained by GBDT algorithms in the present embodiment, corresponding raw score probability cumulative distribution curve is such as
Shown in Fig. 7, the probability cumulative distribution curve of the target score after mapping is then as shown in Figure 8.Fig. 7 and Fig. 8 ordinate is represented
Probability accumulated value, abscissa represents the score value being sorted in ascending order.
In one embodiment, step 606 includes:Raw score probability accumulated value is mapped as to be more than raw score probability
The minimum target score value probability accumulated value of accumulated value.
In one embodiment, step 606 includes:Obtain general less than the maximum target score value of raw score probability accumulated value
Rate accumulated value and more than the average value between the minimum target score value probability accumulated value of raw score probability accumulated value, by raw score
Probability accumulated value is mapped as average value.Here average value can be arithmetic mean of instantaneous value or weighted average, weighted average
Weight can be arranged as required to.
In the present embodiment, according to target score probability cumulative distribution, preferential lookup is equal with raw score probability accumulated value
Target score probability accumulated value mapped, search the target close with raw score probability accumulated value again if not finding
Score value probability accumulated value is mapped.The mapping of more accurate isotonic regression can be reached with less amount of calculation, it is simple, high
Imitate and accurate.
In one embodiment, target score is to carry out the general of probabilistic forecasting acquisition to forecast sample collection according to forecast model
Rate value, and forecast model is according to training sample set and by the progress of logistic regression (Logistic Regression, LR) algorithm
What training was obtained.In the present embodiment, raw score is mapped directly into probable value, can be directly according to probable value in recommended
Recommended, it is to avoid target score is again mapped as to the complicated processes of probable value.And use logistic regression Algorithm for Training to obtain
Forecast model the probability distribution of acquisition is predicted to forecast sample collection closer to real probability distribution.
As shown in figure 9, in one embodiment, the object recommendation method also includes generation target score probability cumulative distribution
The step of, specifically include following steps:
Step 902, the full dose training sample set and full dose forecast sample collection under existing object recommendation scene are obtained.
Specifically, server can recommend existing object all samples under scene to be divided into two parts, and a part is as complete
Training sample set is measured, another part is used as full dose forecast sample collection.Full dose training sample set includes Positive training sample and negative training
Sample.
Step 904, according to full dose training sample set and pass through logistic regression Algorithm for Training obtain forecast model.
Specifically, server can use full dose training sample set, and the training sample that full dose training sample is concentrated extracts special
After levying, it is trained according to the feature of extraction and by logistic regression algorithm, obtains forecast model.General full dose training sample set
Middle training samples number is very more, and time consumption for training is long, but the probability distribution of the score value of the forecast model prediction of training acquisition is very accurate
Really.
Step 906, score value prediction is carried out to full dose forecast sample collection according to forecast model, obtains target score probability point
Cloth.
Specifically, server uses the feature identical feature extraction mode with extracting training sample, from the pre- test sample of full dose
All forecast samples in this extract feature, and are input to the forecast model trained, obtain the target point of each forecast sample
Value, target score probability distribution is obtained to the target score statistical probability predicted.
Step 908, corresponding target score probability cumulative distribution is obtained according to target score probability distribution.
Specifically, server can travel through the target score in ranking results, obtain after by target score according to ascending sort
Take the corresponding probable value of the target score currently traveled through, thus by the ranking results since the first to the target currently traveled through
All probable values of score value are added, and obtain the corresponding target score probability accumulated value of target score currently traveled through, according to all
The corresponding target score probability accumulated value of target score constitutes target score probability distribution.
In the present embodiment, forecast model is obtained according to full dose training sample set and by logistic regression Algorithm for Training, so
The score value that the forecast model trained is predicted is distributed close to true probability, and in recommended, recommendation results will be more accurate.
As shown in Figure 10, in one embodiment, the object recommendation method also includes generation target score probability iterated integral
The step of cloth, specifically include following steps:
Step 1002, the full dose training sample set and full dose forecast sample collection under existing object recommendation scene are obtained.
Step 1004, uniform sampling is carried out to full dose training sample set, according to the training sample of sampling acquisition and by patrolling
Collect regression algorithm training and obtain forecast model.
In one embodiment, step 1004 includes:The full dose Negative training sample that full dose training sample is concentrated is carried out equal
The Negative training sample of even sampling, the full dose Positive training sample concentrated according to full dose training sample and sampling acquisition is simultaneously returned by logic
Algorithm for Training is returned to obtain forecast model.
Full dose training sample concentrates the quantity of Negative training sample to be much larger than the quantity of Positive training sample, only retains full dose and just instructs
White silk sample can ensure the forecasting accuracy of the forecast model trained, and full dose Negative training sample progress uniform sampling can be carried
High training effectiveness.Such as full dose training sample concentrates training sample sum to be 100,000,000, wherein Positive training sample 1,000,000, bears training sample
This 99,000,000.If all training samples are all used for train forecast model, the requirement to computing capability can be higher.Therefore it is right
Positive training sample all risk insurance is stayed, and uniform sampling is carried out in 10% ratio to Negative training sample, that is, from 99,000,000 Negative training samples
9,900,000 training samples are extracted according to certain rule in the inside.
In one embodiment, step 1004 includes:The full dose Positive training sample and full dose concentrated to full dose training sample
Negative training sample carries out uniform sampling according to different sample rates respectively, the Positive training sample obtained according to sampling and negative training sample
This simultaneously obtains forecast model by logistic regression Algorithm for Training.Full dose just can trained respectively when training sample set is very huge
Sample and full dose Negative training sample carry out uniform sampling, to improve training effectiveness.
Step 1006, forecast model is calibrated.
In one embodiment, if full dose Positive training sample is retained, full dose Negative training sample carries out equal according to default sample rate r
Even sampling., can be by forecast model then when being calibrated to forecast modelIt is revised as
Wherein xiFor the feature of input, wiThe weight being characterized, k and c are constant parameter, and r is the sample rate of Negative training sample, and y is prediction
Middle score value.
In one embodiment, the logical function of forecast model can be modified to realize the calibration of forecast model.Tool
The logical function of body script is p=1/ (1+e-y), amended logical function is p '=p/ (p+ (1-p)/r), and wherein y is prediction
Middle score value, pass through amended logical function p ' export target score.
Step 1008, score value prediction is carried out to full dose forecast sample collection according to the forecast model of calibration, obtains target score
Probability distribution.
Step 1010, corresponding target score probability cumulative distribution is obtained according to target score probability distribution.
In the present embodiment, training effectiveness is improved by carrying out uniform sampling to full dose training sample set, and by pre-
The calibration for surveying model causes the target score probability distribution that final prediction is obtained to meet real probability distribution.Reference picture 11, it is bent
Line 1 is the target score probability density curve before calibration, and curve 2 is the target score probability density curve after calibration, and curve 2 is more
Meet real target score probability distribution.Abscissa is the value of target score in Figure 11, and ordinate is corresponding probable value.
As shown in figure 12, in one embodiment there is provided a kind of recommendation information processing method, the present embodiment is in this way
Illustrated applied to the terminal 110 in above-mentioned Fig. 1 and Fig. 2.This method specifically includes following steps:
Step 1202, the recommendation information of recommended is received;Recommendation information is picked out according to the target score of object,
Target score is that being scored object and after obtaining raw score, default target point under scene is recommended according to existing object
Value probability cumulative distribution maps raw score isotonic regression.
Specifically, server can obtain the raw score of each object in candidate target set;Obtain existing object and recommend field
Default target score probability cumulative distribution under scape;According to target score probability cumulative distribution, raw score isotonic regression is reflected
Penetrate as target score;Corresponding object is selected from candidate target set according to target score;Recommendation is sent to terminal to pick out
Object recommendation information.Recommendation information includes object identity, can also include object factory text and/or description picture, also
The corresponding target score of object each picked out can be included.
In one embodiment, server can obtain each original by raw score according to ascending sort according to ranking results
The corresponding raw score probability accumulated value of score value;According to target score probability cumulative distribution, raw score probability accumulated value is reflected
Penetrate the target score probability accumulated value to be matched with raw score probability accumulated value;Raw score is mapped as general with target score
The corresponding target score of rate accumulated value.
In one embodiment, server can be searched and tire out with raw score probability according to target score probability cumulative distribution
The equal target score probability accumulated value of product value;If finding, the mesh that raw score probability accumulated value is mapped as finding
Mark score value probability accumulated value;If not finding, the target score probability that raw score probability accumulated value is mapped as finding
Accumulated value.
In one embodiment, if server is not found, raw score probability accumulated value can be mapped as with it is original
The immediate target score probability accumulated value of score value probability accumulated value;Or, raw score probability accumulated value is mapped as being less than
The maximum target score value probability accumulated value of raw score probability accumulated value;Or, raw score probability accumulated value is mapped as greatly
In the minimum target score value probability accumulated value of raw score probability accumulated value;Or, obtain and be less than raw score probability accumulated value
Maximum target score value probability accumulated value and more than between the minimum target score value probability accumulated value of raw score probability accumulated value
Average value, average value is mapped as by raw score probability accumulated value.
In one embodiment, target score is to carry out the general of probabilistic forecasting acquisition to forecast sample collection according to forecast model
Rate value, and forecast model is according to training sample set and is trained acquisition by logistic regression algorithm.
In one embodiment, the full dose training sample set and full dose that server can be obtained under existing object recommendation scene are pre-
Survey sample set;Forecast model is obtained according to full dose training sample set and by logistic regression Algorithm for Training;According to forecast model pair
Full dose forecast sample collection carries out score value prediction, obtains target score probability distribution;Obtain corresponding according to target score probability distribution
Target score probability cumulative distribution.
In one embodiment, the full dose training sample set and full dose that server can be obtained under existing object recommendation scene are pre-
Survey sample set;Uniform sampling is carried out to full dose training sample set, the training sample obtained according to sampling is simultaneously calculated by logistic regression
Method training obtains forecast model;Forecast model is calibrated;Full dose forecast sample collection is carried out according to the forecast model of calibration
Score value is predicted, obtains target score probability distribution;Corresponding target score probability accumulation is obtained according to target score probability distribution
Distribution.
Step 1204, recommendation information is ranked up according to the target score of object.
Step 1206, recommendation information is shown according to clooating sequence.
Specifically, terminal can be ranked up according to the order of target score from big to small to corresponding recommendation information, and root
According to clooating sequence, recommendation information is shown according to the order of corresponding target score from big to small.
Terminal can also obtain the operational order to the recommendation information of displaying, and recommendation information is rung according to operational order
Should.If being user than object, then terminal can initiate plusing good friend request to server;If object is commodity, terminal can be to clothes
Business device initiates goods purchase request;If object is application program, terminal can initiate application program download request etc. to server.
Above-mentioned recommendation information processing method, carries out scoring by object and obtains corresponding original score value, according still further to current right
As recommending default target score probability cumulative distribution under scene that original score value isotonic regression is mapped into target score, so that
Recommend corresponding object according to target score., can order-preserving even if obtaining raw score using different scoring algorithms
It is mapped to the target score for meeting target score probability cumulative distribution, the probability distribution of the target score used in such recommended
It is accurate, the recommendation results further according to target score recommended are also accurately.
As shown in figure 13, obtained in one embodiment there is provided a kind of object recommendation device 1300, including raw score
Module 1301, target score probability cumulative distribution acquisition module 1302, mapping block 1303, Choosing module 1304 and recommending module
1305。
Raw score acquisition module 1301, the raw score for obtaining each object in candidate target set.
Target score probability cumulative distribution acquisition module 1302, default target under scene is recommended for obtaining existing object
Score value probability cumulative distribution.
Mapping block 1303, for according to target score probability cumulative distribution, raw score isotonic regression to be mapped as into mesh
Mark score value.
Choosing module 1304, for selecting corresponding object from candidate target set according to target score.
Recommending module 1305, for the object for recommending to pick out.
Above-mentioned object recommendation device 1300, carries out scoring by object and obtains corresponding original score value, according still further to current right
As recommending default target score probability cumulative distribution under scene that original score value isotonic regression is mapped into target score, so that
Recommend corresponding object according to target score., can order-preserving even if obtaining raw score using different scoring algorithms
It is mapped to the target score for meeting target score probability cumulative distribution, the probability distribution of the target score used in such recommended
It is accurate, the recommendation results further according to target score recommended are also accurately.
As shown in figure 14, in one embodiment, mapping block 1303 includes:Raw score probability accumulated value acquisition module
1303a, probability accumulated value mapping block 1303b and score value mapping block 1303c.
Raw score probability accumulated value acquisition module 1303a, for raw score, according to ascending sort, to be tied according to sequence
Fruit obtains the corresponding raw score probability accumulated value of each raw score.
Probability accumulated value mapping block 1303b, for according to target score probability cumulative distribution, raw score probability to be tired out
Product value is mapped as the target score probability accumulated value matched with raw score probability accumulated value.
Score value mapping block 1303c, for raw score to be mapped as into target corresponding with target score probability accumulated value
Score value.
In the present embodiment, raw score is after ascending sort, and the raw score probability accumulated value of acquisition can reflect
The raw score for going out to be less than or equal to current raw score accounts for the ratio of all raw scores, and target score probability accumulated value is also anti-
The target score mirrored less than or equal to corresponding target score accounts for the ratio of all target scores.So raw score probability is tired out
After product value and the matching of target score probability accumulated value, corresponding raw score is mapped as corresponding target score, just to reflect
Target score after penetrating can both retain the magnitude relationship of raw score, and default target score probability distribution can be met again,
And amount of calculation is small.
In one embodiment, probability accumulated value mapping block 1303b is specifically for according to target score probability iterated integral
Cloth, searches the target score probability accumulated value equal with raw score probability accumulated value;If finding, by raw score probability
Accumulated value is mapped as the target score probability accumulated value found;If not finding, raw score probability accumulated value is mapped
For with the immediate target score probability accumulated value of raw score probability accumulated value;Or, raw score probability accumulated value is reflected
Penetrate as the maximum target score value probability accumulated value less than raw score probability accumulated value;Or, by raw score probability accumulated value
It is mapped as the minimum target score value probability accumulated value more than raw score probability accumulated value;Or, obtain general less than raw score
The maximum target score value probability accumulated value of rate accumulated value and more than raw score probability accumulated value minimum target score value probability tire out
Average value between product value, average value is mapped as by raw score probability accumulated value.
In the present embodiment, according to target score probability cumulative distribution, preferential lookup is equal with raw score probability accumulated value
Target score probability accumulated value mapped, search the target close with raw score probability accumulated value again if not finding
Score value probability accumulated value is mapped.The mapping of more accurate isotonic regression can be reached with less amount of calculation, it is simple, high
Imitate and accurate.
In one embodiment, target score is to carry out the general of probabilistic forecasting acquisition to forecast sample collection according to forecast model
Rate value, and forecast model is according to training sample set and is trained acquisition by logistic regression algorithm.
As shown in figure 15, in one embodiment, object recommendation device 1300 also includes:Full dose training sample set obtains mould
Block 1306, training module 1307, prediction module 1308 and probability cumulative distribution generation module 1309.
Full dose training sample set acquisition module 1306, for obtaining the full dose training sample set under existing object recommendation scene
With full dose forecast sample collection.
Training module 1307, for obtaining prediction mould according to full dose training sample set and by logistic regression Algorithm for Training
Type.
Prediction module 1308, for carrying out score value prediction to full dose forecast sample collection according to forecast model, obtains target point
It is worth probability distribution.
Probability cumulative distribution generation module 1309, it is general for obtaining corresponding target score according to target score probability distribution
Rate cumulative distribution.
In the present embodiment, forecast model is obtained according to full dose training sample set and by logistic regression Algorithm for Training, so
The score value that the forecast model trained is predicted is distributed close to true probability, and in recommended, recommendation results will be more accurate.
As shown in figure 16, in one embodiment, object recommendation device 1300 also includes:Full dose training sample set obtains mould
Block 1306, training module 1307, prediction module 1308, probability cumulative distribution generation module 1309 and calibration module 1310.
Full dose training sample set acquisition module 1306, for obtaining the full dose training sample set under existing object recommendation scene
With full dose forecast sample collection.
Training module 1307, for carrying out uniform sampling to full dose training sample set, the training sample obtained according to sampling
And forecast model is obtained by logistic regression Algorithm for Training.
Calibration module 1310, is calibrated to forecast model.
Prediction module 1308, carries out score value prediction to full dose forecast sample collection for the forecast model according to calibration, obtains
Target score probability distribution.
Probability cumulative distribution generation module 1309, it is general for obtaining corresponding target score according to target score probability distribution
Rate cumulative distribution.
In the present embodiment, training effectiveness is improved by carrying out uniform sampling to full dose training sample set, and by pre-
The calibration for surveying model causes the target score probability distribution that final prediction is obtained to meet real probability distribution.
In one embodiment, training module 1307 is specifically for the full dose Negative training sample concentrated to full dose training sample
Uniform sampling is carried out, the Negative training sample that the full dose Positive training sample and sampling concentrated according to full dose training sample are obtained simultaneously passes through
Logistic regression Algorithm for Training obtains forecast model.
As shown in figure 17, in one embodiment there is provided a kind of recommendation information processing unit 1700, including:Recommendation
Cease receiving module 1701, order module 1702 and recommendation information display module 1703.
Recommendation information receiving module 1701, the recommendation information for receiving recommended;Recommendation information is according to object
What target score was picked out, target score is after obtaining raw score, to be recommended being scored object according to existing object
Default target score probability cumulative distribution maps raw score isotonic regression under scene.
Order module 1702, is ranked up for the target score according to object to recommendation information.
Recommendation information display module 1703, for showing recommendation information according to clooating sequence.
Above-mentioned recommendation information processing unit 1700, carries out scoring by object and obtains corresponding original score value, according still further to working as
Original score value isotonic regression is mapped to target score by default target score probability cumulative distribution under preceding object recommendation scene,
So as to recommend corresponding object according to target score., can even if obtaining raw score using different scoring algorithms
Order preserving map is to the target score for meeting target score probability cumulative distribution, the probability of the target score used in such recommended
Distribution is accurate, and the recommendation results further according to target score recommended are also accurately.
In one embodiment, target score is to carry out the general of probabilistic forecasting acquisition to forecast sample collection according to forecast model
Rate value, and forecast model is according to training sample set and is trained acquisition by logistic regression algorithm.
One of ordinary skill in the art will appreciate that realize all or part of flow in above-described embodiment method, being can be with
The hardware of correlation is instructed to complete by computer program, the computer program can be stored in embodied on computer readable storage Jie
In matter, the program is upon execution, it may include such as the flow of the embodiment of above-mentioned each method.Wherein, foregoing storage medium can be
The non-volatile memory mediums such as magnetic disc, CD, read-only memory (Read-Only Memory, ROM), or random storage note
Recall body (Random Access Memory, RAM) etc..
Each technical characteristic of above example can be combined arbitrarily, to make description succinct, not to above-described embodiment
In each technical characteristic it is all possible combination be all described, as long as however, the combination of these technical characteristics be not present lance
Shield, is all considered to be the scope of this specification record.
Above example only expresses the several embodiments of the present invention, and it describes more specific and detailed, but can not
Therefore it is construed as limiting the scope of the patent.It should be pointed out that for the person of ordinary skill of the art,
On the premise of not departing from present inventive concept, various modifications and improvements can be made, these belong to protection scope of the present invention.
Therefore, the protection domain of patent of the present invention should be determined by the appended claims.
Claims (18)
1. a kind of object recommendation method, methods described includes:
Obtain the raw score of each object in candidate target set;
Obtain existing object and recommend default target score probability cumulative distribution under scene;
According to the target score probability cumulative distribution, the raw score isotonic regression is mapped as target score;
Corresponding object is selected from the candidate target set according to the target score;
Recommend the object picked out.
2. according to the method described in claim 1, it is characterised in that described according to the target score probability cumulative distribution, will
The raw score isotonic regression is mapped as target score, including:
By the raw score according to ascending sort, the corresponding raw score of each raw score is obtained according to ranking results general
Rate accumulated value;
According to the target score probability cumulative distribution, the raw score probability accumulated value is mapped as and the raw score
The target score probability accumulated value of probability accumulated value matching;
The raw score is mapped as target score corresponding with the target score probability accumulated value.
3. method according to claim 2, it is characterised in that described according to the target score probability cumulative distribution, will
The raw score probability accumulated value is mapped as the target score probability accumulated value matched with the raw score probability accumulated value,
Including:
According to the target score probability cumulative distribution, the target score equal with the raw score probability accumulated value is searched general
Rate accumulated value;
If finding, the target score probability accumulated value that the raw score probability accumulated value is mapped as finding;
If not finding,
The raw score probability accumulated value is mapped as general with the immediate target score of raw score probability accumulated value
Rate accumulated value;Or,
The raw score probability accumulated value is mapped as general less than the maximum target score value of the raw score probability accumulated value
Rate accumulated value;Or,
The raw score probability accumulated value is mapped as general more than the minimum target score value of the raw score probability accumulated value
Rate accumulated value;Or,
Obtain less than the maximum target score value probability accumulated value of the raw score probability accumulated value and more than the raw score
Average value between the minimum target score value probability accumulated value of probability accumulated value, institute is mapped as by the raw score probability accumulated value
State average value.
4. according to the method described in claim 1, it is characterised in that the target score is to forecast sample according to forecast model
Collection carries out the probable value of probabilistic forecasting acquisition, and the forecast model is according to training sample set and entered by logistic regression algorithm
Row training is obtained.
5. according to the method described in claim 1, it is characterised in that methods described also includes:
Obtain the full dose training sample set and full dose forecast sample collection under existing object recommendation scene;
Forecast model is obtained according to the full dose training sample set and by logistic regression Algorithm for Training;
Score value prediction is carried out to full dose forecast sample collection according to the forecast model, target score probability distribution is obtained;
Corresponding target score probability cumulative distribution is obtained according to the target score probability distribution.
6. according to the method described in claim 1, it is characterised in that methods described also includes:
Obtain the full dose training sample set and full dose forecast sample collection under existing object recommendation scene;
Uniform sampling is carried out to the full dose training sample set, the training sample obtained according to sampling simultaneously passes through logistic regression algorithm
Training obtains forecast model;
The forecast model is calibrated;
Score value prediction is carried out to full dose forecast sample collection according to the forecast model of calibration, target score probability distribution is obtained;
Corresponding target score probability cumulative distribution is obtained according to the target score probability distribution.
7. method according to claim 6, it is characterised in that described uniformly to be adopted to the full dose training sample set
Sample, the training sample obtained according to sampling simultaneously obtains forecast model by logistic regression Algorithm for Training, including:
Uniform sampling is carried out to the full dose Negative training sample that the full dose training sample is concentrated, according to the full dose training sample set
In full dose Positive training sample and sampling obtain Negative training sample and pass through logistic regression Algorithm for Training obtain forecast model.
8. a kind of recommendation information processing method, methods described includes:
Receive the recommendation information of recommended;The recommendation information is picked out according to the target score of the object, described
Target score is that being scored the object and after obtaining raw score, default mesh under scene is recommended according to existing object
Mark score value probability cumulative distribution maps the raw score isotonic regression;
The recommendation information is ranked up according to the target score of the object;
The recommendation information is shown according to clooating sequence.
9. method according to claim 8, it is characterised in that the target score is to forecast sample according to forecast model
Collection carries out the probable value of probabilistic forecasting acquisition, and the forecast model is according to training sample set and entered by logistic regression algorithm
Row training is obtained.
10. a kind of object recommendation device, it is characterised in that described device includes:
Raw score acquisition module, the raw score for obtaining each object in candidate target set;
Target score probability cumulative distribution acquisition module, default target score probability under scene is recommended for obtaining existing object
Cumulative distribution;
Mapping block, for according to the target score probability cumulative distribution, the raw score isotonic regression to be mapped as into mesh
Mark score value;
Choosing module, for selecting corresponding object from the candidate target set according to the target score;
Recommending module, for the object for recommending to pick out.
11. device according to claim 10, it is characterised in that the mapping block includes:
Raw score probability accumulated value acquisition module, for the raw score, according to ascending sort, to be obtained according to ranking results
Obtain the corresponding raw score probability accumulated value of each raw score;
Probability accumulated value mapping block, for according to the target score probability cumulative distribution, the raw score probability to be tired out
Product value is mapped as the target score probability accumulated value matched with the raw score probability accumulated value;
Score value mapping block, for the raw score to be mapped as into target corresponding with the target score probability accumulated value point
Value.
12. device according to claim 11, it is characterised in that the probability accumulated value mapping block is specifically for basis
The target score probability cumulative distribution, searches the target score probability accumulation equal with the raw score probability accumulated value
Value;If finding, the target score probability accumulated value that the raw score probability accumulated value is mapped as finding;If not looking into
Find, be then mapped as the raw score probability accumulated value and the immediate target score of raw score probability accumulated value
Probability accumulated value;Or, the raw score probability accumulated value is mapped as to be less than the raw score probability accumulated value most
Big target score probability accumulated value;Or, the raw score probability accumulated value is mapped as to be more than the raw score probability
The minimum target score value probability accumulated value of accumulated value;Or, obtain the maximum target less than the raw score probability accumulated value
Score value probability accumulated value and more than the average value between the minimum target score value probability accumulated value of the raw score probability accumulated value,
The raw score probability accumulated value is mapped as the average value.
13. device according to claim 10, it is characterised in that the target score is to pre- test sample according to forecast model
This collection carries out the probable value of probabilistic forecasting acquisition, and the forecast model is according to training sample set and by logistic regression algorithm
It is trained acquisition.
14. device according to claim 10, it is characterised in that described device also includes:
Full dose training sample set acquisition module, recommends full dose training sample set and full dose under scene pre- for obtaining existing object
Survey sample set;
Training module, for obtaining forecast model according to the full dose training sample set and by logistic regression Algorithm for Training;
Prediction module, for carrying out score value prediction to full dose forecast sample collection according to the forecast model, obtains target score general
Rate is distributed;
Probability cumulative distribution generation module, tires out for obtaining corresponding target score probability according to the target score probability distribution
Integrate cloth.
15. device according to claim 10, it is characterised in that described device also includes:
Full dose training sample set acquisition module, recommends full dose training sample set and full dose under scene pre- for obtaining existing object
Survey sample set;
Training module, for carrying out uniform sampling to the full dose training sample set, the training sample obtained according to sampling simultaneously leads to
Cross logistic regression Algorithm for Training and obtain forecast model;
Calibration module, is calibrated to the forecast model;
Prediction module, carries out score value prediction to full dose forecast sample collection for the forecast model according to calibration, obtains target score
Probability distribution;
Probability cumulative distribution generation module, tires out for obtaining corresponding target score probability according to the target score probability distribution
Integrate cloth.
16. device according to claim 15, it is characterised in that the training module to the full dose specifically for training
Full dose Negative training sample in sample set carries out uniform sampling, the full dose Positive training sample concentrated according to the full dose training sample
With sampling obtain Negative training sample and pass through logistic regression Algorithm for Training obtain forecast model.
17. a kind of recommendation information processing unit, it is characterised in that described device includes:
Recommendation information receiving module, the recommendation information for receiving recommended;The recommendation information is according to the object
What target score was picked out, the target score be the object is scored and after obtaining raw score, according to current
Default target score probability cumulative distribution maps the raw score isotonic regression under object recommendation scene;
Order module, is ranked up for the target score according to the object to the recommendation information;
Recommendation information display module, for showing the recommendation information according to clooating sequence.
18. device according to claim 17, it is characterised in that the target score is to pre- test sample according to forecast model
This collection carries out the probable value of probabilistic forecasting acquisition, and the forecast model is according to training sample set and by logistic regression algorithm
It is trained acquisition.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610154736.3A CN107203558B (en) | 2016-03-17 | 2016-03-17 | Object recommendation method and device, and recommendation information processing method and device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610154736.3A CN107203558B (en) | 2016-03-17 | 2016-03-17 | Object recommendation method and device, and recommendation information processing method and device |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107203558A true CN107203558A (en) | 2017-09-26 |
CN107203558B CN107203558B (en) | 2021-03-09 |
Family
ID=59904136
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610154736.3A Active CN107203558B (en) | 2016-03-17 | 2016-03-17 | Object recommendation method and device, and recommendation information processing method and device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107203558B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108898591A (en) * | 2018-06-22 | 2018-11-27 | 北京小米移动软件有限公司 | Methods of marking and device, electronic equipment, the readable storage medium storing program for executing of picture quality |
CN109816187A (en) * | 2017-11-21 | 2019-05-28 | 财付通支付科技有限公司 | Information processing method, device, computer equipment and storage medium |
CN111382349A (en) * | 2018-12-29 | 2020-07-07 | 广州市百果园网络科技有限公司 | Information recommendation method and device, computer equipment and storage medium |
CN112925976A (en) * | 2021-01-29 | 2021-06-08 | 北京达佳互联信息技术有限公司 | Rating data distribution method, device, server and storage medium |
CN113256743A (en) * | 2021-06-16 | 2021-08-13 | 图兮数字科技(北京)有限公司 | Image processing method and device, electronic equipment and readable storage medium |
CN114528482A (en) * | 2022-01-25 | 2022-05-24 | 北京三快在线科技有限公司 | Method and device for determining recommended object, electronic equipment and storage medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102508870A (en) * | 2011-10-10 | 2012-06-20 | 南京大学 | Individualized recommending method in combination of rating data and label data |
CN102594905A (en) * | 2012-03-07 | 2012-07-18 | 南京邮电大学 | Method for recommending social network position interest points based on scene |
US20130344964A1 (en) * | 2012-06-22 | 2013-12-26 | J. Nathaniel Sloan | Method and device for fantasy sports roster recommendations |
CN103927347A (en) * | 2014-04-01 | 2014-07-16 | 复旦大学 | Collaborative filtering recommendation algorithm based on user behavior models and ant colony clustering |
US20150178293A1 (en) * | 2011-02-14 | 2015-06-25 | Microsoft Technology Licensing, Llc | Change invariant scene recognition by an agent |
-
2016
- 2016-03-17 CN CN201610154736.3A patent/CN107203558B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150178293A1 (en) * | 2011-02-14 | 2015-06-25 | Microsoft Technology Licensing, Llc | Change invariant scene recognition by an agent |
CN102508870A (en) * | 2011-10-10 | 2012-06-20 | 南京大学 | Individualized recommending method in combination of rating data and label data |
CN102594905A (en) * | 2012-03-07 | 2012-07-18 | 南京邮电大学 | Method for recommending social network position interest points based on scene |
US20130344964A1 (en) * | 2012-06-22 | 2013-12-26 | J. Nathaniel Sloan | Method and device for fantasy sports roster recommendations |
CN103927347A (en) * | 2014-04-01 | 2014-07-16 | 复旦大学 | Collaborative filtering recommendation algorithm based on user behavior models and ant colony clustering |
Non-Patent Citations (3)
Title |
---|
KUMARAVEL, ARTHI 等: "A Model to Predict the Severity of Acute Pancreatitis Based on Serum Level of Amylase and Body Mass Index", 《CLINICAL GASTROENTEROLOGY AND HEPATOLOGY》 * |
刘树栋 等: "基于位置的社会化网络推荐系统", 《计算机学报》 * |
许景楠: "基于评论和评分的个性化推荐算法研究", 《中国优秀硕士论文全文数据库 信息科技辑》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109816187A (en) * | 2017-11-21 | 2019-05-28 | 财付通支付科技有限公司 | Information processing method, device, computer equipment and storage medium |
CN108898591A (en) * | 2018-06-22 | 2018-11-27 | 北京小米移动软件有限公司 | Methods of marking and device, electronic equipment, the readable storage medium storing program for executing of picture quality |
CN111382349A (en) * | 2018-12-29 | 2020-07-07 | 广州市百果园网络科技有限公司 | Information recommendation method and device, computer equipment and storage medium |
CN111382349B (en) * | 2018-12-29 | 2023-08-15 | 广州市百果园网络科技有限公司 | Information recommendation method, device, computer equipment and storage medium |
CN112925976A (en) * | 2021-01-29 | 2021-06-08 | 北京达佳互联信息技术有限公司 | Rating data distribution method, device, server and storage medium |
CN112925976B (en) * | 2021-01-29 | 2023-11-21 | 北京达佳互联信息技术有限公司 | Method, device, server and storage medium for allocating denomination data |
CN113256743A (en) * | 2021-06-16 | 2021-08-13 | 图兮数字科技(北京)有限公司 | Image processing method and device, electronic equipment and readable storage medium |
CN113256743B (en) * | 2021-06-16 | 2022-09-02 | 图兮数字科技(北京)有限公司 | Image processing method and device, electronic equipment and readable storage medium |
CN114528482A (en) * | 2022-01-25 | 2022-05-24 | 北京三快在线科技有限公司 | Method and device for determining recommended object, electronic equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN107203558B (en) | 2021-03-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107203558A (en) | Object recommendation method and apparatus, recommendation information treating method and apparatus | |
CN106651519B (en) | Personalized recommendation method and system based on label information | |
CN103902538B (en) | Information recommending apparatus and method based on decision tree | |
Chen et al. | Using fruit fly optimization algorithm optimized grey model neural network to perform satisfaction analysis for e-business service | |
CN104424296B (en) | Query word sorting technique and device | |
CN111191732A (en) | Target detection method based on full-automatic learning | |
CN106815192A (en) | Model training method and device and sentence emotion identification method and device | |
CN108133013A (en) | Information processing method, device, computer equipment and storage medium | |
CN108427708A (en) | Data processing method, device, storage medium and electronic device | |
CN106600052A (en) | User attribute and social network detection system based on space-time locus | |
CN102841946A (en) | Commodity data retrieval sequencing and commodity recommendation method and system | |
CA3059929C (en) | Text searching method, apparatus, and non-transitory computer-readable storage medium | |
CN106844407A (en) | Label network production method and system based on data set correlation | |
CN110019163A (en) | Method, system, equipment and the storage medium of prediction, the recommendation of characteristics of objects | |
CN107016122A (en) | Knowledge recommendation method based on time-shift | |
CN107918657A (en) | The matching process and device of a kind of data source | |
CN108052625A (en) | A kind of entity sophisticated category method | |
CN107153656A (en) | A kind of information search method and device | |
CN107481054A (en) | The push of hotel's favor information and device, electronic equipment, storage medium | |
CN110532351A (en) | Recommend word methods of exhibiting, device, equipment and computer readable storage medium | |
CN110019790A (en) | Text identification, text monitoring, data object identification, data processing method | |
CN106919588A (en) | A kind of application program search system and method | |
CN109447110A (en) | The method of the multi-tag classification of comprehensive neighbours' label correlative character and sample characteristics | |
CN108932648A (en) | A kind of method and apparatus for predicting its model of item property data and training | |
CN108897750A (en) | Merge the personalized location recommendation method and equipment of polynary contextual information |
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 |