CN104951441A - Method and device for sequencing objects - Google Patents
Method and device for sequencing objects Download PDFInfo
- Publication number
- CN104951441A CN104951441A CN201410112060.2A CN201410112060A CN104951441A CN 104951441 A CN104951441 A CN 104951441A CN 201410112060 A CN201410112060 A CN 201410112060A CN 104951441 A CN104951441 A CN 104951441A
- Authority
- CN
- China
- Prior art keywords
- keyword
- user
- feature
- behavior
- current preference
- 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.)
- Pending
Links
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention relates to a method for sequencing objects. The method includes the steps of obtaining keywords, keyword appearance identification sequences and keyword corresponding behavior types related to later behaviors of a user, determining features of the objects currently preferred by the user through a preset user current preference prediction model according to the keywords, keyword appearance identification sequences and keyword corresponding behavior types, and adjusting sequencing factors of the to-be-sequenced objects related to the user current preference according to the determined user current preference so that sequencing of the to-be-sequenced objects can be influenced. By means of the scheme, the relations between all feature data and features of final preference are analyzed with the aid of recent behavior feature data, the features of the objects of the user behavior preference can be more accurately determined, the sequencing of search behavior results corresponding to the user is guided through the preference features, and a more individual object sequencing result is provided for the user.
Description
Technical field
The application relates to internet arena, more specifically, relates to a kind of method of sorting to object and device.
Background technology
At present, there are some and predicted according to carrying out user preference the method that data search result is provided to user, these method major parts utilize the concrete behavior of user (such as: click) to carry out compute user preferences in conjunction with the individualized feature of user, then provides the data processed result of personalized ordering to user according to the user preference calculated.Such as, user, when e-commerce website is searched for, is mainly divided into two classes to the personalized search results relevant individualized feature that sorts: a class is the build-in attribute feature of user, as sex, age, region etc.; Another kind of is the behavioural characteristic of user, as purchasing power, classification preference, keyword etc.
But, along with people live require raising, derived multiple personal demand.Weigh the dimension of these individual demands except intrinsic sex, age, purchasing power etc., also comprise the interest preference that some are more abstract, the interest preference of this abstract can use the genre preference of user to describe.In addition, due to the polytrope of user behavior and interest, the genre preference predicting user in real time is also needed.
In the prior art, carry out personalized ordering owing to depending on the concrete behavior of user to data processed result, therefore, if user's liveness is not enough, likely cause the basic data of carrying out type of preferences calculating comparatively sparse, generalization ability is poor.In addition, if keyword granularity is meticulous, also the problems such as a series of data expansion may be brought.Meanwhile, existing method seldom utilizes the future behaviour of real-time behavior sequence to user of user to predict, results through the user individual the data precision that existing technical scheme obtains not high.Therefore, the individuation data result obtained by the technical scheme that this generalization ability is not strong can not embody the interest preference of user exactly, namely, genre preference, thus cause that the data processed result efficiency being supplied to respective user is low, poor accuracy, the inadequate hommization of sequence of user behavior result, reduces the experience of user.
Summary of the invention
The fundamental purpose of the application is, for above-mentioned defect, the technology sorted to object is provided, to solve the problem that the basic data of carrying out type of preferences calculating is comparatively sparse, generalization ability is poor caused because user's liveness is not enough, and the problem of a series of data expansions synonym problems in keyword can avoided and bring because keyword granularity is meticulous.Meanwhile, carry out modeling by the real-time behavior sequence of user, thus the preference that real-time estimate user is current more accurately, promote the accuracy of data processed result, improve the experience of user.
According to the first aspect of the application, provide a kind of method that object is sorted, it is characterized in that, comprising: the behavior type corresponding to identifier and keyword appears in the keyword that the recent behavior obtaining user relates to, keyword; There is the behavior type corresponding to identifier and keyword according to described keyword, keyword, use the current preference forecast model of user set up in advance, determine the feature of the object of the current preference of user; And according to the current preference of the user determined, adjust the ranking factor for the treatment of ranked object relevant to the current preference of described user, to affect the sequence treating ranked object.
According to the second aspect of the application, providing a kind of device sorted to object, it is characterized in that, comprising: acquisition module, there is the behavior type corresponding to identifier and keyword in the keyword that the recent behavior obtaining user relates to, keyword; Determination module, for there is the behavior type corresponding to identifier and keyword according to described keyword, keyword, uses the current preference forecast model of user set up in advance, determines the feature of the object of the current preference of user; And adjusting module, for according to the current preference of the user determined, adjust the ranking factor for the treatment of ranked object relevant to the current preference of described user, to affect the sequence treating ranked object.
Compared with prior art, according to the technical scheme of the application, can when there is behavior operation in user, by (as: clicking the behavior of user's active user, collection, search) keyword carry out abstract, form the preference of extensive relevant stylistic category, more accurately the feature of the object of the current preference of user is predicted, thus avoid the basic data that causes because user's liveness is not enough in the prior art sparse, the problem of generalization ability difference, synonym problems in keyword, and the defect such as a series of data expansions to bring because keyword granularity is meticulous, simultaneously, modeling is carried out by the real-time behavior sequence of user, and then make the sequence hommization more of the real-time behavior outcome of user, improve the experience of user.
Accompanying drawing explanation
Accompanying drawing described herein is used to provide further understanding of the present application, and form a application's part, the schematic description and description of the application, for explaining the application, does not form the improper restriction to the application.In the accompanying drawings:
Fig. 1 is the process flow diagram to the method that object sorts according to the application's embodiment;
Fig. 2 is the process flow diagram setting up the method for the current preference forecast model of user according to the application's embodiment;
Fig. 3 is the process flow diagram to the method that object sorts of the more specific embodiment according to the application; And
Fig. 4 is the block diagram to the device that object sorts according to the application's embodiment.
Embodiment
The main thought of the application is, by determining the feature of the object of the current preference of user in conjunction with the recent behavioural characteristic data of user before current point in time, and adjust the relevant ranking factor treating ranked object of preference current to this according to the current preference of the user determined.This programme is by the relation between each characteristic of recent behavioural characteristic data analysis and the feature of final preference, can determine the feature of the object of user behavior preference more accurately, and the sequence of the search behavior result instructing user corresponding by the feature of this preference, provide more personalized object order result to user.
For describing the scheme of the application, below by for user behavior amount huge and intelligible e-commerce platform, be specifically described.
Conveniently hereafter describe, first introductory section terminological interpretation.
Stylistic category: a kind of abstractness of representative object characteristic describes, as the American-European style under the women's dress in e-commerce platform, Japan and Korea S's style etc., the European style in household classification, pastoralism etc.
User behaviors log: on e-commerce website, the various operations of user all can go on record and be used as user journal, specifically comprise: search, category browses, check object, the time series relation of operation (as bought/collection) on object details page and these behaviors.
Historical behavior: if this behavior occurs in the time in the past, be called historical behavior.
Recent behavior: the behavior of the default behavior number of times before current point in time.
For making the object of the application, technical scheme and advantage clearly, below in conjunction with the application's specific embodiment and corresponding accompanying drawing, technical scheme is clearly and completely described.Obviously, described embodiment is only some embodiments of the present application, instead of whole embodiments.Based on the embodiment in the application, those of ordinary skill in the art are not making the every other embodiment obtained under creative work prerequisite, all belong to the scope of the application's protection.
Fig. 1 is the process flow diagram to the method 100 that object sorts according to the application's embodiment.As shown in Figure 1, method 100 starts from step 101.
In step 101, there is the behavior type corresponding to identifier and keyword in the keyword that the recent behavior obtaining user relates to, keyword.
Specifically, the behavioral data before user's current point in time can be obtained from online user behaviors log.After determining current time point, according to presetting behavior number of times, the behavior sequence of these behaviors before current point in time can be defined as recent behavior.Such as, using 7 behaviors before from current point in time to this time point as recent behavior.
Extract the behavior record of recent behavior in user behaviors log after, the behavior can be obtained according to pre-defined rule and record the data object (such as document) related to, by obtaining one or more keyword to the semantic analysis of described data object.Such as, at e-commerce platform, can obtain searching for from journal file, category browses, check object, the time series relation of operation (as bought/collection) on object details page and these behaviors.Based on the relevant textual information of these behavior record referents, according to pre-defined rule, one or more keyword can be obtained.Meanwhile, need to obtain the behavior type corresponding with each keyword, as click, search, collection etc.The keyword that the recent behavior of user relates to also can be the label on a certain characteristic attribute of the recent behavior referent of user.
In addition, by obtaining one or more keywords of recent behavior record referent, whether can occur according in each behavior referent of each keyword in the recent behavior of user, obtaining keyword and occur identifier.
Such as: at e-commerce platform, using from current point in time to 5 before, time click behavior is as recent behavior, and the keyword that the object clicked comprises is respectively:
The keyword that the object clicked for the first time comprises: sweet, American-European, textile, cartoon, stand in Europe }
The keyword that the object that second time is clicked comprises: { sweet, cartoon, bowknot }
The keyword that the object clicked for the third time comprises: { sweet, textile, human skeleton }
The keyword that the object clicked for 4th time comprises: { American-European, station, Europe, human skeleton }
The keyword that the object clicked for 5th time comprises: { American-European, textile, human skeleton }
In the present embodiment, can set up for each keyword the record occurring situation in each behavior referent of this keyword in the recent behavior of user.When this keyword occurs in certain behavior, then to should time behavior once identify in this record.The record formed can be called the appearance identifier of keyword.Describedly occur that identifier can be represented by the mode of Serial No..Describedly occur that identifier comprises one or more flag, each flag is corresponding with each behavior of user respectively, and whether each flag is respectively used to identify this keyword and occurs in a behavior referent of correspondence.
Such as, whether occur in each click behavior referent according to above-mentioned each keyword (after duplicate removal), as occurred, then in the appearance identifier of this keyword to should the flag of time click behavior identify with numeral " 1 ", as do not occurred, then in the appearance identifier of this keyword to should the flag of time click behavior identify with digital " 0 ".Like this, the appearance identifier of above-mentioned each keyword can be obtained:
Sweet: " 11100 "
American-European: " 10011 "
Textile: " 10101 "
Cartoon: " 11000 "
Station, Europe: " 10010 "
Human skeleton: " 00111 "
Bowknot: " 01000 "
In step 102, the behavior type corresponding to identifier and keyword is there is according to described keyword, keyword, use the user current preference forecast model relevant with user behavior referent feature set up in advance, determine the feature of the object of the current preference of user.
Specifically, to the feature of user behavior referent can be the label of this object on a certain attribute preset.Such as, when carrying out feature interpretation to merchandise items, stylistic category can be used to represent a kind of abstract characteristics of this merchandise items.Here stylistic category is that a kind of abstractness of representative object characteristic describes, as the American-European style under the women's dress in e-commerce platform, Japan and Korea S's style etc., and the European style in household classification, pastoralism etc.Keyword, the behavior type corresponding with keyword and the keyword that the user's historical behavior by obtaining according to user's history log can be related to occur that the parameter combinations such as identifier are as feature, using user's historical behavior for the feature of referent as target, set up the user current preference forecast model relevant with user behavior referent feature, as the probability model utilizing the training of maximum entropy disaggregated model to obtain.The keyword that described user current preference forecast model uses user, the appearance identifier of keyword in nearest behavior and the behavior type etc. corresponding to keyword are associated between keyword feature and the feature of user behavior referent relation.
According to the feature of the keyword that user uses in the recent period, use described user current preference forecast model can obtain the current preference of user.
In step 103, according to the current preference of the user determined, adjust the ranking factor for the treatment of ranked object relevant to the current preference of user, to affect the sequence treating ranked object.
Specifically, after the feature obtaining the current preference of user, to the feature of the current preference of user and can treat that the feature of ranked object carries out similarity respectively or correlativity judges, according to judged result, the ranking factor relevant to the feature of the current preference of user treating ranked object adjusts, thus the sequence of ranked object is treated in impact.
According to an embodiment of the application, can treat that the sequence of ranked object promotes by what there is the feature of the current preference of user.
So far, the process flow diagram to the method 100 that object sorts according to the application's embodiment is described.The method 100 of the present embodiment is by determining the feature of the current preference of user in conjunction with the characteristic of the recent behavior of user before current point in time, and the preference current according to the user determined adjusts the relevant ranking factor treating ranked object of preference current to this.Compared with prior art, the embodiment of the present application can be current to user more accurately preference carry out anticipation, and the sequence of the Search Results instructing user's current behavior corresponding by the preference that user is current, provide more personalized to user, the object order result of hommization.
In addition, the current preference forecast model of the user used in step 102 can preset, and also can be obtained by the training of machine learning model.
According to an embodiment of the application, the current preference forecast model of described user is set up by following steps: the feature obtaining the user's historical behavior referent in database; And obtain according to the user's historical behavior in a period of time of presetting the stylistic category that behavior type corresponding to identifier, keyword and object appear in keyword, keyword, carry out the model training of machine learning, set up the preference forecast model of the current preference of user.
The process of establishing of the current preference forecast model of user will be described in more detail below.
Fig. 2 is the process flow diagram setting up the method 200 of the current preference forecast model of user according to the embodiment of the present application.As shown in Figure 2, method 200 starts from step 201.
First, in step 201, set up user-keyword matrix according to the keyword that the user's historical behavior in section in those years relates to.
Specifically, the historical behavior record of the user in the past period can be obtained by user log files, the keyword comprised according to the object in the historical behavior record in this time period involved by each behavior or genre labels, can form user-keyword matrix.
Such as, at e-commerce platform, obtain the behavior record from current time to past in month by user's history log, with the click behavior of this digging user-keyword, form the matrix A of user-keyword, as follows:
Matrix A:
| Keyword 1 | Keyword 2 | Keyword 3 | Keyword 4 | |
| User 1 | W11 | W12 | W13 | W14 |
| User 2 | W21 | W22 | W23 | w24 |
| User 3 | W31 | W32 | W33 | W34 |
From matrix A, from current to when in the previous moon, have 3 users and click behavior occurs, relate to 4 keywords altogether, the number of times of W11 representative of consumer 1 Key Words 1, the number of times of W12 representative of consumer 1 Key Words 2, by that analogy.Like this, the matrix A of user-keyword is just established.
By this method, the user-keyword matrix can set up in any time period.
Then, entering step 202, according to user-keyword matrix, forming one or more cluster by carrying out cluster to keyword, the corresponding a kind of theme of each cluster, calculates the probability that each keyword belongs to each theme.
In this step, can to be dived semantic analysis by PLSA(probability) method carries out cluster to the keyword that step 201 obtains, that is, user and keyword are mapped to an implicit theme spatially, this implicit theme space is exactly style, thus can obtain the probability matrix that each keyword belongs to each theme.
Such as, at e-commerce platform, using the stylistic category of commodity as described theme.There are two stylistic categories, can obtain by user-keyword matrix the probability matrix that following each keyword belongs to each stylistic category (i.e. theme) by PLSA method:
Style 1:
Undergarment covering the chest and abdomen 0.200316
Wrap up in chest 0.118473
Hang neck 0.109674
Undergarment covering the chest and abdomen skirt 0.0668498
Suspender belt 0.0580471
Be backless 0.0502866
Intersect 0.0444334
Group is to 0.0325834
Suspender skirt 0.032397
Wrap up in chest skirt 0.0302184
Style 2:
Gloomy female 0.128117
Literature and art 0.0967896
Forest 0.0865181
Forest is 0.0780099
Linen-cotton 0.0725525
Japanese 0.0412643
Gloomy is 0.0333173
Literature and art model 0.0331822
Design of scattered small flowers and plants 0.0215523
In step 203, obtain each keyword and the frequency of occurrences thereof, to obtain the keyword vector of object.
In this step, first, by extracting keywords such as the title of each object and attributes, this object keyword vector can be formed by the keyword obtained and its frequency occurred.
In step 204, each keyword related to according to each object belongs to probability and this keyword frequency of occurrences of each theme, calculates the similarity of the keyword vector in the keyword vector cluster corresponding to each theme that this object relates to.
Below, step 203 is illustrated to 204.
Such as, at e-commerce platform, suppose the keyword vector A=[suspender skirt: 1 of certain commodity title and attribute, literature and art: 1, forest: 1, gloomy female: 1], the keyword vector B1 of style 1, style 2, B2 is as above in example shown in style 1 and style 2, then the similarity of the two can use formula (1) to calculate:
By formula (1), the similarity similarity (A, B2) of the keyword of the similarity similarity (A, B1) of the keyword of these commodity and the keyword of style 1 and the keyword of these commodity and style 2 can be drawn.
Afterwards, in step 205, get the feature tag of the theme corresponding to similarity maximal value as this object.
By using the method for step 201 to 205, the feature tag of all objects related in database can be obtained.
So far, complete the excavation to the stylistic category of object under line, that is, determine the feature tag of each object.
In step 206, the user's historical behavior in section in those years, choose sample according to predefined action number of times.Specifically, using the behavior sequence in the user's historical behavior in section in those years as sample data, sample can be chosen according to predefined action number of times.Specifically, sample extraction can be carried out according to the rule of getting some behaviors number of times at every turn.Such as, at e-commerce platform, the behavior sequence in 48 hours before choosing in those years, therefrom chooses sample characteristics.Specifically, such as, the window extracting feature from the behavior of first in this time period, can extract 8 behaviors as a training sample each time, feature is extracted, using the feature of the object involved by the 8th behavior as training objective from wherein front 7 behaviors.After completing the feature extraction of this behavior sequence, window slides backward one (behavior), continues to extract next sample, until extracted all samples.
Afterwards, in step 207, from each sample extracted, obtain keyword, the appearance identifier of keyword and the behavior type corresponding to keyword.
Such as, each extraction 8 behaviors are as a training sample, in each sample, extract keyword, the appearance identifier of keyword and the behavior type corresponding to keyword, using the feature (genre preference) of the object involved by the 8th behavior as target in 7 behaviors in the past.Here, behavior type be in user behavior history log user to the behavior type of involved object, such as, click, search etc.
Below, the method for feature extraction is illustrated.
Such as, at e-commerce platform, in the past sometime before 48 hours in carry out sample and choose, suppose to choose 8 behaviors each time as a sample.As in the sample chosen, user there occurs 8 behaviors, and the keyword of wherein front 7 behaviors is respectively:
The keyword that the object clicked for the first time comprises: sweet, American-European, textile, cartoon, stand in Europe }
The keyword that the object that second time is collected comprises: { American-European, surplus }
The keyword that the object clicked for the third time comprises: { sweet, cartoon, bowknot }
The keyword that the object clicked for 4th time comprises: { sweet, textile, human skeleton }
The keyword that the object clicked for 5th time comprises: { American-European, station, Europe, human skeleton }
The keyword that the object collected for 6th time comprises: { Korea Spro's version, surplus }
The keyword that the object clicked for 7th time comprises: { American-European, textile, human skeleton }
Wherein, the keyword that the object that can obtain clicking comprises is as follows:
The keyword that the object clicked for the first time comprises: sweet, American-European, textile, cartoon, stand in Europe }
The keyword that the object clicked for the third time comprises: { sweet, cartoon, bowknot }
The keyword that the object clicked for 4th time comprises: { sweet, textile, human skeleton }
The keyword that the object clicked for 5th time comprises: { American-European, station, Europe, human skeleton }
The keyword that the object clicked for 7th time comprises: { American-European, textile, human skeleton }
Such as, whether occur in each click behavior referent according to above-mentioned each keyword (after duplicate removal), as occurred, then in the appearance identifier of this keyword to should the flag of time click behavior identify with numeral " 1 ", as do not occurred, then in the appearance identifier of this keyword to should the flag of time click behavior identify with digital " 0 ".Like this, the appearance identifier of above-mentioned each keyword can be obtained:
Sweet: " 11100 "
American-European: " 10011 "
Textile: " 10101 "
Cartoon: " 11000 "
Station, Europe: " 10010 "
Human skeleton: " 00111 "
Bowknot: " 01000 "
The keyword that the object in like manner can collected comprises is as follows:
The keyword that the object that second time is collected comprises: { American-European, surplus }
The keyword that the object collected for 6th time comprises: { Korea Spro's version, surplus }
Whether occur in each click behavior referent according to above-mentioned each keyword (after duplicate removal), as occurred, then in the appearance identifier of this keyword to should the flag of time click behavior identify with numeral " 1 ", as do not occurred, then in the appearance identifier of this keyword to should the flag of time click behavior identify with digital " 0 ".Like this, the appearance identifier of each keyword of above-mentioned collection object can be obtained:
American-European: " 10 "
Surplus: " 11 "
Korea Spro's version: " 01 "
In step 208, using the combination of the behavior type corresponding to the appearance identifier of above-mentioned keyword, keyword and keyword as sample characteristics, using the feature of the object corresponding with sample characteristics as target, carry out the training of machine learning model, such as, utilize maximum entropy disaggregated model to carry out the training of machine learning model, the current preference forecast model of user can be obtained.
As a feature of model training sample after the appearance identifier of each keyword itself, this keyword and the behavior type corresponding to this keyword combine.Such as, in upper example, the appearance identifier of keyword " human skeleton " is " 00111 ", then " human skeleton #00111# clicks " is as a feature of this sample, and " Korea Spro version #01# collects " is as a sample characteristics.With the feature (stylistic category) of the object of the next behavior corresponding with above-mentioned 7 behaviors for target.The impact of different behavior types on postorder favorites style can be learnt out by this method.
In behavior sequence samples, choose sample characteristics successively, until window moves to last behavior, like this, can obtain characteristic set is { f1, f2, f3 ..., fn}, the feature of the object corresponding with sample characteristics, as target, can obtain goal set for { s1, s2, s3 ..., sm}.
According to an embodiment of the application, can using the combination of the behavior type corresponding to the appearance identifier of keyword itself, this keyword and this keyword as sample characteristics, using the feature of the object corresponding to sample characteristics as target, training maximum entropy disaggregated model.
Suppose characteristic set for f1, f2, f3 ..., fn}, goal set be s1, s2, s3 ..., sm}, then model can obtain the weight of each feature in each target.If model is matrix M, then the weight of Mij=i-th feature in a jth target.
So far, the process of model training is completed.The process of this model training is by user behavior operation on user behavior record artificial line under line, occur that the behavior type etc. corresponding to identifier, keyword combines as sample characteristics using keyword, keyword, using the feature of the object corresponding to sample characteristics as target, by model training, obtain each feature weight.
Fig. 3 is the process flow diagram to the method 300 that object sorts of the more specific embodiment according to the application.Method 300 starts from step 301.
First, in step 301, from recent behavior, obtain the behavior type corresponding to keyword, keyword.
The behavior of the preset times of user before current point in time is recent behavior.Can obtain the behavior record of these behaviors from real-time logs, therefrom extracting keywords and behavior type corresponding to keyword are as feature, and the extracting method of feature is similar with the method extracting feature during the training current preference pattern of user.Specifically, when will train the current preference forecast model of user, during the behavior number of times of the feature extraction of each training sample preference current as real-time online prediction user, need the behavior number of times before the current point in time chosen.Such as, the behavior sequence of seven behaviors of this user before choosing current point in time, therefrom extracts the behavior type corresponding to keyword, keyword, as feature.
Then, in step 302, whether occur in each behavior of recent behavior according to keyword, obtain the appearance identifier of keyword.To the description of this step and step 206 similar, do not repeat them here.
Afterwards, in step 303, occur that the combination of the behavior type corresponding to identifier and keyword is as feature, using the feature of the object of current preference as target using keyword, keyword, use the current preference forecast model of user, obtain the probability that each feature belongs to the feature of each object.
In step 304, the object belonging to the current preference of probability calculation user of the feature of each object according to each feature belongs to the probability of each feature tag.
In step 305, get the feature tag of the feature tag corresponding to maximum probability value as the object of the current preference of user.
Below, step 301 is illustrated to 305.
Such as, at e-commerce platform, obtained the real-time behavior of user by online real-time logs analytic system, suppose that the keyword chosen in 7 behaviors of user before current point in time is respectively:
The keyword that the object clicked for the first time comprises: sweet, American-European, textile, cartoon, stand in Europe }
The keyword that the object that second time is collected comprises: { American-European, surplus }
The keyword that the object clicked for the third time comprises: { sweet, cartoon, bowknot }
The keyword that the object clicked for 4th time comprises: { sweet, textile, human skeleton }
The keyword that the object clicked for 5th time comprises: { American-European, station, Europe, human skeleton }
The keyword that the object collected for 6th time comprises: { Korea Spro's version, surplus }
The keyword that the object clicked for 7th time comprises: { American-European, textile, human skeleton }
According to above behavior sequence, occur that the combination of the behavior type corresponding to identifier and keyword is as feature using keyword, keyword, following feature F1 to F10 can be obtained:
F1: sweet #11100# clicks
F2: American-European #10011# clicks
F3: textile #10101# clicks
F4: cartoon #11000# clicks
F5: station, Europe #10010# clicks
F6: human skeleton #00111# clicks
F7: bowknot #01000# clicks
F8: Korea Spro version #01# collects
F9: American-European #10# collection
F10: surplus #11# collects
Suppose feature tag class object: the sweet style of S1=, the American-European style of S2=, using above-mentioned feature as feature, by using the current preference forecast model of user set up in advance, can draw following matrix B, this matrix shows the probability that each feature belongs to each feature tag:
Matrix B:
| F1 | F2 | F3 | F4 | F5 | F6 | F7 | F8 | F9F10 | |
| S1 | 0.2 | -0.5 | 0.05 | 0.01 | -0.2 | -0.7 | 1.0 | 0.01 | -0.60.01 |
| S2 | -0.7 | 0.5 | 0.02 | 0.01 | 0.6 | 1.0 | -0.9 | -0.1 | 0.50.01 |
Afterwards, the mark of each feature tag of maximum entropy probabilistic forecasting formulae discovery can be used, and obtain the feature tag of point the highest feature tag as the object of the current preference of user.
Maximum entropy probabilistic forecasting formula is:
Scorej=e^ ∑ fiMij (formula 2)
In formula 2, fi be characteristic set f1, f3 ..., the feature (i=1 ~ k) in fk}, the weight of Mij=i-th feature in a jth target.
In this example, predict that the probability of the object of the sweet style of current click S1 is:
P(S1)=2.7^(0.2-0.5+0.05+0.01-0.2-0.7+1.0+0.01-0.6+0.01)=0.49
Predict that the probability of the object of the American-European style of current click S2 is:
P(S2)=2.7^(-0.7+0.5+0.02+0.01+0.6+1.0-0.9-0.1+0.5+0.01)=2.54
Therefore, measurable (determination) goes out the feature of the object of the current preference of user is American-European style.
In step 306, according to the current preference of the user determined, judge to treat whether ranked object has the feature of the current preference of user.
When having the feature of the current preference of user until ranked object, perform step 307, promote the sequence treating ranked object.
When not having the feature of the current preference of user until ranked object, perform step 308, do not promote the sequence treating ranked object.
According to an embodiment of the application, the ranking factor relevant to the current preference of user can by the mode of linear weighted function and other ranking factor actings in conjunction, thus determine to treat the ranking results that ranked object is final.
So far, the process flow diagram to the method 300 that object sorts according to the application's more specific embodiment is described.Compared with method 100, the method 300 provides the detailed step of real-time estimate stylistic category on line, the method 300 can be determined the feature of the object of the current preference of user equally exactly, and the sequence of user behavior result is instructed by this feature, provide more personalized object order result to user.
Fig. 4 is the block diagram to the device that object sorts according to the application's embodiment.
As shown in Figure 4, device 400 can comprise: acquisition module 401, and the behavior type corresponding to identifier and keyword appears in the keyword that the recent behavior obtaining user relates to, keyword; Determination module 402, for there is the behavior type corresponding to identifier and keyword according to described keyword, keyword, uses the current preference forecast model of user set up in advance, determines the feature of the object of the current preference of user; And adjusting module 403, for according to the current preference of the user determined, adjust the ranking factor for the treatment of ranked object relevant to the current preference of described user, to affect the sequence treating ranked object.
According to an embodiment of the application, the current preference forecast model of described user can pass through to set up with lower module: labeling module, for marking the feature tag of object; And set up module, for obtaining keyword according to the user's historical behavior in section in those years, there is corresponding to identifier, keyword behavior type in keyword, carries out the model training of machine learning, set up the current preference forecast model of user.
According to an embodiment of the application, described determination module 402 can comprise: set up submodule, and user-keyword matrix set up in the keyword related to for the user's historical behavior in basis in those years section; First calculating sub module, for according to described user-keyword matrix, forms one or more cluster by carrying out cluster to keyword, and the corresponding a kind of theme of each cluster, calculates the probability that each keyword belongs to each theme; Obtain submodule, for obtaining each keyword and the frequency of occurrences thereof, to obtain the keyword vector of object; Second calculating sub module, each keyword for relating to according to each object belongs to the probability of each theme and this keyword frequency of occurrences, calculates the similarity of the keyword vector in the keyword vector cluster corresponding to each theme that described object relates to; And value submodule, for getting the feature tag of the theme corresponding to similarity maximal value as described object.
According to an embodiment of the application, described module of setting up can comprise (not shown): choose submodule, in the user's historical behavior in section in those years, chooses sample according to predefined action number of times; Extract submodule, for obtaining the behavior type corresponding to keyword, the appearance identifier of keyword and keyword from each sample extracted; And training submodule, for using the combination of the behavior type corresponding to the appearance identifier of above-mentioned keyword, keyword and keyword as sample characteristics, using the feature of the object corresponding with sample characteristics as target, carry out the training of machine learning model, to obtain the current preference forecast model of user.
According to an embodiment of the application, described acquisition module 401 can comprise: obtain submodule, for obtaining the behavior type corresponding to keyword, keyword from described recent behavior; And obtain submodule, for whether occurring according to each behavior of keyword in described recent behavior, obtain the appearance identifier of keyword.
According to an embodiment of the application, described determination module 402 can comprise: use submodule, for occurring that the combination of the behavior type corresponding to identifier and keyword is as feature using described keyword, keyword, using the feature of the object of the current preference of user as target, use the current preference forecast model of user set up in advance, obtain the probability that each feature belongs to the feature of each object; 3rd calculating sub module, the feature for the object belonging to the current preference of probability calculation user of the feature of each object according to each feature belongs to the probability of each feature tag; And label determination submodule, for getting the feature tag of the feature tag corresponding to maximum probability value as the object of the current preference of user.
According to an embodiment of the application, described adjusting module 403 can comprise: judge module, for according to the current preference of the user determined, judges to treat whether ranked object has the feature of the current preference of user; First processing module, for when having the feature of the current preference of user until ranked object, treats described in lifting that ranked object sorts; And second processing module, for when not having the feature of the current preference of user until ranked object, described in not promoting, treat the sequence of ranked object.
The function realized due to the device of the present embodiment is substantially corresponding to the embodiment of the method shown in earlier figures 1 to Fig. 3, therefore not detailed part in the description of the present embodiment, see the related description in previous embodiment, can not repeat at this.
In one typically configuration, computing equipment comprises one or more processor (CPU), input/output interface, network interface and internal memory.
Internal memory may comprise the volatile memory in computer-readable medium, and the forms such as random access memory (RAM) and/or Nonvolatile memory, as ROM (read-only memory) (ROM) or flash memory (flashRAM).Internal memory is the example of computer-readable medium.
Computer-readable medium comprises permanent and impermanency, removable and non-removable media can be stored to realize information by any method or technology.Information can be computer-readable instruction, data structure, the module of program or other data.The example of the storage medium of computing machine comprises, but be not limited to phase transition internal memory (PRAM), static RAM (SRAM), dynamic RAM (DRAM), the random access memory (RAM) of other types, ROM (read-only memory) (ROM), Electrically Erasable Read Only Memory (EEPROM), fast flash memory bank or other memory techniques, read-only optical disc ROM (read-only memory) (CD-ROM), digital versatile disc (DVD) or other optical memory, magnetic magnetic tape cassette, tape magnetic rigid disk stores or other magnetic storage apparatus or any other non-transmitting medium, can be used for storing the information can accessed by computing equipment.According to defining herein, computer-readable medium does not comprise temporary computer readable media (transitory media), as data-signal and the carrier wave of modulation.
Also it should be noted that, term " comprises ", " comprising " or its any other variant are intended to contain comprising of nonexcludability, thus make to comprise the process of a series of key element, method, commodity or equipment and not only comprise those key elements, but also comprise other key elements clearly do not listed, or also comprise by the intrinsic key element of this process, method, commodity or equipment.When not more restrictions, the key element limited by statement " comprising ... ", and be not precluded within process, method, commodity or the equipment comprising described key element and also there is other identical element.
It will be understood by those skilled in the art that the embodiment of the application can be provided as method, system or computer program.Therefore, the application can adopt the form of complete hardware embodiment, completely software implementation or the embodiment in conjunction with software and hardware aspect.And the application can adopt in one or more form wherein including the upper computer program implemented of computer-usable storage medium (including but not limited to magnetic disk memory, CD-ROM, optical memory etc.) of computer usable program code.
The foregoing is only the embodiment of the application, be not limited to the application.To those skilled in the art, the application can have various modifications and variations.Any amendment done within all spirit in the application and principle, equivalent replacement, improvement etc., within the right that all should be included in the application.
Claims (14)
1. to the method that object sorts, it is characterized in that, comprising:
There is the behavior type corresponding to identifier and keyword in the keyword that the recent behavior obtaining user relates to, keyword;
There is the behavior type corresponding to identifier and keyword according to described keyword, keyword, use the current preference forecast model of user set up in advance, determine the feature of the object of the current preference of user; And
According to the current preference of the user determined, adjust the ranking factor for the treatment of ranked object relevant to the current preference of described user, to affect the sequence treating ranked object.
2. method according to claim 1, is characterized in that, the current preference forecast model of described user is set up by following steps:
The feature tag of mark object; And
Obtain keyword according to the user's historical behavior in section in those years, behavior type that keyword occurs corresponding to identifier, keyword, carry out the model training of machine learning, set up the current preference forecast model of user.
3. method according to claim 2, is characterized in that, determines that the feature tag of object comprises:
User-keyword matrix is set up according to the keyword that the user's historical behavior in section in those years relates to;
According to described user-keyword matrix, form one or more cluster by carrying out cluster to keyword, the corresponding a kind of theme of each cluster, calculates the probability that each keyword belongs to each theme;
Obtain each keyword and the frequency of occurrences thereof, to obtain the keyword vector of object;
The each keyword related to according to each object belongs to probability and this keyword frequency of occurrences of each theme, calculates the similarity of the keyword vector in the keyword vector cluster corresponding to each theme that described object relates to; And
Get the feature tag of the theme corresponding to similarity maximal value as described object.
4. method according to claim 2, it is characterized in that, obtain keyword according to the user's historical behavior in section in those years, behavior type that keyword occurs corresponding to identifier, keyword, carry out the model training of machine learning, set up the current preference forecast model of user, comprising:
User's historical behavior in section in those years, choose sample according to predefined action number of times;
Keyword, the appearance identifier of keyword and the behavior type corresponding to keyword is extracted from each sample extracted; And
Using the combination of the behavior type corresponding to the appearance identifier of above-mentioned keyword, keyword and keyword as sample characteristics, using the feature of the object corresponding with sample characteristics as target, carry out the training of machine learning model, to obtain the current preference forecast model of user.
5. according to the method one of claim 1-4 Suo Shu, it is characterized in that, there is the behavior type corresponding to identifier and keyword in the keyword that the recent behavior obtaining user relates to, keyword, comprising:
The behavior type corresponding to keyword, keyword is obtained from described recent behavior; And
Whether occur according to each behavior of keyword in described recent behavior, obtain the appearance identifier of keyword.
6. according to the method one of claim 1-4 Suo Shu, it is characterized in that, according to the behavior type corresponding to the appearance identifier of described keyword, keyword and keyword, use the current preference forecast model of user, determine the feature of the object of the current preference of user, comprising:
Occur that the combination of the behavior type corresponding to identifier and keyword is as feature using described keyword, keyword, using the feature of the object of the current preference of user as target, use the current preference forecast model of user set up in advance, obtain the probability that each feature belongs to the feature of each object;
The feature belonging to the object of the current preference of probability calculation user of the feature of each object according to each feature belongs to the probability of each feature tag; And
Get the feature tag of the feature tag corresponding to maximum probability value as the object of the current preference of user.
7. according to the method one of claim 1-4 Suo Shu, it is characterized in that, according to the current preference of the user determined, adjust the ranking factor for the treatment of ranked object relevant to the current preference of described user, comprising:
According to the current preference of the user determined, judge to treat whether ranked object has the feature of the current preference of user;
When having the feature of the current preference of user until ranked object, described in lifting, treat that ranked object sorts; And
When not having the feature of the current preference of user until ranked object, described in not promoting, treat the sequence of ranked object.
8. to the device that object sorts, it is characterized in that, comprising:
Acquisition module, there is the behavior type corresponding to identifier and keyword in the keyword that the recent behavior obtaining user relates to, keyword;
Determination module, for there is the behavior type corresponding to identifier and keyword according to described keyword, keyword, uses the current preference forecast model of user set up in advance, determines the feature of the object of the current preference of user; And
Adjusting module, for according to the current preference of the user determined, adjusts the ranking factor for the treatment of ranked object relevant to the current preference of described user, to affect the sequence treating ranked object.
9. device according to claim 8, is characterized in that, the current preference forecast model of described user is by setting up with lower module:
Labeling module, for marking the feature tag of object; And
Set up module, for obtaining keyword according to the user's historical behavior in section in those years, there is corresponding to identifier, keyword behavior type in keyword, carries out the model training of machine learning, set up the current preference forecast model of user.
10. device according to claim 9, is characterized in that, described determination module comprises:
Set up submodule, user-keyword matrix set up in the keyword related to for the user's historical behavior in basis in those years section;
First calculating sub module, for according to described user-keyword matrix, forms one or more cluster by carrying out cluster to keyword, and the corresponding a kind of theme of each cluster, calculates the probability that each keyword belongs to each theme;
Obtain submodule, for obtaining each keyword and the frequency of occurrences thereof, to obtain the keyword vector of object;
Second calculating sub module, each keyword for relating to according to each object belongs to the probability of each theme and this keyword frequency of occurrences, calculates the similarity of the keyword vector in the keyword vector cluster corresponding to each theme that described object relates to; And
Value submodule, for getting the feature tag of the theme corresponding to similarity maximal value as described object.
11. devices according to claim 9, is characterized in that, described module of setting up comprises:
Choose submodule, in the user's historical behavior in section in those years, choose sample according to predefined action number of times;
Extract submodule, for obtaining the behavior type corresponding to keyword, the appearance identifier of keyword and keyword from each sample extracted; And
Training submodule, for using the combination of the behavior type corresponding to the appearance identifier of above-mentioned keyword, keyword and keyword as sample characteristics, using the feature of the object corresponding with sample characteristics as target, carry out the training of machine learning model, to obtain the current preference forecast model of user.
12. one of-11 described devices according to Claim 8, it is characterized in that, described acquisition module comprises:
Obtain submodule, for obtaining the behavior type corresponding to keyword, keyword from described recent behavior; And
Obtaining submodule, for whether occurring according to each behavior of keyword in described recent behavior, obtaining the appearance identifier of keyword.
13. one of-11 described devices according to Claim 8, it is characterized in that, described determination module comprises:
Use submodule, for occurring that the combination of the behavior type corresponding to identifier and keyword is as feature using described keyword, keyword, using the feature of the object of the current preference of user as target, use the current preference forecast model of user set up in advance, obtain the probability that each feature belongs to the feature of each object;
3rd calculating sub module, the feature for the object belonging to the current preference of probability calculation user of the feature of each object according to each feature belongs to the probability of each feature tag; And
Label determination submodule, for getting the feature tag of the feature tag corresponding to maximum probability value as the object of the current preference of user.
14. one of-11 described devices according to Claim 8, it is characterized in that, described adjusting module comprises:
Judge module, for according to the current preference of the user determined, judges to treat whether ranked object has the feature of the current preference of user;
First processing module, for when having the feature of the current preference of user until ranked object, treats described in lifting that ranked object sorts; And
Second processing module, for when not having the feature of the current preference of user until ranked object, treats the sequence of ranked object described in not promoting.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410112060.2A CN104951441A (en) | 2014-03-24 | 2014-03-24 | Method and device for sequencing objects |
| HK15111890.1A HK1211111A1 (en) | 2014-03-24 | 2015-12-03 | Method and apparatus for sorting objects |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410112060.2A CN104951441A (en) | 2014-03-24 | 2014-03-24 | Method and device for sequencing objects |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104951441A true CN104951441A (en) | 2015-09-30 |
Family
ID=54166104
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410112060.2A Pending CN104951441A (en) | 2014-03-24 | 2014-03-24 | Method and device for sequencing objects |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN104951441A (en) |
| HK (1) | HK1211111A1 (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105718565A (en) * | 2016-01-20 | 2016-06-29 | 北京京东尚科信息技术有限公司 | Data warehouse model construction method and construction apparatus |
| CN106095842A (en) * | 2016-06-01 | 2016-11-09 | 腾讯科技（深圳）有限公司 | Online course searching method and device |
| CN107729422A (en) * | 2017-09-27 | 2018-02-23 | 广州市万表科技股份有限公司 | A kind of personality method of testing and system based on commodity identification |
| CN108304407A (en) * | 2017-01-12 | 2018-07-20 | 阿里巴巴集团控股有限公司 | A kind of method and system carrying out object order |
| WO2019072128A1 (en) * | 2017-10-09 | 2019-04-18 | 北京京东尚科信息技术有限公司 | Object identification method and system therefor |
| CN109785034A (en) * | 2018-11-13 | 2019-05-21 | 北京码牛科技有限公司 | User's portrait generation method, device, electronic equipment and computer-readable medium |
| CN110019796A (en) * | 2017-11-10 | 2019-07-16 | 北京信息科技大学 | A kind of user version information analysis method and device |
| CN110188282A (en) * | 2019-05-31 | 2019-08-30 | 广州华多网络科技有限公司 | The method and apparatus for determining preference-score |
| CN114282119A (en) * | 2022-03-03 | 2022-04-05 | 深圳昊通技术有限公司 | Scientific and technological information resource retrieval method and system based on heterogeneous information network |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101055587A (en) * | 2007-05-25 | 2007-10-17 | 清华大学 | Search engine retrieving result reordering method based on user behavior information |
| CN101079824A (en) * | 2006-06-15 | 2007-11-28 | 腾讯科技（深圳）有限公司 | A generation system and method for user interest preference vector |
| CN101923545A (en) * | 2009-06-15 | 2010-12-22 | 北京百分通联传媒技术有限公司 | Method for recommending personalized information |
| CN102663627A (en) * | 2012-04-26 | 2012-09-12 | 焦点科技股份有限公司 | Personalized recommendation method |
| CN102999634A (en) * | 2012-12-18 | 2013-03-27 | 百度在线网络技术（北京）有限公司 | User navigation recommending method and system based on browser data as well as cloud server |
| CN103226618A (en) * | 2013-05-21 | 2013-07-31 | 焦点科技股份有限公司 | Related word extracting method and system based on data market mining |
| CN104143001A (en) * | 2014-08-01 | 2014-11-12 | 百度在线网络技术（北京）有限公司 | Search term recommending method and device |
-
2014
- 2014-03-24 CN CN201410112060.2A patent/CN104951441A/en active Pending
-
2015
- 2015-12-03 HK HK15111890.1A patent/HK1211111A1/en unknown
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101079824A (en) * | 2006-06-15 | 2007-11-28 | 腾讯科技（深圳）有限公司 | A generation system and method for user interest preference vector |
| CN101055587A (en) * | 2007-05-25 | 2007-10-17 | 清华大学 | Search engine retrieving result reordering method based on user behavior information |
| CN101923545A (en) * | 2009-06-15 | 2010-12-22 | 北京百分通联传媒技术有限公司 | Method for recommending personalized information |
| CN102663627A (en) * | 2012-04-26 | 2012-09-12 | 焦点科技股份有限公司 | Personalized recommendation method |
| CN102999634A (en) * | 2012-12-18 | 2013-03-27 | 百度在线网络技术（北京）有限公司 | User navigation recommending method and system based on browser data as well as cloud server |
| CN103226618A (en) * | 2013-05-21 | 2013-07-31 | 焦点科技股份有限公司 | Related word extracting method and system based on data market mining |
| CN104143001A (en) * | 2014-08-01 | 2014-11-12 | 百度在线网络技术（北京）有限公司 | Search term recommending method and device |
Non-Patent Citations (2)
| Title |
|---|
| 车高营 等: ""基于序列模式的用户浏览行为提取与分析"", 《计算机技术与发展》 * |
| 郭韦昱: ""基于用户行为分析的个性化推荐系统设计与实现"", 《中国优秀硕士学位论文全文数据库·信息科技辑》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105718565A (en) * | 2016-01-20 | 2016-06-29 | 北京京东尚科信息技术有限公司 | Data warehouse model construction method and construction apparatus |
| CN105718565B (en) * | 2016-01-20 | 2019-07-02 | 北京京东尚科信息技术有限公司 | The construction method and construction device of data warehouse model |
| CN106095842A (en) * | 2016-06-01 | 2016-11-09 | 腾讯科技（深圳）有限公司 | Online course searching method and device |
| CN106095842B (en) * | 2016-06-01 | 2021-06-25 | 腾讯科技（深圳）有限公司 | Online course searching method and device |
| CN108304407A (en) * | 2017-01-12 | 2018-07-20 | 阿里巴巴集团控股有限公司 | A kind of method and system carrying out object order |
| CN108304407B (en) * | 2017-01-12 | 2022-02-25 | 阿里巴巴集团控股有限公司 | Method and system for sequencing objects |
| CN107729422A (en) * | 2017-09-27 | 2018-02-23 | 广州市万表科技股份有限公司 | A kind of personality method of testing and system based on commodity identification |
| WO2019072128A1 (en) * | 2017-10-09 | 2019-04-18 | 北京京东尚科信息技术有限公司 | Object identification method and system therefor |
| CN110019796A (en) * | 2017-11-10 | 2019-07-16 | 北京信息科技大学 | A kind of user version information analysis method and device |
| CN109785034A (en) * | 2018-11-13 | 2019-05-21 | 北京码牛科技有限公司 | User's portrait generation method, device, electronic equipment and computer-readable medium |
| CN110188282A (en) * | 2019-05-31 | 2019-08-30 | 广州华多网络科技有限公司 | The method and apparatus for determining preference-score |
| CN114282119A (en) * | 2022-03-03 | 2022-04-05 | 深圳昊通技术有限公司 | Scientific and technological information resource retrieval method and system based on heterogeneous information network |
Also Published As
| Publication number | Publication date |
|---|---|
| HK1211111A1 (en) | 2016-05-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104951441A (en) | Method and device for sequencing objects | |
| CN104679743B (en) | A kind of method and device of the preference pattern of determining user | |
| CN104899273B (en) | A kind of Web Personalization method based on topic and relative entropy | |
| CN104778176A (en) | Data search processing method and device | |
| CN103020164B (en) | Semantic search method based on multi-semantic analysis and personalized sequencing | |
| CN102193936B (en) | Data classification method and device | |
| CN104199857B (en) | A kind of tax document hierarchy classification method based on multi-tag classification | |
| CN105893609B (en) | A kind of mobile APP recommended method based on weighted blend | |
| CN104866474B (en) | Individuation data searching method and device | |
| CN102279851B (en) | Intelligent navigation method, device and system | |
| CN107632984A (en) | A kind of cluster data table shows methods, devices and systems | |
| CN109800350A (en) | A kind of Personalize News recommended method and system, storage medium | |
| CN105718579A (en) | Information push method based on internet-surfing log mining and user activity recognition | |
| Seret et al. | A new SOM-based method for profile generation: Theory and an application in direct marketing | |
| CN104035927A (en) | User behavior-based search method and system | |
| CN105302810A (en) | Information search method and apparatus | |
| CN102521248A (en) | Network user classification method and device | |
| CN103473317A (en) | Method and equipment for extracting keywords | |
| CN105069077A (en) | Search method and device | |
| CN101706812B (en) | Method and device for searching documents | |
| CN111860981B (en) | Enterprise national industry category prediction method and system based on LSTM deep learning | |
| CN108959580A (en) | A kind of optimization method and system of label data | |
| Rao et al. | A machine learning approach to classify news articles based on location | |
| Wang et al. | Multi-task multimodal learning for disaster situation assessment | |
| CN105677825A (en) | Analysis method for client browsing operation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| REG | Reference to a national code |
Ref country code: HK Ref legal event code: DE Ref document number: 1211111 Country of ref document: HK |
|
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20150930 |
|
| REG | Reference to a national code |
Ref country code: HK Ref legal event code: WD Ref document number: 1211111 Country of ref document: HK |