CN109558544A - Sort method and device, server and storage medium - Google Patents
Sort method and device, server and storage medium Download PDFInfo
- Publication number
- CN109558544A CN109558544A CN201811520392.9A CN201811520392A CN109558544A CN 109558544 A CN109558544 A CN 109558544A CN 201811520392 A CN201811520392 A CN 201811520392A CN 109558544 A CN109558544 A CN 109558544A
- Authority
- CN
- China
- Prior art keywords
- real
- data
- model
- time
- lower single
- 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
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present embodiments relate to technical field of information recommendation, a kind of sort method and device, server and storage medium are disclosed.This method comprises: generating real-time training sample；The model parameter of lower single model and click model is updated according to real-time training sample；Obtain real-time stream；Lower single model and click model according to real-time stream and based on update are predicted to obtain ranking results.Order models are updated the present invention is based on real-time training sample and obtain ranking results, so that ranking results can reflect the interests change of user or group in real time, are significantly improved the wish that user clicks purchase, are improved sequence performance.
Description
Technical field
The present invention relates to technical field of information recommendation, in particular to a kind of sort method and device, server and storage are situated between
Matter.
Background technique
When user is intended to clear, we lead to Usual Search Engines to solve the problem of information overload of Internet era, but
When the intention of user is indefinite or is difficult to use clearly semantic meaning representation, search engine is with regard to helpless.At this point, by user
Its intention of the analysis and understanding of behavior is its push personalization as a result, becoming as a kind of better choice.Existing take-away scene
In, background server has a special order models that can carry out intelligence according to information such as the businessmans and food product that user chooses
Recommend sequence, the difference of businessman and food product that businesses lists on front-end software interface are chosen according to user and more
Newly.
Inventor find the relevant technologies the prior art has at least the following problems: existing order models utilize historical data, such as by
The behavioral data (businessman, food product for choosing etc.) of the user of the previous day is updated order models, so as to subsequent, such as
Recommend the ranking results (such as businesses lists) updated for user within two days.Therefore, there are certain lag for existing sort method
Property.
Summary of the invention
Embodiment of the present invention is designed to provide a kind of sort method and device, server and storage medium, is based on
Real-time training sample updates order models and obtains ranking results, so that ranking results can reflect user or group in real time
The interests change of body significantly improves the wish that user clicks purchase, improves sequence performance.
In order to solve the above technical problems, embodiments of the present invention provide a kind of sort method, comprising:
Generate real-time training sample；
The model parameter of lower single model and click model is updated according to the real-time training sample；
Obtain real-time stream；
Lower single model and click model according to the real-time stream and based on update are predicted to obtain sequence knot
Fruit.
Embodiments of the present invention additionally provide a kind of collator, comprising:
Generation module, for generating real-time training sample；
Update module, for updating the model parameter of lower single model and click model according to the real-time training sample；
Module is obtained, for obtaining real-time stream；
Prediction module, for according to the real-time stream and based on update lower single model and click model it is pre-
Measure ranking results.
Embodiments of the present invention additionally provide a kind of server, comprising: memory and processor, memory storage calculate
Machine program, processor run the computer program to realize:
Generate real-time training sample；
The model parameter of lower single model and click model is updated according to the real-time training sample；
Obtain real-time stream；
Lower single model and click model according to the real-time stream and based on update are predicted to obtain sequence knot
Fruit.
Embodiments of the present invention additionally provide a kind of storage medium, for storing computer-readable program, the calculating
Machine readable program is used to execute sort method as described above for computer.
Embodiment of the present invention in terms of existing technologies, generates real-time training sample to update lower single model and point
The model parameter of model is hit, and predicts to obtain ranking results based on real-time stream, lower single model of update and click model.
Since real-time training sample can embody the interests change of user or group in time, the ranking results based on it can be real
The interests change of Shi Fanying user or group improves sequence performance so as to significantly improve the wish that user clicks purchase,
And it is more acurrate based on the ranking results that click model and lower single model obtain prediction.
It is described to generate real-time training sample as one embodiment, it specifically includes: collecting sample data；To the institute of acquisition
Sample data is stated to be screened to obtain candidate samples data；The candidate samples data include: business diary and user behavior
Data；It polymerize the business diary and the user behavior data to obtain real-time training sample.
As one embodiment, the business diary includes exposure frequency, and the user behavior data includes number of clicks
And lower single number；The sample data of described pair of acquisition is screened to obtain candidate samples data, is specifically included: being determined institute
State exposure frequency in business diary in corresponding user behavior data number of clicks and the ratio between lower single number whether
Within the scope of preset ratio；If within the scope of the preset ratio, by the business diary and its corresponding user behavior data
As the candidate samples data.
As one embodiment, the sample data of described pair of acquisition is screened to obtain candidate samples data, specifically
It comprises determining that the click in the user behavior data and whether the corresponding display location that places an order is less than preset value, if
Less than preset value, then using the user behavior data and its corresponding business diary as the candidate samples data.
As one embodiment, the sample data of described pair of acquisition is screened to obtain candidate samples data, specifically
Comprise determining that whether the sample data is invalid data；If invalid data, then the invalid data of preset ratio is abandoned.
As one embodiment, the sample data of described pair of acquisition is screened to obtain candidate samples data, specifically
It include: visitor's unit price in the detection sample data；Objective unit price is greater than to the sample data for presetting unit price as the candidate sample
Notebook data.
As one embodiment, further includes: the purchasing price of prediction user in real time；It is described according to the real-time stream simultaneously
Lower single model and click model based on update are predicted to obtain after ranking results, further includes: according to prediction
Purchasing price is adjusted the ranking results.
It is described that the ranking results are adjusted according to the purchasing price of prediction as one embodiment, specifically
It include: the contingent order price for obtaining the trade company in the ranking results；According to the purchasing price and the ginseng of prediction
Monovalent lattice are examined and corrected to be adjusted the ranking results.
As one embodiment, the lower single model and click according to the real-time stream and based on update
Model prediction obtains after ranking results, further includes: obtains reserved sorting position；It is adjusted based on the reserved sorting position
The ranking results.
As one embodiment, the real-time stream includes: browsing pages and click.
As one embodiment, the model that lower single model and click model are updated according to the real-time training sample
Parameter specifically includes: being trained to obtain respectively to lower single model and click model according to the real-time training sample
New model parameter；Stability is carried out to the new model parameter and verifies to obtain the candidate family that stability meets preset condition
Parameter；The candidate family parameter is periodically loaded to lower single model and click model.
Detailed description of the invention
Fig. 1 is the flow chart of first embodiment sort method according to the present invention；
Fig. 2 is the flow chart of second embodiment sort method according to the present invention；
Fig. 3 is the flow chart of third embodiment sort method according to the present invention；
Fig. 4 is the flow chart of the 4th embodiment sort method according to the present invention；
Fig. 5 is the structural schematic diagram of the 5th embodiment collator according to the present invention；
Fig. 6 is the structural schematic diagram of sixth embodiment server according to the present invention.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with attached drawing to the present invention
Each embodiment be explained in detail.However, it will be understood by those skilled in the art that in each embodiment party of the present invention
In formula, many technical details are proposed in order to make reader more fully understand the present invention.But even if without these technical details
And various changes and modifications based on the following respective embodiments, claimed technical solution of the invention also may be implemented.
The first embodiment of the present invention is related to a kind of sort methods, can be applied to server side.This method comprises: raw
At real-time training sample, the model parameter of lower single model and click model is updated according to real-time training sample, obtains number in real time
According to stream；And lower single model according to real-time stream and based on update and click model are predicted to obtain ranking results.This hair
Bright embodiment in terms of existing technologies, generates real-time training sample to update the model of lower single model and click model
Parameter, and predict to obtain ranking results based on real-time stream, lower single model of update and click model.Due to training in real time
Sample can embody the interests change of user or group in time, thus based on its ranking results can reflect in real time user or
The interests change of person group improves sequence performance so as to significantly improve the wish that user clicks purchase, and is based on clicking
The ranking results that model and lower single model obtain prediction are more acurrate.
The realization details of the sort method of present embodiment is specifically described below, the following contents is only for convenience of reason
The realization details provided is provided, the necessary of this programme is not implemented.Referring to FIG. 1, the sort method of present embodiment specifically includes
Step 101 is to step 104.
Step 101: generating real-time training sample.
Step 101 may include: real-time acquisition and tissue samples data, be polymerize to obtain reality to the sample data of tissue
When training sample.Wherein, sample data includes: the business diary of rear end and the user behavior data of front end feedback.Instruction in real time
Practicing sample is by the business diary that rear end is launched in real time and the model training that front end is associated its feedback data
Sample.Wherein, business diary is for example including the ranking results pushed in real time from the background, such as recommends businesses lists.User behavior number
According to the Real-time Feedback for referring to business diary of the user to push, such as purchase Business Information and food product information, click businessman's letter
The behavioral data of breath and the evaluation information etc. of food product information, businessman and food product.In practical applications, sample data can be with
According to the second, the real-time of grade is acquired, i.e., a business diary and the acquisition of user behavior data are completed in the several seconds.Wherein
The data source of business diary can be server, and user behavior data source can be mobile terminal and PC
The user terminal of (Personal Computer, PC) etc..Wherein, backend services log and user behavior data are to acquire respectively
It obtains.Present embodiment is not particularly limited the acquisition mode of business diary and user behavior data.Acquisition in real time
Sample data can be cached in case generating real-time training sample.Pass through the business diary that will be acquired and user behavior number
According to it is corresponding associate and (polymerize) real-time training sample can be obtained, thus for lower single model and click model in real time more
Newly provide precondition.
The acquisition of sample data is illustrated below as follows:
The data for determining acquisition first are business diary or user behavior data, if business diary, it is determined that acquisition
Data whether there is paging, paging if it exists caches after then merging the sample data of paging；Paging if it does not exist,
The directly data of caching acquisition.If the data of acquisition are user behavior data, the user behavior data of acquisition is cached, then really
It is fixed whether to be cached with business diary corresponding with the user behavior data acquired, if it does not exist with the user behavior data pair of caching
The business diary answered then abandons the user behavior data of acquisition, if it exists business corresponding with the user behavior data of caching
Log then retains the user behavior data of acquisition.
Specifically, simultaneously tissue samples data can be acquired according to the number of sequencing requests.Wherein, rear end is directed to every minor sort
Request can generate a unique rankid, represent unique a certain minor sort as a result, for example unique some recommendation businessman's column
Table information.It in present embodiment, can be acquired according to rankid and tissue samples data, i.e., according to each ranking results by business
Log and user behavior data are mapped.It is so without being limited thereto, such as can also be adopted according to certain time for same user
Collect simultaneously tissue samples data, such as the corresponding sample data of multiple rankid can be acquired tissue together.
Wherein, the recommendation businesses lists information obtained for a sequencing requests, in capturing service log, due to one
Ranking results may show several times, although generally can first show one for example, user nearby might have tens dining rooms
It is a recommend businesses lists a part, such as show ranking near more than preceding 10 a businessmans, and user glide continue browsing should
When recommending businesses lists, the remainder for showing the ranking results will continue to.And in capturing service log, repeatedly show
Content can be slit into sheets caching.Therefore, in organization business log, determine collected multiple data there are when paging, such as
When multiple data correspond to same rankid, cached after just merging corresponding to the data of the paging of same rankid, i.e., it will be same
It is cached after the corresponding multistage business diary splicing of rankid.
It, can be within a preset time when determining whether to be cached with business diary corresponding with the user behavior data of acquisition
Determine whether to be polled to corresponding business diary, if not being polled to corresponding business diary, abandons the sample data of acquisition, if
It is polled to corresponding business diary, then retains the user behavior data of acquisition.In collecting sample data, same rankid is corresponding
User behavior data and the acquisition time of business diary there may be successive, it is also possible to user behavior data pair can not be collected
The business diary answered.Therefore, it for the user behavior data for not yet collecting corresponding business diary, can first be cached,
Then whether automatic regular polling collects corresponding business diary within a preset time.Wherein, preset time is, for example, 1 second, so
It is without being limited thereto.If reach preset time and be not polled to corresponding business diary, due to not collecting and the user behavior
The corresponding business diary of data, so abandon the user behavior data.If being polled to corresponding business diary within a preset time,
Or collect corresponding business diary prior to user behavior data, then retain the user behavior data in case subsequent polymerization generates
Real-time training sample.
Step 102: the model parameter of lower single model and click model is updated according to real-time training sample.
Click model and lower single model are trained and respectively obtain the new of the two according to real-time training sample online
Model parameter, and new model parameter is loaded onto the lower single model for being used to respond online sequencing requests and click in real time
Model.
Step 103: obtaining real-time stream.
Wherein, real-time stream includes but is not limited to: browsing pages and click etc..
Step 104: lower single model and click model according to real-time stream and based on update are predicted to obtain sequence knot
Fruit.
The sequencing requests of user are obtained in real time, and based on the updated click model of parameter and lower single model to user
Intention is predicted to obtain ranking results.For example, being distinguished by click model and lower single model pre- when predicting businessman
The corresponding score of some businessman is measured, then weighted sum obtains the gross score of the businessman, is then ranked up according to gross score
The businesses lists information of recommendation is pushed on user terminal by the businesses lists information that recommendation can be obtained, and user can be timely
Know that current quality is preferable, the higher businessman of temperature and food product etc..
For example, if make a reservation number of certain dining room by the previous day is seldom, therefore the meal is not recommended to user
The Room, but trendy food product is released due to the dining room suddenly and makes number sharp increase of making a reservation, then present embodiment due to that can adopt in real time
Collect business diary and user behavior data and construct real-time training sample, update model parameter, so that updated model can be with
Recommend this family dining room in time for user.
Present embodiment can acquire user in real time and generate real-time training sample to the feedback data of online ranking results,
And the model parameter of lower single model and click model is updated according to real-time training sample, while obtaining real-time stream, it is needing
When sorting, lower single model and click model based on real-time update predict ranking results.It can be seen that present embodiment
Online ranking results, the acquisition of sample data and the training of model parameter form closure data flow in real time in sort method,
The sort method updated to the order models real-time online of the lower single model of realization and click model etc..Present embodiment passes through
Real-time training sample is generated, existing order models are updated using newest sample increment iterative, so that order models timely feedback sample
This changes in distribution improves model sequence performance.
Second embodiment of the present invention is related to a kind of sort method, as shown in Fig. 2, present embodiment includes step 201
To step 206.
Step 201: collecting sample data.
The acquisition of sample data and tissue can refer to aforementioned embodiments, and details are not described herein again.
Step 202: the sample data of acquisition being screened to obtain candidate samples data.
Wherein, step 202 can using it is following any one or combinations thereof by the way of obtain candidate samples data, this right reality
Mode is applied to be not particularly limited the screening mode of candidate samples data.
Candidate samples data screening mode one:
Wherein, business diary includes exposure frequency, and user behavior data includes number of clicks and lower single number.It can incite somebody to action
The corresponding business diary of same rankid and user behavior data are associated, such as by the exposure frequency and use in business diary
Number of clicks and the lower corresponding association of single number in the behavioral data of family.Determine exposure frequency in business diary and corresponding use
Whether the ratio between number of clicks and lower single number in the behavioral data of family are within the scope of preset ratio, if in preset ratio range
It is interior, then using business diary and its corresponding user behavior data as candidate samples data, if not within the scope of preset ratio,
Abandon the sample data.Wherein, preset ratio range be, for example, (110~90): 10:1, exposure frequency, number of clicks and under
The ratio of single number is preferably 100:10:1, so without being limited thereto.In this way, can prevent real-time training sample from relatively large deviation occur,
It is consistent with the probability distribution of truthful data as far as possible, to improve the accuracy of click model and lower single model.
Candidate samples data screening mode two:
It determines the click in user behavior data and whether the corresponding display location that places an order is less than preset value, if small
In preset value, then using user behavior data and its corresponding business diary as candidate samples data, while display location is abandoned
More than or equal to the user behavior data of preset value and its corresponding business diary.Preset value is, for example, 100, is so not limited to
This.It may be cheating data that the excessive user of display location, which clicks and places an order, influences click model and point single model updates
Accuracy.
Candidate samples data screening mode three:
Determine whether sample data is invalid data, if invalid data, then abandons the invalid data of preset ratio.In vain
Data can be the sample data that no user clicks.Such as 20% invalid data can be selected at random using pseudo-random function
It discards.In this way, the real-time of model training not only can be improved, the precision of model can also be improved.
Candidate samples data screening mode four:
Visitor's unit price in sample data is detected, objective unit price is greater than the sample data for presetting unit price as candidate samples number
According to discarding visitor's unit price, which is less than or equal to, presets monovalent sample data.Objective unit price is the order amount of money of the single order of user.This
Sample can be such that the training of model focuses on the higher order of price.
Step 203: business diary and user behavior data being polymerize to obtain real-time training sample.
Step 204: the model parameter of lower single model and click model is updated according to real-time training sample.
Step 205: obtaining real-time stream.
Step 206: lower single model and click model according to real-time stream and based on update are predicted to obtain sequence knot
Fruit.
Present embodiment can obtain candidate samples by one or more kinds of screening modes compared with aforementioned embodiments
Data, to be conducive to improve real-time, the accuracy of model modification.
Third embodiment of the present invention is related to a kind of sort method, as shown in figure 3, present embodiment includes step 301
To step 306.
Step 301: generating real-time training sample.
Step 302: the model parameter of lower single model and click model is updated according to real-time training sample.
Step 303: obtaining real-time stream.
Step 304: lower single model and click model according to real-time stream and based on update are predicted to obtain sequence knot
Fruit.
Step 305: obtaining reserved sorting position.
Wherein, reserved sorting position includes but is not limited to 3,4,7,8,9.Reserved sorting position can give dealer
Family is locked as advertisement position.
Step 306: ranking results are adjusted based on reserved sorting position.
It should be noted that in one example, can be combined by individually or by a manner of reserved sorting position
Mode ranking results are adjusted.Specifically, the purchasing price of user can be predicted in real time, after the step 304, according to
The purchasing price of prediction is adjusted ranking results.
The mode that ranking results are adjusted is illustrated below to according to the purchasing price of prediction below: can basis
The affiliated section of the purchasing price of the user of prediction adjusts ranking results, for example, predict the purchasing price of user 20~25 yuan it
Between when, the weight of the appropriate trade company between 20~25 yuan that increases food product price.It can certainly be according to the purchasing price of user
The consumption grade of user is divided, for example, by the consumption grade classification of user be 5 grades, wherein the 2nd grade is for example
Medium consumption grade is corresponded to, after the consumption grade for determining user is medium consumption grade, increasing consumption grade is medium disappear
Take the weight of the trade company of grade.
Further, it is also possible in the ranking results of prediction or the base of increase consumption grade and the weight of the matched trade company of user
On plinth, ranking results are further finely tuned.Specifically, the contingent order price for obtaining the trade company in ranking results, according to prediction
Purchasing price and contingent order price ranking results are adjusted.Wherein, the contingent order price of trade company is similar to trade company
Consumption grade, such as the consumption of some middle and high end trade companies is higher ranked, and during the consumption grade of some fast food brands is generally
Ranking results can also be finely adjusted according to the consumption grade sequence from high to low of trade company Deng, present embodiment.Such as
When including 3 middle and high end trade companies and 20 fast food trade companies in the merchant list of recommendation, middle and high end trade company can be properly increased
Display location.
Therefore, present embodiment, can be by being finely adjusted ranking results, thus favorably compared with aforementioned embodiments
In the income for improving platform.
4th embodiment of the invention is related to a kind of sort method, as shown in figure 4, present embodiment includes step 401
To step 406.
Step 401: generating real-time training sample.
Step 402: lower single model and click model being trained to obtain new mould respectively according to real-time training sample
Shape parameter.
Wherein, the real-time training sample of generation can be put into message queue, after on-line training engine receives message, to placing an order
Model and click model are trained to obtain new model parameter respectively, and snapshot stores new model parameter.Wherein, single-mode
Type and the every update of click model can once respectively obtain the new model parameter of a set of (an also referred to as version), and snapshot is deposited
The model parameter of Chu Xin is to store to the model parameter of lower single model and the respective new version of click model.
Step 403: stability being carried out to new model parameter and verifies to obtain the candidate family that stability meets preset condition
Parameter.
In some cases, the fluctuation of the ranking results based on new model parameter can be bigger, i.e., new model parameter
Stability may be poor.Therefore, after obtaining new model parameter, it can be verified to determine its stability
Whether preset condition is met, as candidate family parameter if its stability meets preset condition, if its stability is discontented
Sufficient preset condition can then abandon the model parameter.The stability verification of model parameter can be tied by the sequence of comparison front and back
The mode of the difference of fruit is realized.For example, A, B, C, D businessman arrange respectively in the ranking results of the model parameter based on previous version
At first four, and in the ranking results of the model parameter based on current version, A, B, C, D businessman do not appear in businessman and recommend column
In table, then it can determine that the fluctuation of the model parameter of current version is larger, be unsatisfactory for preset condition.Present embodiment is for candidate mould
The determination of stability mode of shape parameter is not particularly limited.
Step 404: periodically loading candidate family parameter to lower single model and click model.
Nearest candidate family parameter can be chosen in the candidate family parameter generated in real time is loaded onto online place an order
In model and click model, and the load of model parameter can be carried out according to predetermined period, such as load one in every 10 seconds
Secondary model parameter.
Step 405: obtaining real-time stream.
Step 406: lower single model and click model according to real-time stream and based on update are predicted to obtain sequence knot
Fruit.
Present embodiment, by carrying out stability verification to new model parameter, can be kept away compared with aforementioned embodiments
Exempt from ranking results fluctuation, advantageously ensures that the quality of sequence.
5th embodiment of the invention is related to a kind of collator, as shown in figure 5, the collator of present embodiment
500 include:
Generation module 501, for generating real-time training sample；
Update module 502, for updating the model parameter of lower single model and click model according to real-time training sample；
Module 503 is obtained, for obtaining real-time stream；And
Prediction module 504, for according to real-time stream and based on update lower single model and click model measure in advance
To ranking results.
Wherein, generation module 501 is specifically used for acquisition and tissue samples data in real time, gathers to the sample data of tissue
Conjunction obtains real-time training sample.Wherein, sample data includes: the business diary of rear end and the user behavior number of front end feedback
According to.Real-time training sample is to be associated business diary that rear end is launched in real time and front end to its feedback data
Model training sample.Wherein, business diary is for example including the ranking results pushed in real time from the background, such as recommends businesses lists.With
Family behavioral data refers to the Real-time Feedback of business diary of the user to push, such as purchase Business Information and food product information, point
Hit the behavioral data of evaluation information of Business Information and food product information, businessman and food product etc..In practical applications, sample
Data can be acquired according to the real-time of second grade, i.e., adopting for a business diary and user behavior data is completed in the several seconds
Collection.Wherein the data source of business diary can be server, and user behavior data source can be mobile terminal and personal electricity
The user terminal of brain (Personal Computer, PC) etc..Wherein, backend services log and user behavior data are to adopt respectively
What collection obtained.Present embodiment is not particularly limited the acquisition mode of business diary and user behavior data.It adopts in real time
The sample data of collection can be cached in case generating real-time training sample.Pass through the business diary that will be acquired and user behavior
Data are corresponding to associate and (polymerize) real-time training sample can be obtained, thus for the real-time of lower single model and click model
Update provides precondition.
The acquisition of sample data is illustrated below as follows:
The data for determining acquisition first are business diary or user behavior data, if business diary, it is determined that acquisition
Data whether there is paging, paging if it exists caches after then merging the sample data of paging；Paging if it does not exist,
The directly data of caching acquisition.If the data of acquisition are user behavior data, the user behavior data of acquisition is cached, then really
It is fixed whether to be cached with business diary corresponding with the user behavior data acquired, if it does not exist with the user behavior data pair of caching
The business diary answered then abandons the user behavior data of acquisition, if it exists business corresponding with the user behavior data of caching
Log then retains the user behavior data of acquisition.
Specifically, simultaneously tissue samples data can be acquired according to the number of sequencing requests.Wherein, rear end is directed to every minor sort
Request can generate a unique rankid, represent unique a certain minor sort as a result, for example unique some recommendation businessman's column
Table information.It in present embodiment, can be acquired according to rankid and tissue samples data, i.e., according to each ranking results by business
Log and user behavior data are mapped.It is so without being limited thereto, such as can also be adopted according to certain time for same user
Collect simultaneously tissue samples data, such as the corresponding sample data of multiple rankid can be acquired tissue together.
Wherein, the recommendation businesses lists information obtained for a sequencing requests, in capturing service log, due to one
Ranking results may show several times, although generally can first show one for example, user nearby might have tens dining rooms
It is a recommend businesses lists a part, such as show ranking near more than preceding 10 a businessmans, and user glide continue browsing should
When recommending businesses lists, the remainder for showing the ranking results will continue to.And in capturing service log, repeatedly show
Content can be slit into sheets caching.Therefore, in organization business log, determine collected multiple data there are when paging, such as
When multiple data correspond to same rankid, cached after just merging corresponding to the data of the paging of same rankid, i.e., it will be same
It is cached after the corresponding multistage business diary splicing of rankid.
It, can be within a preset time when determining whether to be cached with business diary corresponding with the user behavior data of acquisition
Determine whether to be polled to corresponding business diary, if not being polled to corresponding business diary, abandons the sample data of acquisition, if
It is polled to corresponding business diary, then retains the user behavior data of acquisition.In collecting sample data, same rankid is corresponding
User behavior data and the acquisition time of business diary there may be successive, it is also possible to user behavior data pair can not be collected
The business diary answered.Therefore, it for the user behavior data for not yet collecting corresponding business diary, can first be cached,
Then whether automatic regular polling collects corresponding business diary within a preset time.Wherein, preset time is, for example, 1 second, so
It is without being limited thereto.If reach preset time and be not polled to corresponding business diary, due to not collecting and the user behavior
The corresponding business diary of data, so abandon the user behavior data.If being polled to corresponding business diary within a preset time,
Or collect corresponding business diary prior to user behavior data, then retain the user behavior data in case subsequent polymerization generates
Real-time training sample.
Update module 502 is specifically used for online being trained click model and lower single model according to real-time training sample
The new model parameter of the two is respectively obtained, and new model parameter is loaded onto real time and is used to respond online sequencing requests
Lower list model and click model.
Prediction module 504 is specifically used for obtaining the sequencing requests of user in real time, and is based on the updated click model of parameter
And lower single model is intended to be predicted to obtain ranking results to user.Wherein, sequencing requests can be from the real time data of acquisition
It is extracted in stream, real-time stream includes but is not limited to: browsing pages and click etc..
Prediction module 504 can be predicted to obtain respectively when predicting businessman by click model and lower single model
The corresponding score of some businessman, then weighted sum obtain the gross score of the businessman, and being then ranked up according to gross score can obtain
To the businesses lists information of recommendation, the businesses lists information of recommendation is pushed on user terminal, user can timely learning work as
Preceding quality is preferable, the higher businessman of temperature and food product etc..
Present embodiment generation module can acquire user in real time and generate to the feedback data of online ranking results in real time
Training sample updates the model parameter of lower single model and click model by update module, simultaneously according to real-time training sample
Real-time stream is obtained by obtaining module, when needing to sort, lower single model and point of the prediction module based on real-time update
Hit model prediction ranking results.It can be seen that in the collator of present embodiment online ranking results, sample data acquisition
And the training of model parameter forms closure data flow in real time, to realize the sequence mould of lower single model and click model etc.
Type real-time online updates.Present embodiment has row by generating real-time training sample, using the update of newest sample increment iterative
Sequence model improves model sequence performance so that order models timely feedback sample distribution variation.
In one example, generation module 501 is specifically used for collecting sample data, screens to the sample data of acquisition
Obtain candidate samples data；Candidate samples data include: business diary and user behavior data；By business diary and user
Behavioral data polymerize to obtain real-time training sample.
In some instances, generation module 501 can obtain candidate samples one of in the following ways or any combination thereof
Data:
Candidate samples data screening mode one:
Wherein, business diary includes exposure frequency, and user behavior data includes number of clicks and lower single number.It can incite somebody to action
The corresponding business diary of same rankid and user behavior data are associated, such as by the exposure frequency and use in business diary
Number of clicks and the lower corresponding association of single number in the behavioral data of family.Determine exposure frequency in business diary and corresponding use
Whether the ratio between number of clicks and lower single number in the behavioral data of family are within the scope of preset ratio, if in preset ratio range
It is interior, then using business diary and its corresponding user behavior data as candidate samples data, if not within the scope of preset ratio,
Abandon the sample data.Wherein, preset ratio range be, for example, (110~90): 10:1, exposure frequency, number of clicks and under
The ratio of single number is preferably 100:10:1, so without being limited thereto.In this way, can prevent real-time training sample from relatively large deviation occur,
It is consistent with the probability distribution of truthful data as far as possible, to improve the accuracy of click model and lower single model.
Candidate samples data screening mode two:
It determines the click in user behavior data and whether the corresponding display location that places an order is less than preset value, if small
In preset value, then using user behavior data and its corresponding business diary as candidate samples data, while display location is abandoned
More than or equal to the user behavior data of preset value and its corresponding business diary.Preset value is, for example, 100, is so not limited to
This.It may be cheating data that the excessive user of display location, which clicks and places an order, influences click model and point single model updates
Accuracy.
Candidate samples data screening mode three:
Determine whether sample data is invalid data, if invalid data, then abandons the invalid data of preset ratio.In vain
Data can be the sample data that no user clicks.Such as 20% invalid data can be selected at random using pseudo-random function
It discards.In this way, the real-time of model training not only can be improved, the precision of model can also be improved.
Candidate samples data screening mode four:
Visitor's unit price in sample data is detected, objective unit price is greater than the sample data for presetting unit price as candidate samples number
According to discarding visitor's unit price, which is less than or equal to, presets monovalent sample data.Objective unit price is the order amount of money of the single order of user.This
Sample can be such that the training of model focuses on the higher order of price.
In one example, collator can also include adjustment module (not shown), and adjustment module can be used in real time
It predicts the purchasing price of user, and is adjusted according to the ranking results that the purchasing price of prediction predicts prediction module 504.It adjusts
The contingent order price of the specific trade company in available ranking results of mould preparation block, and according to the purchasing price and reference of prediction
Order price is adjusted ranking results.Adjustment module can also obtain reserved sorting position, and based on reserved sequence
Position adjusts ranking results.
Update module 502 specifically can be used for carrying out lower single model and click model respectively according to real-time training sample
Training obtains new model parameter；Stability is carried out to new model parameter and verifies to obtain the candidate that stability meets preset condition
Model parameter；Candidate family parameter is periodically loaded to lower single model and click model.Specifically, generation module 501 is raw
At real-time training sample can be put into message queue, after on-line training engine receives message, to lower single model and click mould
Type is trained to obtain new model parameter respectively, and snapshot stores new model parameter.Wherein, lower single model and click model
Every update can once respectively obtain the new model parameter of a set of (an also referred to as version), and snapshot stores new model parameter
It is to be stored to the model parameter of lower single model and the respective new version of click model.In some cases, based on new
The fluctuation of ranking results of model parameter can be bigger, i.e., the stability of new model parameter may be poor.Therefore, exist
After obtaining new model parameter, it can be verified to determine whether its stability meets preset condition, if its stability
Meet preset condition then as candidate family parameter, if its stability is unsatisfactory for preset condition, the model can be abandoned
Parameter.The stability verification of model parameter can be realized by way of the difference of comparison front and back ranking results.For example, being based on
In the ranking results of the model parameter of previous version, A, B, C, D businessman come first four respectively, and the model based on current version
In the ranking results of parameter, A, B, C, D businessman are not appeared in businessman's recommendation list, then can determine the model ginseng of current version
Several fluctuations is larger, is unsatisfactory for preset condition.Present embodiment, which does not do the determination of stability mode of candidate family parameter, to be had
Body limitation.Nearest candidate family parameter can be chosen in the candidate family parameter generated in real time is loaded onto online single-mode
In type and click model, and the load of model parameter can be carried out according to predetermined period, such as load in every 10 seconds is primary
Model parameter.
Present embodiment generation module can acquire user in real time and generate to the feedback data of online ranking results in real time
Training sample updates the model parameter of lower single model and click model by update module, simultaneously according to real-time training sample
Real-time stream is obtained by obtaining module, when needing to sort, lower single model and point of the prediction module based on real-time update
Hit model prediction ranking results.It can be seen that in the collator of present embodiment online ranking results, sample data acquisition
And the training of model parameter forms closure data flow in real time, to realize the sequence mould of lower single model and click model etc.
Type real-time online updates.Present embodiment has row by generating real-time training sample, using the update of newest sample increment iterative
Sequence model improves model sequence performance so that order models timely feedback sample distribution variation.Also, one kind can also be passed through
Or a variety of screening modes obtain candidate samples data, to be conducive to improve real-time, the accuracy of model modification；Simultaneously also
It can be by being finely adjusted to ranking results, to be conducive to improve the income of platform；Further, it is also possible to by new model
Parameter carries out stability verification, can advantageously ensure that the quality of sequence to avoid ranking results fluctuation.
Sixth embodiment of the invention is related to a kind of server, can be independent server, is also possible to server
Group.As shown in fig. 6, the server includes memory 602 and processor 601；
Wherein, the memory 602 is stored with the instruction that can be executed by least one described processor 601, described instruction
It is executed by least one described processor 601 to realize: generating real-time training sample；Under being updated according to the real-time training sample
The model parameter of single model and click model；Obtain real-time stream；According to the real-time stream and the institute based on update
It states lower single model and click model is predicted to obtain ranking results.
One or more processors 601 and memory 602, in Fig. 6 by taking a processor 601 as an example.Processor 601,
Memory 602 can be connected by bus or other modes, in Fig. 6 for being connected by bus.Memory 602 is used as one
Kind non-volatile computer readable storage medium storing program for executing, it is executable to can be used for storing non-volatile software program, non-volatile computer
Program and module.Non-volatile software program, instruction and mould of the processor 601 by operation storage in the memory 602
Block realizes above-mentioned sort method thereby executing the various function application and data processing of equipment.
Memory 602 may include storing program area and storage data area, wherein storing program area can store operation system
Application program required for system, at least one function；It storage data area can stored filter device etc..In addition, memory 602 can wrap
High-speed random access memory is included, can also include nonvolatile memory, for example, at least disk memory, a flash memories
Part or other non-volatile solid state memory parts.In some embodiments, it includes relative to processor that memory 602 is optional
601 remotely located memories, these remote memories can pass through network connection to external equipment.The example packet of above-mentioned network
Include but be not limited to internet, intranet, local area network, mobile radio communication and combinations thereof.
One or more module stores in the memory 602, when being executed by one or more processor 601, holds
Sort method in the above-mentioned any means embodiment of row.
Present embodiment can acquire user in real time and generate real-time training sample to the feedback data of online ranking results,
And the model parameter of lower single model and click model is updated according to real-time training sample, while obtaining real-time stream, it is needing
When sorting, lower single model and click model based on real-time update predict ranking results.It can be seen that present embodiment
Online ranking results, the acquisition of sample data and the training of model parameter form closure data flow in real time in sort method,
The sort method updated to the order models real-time online of the lower single model of realization and click model etc..Present embodiment passes through
Real-time training sample is generated, existing order models are updated using newest sample increment iterative, so that order models timely feedback sample
This changes in distribution improves model sequence performance.
Above equipment can be performed embodiment of the present invention provided by method, have the corresponding functional module of execution method and
Beneficial effect, the not technical detail of detailed description in the present embodiment, reference can be made to method provided by embodiment of the present invention.
7th embodiment of the invention is related to a kind of non-volatile memory medium, for storing computer-readable program,
The computer-readable program is used to execute above-mentioned all or part of embodiment of the method for computer.
That is, it will be understood by those skilled in the art that implement the method for the above embodiments be can be with
Relevant hardware is instructed to complete by program, which is stored in a storage medium, including some instructions are to make
It obtains an equipment (can be single-chip microcontroller, chip etc.) or processor (processor) executes side described in each embodiment of the present invention
The all or part of the steps of method.And storage medium above-mentioned includes: USB flash disk, mobile hard disk, read-only memory (ROM, Read-Only
Memory), random access memory (RAM, Random Access Memory), magnetic or disk etc. are various can store journey
The medium of sequence code.
It will be understood by those skilled in the art that the respective embodiments described above are to realize specific embodiments of the present invention,
And in practical applications, can to it, various changes can be made in the form and details, without departing from the spirit and scope of the present invention.
The embodiment of the present application discloses a kind of sort method of A1., comprising:
Generate real-time training sample；
The model parameter of lower single model and click model is updated according to the real-time training sample；
Obtain real-time stream；
Lower single model and click model according to the real-time stream and based on update are predicted to obtain sequence knot
Fruit.
A2. sort method as described in a1, it is described to generate real-time training sample, it specifically includes:
Collecting sample data；
The sample data of acquisition is screened to obtain candidate samples data；The candidate samples data include: industry
Business log and user behavior data；
It polymerize the business diary and the user behavior data to obtain real-time training sample.
A3. sort method, the business diary include exposure frequency as described in A2, and the user behavior data includes a little
Hit number and lower single number；
The sample data of described pair of acquisition is screened to obtain candidate samples data, is specifically included:
It determines exposure frequency in the business diary and the number of clicks in corresponding user behavior data and places an order
Whether the ratio between number is within the scope of preset ratio；If within the scope of the preset ratio, by the business diary and its correspondence
User behavior data as the candidate samples data.
A4. sort method, the sample data of described pair of acquisition are screened to obtain candidate samples number as described in A2
According to specifically including:
It determines the click in the user behavior data and whether the corresponding display location that places an order is less than preset value,
If being less than preset value, using the user behavior data and its corresponding business diary as the candidate samples data.
A5. sort method, the sample data of described pair of acquisition are screened to obtain candidate samples number as described in A2
According to specifically including:
Determine whether the sample data is invalid data；If invalid data, then the invalid data of preset ratio is abandoned.
A6. sort method, the sample data of described pair of acquisition are screened to obtain candidate samples number as described in A2
According to specifically including:
Detect visitor's unit price in the sample data；
Objective unit price is greater than the sample data for presetting unit price as the candidate samples data.
A7. sort method as described in a1, further includes:
The purchasing price of prediction user in real time；
Lower single model and click model according to the real-time stream and based on update is predicted to be arranged
After sequence result, further includes:
The ranking results are adjusted according to the purchasing price of prediction.
A8. the sort method as described in A7, it is described that the ranking results are adjusted according to the purchasing price of prediction
It is whole, it specifically includes:
Obtain the contingent order price of the trade company in the ranking results；
The ranking results are adjusted according to the purchasing price of prediction and the contingent order price.
A9. sort method as described in a1, lower single model according to the real-time stream and based on update
And after click model predicts to obtain ranking results, further includes:
Obtain reserved sorting position；
The ranking results are adjusted based on the reserved sorting position.
A10. sort method as described in a1, the real-time stream include: browsing pages and click.
A11. the sort method as described in any one of A1 to A10, it is described to be placed an order according to the real-time training sample update
The model parameter of model and click model, specifically includes:
Lower single model and click model are trained to obtain new mould respectively according to the real-time training sample
Shape parameter；
Stability is carried out to the new model parameter and verifies to obtain the candidate family parameter that stability meets preset condition；
The candidate family parameter is periodically loaded to lower single model and click model.
The embodiment of the present application also discloses a kind of collator of B1., comprising:
Generation module, for generating real-time training sample；
Update module, for updating the model parameter of lower single model and click model according to the real-time training sample；
Module is obtained, for obtaining real-time stream；
Prediction module, for according to the real-time stream and based on update lower single model and click model it is pre-
Measure ranking results.
The embodiment of the present application also discloses a kind of server of C1., comprising: memory and processor, memory storage calculate
Machine program, processor run the computer program to realize:
Generate real-time training sample；
The model parameter of lower single model and click model is updated according to the real-time training sample；
Obtain real-time stream；
Lower single model and click model according to the real-time stream and based on update are predicted to obtain sequence knot
Fruit.
C2. the server as described in C1, the processor are also used to execute the sequence side as described in any one of A2 to A11
Method.
The embodiment of the present application also discloses a kind of computer readable storage medium of D1., is stored with computer program, the meter
Calculation machine program is executed by processor the sort method as described in any one of A1 to A11.
Claims (10)
1. a kind of sort method characterized by comprising
Generate real-time training sample；
The model parameter of lower single model and click model is updated according to the real-time training sample；
Obtain real-time stream；
Lower single model and click model according to the real-time stream and based on update are predicted to obtain ranking results.
2. sort method according to claim 1, which is characterized in that it is described to generate real-time training sample, it specifically includes:
Collecting sample data；
The sample data of acquisition is screened to obtain candidate samples data；The candidate samples data include: business day
Will and user behavior data；
It polymerize the business diary and the user behavior data to obtain real-time training sample.
3. sort method according to claim 2, which is characterized in that the business diary includes exposure frequency, the use
Family behavioral data includes number of clicks and lower single number；
The sample data of described pair of acquisition is screened to obtain candidate samples data, is specifically included:
Determine exposure frequency in the business diary in corresponding user behavior data number of clicks and lower single number
The ratio between whether within the scope of preset ratio；If within the scope of the preset ratio, by the business diary and its corresponding use
Family behavioral data is as the candidate samples data.
4. sort method according to claim 2, which is characterized in that the sample data of described pair of acquisition is screened
Candidate samples data are obtained, are specifically included:
It determines the click in the user behavior data and whether the corresponding display location that places an order is less than preset value, if small
In preset value, then using the user behavior data and its corresponding business diary as the candidate samples data.
5. sort method according to claim 2, which is characterized in that the sample data of described pair of acquisition is screened
Candidate samples data are obtained, are specifically included:
Determine whether the sample data is invalid data；If invalid data, then the invalid data of preset ratio is abandoned.
6. sort method according to claim 2, which is characterized in that the sample data of described pair of acquisition is screened
Candidate samples data are obtained, are specifically included:
Detect visitor's unit price in the sample data；
Objective unit price is greater than the sample data for presetting unit price as the candidate samples data.
7. sort method according to claim 1, which is characterized in that further include:
The purchasing price of prediction user in real time；
Lower single model and click model according to the real-time stream and based on update is predicted to obtain sequence knot
After fruit, further includes:
The ranking results are adjusted according to the purchasing price of prediction.
8. a kind of collator characterized by comprising
Generation module, for generating real-time training sample；
Update module, for updating the model parameter of lower single model and click model according to the real-time training sample；
Module is obtained, for obtaining real-time stream；
Prediction module, for according to the real-time stream and based on update lower single model and click model measure in advance
To ranking results.
9. a kind of server characterized by comprising memory and processor, memory store computer program, processor fortune
The row computer program is to realize:
Generate real-time training sample；
The model parameter of lower single model and click model is updated according to the real-time training sample；
Obtain real-time stream；
Lower single model and click model according to the real-time stream and based on update are predicted to obtain ranking results.
10. a kind of storage medium, which is characterized in that for storing computer-readable program, the computer-readable program is used for
The sort method as described in any one of claims 1 to 7 is executed for computer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811520392.9A CN109558544B (en) | 2018-12-12 | 2018-12-12 | Sorting method and device, server and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811520392.9A CN109558544B (en) | 2018-12-12 | 2018-12-12 | Sorting method and device, server and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109558544A true CN109558544A (en) | 2019-04-02 |
| CN109558544B CN109558544B (en) | 2021-04-27 |
Family
ID=65869639
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811520392.9A Active CN109558544B (en) | 2018-12-12 | 2018-12-12 | Sorting method and device, server and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109558544B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110310168A (en) * | 2019-05-17 | 2019-10-08 | 北京小度信息科技有限公司 | Method of adjustment, device, server and the storage medium of model |
| CN110363590A (en) * | 2019-07-16 | 2019-10-22 | 深圳乐信软件技术有限公司 | A kind of advertisement recommended method, device, terminal and storage medium |
| CN110674408A (en) * | 2019-09-30 | 2020-01-10 | 北京三快在线科技有限公司 | Service platform, and real-time generation method and device of training sample |
| CN111931056A (en) * | 2020-08-17 | 2020-11-13 | 北京小川科技有限公司 | Push content recommendation method and device |
| WO2020233432A1 (en) * | 2019-05-20 | 2020-11-26 | 阿里巴巴集团控股有限公司 | Method and device for information recommendation |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103412882A (en) * | 2013-07-18 | 2013-11-27 | 百度在线网络技术（北京）有限公司 | Method and device for distinguishing consumption intention |
| US20130339350A1 (en) * | 2012-06-18 | 2013-12-19 | Alibaba Group Holding Limited | Ranking Search Results Based on Click Through Rates |
| CN105260471A (en) * | 2015-10-19 | 2016-01-20 | 广州唯品会信息科技有限公司 | Training method and system of commodity personalized ranking model |
| CN106168975A (en) * | 2016-07-12 | 2016-11-30 | 精硕世纪科技（北京）有限公司 | The acquisition methods of targeted customer's concentration and device |
| WO2016197854A1 (en) * | 2015-06-09 | 2016-12-15 | 阿里巴巴集团控股有限公司 | Service object information processing, credential information processing method and apparatus |
| CN107194430A (en) * | 2017-05-27 | 2017-09-22 | 北京三快在线科技有限公司 | A kind of screening sample method and device, electronic equipment |
| CN107423355A (en) * | 2017-05-26 | 2017-12-01 | 北京三快在线科技有限公司 | Information recommendation method and device, electronic equipment |
| CN107657506A (en) * | 2017-09-06 | 2018-02-02 | 北京五八到家信息技术有限公司 | The intelligent recommendation method and commending system of a kind of service commodity |
| CN108334575A (en) * | 2018-01-23 | 2018-07-27 | 北京三快在线科技有限公司 | A kind of recommendation results sequence modification method and device, electronic equipment |
| CN108733277A (en) * | 2018-05-22 | 2018-11-02 | 江西午诺科技有限公司 | Application icon aligning method, device, readable storage medium storing program for executing and mobile terminal |
-
2018
- 2018-12-12 CN CN201811520392.9A patent/CN109558544B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130339350A1 (en) * | 2012-06-18 | 2013-12-19 | Alibaba Group Holding Limited | Ranking Search Results Based on Click Through Rates |
| CN103412882A (en) * | 2013-07-18 | 2013-11-27 | 百度在线网络技术（北京）有限公司 | Method and device for distinguishing consumption intention |
| WO2016197854A1 (en) * | 2015-06-09 | 2016-12-15 | 阿里巴巴集团控股有限公司 | Service object information processing, credential information processing method and apparatus |
| CN105260471A (en) * | 2015-10-19 | 2016-01-20 | 广州唯品会信息科技有限公司 | Training method and system of commodity personalized ranking model |
| CN106168975A (en) * | 2016-07-12 | 2016-11-30 | 精硕世纪科技（北京）有限公司 | The acquisition methods of targeted customer's concentration and device |
| CN107423355A (en) * | 2017-05-26 | 2017-12-01 | 北京三快在线科技有限公司 | Information recommendation method and device, electronic equipment |
| CN107194430A (en) * | 2017-05-27 | 2017-09-22 | 北京三快在线科技有限公司 | A kind of screening sample method and device, electronic equipment |
| CN107657506A (en) * | 2017-09-06 | 2018-02-02 | 北京五八到家信息技术有限公司 | The intelligent recommendation method and commending system of a kind of service commodity |
| CN108334575A (en) * | 2018-01-23 | 2018-07-27 | 北京三快在线科技有限公司 | A kind of recommendation results sequence modification method and device, electronic equipment |
| CN108733277A (en) * | 2018-05-22 | 2018-11-02 | 江西午诺科技有限公司 | Application icon aligning method, device, readable storage medium storing program for executing and mobile terminal |
Non-Patent Citations (2)
| Title |
|---|
| TAESUP MOON等: "《An Online Learning Framework for Refining Recency Search Results with User Click Feedback》", 《ACM TRANSACTIONS ON INFORMATION SYSTEMS》 * |
| 曾宪宇等: "《用户在线购买预测:一种基于用户操作序列和选择模型的方法》", 《计算机研究与发展》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110310168A (en) * | 2019-05-17 | 2019-10-08 | 北京小度信息科技有限公司 | Method of adjustment, device, server and the storage medium of model |
| WO2020233432A1 (en) * | 2019-05-20 | 2020-11-26 | 阿里巴巴集团控股有限公司 | Method and device for information recommendation |
| CN110363590A (en) * | 2019-07-16 | 2019-10-22 | 深圳乐信软件技术有限公司 | A kind of advertisement recommended method, device, terminal and storage medium |
| CN110674408A (en) * | 2019-09-30 | 2020-01-10 | 北京三快在线科技有限公司 | Service platform, and real-time generation method and device of training sample |
| CN110674408B (en) * | 2019-09-30 | 2021-06-04 | 北京三快在线科技有限公司 | Service platform, and real-time generation method and device of training sample |
| CN111931056A (en) * | 2020-08-17 | 2020-11-13 | 北京小川科技有限公司 | Push content recommendation method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109558544B (en) | 2021-04-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109558544A (en) | Sort method and device, server and storage medium | |
| CN105701216B (en) | A kind of information-pushing method and device | |
| JP6211605B2 (en) | Ranking search results based on click-through rate | |
| CN103714084B (en) | The method and apparatus of recommendation information | |
| Steck | Item popularity and recommendation accuracy | |
| CN105653562B (en) | The calculation method and device of correlation between a kind of content of text and inquiry request | |
| CN102053971B (en) | Recommending method and equipment for sequencing-oriented collaborative filtering | |
| CN108596695B (en) | Entity pushing method and system | |
| CN105184618A (en) | Commodity individual recommendation method for new users and system | |
| Lommatzsch | Real-time news recommendation using context-aware ensembles | |
| CN102932206B (en) | The method and system of monitoring website access information | |
| CN108550068A (en) | A kind of individual commodity recommendation method and system based on user behavior analysis | |
| CN108921398B (en) | Shop quality evaluation method and device | |
| CN110163647A (en) | A kind of data processing method and device | |
| CN102663617A (en) | Method and system for prediction of advertisement clicking rate | |
| CN105765573A (en) | Improvements in website traffic optimization | |
| CN109767270A (en) | The old information recommendation method of housing support and system are deposited based on artificial intelligence | |
| CN110490625A (en) | User preference determines method and device, electronic equipment, storage medium | |
| CN108319646B (en) | Vehicle source searching method and device based on user historical behaviors | |
| CN110334356A (en) | Article matter method for determination of amount, article screening technique and corresponding device | |
| CN106919575A (en) | application program searching method and device | |
| CN111061979B (en) | User tag pushing method and device, electronic equipment and medium | |
| WO2023000491A1 (en) | Application recommendation method, apparatus and device, and computer-readable storage medium | |
| CN105740434B (en) | Network information methods of marking and device | |
| JP2008234338A (en) | Season degree analysis system, in-season degree analysis method, and season degree analysis program |
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 |