CN115564511A - CTR position depolarization method combining adjacent positions and double history sequences - Google Patents

CTR position depolarization method combining adjacent positions and double history sequences Download PDF

Info

Publication number
CN115564511A
CN115564511A CN202211038536.3A CN202211038536A CN115564511A CN 115564511 A CN115564511 A CN 115564511A CN 202211038536 A CN202211038536 A CN 202211038536A CN 115564511 A CN115564511 A CN 115564511A
Authority
CN
China
Prior art keywords
information
adjacent
side information
user
pooling
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
Application number
CN202211038536.3A
Other languages
Chinese (zh)
Inventor
韩弘炀
傅剑文
陈心童
章建森
周文彬
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tianyi Electronic Commerce Co Ltd
Original Assignee
Tianyi Electronic Commerce Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Tianyi Electronic Commerce Co Ltd filed Critical Tianyi Electronic Commerce Co Ltd
Priority to CN202211038536.3A priority Critical patent/CN115564511A/en
Priority to PCT/CN2022/136489 priority patent/WO2024045394A1/en
Publication of CN115564511A publication Critical patent/CN115564511A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/06—Buying, selling or leasing transactions
    • G06Q30/0601—Electronic shopping [e-shopping]
    • G06Q30/0631—Item recommendations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Databases & Information Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Finance (AREA)
  • Accounting & Taxation (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Strategic Management (AREA)
  • Biomedical Technology (AREA)
  • Development Economics (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a CTR position depolarization method combining adjacent positions and double history sequences. The invention provides a CTR position depolarization method combining adjacent positions and a double history sequence aiming at eliminating position offset in a recommendation system. On the basis, the positive and negative feedback information of the user history is utilized, the information about the preference and aversion of the user can be more fully learned, and the model effect is improved. Independent network learning is designed, the influence of position factors on results can be eliminated, and the method is convenient to use in actual model inference and more flexibly deploy on line.

Description

CTR position depolarization method combining adjacent positions and double history sequences
Technical Field
The invention relates to the field of recommendation systems, in particular to a CTR (program traffic indicator) position depolarization method combining adjacent positions and a double history sequence.
Background
The recommendation form of thousands of people and thousands of faces plays a role in scenes of E-commerce, videos, news and other life, and a lot of information is finely selected and distributed to users through intelligent result display, so that the users can see articles which are more interesting. The recommendation model is used for judging the preference of the user for recommended articles based on learning the interaction data of the user in the scene, so that the satisfactory articles are recommended for the user. However, the user's interaction data is affected by the overall presentation environment, e.g., the material at position 3 is more easily noticed than the material at position 6, and the material with which the user has historically interacted is more relevant to position 3 but less relevant to the first 2 and more easily ignores the first 2 material. Therefore, the clicking behavior of the user is derived from the fact that the user is more interested in the material and the display position and the environment of the material can enable the user to pay attention to the material. However, most current recommendation systems only address the interests of the user, and therefore result in a prediction bias due to the location information. Even if the position factor is considered in a small amount of scenes, only the position information corresponding to the material is considered, and the influence of the adjacent position of the material and the material displayed by the adjacent position on the material in combination with the historical behavior sequence information of the user is not considered.
Aiming at the practical background, the invention provides a CTR position depolarization method combining adjacent positions and double history sequences, which can effectively utilize user history feedback information and adjacent environment information to eliminate prediction deviation, improve recommendation accuracy, separate calculation of position factors on a model structure level, and facilitate use and online flexible deployment when an actual model is deduced.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a CTR position depolarization method combining adjacent positions and a double history sequence.
The invention provides the following technical scheme:
the invention provides a CTR position depolarization method combining adjacent positions and double history sequences, which comprises the following steps:
1. an acquisition module:
method for inputting exposure and click data of user in recommended scene by using streaming technologyAcquiring and storing the rows into hive and recording the rows as table; marking the exposed and un-clicked sample in the data as 0 and the clicked sample as 1 by utilizing SQL to obtain the ith sample label y i And splice the material display position pos i And set of adjacent material positions
Figure BDA0003819826950000021
Material and its side information (such as material category, etc.) item i And adjacent material and side information set thereof
Figure BDA0003819826950000022
Other attributes x i (ii) a At most m click sequences in t days of history of the user u
Figure BDA0003819826950000023
And expose non-clicked sequence
Figure BDA0003819826950000024
It should be noted that data acquisition techniques and storage formats include, but are not limited to, those described above;
2. an adjacent position information extraction module:
1) Imbedding transformation
For samples in the table, showing the position pos of the material i Set of adjacent material positions in display page
Figure BDA0003819826950000025
Material and side information item thereof i And the collection of adjacent materials and side information in the display page
Figure BDA0003819826950000026
Positive feedback sequence
Figure BDA0003819826950000027
And negative feedback sequence
Figure BDA0003819826950000028
Respectively converted into corresponding embedding forms pemb i 、
Figure BDA0003819826950000029
iemb i 、
Figure BDA00038198269500000210
And
Figure BDA00038198269500000211
2) Positive and negative feedback sequence posing calculation
For the material and the side information of the sample i, calculating the positive and negative feedback sequences and posing to obtain the preference and aversion information pooling posiool corresponding to the material and the side information according to the historical feedback information of the user u,i And negpool u,i Wherein S (-) is an MLP network with output dimension 1;
Figure BDA00038198269500000212
Figure BDA00038198269500000213
Figure BDA00038198269500000214
Figure BDA00038198269500000215
similarly, for the adjacent material and the side information thereof of the sample i, the preference and the aversion information pospoool of the historical feedback information of the user corresponding to the adjacent material and the side information thereof can be obtained u,i And negpool u,i ；
Figure BDA0003819826950000031
Figure BDA0003819826950000032
Figure BDA0003819826950000033
Figure BDA0003819826950000034
3) Neighbor location information extraction
Calculating the attention scores corresponding to the adjacent positions of each sample material
Figure BDA0003819826950000035
Pooling of npools in Adjacent positions i Materials at adjacent positions and side information pooling niool thereof i Positive feedback sequence pooling npospool at adjacent positions i And adjacent position negative feedback sequence pooling nnegpool i ；
concat i ＝concat(pemb i ,iemb i ,pospool u,i ,negpool u,i )
Figure BDA0003819826950000036
Figure BDA0003819826950000037
Figure BDA0003819826950000038
Figure BDA0003819826950000039
Figure BDA00038198269500000310
Figure BDA00038198269500000311
3. Model building module
1) Location network learning
Separately entering pemb in location network i And nppool i Using the same multi-layer deep network learning, the output results are pres respectively i And npres i ，W 1 ,…,W H And B 1 ,…,B H Respectively representing the weight and the bias parameter of learning, wherein S (-) is a multilayer depth network with an output dimension of 1;
pres i ＝S(pemb i ,W 1 ,…,W H ,B 1 ,…,B H )
npres i ＝S(nppool i ,W 1 ,…,W H ,B 1 ,…,B H )；
2) Other information web learning
X is to be i Conversion into the corresponding embedding form xemb i Separately entering iconcat in other information networks i And niconcat i The same multilayer deep network learning is used, and the output result is ires i And nires i ，W’ 1 ,…,W’ H And B' 1 ,…,B’ H Respectively representing the learned weight and the bias parameter;
iconcat i ＝concat(iemb i ,pospool u,i ,negpool u,i ,xemb i )
niconcat i ＝concat(nipool i ,npospool i ,nnegpool i ,xemb i )
ires i ＝S(iconcat i ,W’ 1 ,…,W’ H ,B’ 1 ,…,B’ H )
nires i ＝S(niconcat i ,W’ 1 ,…,W’ H ,B’ 1 ,…,B’ H )；
4. model prediction module
During training
Figure BDA0003819826950000041
Loss function
Figure BDA0003819826950000042
Figure BDA0003819826950000043
Only the materials and the side information, the positive and negative feedback sequences and other information are input during the inference,
Figure BDA0003819826950000044
and storing a recommendation result table result in hive; and outputting the recommendation interface to the front end for display by the interface.
Compared with the prior art, the invention has the following beneficial effects:
the traditional recommendation model only aims at the interest of the user in the materials, so that prediction deviation generated because the position information does not participate in modeling is introduced during model construction, and the recommendation result of the user is biased. Although the position information is considered in a small amount of scenes, only the position information corresponding to the material is considered, and the influence of the adjacent position of the material and the material displayed by the adjacent position on the material in combination with the historical behavior sequence information of the user is not considered.
The invention provides a CTR position depolarization method combining adjacent positions and a double history sequence aiming at eliminating position offset in a recommendation system. On the basis, the positive and negative feedback information of the user history is utilized, the information about the preference and aversion of the user can be more fully learned, and the model effect is improved. Independent network learning is designed, the influence of position factors on results can be eliminated, and the method is convenient to use in actual model inference and more flexibly deploy on line.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:
FIG. 1 is a schematic diagram of the CTR position depolarization algorithm training and inference in conjunction with neighboring positions and a dual history sequence in accordance with the present invention;
fig. 2 is a flow chart of an implementation of the present invention.
Detailed Description
The preferred embodiments of the present invention will be described in conjunction with the accompanying drawings, and it will be understood that they are described herein for the purpose of illustration and explanation and not limitation. Wherein like reference numerals refer to like parts throughout.
Example 1
Referring to fig. 1-2, the present invention provides a CTR position depolarization method combining neighboring positions and a dual history sequence, comprising the following steps:
1. an acquisition module:
acquiring exposure and click data of a user in a recommended scene by using a streaming technology, and storing the data into hive to be recorded as table; marking the exposed and un-clicked sample in the data as 0 and marking the clicked sample as 1 by utilizing SQL to obtain the ith sample label y i And splicing the material display positions pos i And set of adjacent material positions
Figure BDA0003819826950000051
Material and its side information (such as material category, etc.) item i And adjacent material and side information set thereof
Figure BDA0003819826950000052
Other attributes x i (ii) a At most m click sequences in the historical t days of the user u
Figure BDA0003819826950000053
And expose non-click sequences
Figure BDA0003819826950000054
It should be noted that data acquisition techniques and storage formats include, but are not limited to, those described above;
2. an adjacent position information extraction module:
1) Imbedding transformation
For samples in the table, showing the position pos of the material i Set of adjacent material positions in display page
Figure BDA0003819826950000061
Material and side information item thereof i And the collection of adjacent materials and side information in the display page
Figure BDA0003819826950000062
Positive feedback sequence
Figure BDA0003819826950000063
And negative feedback sequence
Figure BDA0003819826950000064
Respectively converted into corresponding embedding form pemb i 、
Figure BDA0003819826950000065
iemb i 、
Figure BDA0003819826950000066
And
Figure BDA0003819826950000067
2) Positive and negative feedback sequence posing calculation
For the material and the side information of the sample i, calculating the positive and negative feedback sequence pooling to obtain the preference and aversion information pooling poppool corresponding to the material and the side information according to the historical feedback information of the user u,i And negpool u,i Wherein S (-) is an MLP network with output dimension 1;
Figure BDA0003819826950000068
Figure BDA0003819826950000069
Figure BDA00038198269500000610
Figure BDA00038198269500000611
similarly, for the adjacent material and the side information thereof of the sample i, the preference and the aversion information pospoool of the historical feedback information of the user corresponding to the adjacent material and the side information thereof can be obtained u,i And negpool u,i ；
Figure BDA00038198269500000612
Figure BDA00038198269500000613
Figure BDA00038198269500000614
Figure BDA00038198269500000615
3) Neighbor location information extraction
Calculating attention scores corresponding to adjacent positions of each sample material
Figure BDA00038198269500000616
Pooling of npools in Adjacent positions i Material in adjacent position and its side information pooling niool i Positive feedback sequence pooling npospool at adjacent positions i And adjacent bitNegative feedback sequence pooling nnegpool i ；
concat i ＝concat(pemb i ,iemb i ,pospool u,i ,negpool u,i )
Figure BDA0003819826950000071
Figure BDA0003819826950000072
Figure BDA0003819826950000073
Figure BDA0003819826950000074
Figure BDA0003819826950000075
Figure BDA0003819826950000076
3. Model building module
1) Location network learning
Separately entering pemb in location network i And nppool i Using the same multi-layer deep network learning, the output results are pres respectively i And npres i ，W 1 ,…,W H And B 1 ,…,B H Respectively representing the weight and the bias parameter of learning, wherein S (-) is a multilayer depth network with an output dimension of 1;
pres i ＝S(pemb i ,W 1 ,…,W H ,B 1 ,…,B H )
npres i ＝S(nppool i ,W 1 ,…,W H ,B 1 ,…,B H )；
2) Other information web learning
X is to be i Conversion into the corresponding embedding form xemb i Separately entering iconcat in other information networks i And niconcat i The same multilayer deep network learning is used, and the output result is ires i And nires i ，W’ 1 ,…,W’ H And B' 1 ,…,B’ H Respectively representing the learned weight and the bias parameter;
iconcat i ＝concat(iemb i ,pospool u,i ,negpool u,i ,xemb i )
niconcat i ＝concat(nipool i ,npospool i ,nnegpool i ,xemb i )
ires i ＝S(iconcat i ,W’ 1 ,…,W’ H ,B’ 1 ,…,B’ H )
nires i ＝S(niconcat i ,W’ 1 ,…,W’ H ,B’ 1 ,…,B’ H )；
4. model prediction module
During training
Figure BDA0003819826950000081
Loss function
Figure BDA0003819826950000082
Figure BDA0003819826950000083
Only the materials and the side information, the positive and negative feedback sequences and other information are input during the inference,
Figure BDA0003819826950000084
and storing a recommendation result table result in hive; and outputting the recommended interface to the front end by the interface for displaying.
Further, examples are as follows:
1. using streaming techniques to target usersThe exposure and click data are collected and stored into hive as table. Obtaining ith sample label y by SQL i And splicing the material display positions pos i And set of adjacent material positions
Figure BDA0003819826950000085
Material and its side information (such as material category, etc.) item i And adjacent material and side information set thereof
Figure BDA0003819826950000086
Other attributes x i (ii) a At most m click sequences in t days of history of the user u
Figure BDA0003819826950000087
And expose non-click sequences
Figure BDA0003819826950000088
2. Converting the sample value into a corresponding imbedding form pemb i 、
Figure BDA0003819826950000089
iemb i 、
Figure BDA00038198269500000810
xemb i 、
Figure BDA00038198269500000811
And
Figure BDA00038198269500000812
pooling pop ool of the material and the side information thereof, and the corresponding preference and aversion information of the adjacent material and the side information thereof according to the historical feedback information of the user u,i 、negpool u,i 、pospool u,i And negpool u,i Where S (-) is the MLP network with output dimension 1.
Figure BDA00038198269500000813
Figure BDA00038198269500000814
Figure BDA00038198269500000815
Figure BDA00038198269500000816
Figure BDA00038198269500000817
Figure BDA00038198269500000818
Figure BDA0003819826950000091
Figure BDA0003819826950000092
3. Calculating neighboring position pooling nppool i Material in adjacent position and its side information pooling nppool i Pooling npospool with adjacent position positive and negative feedback sequences i And nnegpool i 。
concat i ＝concat(pemb i ,iemb i ,pospool u,i ,negpool u,i )
Figure BDA0003819826950000093
Figure BDA0003819826950000094
Figure BDA0003819826950000095
Figure BDA0003819826950000096
Figure BDA0003819826950000097
Figure BDA0003819826950000098
4. Using location independent network computing
pres i ＝S(pemb i ,W 1 ,…,W H ,B 1 ,…,B H )
npres i ＝S(nppool i ,W 1 ,…,W H ,B 1 ,…,B H )
5. Independent network computing using other information
iconcat i ＝concat(iemb i ,pospool u,i ,negpool u,i ,xemb i )
niconcat i ＝concat(nipool i ,npospool i ,nnegpool i ,xemb i )
ires i ＝S(iconcat i ,W’ 1 ,…,W’ H ,B’ 1 ,…,B’ H )
nires i ＝S(niconcat i ,W’ 1 ,…,W’ H ,B’ 1 ,…,B’ H )
6. To be provided with
Figure BDA0003819826950000099
Loss function
Figure BDA00038198269500000910
Figure BDA00038198269500000911
And (5) training the model.
7. Converting materials to be predicted and side information, positive and negative feedback sequences and other information thereof into corresponding imbedding form iemb j 、pospool u,j 、negpool u,j And xemb j . Computing with trained models
ires i ＝S(iconcat i ,W’ 1 ,…,W’ H ,B’ 1 ,…,B’ H ) And
Figure BDA0003819826950000101
the recommended result is stored in hive and the result is pushed to hbase. And outputting the recommendation interface to the front end for display by the interface.
The invention has the following characteristics:
1. the method for eliminating the position offset of the recommended model is provided, corresponding information is extracted by using the material display position, the adjacent material, the historical behavior sequence pooling and the historical behavior sequence pooling of the adjacent material, and the model accuracy can be effectively improved.
2. The method for extracting the preference and aversion information of the user by using the double history sequence is provided, so that the positive and negative feedback information of the user can be learned more accurately in the model, and the recommendation effect of the model is effectively improved.
Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that modifications may be made to the embodiments described above, or equivalents may be substituted for elements thereof. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (1)

1. A CTR position depolarization method combining neighboring positions and a dual history sequence is characterized by comprising the following steps:
1. an acquisition module:
acquiring exposure and click data of a user in a recommended scene by using a streaming technology, and storing the data into hive to be recorded as table; marking the exposed and un-clicked sample in the data as 0 and the clicked sample as 1 by utilizing SQL to obtain the ith sample label y i And splice the material display position pos i And set of adjacent material positions
Figure FDA0003819826940000011
Material and its side information (such as material category, etc.) item i And adjacent material and side information set thereof
Figure FDA0003819826940000012
Other attributes x i (ii) a At most m click sequences in t days of history of the user u
Figure FDA0003819826940000013
And expose non-clicked sequence
Figure FDA0003819826940000014
It should be noted that data acquisition techniques and storage formats include, but are not limited to, those described above;
2. an adjacent position information extraction module:
1) Imbedding transformation
For the sample in the table, show the position pos of the material i Set of adjacent material positions in display page
Figure FDA0003819826940000015
Material and side information item thereof i And the collection of adjacent materials and side information in the display page
Figure FDA0003819826940000016
Positive feedback sequence
Figure FDA0003819826940000017
And negative feedback sequence
Figure FDA0003819826940000018
Respectively converted into corresponding embedding forms pemb i 、
Figure FDA0003819826940000019
iemb i 、
Figure FDA00038198269400000110
And
Figure FDA00038198269400000111
2) Positive and negative feedback sequence posing calculation
For the material and the side information of the sample i, calculating the positive and negative feedback sequence pooling to obtain the preference and aversion information pooling poppool corresponding to the material and the side information according to the historical feedback information of the user u,i And negpool u,i Wherein S (-) is an MLP network with output dimension 1;
Figure FDA00038198269400000112
Figure FDA00038198269400000113
Figure FDA00038198269400000114
Figure FDA00038198269400000115
in the same way, forThe adjacent materials and the side information thereof of the sample i can obtain the preference and aversion information poppool of the historical feedback information of the user corresponding to the adjacent materials and the side information thereof u,i And negpool u,i ；
Figure FDA0003819826940000021
Figure FDA0003819826940000022
Figure FDA0003819826940000023
Figure FDA0003819826940000024
3) Neighbor location information extraction
Calculating the attention scores corresponding to the adjacent positions of each sample material
Figure FDA0003819826940000025
Pooling of npools in Adjacent positions i Material in adjacent position and its side information pooling niool i Positive feedback sequence pooling npospool at adjacent positions i And adjacent position negative feedback sequence pooling nnegpool i ；
concat i ＝concat(pemb i ,iemb i ,pospool u,i ,negpool u,i )
Figure FDA0003819826940000026
Figure FDA0003819826940000027
Figure FDA0003819826940000028
Figure FDA0003819826940000029
Figure FDA00038198269400000210
Figure FDA00038198269400000211
3. Model building module
1) Location network learning
Separately entering pemb in location network i And nppool i Using the same multilayer deep network learning, the output result is pres respectively i And npres i ，W 1 ,…,W H And B 1 ,…,B H Respectively representing the weight and the bias parameter of learning, wherein S (-) is a multilayer deep network with an output dimension of 1;
pres i ＝S(pemb i ,W 1 ,…,W H ,B 1 ,…,B H )
npres i ＝S(nppool i ,W 1 ,…,W H ,B 1 ,…,B H )；
2) Other information web learning
X is to be i Conversion into the corresponding embedding form xemb i Separately entering iconcat in other information networks i And niconcat i The same multilayer deep network learning is used, and the output result is ires i And nires i ，W’ 1 ,…,W’ H And B' 1 ,…,B’ H Respectively represent the weights of learningAnd a bias parameter;
iconcat i ＝concat(iemb i ,pospool u,i ,negpool u,i ,xemb i )
niconcat i ＝concat(nipool i ,npospool i ,nnegpool i ,xemb i )
ires i ＝S(iconcat i ,W′ 1 ,…,W′ H ,B′ 1 ,…,B′ H )
nires i ＝S(niconcat i ,W′ 1 ,…,W′ H ,B′ 1 ,…,B′ H )；
4. model prediction module
During training
Figure FDA0003819826940000031
Loss function
Figure FDA0003819826940000032
Figure FDA0003819826940000033
Only the materials and the side information, the positive and negative feedback sequences and other information are input during the inference,
Figure FDA0003819826940000034
and storing a recommendation result table result in hive; and outputting the recommendation interface to the front end for display by the interface.
CN202211038536.3A 2022-08-29 2022-08-29 CTR position depolarization method combining adjacent positions and double history sequences Pending CN115564511A (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202211038536.3A CN115564511A (en) 2022-08-29 2022-08-29 CTR position depolarization method combining adjacent positions and double history sequences
PCT/CN2022/136489 WO2024045394A1 (en) 2022-08-29 2022-12-05 Ctr position offset elimination method combining adjacent positions and double historical sequences

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211038536.3A CN115564511A (en) 2022-08-29 2022-08-29 CTR position depolarization method combining adjacent positions and double history sequences

Publications (1)

Publication Number Publication Date
CN115564511A true CN115564511A (en) 2023-01-03

Family

ID=84738297

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211038536.3A Pending CN115564511A (en) 2022-08-29 2022-08-29 CTR position depolarization method combining adjacent positions and double history sequences

Country Status (2)

Country Link
CN (1) CN115564511A (en)
WO (1) WO2024045394A1 (en)

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20100153370A1 (en) * 2008-12-15 2010-06-17 Microsoft Corporation System of ranking search results based on query specific position bias
CN112487278A (en) * 2019-09-11 2021-03-12 华为技术有限公司 Training method of recommendation model, and method and device for predicting selection probability
CN111177575B (en) * 2020-04-07 2020-07-24 腾讯科技（深圳）有限公司 Content recommendation method and device, electronic equipment and storage medium

Also Published As

Publication number Publication date
WO2024045394A1 (en) 2024-03-07

Similar Documents

Publication Publication Date Title
Schröter et al. Citizen science for assessing ecosystem services: Status, challenges and opportunities
CN107105320B (en) A kind of Online Video temperature prediction technique and system based on user emotion
Stafford et al. Eu-social science: the role of internet social networks in the collection of bee biodiversity data
CN108268441B (en) Sentence similarity calculation method, device and system
CN101754056A (en) Digital content inventory management system supporting automatic mass data processing and the method thereof
CN114565826A (en) Agricultural pest and disease identification and diagnosis method, system and device
Rong et al. Pest Identification and Counting of Yellow Plate in Field Based on Improved Mask R‐CNN
CN108806355B (en) Painting and calligraphy art interactive education system
CN108182597A (en) A kind of clicking rate predictor method based on decision tree and logistic regression
CN112528010B (en) Knowledge recommendation method and device, computer equipment and readable storage medium
CN112967112A (en) Electronic commerce recommendation method for self-attention mechanism and graph neural network
Mesaglio et al. Recognition and completeness: two key metrics for judging the utility of citizen science data
JP2023545896A (en) Feature combination recommendation algorithm framework that combines sequence information
CN112036659A (en) Social network media information popularity prediction method based on combination strategy
CN113239159A (en) Cross-modal retrieval method of videos and texts based on relational inference network
CN110704510A (en) User portrait combined question recommendation method and system
CN112000889A (en) Information gathering and presenting system
CN112613548A (en) User customized target detection method, system and storage medium based on weak supervised learning
CN113065342B (en) Course recommendation method based on association relation analysis
CN113361928A (en) Crowdsourcing task recommendation method based on special-pattern attention network
CN113095883B (en) Video payment user prediction method and system based on deep cross attention network
CN111753151B (en) Service recommendation method based on Internet user behavior
CN115564511A (en) CTR position depolarization method combining adjacent positions and double history sequences
Karthikeyan et al. Machine learning techniques application: social media, agriculture, and scheduling in distributed systems
CN116542687A (en) Digital collection attribute information analysis processing method and device

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination