CN115809374B - Method, system, device and storage medium for correcting mainstream deviation of recommendation system - Google Patents

Method, system, device and storage medium for correcting mainstream deviation of recommendation system Download PDF

Info

Publication number
CN115809374B
CN115809374B CN202310104256.6A CN202310104256A CN115809374B CN 115809374 B CN115809374 B CN 115809374B CN 202310104256 A CN202310104256 A CN 202310104256A CN 115809374 B CN115809374 B CN 115809374B
Authority
CN
China
Prior art keywords
user
mainstream
dynamic
vector
interaction
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.)
Active
Application number
CN202310104256.6A
Other languages
Chinese (zh)
Other versions
CN115809374A (en
Inventor
张海仙
张宽易
谢敏
张懿
谌祖港
黄粱可汗
李欣洋
尚文一
尹腾
杨雨奇
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.)
Sichuan University
Original Assignee
Sichuan University
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 Sichuan University filed Critical Sichuan University
Priority to CN202310104256.6A priority Critical patent/CN115809374B/en
Publication of CN115809374A publication Critical patent/CN115809374A/en
Application granted granted Critical
Publication of CN115809374B publication Critical patent/CN115809374B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Abstract

The invention discloses a method, a system, equipment and a storage medium for correcting mainstream deviation of a recommendation system, belongs to mainstream deviation correction of the recommendation system in the field of data recommendation, and aims to solve the technical problems of low overall recommendation accuracy and low recommendation fairness caused by the fact that the mainstream deviation is not fully considered by the recommendation system in the prior art. According to the method, the mainstream score of the user is calculated and weighted, then the weighted user interaction data is used as training data of model training, through the method, the mainstream degree of the user can be considered in the reconstruction of the model, the phenomenon of excessively recommending popular articles can be avoided, the influence of mainstream deviation on the recommendation system can be effectively reduced, the effect of the recommendation system on a wider user group is improved, higher fairness is realized, the overall recommendation accuracy of the recommendation system is improved, and the recommendation fairness of the recommendation system is higher.

Description

Method, system, device and storage medium for correcting mainstream deviation of recommendation system
Technical Field
The invention belongs to the technical field of data recommendation, relates to the field of mainstream deviation correction of a recommendation system, and particularly relates to a method, a system, equipment and a storage medium for correcting mainstream deviation of the recommendation system.
Background
With the expansion of the scale of the internet and information systems, the amount of information generated by the internet and information systems also shows an explosive growth situation, and the difficulty of a user in retrieving the content desired by the user from mass information is greatly improved. At present, various websites use a recommendation system in the background, the recommendation system calculates candidate products most suitable for being recommended to a user according to the access characteristics of the user, and then the candidate products are displayed to the user for selection. As a technology capable of effectively solving information overload, the recommendation system can filter out the most interesting part of the user from massive contents according to personalized requirements of different users, so that the recommendation system is widely applied to the fields of e-commerce, video and audio entertainment, accurate advertisement delivery and the like. The collaborative filtering method is a common method in a recommendation system, and can be divided into user-based collaborative filtering and article-based collaborative filtering according to different targets. The collaborative filtering based on the users takes the users as the center, and the idea is to recommend articles similar to the users to the target users; item-based collaborative filtering is item-centric, with the idea being to recommend similar items to a target item to a user who likes the current item. However, due to the influence of the user crowd effect, the platform display mechanism, the quality difference of the articles, and other factors, a phenomenon that the interaction of many users is concentrated on a small part of the articles easily occurs. The traditional collaborative filtering method is easy to learn the deviation in the training process, so that the small part of popular items tend to be recommended, and the large part of the popular items are difficult to obtain the recommended opportunity, so that the recommendation result cannot reflect the real preference of the user. This causes that the recommendation effect received by the mainstream users who like to pursue the hot spot is often very good, while the recommendation effect received by the rest of the broader user groups is not satisfactory, and the received recommendation effect of different users is greatly different due to different mainstream degrees, which is the mainstream deviation phenomenon in the recommendation system.
Most of the existing deviation correcting schemes start from the perspective of articles, namely, the mainstream deviation is indirectly reduced by correcting the popularity deviation of the articles in a recommendation system. The invention patent application with the application number of CN202110218946.5 discloses a causal reasoning method for correcting popularity deviation of a recommendation system, which comprises the following steps: acquiring a matching score of a user and an article in a current recommendation system; predicting an item score according to the popularity of the item, and predicting a user score according to the preference of the user; and aggregating the matching scores of the user and the articles, the article scores and the user scores, predicting the matching scores of the user and the articles, and removing the influence caused by the popularity deviation to obtain the final matching scores of the user and the articles. The method is a model-independent counterfactual reasoning framework, can be suitable for various recommendation systems, improves the recommendation performance of the recommendation system by eliminating the popularity deviation, and can provide high-quality and accurate personalized recommendation content for users. The method is the same as other article-based collaborative filtering methods, mainly aims at improving the phenomenon that recommendation is concentrated on a small part of popular articles, reduces the influence of the popular articles on the overall model recommendation decision in the training process by adopting modes such as inverse tendency fraction weighting, and the like, and simultaneously gives higher weight to the long-tail articles to increase the recommendation probability of the long-tail articles.
In recent years, there are also methods for removing mainstream deviation from the perspective of users, such as adjusting weights of different users in a training process, training a model separately for user groups with different preferences, and the like, so as to enhance capturing capability of the model for preferences of a specific user group. The invention patent application with the application number of CN201911056270.3 discloses a recommendation list re-ranking method for improving the diversity of a recommendation system. The method is the same as other collaborative filtering based on the user, different requirements of the user on diversity of the recommendation list can be considered, so that the recommended articles are more fit for real feeling of people, grading deviation of different users on the same article is also considered, the diversity is properly improved on the balance of accuracy and diversity, and the influence on the accuracy is small.
The method for correcting the mainstream deviation of the recommendation system can actually expand the recommendation range of the recommendation system, so that the recommendation system can not be limited to a part of popular articles, but can take care of some long-tail articles, and the fairness problem in article recommendation is solved to a certain extent. However, this does not mean that these long-tailed items can be recommended to the appropriate users, but rather reduces the accuracy of the recommendation system if recommended to mainstream users who prefer to pursue hot spots. Therefore, a method of correcting the deviation of popularity alone does not necessarily play a positive role in correcting the deviation of the mainstream. The existing method for directly correcting the mainstream deviation also has a certain problem, and the effect of the part of users is easily damaged by reducing the weight of the mainstream user in the training process, so that the overall accuracy of the recommendation system is reduced; the method for training different models separately for different user groups also has problems, and the division of the user groups, the training of a plurality of models and the consumption during integration all make the method difficult to realize in the actual production environment. In addition, the existing method for correcting the mainstream deviation does not consider the characteristic of the mainstream change, a group of users belonging to the mainstream at present are not necessarily mainstream users in the past, and a group of users which are not mainstream in the past can also become mainstream users in the future due to pursuit of hot spots.
Disclosure of Invention
The invention aims to: in order to solve the technical problems of low overall recommendation accuracy and low recommendation fairness caused by the fact that a recommendation system does not fully consider mainstream deviation in the prior art, the invention provides a method, a system, equipment and a storage medium for correcting the mainstream deviation of the recommendation system.
The invention specifically adopts the following technical scheme for realizing the purpose:
a method for correcting mainstream deviation of a recommendation system comprises the following steps:
step S1, data collection and processing
Obtaining user information, article information and user article interaction information in a recommendation system, and respectively constructing user co-occurrence vectors
Figure SMS_1
The co-occurrence vector of the object is greater or less than>
Figure SMS_2
；
Step S2, calculating the mainstream score
According to co-occurrence vector of articles
Figure SMS_4
Calculating the total number of times of interaction of the object>
Figure SMS_8
(ii) a Based on the user co-occurrence vector->
Figure SMS_10
Calculating the total number of user interactions>
Figure SMS_6
(ii) a Based on total number of times of interaction of items>
Figure SMS_9
Total number of user interactions->
Figure SMS_11
Item category, count user->
Figure SMS_13
Is greater than or equal to the dynamic main flow degree score of>
Figure SMS_5
(ii) a According to user>
Figure SMS_7
Dynamic mainstream level score of
Figure SMS_12
Computing stationObtaining the global dynamic mainstream degree score by the average value of the dynamic mainstream degree scores of the users
Figure SMS_14
And forming the global dynamic mainstream degree scores of all the article categories into a global dynamic mainstream degree vector
Figure SMS_3
；
S3, constructing a dynamic mainstream degree characteristic model
Constructing a dynamic mainstream degree characteristic model based on a three-layer perceptron MLP model, wherein the first two layers of the dynamic mainstream degree characteristic model use a ReLU function as an activation function, and the last layer of the dynamic mainstream degree characteristic model uses a softmax activation function; with user information vectors
Figure SMS_15
And the global dynamic main flow degree vector ≥ output in step S2>
Figure SMS_16
Spliced and used as input of a dynamic mainstream degree characteristic model>
Figure SMS_17
Outputting a dynamic mainstream characteristic hidden vector by the dynamic mainstream degree characteristic model;
s4, constructing a collaborative filtering module
Constructing a collaborative filtering module comprising an encoder and a decoder;
the encoder is constructed by adopting a three-layer perceptron MLP model; user interaction data
Figure SMS_19
An input encoder that calculates user interaction data &>
Figure SMS_22
And generates &, respectively>
Figure SMS_24
Mean and>
Figure SMS_20
a variance constituting a mean vector &' for the user>
Figure SMS_21
And the variance vector pick>
Figure SMS_25
Wherein both vectors are in the t dimension, forming a mean vector ≥ for the user>
Figure SMS_26
And the variance vector pick>
Figure SMS_18
Wherein both vectors are t-dimensional, and then generating a h-dimensional user interaction hidden vector ≥ by random sampling>
Figure SMS_23
；
The decoder is constructed by adopting a four-layer perceptron MLP model, the first three layers of activation functions of the decoder are tanh functions, and the last layer of activation functions of the decoder are softmax functions; step S3, outputting the dynamic mainstream characteristic hidden vector and the user interaction hidden vector output by the encoder
Figure SMS_27
The decoder output reconstructs the user interaction data ≥ as input from the decoder>
Figure SMS_28
And reconstructing the dynamic mainstream feature vector->
Figure SMS_29
Reconstructing a dynamic mainstream feature vector>
Figure SMS_30
For completing the reconstruction of the decoder;
step S5, recommendation result generation
Inputting the user number to be predicted according to the collaborative filtering module obtained by the training completion in the step S4Previously observed user interaction data
Figure SMS_31
The encoder outputs a user interaction hidden vector ≥>
Figure SMS_32
(ii) a Then the user interaction is hidden by the vector->
Figure SMS_33
And inputting the dynamic mainstream characteristic hidden vector output by the step S3 into a decoder, and outputting reconstructed user interaction data by the decoder.
Further, in step S1, an arbitrary user is constructed according to the user item interaction information
Figure SMS_34
User co-occurrence vector with all items>
Figure SMS_35
Any item/device is constructed according to user and item interaction information>
Figure SMS_36
Co-occurrence vector with all items>
Figure SMS_37
；
Wherein, the first and the second end of the pipe are connected with each other,
Figure SMS_39
represents the total number of users, and>
Figure SMS_43
indicates the total number of items, based on the number of items present in the container>
Figure SMS_46
Indicates a item +>
Figure SMS_38
Indicates the fifth->
Figure SMS_42
Each item>
Figure SMS_44
Represents a user>
Figure SMS_47
And (iv) an article>
Figure SMS_40
In a predetermined interaction condition, based on the presence of a predetermined condition in the system, and>
Figure SMS_41
indicates that the article is present>
Figure SMS_45
And the user->
Figure SMS_48
The interaction scenario of (2).
Further, in step S2, according to the article
Figure SMS_49
Co-occurrence vector with an item->
Figure SMS_50
And calculating the total number of times of interaction of the article>
Figure SMS_51
；
According to the user
Figure SMS_52
Co-occurrence with user vector &>
Figure SMS_53
Calculating the total number of times of user interaction, based on the calculated total number of times of user interaction>
Figure SMS_54
；
According to the total number of interaction times of the articles
Figure SMS_56
Total number of user interactions->
Figure SMS_59
Articles, and the likeDetermining whether the user is present or not>
Figure SMS_62
Is greater than or equal to the dynamic main flow degree score of>
Figure SMS_55
(ii) a In case the number of interactions is not sensitive, it is ≥ for the class>
Figure SMS_58
Article of, user
Figure SMS_61
Is greater than or equal to the dynamic main flow degree score of>
Figure SMS_64
Calculating according to the formula (1); in the case of sensitive number of interactions, a @ class>
Figure SMS_57
User->
Figure SMS_60
Is greater than or equal to the dynamic main flow degree score of>
Figure SMS_63
Calculating according to the formula (2);
Figure SMS_65
（1）
Figure SMS_66
（2）
according to the user
Figure SMS_67
Dynamic mainstream level score of>
Figure SMS_68
Calculating the average of the dynamic mainstream level scores of all users to obtain a global dynamic mainstream level score>
Figure SMS_69
The calculation formula is as follows:
Figure SMS_70
then, the global dynamic mainstream degree scores of all the article categories are divided
Figure SMS_71
Constitute a dimension of->
Figure SMS_72
Is determined by the global dynamic prevailing degree vector->
Figure SMS_73
Expressed as:
Figure SMS_74
wherein the content of the first and second substances,
Figure SMS_76
、/>
Figure SMS_79
each represents a time +>
Figure SMS_81
Representing a hyper-parameter (for controlling a logarithmic curve)>
Figure SMS_75
Indicates that the article is present>
Figure SMS_78
Belongs to the category->
Figure SMS_82
，/>
Figure SMS_83
Represents a total number of users, <' > based on>
Figure SMS_77
Represents a collection of all users, and>
Figure SMS_80
representing the total number of item categories.
Furthermore, a co-occurrence vector weighting process is also performed, and the weighting process is described as: user will be connected
Figure SMS_84
To belong to the category
Figure SMS_85
In combination with a sun or a sun/sun unit>
Figure SMS_86
In the interaction->
Figure SMS_87
Multiply by user>
Figure SMS_88
Is in the category->
Figure SMS_89
On a dynamic main flow degree score->
Figure SMS_90
；
For user co-occurrence vectors
Figure SMS_91
Each item in the list is weighted, and the co-occurrence vector of the whole user is obtained after the weighting is finished
Figure SMS_92
Normalization using a softmax function results in user interaction data ≥ being used for input to the collaborative filtering module>
Figure SMS_93
：
Figure SMS_94
Wherein any of the user interaction data
Figure SMS_95
Has a value range of [0,1]。
Further, in step S3, the user information vector
Figure SMS_96
Expressed as:
Figure SMS_97
wherein, the first and the second end of the pipe are connected with each other,
Figure SMS_98
indicates that the user is pickand place>
Figure SMS_99
In the age database, based on age information, based on the age value of the subject>
Figure SMS_100
Indicates that the user is pickand place>
Figure SMS_101
The binary sex information of (2);
input of dynamic mainstream degree characteristic model
Figure SMS_102
Expressed as:
Figure SMS_103
；
wherein the content of the first and second substances,
Figure SMS_104
representing a vector stitching operation.
Further, in step S4, the loss function of the filter module is cooperated
Figure SMS_105
Is divided into a reconstruction target loss->
Figure SMS_106
Approximate loss of distribution>
Figure SMS_107
And dynamic mainstream feature vector approximation loss>
Figure SMS_108
Three portions, the loss function>
Figure SMS_109
The calculation formula of (c) is:
Figure SMS_110
reconstructing object losses
Figure SMS_111
The calculation formula of (c) is:
Figure SMS_112
distribution approximation loss
Figure SMS_113
The calculation formula of (c) is:
Figure SMS_114
dynamic mainstream feature vector approximation loss
Figure SMS_115
The calculation formula of (2) is as follows:
Figure SMS_116
wherein the content of the first and second substances,
Figure SMS_126
represents a user interaction hidden vector, <' > or>
Figure SMS_118
Represents user interaction data, and->
Figure SMS_121
Represents a posterior distribution of each user data sample, <' > based on the data sample>
Figure SMS_127
Represents a variation profile, a variation profile>
Figure SMS_133
And the posterior distribution->
Figure SMS_131
In the approximation that the difference between the first and second values,
Figure SMS_134
represents->
Figure SMS_130
Is paired and/or matched>
Figure SMS_132
Is desired, is based on>
Figure SMS_119
Is hyperparameter, is greater than or equal to>
Figure SMS_125
Is a user interaction hidden vector->
Figure SMS_120
KL represents KL divergence, and>
Figure SMS_122
represents a prior distribution, <' > or>
Figure SMS_123
Represents a variance vector, < > based on the variance>
Figure SMS_129
Represents the square of the mean vector, <' > is selected>
Figure SMS_117
Is hyperparameter, is greater than or equal to>
Figure SMS_124
Represents the original dynamic mainstream feature vector, <' > or>
Figure SMS_128
Representing the reconstructed dynamic mainstream feature vector.
A system for correcting deviations in a recommendation system mainstream comprising:
the data collection and processing module is used for acquiring the user information, the article information and the user article interaction information in the recommendation system and respectively constructing the user co-occurrence vectors
Figure SMS_135
The co-occurrence vector of the object is greater or less than>
Figure SMS_136
；
A mainstream score calculation module for calculating a mainstream score according to the co-occurrence vector of the article
Figure SMS_138
Calculate the total number of times of the interaction of the article->
Figure SMS_143
(ii) a Based on the user co-occurrence vector->
Figure SMS_145
Calculating the total number of times of the user's interaction>
Figure SMS_140
(ii) a Based on the total number of times of interaction of the article>
Figure SMS_141
Total number of user interactions->
Figure SMS_144
Item category, count user->
Figure SMS_147
Dynamic primary flow degree ofNumber/device>
Figure SMS_139
(ii) a According to the user
Figure SMS_142
Is greater than or equal to the dynamic main flow degree score of>
Figure SMS_146
Calculating the average value of the dynamic mainstream degree scores of all the users to obtain a global dynamic mainstream degree score->
Figure SMS_148
And the global dynamic mainstream level scores of all the item classes are combined into a global dynamic mainstream level vector->
Figure SMS_137
；
The dynamic mainstream degree characteristic model building module is used for building a dynamic mainstream degree characteristic model based on a three-layer perceptron MLP model, the ReLU function is used as an activation function in the first two layers of the dynamic mainstream degree characteristic model, and the softmax activation function is used in the last layer of the dynamic mainstream degree characteristic model; with user information vectors
Figure SMS_149
And a global dynamic mainstream level vector output by the mainstream score calculation module>
Figure SMS_150
Spliced and used as input of a dynamic mainstream degree characteristic model>
Figure SMS_151
Outputting a dynamic mainstream characteristic hidden vector by the dynamic mainstream degree characteristic model;
the collaborative filtering module construction module is used for constructing a collaborative filtering module comprising an encoder and a decoder;
the encoder is constructed by adopting a three-layer perceptron MLP model; user interaction data
Figure SMS_152
An input encoder that calculates user interaction data &>
Figure SMS_153
And generates &, respectively>
Figure SMS_154
Mean and->
Figure SMS_155
Variance, forming a mean vector ≥ of users>
Figure SMS_156
And the variance vector pick>
Figure SMS_157
Wherein the two vectors are in t dimension, and then a user interactive hidden vector ^ in h dimension is generated through random sampling>
Figure SMS_158
；
The decoder is constructed by adopting a four-layer perceptron MLP model, the first three layers of activation functions of the decoder are tanh functions, and the last layer of activation functions of the decoder are softmax functions; dynamic mainstream characteristic hidden vectors output by the dynamic mainstream degree characteristic model building module and user interaction hidden vectors output by the encoder
Figure SMS_159
The decoder output reconstructs the user interaction data ≥ as input from the decoder>
Figure SMS_160
And reconstructing a dynamic mainstream feature vector>
Figure SMS_161
Reconstructing the dynamic mainstream feature vector->
Figure SMS_162
For completing the reconstruction of the decoder;
recommendation generation module for rootInputting user interaction data currently observed by a user to be predicted according to a collaborative filtering module obtained by training of a collaborative filtering module construction module
Figure SMS_163
The encoder outputs a user interaction hidden vector ≥>
Figure SMS_164
(ii) a Then the user interaction is hidden by the vector->
Figure SMS_165
And inputting the dynamic mainstream characteristic hidden vector output by the dynamic mainstream degree characteristic model building module into a decoder, and outputting reconstructed user interaction data by the decoder.
A computer device comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of the above method.
A computer-readable storage medium, in which a computer program is stored which, when executed by a processor, causes the processor to carry out the steps of the above-mentioned method.
The invention has the following beneficial effects:
1. in the invention, in the process of dynamically correcting the mainstream deviation, the mainstream score of the user is calculated and weighted, and then the weighted user interaction data is used as the training data of model training.
2. In the invention, a collaborative filtering module based on an asymmetric variational self-encoder is constructed, and the capability of capturing and utilizing dynamic mainstream characteristics of a model is enhanced through asymmetric structural design and introduction of dynamic mainstream characteristic vectors.
3. In the invention, two scenes of sensitive interaction times and insensitive interaction times are fully considered, a method for dynamically calculating the mainstream degree score of the user is provided, the mainstream degree score can be used as input data of a variational self-encoder through a weighting normalization process, the influence of mainstream deviation is fully considered, the overall recommendation accuracy of the recommendation system is improved, and the recommendation fairness of the recommendation system is higher.
Drawings
FIG. 1 is a schematic flow diagram of the present invention;
FIG. 2 is a detailed flow diagram of the present invention;
FIG. 3 is a schematic diagram of the structure of the collaborative filtering module according to the present invention.
Detailed Description
Example 1
The embodiment provides a method for correcting a mainstream deviation of a recommendation system, a flow diagram of which is shown in fig. 1, and the method comprises the steps of data collection and processing, mainstream score calculation, dynamic mainstream degree feature model construction, collaborative filtering module construction and recommendation result generation. The detailed flow diagram of the method is shown in fig. 2, and specifically includes:
step S1, data collection and processing
Obtaining user information, article information and user article interaction information in a recommendation system, and respectively constructing user co-occurrence vectors
Figure SMS_166
The co-occurrence vector of the object is greater or less than>
Figure SMS_167
。
Constructing any user according to user article interaction information
Figure SMS_169
User co-occurrence vector with all items
Figure SMS_173
，/>
Figure SMS_176
Indicates that the user is pickand place>
Figure SMS_171
And article>
Figure SMS_174
The interaction scenario of (2). In scenarios where the number of interactions is not sensitive (only regarding whether there are interactions, but not regarding the number of interactions, such as movie recommendations, book recommendations, etc.), if an explicit interaction is generated, then ≧>
Figure SMS_177
Otherwise->
Figure SMS_179
(ii) a In sensitive cases of number of interactions (concern about whether there is an interaction, also concerning the number of interactions, such as music recommendations, short video recommendations, etc.), if an explicit interaction is generated ≧>
Figure SMS_170
Wherein->
Figure SMS_172
Indicates that the user is pickand place>
Figure SMS_175
And article>
Figure SMS_178
The number of interactions; otherwise->
Figure SMS_168
。
Constructing any article according to user article interaction information
Figure SMS_181
Co-occurrence vector with all articles
Figure SMS_185
，/>
Figure SMS_188
Indicating a goods>
Figure SMS_180
And the user->
Figure SMS_186
The interaction scenario of (2). In sensitive interaction times scenarios (only regarding whether there is an interaction, but not regarding the number of interactions, such as movie recommendations, book recommendations, etc.), if an explicit interaction is generated, then ≧>
Figure SMS_187
Otherwise->
Figure SMS_190
. In sensitive cases of number of interactions (concern about whether there is an interaction, also concerning the number of interactions, such as music recommendations, short video recommendations, etc.), if an explicit interaction is generated ≧>
Figure SMS_183
Wherein->
Figure SMS_184
Indicates that the article is present>
Figure SMS_189
And the user->
Figure SMS_191
The number of interactions; otherwise->
Figure SMS_182
。
Wherein the content of the first and second substances,
Figure SMS_192
represents the total number of users, and>
Figure SMS_193
indicates the total number of items, based on the number of items present in the container>
Figure SMS_194
Indicates a substance is present>
Figure SMS_195
Indicates the fifth->
Figure SMS_196
Each item>
Figure SMS_197
Representing the user.
Data cleaning is mainly to filter out part of users and articles according to a threshold (for example, to filter out articles with interaction times smaller than a certain threshold), and the purpose is to clear abnormal data to ensure normal operation of a recommendation process.
Step S2, calculating the mainstream score
The step mainly quantizes the user and the global mainstream degree, so that a dynamic mainstream degree feature vector can be generated in the step S3 conveniently. Since the concept of mainstream level involves both individual users and overall users, the mainstream level scores of individual users and global can be calculated separately here.
According to the article
Figure SMS_198
Co-occurrence vector with the item->
Figure SMS_199
The total number of interactions of the item can be calculated>
Figure SMS_200
. Since the degree of main flow is a dynamically changing concept, the number of item interactions can be filtered over time, if only the occurrence is taken into account>
Figure SMS_201
Interactions within a time period, then only the release date is considered @>
Figure SMS_202
Previous article with total number of article interactions of
Figure SMS_203
In which arbitrary objects interactNumber of times>
Figure SMS_204
Only data within this time period is considered.
According to the user
Figure SMS_205
Co-occurrence vector with user->
Figure SMS_206
The total number of interactions by the user can be calculated>
Figure SMS_207
. Since the degree of main flow is a dynamically changing concept, the number of user interactions can be filtered over time, if only the occurrence is taken into account>
Figure SMS_208
And (3) the total interaction times of the user in the time period are as follows: />
Figure SMS_209
Wherein arbitrarily->
Figure SMS_210
Considering only the interaction data during this time period, the release date is @>
Figure SMS_211
The subsequent item interaction data are all set to 0.
According to the idea of collaborative filtering, the degree of user's mainstream depends on whether their interactive items are interacted with by other users. Meanwhile, the method considers that the preference degrees of users to different categories of articles are inconsistent, so the influence of the article category factor is also considered in the calculation process of the mainstream degree score of the user. According to the total number of interactions of the article
Figure SMS_214
Total number of user interactions->
Figure SMS_217
Articles, and the likeDetermining whether the user is present or not>
Figure SMS_220
Is greater than or equal to the dynamic main flow degree score of>
Figure SMS_212
. In case the number of interactions is not sensitive, it is ≥ for the class>
Figure SMS_216
User->
Figure SMS_219
Is greater than or equal to the dynamic main flow degree score of>
Figure SMS_221
Calculating according to the formula (1); in the case of sensitive number of interactions, a @ class>
Figure SMS_213
User->
Figure SMS_215
Dynamic mainstream degree score of
Figure SMS_218
Calculating according to the formula (2);
Figure SMS_222
（1）
Figure SMS_223
（2）
wherein the content of the first and second substances,
Figure SMS_225
will be/are>
Figure SMS_228
In a time period, the user->
Figure SMS_232
All items that have interacted add up to the number of interactions of all users. Since the power law distribution phenomenon exists in the interaction records of real-world items (namely a small part of items occupy most of interactions, and a large part of items have no interactions), the total interaction times of each item and other users are combined by using a logarithmic function>
Figure SMS_226
Performing inhibition, and counting at the bottom>
Figure SMS_229
Is a hyper-parameter. In case the number of interactions is not sensitive, the user's dynamic mainstream score is ≥ based on ≥ the number of interactions>
Figure SMS_230
Each interaction record is given the same weight. In the case where the number of interactions is sensitive, the user's dynamic mainstream score is ≧ or>
Figure SMS_233
Assigning a different weight, based on the number of interactive records, to each of the interactive records>
Figure SMS_224
Represents a user>
Figure SMS_227
And article>
Figure SMS_231
This means that the greater the number of user interactions, the greater the weight of the item in the user mainstream level score evaluation.
According to the user
Figure SMS_234
Is greater than or equal to the dynamic main flow degree score of>
Figure SMS_235
To for category is ^ er>
Figure SMS_236
Is determined, the global dynamic mainstream degree score->
Figure SMS_237
Is the average of all users' dynamic mainstream level scores, so a global dynamic mainstream level score ≧>
Figure SMS_238
The calculation formula of (c) is:
Figure SMS_239
then, the global dynamic mainstream degree scores of all the article categories are divided
Figure SMS_240
Constitute a dimension of->
Figure SMS_241
Is determined by the global dynamic prevailing degree vector->
Figure SMS_242
Expressed as:
Figure SMS_243
wherein the content of the first and second substances,
Figure SMS_245
、/>
Figure SMS_248
all indicate a time>
Figure SMS_250
Represents a hyper-parameter (for controlling a logarithmic curve), -is>
Figure SMS_246
Indicates that the article is present>
Figure SMS_249
Belongs to the category->
Figure SMS_251
，/>
Figure SMS_252
Represents the total number of users, and>
Figure SMS_244
represents a collection of all users, and>
Figure SMS_247
representing the total number of item categories.
And then weighting the co-occurrence vectors of the users according to the obtained mainstream degree scores of the users, namely introducing the mainstream information of the users into the co-occurrence vectors. The weighting process is described as: user will be
Figure SMS_253
For belonging to the category of->
Figure SMS_254
Is based on>
Figure SMS_255
In the interaction->
Figure SMS_256
Multiplied by user pick>
Figure SMS_257
Is in the category->
Figure SMS_258
Dynamic mainstream level score of &>
Figure SMS_259
；
For user co-occurrence vectors
Figure SMS_260
Each item in the list is weighted, and the co-occurrence vector of the whole user is obtained after the weighting is finished
Figure SMS_261
Using softmax functionRow normalization to obtain user interaction data ≥ for input into the collaborative filtering module>
Figure SMS_262
：
Figure SMS_263
Wherein any of the user interaction data
Figure SMS_264
Has a value range of [0,1]。
User interaction data obtained here
Figure SMS_265
As well as the input of the encoder followed by the collaborative filtering module.
S3, constructing a dynamic mainstream degree characteristic model
Can be calculated according to the step (2) in
Figure SMS_266
Time period arbitrary user>
Figure SMS_269
Respect to any type of item set >>
Figure SMS_272
Is greater than or equal to the prevailing degree score of>
Figure SMS_268
. For any user->
Figure SMS_270
For all categories of items, the main flow degree scores are calculated, which can form a ^ based on>
Figure SMS_271
Vector of dimensions, denoted as
Figure SMS_273
The vector may characterize the user
Figure SMS_267
The mainstream degree of all the item categories, the larger the score of the mainstream degree of an item category is, the more the user prefers popular items in the category, and the smaller the score is, the more the user prefers popular items in the category.
In order to enable the collaborative filtering model in step S4 to make full use of the user information and the global information, a dynamic mainstream feature expression based on the user information and the global information needs to be obtained. A dynamic mainstream level feature model is thus constructed here. And constructing a dynamic mainstream degree feature model based on the three-layer perceptron MLP model, wherein the dynamic mainstream degree feature model is used for extracting key features in user information and global information and reducing the original input into a hidden vector with a lower dimensionality. The first two layers of the dynamic mainstream degree characteristic model use a ReLU function as an activation function, and the last layer of the dynamic mainstream degree characteristic model uses a softmax activation function. With user information vectors
Figure SMS_274
And the global dynamic main flow degree vector ≥ output in step S2>
Figure SMS_275
Spliced and used as input of a dynamic mainstream degree characteristic model>
Figure SMS_276
。
The user information vector
Figure SMS_277
Expressed as:
Figure SMS_278
wherein the content of the first and second substances,
Figure SMS_279
indicates that the user is pickand place>
Figure SMS_280
Is quantified age information and is greater or less than>
Figure SMS_281
Indicates that the user is pickand place>
Figure SMS_282
The binary gender information of (1);
inputting the dynamic mainstream degree characteristic model
Figure SMS_283
Expressed as:
Figure SMS_284
；
wherein the content of the first and second substances,
Figure SMS_285
representing vector stitching operations
Input device
Figure SMS_286
And outputting the dynamic mainstream characteristic hidden vector after dimensionality reduction of the three-layer perceptron of the dynamic mainstream degree characteristic model.
S4, constructing a collaborative filtering module
An asymmetric variational self-encoder (VAE) is used as a main structure of the collaborative filtering model, and the structure is shown in fig. 3. The method adopts an asymmetric variational self-encoder to carry out collaborative filtering, and aims to add extra dynamic mainstream degree information into a hidden layer between an encoder and a decoder, so that the decoder can directly utilize personal information of a user and global mainstream degree during decoding.
The collaborative filtering module comprises an encoder and a decoder;
the encoder is constructed by adopting a three-layer perceptron MLP model. User interaction data
Figure SMS_287
An input encoder that calculates user interaction data &>
Figure SMS_291
And generates £ respectively>
Figure SMS_293
Mean and->
Figure SMS_288
Variance, forming a mean vector ≥ of users>
Figure SMS_290
And the variance vector pick>
Figure SMS_292
Wherein both vectors are t-dimensional, and then generating a h-dimensional user interaction hidden vector ≥ by random sampling>
Figure SMS_294
，
Figure SMS_289
. Since the network cannot perform back propagation due to the random sampling method, the re-parameterization method is adopted to complete the sampling process.
The decoder is constructed by adopting a four-layer perceptron MLP model, the first three layers of activation functions of the decoder are tanh functions, and the last layer of activation functions of the decoder are softmax functions and are used for generating probability distribution
Figure SMS_295
. The dynamic mainstream characteristic hidden vector output in the step S3 and the user interaction hidden vector output by the encoder->
Figure SMS_296
The decoder output reconstructs the user interaction data ≥ as input from the decoder>
Figure SMS_297
And reconstructing the dynamic mainstream feature vector->
Figure SMS_298
Reconstruction of dynamic mainstream featuresSign vector->
Figure SMS_299
For completing the reconstruction of the decoder.
The variational autoencoder reasoning process is as follows, assuming the user
Figure SMS_300
Corresponding user interaction hidden vector ≥>
Figure SMS_301
Compliance
Figure SMS_302
Is normally distributed. Based on the recommender system interaction data characteristic, it is assumed that user interaction data entered into the encoder is @>
Figure SMS_303
Obey probability is->
Figure SMS_304
The likelihood function of the polynomial distribution of (2) is as follows:
Figure SMS_305
wherein the content of the first and second substances,
Figure SMS_306
indicating that the vector is hidden by a user interaction>
Figure SMS_307
Determined and/or>
Figure SMS_308
Probability of individual item interaction>
In order to enable the network to learn the parameters by back-propagation, a posteriori distributions for each user data sample must be found
Figure SMS_309
Since this posterior distribution is not easy to find, it is used here by using a variation inferenceDistribution of variation
Figure SMS_318
To approximate>
Figure SMS_319
. Hypothesis->
Figure SMS_310
Satisfies a Gaussian distribution->
Figure SMS_312
Wherein->
Figure SMS_314
Is a variance vector pick>
Figure SMS_316
The diagonal covariance matrix of (2). The optimization goal of the network at this point is that the optimization parameter generates a mean vector ≥ then>
Figure SMS_311
And the variance vector pick>
Figure SMS_313
So that the differentiation profile->
Figure SMS_315
And the posterior distribution->
Figure SMS_317
As similar as possible.
Wherein the reparameterization method operates as follows, assuming noise
Figure SMS_320
Obey a normal distribution>
Figure SMS_321
Hidden vector of user interaction
Figure SMS_322
Can be determined by the variance vector>
Figure SMS_323
Mean vector @>
Figure SMS_324
And noise are linearly combined, so that the network can learn. The reparameterization formula is as follows:
Figure SMS_325
wherein the content of the first and second substances,
Figure SMS_326
is the standard deviation;
unlike the standard variational self-encoder network, an asymmetric structure is used to obtain the user interaction implicit vector
Figure SMS_327
After that, the input generated in step S3 is ^ ed>
Figure SMS_328
Spliced on/in>
Figure SMS_329
And then fed into the decoder. The generating part of the decoder may be divided into reconstructing user interaction data ≥>
Figure SMS_330
And reconstructing the dynamic mainstream feature vector->
Figure SMS_331
。
In conclusion, loss function of collaborative filtering model based on asymmetric variational self-encoder
Figure SMS_332
Is divided into a reconstruction target loss->
Figure SMS_333
Approximate loss of distribution>
Figure SMS_334
And a dynamic mainstream feature vector approximation loss>
Figure SMS_335
Three sections, loss function>
Figure SMS_336
The calculation formula of (c) is:
Figure SMS_337
reconstructing object loss
Figure SMS_338
The purpose of this is to make the reconstructed user interaction data output by the decoder as identical as possible to the user interaction data input to the encoder, the calculation formula being:
Figure SMS_339
distribution approximation loss
Figure SMS_340
With the aim of making the variational distribution->
Figure SMS_341
As close as possible to a posterior distribution>
Figure SMS_342
For measuring the approximation degree of two distributions, the calculation formula is:
Figure SMS_343
to convert the original dynamic mainstream feature vector
Figure SMS_344
And reconstructing the dynamic mainstream feature vector->
Figure SMS_345
As an objective function, the negative number of which is taken as a loss term, with the aim of reconstructing the dynamic mainstream feature vector ≥ s>
Figure SMS_346
Is directionally correlated with the original dynamic main flow feature vector->
Figure SMS_347
As close as possible, losses are approximated by dynamic mainstream feature vectors>
Figure SMS_348
The method can enable a decoder to complete the reconstruction process by using the dynamic mainstream characteristics as much as possible. Approximate loss of a dynamic mainstream feature vector>
Figure SMS_349
The calculation formula of (2) is as follows:
Figure SMS_350
wherein the content of the first and second substances,
Figure SMS_359
represents a user-interactive hidden vector, < '> based on a user's interaction>
Figure SMS_353
Represents user interaction data, and->
Figure SMS_355
Represents a posterior distribution of each user data sample, <' > based on the data sample>
Figure SMS_354
Represents a variational distribution, a variational distribution->
Figure SMS_357
And the posterior distribution->
Figure SMS_360
In the approximation that the difference between the first and second values,
Figure SMS_363
represents->
Figure SMS_361
Is paired and/or matched>
Figure SMS_365
Is desired, is based on>
Figure SMS_351
Is a hyper-parameter for controlling the penalty degree of the distribution similarity degree on the whole objective function, and is used for judging whether the distribution similarity degree is greater than or equal to the penalty degree>
Figure SMS_358
Is a user interaction hidden vector->
Figure SMS_364
KL represents KL divergence, ->
Figure SMS_367
Represents a prior distribution, <' > or>
Figure SMS_366
Represents a variance vector, < > based on the variance>
Figure SMS_368
Represents the square of the mean vector>
Figure SMS_352
Is hyperparameter, is greater than or equal to>
Figure SMS_356
Represents the original dynamic mainstream feature vector, <' > or>
Figure SMS_362
Representing the reconstructed dynamic mainstream feature vector.
Step S5, recommendation result generation
Inputting the user to be predicted according to the collaborative filtering module obtained by the training completion of the step S4
Figure SMS_369
Currently observed user interaction data &>
Figure SMS_370
The encoder outputs a mean vector &>
Figure SMS_371
And variance vector +>
Figure SMS_372
Then according to the formula->
Figure SMS_373
Calculating to obtain a user interaction hidden vector->
Figure SMS_374
(ii) a Then the user interaction is hidden by the vector->
Figure SMS_375
Inputting the dynamic mainstream characteristic hidden vector output in the step S3 into a decoder, and outputting the reconstructed user interaction data by the decoder, wherein the dimensionality of the user interaction data is n-dimensionality, the dimensionality is the same as the quantity of all articles, and the value of each dimensionality is [0,1 ]]In the meantime.
Aiming at the obtained reconstructed user interaction data, firstly, eliminating articles which do not meet the time requirement, and only considering
Figure SMS_376
And a previous time period, then any issue time is>
Figure SMS_377
And setting the value of the dimension where the subsequent article serial number is positioned as 0. Second, items that have already appeared in the history are culled, and the currently observed interaction is->
Figure SMS_378
And recording the dimension serial numbers with the values not being 0, and setting the dimensions of the reconstructed user interaction data as 0. And finally, sorting the reconstructed user interaction data from large to small, wherein the dimension serial number of top-N is the user to be predicted->
Figure SMS_379
Front ofAnd N item recommendation lists.
Example 2
The embodiment provides a system for correcting mainstream deviation of a recommendation system, which comprises a data collection and processing module, a mainstream score calculation module, a dynamic mainstream degree feature model construction module, a collaborative filtering module construction module and a recommendation result generation module, wherein the specific content of each module is as follows:
the data collection and processing module is used for acquiring the user information, the article information and the user article interaction information in the recommendation system and respectively constructing the user co-occurrence vectors
Figure SMS_380
The co-occurrence vector of the object is greater or less than>
Figure SMS_381
。
Constructing any user according to user article interaction information
Figure SMS_382
User co-occurrence vector with all items
Figure SMS_387
，/>
Figure SMS_389
Indicates that the user is pickand place>
Figure SMS_385
And article>
Figure SMS_386
The interaction scenario of (2). In scenarios where the number of interactions is not sensitive (only regarding whether there are interactions, but not regarding the number of interactions, such as movie recommendations, book recommendations, etc.), if an explicit interaction is generated, then ≧>
Figure SMS_391
Otherwise->
Figure SMS_393
(ii) a In case of sensitive number of interactions (off)Note if there is an interaction, also note the number of interactions, such as music recommendations, short video recommendations, etc.), if an explicit interaction results ≦>
Figure SMS_383
Wherein->
Figure SMS_388
Represents a user>
Figure SMS_390
And article>
Figure SMS_392
The number of interactions; or else>
Figure SMS_384
。
Constructing any article according to user article interaction information
Figure SMS_395
Co-occurrence vector with all articles
Figure SMS_398
，/>
Figure SMS_402
Indicating a goods>
Figure SMS_397
And the user->
Figure SMS_399
The interaction scenario of (2). In sensitive interaction times scenarios (only regarding whether there is an interaction, but not regarding the number of interactions, such as movie recommendations, book recommendations, etc.), if an explicit interaction is generated, then ≧>
Figure SMS_403
Otherwise->
Figure SMS_405
. In the case of sensitive interaction times (concerning whether there is interaction, also concerning the number of interactions)E.g., music recommendation, short video recommendation, etc.)), if an explicit interaction is generated ≦ device>
Figure SMS_394
In which>
Figure SMS_400
Indicates that the article is present>
Figure SMS_401
And the user->
Figure SMS_404
The number of interactions; otherwise->
Figure SMS_396
。
Wherein the content of the first and second substances,
Figure SMS_406
represents a total number of users, <' > based on>
Figure SMS_407
Indicates the total number of items, based on the number of items present in the container>
Figure SMS_408
Indicates a substance is present>
Figure SMS_409
Indicates the fifth->
Figure SMS_410
Each item>
Figure SMS_411
Representing the user.
Data cleaning is mainly to filter out part of users and articles according to a threshold (for example, to filter out articles with interaction times smaller than a certain threshold), and the purpose is to clear abnormal data to ensure normal operation of a recommendation process.
And the mainstream score calculating module is mainly used for quantizing the user and the global mainstream degree and facilitating the generation of the dynamic mainstream degree characteristic vector in the dynamic mainstream degree characteristic model building module. Since the concept of mainstream level involves both individual users and overall users, the mainstream level scores of individual users and global can be calculated separately here.
According to the article
Figure SMS_412
Co-occurrence vector with the item->
Figure SMS_413
The total number of interactions of the item can be calculated>
Figure SMS_414
. Since the degree of main flow is a dynamically changing concept, the number of item interactions can be filtered over time, if only the occurrence is taken into account>
Figure SMS_415
Interactions within a time period, then only the release date is considered @>
Figure SMS_416
Previous article with total number of article interactions of
Figure SMS_417
Wherein any item has been exchanged a number of times>
Figure SMS_418
Only data within this time period is considered.
According to the user
Figure SMS_419
Co-occurrence vector with user->
Figure SMS_420
The total number of interactions by the user can be calculated>
Figure SMS_421
. Since the degree of main flow is a dynamically changing concept, the number of user interactions can be filtered over time, for example only taking into account the occurrence &>
Figure SMS_422
And (3) the total interaction times of the user in the time period are as follows: />
Figure SMS_423
In which optionally>
Figure SMS_424
Considering only the interaction data during this time period, the release date is @>
Figure SMS_425
The subsequent item interaction data are all set to 0.
According to the idea of collaborative filtering, the degree of user's mainstream depends on whether their interactive items are interacted with by other users. Meanwhile, the method considers that the preference degrees of users to different categories of articles are inconsistent, so the influence of the article category factor is also considered in the calculation process of the mainstream degree score of the user. According to the total number of interactions of the article
Figure SMS_426
Total number of user interactions->
Figure SMS_430
Item category, count user->
Figure SMS_432
Is greater than or equal to the dynamic main flow degree score of>
Figure SMS_427
. In case the number of interactions is not sensitive, it is ≥ for the class>
Figure SMS_429
User->
Figure SMS_433
Is greater than or equal to the dynamic main flow degree score of>
Figure SMS_435
Calculating according to the formula (1); in case of sensitive interaction timesDown, for a category->
Figure SMS_428
User->
Figure SMS_431
Dynamic mainstream degree score of
Figure SMS_434
Calculating according to the formula (2);
Figure SMS_436
（1）
Figure SMS_437
（2）
wherein the content of the first and second substances,
Figure SMS_440
will->
Figure SMS_442
In a time period, the user->
Figure SMS_445
All items that have interacted add up to the number of interactions of all users. Since the power law distribution phenomenon exists in the interaction records of real-world items (namely a small part of items occupy most of interactions, and a large part of items have no interactions), the total interaction times of each item and other users are combined by using a logarithmic function>
Figure SMS_438
Performing inhibition, and counting at the bottom>
Figure SMS_441
Is a hyper-parameter. In case the number of interactions is not sensitive, the user's dynamic mainstream score is ≥ based on ≥ the number of interactions>
Figure SMS_444
Each interaction record is given the same weight. In thatWith sensitive number of interactions, a user's dynamic mainstream score in &>
Figure SMS_447
Assigning a different weight, based on the number of interactive records, to each of the interactive records>
Figure SMS_439
Indicates that the user is pickand place>
Figure SMS_443
And article>
Figure SMS_446
This means that the greater the number of user interactions, the greater the weight in the evaluation of the degree of item occupancy in the score of the degree of mainstream of the user.
According to the user
Figure SMS_448
Is greater than or equal to the dynamic main flow degree score of>
Figure SMS_449
For a category is->
Figure SMS_450
Is determined, the global dynamic mainstream degree score->
Figure SMS_451
Is the average of all users' dynamic mainstream level scores, so the global dynamic mainstream level score ≧>
Figure SMS_452
The calculation formula of (2) is as follows:
Figure SMS_453
then, the global dynamic mainstream degree scores of all the article categories are divided
Figure SMS_454
Constitute a dimension of->
Figure SMS_455
Is determined by the global dynamic prevailing degree vector->
Figure SMS_456
Expressed as:
Figure SMS_457
wherein the content of the first and second substances,
Figure SMS_458
、/>
Figure SMS_461
all indicate a time>
Figure SMS_464
Represents a hyper-parameter (for controlling a logarithmic curve), -is>
Figure SMS_459
Indicates that the article is present>
Figure SMS_463
Belongs to the category->
Figure SMS_465
，/>
Figure SMS_466
Represents the total number of users, and>
Figure SMS_460
representing a total set of users, <' > based on>
Figure SMS_462
Representing the total number of item categories.
And then weighting the co-occurrence vectors of the users according to the obtained mainstream degree scores of the users, namely introducing the mainstream information of the users into the co-occurrence vectors. The weighting process is described as: user will be
Figure SMS_467
For belonging to the category of->
Figure SMS_468
Is based on>
Figure SMS_469
In the interaction->
Figure SMS_470
Multiplied by user pick>
Figure SMS_471
In category +>
Figure SMS_472
Dynamic mainstream level score of &>
Figure SMS_473
；
For user co-occurrence vectors
Figure SMS_474
Each item in the list is weighted, and the co-occurrence vector of the whole user is obtained after the weighting is finished
Figure SMS_475
Normalization using a softmax function resulting in user interaction data &forinput to the collaborative filtering module>
Figure SMS_476
：
Figure SMS_477
Wherein any one of the user interaction data
Figure SMS_478
Has a value range of [0,1]。
User interaction data obtained here
Figure SMS_479
Also as a subsequent collaborative filtering moduleThe input of the device.
And the dynamic mainstream degree characteristic model building module is used for building a dynamic mainstream degree characteristic model.
Can be calculated according to the mainstream score calculation module when
Figure SMS_480
Subscriber arbitrarily for a time period>
Figure SMS_484
Respect to any type of item set >>
Figure SMS_486
Is greater than or equal to the prevailing degree score of>
Figure SMS_481
. For any user->
Figure SMS_483
Calculating its mainstream score for all of the categories of items, which can constitute a ≧ or { (R) }>
Figure SMS_485
Vector of dimensions, denoted as
Figure SMS_487
The vector may characterize the user
Figure SMS_482
The mainstream degree of all the item categories, the larger the score of the mainstream degree of an item category is, the more the user prefers popular items in the category, and the smaller the score is, the more the user prefers popular items in the category.
In order to enable the collaborative filtering model of the collaborative filtering module building module to fully utilize the user information and the global information, a dynamic mainstream feature expression based on the user information and the global information needs to be obtained. A dynamic mainstream level feature model is thus constructed here. Constructing a dynamic mainstream degree characteristic model based on a three-layer perceptron MLP model, extracting key characteristics in user information and global information,and reduces the original input to a hidden vector of lower dimensionality. The first two layers of the dynamic mainstream degree characteristic model use a ReLU function as an activation function, and the last layer of the dynamic mainstream degree characteristic model uses a softmax activation function. With user information vectors
Figure SMS_488
And a global dynamic mainstream degree vector output by the mainstream score calculation module>
Figure SMS_489
Spliced and used as input of a dynamic mainstream degree characteristic model>
Figure SMS_490
。
The user information vector
Figure SMS_491
Expressed as: />
Figure SMS_492
Wherein the content of the first and second substances,
Figure SMS_493
indicates that the user is pickand place>
Figure SMS_494
Is quantified age information and is greater or less than>
Figure SMS_495
Indicates that the user is pickand place>
Figure SMS_496
The binary sex information of (2);
inputting the dynamic mainstream degree characteristic model
Figure SMS_497
Expressed as:
Figure SMS_498
；
wherein, the first and the second end of the pipe are connected with each other,
Figure SMS_499
representing vector stitching operations
Input device
Figure SMS_500
And outputting the dynamic mainstream characteristic hidden vector after dimensionality reduction of the three-layer perceptron of the dynamic mainstream degree characteristic model.
The collaborative filtering module construction module is used for constructing a collaborative filtering module, and adopts an asymmetric variational self-encoder (VAE) as a main structure of a collaborative filtering model, and the structure of the collaborative filtering module is shown in FIG. 3. The method adopts an asymmetric variational self-encoder to carry out collaborative filtering, and aims to add extra dynamic mainstream degree information into a hidden layer between an encoder and a decoder, so that the decoder can directly utilize personal information of a user and global mainstream degree during decoding.
The collaborative filtering module comprises an encoder and a decoder;
the encoder is constructed by adopting a three-layer perceptron MLP model. User interaction data
Figure SMS_502
An input encoder that calculates user interaction data &>
Figure SMS_505
And generates £ respectively>
Figure SMS_507
Mean and->
Figure SMS_501
Variance, forming a mean vector ≥ of users>
Figure SMS_504
And the variance vector pick>
Figure SMS_506
Wherein the two vectors are both in t dimension, and then randomly samplingGenerating a user-interaction hidden vector ≥ of h-dimension>
Figure SMS_508
，
Figure SMS_503
. Since the network cannot perform back propagation due to the random sampling method, the re-parameterization method is adopted to complete the sampling process.
The decoder is constructed by adopting a four-layer perceptron MLP model, the first three layers of activation functions of the decoder are tanh functions, and the last layer of activation functions of the decoder are softmax functions and are used for generating probability distribution
Figure SMS_509
. Dynamic mainstream characteristic hidden vector output by the dynamic mainstream degree characteristic model building module and user interaction hidden vector/device output by the encoder>
Figure SMS_510
The decoder output reconstructs the user interaction data ≥ as input from the decoder>
Figure SMS_511
And reconstructing a dynamic mainstream feature vector>
Figure SMS_512
Reconstructing the dynamic mainstream feature vector->
Figure SMS_513
For completing the reconstruction of the decoder.
The variational autoencoder reasoning process is as follows, assuming the user
Figure SMS_514
Corresponding user interaction hidden vector ≥>
Figure SMS_515
Compliance
Figure SMS_516
Is normally distributed. Interacting data according to a recommendation systemFeature, presuming user interaction data entered into the encoder>
Figure SMS_517
Obey probability is->
Figure SMS_518
The likelihood function of the polynomial distribution of (2) is as follows:
Figure SMS_519
wherein the content of the first and second substances,
Figure SMS_520
indicating that the vector is hidden by a user interaction>
Figure SMS_521
Determined and/or>
Figure SMS_522
Probability of individual item interaction;
in order to enable the network to learn the parameters by back-propagation, a posteriori distributions for each user data sample must be found
Figure SMS_524
Since this posterior distribution is not easy to find, the variation distribution is used here by means of variation deduction
Figure SMS_526
To approximate>
Figure SMS_528
. Hypothesis->
Figure SMS_523
Satisfies a Gaussian distribution->
Figure SMS_531
Wherein->
Figure SMS_532
Is a variance vector pick>
Figure SMS_533
The diagonal covariance matrix of (2). The optimization goal of the network at this point is that the optimization parameter generates a mean vector ≥ then>
Figure SMS_525
And the variance vector pick>
Figure SMS_527
So that the differentiation profile->
Figure SMS_529
And the posterior distribution->
Figure SMS_530
As similar as possible.
Wherein the reparameterization method operates as follows, assuming noise
Figure SMS_534
Obey a normal distribution>
Figure SMS_535
Hidden vector of user interaction
Figure SMS_536
Can be determined by the variance vector>
Figure SMS_537
The mean vector->
Figure SMS_538
And noise are linearly combined so that the network can learn. The reparameterization formula is as follows:
Figure SMS_539
wherein the content of the first and second substances,
Figure SMS_540
is the standard deviation;
unlike standard variational self-encoder networks, this isThe asymmetric structure is used in the method, and the user interaction hidden vector is obtained
Figure SMS_541
Thereafter, the input generated in the dynamic mainstream severity feature model building module is ≥ based on the input ≥ generated in the dynamic mainstream severity feature model building module>
Figure SMS_542
Spliced on/in>
Figure SMS_543
And then fed into the decoder. The generating part of the decoder may be divided into reconstructing user interaction data &>
Figure SMS_544
And reconstructing the dynamic mainstream feature vector->
Figure SMS_545
。
In conclusion, loss function of collaborative filtering model based on asymmetric variational self-encoder
Figure SMS_546
Is divided into a reconstruction target loss->
Figure SMS_547
The distribution is approximately lost>
Figure SMS_548
And a dynamic mainstream feature vector approximation loss>
Figure SMS_549
Three portions, the loss function>
Figure SMS_550
The calculation formula of (2) is as follows:
Figure SMS_551
reconstructing object losses
Figure SMS_552
The aim of this is to make the reconstructed user interaction data output by the decoder as identical as possible to the user interaction data input to the encoder, the calculation formula being:
Figure SMS_553
distribution approximation loss
Figure SMS_554
With the aim of making the variational distribution->
Figure SMS_555
As close as possible to a posterior distribution>
Figure SMS_556
For measuring the approximation degree of two distributions, the calculation formula is:
Figure SMS_557
to convert the original dynamic mainstream feature vector
Figure SMS_558
And reconstructed dynamic mainstream feature vector>
Figure SMS_559
As an objective function, the negative number of which is taken as a loss term, with the aim of reconstructing the dynamic mainstream feature vector ≥ s>
Figure SMS_560
Is directionally correlated with the original dynamic main flow feature vector->
Figure SMS_561
As close as possible, the loss is approximated by a dynamic main flow feature vector>
Figure SMS_562
The method can enable a decoder to complete the reconstruction process by utilizing the dynamic mainstream characteristics as much as possible. Approximate loss of a dynamic mainstream feature vector>
Figure SMS_563
The calculation formula of (2) is as follows:
Figure SMS_564
wherein, the first and the second end of the pipe are connected with each other,
Figure SMS_574
represents a user interaction hidden vector, <' > or>
Figure SMS_566
Represents user interaction data, and->
Figure SMS_568
Represents a posterior distribution of each user data sample, based on the posterior distribution of the user data sample>
Figure SMS_572
Represents a variational distribution, a variational distribution->
Figure SMS_577
And the posterior distribution->
Figure SMS_575
In the approximation that the difference between the first and second values,
Figure SMS_580
represents->
Figure SMS_573
To (X)>
Figure SMS_576
Is desired, is based on>
Figure SMS_565
Is a hyperparameter for controlling the penalty of distribution similarity on the overall objective function, and>
Figure SMS_569
is a user interaction hidden vector->
Figure SMS_578
KL represents KL divergence, and>
Figure SMS_579
represents a prior distribution, <' > or>
Figure SMS_581
Represents a variance vector, < > based on the variance>
Figure SMS_582
Represents the square of the mean vector>
Figure SMS_567
Is hyperparameter, is greater than or equal to>
Figure SMS_570
Represents the original dynamic mainstream feature vector>
Figure SMS_571
Representing the reconstructed dynamic mainstream feature vector.
A recommendation result generation module for inputting the user to be predicted according to the collaborative filtering module obtained by the collaborative filtering module construction module after training
Figure SMS_583
Currently observed user interaction data ≧>
Figure SMS_584
The encoder outputs a mean vector->
Figure SMS_585
Sum variance vector
Figure SMS_586
Then according to the formula->
Figure SMS_587
Calculating to obtain a user interaction hidden vector>
Figure SMS_588
(ii) a Then interact with the userHidden vector->
Figure SMS_589
Inputting the dynamic mainstream characteristic hidden vector output by the dynamic mainstream degree characteristic model building module into a decoder, outputting the reconstructed user interaction data by the decoder, wherein the dimensionality of the user interaction data is n-dimensional and is the same as the quantity of all articles, and the value of each dimensionality is [0,1 ]]In the meantime.
Aiming at the obtained reconstructed user interaction data, firstly, eliminating articles which do not meet the time requirement, and only considering
Figure SMS_590
And a previous time period, then any issue time is @>
Figure SMS_591
And setting the value of the dimension where the subsequent article serial number is positioned as 0. Secondly, items which have already appeared in the history data are eliminated, and the currently observed interaction situation is combined>
Figure SMS_592
And recording the dimension serial numbers with the values not being 0, and setting the dimensions of the reconstructed user interaction data as 0. And finally, sorting the reconstructed user interaction data from large to small, wherein the dimension serial number of top-N is the user to be predicted->
Figure SMS_593
The top N item recommendation lists.
Example 3
The present embodiment provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, causes the processor to execute the steps of the above method for correcting a mainstream deviation of a recommendation system.
The computer device may be a desktop computer, a notebook, a palm computer, a cloud server, or other computing devices. The computer equipment can carry out man-machine interaction with a user in a keyboard mode, a mouse mode, a remote controller mode, a touch panel mode or a voice control equipment mode.
The memory includes at least one type of readable storage medium including a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., SD or D interface display memory, etc.), a Random Access Memory (RAM), a Static Random Access Memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the storage may be an internal storage unit of the computer device, such as a hard disk or a memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) Card, a Flash memory Card (Flash Card), and the like provided on the computer device. Of course, the memory may also include both internal and external storage devices of the computer device. In this embodiment, the memory is usually used for storing an operating system and various types of application software installed on the computer device, for example, program codes of the method for correcting the mainstream deviation of the recommendation system, and the like. In addition, the memory may also be used to temporarily store various types of data that have been output or are to be output.
The processor may be a Central Processing Unit (CPU), controller, microcontroller, microprocessor, or other data Processing chip in some embodiments. The processor is typically used to control the overall operation of the computer device. In this embodiment, the processor is configured to execute the program code stored in the memory or process data, for example, execute the program code of the method for correcting the deviation of the main stream of the recommendation system.
Example 4
The present embodiment provides a computer-readable storage medium, in which a computer program is stored, which, when being executed by a processor, causes the processor to carry out the steps of the above-mentioned method for correcting a deviation of a mainstream of a recommendation system.
Wherein the computer readable storage medium stores an interface display program executable by at least one processor to cause the at least one processor to perform the steps of the method for correcting a recommended system mainstream deviation as described above.
Through the above description of the embodiments, those skilled in the art will clearly understand that the method of the above embodiments can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware, but in many cases, the former is a better implementation manner. Based on such understanding, the technical solution of the present application or portions thereof that contribute to the prior art may be embodied in the form of a software product, where the computer software product is stored in a storage medium (e.g., ROM/RAM, magnetic disk, optical disk), and includes several instructions for enabling a terminal device (e.g., a mobile phone, a computer, a server, or a network device) to execute the method for correcting mainstream deviation of a recommendation system according to the embodiment of the present application.

Claims (7)

1. A method for correcting deviations in a recommendation system mainstream comprising the steps of:
step S1, data collection and processing
Obtaining user information, article information and user article interaction information in a recommendation system, and respectively constructing user co-occurrence vectors
Figure QLYQS_1
The co-occurrence vector of the object is greater or less than>
Figure QLYQS_2
；
Step S2, calculating the mainstream score
At [ t1, t2]Within a time period, according to the co-occurrence vector of the articles
Figure QLYQS_4
Calculate the total number of times of the interaction of the article->
Figure QLYQS_9
(ii) a Based on the user co-occurrence vector->
Figure QLYQS_11
Calculating the total number of times of the user's interaction>
Figure QLYQS_5
(ii) a Based on the total number of times of interaction of the article>
Figure QLYQS_8
Total number of user interactions->
Figure QLYQS_10
Item category, calculating user &>
Figure QLYQS_13
Dynamic mainstream level score of>
Figure QLYQS_3
(ii) a According to user>
Figure QLYQS_7
Dynamic mainstream level score of>
Figure QLYQS_12
Calculating the average of the dynamic mainstream level scores of all users to obtain a global dynamic mainstream level score>
Figure QLYQS_14
And the global dynamic mainstream level scores of all the item classes are combined into a global dynamic mainstream level vector->
Figure QLYQS_6
；
And carrying out co-occurrence vector weighting processing, wherein the weighting processing is described as follows: user will be connected
Figure QLYQS_15
For belonging to the category of->
Figure QLYQS_16
Is based on>
Figure QLYQS_17
In the interaction->
Figure QLYQS_18
Multiplied by user pick>
Figure QLYQS_19
In category +>
Figure QLYQS_20
On a dynamic main flow degree score->
Figure QLYQS_21
；
For user co-occurrence vectors
Figure QLYQS_22
Each term appearing in (a) is weighted and the entire user co-occurrence vector is then evaluated against>
Figure QLYQS_23
Normalization using a softmax function results in user interaction data ≥ being used for input to the collaborative filtering module>
Figure QLYQS_24
：
Figure QLYQS_25
Wherein any of the user interaction data
Figure QLYQS_26
Has a value range of [0,1]；
S3, constructing a dynamic mainstream degree characteristic model
Dynamic mainstream constructed based on three-layer perceptron MLP modelThe degree characteristic model, the ReLU function is used as an activation function in the first two layers of the dynamic mainstream degree characteristic model, and the softmax activation function is used in the last layer of the dynamic mainstream degree characteristic model; with user information vectors
Figure QLYQS_27
And the global dynamic main flow degree vector ≥ output in step S2>
Figure QLYQS_28
Spliced and used as input of a dynamic mainstream degree characteristic model>
Figure QLYQS_29
Outputting a dynamic mainstream characteristic hidden vector by the dynamic mainstream degree characteristic model;
s4, constructing a collaborative filtering module
Constructing a collaborative filtering module comprising an encoder and a decoder;
the encoder is constructed by adopting a three-layer perceptron MLP model; user interaction data
Figure QLYQS_30
An input encoder that calculates user interaction data &>
Figure QLYQS_31
And generates &, respectively>
Figure QLYQS_32
Mean and->
Figure QLYQS_33
A variance constituting a mean vector &' for the user>
Figure QLYQS_34
And the variance vector pick>
Figure QLYQS_35
Where both vectors are in the t dimension, and then by randomizationSampling to generate a user-interaction hidden vector ≥ h-dimension>
Figure QLYQS_36
；
The decoder is constructed by adopting a four-layer perceptron MLP model, the first three layers of activation functions of the decoder are tanh functions, and the last layer of activation functions of the decoder are softmax functions; step S3, outputting the dynamic mainstream characteristic hidden vector and the user interaction hidden vector output by the encoder
Figure QLYQS_37
As input to a decoder, the decoder output reconstructs user interaction data>
Figure QLYQS_38
And reconstructing the dynamic mainstream feature vector->
Figure QLYQS_39
Reconstructing the dynamic mainstream feature vector->
Figure QLYQS_40
For completing the reconstruction of the decoder;
step S5, recommendation result generation
Inputting user interaction data currently observed by the user to be predicted according to the collaborative filtering module obtained by training in the step S4
Figure QLYQS_41
The encoder outputs a user interaction hidden vector ≥>
Figure QLYQS_42
(ii) a Then the user interaction is hidden by the vector->
Figure QLYQS_43
Inputting the dynamic mainstream characteristic hidden vector output in the step S3 into a decoder, and outputting reconstructed user interaction data by the decoder;
in step S1, optional use is constructed according to user article interaction informationHousehold
Figure QLYQS_44
User co-occurrence vector with all items
Figure QLYQS_45
And any item is constructed according to the item interaction information of the user>
Figure QLYQS_46
Co-occurrence vector with all items>
Figure QLYQS_47
；
Wherein the content of the first and second substances,
Figure QLYQS_49
represents a total number of users, <' > based on>
Figure QLYQS_51
Indicates the total number of items, based on the number of items present in the container>
Figure QLYQS_55
Indicates a substance is present>
Figure QLYQS_50
Indicates the fifth->
Figure QLYQS_52
Each item>
Figure QLYQS_56
Representing a user
Figure QLYQS_58
And article>
Figure QLYQS_48
In a predetermined condition, or a predetermined condition, based on the presence of a predetermined condition>
Figure QLYQS_53
Indicates that the article is present>
Figure QLYQS_54
And user>
Figure QLYQS_57
The interaction scenario of (2).
2. A method for correcting deviations in a recommendation system mainstream according to claim 1, wherein: in step S2, according to the article
Figure QLYQS_59
Co-occurrence vector with the item->
Figure QLYQS_60
Calculating the total number of article interactions
Figure QLYQS_61
；
According to the user
Figure QLYQS_62
Co-occurrence vector with user->
Figure QLYQS_63
And the total number of times of interaction of the user is calculated,
Figure QLYQS_64
；
according to the total number of interaction times of the articles
Figure QLYQS_67
Total number of user interactions->
Figure QLYQS_69
Item category, count user->
Figure QLYQS_71
Is greater than or equal to the dynamic main flow degree score of>
Figure QLYQS_66
(ii) a In case the number of interactions is not sensitive, it is ≥ for the class>
Figure QLYQS_68
User->
Figure QLYQS_73
Dynamic mainstream level score of>
Figure QLYQS_74
Calculating according to the formula (1); in the case of sensitive number of interactions, a @ class>
Figure QLYQS_65
User->
Figure QLYQS_70
Is greater than or equal to the dynamic main flow degree score of>
Figure QLYQS_72
Calculating according to the formula (2);
Figure QLYQS_75
（1）
Figure QLYQS_76
（2）
according to the user
Figure QLYQS_77
Dynamic mainstream level score of>
Figure QLYQS_78
Calculating the average value of the dynamic mainstream degree scores of all the users to obtain a global dynamic mainstream degree score->
Figure QLYQS_79
The calculation formula is as follows:
Figure QLYQS_80
then, the global dynamic mainstream degree scores of all the article categories are divided
Figure QLYQS_81
Constitute a dimension of->
Figure QLYQS_82
Is determined by the global dynamic prevailing degree vector->
Figure QLYQS_83
Expressed as:
Figure QLYQS_84
wherein, the first and the second end of the pipe are connected with each other,
Figure QLYQS_87
、/>
Figure QLYQS_88
all indicate a time>
Figure QLYQS_91
Indicates a hyper-parameter->
Figure QLYQS_86
Indicating a goods>
Figure QLYQS_89
Belongs to the category->
Figure QLYQS_92
，/>
Figure QLYQS_93
Represents the total number of users, and>
Figure QLYQS_85
representing a total set of users, <' > based on>
Figure QLYQS_90
Representing the total number of categories of items.
3. A method for correcting deviations in a recommendation system mainstream according to claim 1, wherein: in step S3, the user information vector
Figure QLYQS_94
Expressed as:
Figure QLYQS_95
wherein the content of the first and second substances,
Figure QLYQS_96
represents a user>
Figure QLYQS_97
Is quantified age information and is greater or less than>
Figure QLYQS_98
Indicates that the user is pickand place>
Figure QLYQS_99
The binary gender information of (1); />
Input of dynamic mainstream degree characteristic model
Figure QLYQS_100
Expressed as:
Figure QLYQS_101
；
wherein, the first and the second end of the pipe are connected with each other,
Figure QLYQS_102
representing a vector stitching operation.
4. A method of correcting deviations in a mainstream of a recommender system as set forth in claim 1, wherein: in step S4, the loss function of the collaborative filtering module
Figure QLYQS_103
Grouping into reconstruction target loss>
Figure QLYQS_104
Approximate loss of distribution>
Figure QLYQS_105
And a dynamic mainstream feature vector approximation loss>
Figure QLYQS_106
Three sections, loss function>
Figure QLYQS_107
The calculation formula of (2) is as follows:
Figure QLYQS_108
reconstructing object losses
Figure QLYQS_109
The calculation formula of (2) is as follows:
Figure QLYQS_110
distribution approximation loss
Figure QLYQS_111
The calculation formula of (2) is as follows:
Figure QLYQS_112
dynamic mainstream eigenvector approximation loss
Figure QLYQS_113
The calculation formula of (2) is as follows:
Figure QLYQS_114
wherein the content of the first and second substances,
Figure QLYQS_122
represents a user-interactive hidden vector, < '> based on a user's interaction>
Figure QLYQS_118
Represents user interaction data, and->
Figure QLYQS_123
Represents a posterior distribution of each user data sample, <' > based on the data sample>
Figure QLYQS_117
Represents a variation profile, a variation profile>
Figure QLYQS_124
And the posterior distribution->
Figure QLYQS_120
Is approximately, is greater than>
Figure QLYQS_127
Represents->
Figure QLYQS_129
To (X)>
Figure QLYQS_132
In a predetermined direction, in a predetermined direction>
Figure QLYQS_115
In the case of hyper-parameters>
Figure QLYQS_131
Is a user interaction hidden vector->
Figure QLYQS_116
KL represents KL divergence, ->
Figure QLYQS_125
Represents an a priori profile, is>
Figure QLYQS_119
Represents a variance vector, < > based on the variance>
Figure QLYQS_126
Represents the square of the mean vector, <' > is selected>
Figure QLYQS_121
Is hyperparameter, is greater than or equal to>
Figure QLYQS_128
Represents the original dynamic mainstream feature vector, <' > or>
Figure QLYQS_130
Representing the reconstructed dynamic mainstream feature vector.
5. A system for correcting deviations in a recommendation system mainstream comprising:
the data collection and processing module is used for acquiring the user information, the article information and the user article interaction information in the recommendation system and respectively constructing the user co-occurrence vectors
Figure QLYQS_133
The co-occurrence vector of the object is greater or less than>
Figure QLYQS_134
；
A mainstream score calculation module for calculating the score at [ t1, t2 ]]Within a time period, according to the co-occurrence vector of the articles
Figure QLYQS_138
Calculating the total number of times of interaction of the object>
Figure QLYQS_139
(ii) a Based on the user co-occurrence vector->
Figure QLYQS_142
Calculating the total number of times of the user's interaction>
Figure QLYQS_137
(ii) a Based on the total number of times of interaction of the article>
Figure QLYQS_140
Total number of user interactions->
Figure QLYQS_145
Item category, count user->
Figure QLYQS_146
Dynamic mainstream degree score of
Figure QLYQS_135
(ii) a According to user>
Figure QLYQS_141
Is greater than or equal to the dynamic main flow degree score of>
Figure QLYQS_143
Calculating the average value of the dynamic mainstream degree scores of all the users to obtain a global dynamic mainstream degree score->
Figure QLYQS_144
And the global dynamic mainstream level scores of all the item classes are combined into a global dynamic mainstream level vector->
Figure QLYQS_136
；
And carrying out co-occurrence vector weighting processing, wherein the weighting processing is described as follows: user will be
Figure QLYQS_147
For belonging to the category of->
Figure QLYQS_148
Is based on>
Figure QLYQS_149
In a mobile communication system situation->
Figure QLYQS_150
Multiply by user>
Figure QLYQS_151
Is in the category->
Figure QLYQS_152
On a dynamic main flow degree score->
Figure QLYQS_153
；
For user co-occurrence vectors
Figure QLYQS_154
Each term appearing in (a) is weighted and the entire user co-occurrence vector is then evaluated against>
Figure QLYQS_155
Normalization using a softmax function results in user interaction data ≥ being used for input to the collaborative filtering module>
Figure QLYQS_156
：
Figure QLYQS_157
Wherein any one of the user interaction data
Figure QLYQS_158
Has a value range of [0,1]；
The dynamic mainstream degree characteristic model building module is used for building a dynamic mainstream degree characteristic model based on a three-layer perceptron MLP model, the ReLU function is used as an activation function in the first two layers of the dynamic mainstream degree characteristic model, and the softmax activation function is used in the last layer of the dynamic mainstream degree characteristic model; with user information vectors
Figure QLYQS_159
And a global dynamic mainstream degree vector output by the mainstream score calculation module>
Figure QLYQS_160
Spliced and used as input of a dynamic mainstream degree characteristic model>
Figure QLYQS_161
Outputting a dynamic mainstream characteristic hidden vector by the dynamic mainstream degree characteristic model;
the collaborative filtering module construction module is used for constructing a collaborative filtering module comprising an encoder and a decoder;
the encoder is constructed by adopting a three-layer perceptron MLP model; user interaction data
Figure QLYQS_162
An input encoder that calculates user interaction data &>
Figure QLYQS_163
And generates &, respectively>
Figure QLYQS_164
Mean and>
Figure QLYQS_165
variance, forming a mean vector ≥ of users>
Figure QLYQS_166
And variance vector +>
Figure QLYQS_167
Wherein both vectors are t-dimensional, and then generating a h-dimensional user interaction hidden vector ≥ by random sampling>
Figure QLYQS_168
；
The decoder is constructed by adopting a four-layer perceptron MLP model, the first three layers of activation functions of the decoder are tanh functions, and the last layer of activation functions of the decoder are softmax functions; dynamic mainstream characteristic hidden vectors output by the dynamic mainstream degree characteristic model building module and user interaction hidden vectors output by the encoder
Figure QLYQS_169
The decoder output reconstructs the user interaction data ≥ as input from the decoder>
Figure QLYQS_170
And reconstructing the dynamic mainstream feature vector->
Figure QLYQS_171
Reconstructing the dynamic mainstream feature vector->
Figure QLYQS_172
For completing the reconstruction of the decoder;
a recommendation result generation module for inputting the user interaction data currently observed by the user to be predicted according to the collaborative filtering module obtained by the collaborative filtering module construction module after training
Figure QLYQS_173
The encoder outputs a user interaction hidden vector ≥>
Figure QLYQS_174
(ii) a Then the user interaction is hidden by the vector->
Figure QLYQS_175
Inputting the dynamic mainstream characteristic hidden vector output by the dynamic mainstream degree characteristic model building module into a decoder, and outputting reconstructed user interaction data by the decoder;
in the data collecting and processing module, any user is constructed according to the user article interaction information
Figure QLYQS_176
User co-occurrence vector with all items pick>
Figure QLYQS_177
And any item is constructed according to the item interaction information of the user>
Figure QLYQS_178
Co-occurrence vector with all items>
Figure QLYQS_179
；
Wherein the content of the first and second substances,
Figure QLYQS_181
represents the total number of users, and>
Figure QLYQS_183
indicates the total number of items, based on the number of items present in the container>
Figure QLYQS_188
Indicates a item +>
Figure QLYQS_182
Indicates the fifth->
Figure QLYQS_185
Each item>
Figure QLYQS_186
Representing a user
Figure QLYQS_190
And article>
Figure QLYQS_180
In a predetermined condition, or a predetermined condition, based on the presence of a predetermined condition>
Figure QLYQS_184
Indicates that the article is present>
Figure QLYQS_187
And the user->
Figure QLYQS_189
The interaction scenario of (2).
6. A computer device, characterized by: comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to carry out the steps of the method according to any one of claims 1 to 4.
7. A computer-readable storage medium characterized by: a computer program is stored which, when being executed by a processor, causes the processor to carry out the steps of the method according to any one of claims 1 to 4.
CN202310104256.6A 2023-02-13 2023-02-13 Method, system, device and storage medium for correcting mainstream deviation of recommendation system Active CN115809374B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310104256.6A CN115809374B (en) 2023-02-13 2023-02-13 Method, system, device and storage medium for correcting mainstream deviation of recommendation system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310104256.6A CN115809374B (en) 2023-02-13 2023-02-13 Method, system, device and storage medium for correcting mainstream deviation of recommendation system

Publications (2)

Publication Number Publication Date
CN115809374A CN115809374A (en) 2023-03-17
CN115809374B true CN115809374B (en) 2023-04-18

Family

ID=85487843

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310104256.6A Active CN115809374B (en) 2023-02-13 2023-02-13 Method, system, device and storage medium for correcting mainstream deviation of recommendation system

Country Status (1)

Country Link
CN (1) CN115809374B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116186421B (en) * 2023-05-04 2023-07-14 中国科学技术大学 Recommendation method, system, equipment and storage medium for eliminating popularity deviation

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113987358A (en) * 2021-11-15 2022-01-28 中国科学技术大学 Training method, recommendation method and recommendation system of recommendation model
CN114861783A (en) * 2022-04-26 2022-08-05 北京三快在线科技有限公司 Recommendation model training method and device, electronic equipment and storage medium
CN114912033A (en) * 2022-05-16 2022-08-16 重庆大学 Knowledge graph-based recommendation popularity deviation adaptive buffering method
CN115129945A (en) * 2022-06-23 2022-09-30 阿里巴巴新加坡控股有限公司 Graph structure contrast learning method, equipment and computer storage medium

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20080097821A1 (en) * 2006-10-24 2008-04-24 Microsoft Corporation Recommendations utilizing meta-data based pair-wise lift predictions
US10289733B2 (en) * 2014-12-22 2019-05-14 Rovi Guides, Inc. Systems and methods for filtering techniques using metadata and usage data analysis
CN108647226B (en) * 2018-03-26 2021-11-02 浙江大学 Hybrid recommendation method based on variational automatic encoder
WO2021038592A2 (en) * 2019-08-30 2021-03-04 Tata Consultancy Services Limited System and method for handling popularity bias in item recommendations
CN112184391B (en) * 2020-10-16 2023-10-10 中国科学院计算技术研究所 Training method of recommendation model, medium, electronic equipment and recommendation model
CN113158024B (en) * 2021-02-26 2022-07-15 中国科学技术大学 Causal reasoning method for correcting popularity deviation of recommendation system
CN114428910A (en) * 2022-01-28 2022-05-03 腾讯科技（深圳）有限公司 Resource recommendation method and device, electronic equipment, product and medium
CN115147192A (en) * 2022-07-29 2022-10-04 华东师范大学 Recommendation method and recommendation system based on double-view-angle deviation correction
CN115438871A (en) * 2022-09-23 2022-12-06 哈尔滨商业大学 Ice and snow scenic spot recommendation method and system integrating preference and eliminating popularity deviation

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113987358A (en) * 2021-11-15 2022-01-28 中国科学技术大学 Training method, recommendation method and recommendation system of recommendation model
CN114861783A (en) * 2022-04-26 2022-08-05 北京三快在线科技有限公司 Recommendation model training method and device, electronic equipment and storage medium
CN114912033A (en) * 2022-05-16 2022-08-16 重庆大学 Knowledge graph-based recommendation popularity deviation adaptive buffering method
CN115129945A (en) * 2022-06-23 2022-09-30 阿里巴巴新加坡控股有限公司 Graph structure contrast learning method, equipment and computer storage medium

Also Published As

Publication number Publication date
CN115809374A (en) 2023-03-17

Similar Documents

Publication Publication Date Title
Catherine et al. Transnets: Learning to transform for recommendation
CN111127142B (en) Article recommendation method based on generalized nerve attention
EP3862893A1 (en) Recommendation model training method, recommendation method, device, and computer-readable medium
CN111461841B (en) Article recommendation method, device, server and storage medium
CN109102127B (en) Commodity recommendation method and device
CN110969516A (en) Commodity recommendation method and device
EP4181026A1 (en) Recommendation model training method and apparatus, recommendation method and apparatus, and computer-readable medium
CN110084670B (en) Shelf commodity combination recommendation method based on LDA-MLP
CN109903103B (en) Method and device for recommending articles
CN111400603A (en) Information pushing method, device and equipment and computer readable storage medium
CN113761359B (en) Data packet recommendation method, device, electronic equipment and storage medium
CN113158024B (en) Causal reasoning method for correcting popularity deviation of recommendation system
CN114202061A (en) Article recommendation method, electronic device and medium based on generation of confrontation network model and deep reinforcement learning
CN110955826A (en) Recommendation system based on improved recurrent neural network unit
CN115809374B (en) Method, system, device and storage medium for correcting mainstream deviation of recommendation system
CN110598120A (en) Behavior data based financing recommendation method, device and equipment
CN112396492A (en) Conversation recommendation method based on graph attention network and bidirectional long-short term memory network
US20240037133A1 (en) Method and apparatus for recommending cold start object, computer device, and storage medium
CN116452293A (en) Deep learning recommendation method and system integrating audience characteristics of articles
CN115544379A (en) Quaternion map convolutional neural network-based recommendation method and device
CN115456656A (en) Method and device for predicting purchase intention of consumer, electronic equipment and storage medium
Chae et al. Adversarial training of deep autoencoders towards recommendation tasks
CN112801743B (en) Commodity recommendation method and device, electronic equipment and storage medium
Peng et al. Credit scoring model in imbalanced data based on cnn-atcn
CN113763111A (en) Article collocation method, device and storage medium

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