CN117252665B - Service recommendation method and device, electronic equipment and storage medium - Google Patents
Service recommendation method and device, electronic equipment and storage medium Download PDFInfo
- Publication number
- CN117252665B CN117252665B CN202311508557.1A CN202311508557A CN117252665B CN 117252665 B CN117252665 B CN 117252665B CN 202311508557 A CN202311508557 A CN 202311508557A CN 117252665 B CN117252665 B CN 117252665B
- Authority
- CN
- China
- Prior art keywords
- sample
- dimension
- target
- behavior
- feature dimension
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 82
- 230000006399 behavior Effects 0.000 claims description 638
- 238000003062 neural network model Methods 0.000 claims description 110
- 239000013598 vector Substances 0.000 claims description 96
- 230000006870 function Effects 0.000 claims description 68
- 238000012512 characterization method Methods 0.000 claims description 67
- 238000012549 training Methods 0.000 claims description 67
- 238000012545 processing Methods 0.000 claims description 47
- 238000007781 pre-processing Methods 0.000 claims description 12
- 235000019580 granularity Nutrition 0.000 claims description 11
- 238000004590 computer program Methods 0.000 claims description 10
- 230000004913 activation Effects 0.000 claims description 9
- 230000003997 social interaction Effects 0.000 claims description 8
- 238000006243 chemical reaction Methods 0.000 claims description 7
- 238000001514 detection method Methods 0.000 claims description 6
- 238000010801 machine learning Methods 0.000 description 11
- 238000010586 diagram Methods 0.000 description 6
- 238000010606 normalization Methods 0.000 description 6
- 238000004891 communication Methods 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 5
- 238000010187 selection method Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 4
- 238000011156 evaluation Methods 0.000 description 4
- 238000012886 linear function Methods 0.000 description 4
- 239000011159 matrix material Substances 0.000 description 4
- 230000009466 transformation Effects 0.000 description 4
- 238000000844 transformation Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 230000003542 behavioural effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000012356 Product development Methods 0.000 description 1
- 238000000546 chi-square test Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000007637 random forest analysis Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Finance (AREA)
- Accounting & Taxation (AREA)
- Strategic Management (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Business, Economics & Management (AREA)
- Evolutionary Biology (AREA)
- Marketing (AREA)
- Economics (AREA)
- Medical Informatics (AREA)
- Development Economics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a service recommendation method, a device, electronic equipment and a storage medium, and relates to the technical field of computers, wherein the method comprises the following steps: inputting identification information of each first sample feature dimension and the target behavior category into a target relation prediction model to obtain a relation prediction value between each first sample feature dimension and the target behavior category; determining a target feature dimension corresponding to the target behavior category based on each relation predicted value; and recommending the service to the target user based on the target feature dimension and the target behavior category. The business recommendation method, the device, the electronic equipment and the storage medium can acquire the association relation between the feature dimension of the different sample and the behavior category of the different sample more accurately and more efficiently, further acquire the association relation between the feature dimension of the different sample and the target behavior category more accurately and more efficiently, and realize more accurate and more efficient business recommendation based on the feature dimension of the target.
Description
Technical Field
The present invention relates to the field of computer technologies, and in particular, to a service recommendation method, a device, an electronic device, and a storage medium.
Background
For a shopping platform, particularly an e-commerce shopping platform, an operator can acquire historical behavior characteristics of a consumer based on historical behaviors from information searching to product purchasing and even after purchasing, and further service recommendation can be provided for the consumer based on the user characteristics and the historical behavior characteristics of the consumer.
It will be appreciated that shopping platforms may generally provide consumers with a variety of different categories, different brands, and different prices of business.
However, in the related art, only the association relationship between the user feature and the historical behavior feature of the consumer and a certain behavior category can be obtained, so that it is difficult to obtain the association relationship between the user feature and the historical behavior feature of the consumer and a plurality of behavior categories, and also difficult to obtain the association relationship between the user feature and the historical behavior feature of the consumer and a newly-built behavior category, which results in low accuracy of service recommendation in the related art.
Disclosure of Invention
The invention provides a service recommendation method, a device, electronic equipment and a storage medium, which are used for solving the problems that in the prior art, only the association relation between the user characteristics and the historical behavior characteristics of a consumer and a certain behavior category can be obtained, the association relation between the user characteristics and the historical behavior characteristics of the consumer and a plurality of behavior categories is difficult to obtain, and the association relation between the user characteristics and the historical behavior characteristics of the consumer and a newly-built behavior category is difficult to obtain, so that the accuracy of service recommendation is low in the related art, and the user characteristics and the historical behavior characteristics of the consumer and the association relation between the user characteristics and the historical behavior characteristics of the consumer and the newly-built behavior category are more accurately obtained.
The invention provides a service recommendation method, wherein the service comprises the following steps: any one of goods, services, and content; the method comprises the following steps:
inputting identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension into a target relation prediction model to obtain a relation prediction value between each first sample feature dimension and the target behavior category output by the target relation prediction model, wherein the target relation prediction model is obtained after training based on a first relation tag value and a second relation tag value, the first relation tag value comprises a relation tag value between each sample feature dimension and each sample behavior category, and the second relation tag value comprises a relation tag value between each second sample feature dimension in each sample feature dimension and the target behavior category;
determining a target feature dimension corresponding to the target behavior category in each first sample feature dimension based on a relation prediction value between each first sample feature dimension and the target behavior category;
based on the target feature dimension and the target behavior category, performing service recommendation on a target user;
Wherein the sample feature dimension comprises a user feature dimension and/or a historical behavior dimension; the user feature dimension comprises a feature dimension for describing individual users; the historical behavior dimension comprises a feature dimension for describing a user's historical behavior; behavior categories are used to describe behaviors of different dimensions and different granularities; the target behavior category is a different behavior category than the sample behavior category.
The invention also provides a service recommending device, wherein the service comprises: any one of goods, services, and content; the device comprises:
the data input module is used for inputting identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension into a target relation prediction model, obtaining a relation prediction value between each first sample feature dimension and the target behavior category output by the target relation prediction model, wherein the target relation prediction model is obtained after training based on a first relation tag value and a second relation tag value, the first relation tag value comprises a relation tag value between each sample feature dimension and each sample behavior category, and the second relation tag value comprises a relation tag value between each second sample feature dimension in each sample feature dimension and the target behavior category;
The feature selection module is used for determining a target feature dimension corresponding to the target behavior category in the first sample feature dimensions based on a relation prediction value between each first sample feature dimension and the target behavior category;
the service recommendation module is used for recommending the service to the target user based on the target feature dimension and the target behavior category;
wherein the sample feature dimension comprises a user feature dimension and/or a historical behavior dimension; the user feature dimension comprises a feature dimension for describing individual users; the historical behavior dimension comprises a feature dimension for describing a user's historical behavior; behavior categories are used to describe behaviors of different dimensions and different granularities; the target behavior category is a different behavior category than the sample behavior category.
The invention also provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the service recommendation method as described in any of the above when executing the program.
The present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the service recommendation method as described in any of the above.
The invention also provides a computer program product comprising a computer program which when executed by a processor implements a service recommendation method as described in any one of the above.
The invention provides a service recommendation method, a device, an electronic device and a storage medium, which are used for sequentially inputting identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension as a group of information into a target relationship prediction model to obtain a relationship prediction value between each first sample feature dimension and the target behavior category which are sequentially output by the target relationship prediction model, determining a target feature dimension corresponding to the target behavior category in each first sample feature dimension based on the relationship prediction value between each first sample feature dimension and the target behavior category, further performing service recommendation on a target user based on the target feature dimension and the target behavior category, wherein the target relationship prediction model is obtained after training based on a first relationship tag value and a second relationship tag value, the first relation tag value comprises a relation tag value between each sample feature dimension and each sample behavior category, the second relation tag value comprises a relation tag value between each second sample feature dimension in each sample feature dimension and the target behavior category, the association relation between different sample feature dimensions and different sample behavior categories can be more accurately and more efficiently obtained, the association relation between different sample feature dimensions and the target behavior categories can be more accurately and more efficiently obtained, the target feature dimension with the strongest association with the target behavior category can be more accurately and more efficiently determined based on the association relation between the different sample feature dimensions and the target behavior category, more accurate and more efficient service recommendation can be realized based on the target feature dimensions, user requirements can be better met, and user perception can be improved.
Drawings
In order to more clearly illustrate the invention or the technical solutions of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.
Fig. 1 is a schematic flow chart of a service recommendation method provided by the invention;
FIG. 2 is a schematic structural diagram of a target relationship prediction model in the service recommendation method provided by the invention;
FIG. 3 is a flowchart of a feature selection step in the service recommendation method provided by the present invention;
fig. 4 is a schematic structural diagram of a service recommendation device provided by the present invention;
fig. 5 is a schematic structural diagram of an electronic device provided by the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
It should be noted that, feature selection (Feature Selection) is an important link of machine learning and feature engineering, and the goal is to select some features that are most important to the model, i.e. find the optimal feature subset, and remove irrelevant and redundant features, so as to reduce complexity of the model, reduce training time, and improve algorithm performance and generalization capability.
Specifically, in the related technology, the user characteristic and the historical behavior characteristic which have the highest correlation with the consumption behavior of the target category can be determined through characteristic selection, and then the machine learning model can be trained based on the consumption behavior of the target category and the user characteristic and the historical behavior characteristic which have the highest correlation with the consumption behavior of the target category, so that the trained machine learning model is obtained.
It should be noted that, in the description of the present invention, behavior categories may be used to describe behaviors of different dimensions and different granularities, for example: the behavior category 1 can be used for describing the consumption behavior of purchasing different types of commodities (such as apparel, mother and infant, electronic digital) and the like; behavior class 2 may be used to describe consumption behavior to purchase different values; behavior category 3 may be used to describe the consumption behavior of purchasing different brands of goods; behavior category 4 may be used to purchase a consumer behavior describing purchasing a commodity of a certain consumer type under a certain consumer brand.
Also for example: the behavior category 5 may be used to describe viewing behavior for viewing different types of video (e.g., animation, movie, television show, short video, etc.); the behavior category 6 may be used to describe viewing behavior for viewing movies of different actors; the behavior categories may be used to describe viewing behavior for viewing videos of different languages.
In the related art, there are many classical feature selection methods, for example, feature selection algorithms can be classified from three aspects of evaluation criteria, search strategies and supervision information, and feature selection is performed by evaluating the relationship between features and sample categories, or data dimension reduction is performed by evaluating the relative importance of the features, so as to achieve the purpose of feature selection; for another example, feature choices may be categorized and compared from both search strategies and evaluation criteria, respectively; for another example, feature subset space may be explored based on a genetic algorithm of heuristic search; for another example, deep learning may be used to select multidimensional features of an image, and the effect of reducing the query time is obtained in image retrieval; for another example, conventional feature extraction may be modeled based on a machine learning model of random forests.
However, most of the above classical feature selection methods are non-deep learning methods, and the above classical feature selection methods perform feature selection only on a single sample category, ignore the internal relations among different sample categories, between different sample categories and features, and between different features, and result in insufficient modeling of the internal relations between existing sample categories and features, and it is difficult to use the existing sample categories and features to assist in feature selection of new sample categories.
In order to solve the problem of the lack of an inherent quantitative relation between a plurality of sample categories and a plurality of features and fully utilize the existing sample categories and the existing feature data to help the new sample category to carry out feature selection, the invention provides a service recommendation method, wherein the method is based on deep learning and a neural network. According to the feature selection method, strong learning capacity of the neural network is utilized, internal relations between sample categories and features, between features and between the sample categories and the features are explored, self attributes of the sample categories and the features are comprehensively modeled, influences of the sample categories and the features on the internal relations of the sample categories and the features in dimensions of the hidden space are comprehensively modeled, and meanwhile a linear relation network and a nonlinear relation network are used for modeling the linear relation and the nonlinear relation of the sample categories and the features respectively. The model can accurately and efficiently estimate the quantitative relation between the new sample category and the new characteristics by using the deep neural network, and can effectively select the characteristics of the new sample according to the quantitative relation. Meanwhile, in order to increase the expression capability of the features, a series of enhancement operations are performed on the features to generate new features, the new features better express the problem essence, the expression capability of the original features of 0 is effectively expanded, and then the task model effect is improved. Finally, from the original features and the enhanced feature sets thereof, valid features are selected for the new sample class. The method fully models the whole relation of the data, and the feature selection has higher accuracy, high efficiency and generalization.
Fig. 1 is a schematic flow chart of a service recommendation method provided by the invention. The service recommendation method of the present invention is described below with reference to fig. 1. As shown in fig. 1, the method includes: step 101, inputting identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension into a target relationship prediction model, and obtaining a relationship prediction value between each first sample feature dimension and the target behavior category output by the target relationship prediction model, wherein the target relationship prediction model is obtained after training based on a first relationship tag value and a second relationship tag value, the first relationship tag value comprises a relationship tag value between each sample feature dimension and each sample behavior category, and the second relationship tag value comprises a relationship tag value between each second sample feature dimension in each sample feature dimension and the target behavior category.
Wherein the sample feature dimension comprises a user feature dimension and/or a historical behavior dimension; the user feature dimension includes a feature dimension for describing an individual user; the historical behavior dimension includes a feature dimension for describing a user's historical behavior; behavior categories are used to describe behaviors of different dimensions and different granularities; the target behavior category is a different behavior category than the sample behavior category.
It should be noted that, the execution body in the embodiment of the present invention is a service recommendation device.
The business in the embodiment of the invention can be embodied in various forms such as goods, services and contents.
Specifically, sample feature dimensions in embodiments of the present invention may include user feature dimensions and/or historical behavior dimensions; the user feature dimensions include feature dimensions that may be used to describe the individual user; the user feature dimensions include feature dimensions that may be used to describe the user's historical behavior.
For example: the sample feature dimension 1 may be an age dimension; the sample feature dimension 2 may be a gender dimension; the sample feature dimension 3 may be an academic dimension; the sample feature dimension 4 may be a region dimension in which the user is located; the sample feature dimension 5 may be a dimension of a user's historical browsing behavior; the sample feature dimension 6 may be a user history click behavior dimension; sample feature dimension 7 may be a user history scoring behavior dimension; the sample feature dimension 8 may be a user historical feedback behavior dimension; the sample feature dimension 9 may be a user history social interaction behavior dimension; the sample feature dimension 10 may be a user historical reading behavior dimension.
Optionally, the user feature dimension includes at least one of an age dimension, a gender dimension, an academic dimension, a locale dimension, a occupation dimension, and a hobby dimension; the historical behavior dimension includes at least one of a user historical browsing behavior dimension, a user historical click behavior dimension, a user historical scoring behavior dimension, a user historical feedback behavior dimension, a user historical social interaction behavior dimension, and a user historical reading behavior dimension.
Embodiments of the invention can be usedRepresenting the number of sample feature dimensions; wherein (1)>Is a positive integer greater than 1.
It should be noted that, in the embodiment of the present invention, it may be thatRandom selection of the individual sample feature dimensions>The sample feature dimension will be +.>The remaining sample feature dimensions in the sample feature dimensions are used as first sample feature dimensions; alternatively, a priori knowledge can be based on +.>Designating +.>The sample feature dimension will be +.>The remaining sample feature data in the sample feature data is taken as a first sample feature dimension.
It will be appreciated that the number of second sample feature dimensions isThe number of the first sample feature dimensions isAnd each.
It should be noted that, in the embodiment of the present invention, the number of the first sample feature dimensions is greater than the number of the second sample feature dimensions, that is,。
it is understood that any first sample feature dimension is also a sample feature dimension. Embodiments of the invention can be usedIndicate->The order of the first sample feature dimension in all sample feature dimensions, i.e. +.>The first sample feature dimension is also +.>A sample feature dimension.
It will be appreciated that the target feature dimension is different from any one sample feature dimension.
The behavior categories in the embodiment of the invention can be used for describing behaviors with different dimensions and different granularities. The target behavior category is a different behavior category than the sample behavior category.
Embodiments of the invention can be usedRepresenting the number of sample behavior categories; wherein (1)>Is a positive integer greater than 1.
It can be appreciated that the sample behavior category in the embodiment of the present invention is not the same as the target behavior category. Also, the number of sample behavior categories is a plurality, and the number of target behavior categories may be one or more.
For the followingThe +.f. in the first sample feature dimension>A first sample characteristic dimension (th->Individual sample feature dimension), will be->A first sample characteristic dimension (th->Individual sample feature dimensions) and the identification information of the target feature dimension are input into the target relationship prediction model, the target relationship prediction model may be applied to +.>A first sample characteristic dimension (th->Individual sample feature dimensions) and target behaviorThe association relation between the categories is predicted, and the +.>A first sample characteristic dimension (th- >Individual sample feature dimension) and target behavior class>。
Wherein,sequentially taking 1,2,3 and …, < >>,/>Is a positive integer greater than 1, +.>Representing the total number of feature dimensions for each sample.
Note that, the firstA first sample characteristic dimension (th->Individual sample feature dimension) and target behavior class>Can be used to describe +.>The correlation between the first sample feature dimension and the target behavior class is strong or weak.
Optionally, in the embodiment of the present invention, the firstA first sample characteristic dimension (th->Individual sample feature dimension) and target behavior class>The larger the->The stronger the correlation between the first sample feature dimension and the target behavior class.
It should be noted that, the target relationship prediction model may be constructed based on a machine learning model, and is obtained by training a first relationship tag value and a second relationship tag value.
As an optional embodiment, the method further includes, before inputting the identification information of each first sample feature dimension of the sample feature dimensions and the identification information of the target behavior class into the target relationship prediction model to obtain the relationship prediction value between each first sample feature dimension and the target behavior class output by the target relationship prediction model: sample feature data corresponding to each sample feature dimension and sample behavior class data corresponding to each sample behavior class are obtained, each sample feature data comprises feature data of a sample user under each sample feature dimension, and each sample behavior class data comprises behavior data of the sample user under each sample behavior class.
It should be noted that, any sample feature data in the embodiments of the present invention may represent a feature of a sample user in a sample feature dimension. Accordingly, in the embodiment of the invention, the number of sample feature data isAnd each.
It is understood that the number of sample users in the embodiment of the present invention may be plural.
In the embodiment of the inventionThe sample feature data may include feature data of a sample user in a plurality of feature dimensions; />The individual sample characteristic data can be associated with->The sample feature dimensions are in one-to-one correspondence.
For the followingThe first part of the sample characteristic data>Sample characteristic data, th->The sample characteristic data can be equal to +.>The sample feature dimensions correspond.
It should be noted that, any sample characteristic data in the embodiment of the present invention may be represented by a data table.
Table 1 shows the first embodimentAnd a data table corresponding to the sample characteristic data. As shown in Table 1, the->The field name of the data table corresponding to the sample characteristic data comprises the user identification and the identification information of the sample characteristic dimension +.>The variable names of the data table corresponding to the sample Feature data comprise identification information (Identity Document, ID) of the sample user and Feature values (features) of the sample user in the sample Feature dimension.
TABLE 1 firstData table corresponding to each sample characteristic data
In the embodiment of the invention, the identity of the sample user can be identified by using the identity information of the sample user, and the characteristic value is used for describing the characteristic of the sample user under a certain sample characteristic dimension.
For example, in the case that the user feature dimension is the gender dimension, in the embodiment of the present invention, a female may be represented by a feature value "0", a male may be represented by a feature value "1", and an unknown may be represented by a feature value "2";
for another example, in the case where the history behavior dimension is the browsing behavior dimension for the commodity a, in the embodiment of the present invention, the feature value "0" may be used to indicate that no browsing behavior occurs, and the feature value "1" may be used to indicate that the browsing behavior occurs;
for another example, when the user feature dimension is the age dimension, in the embodiment of the present invention, the feature value "1" may be used to indicate that the user is less than 20 years old, the feature value "2" may be used to indicate that the user is between 20 years old and 30 years old, the feature value "3" may be used to indicate that the user is between 30 years old and 40 years old, the feature value "4" may be used to indicate that the user is between 40 years old and 50 years old, and the feature value "5" may be used to indicate that the user is over 60 years old.
As an alternative embodiment, sample feature data corresponding to each sample feature dimension is acquired: and acquiring the characteristic data of the sample user in each original characteristic dimension as first original sample characteristic data.
Specifically, in the embodiment of the invention, the feature data of the sample user under a plurality of original feature dimensions can be obtained from the database by a data query method and used as the first original sample feature data.
Performing data preprocessing on the first original sample characteristic data, and obtaining second original sample characteristic data according to a data preprocessing result;
wherein the data preprocessing includes at least one of data format detection, outlier processing, duplicate value processing, and missing value processing.
Specifically, after the first original sample feature data is acquired, the second original sample feature data may be obtained by performing data format detection, outlier processing, repeated value processing, and missing value processing on the first original sample feature data.
Performing feature enhancement processing on the second original sample feature data to obtain feature data of a sample user in each sample feature dimension as each sample feature data;
the feature enhancement processing comprises at least one of function conversion processing, feature scaling processing, dimensionless processing, numerical feature sub-bucket and feature cross combination.
It can be understood that the enhancement operation is performed on the feature data in the original feature dimension, so that the feature data in the new feature dimension can be generated, the feature data in the new feature dimension can better express the feature, the expression capability of the feature data in the original feature dimension can be effectively increased, and further, the model effect obtained through training is improved.
Therefore, after the second original sample feature data is obtained, the second original sample feature data may be subjected to feature enhancement processing, and feature data in a new feature dimension may be generated by performing feature enhancement processing on the second original sample feature data, so that the feature data in the new feature dimension and the second original sample feature data may be determined to be sample feature data, where the sample feature data may include feature data in the original feature dimension and feature data in the new feature dimension.
Specifically, after the second original sample feature data is obtained, feature enhancement processing can be performed on the second original sample feature data through at least one of function conversion, feature scaling, dimensionless processing, numerical feature sub-bucket and feature cross combination, and then the sample feature data can be obtained according to the feature enhancement processing result.
Alternatively, a function ofThe conversion process may include maximum minimum normalization (Min-Max Normalization). Wherein, the maximum value and minimum value normalization is to scale the numerical characteristic data to 0-1, and the normalized new characteristic valueThe method can be calculated by the following formula:
Wherein,and->Respectively representing the maximum value and the minimum value of the characteristic; />Representing the original value of the feature.
Optionally, the feature scaling process may include feature normalization (Feature Standardization). Wherein, the feature normalization is to convert the original feature value into a new feature value after normalization under the same dimensionThe method can be calculated by the following formula:
wherein,is the average value of the characteristic data; />Standard deviation representing characteristic data; />Representing the original value of the feature.
Numerical feature sub-buckets (Numerical Feature Binning) are one way to discretize successive numerical features into different buckets or bins. It may divide a continuous range of values into several discrete intervals, thereby converting continuous data into discrete data. For example, the age dimension is a numerical user feature dimension, and the ages can be equally divided into 10 sections from 0 to 100, and the user ages are divided into corresponding sections.
Feature Cross-combining (Feature Cross) is a technique used in machine learning to combine multiple features. Through feature cross-combining, new features can be created to better represent the relationships and interactions between the original features. For example, the age bins and gender characteristics may be crossed to form male characteristics of 20-30 years old.
As an optional embodiment, obtaining sample behavior class data corresponding to each sample behavior class includes: and acquiring behavior data of the sample user under each original sample behavior category as first original sample behavior category data.
Specifically, in the embodiment of the invention, the behavior data of the sample user under a plurality of original sample behavior categories can be obtained from the database by a data query method and used as the first original sample behavior category data.
Performing data preprocessing on the first original sample behavior category data, and obtaining second original sample behavior category data according to a data preprocessing result;
wherein the preprocessing includes at least one of data format detection, outlier processing, duplicate value processing, and missing value processing.
Specifically, after the first original sample behavior class data is acquired, the second original sample behavior class data may be obtained by performing data format detection, outlier processing, repeated value processing, and missing value processing on the first original sample behavior class data.
Performing feature enhancement processing on the second original sample behavior category data to obtain feature data of a sample user under each sample behavior category as each sample behavior category data;
The feature enhancement processing comprises at least one of function conversion processing, feature scaling processing, dimensionless processing, numerical feature sub-bucket and feature cross combination.
It should be noted that, in the embodiment of the present invention, specific steps of performing feature enhancement processing on the second original sample behavior category data may be referred to the content of each embodiment, which is not described in detail in the embodiment of the present invention.
According to the embodiment of the invention, the sample characteristic data and the sample behavior category data with more expressive capacity can be obtained by carrying out data processing and characteristic enhancement on the characteristic data of the sample user in each original characteristic dimension and the behavior data of the sample user in each original sample category, so that the accuracy rate of service recommendation can be further improved.
And determining part of sample characteristic data in each sample characteristic data as first sample characteristic data, and determining the rest sample characteristic data as second sample characteristic data.
Specifically, obtainAfter the individual sample characteristic data, one can be +.>Random selection of +.>The sample characteristic data will be +.>The remaining sample characteristic data in the sample characteristic data are used as first sample characteristic data; alternatively, a priori knowledge can be based on +. >Designating +.>Number of sample featuresAccording to the second sample characteristic data, will +.>The remaining sample feature data in the sample feature data is taken as first sample feature data.
It will be appreciated that the second sample feature data is in the amount ofThe number of the first sample characteristic data isAnd each.
It will be appreciated that in embodiments of the present invention the number of second sample feature dimensions is greater than the number of first sample feature dimensions.
And determining the sample characteristic dimension corresponding to each first sample characteristic data as each first sample characteristic dimension, and determining the sample characteristic dimension corresponding to each second sample characteristic data as each second sample characteristic dimension.
As an alternative embodiment, the first relationship tag value is obtained based on the steps of:
based on each sample feature data and each sample behavior class data, a relationship tag value between each sample feature dimension and each sample behavior class is obtained as a first relationship tag value.
As an alternative embodiment, the second relationship tag value is obtained based on the steps of:
and acquiring a relationship tag value between each second sample feature dimension and the target behavior class based on each second sample feature data and the target behavior class data, wherein the target behavior class data comprises behavior data of the sample user under the target behavior class as a second relationship tag value.
In particular, the method comprises the steps of,the sample behavior category data may include sample user +.>Behavior data of individual sample behavior categories, any sample behavior category data can represent the behavior of a sample user in one sample behavior category; />The individual sample behavior class data can be associated with +.>The individual sample behavior categories are in one-to-one correspondence. For->Sample behavior class data +.>Sample behavioral class data, th->Sample behavioral category data and +.>The individual sample behavior categories correspond.
It should be noted that any one of the behavior category data in the embodiments of the present invention may be represented by a data table. The target behavior category data may also be represented by a data table.
Table 2 shows the first embodimentAnd a data table corresponding to the sample behavior class data. As shown in Table 2, the +.>The field name of the data table corresponding to the sample behavior class data comprises the user identification and the identification information of the sample behavior class +.>The variable names of the data table corresponding to the sample behavior category data comprise the identity identification information of the sample userInformation (Identity Document, ID) and a Label value (Label) corresponding to the ID of the sample user and the sample behavior class.
TABLE 2 No. 2Data table corresponding to each sample behavior category data
It should be noted that, in the embodiment of the present invention, the tag value may be used to describe the behavior of a certain sample behavior class.
For example, in the case that the sample behavior class is used for judging whether the consumption behavior occurs in a certain major promotion, the embodiment of the invention can be used for indicating that no behavior occurs in the major promotion and indicating that no behavior occurs in the major promotion by using the tag value of "0";
for another example, in the case where the sample behavior class is used to describe the consumption behavior of different consumption values, in the embodiment of the present invention, the tag value "0" may be used to indicate that the consumption amount is 0 yuan, the tag value "1" is used to indicate that the consumption amount is within [1,100 ], the tag value "2" is used to indicate that the consumption amount is within [100, 1000), and the tag value "3" is used to indicate that the consumption amount is not less than 1000 yuan;
for another example, in the case where the sample behavior category is used to describe the consumption behavior of purchasing brand E and footwear, the embodiment of the present invention may be used for a tag value of "0" to indicate the behavior of not purchasing brand E shoes, and a tag value of "1" to indicate the behavior of purchasing brand E shoes.
The target behavior category data may include behavior data of the sample user in a target behavior category, and the target behavior category data may represent behavior of the sample user in the target behavior category.
Table 3 is a data table corresponding to the target behavior class data. As shown in table 3, the field names of the data table corresponding to the target behavior category data include the user identification and the identification information of the target behavior category, and the variable names of the data table corresponding to the target behavior category data include the identification information (Identity Document, ID) of the sample user and the tag value (Label) corresponding to the ID of the sample user and the target behavior category.
Table 3 data table corresponding to target behavior class data
It should be noted that, in the embodiment of the present invention, the tag value may be used to describe the behavior of the target behavior class.
For example, in the case where the target behavior class is used to describe the consumption behavior of brand E and the consumption amount, in the embodiment of the present invention, the consumption amount of brand E may be 0 element, the consumption amount of brand E may be [1,100 ] by the label value "1", the consumption amount of brand E may be [10,1000 ] by the label value "2", and the consumption amount of brand E may be not less than 1000 element by the label value "3".
Based on the identity information of the sample user, each sample characteristic data and each sample behavior category data can be spliced to obtain the corresponding relation among the identity information of the sample user, the sample characteristic data and the sample behavior category data.
Based on the identity information of the sample user, the first and second sample characteristic data and the target behavior category data can be spliced to obtain the corresponding relation among the identity information of the sample user, the second sample characteristic data and the target behavior category data.
In the embodiment of the invention, the corresponding relation among the identity information of the sample user, the sample characteristic data and the sample behavior class data can be expressed in a tabular mode. Table 4 is a correspondence table among the identity information of the sample user, the sample feature data and the sample behavior class data.
Table 4 correspondence table between identification information of sample user, sample feature data and sample behavior class data
In the embodiment of the invention, the quantitative relation among the identity information of the sample user, the second sample characteristic data and the target behavior class data can be expressed in a tabular mode. Table 5 is a table of correspondence between the identity information of the sample user, the second sample feature data, and the target sample behavior class data.
Table 5 correspondence table between identification information of sample user, second sample feature data and target behavior class data
Based on the correspondence between the identity information of the sample user, the sample feature data and the sample behavior class data, a relationship tag value between each sample feature dimension and each sample behavior class may be calculated as a first relationship tag value.
It should be noted that in the embodiments of the present invention, it is possible to useIndicate->Sample behavior category and->Relationship tag values between individual sample feature dimensions. First->Sample behavior category and->Relation tag value between individual sample feature dimensions +.>Can be used to describe +.>Sample behavior category and->The correlation between the individual sample feature dimensions is strong or weak.
Based on the corresponding relationship among the identity information of the sample user, the second sample feature data and the target sample behavior class data, a relationship tag value between each second sample feature data and the target behavior class data can be calculated and used as a second relationship tag value.
It is understood that any of the second sample feature dimensions is also a sample feature dimension. Embodiments of the invention can be used Indicate->The order of the second sample feature dimension in all sample feature dimensions, i.e. +.>The second sample feature dimension is also +.>A sample feature dimension.
Embodiments of the invention can be usedRepresenting the target behavior category and->A second sample characteristic dimension (th->Individual sample feature dimensions). Target behavior category and->Second sample characteristicsDimension (th->Individual sample feature dimension) a relationship tag value +.>Can be used to describe the above +.>A second sample characteristic dimension (th->Individual sample feature dimensions) and the target behavior class.
In the embodiment of the invention, the first relationship tag value and the second relationship tag value can be expressed in a table form. The first and second relationship tag values are shown in table 6.
TABLE 6 data sheet of first and second relationship tag values
The embodiment of the invention can be calculated in various modesSample behavior category and->Relation tag value between individual sample feature dimensions +.>And target behavior category data and +.>A second sample characteristic dimension (th->Individual sample feature dimension) a relationship tag value +. >
For example, in the embodiment of the present invention, the calculation may be performed by using machine learning model evaluation methods such as ACU (Area Under the Curve), variance, chi-square test, correlation coefficient, mutual information, tree model, etcSample behavior category and->Relation tag value between individual sample feature dimensions +.>And target behavior category data and +.>A second sample characteristic dimension (firstIndividual sample feature dimension) a relationship tag value +.>。
The above-mentioned evaluation methods analyze the relationship between the category and the feature dimension from a certain angle, but all the calculations must be performed by the same calculation method. The quantitative relation has different requirements on the data forms of the category and the characteristic dimension, and some metrics are discrete data and some metrics are continuous data. However, the category and feature dimension may be any data form, so that the discrete data form may be converted into a continuous data form or the continuous data form may be converted into a discrete data form, such as a barrel discretization, according to the requirement of the quantitative relation on the data form, where the data conversion does not affect the method set forth in the present invention.
Step 102, determining a target feature dimension corresponding to the target behavior category in the first sample feature dimensions based on the relation prediction value between each first sample feature dimension and the target behavior category.
Specifically, after the identification information of each first feature dimension and the identification information of the target behavior class are sequentially input into the target relationship prediction model, a relationship prediction value between each first feature dimension and the target behavior class may be obtained.
Based on the relation prediction value between each sample feature dimension and the target behavior category, one or more first sample feature dimensions with highest correlation with the target behavior category can be determined in each first sample feature dimension through a mathematical statistics calculation and other modes, and then the one or more first sample feature dimensions can be determined as target feature dimensions corresponding to the target behavior category.
As an optional embodiment, determining, based on the relationship prediction value between each first sample feature dimension and the target behavior class, the target feature dimension corresponding to the target behavior class in each first sample feature dimension includes: sequencing the feature dimensions of each first sample according to the relation predicted value between the feature dimension of each first sample and the target behavior class;
and selecting a preset number of first sample feature dimensions from the first sample feature dimensions according to the sorting result as target feature dimensions corresponding to the target behavior category.
It should be noted that, in the embodiment of the present invention, the preset number may be determined based on a priori knowledge and/or actual situations. The specific values of the preset number in the embodiment of the present invention are not limited.
And 103, recommending the service to the target user based on the target feature dimension and the target behavior category.
It will be appreciated that in embodiments of the present invention, the number of target users may be one or more.
Specifically, after the target feature dimension corresponding to the target behavior category is determined, feature data of the target feature dimension of the target user can be obtained through data query, user input and other modes.
After the feature data of the target feature dimension of the target user is obtained, service recommendation can be performed on the target user through machine learning, condition judgment and other modes based on the feature data of the target feature dimension of the target user.
As an alternative embodiment, making a service recommendation to a target user based on a target feature dimension and a target behavior class, includes: and determining whether to push the target business corresponding to the target behavior category for the target user based on the feature data of the target user target feature dimension.
It is understood that different behavior categories may correspond to different categories, different brands, or different priced businesses. For example: behavior category 1 may correspond to business of the mother and infant class; behavior class 2 may correspond to traffic with a value exceeding 1000 yuan; behavior class 3 may correspond to a business of brand E; behavior category 4 may correspond to a commodity of the footwear class in brand E.
Specifically, after the feature data of the target feature dimension of the target user is obtained, whether to push the target service corresponding to the target behavior class for the target user can be determined through a machine learning mode, a condition judgment mode and the like based on the feature data of the target feature dimension of the target user.
In the number of sample feature dimensionsIn large cases, the target behavior category +.>The selection of features is time consuming because the sample feature data and sample behavior class data are stored in separate databases, as shown in tables 1 and 2. Splicing each sample feature data and each sample behavior class data requires a lot of resources and time.
Therefore, in the embodiment of the invention, the relationship tag value between each sample feature dimension and each sample behavior category is firstly calculated offline and used as the first relationship tag value, and the first relationship tag value is stored. In the target behavior classWhen the feature selection is performed, a small amount of the feature selection is selectedCalculating the relation label between each second sample feature dimension and the target behavior category by the second sample feature data, performing model training based on the first relation label value and the second relation label value to obtain a target relation prediction model, and predicting the target behavior category and the rest based on the target relation prediction model >A predicted value of the relation between the first sample feature dimensions and based on the target behavior class and the residual ∈ ->The relation predicted value among the first sample feature dimensions determines the target feature dimensions corresponding to the target behavior categories, so that the feature selection efficiency can be greatly improved.
According to the embodiment of the invention, the identification information of each first sample characteristic dimension and the identification information of the target behavior category are sequentially input into the target relation prediction model as a group of information, after the relation prediction value between each first sample characteristic dimension and the target behavior category, which are sequentially output by the target relation prediction model, is obtained, the relation prediction value between each second sample characteristic dimension and the target behavior category is based on the relation prediction value between each first sample characteristic dimension, the target characteristic dimension corresponding to the target behavior category is determined in each first sample characteristic dimension, and then the service recommendation is carried out on the target user based on the target characteristic dimension and the target behavior category, wherein the target relation prediction model is obtained after training based on the first relation label value and the second relation label value, the first relation label value comprises the relation label value between each sample characteristic dimension and each sample behavior category, and the second relation label value between each second sample characteristic dimension and the target behavior category can be more accurately and more efficiently obtained, the association relation between the different sample characteristic dimension and the different sample behavior category can be more accurately and more efficiently obtained, the association between the different sample characteristic dimension and the target behavior category can be more efficiently obtained, the target user can be more efficiently and more accurately perceived, the association between the target user and the target user can be more accurately and more accurately based on the relation between the target characteristics and the target behavior category can be more accurately perceived, and more accurately.
Fig. 2 is a schematic structural diagram of a target relationship prediction model in the service recommendation method provided by the invention. As shown in fig. 2, the target relation prediction model includes: an input layer 201, a feature characterization layer 202, and an output layer 203;
inputting the identification information of each first sample feature dimension and the identification information of the target behavior category in each sample feature dimension into a target relation prediction model to obtain a relation prediction value between each first sample feature dimension and the target behavior category output by the target relation prediction model, wherein the method comprises the following steps: the identification information of each first sample feature dimension and the identification information of the target behavior category are input into the input layer 201, and a first single-hot encoding sparse vector corresponding to the target behavior category and a second single-hot encoding sparse vector corresponding to each first sample feature dimension output by the input layer 201 are obtained.
Specifically, the firstA first sample characteristic dimension (th->A sample feature dimension) is input into the input layer 201, the input layer 201 may input +.>A first sample characteristic dimension (th->Individual sample feature dimension) is converted into a single-hot encoded sparse vector, so that the +.o. output by the input layer 201 can be obtained >A first sample characteristic dimension (th->Individual sample feature dimension) corresponding second one-hot encoded sparse vector>。
Note that, the firstA first sample characteristic dimension (th->Individual sample feature dimension) corresponding second one-hot encoded sparse vector>Can be +.>. First->A first sample characteristic dimension (th->Individual sample feature dimension) corresponding second one-hot encoded sparse vector>The method comprises the steps of carrying out a first treatment on the surface of the Wherein->A first sample characteristic dimension (th->Individual sample feature dimension) corresponding second one-hot encoded sparse vector>Middle->Each position is 1 and the remaining positions are 0.
After the identification information of the target behavior category is input into the input layer 201, the input layer 201 may convert the identification information of the target behavior category into a sparse vector of one-hot encoding, so as to obtain a sparse vector of a first one-hot encoding corresponding to the target behavior category output by the input layer 201。
It should be noted that, the first single thermal encoding sparse vector corresponding to the target behavior classMay be of length of. First one-hot encoded sparse vector corresponding to target behavior class +.>The method comprises the steps of carrying out a first treatment on the surface of the Wherein, the first independent heat coding sparse vector corresponding to the target behavior category is->Middle->Each position is 1 and the remaining positions are 0.
As an alternative embodiment, the input layer 201 includes: a behavior category input layer 204 and a feature dimension input layer 205;
inputting the identification information of each first sample feature dimension and the identification information of the target behavior category into the input layer 201, and obtaining a first single-hot encoding sparse vector corresponding to the target behavior category output by the input layer 201 and a single-hot encoding sparse vector corresponding to each first sample feature dimension, wherein the method comprises the following steps: the identification information of each first sample feature dimension is input into a feature dimension input layer 205, a second single-hot encoding sparse vector output by the feature dimension input layer 205 is obtained, the identification information of the target behavior type is input into a behavior type input layer 204, and a first single-hot encoding sparse vector output by the behavior type input layer 204 is obtained.
Specifically, the firstA first sample characteristic dimension (th->A sample feature dimension) is input to the feature dimension input layer 205, the feature dimension input layer 205 may input +.>A first sample characteristic dimension (th->Individual sample feature dimension) is converted into a single-hot encoded sparse vector, so that the +.o. output by the feature dimension input layer 205 can be obtained>A first sample characteristic dimension (th- >Individual sample feature dimension) corresponding second one-hot encoded sparse vector>。
After the identification information of the target behavior class is input into the behavior class input layer 204, the behavior class input layer 204 may convert the identification information of the target behavior class into a single-hot encoded sparse vector, so as to obtain a first single-hot encoded sparse vector corresponding to the target behavior class output by the behavior class input layer 204。
The first single-hot encoded sparse vector and the second single-hot encoded sparse vector are input to the feature characterization layer 202, and each feature data output by the feature characterization layer 202 is obtained.
Specifically, the firstA first sample characteristic dimension (th->Individual sample feature dimension) corresponding second one-hot encoded sparse vector>First one-hot encoded sparse vector corresponding to target behavior class +.>After inputting the feature characterization layer 202, the feature characterization layer 202 may be based on +.>A first sample characteristic dimension (th->Individual sample feature dimension) corresponding second one-hot encoded sparse vector>First one-hot encoded sparse vector corresponding to target behavior class +.>For->A first sample characteristic dimension (th->Individual sample feature dimension) and the target behavior category, so that the +.f output by the feature characterization layer 202 can be obtained and output >And characteristic data.
As an alternative embodiment, feature characterization layer 202 includes an attribute layer 206, an embedding layer 207, and a relational network layer 208;
the first and second single-hot encoded sparse vectors are input to the feature characterization layer 202 to obtain each feature data output by the feature characterization layer 202, including: the first single-hot encoding sparse vector and the second single-hot encoding sparse vector are input into the attribute layer 206, a first attribute layer representation corresponding to the target behavior category output by the attribute layer 206 and a second attribute layer representation corresponding to each first sample feature dimension are obtained, the first single-hot encoding sparse vector and the second single-hot encoding sparse vector are input into the embedded layer 207, and a first embedded layer representation corresponding to the target behavior category output by the embedded layer 207 and a second embedded layer representation corresponding to each first sample feature dimension are obtained.
It should be noted that, in the embodiment of the present invention, the attribute layer 206 is a linear component, which is simple in form and has good scalability and interpretability.
Specifically, the firstA first sample characteristic dimension (th->Individual sample feature dimension) corresponding second one-hot encoded sparse vector>First one-hot encoded sparse vector corresponding to target behavior class +. >After inputting property layer 206, the +.th of the output of property layer 206 can be obtained>A first sample characteristic dimension (th->Individual sample feature dimension) corresponding second attribute layer characterization +.>And the first attribute layer representation corresponding to the target behavior class +.>。
The embedding layer 207 in the embodiment of the present invention may convert the one-hot encoded sparse vector into a dense vector representation of a preset dimension.
Will be the firstA first sample characteristic dimension (th->Individual sample feature dimensions) corresponding second one-hot encoded sparse vectorsFirst one-hot encoded sparse vector corresponding to target behavior class +.>After inputting the embedded layer 207, the +.th of the output of the embedded layer 207 can be obtained>A first sample characteristic dimension (th->Individual sample feature dimension) corresponding second embedded layer characterization +.>And the first embedded layer representation corresponding to the target behavior class +.>。
Note that, the firstA first sample characteristic dimension (th->Individual samplesFeature dimension) corresponding embedded layer characterization +.>First embedded layer characterization corresponding to the target behavior category +.>Is the same.
As an alternative embodiment, the attribute layer 206 includes: a behavior category attribute layer 209 and a feature dimension attribute layer 210;
first independent-heat coding sparse vector and second independent-heat coding sparse vector corresponding to target behavior category The independent heat coding sparse vector corresponding to the first sample feature dimension is input into the attribute layer 206, and the first attribute layer representation and the +.>Attribute layer characterization corresponding to the first sample feature dimension includes: the first single thermal encoding sparse vector is input into the behavior category attribute layer 209 to obtain a first attribute layer representation corresponding to the target behavior category output by the behavior category attribute layer 209, the second single thermal encoding sparse vector is input into the feature dimension attribute layer 210 to obtain a second attribute layer representation corresponding to each first sample feature dimension output by the feature dimension attribute layer 210.
Specifically, the firstA first sample characteristic dimension (th->Individual sample feature dimension) corresponding second one-hot encoded sparse vector>After the feature dimension attribute layer 210 is input, the feature dimension attribute layer 210 may be calculated based on the following formula to obtain +.>A first sample characteristic dimension (th->Individual sample feature dimension) corresponding second attribute layer characterization +.>:
Wherein,and->Weights and offsets representing linear transformations, respectively, are of size +.>And->;/>Representing a first linear function.
The first independent-heat coding sparse vector corresponding to the target behavior category After the behavior category attribute layer 209 is input, the behavior category attribute layer 209 may calculate the first attribute layer representation corresponding to the target behavior category based on the following formula ∈ ->:
Wherein,and->Weights and offsets representing linear transformations, respectively, are of size +.>And->;/>Representing a second linear function.
As an alternative embodiment, the embedded layer 207 includes: a behavior category embedding layer 211 and a feature dimension embedding layer 212;
first independent-heat coding sparse vector and second independent-heat coding sparse vector corresponding to target behavior categoryThe second single-hot encoding sparse vector corresponding to the characteristic dimension of the first sample is input into the embedding layer 207 to obtain +.>The embedded layer representation corresponding to the first sample feature dimension and the first embedded layer representation corresponding to the target behavior category comprise: inputting a first independent thermal coding sparse vector corresponding to a target behavior type into the behavior type embedding layer 211, obtaining a first embedding layer representation corresponding to the target behavior type output by the behavior type embedding layer 211, and adding +.>The second single thermal encoding sparse vector corresponding to the first sample feature dimension is input into the feature dimension embedding layer 212 to obtain the +.>The embedded layer characterization corresponding to the first sample feature dimension.
Specifically, the firstA first sample characteristic dimension (th->Individual sample feature dimension) corresponding second one-hot encoded sparse vector>After inputting the feature dimension embedding layer 212, the feature dimension embedding layer 212 may calculate the +.>A first sample characteristic dimension (th->Individual sample feature dimension) corresponding second embedded layer characterization +.>:/>
Wherein,representing the characterization function in feature dimension embedding layer 212.
The first independent-heat coding sparse vector corresponding to the target behavior categoryAfter the behavior class embedding layer 211 is input, the behavior class embedding layer 211 may calculate a first embedding layer representation corresponding to the target behavior class based on the following formula ∈ ->:
Wherein,representing the characterization function in behavior category embedding layer 211.
Note that, the firstA first sample characteristic dimension (th->Individual sample feature dimension) corresponding second embedded layer characterization +.>First embedded layer characterization corresponding to target behavior class +.>Are all +.>;/>Indicate->A first sample characteristic dimension (th->Individual sample feature dimension) corresponding embedded layer characterization +.>First embedded layer characterization corresponding to target behavior class +.>Is a dimension of (c).
Inputting the second embedded layer representation and the first embedded layer representation into the relational network layer 208 to obtain a relational vector between each first sample feature dimension and the target behavior class output by the relational network layer 208;
Each feature data comprises a first attribute layer representation and an embedded layer representation corresponding to the target behavior category, an attribute layer representation and an embedded layer representation corresponding to each first sample feature dimension, and a relation vector between each first sample feature dimension and the target behavior category.
Specifically, the firstA first sample characteristic dimension (th->Individual sample feature dimension) corresponding second embedded layer characterization +.>First embedded layer characterization corresponding to target behavior class +.>After entering the relational network layer 208, the relational network layer 208 may be based on acquiring the +.>A first sample characteristic dimension (th->Individual sample feature dimension) corresponding second embedded layer characterization +.>First embedded layer characterization corresponding to target behavior class +.>First->A first sample characteristic dimension (th->Individual sample feature dimension) and the target behavior category, and further can obtain the +.f output by the relational network layer 208>A relationship vector between the first sample feature dimension and the target behavior class.
As an alternative embodiment, the relationship network layer 208 includes a linear relationship network layer 213 and a nonlinear relationship network layer 214;
inputting the first embedded layer representation and the second embedded layer representation into the relational network layer 208 to obtain a relationship vector between each first sample feature dimension and the target behavior class output by the relational network layer 208, comprising: inputting the first embedded layer representation and the second embedded layer representation into the linear relation network layer 213, and obtaining a linear relation representation between a target behavior category output by the linear relation network layer 213 and each first sample feature dimension;
The linear relationship characterization is input into the nonlinear relationship network layer 214, and nonlinear relationship characterization between the target behavior class output by the nonlinear relationship network layer 214 and each first sample feature dimension is obtained.
The relation vector between each first sample feature dimension and the target behavior category comprises a linear relation characterization and a nonlinear relation characterization between each first sample feature dimension and the target behavior category.
In particular, the linear relation network in the embodiment of the invention can use the characterization vector to carry out the following stepsThe first sample feature dimension is associated with a target behavior class, and the relationship between the first sample feature dimension and the target behavior class is modeled linearly. In the above linear relationship, each dimension is independent, so the linear relationship network in the embodiment of the invention can divide +.>A first sample characteristic dimension (th->Individual sample feature dimension) corresponding second embedded layer characterization +.>First embedded layer characterization corresponding to target behavior classThe corresponding dimensions are multiplied directly to maintain dimensional independence. />
Will be the firstA first sample characteristic dimension (th->Individual sample feature dimension) corresponding second embedded layer characterization +.>First embedded layer characterization corresponding to target behavior class +. >After inputting the linear relation network, the linear relation network can calculate the target behavior category and the ++>A first sample characteristic dimension (th->Individual sample feature dimensions) are characterized by a linear relationship ∈>:
Wherein,element-wise multiplication of the representative vector; target behavior category and->Characterization of the linear relation between the characteristic dimensions of the first sample +.>Is +.>。
The nonlinear network in the embodiment of the invention can be used in the following stepA first sample characteristic dimension (th->Individual sample feature dimension) corresponding embedded layer characterization +.>First embedded layer characterization corresponding to target behavior class +.>The hidden layer is added on the basis of splicing, so that interaction between sample behavior categories and sample feature dimensions can be learned based on a standard multi-layer perceptron (Multilayer Perceptron, MLP), and the target relation prediction model has higher nonlinear modeling capacity.
Will be the firstA first sample characteristic dimension (th->Individual sample feature dimension) corresponding second embedded layer characterization +.>First embedded layer characterization corresponding to target behavior class +.>After the nonlinear relation network is input, the nonlinear relation network can calculate and obtain the target behavior category and the ++ >Characterization of non-linear relationship between first sample feature dimensions:
I.e.
Wherein,、/>and->Represents->A weight matrix, a bias matrix, and an activation function for the layer perceptron. For the activation function of each layer of perceptrons, the invention can be implemented in +.>、/>And->Is selected from the group consisting of a plurality of combinations of the above.
Obtain the output of feature dimension attribute layer 210A first sample characteristic dimension (th->Individual sample feature dimension) corresponding second attribute layer characterization +.>First attribute layer representation corresponding to target behavior class output by behavior class attribute layer 209>The +.f of the feature dimension embedding layer 212 output>A first sample characteristic dimension (th->Individual sample feature dimension) corresponding embedded layer characterization +.>Targets output by behavior class embedding layer 211 first embedded layer representation corresponding to behavior category +.>Target behavior category and +.>Characterization of the linear relation between the characteristic dimensions of the first sample +.>And the target behavior class and the +.>Characterization of the nonlinear relation between the characteristic dimensions of the first sample +.>Afterwards, +.>、/>、/>、/>、/>And->Is determined as +.>And characteristic data.
Each feature data is input to the output layer 203, and a predicted value of the relationship between each first sample feature dimension and the target behavior class output by the output layer 203 is obtained.
Specifically, obtain the firstAfter the characteristic data +.>And a feature data input output layer 203.
The output layer 203 will、/>、/>、/>、/>And->After splicing, canOutputting +.>Predicted value of relation between the first sample feature dimension and the target behavior class +.>Taking one layer of MLP as an example, the output layer 203 can obtain the +.>A first sample characteristic dimension (th->Individual sample feature dimension) and target behavior class>:
Wherein,、/>and->Representing a weight matrix, a bias matrix and an activation function; />Representing an activation function, activation function->The forms of sigmoid, tanh, etc. may be selected according to the difference between the sample behavior class and the quantitative relationship form of the sample feature dimension, or the activation function may be selected not to be used.
The target relation prediction model in the embodiment of the invention comprises an input layer, a characteristic characterization layer and an output layer, and can more accurately predict the association relation between each first sample characteristic dimension and the target behavior category based on the association relation between different sample behavior categories and different characteristic dimensions, and can provide a more accurate data basis for service recommendation.
As an alternative embodiment, the target relationship prediction model is trained based on the following steps: training the initial neural network model based on the first relation tag value to obtain a pre-trained neural network model;
training the pre-trained neural network model based on the second relationship label value to obtain a target relationship prediction model.
According to the embodiment of the invention, the initial neural network model is trained based on the first relation tag value, so that the initial neural network model learns the association relation between different sample feature dimensions and different sample behavior categories, after the pre-training neural network model is obtained, the pre-training neural network model is trained based on the second relation tag value, so that the pre-training neural network model learns the relation between each second sample feature dimension and the target behavior category, a target relation prediction model is obtained, and the training efficiency of the target relation prediction model can be improved.
As an alternative embodiment, training the initial neural network model based on the first relationship tag value to obtain a pre-trained neural network model includes: will be the firstIdentification information of individual sample feature dimension and +. >Identification information of each sample behavior category is input into an initial neural network model to obtain the +.>Sample feature dimension and the firstPredicted value of the relation between the individual sample behavior categories, < >>Is greater than zero and less than or equal to->Positive integer of>Is greater than zero and less than or equal to->Positive integer of>And->Are positive integers greater than 1;
based on the firstSample feature dimension and->Predicted value of relation between individual sample behavior categories and the first relation tag value +.>Sample feature dimension and->Calculating a loss function value of the initial neural network model according to the relation tag values among the sample behavior categories;
updating model parameters and updating of the initial neural network model in the event that it is determined that the initial neural network model is not converged based on the loss function value of the initial neural network modelAnd/or +.>Repeating the step of calculating the loss function value of the initial neural network model until the initial neural network model is determined to converge based on the loss function value of the initial neural network model, and obtaining the pre-training neural network model.
It should be noted that, in the embodiment of the present invention, the initial neural network model, the pre-training neural network model, and the target relationship prediction model are different stages of the same model structure. The initial neural network model is a model structure before training, the pre-training neural network model is a model structure in training, and the target relation prediction model is a well-trained model structure.
Specifically, the firstAfter identification information of each sample feature dimension is input to the feature dimension input layer 205, the feature dimension input layer 205 may add +.>Identification information of the individual sample feature dimension is converted into +.>One-hot encoded sparse vector corresponding to each sample feature dimension>。
Wherein, the firstOne-hot encoded sparse vector corresponding to each sample feature dimension>Can be +.>. First->One-hot encoded sparse vector corresponding to each sample feature dimension>The method comprises the steps of carrying out a first treatment on the surface of the Wherein->One-hot encoded sparse vector corresponding to each sample feature dimension>Middle->Each position is 1 and the remaining positions are 0.
Will be the firstAfter the identification information of the individual sample behavior categories is input to the behavior category input layer 204, the behavior category input layer 204 may add +.>Identification information of the individual sample behavior categories is converted into +.>Independent heat coding sparse vector corresponding to each sample behavior category>。
Wherein, the firstIndependent heat coding sparse vector corresponding to each sample behavior category>Can be +.>. First->Independent heat coding sparse vector corresponding to each sample behavior category>The method comprises the steps of carrying out a first treatment on the surface of the Wherein->Independent heat coding sparse vector corresponding to each sample behavior category>Middle->Each position is 1 and the remaining positions are 0.
Will be the firstOne-hot encoded sparse vector corresponding to each sample feature dimension >After the feature dimension attribute layer 210 is input, the feature dimension attribute layer 210 may be calculated based on the following formula to obtain +.>Attribute layer characterization corresponding to individual sample feature dimensions:
Wherein,and->Weights and offsets representing linear transformations, respectively, are of size +.>And->;/>Representing a first linear function.
Will be the firstIndependent heat coding sparse vector corresponding to each sample behavior category>After the behavior category attribute layer 209 is input, the behavior category attribute layer 209 may calculate the first attribute layer representation corresponding to the target behavior category based on the following formula ∈ ->:
Wherein,and->Weights and offsets representing linear transformations, respectively, are of size +.>And->;/>Representing a second linear function.
Will be the firstOne-hot encoded sparse vector corresponding to each sample feature dimension>After inputting the feature dimension embedding layer 212, the feature dimension is embeddedThe in-layer 212 may be calculated based on the following formula>Embedded layer characterization corresponding to individual sample feature dimensions:
。
Will be the firstIndependent heat coding sparse vector corresponding to each sample behavior category>After the behavior class embedding layer 211 is input, the behavior class embedding layer 211 may calculate +.>Embedded layer characterization corresponding to individual sample behavior categories:
。
Note that, the first Embedded layer characterization corresponding to each sample feature dimension +.>And->Embedded layer characterization corresponding to each sample behavior category +.>Are all +.>。
Will be the firstEmbedded layer characterization corresponding to each sample feature dimension +.>And->Embedded layer characterization corresponding to each sample behavior category +.>After inputting the linear relation network, the linear relation network can be calculated by the following formula to obtain the +.>Sample feature dimension and->Characterization of the linear relationship between the individual sample behavior classes +.>:
。
Wherein, the firstSample feature dimension and->Characterization of the linear relationship between the individual sample behavior classes +.>Is +.>。
Will be the firstEmbedded layer characterization corresponding to each sample feature dimension +.>And->Embedded layer characterization corresponding to each sample behavior category +.>After inputting the nonlinear relation network, the nonlinear relation network can calculate the ++>Sample feature dimension and->Characterization of the non-linear relationship between the individual sample behavior categories +.>:
I.e.
。
The input layer 201 will、/>、/>、/>、/>And->Thereafter, the +.>Sample feature dimension and->Predicted value of the relation between the individual sample behavior categories +.>Taking one layer of MLP as an example, the output layer 203 can obtain the +.>Sample feature dimension and- >Predicted value of the relation between the individual sample behavior categories +.>:
。
Acquisition of the firstSample feature dimension and->Predicted value of the relation between the individual sample behavior categories +.>Thereafter, the first relation tag value may be based on +.>Sample feature dimension and->Relation tag value between individual sample behavior categories +.>And->Sample feature dimension and->Predicted value of the relation between the individual sample behavior categories +.>And calculating to obtain the loss function value of the initial neural network model.
Based on the loss function value of the initial neural network model, whether the initial neural network model is converged or not can be judged, and model parameters and updates of the initial neural network model can be updated under the condition that the initial neural network model is not convergedAnd/or +.>Repeating the step of calculating the loss function value of the initial neural network model until the initial neural network model converges to obtain the pre-training neural network model.
Optionally, in an embodiment of the present inventionAnd->The initial value of (2) may be 1, the model parameters of the initial neural network model are updated every time,/->Or->The value of (2) is alternately increased by 1.
In the embodiment of the inventionAnd->The respective inherent bias properties of sample behavior class and feature dimension are characterized, +. >And->The importance of the sample behavior category and the feature dimension on different dimensions of the hidden space is characterized, and the sample behavior category and the feature dimension are +.>Modeling the linear relation of the two by using a linear relation network, wherein the two characteristics still hold the linear independence of the dimensions after the corresponding dimensions are multiplied in the hidden space; />The nonlinear relation between the two is modeled by using a nonlinear relation network, and the complex association between the two is learned through the nonlinear modeling capability of the multi-layer perceptron.
According to the initial neural network model in the embodiment of the invention, the internal relations between different sample behaviors and different characteristic dimensions are comprehensively modeled, and the initial neural network model is trained by continuously inputting identification information of different sample behavior categories, identification information of different sample characteristic dimensions and relation label values between the identification information and the identification information of different sample behavior categories and the identification information of different sample characteristic dimensions and the relation label values between the identification information and the identification information, and by utilizing a Loss function and a gradient back propagation mechanism.
As an alternative embodiment, training the pre-trained neural network model based on the second relationship tag value to obtain the target relationship prediction model includes: will be the firstThe identification information of the characteristic dimension of the second sample and the identification information of the target behavior category are input into a pre-training neural network model to obtain the +. >Predicted value of relation between second sample feature dimension and target behavior class,/for each of the target behavior classes>Is greater than zero and less than or equal to->Is a positive integer of (2);
based on the firstPredicted values of the relation between the second sample feature dimension and the target behavior category and the second relation tag value +.>Calculating a loss function value of the pre-training neural network model according to the relation tag values between the second sample feature dimensions and the target behavior categories; />
Updating model parameters and updating of the pre-trained neural network model in case it is determined that the pre-trained neural network model is not converged based on the loss function value of the pre-trained neural network modelRepeating the step of calculating the loss function value of the pre-training neural network model until the pre-training neural network model is determined to converge based on the loss function value of the pre-training neural network model, and obtaining the target relation prediction model.
Will be the firstA second sample characteristic dimension (th->Individual sample feature dimensions)After the identification information of the identification information and the target behavior category is input into the pre-training neural network model, the +.>A second sample characteristic dimension (th->Individual sample feature dimension) and target behavior class >。
It should be noted that the pre-trained neural network model is calculated to obtain the firstA second sample characteristic dimension (th->Individual sample feature dimension) and target behavior class>The specific steps of the above embodiments may be referred to in the above embodiments, and are not described in detail in the embodiments of the present invention.
Acquisition of the firstA second sample characteristic dimension (th->Individual sample feature dimension) and target behavior class>Thereafter, the +.>A relationship tag value between the second sample feature dimension and the target behavior class +.>And->Sample feature dimension and->Predicted value of the relation between the individual sample behavior categories +.>And calculating to obtain the loss function value of the pre-trained neural network model.
Based on the loss function value of the pre-training neural network model, whether the pre-training neural network model is converged or not can be judged, and model parameters and updates of the pre-training neural network model can be updated under the condition that the pre-training neural network model is not convergedRepeating the step of calculating the loss function value of the pre-training neural network model until the pre-training neural network model converges to obtain the target relation prediction model.
Optionally, in an embodiment of the present inventionThe initial value of (2) may be 1, the model parameters of the pre-trained neural network model are updated every time,/->The value of (2) is increased by 1.
As an alternative embodiment, the loss function of the initial neural network model is determined based on the tag type of the first relationship tag value.
As an alternative embodiment, in case the tag type of the first relational tag value is a classification tag, the loss function of the initial neural network model is a classification loss function; wherein the classification loss function comprises a cross entropy loss function.
Specifically, the tag type with the tag value in the first relation isIn the case of classification labels (AUC), the loss function of the initial neural network modelThe expression can be expressed by the following formula:
as an alternative embodiment, in case the label type of the first relational label value is a correlation coefficient label or a mutual information label, the loss function of the initial neural network model is a regression loss function; wherein the regression loss function includes a mean square error loss.
Specifically, when the first relationship tag value is a correlation coefficient or mutual information, the activation function may be selected not to be used,/>Loss function of initial neural network model +. >The mean square error loss is as follows:
。
as an alternative embodiment, in case the tag type of the first relational tag value is a classification tag, the activation function of the initial neural network model isThe function is activated. />
In particular, in the case where the tag type for which the first relationship tag value is a class tag (AUC),。
in order to facilitate understanding of the feature selection step in the service recommendation method provided by the present invention, the feature selection step in the service recommendation method provided by the present invention is described below by way of an example. Fig. 3 is a flowchart of a feature selection step in the service recommendation method provided by the present invention. As shown in fig. 3, the feature selection step in the service recommendation method provided by the present invention includes: step 301, obtaining feature data of a sample user in each original feature dimension and behavior data of the sample user in each original sample category;
step 302, performing data processing and feature enhancement on the feature data of the sample user in each original feature dimension and the behavior data of the sample user in each original sample category to obtainSample characteristic data and->Sample behavior class data;
step 303, calculating a relationship tag value between each sample feature dimension and each sample behavior category as a first relationship tag value;
Step 304, atSample characteristic data selection->The sample feature data are used as second sample feature data, and a relationship tag value between each second sample feature data and the target behavior category data is calculated and used as a second relationship tag value;
step 305, constructing an initial neural network model, and performing model training based on the first relationship tag value and the second relationship tag value to obtain a target relationship prediction model;
step 306, obtaining a relationship predicted value between each first sample feature dimension and a target behavior class based on the target relationship model;
step 307, determining one or more first sample feature dimensions with highest correlation with the target behavior category in the first sample feature dimensions based on the relation prediction value between each sample feature dimension and the target behavior category, and taking the one or more first sample feature dimensions as target feature dimensions corresponding to the target behavior category.
Aiming at the problem of feature selection, the invention designs a brand-new feature selection method based on deep learning, and the method utilizes the relation between the existing sample category and the features to estimate the quantitative relation between the new sample category and the features, so that the feature selection is carried out according to the size of the estimated quantitative relation, and the method effectively improves the accuracy, generalization and efficiency of the feature selection.
The invention provides a deep learning pre-estimation model of quantitative relation between sample category and characteristic, which can respectively describe the inherent deviation attribute of the sample category and the characteristic, the importance of the sample category and the characteristic in different dimensions of a hidden space, and simultaneously, a linear relation network and a nonlinear relation network are designed to respectively model the linear relation and the nonlinear relation between the sample category and the characteristic.
The model comprehensively models the internal relation between sample class data and feature data, has strong modeling capacity, can effectively estimate the quantitative relation between the sample class and the feature, and has strong generalization performance compared with the traditional method.
Finally, most of the calculated amount of the model can be calculated in advance and reused, and the whole flow has high efficiency.
The invention also adopts a feature enhancement mechanism, can carry out a series of enhancement operations on the original features to generate new features, effectively expands the expression capacity of the features, and finally can select the features more suitable for new sample types.
The feature selection step in the invention can effectively estimate the correlation and quantitative relation between the new sample category and the feature, has higher accuracy, generalization and efficiency, can be used in the feature data processing stage of a server and a machine learning platform, can automatically select and intelligently select the floor features, enriches the product line function, and can effectively support the construction of the server and the machine learning platform. Meanwhile, the method can be further expanded into various scenes facing data selection, and can be applied to the product development aspect of the MetaEngine supporting the metauniverse server.
Fig. 4 is a schematic structural diagram of a service recommendation device provided by the invention. The service recommending device provided by the present invention will be described below with reference to fig. 4, and the service recommending device described below and the service recommending method provided by the present invention described above may be referred to correspondingly. As shown in fig. 4, a data input module 401, a feature selection module 402, and a service recommendation module 403.
The data input module 401 is configured to input identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension into a target relationship prediction model, obtain a relationship prediction value between each first sample feature dimension and the target behavior category output by the target relationship prediction model, and obtain the target relationship prediction model after training based on a first relationship tag value and a second relationship tag value, where the first relationship tag value includes a relationship tag value between each sample feature dimension and each sample behavior category, and the second relationship tag value includes a relationship tag value between each second sample feature dimension in each sample feature dimension and the target behavior category;
the feature selection module 402 determines a target feature dimension corresponding to the target behavior category from among the first sample feature dimensions based on a relationship prediction value between each of the first sample feature dimensions and the target behavior category;
A service recommendation module 403, configured to recommend a service to a target user based on the target feature dimension and the target behavior class;
wherein the sample feature dimension comprises a user feature dimension and/or a historical behavior dimension; the user feature dimensions include feature dimensions for describing the individual users, including: at least one of an age dimension, a gender dimension, an academic dimension, a locale dimension, a occupation dimension, and a hobby dimension; the historical behavior dimension comprises a characteristic dimension for describing the historical behavior of the user, and comprises at least one of a dimension of the historical browsing behavior of the user, a dimension of the historical clicking behavior of the user, a dimension of the scoring behavior of the user, a dimension of the feedback behavior of the user, a dimension of the social interaction behavior of the user and a dimension of the reading behavior of the user;
behavior categories are used to describe behaviors of different dimensions and different granularities; the target behavior category is a different behavior category than the sample behavior category.
Specifically, the data input module 401, the feature selection module 402, and the service recommendation module 403 are electrically connected.
According to the service recommendation device in the embodiment of the invention, the identification information of each first sample feature dimension and the identification information of the target behavior category are sequentially input into the target relationship prediction model as a group of information, after the relationship prediction value between each first sample feature dimension and the target behavior category, which are sequentially output by the target relationship prediction model, is obtained, the relationship prediction value between each second sample feature dimension and the target behavior category is based on the relationship prediction value between each first sample feature dimension, the target feature dimension corresponding to the target behavior category is determined in each first sample feature dimension, and then the service recommendation is carried out on the target user based on the target feature dimension and the target behavior category, wherein the target relationship prediction model is obtained after training based on the first relationship label value and the second relationship label value, the first relationship label value comprises the relationship label value between each sample feature dimension and each sample behavior category, the second relationship label value comprises the relationship label value between each sample feature dimension and the target behavior category, the association relationship between different sample feature dimensions and the target behavior category can be more accurately and efficiently obtained, the association relationship between different sample feature dimensions and different behavior categories can be more accurately and efficiently obtained, the association between different sample feature dimensions and different target behavior categories can be more efficiently obtained, the association between the target feature dimensions and the target feature categories can be more accurately and more accurately obtained, and the association between the target feature categories can be more accurately and more accurately based on the association between the target feature and the target feature categories and the target behavior categories.
Fig. 5 illustrates a physical schematic diagram of an electronic device, as shown in fig. 5, which may include: a processor (processor) 501, a communication interface (Communications Interface) 502, a memory (memory) 503 and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 communicate with each other via the communication bus 504. The processor 501 may invoke logic instructions in the memory 503 to perform a business recommendation method comprising: inputting identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension into a target relation prediction model to obtain a relation prediction value between each first sample feature dimension and the target behavior category output by the target relation prediction model, wherein the target relation prediction model is obtained after training based on a first relation tag value and a second relation tag value, the first relation tag value comprises a relation tag value between each sample feature dimension and each sample behavior category, and the second relation tag value comprises a relation tag value between each second sample feature dimension in each sample feature dimension and the target behavior category; determining a target feature dimension corresponding to the target behavior category in the first sample feature dimensions based on a relation prediction value between each first sample feature dimension and the target behavior category; based on the target feature dimension and the target behavior category, carrying out service recommendation on the target user; wherein the sample feature dimension comprises a user feature dimension and/or a historical behavior dimension; the user feature dimensions include feature dimensions for describing the individual users, including: at least one of an age dimension, a gender dimension, an academic dimension, a locale dimension, a occupation dimension, and a hobby dimension; the historical behavior dimension comprises a characteristic dimension for describing the historical behavior of the user, and comprises at least one of a dimension of the historical browsing behavior of the user, a dimension of the historical clicking behavior of the user, a dimension of the scoring behavior of the user, a dimension of the feedback behavior of the user, a dimension of the social interaction behavior of the user and a dimension of the reading behavior of the user; behavior categories are used to describe behaviors of different dimensions and different granularities; the target behavior category is a different behavior category than the sample behavior category.
Further, the logic instructions in the memory 503 described above may be implemented in the form of software functional units and may be stored in a computer readable storage medium when sold or used as a stand alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
In another aspect, the present invention also provides a computer program product comprising a computer program stored on a non-transitory computer readable storage medium, the computer program comprising program instructions which, when executed by a computer, enable the computer to perform a service recommendation method provided by the above methods, the method comprising: inputting identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension into a target relation prediction model to obtain a relation prediction value between each first sample feature dimension and the target behavior category output by the target relation prediction model, wherein the target relation prediction model is obtained after training based on a first relation tag value and a second relation tag value, the first relation tag value comprises a relation tag value between each sample feature dimension and each sample behavior category, and the second relation tag value comprises a relation tag value between each second sample feature dimension in each sample feature dimension and the target behavior category; determining a target feature dimension corresponding to the target behavior category in the first sample feature dimensions based on a relation prediction value between each first sample feature dimension and the target behavior category; based on the target feature dimension and the target behavior category, carrying out service recommendation on the target user; wherein the sample feature dimension comprises a user feature dimension and/or a historical behavior dimension; the user feature dimensions include feature dimensions for describing the individual users, including: at least one of an age dimension, a gender dimension, an academic dimension, a locale dimension, a occupation dimension, and a hobby dimension; the historical behavior dimension comprises a characteristic dimension for describing the historical behavior of the user, and comprises at least one of a dimension of the historical browsing behavior of the user, a dimension of the historical clicking behavior of the user, a dimension of the scoring behavior of the user, a dimension of the feedback behavior of the user, a dimension of the social interaction behavior of the user and a dimension of the reading behavior of the user; behavior categories are used to describe behaviors of different dimensions and different granularities; the target behavior category is a different behavior category than the sample behavior category.
In yet another aspect, the present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, is implemented to perform the above-mentioned provided service recommendation methods, the method comprising: inputting identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension into a target relation prediction model to obtain a relation prediction value between each first sample feature dimension and the target behavior category output by the target relation prediction model, wherein the target relation prediction model is obtained after training based on a first relation tag value and a second relation tag value, the first relation tag value comprises a relation tag value between each sample feature dimension and each sample behavior category, and the second relation tag value comprises a relation tag value between each second sample feature dimension in each sample feature dimension and the target behavior category; determining a target feature dimension corresponding to the target behavior category in the first sample feature dimensions based on a relation prediction value between each first sample feature dimension and the target behavior category; based on the target feature dimension and the target behavior category, carrying out service recommendation on the target user; wherein the sample feature dimension comprises a user feature dimension and/or a historical behavior dimension; the user feature dimensions include feature dimensions for describing the individual users, including: at least one of an age dimension, a gender dimension, an academic dimension, a locale dimension, a occupation dimension, and a hobby dimension; the historical behavior dimension comprises a characteristic dimension for describing the historical behavior of the user, and comprises at least one of a dimension of the historical browsing behavior of the user, a dimension of the historical clicking behavior of the user, a dimension of the scoring behavior of the user, a dimension of the feedback behavior of the user, a dimension of the social interaction behavior of the user and a dimension of the reading behavior of the user; behavior categories are used to describe behaviors of different dimensions and different granularities; the target behavior category is a different behavior category than the sample behavior category.
The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.
From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (20)
1. A service recommendation method, wherein the service comprises: any one of goods, services, and content; the method comprises the following steps:
inputting identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension into a target relation prediction model to obtain a relation prediction value between each first sample feature dimension and the target behavior category output by the target relation prediction model, wherein the target relation prediction model is obtained after training based on a first relation tag value and a second relation tag value, the first relation tag value comprises a relation tag value between each sample feature dimension and each sample behavior category, and the second relation tag value comprises a relation tag value between each second sample feature dimension in each sample feature dimension and the target behavior category;
Determining a target feature dimension corresponding to the target behavior category in each first sample feature dimension based on a relation prediction value between each first sample feature dimension and the target behavior category;
based on the target feature dimension and the target behavior category, performing service recommendation on a target user;
wherein the sample feature dimension comprises a user feature dimension and/or a historical behavior dimension; the user feature dimension comprises a feature dimension for describing individual users; the historical behavior dimension comprises a feature dimension for describing a user's historical behavior; behavior categories are used to describe behaviors of different dimensions and different granularities; the target behavior category and the sample behavior category are different behavior categories;
before the identification information of each first sample feature dimension and the identification information of the target behavior category in each sample feature dimension are input into the target relation prediction model to obtain the relation prediction value between each first sample feature dimension and the target behavior category output by the target relation prediction model, the method further comprises:
acquiring sample feature data corresponding to each sample feature dimension and sample behavior category data corresponding to each sample behavior category, wherein each sample feature data comprises feature data of a sample user under each sample feature dimension, and each sample behavior category data comprises behavior data of the sample user under each sample behavior category;
Determining part of sample characteristic data in each sample characteristic data as first sample characteristic data, and determining the rest sample characteristic data as second sample characteristic data;
determining a sample feature dimension corresponding to each first sample feature data as each first sample feature dimension, and determining each sample feature dimension corresponding to each second sample feature data as each second sample feature dimension;
wherein the user feature dimension includes at least one of an age dimension, a gender dimension, an academic dimension, a locale dimension, a occupation dimension, and a hobby dimension; the historical behavior dimension comprises at least one of a user historical browsing behavior dimension, a user historical clicking behavior dimension, a user historical scoring behavior dimension, a user historical feedback behavior dimension, a user historical social interaction behavior dimension and a user historical reading behavior dimension;
the target relation prediction model is trained based on the following steps: training the initial neural network model based on the first relation tag value to obtain a pre-trained neural network model;
training the pre-trained neural network model based on the second relationship label value to obtain the target relationship prediction model;
Training the initial neural network model based on the first relation tag value to obtain a pre-trained neural network model, wherein the training comprises the following steps:
will be the firstDimension of individual sample featuresIdentification information and->Identification information of each sample behavior category is input into the initial neural network model to obtain the +.>A sample feature dimension and said +.>Predicted value of the relation between the individual sample behavior categories, < >>Is greater than zero and less than or equal to->Positive integer of>Is greater than zero and less than or equal to->Positive integer of>Is a positive integer greater than 1, +.>Representing a total number of each of the sample feature dimensions; />Representing the number of sample behavior categories;
based on the firstA sample feature dimension and said +.>A predicted value of a relation between the sample behavior categories and the first relation tag value>A sample feature dimension and said +.>Calculating a loss function value of the initial neural network model according to the relation tag values among the sample behavior categories;
updating model parameters and updating of the initial neural network model in the event that it is determined that the initial neural network model is not converged based on the loss function value of the initial neural network model And/or +.>Repeating the step of calculating the loss function value of the initial neural network model until the initial neural network model is determined to converge based on the loss function value of the initial neural network model, so as to obtain the pre-training neural network model;
training the pre-training neural network model based on the second relationship label value to obtain the target relationship prediction model, including:
will be the firstIdentification information of a second sample feature dimension and identification information of a target behavior category are input into the pre-training neural network model to obtain the +.>Predicted values of the relationship between the second sample feature dimension and the target behavior class, ++>Is greater than zero and less than or equal to->Positive integer of>Representing a number of the second sample feature dimensions;
based on the firstPredicted values of the relationship between a second sample feature dimension and the target behavior class and the second relationship tag value +.>Calculating a loss function value of the pre-training neural network model according to the relation tag values between the second sample feature dimension and the target behavior class;
updating model parameters and updating of the pre-trained neural network model in the event that it is determined that the pre-trained neural network model is not converging based on the loss function values of the pre-trained neural network model Repeating the step of calculating the loss function value of the pre-training neural network model until the pre-training neural network model is determined to converge based on the loss function value of the pre-training neural network model, and obtaining the target relation prediction model.
2. The business recommendation method according to claim 1, wherein the target relation prediction model comprises: an input layer, a feature characterization layer, and an output layer;
inputting identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension into a target relationship prediction model, and obtaining a relationship prediction value between each first sample feature dimension and the target behavior category output by the target relationship prediction model, wherein the method comprises the following steps:
inputting the identification information of each first sample feature dimension and the identification information of the target behavior category into the input layer to obtain a first single-hot coding sparse vector corresponding to the target behavior category output by the input layer and a second single-hot coding sparse vector corresponding to each first sample feature dimension;
inputting the first single thermal coding sparse vector and the second single thermal coding sparse vector into the characteristic characterization layer to obtain each characteristic data output by the characteristic characterization layer;
And inputting each characteristic data into the output layer to obtain a relation prediction value between each first sample characteristic dimension output by the output layer and the target behavior class.
3. The service recommendation method according to claim 2, wherein the input layer comprises: a behavior category input layer and a feature dimension input layer;
inputting the identification information of each first sample feature dimension and the identification information of the target behavior category into the input layer to obtain a first single-hot encoding sparse vector corresponding to the target behavior category output by the input layer and a second single-hot encoding sparse vector corresponding to each first sample feature dimension, wherein the method comprises the following steps:
inputting the identification information of each first sample feature dimension into the feature dimension input layer to obtain the second single-hot encoding sparse vector output by the feature dimension input layer,
and inputting the identification information of the target behavior category into the behavior category input layer to obtain the first single-hot coding sparse vector output by the behavior category input layer.
4. The business recommendation method according to claim 2, wherein the feature characterization layer comprises an attribute layer, an embedding layer and a relational network layer;
The inputting the first single-hot encoded sparse vector and the second single-hot encoded sparse vector to the feature characterization layer, obtaining each feature data output by the feature characterization layer, includes:
inputting the first single thermal encoding sparse vector and the second single thermal encoding sparse vector into the attribute layer to obtain a first attribute layer representation corresponding to the target behavior category output by the attribute layer and a second attribute layer representation corresponding to each first sample feature dimension,
inputting the first single-hot encoding sparse vector and the second single-hot encoding sparse vector into the embedded layer to obtain a first embedded layer representation corresponding to the target behavior category output by the embedded layer and a second embedded layer representation corresponding to each first sample feature dimension;
inputting the first embedded layer representation and the second embedded layer representation into the relational network layer to obtain a relation vector between each first sample feature dimension output by the relational network layer and the target behavior class;
the feature data comprises a first attribute layer representation and a second embedded layer representation corresponding to the target behavior category, a second attribute layer representation and a second embedded layer representation corresponding to each first sample feature dimension, and a relation vector between each first sample feature dimension and the target behavior category.
5. The service recommendation method according to claim 4, wherein the attribute layer comprises: a behavior category attribute layer and a feature dimension attribute layer;
inputting the first single-hot encoding sparse vector and the second single-hot encoding sparse vector into the attribute layer to obtain a first attribute layer representation corresponding to the target behavior category output by the attribute layer and a second attribute layer representation corresponding to each first sample feature dimension, wherein the method comprises the following steps:
inputting the first single thermal encoding sparse vector into the behavior category attribute layer to obtain the first attribute layer representation output by the behavior category attribute layer,
and inputting the second independent heat coding sparse vector into the characteristic dimension attribute layer to obtain the second attribute layer representation output by the characteristic dimension attribute layer.
6. The service recommendation method according to claim 4, wherein the embedding layer comprises: a behavior category embedding layer and a feature dimension embedding layer;
inputting the first single-hot encoded sparse vector and the second single-hot encoded sparse vector into the embedded layer to obtain a first embedded layer representation corresponding to the target behavior class and a second embedded layer representation corresponding to each first sample feature dimension output by the embedded layer, wherein the method comprises the following steps:
Inputting the first single thermal encoding sparse vector into the behavior class embedding layer to obtain the first embedding layer representation output by the behavior class embedding layer,
and inputting the second independent thermal coding sparse vector into the characteristic dimension embedded layer to obtain the second embedded layer representation output by the characteristic dimension embedded layer.
7. The service recommendation method according to claim 4, wherein the relational network layer comprises a linear relational network layer and a nonlinear relational network layer;
inputting the first embedded layer representation and the second embedded layer representation into the relational network layer to obtain a relationship vector between each first sample feature dimension and the target behavior class output by the relational network layer, comprising:
inputting the first embedded layer representation and the second embedded layer representation into the linear relation network layer to obtain linear relation representations between the target behavior category output by the linear relation network layer and each first sample feature dimension;
inputting the linear relation representation into the nonlinear relation network layer, and obtaining nonlinear relation representation between the target behavior category output by the nonlinear relation network layer and each first sample feature dimension;
The relation vector between each first sample feature dimension and the target behavior category comprises linear relation characterization and nonlinear relation characterization between each first sample feature dimension and the target behavior category.
8. The service recommendation method according to claim 1, wherein the obtaining sample feature data corresponding to each sample feature dimension includes:
acquiring feature data of the sample user under each original feature dimension as first original sample feature data;
performing data preprocessing on the first original sample characteristic data, and obtaining second original sample characteristic data according to a data preprocessing result;
performing feature enhancement processing on the second original sample feature data to obtain feature data of the sample user in each sample feature dimension as each sample feature data;
wherein the data preprocessing includes at least one of data format detection, outlier processing, duplicate value processing, and missing value processing; the feature enhancement processing comprises at least one of function conversion processing, feature scaling processing, dimensionless processing, numerical feature sub-bucket and feature cross combination.
9. The service recommendation method according to claim 1, wherein the obtaining sample feature data corresponding to each sample feature dimension includes:
acquiring behavior data of the sample user under each original sample behavior category as first original sample behavior category data;
performing data preprocessing on the first original sample behavior category data, and obtaining second original sample behavior category data according to a data preprocessing result;
performing feature enhancement processing on the second original sample behavior category data to obtain feature data of the sample user under each sample behavior category as each sample behavior category data;
wherein the data preprocessing includes at least one of data format detection, outlier processing, duplicate value processing, and missing value processing; the feature enhancement processing comprises at least one of function conversion processing, feature scaling processing, dimensionless processing, numerical feature sub-bucket and feature cross combination.
10. The service recommendation method according to claim 1, wherein the first relationship tag value is obtained based on the steps of:
And acquiring a relationship tag value between each sample feature dimension and each sample behavior category as the first relationship tag value based on each sample feature data and each sample behavior category data.
11. The service recommendation method according to claim 1, wherein the second relationship tag value is obtained based on the steps of:
and acquiring a relationship tag value between each second sample feature dimension and the target behavior class based on each second sample feature data and target behavior class data, wherein the target behavior class data comprises behavior data of the sample user under the target behavior class, and the relationship tag value is used as the second relationship tag value.
12. The service recommendation method according to claim 1, wherein determining, among the first sample feature dimensions, a target feature dimension corresponding to the target behavior category based on a relationship prediction value between each of the first sample feature dimensions and the target behavior category, comprises:
sorting the first sample feature dimensions according to a relation prediction value between each first sample feature dimension and the target behavior category;
And selecting a preset number of first sample feature dimensions from the first sample feature dimensions according to the sorting result, wherein the first sample feature dimensions are used as target feature dimensions corresponding to the target behavior categories.
13. The service recommendation method according to claim 1, wherein the performing service recommendation on the target user based on the target feature dimension and the target behavior class includes:
and determining whether to push the target service corresponding to the target behavior class for the target user based on the feature data of the target feature dimension of the target user.
14. The business recommendation method of claim 1 wherein the loss function of the initial neural network model is determined based on a tag type of the first relationship tag value.
15. The traffic recommendation method according to claim 14, wherein in case the tag type of the first relationship tag value is a classification tag, the loss function of the initial neural network model is a classification loss function; wherein the classification loss function comprises a cross entropy loss function.
16. The service recommendation method according to claim 14, wherein in the case where the tag type of the first relationship tag value is a correlation coefficient tag or a mutual information tag, the loss function of the initial neural network model is a regression loss function; wherein the regression loss function includes a mean square error loss.
17. The traffic recommendation method according to claim 16, wherein in case the tag type of the first relationship tag value is a classification tag, the activation function of the initial neural network model isThe function is activated.
18. A service recommendation device, wherein the service comprises: any one of goods, services, and content; the device comprises:
the data input module is used for inputting identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension into a target relation prediction model, obtaining a relation prediction value between each first sample feature dimension and the target behavior category output by the target relation prediction model, wherein the target relation prediction model is obtained after training based on a first relation tag value and a second relation tag value, the first relation tag value comprises a relation tag value between each sample feature dimension and each sample behavior category, and the second relation tag value comprises a relation tag value between each second sample feature dimension in each sample feature dimension and the target behavior category;
The feature selection module is used for determining a target feature dimension corresponding to the target behavior category in the first sample feature dimensions based on a relation prediction value between each first sample feature dimension and the target behavior category;
the service recommendation module is used for recommending the service to the target user based on the target feature dimension and the target behavior category;
wherein the sample feature dimension comprises a user feature dimension and/or a historical behavior dimension; the user feature dimension comprises a feature dimension for describing individual users; the historical behavior dimension comprises a feature dimension for describing a user's historical behavior; behavior categories are used to describe behaviors of different dimensions and different granularities; the target behavior category and the sample behavior category are different behavior categories;
the data input module inputs identification information of each first sample feature dimension and identification information of a target behavior category in each sample feature dimension into a target relationship prediction model, and before obtaining a relationship prediction value between each first sample feature dimension and the target behavior category output by the target relationship prediction model, the data input module is further configured to:
Acquiring sample feature data corresponding to each sample feature dimension and sample behavior category data corresponding to each sample behavior category, wherein each sample feature data comprises feature data of a sample user under each sample feature dimension, and each sample behavior category data comprises behavior data of the sample user under each sample behavior category;
determining part of sample characteristic data in each sample characteristic data as first sample characteristic data, and determining the rest sample characteristic data as second sample characteristic data;
determining a sample feature dimension corresponding to each first sample feature data as each first sample feature dimension, and determining each sample feature dimension corresponding to each second sample feature data as each second sample feature dimension;
wherein the user feature dimension includes at least one of an age dimension, a gender dimension, an academic dimension, a locale dimension, a occupation dimension, and a hobby dimension; the historical behavior dimension comprises at least one of a user historical browsing behavior dimension, a user historical clicking behavior dimension, a user historical scoring behavior dimension, a user historical feedback behavior dimension, a user historical social interaction behavior dimension and a user historical reading behavior dimension;
The target relation prediction model is trained based on the following steps:
training the initial neural network model based on the first relation tag value to obtain a pre-trained neural network model;
training the pre-trained neural network model based on the second relationship label value to obtain the target relationship prediction model;
training the initial neural network model based on the first relation tag value to obtain a pre-trained neural network model, wherein the training comprises the following steps:
will be the firstIdentification information of individual sample feature dimension and +.>Identification information of each sample behavior category is input into the initial neural network model to obtain the +.>A sample feature dimension and said +.>Predicted value of the relation between the individual sample behavior categories, < >>Is greater than zero and less than or equal to->Positive integer of>Is greater than zero and less than or equal to->Positive integer of>Is a positive integer greater than 1, +.>Representing a total number of each of the sample feature dimensions; />Representing the number of sample behavior categories;
based on the firstA sample feature dimension and said +.>A predicted value of a relation between the sample behavior categories and the first relation tag value >A sample feature dimension and said +.>Calculating a loss function value of the initial neural network model according to the relation tag values among the sample behavior categories;
updating model parameters and updating of the initial neural network model in the event that it is determined that the initial neural network model is not converged based on the loss function value of the initial neural network modelAnd/or +.>Repeating the step of calculating the loss function value of the initial neural network model until the initial neural network model is determined to converge based on the loss function value of the initial neural network model, so as to obtain the pre-training neural network model;
training the pre-training neural network model based on the second relationship label value to obtain the target relationship prediction model, including:
will be the firstIdentification information of a second sample feature dimension and identification information of a target behavior category are input into the pre-training neural network model to obtain the +.>Predicted values of the relationship between the second sample feature dimension and the target behavior class, ++>Is greater than zero and less than or equal to->Positive integer of>Representing a number of the second sample feature dimensions;
Based on the firstPredicted values of the relationship between a second sample feature dimension and the target behavior class and the second relationship tag value +.>Calculating a loss function value of the pre-training neural network model according to the relation tag values between the second sample feature dimension and the target behavior class;
updating model parameters and updating of the pre-trained neural network model in the event that it is determined that the pre-trained neural network model is not converging based on the loss function values of the pre-trained neural network modelRepeating the step of calculating the loss function value of the pre-training neural network model until the pre-training neural network model is determined to converge based on the loss function value of the pre-training neural network model, and obtaining the target relation prediction model.
19. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the service recommendation method according to any one of claims 1 to 17 when the program is executed by the processor.
20. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the steps of the service recommendation method according to any of claims 1 to 17.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311508557.1A CN117252665B (en) | 2023-11-14 | 2023-11-14 | Service recommendation method and device, electronic equipment and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311508557.1A CN117252665B (en) | 2023-11-14 | 2023-11-14 | Service recommendation method and device, electronic equipment and storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN117252665A CN117252665A (en) | 2023-12-19 |
CN117252665B true CN117252665B (en) | 2024-02-20 |
Family
ID=89137128
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202311508557.1A Active CN117252665B (en) | 2023-11-14 | 2023-11-14 | Service recommendation method and device, electronic equipment and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117252665B (en) |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112633962A (en) * | 2020-12-03 | 2021-04-09 | 北京道隆华尔软件股份有限公司 | Service recommendation method and device, computer equipment and storage medium |
CN113268656A (en) * | 2021-04-15 | 2021-08-17 | 北京沃东天骏信息技术有限公司 | User recommendation method and device, electronic equipment and computer storage medium |
CN113626719A (en) * | 2021-10-12 | 2021-11-09 | 腾讯科技(深圳)有限公司 | Information recommendation method, device, equipment, storage medium and computer program product |
CN115048855A (en) * | 2022-05-06 | 2022-09-13 | 南宁师范大学 | Click rate prediction model, training method and application device thereof |
CN116340643A (en) * | 2023-05-29 | 2023-06-27 | 苏州浪潮智能科技有限公司 | Object recommendation adjustment method and device, storage medium and electronic equipment |
CN116737334A (en) * | 2023-05-31 | 2023-09-12 | 嵩山实验室 | Task scheduling and data set label updating method and device and electronic equipment |
-
2023
- 2023-11-14 CN CN202311508557.1A patent/CN117252665B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112633962A (en) * | 2020-12-03 | 2021-04-09 | 北京道隆华尔软件股份有限公司 | Service recommendation method and device, computer equipment and storage medium |
CN113268656A (en) * | 2021-04-15 | 2021-08-17 | 北京沃东天骏信息技术有限公司 | User recommendation method and device, electronic equipment and computer storage medium |
CN113626719A (en) * | 2021-10-12 | 2021-11-09 | 腾讯科技(深圳)有限公司 | Information recommendation method, device, equipment, storage medium and computer program product |
CN115048855A (en) * | 2022-05-06 | 2022-09-13 | 南宁师范大学 | Click rate prediction model, training method and application device thereof |
CN116340643A (en) * | 2023-05-29 | 2023-06-27 | 苏州浪潮智能科技有限公司 | Object recommendation adjustment method and device, storage medium and electronic equipment |
CN116737334A (en) * | 2023-05-31 | 2023-09-12 | 嵩山实验室 | Task scheduling and data set label updating method and device and electronic equipment |
Also Published As
Publication number | Publication date |
---|---|
CN117252665A (en) | 2023-12-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP7360497B2 (en) | Cross-modal feature extraction method, extraction device, and program | |
CN108648049B (en) | Sequence recommendation method based on user behavior difference modeling | |
CN110717098B (en) | Meta-path-based context-aware user modeling method and sequence recommendation method | |
CN111553759A (en) | Product information pushing method, device, equipment and storage medium | |
CN113256367B (en) | Commodity recommendation method, system, equipment and medium for user behavior history data | |
US11640634B2 (en) | Deep learning based visual compatibility prediction for bundle recommendations | |
CN112650929B (en) | Graph neural network recommendation method integrating comment information | |
CN112800344B (en) | Deep neural network-based movie recommendation method | |
CN112529071B (en) | Text classification method, system, computer equipment and storage medium | |
CN110727855A (en) | Personalized recommendation method based on improved factorization machine | |
CN114077661A (en) | Information processing apparatus, information processing method, and computer readable medium | |
CN114358657A (en) | Post recommendation method and device based on model fusion | |
CN107169830B (en) | Personalized recommendation method based on clustering PU matrix decomposition | |
CN111178986A (en) | User-commodity preference prediction method and system | |
CN117574915A (en) | Public data platform based on multiparty data sources and data analysis method thereof | |
CN114842247B (en) | Characteristic accumulation-based graph convolution network semi-supervised node classification method | |
CN117252665B (en) | Service recommendation method and device, electronic equipment and storage medium | |
Zhang et al. | A two-stage rating prediction approach based on matrix clustering on implicit information | |
CN115344794A (en) | Scenic spot recommendation method based on knowledge map semantic embedding | |
CN114936901A (en) | Visual perception recommendation method and system based on cross-modal semantic reasoning and fusion | |
Pan et al. | A movie recommendation model combining time information and probability matrix factorisation | |
Hmaidi et al. | Anime Link Prediction Using Improved Graph Convolutional Networks | |
Monemian | A neuroevolutionary neural network-based collaborative filtering recommendation system | |
CN113255891B (en) | Method, neural network model and device for processing event characteristics | |
CN116957740B (en) | Agricultural product recommendation system based on word characteristics |
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 |