CN113742594A - Recommendation system recall method and device - Google Patents

Recommendation system recall method and device Download PDF

Info

Publication number
CN113742594A
CN113742594A CN202111088806.7A CN202111088806A CN113742594A CN 113742594 A CN113742594 A CN 113742594A CN 202111088806 A CN202111088806 A CN 202111088806A CN 113742594 A CN113742594 A CN 113742594A
Authority
CN
China
Prior art keywords
user
vector
project
item
output
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202111088806.7A
Other languages
Chinese (zh)
Other versions
CN113742594B (en
Inventor
卢伟
李瑞男
何聪聪
董志刚
尹莉莉
樊泽泽
麻聃
张博
王宇虹
宋柯欣
马慧慧
赵军艳
杜敏琪
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Bank of China Ltd
Original Assignee
Bank of China Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Bank of China Ltd filed Critical Bank of China Ltd
Priority to CN202111088806.7A priority Critical patent/CN113742594B/en
Publication of CN113742594A publication Critical patent/CN113742594A/en
Application granted granted Critical
Publication of CN113742594B publication Critical patent/CN113742594B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

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

Abstract

The invention discloses a method and a device for recalling a recommendation system, which relate to the technical field of artificial intelligence, and comprise the following steps: respectively carrying out discretization and vectorization processing on the user characteristics and the project characteristics, respectively inputting the user characteristics and the project characteristics into a multilayer fully-connected neural network to obtain a deep learning output vector, and accumulating the discrete project characteristics of the user characteristics and the project characteristics to obtain a linear model output vector; splicing the deep learning output vector and the linear model output vector to obtain an output total vector, performing dot product multiplication on the output total vector of the user characteristics and the output total vector of the project characteristics to obtain a calculation model of the matching degree between the user characteristics and the project characteristics, and calculating the matching degree between the user characteristics of the recall target user and the project characteristics of the recall target project; and performing recall operation according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item. The method and the system can improve the recall accuracy and recall efficiency of the recommendation system.

Description

Recommendation system recall method and device
Technical Field
The invention relates to the technical field of artificial intelligence, in particular to a method and a device for recalling a recommendation system.
Background
This section is intended to provide a background or context to the embodiments of the invention that are recited in the claims. The description herein is not admitted to be prior art by inclusion in this section.
With the social development and the popularization of the internet, the information quantity is increased more and more quickly, people find information and people find information, and the recommendation system is intelligent and personalized to extract useful information from mass information and actively push the useful information to people.
The recommendation system needs to find out matching information from mass data information for personalized pushing, generally comprises two links of recall and sequencing in consideration of time overhead, and the recall link is responsible for quickly finding n pieces of information which are possibly matched from the mass information. Vector retrieval is one way to recall links.
Since the deep learning model currently used for recommendation system recall is generally a double tower architecture, the output of the double tower architecture is in the form of vectors, and then vector retrieval is performed. The vector form limits the capability of a deep learning model, and some more complex models cannot be used, so that the capability of a model recalled by a recommendation system is limited, and the accuracy of the recall of the recommendation system is reduced; meanwhile, the linear characteristic vector cannot be calculated by the current deep learning model for recommendation system recall, so that the recommendation system recall efficiency is reduced.
Disclosure of Invention
The embodiment of the invention provides a recommendation system recall method, which is used for improving the accuracy and the working efficiency of recommendation system recall and comprises the following steps:
discretizing and vectorizing the user features to obtain a plurality of user discrete feature vectors of the user features; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics;
discretizing and vectorizing the project characteristics to obtain a plurality of project discrete characteristic vectors of the project characteristics; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics;
performing dot product multiplication on the output total vector of the user characteristic and the output total vector of the project characteristic to obtain a calculation model of the matching degree between the user characteristic and the project characteristic;
inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model to obtain the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item;
and performing recall operation according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item.
The embodiment of the invention also provides a device for recalling the recommendation system, which is used for improving the accuracy and the working efficiency of the recall of the recommendation system, and the device comprises:
the user characteristic processing module is used for carrying out discretization and vectorization processing on the user characteristics to obtain a plurality of user discrete characteristic vectors of the user characteristics; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics;
the project feature processing module is used for carrying out discretization and vectorization processing on the project features to obtain a plurality of project discrete feature vectors of the project features; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics;
the calculation model modeling module of the matching degree between the user characteristics and the project characteristics is used for multiplying the output total vector of the user characteristics and the output total vector of the project characteristics by dot product to obtain a calculation model of the matching degree between the user characteristics and the project characteristics;
the matching degree calculation module is used for inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model to obtain the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item;
and the recommendation system recall module is used for recalling operation according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item.
The embodiment of the invention also provides computer equipment which comprises a memory, a processor and a computer program which is stored on the memory and can run on the processor, wherein the processor executes the computer program to realize the method for recalling the recommendation system.
An embodiment of the present invention further provides a computer-readable storage medium, where a computer program for executing the method for recalling a recommendation system is stored in the computer-readable storage medium.
In the embodiment of the invention, discretization and vectorization processing are carried out on the user characteristics to obtain a plurality of user discrete characteristic vectors of the user characteristics; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics; discretizing and vectorizing the project characteristics to obtain a plurality of project discrete characteristic vectors of the project characteristics; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics; performing dot product multiplication on the output total vector of the user characteristic and the output total vector of the project characteristic to obtain a calculation model of the matching degree between the user characteristic and the project characteristic; inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model to obtain the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item; according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item, the recall operation is carried out, so that the accurate calculation of the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item can be realized by establishing a calculation model of the matching degree between the user characteristics and the item characteristics, the calculation of linear discrete feature vectors can be realized by determining linear output vectors, the problem that the recall capability of a recommendation system is limited because the traditional deep learning model can only carry out vector form calculation in the prior art is solved, the explicit second-order crossing of the discrete feature vectors can be realized through the linear output vectors, and the recall accuracy of the recommendation system is improved; meanwhile, by means of linear output vectors, the recommendation system recall has memory capacity brought by feature cross combination based on explicit type, and recommendation system recall efficiency is improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts. In the drawings:
FIG. 1 is a flowchart illustrating a method for recommending system recalls in an embodiment of the present invention;
FIG. 2 is a diagram illustrating an exemplary method for recommending system recalls in an embodiment of the present invention;
FIG. 3 is a schematic diagram of an apparatus for recommending system recalls in an embodiment of the present invention;
FIG. 4 is a diagram illustrating an exemplary embodiment of an apparatus for recommending system recalls;
FIG. 5 is a diagram illustrating an exemplary computer apparatus for recommending system recalls in accordance with an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings. The exemplary embodiments and descriptions of the present invention are provided to explain the present invention, but not to limit the present invention.
At the present stage, with social development and popularization of the internet, the information quantity is increasing more and more quickly, people find information and change into information finding people, and a recommendation system is intelligent and personalized, useful information is extracted from mass information and is actively pushed to people.
The recommendation system needs to find out matching information from mass data information for personalized pushing, and generally comprises two links of recall and sequencing in consideration of time overhead, wherein the recall link is responsible for quickly finding n pieces of information which are possibly matched from the mass information. Vector retrieval is one way to recall links.
Currently, vector retrieval requires that User and Item be respectively characterized as vectors, but the form limits the use of more complex deep learning models in the recall link. In the sequencing link of the recommendation system, the deep FM model is widely used, and the deep learning model and the FM model are combined, so that the deep learning model has the generalization capability of deep learning and the memory capability of a linear model, and the model performance is enhanced. However, since the recall link of the recommendation system limits the output structure to be a vector, the model cannot be directly used for the recall link of the recommendation system. The deep learning model currently used for recommendation system recalls is typically a two-tower architecture that outputs the form of vectors, which are then subjected to vector retrieval. The vector form limits the capability of deep learning models, and can not use more complex models (such as the DeepFM model which is commonly used in the sequencing link of the recommendation system, can not be used for recalling by the recommendation system), so the capability of the model recalled by the recommendation system is limited. The existing scheme only uses a deep learning model and has strong generalization capability, but the structure cannot be combined with a linear model to strengthen the memory capability of the linear model.
In summary, since the deep learning model currently used for recommendation system recall is generally a double tower architecture, the output form of the double tower architecture is a vector, and then vector retrieval is performed. The vector form limits the capability of a deep learning model, and some more complex models cannot be used, so that the capability of a model recalled by a recommendation system is limited, and the accuracy of the recall of the recommendation system is reduced; meanwhile, the linear characteristic vector cannot be calculated by the current deep learning model for recommendation system recall, so that the recommendation system recall efficiency is reduced.
In order to solve the above problem, an embodiment of the present invention provides a recommendation system recall method for improving accuracy and work efficiency of recommendation system recall, where as shown in fig. 1, the method may include:
step 101: discretizing and vectorizing the user features to obtain a plurality of user discrete feature vectors of the user features; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics;
step 102: discretizing and vectorizing the project characteristics to obtain a plurality of project discrete characteristic vectors of the project characteristics; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics;
step 103: performing dot product multiplication on the output total vector of the user characteristic and the output total vector of the project characteristic to obtain a calculation model of the matching degree between the user characteristic and the project characteristic;
step 104: inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model to obtain the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item;
step 105: and performing recall operation according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item.
In the embodiment of the invention, discretization and vectorization processing are carried out on the user characteristics to obtain a plurality of user discrete characteristic vectors of the user characteristics; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics; discretizing and vectorizing the project characteristics to obtain a plurality of project discrete characteristic vectors of the project characteristics; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics; performing dot product multiplication on the output total vector of the user characteristic and the output total vector of the project characteristic to obtain a calculation model of the matching degree between the user characteristic and the project characteristic; inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model to obtain the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item; according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item, the recall operation is carried out, so that the accurate calculation of the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item can be realized by establishing a calculation model of the matching degree between the user characteristics and the item characteristics, the calculation of linear discrete feature vectors can be realized by determining linear output vectors, the problem that the recall capability of a recommendation system is limited because the traditional deep learning model can only carry out vector form calculation in the prior art is solved, the explicit second-order crossing of the discrete feature vectors can be realized through the linear output vectors, and the recall accuracy of the recommendation system is improved; meanwhile, by means of linear output vectors, the recommendation system recall has memory capacity brought by feature cross combination based on explicit type, and recommendation system recall efficiency is improved.
When the method is specifically implemented, firstly, discretization and vectorization processing are carried out on user characteristics to obtain a plurality of user discrete characteristic vectors of the user characteristics; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; and carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics.
In one embodiment, the vector accumulation may be performed on a plurality of user discrete feature vectors of the user features according to the following formula to obtain a linear output vector of the user features:
Figure BDA0003266510220000061
wherein, UFMA linear output vector representing a user feature; u shapeiRepresenting the ith user discrete feature vector; m represents the number of user discrete feature vectors.
In one embodiment, the deep learning output vector and the linear output vector of the user feature may be vector-spliced according to the following formula to obtain an output total vector of the user feature:
Figure BDA0003266510220000071
wherein, UFMA linear output vector representing a user feature; u shapeDeepA deep learning output vector representing a user feature; u represents the output total vector of the user features;
Figure BDA0003266510220000072
representing a vector stitching operation.
In the embodiment, the linear output vector of the user characteristics is determined, so that the calculation of the linear user discrete characteristic vector can be realized, the problem that the recall capability of the recommendation system is limited due to the fact that the traditional deep learning model can only perform vector form calculation in the prior art is solved, the purpose of performing explicit second-order intersection on the user discrete characteristic vector of each characteristic is realized through the linear output vector of the user characteristics, and the recall accuracy of the recommendation system is improved; meanwhile, the recall of the recommendation system has the memory capacity brought by the cross combination of the features based on the display through the linear output vector of the user features, and the recall efficiency of the recommendation system is improved.
In specific implementation, discretization and vectorization processing are carried out on the user characteristics to obtain a plurality of user discrete characteristic vectors of the user characteristics; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics, and then carrying out discretization and vectorization on the project characteristics to obtain a plurality of project discrete characteristic vectors of the project characteristics; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; and carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics.
In one embodiment, a plurality of item discrete feature vectors of item features may be vector accumulated to obtain a linear output vector of item features according to the following formula:
Figure BDA0003266510220000073
wherein, IFMA linear output vector representing a feature of the item; i isjRepresenting a j-th item discrete feature vector; n represents the number of discrete feature vectors of the item.
In one embodiment, the deep learning output vector and the linear output vector of the project feature may be vector-spliced to obtain an output total vector of the project feature according to the following formula:
Figure BDA0003266510220000074
wherein, IFMA linear output vector representing a feature of the item; i isDeepA deep learning output vector representing a feature of the project; i respectively represents the output total vector of the project characteristics;
Figure BDA0003266510220000075
representing a vector stitching operation.
In the embodiment, the linear output vector of the project features is determined, so that the calculation of the linear project discrete feature vector can be realized, the problem that the recall capability of the recommendation system is limited due to the fact that the traditional deep learning model can only perform vector form calculation in the prior art is solved, the purpose of performing explicit second-order intersection on the project discrete feature vector of each feature is realized through the linear output vector of the project features, and the recall accuracy of the recommendation system is improved; meanwhile, by means of the linear output vector of the project features, the recall of the recommendation system can have the memory capacity brought by the feature cross combination based on the display, and the recall efficiency of the recommendation system is improved.
In specific implementation, discretization and vectorization processing are carried out on the project characteristics to obtain a plurality of project discrete characteristic vectors of the project characteristics; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; and carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics, and carrying out dot product multiplication on the output total vector of the user characteristics and the output total vector of the project characteristics to obtain a calculation model of the matching degree between the user characteristics and the project characteristics.
In one embodiment, the calculation model of the matching degree between the user feature and the item feature may be obtained by multiplying the total output vector of the user feature and the total output vector of the item feature by a dot product according to the following formula:
Figure BDA0003266510220000081
wherein, the Score represents a calculation model of the matching degree between the user characteristic and the item characteristic; u shapeiAnd IiRespectively representing the ith user discrete feature vector and the jth item discrete feature vector; m and n respectively represent the number of user discrete feature vectors and project discrete feature vectors;
Figure BDA0003266510220000082
wherein, UFMAnd IFMA linear output vector representing user features and a linear output vector representing item features, U, respectivelyDeepAnd IDeepA deep learning output vector representing a user feature and a deep learning output vector representing a project feature, respectively, U and I represent an output total vector of the user feature and an output total vector of the project feature, respectively,
Figure BDA0003266510220000083
representing a vector splicing operation;
Figure BDA0003266510220000084
Figure BDA0003266510220000085
wherein, UiAnd IjRespectively representing the ith user discrete feature vector and the jth item discrete feature vector.
In the above embodiment, the accurate calculation of the matching degree between the user characteristic of the recall target user and the item characteristic of the recall target item is realized by establishing a calculation model of the matching degree between the user characteristic and the item characteristic, and the splicing operation of the FM (factorization machine model, which is a machine learning model based on matrix decomposition, i.e., a linear model) accumulated summation vector, the deep learning output vector and the FM accumulated summation vector is obtained by the accumulated summation operation of each discrete characteristic, because the two steps realize the addition of the linear model FM model, and the combination of the deep learning model and the FM model is realized. According to the calculation model of the matching degree between the user characteristics and the project characteristics, the linear M model is additionally added, and explicit second-order crossing of the characteristics can be performed, so that the model not only has the generalization capability of a deep learning model, but also has the memory capability brought by explicit characteristic combination and crossing, and the performance of the model is improved.
In specific implementation, after the output total vectors of the user features and the output total vectors of the item features are multiplied by dot products to obtain a calculation model of the matching degree between the user features and the item features, the user features of the recall target user and the item features of the recall target item are input into the calculation model to obtain the matching degree between the user features of the recall target user and the item features of the recall target item.
In the above-described embodiment, by using the user characteristics of the recall target user and the item characteristics of the recall target item, the calculation of the degree of matching between the user characteristics of the recall target user and the item characteristics of the recall target item can be realized, which facilitates the recall operation in the subsequent steps.
In specific implementation, after the user characteristics of the recall target user and the item characteristics of the recall target item are input into the calculation model to obtain the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item, the recall operation is performed according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item.
In the above embodiment, the combination model of the deep learning model and the linear model FM model is used in the recall link of the recommendation system, where the FM model can perform explicit second-order intersection of each feature, so that the model has not only the generalization capability of the deep learning model, but also the memory capability brought by explicit-based feature cross combination, thereby improving the performance of the model, finally improving the recall accuracy of the recommendation system, and also improving the recall efficiency of the recommendation system.
In specific implementation, the method for recalling a recommendation system provided by the embodiment of the present invention may further include:
according to the cross entropy loss function, correcting the matching degree calculation model of the user characteristic and the project characteristic to obtain a corrected matching degree calculation model of the user characteristic and the project characteristic;
inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model, including:
and inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the corrected calculation model of the matching degree between the user characteristics and the item characteristics.
In an embodiment, the cross-entropy loss function may be a softmax cross-entropy loss function.
In the embodiment, the cross entropy loss function is used for correcting the calculation model of the matching degree of the user characteristics and the project characteristics, so that the training of the calculation model of the matching degree between the user characteristics and the project characteristics can be realized, the accuracy of the calculation model of the matching degree between the user characteristics and the project characteristics is improved, and the recall accuracy and the recall efficiency of the recommendation system can be indirectly improved.
In specific implementation, the method for recalling a recommendation system provided by the embodiment of the present invention may further include:
and calculating and sending loss of the calculation model of the matching degree between the user characteristics and the project characteristics.
In the embodiment, the loss of the calculation model of the matching degree between the user characteristics and the project characteristics is calculated and sent, so that the calculation of the model precision is realized by the staff according to the loss of the model.
A specific embodiment is given below, and a specific application of the method of the present invention is described with reference to fig. 2, where the method in this embodiment may include the following steps:
1. discretizing the User-side features (i.e., the User features described above) and the Item-side features (i.e., the Item features described above), and then assigning a vector (i.e., DenseEmbellings in FIG. 2) to the value of each discretized feature (i.e., the SparseFeatures in FIG. 2, i.e., the User discrete features and the Item discrete features described above), resulting in a plurality of User discrete feature vectors and a plurality of Item discrete feature vectors as described above.
2. Respectively splicing vectors corresponding to discretization features of a User side and an Item side to serve as the input of a multilayer fully-connected neural network, wherein the output vector of the multilayer fully-connected neural network serves as the output vector (DeepEmbodings) of the deep learning part of the model, namely UDeepAnd IDeep. The multilayer fully-connected neural network is a deep learning model;
3. meanwhile, vectors corresponding to all features of the User side and the Item side are respectively subjected to vector accumulation, and the accumulated result is used as an output vector (namely FMEmbellings in FIG. 2) of the FM part of the model, namely UFMAnd IFM：
Figure BDA0003266510220000101
Figure BDA0003266510220000102
Wherein, UiAnd IiVectors corresponding to the ith discretization features of the User side and the Item side respectively; m and n represent the number of discrete feature vectors on the User side and Item side, respectively.
4. Respectively splicing the output vectors (DeepEmbeddings) of the deep learning parts at the User side and the Item side and the linear output vector (FMEmbodings) of the FM part to obtain the total model output vectors at the User side and the Item side, namely the total model output vectors at the User side and the Item side
Figure BDA0003266510220000103
Figure BDA0003266510220000104
Wherein,
Figure BDA0003266510220000111
representing a vector stitching operation.
5. And multiplying the output total vectors of the models of the User and the Item by a dot product to obtain the matching degree score of the User and the Item. The Item side not only contains the discretization feature vector of the positive sample, but also contains the discretization feature vectors of a plurality of negative samples, and model training can be carried out by using a softmax cross entropy loss function.
The degree of match between User and Item is scored as
Figure BDA0003266510220000112
Wherein,
Figure BDA0003266510220000113
it is the FM model, i.e. the explicit second-order intersection, U, of the User side and the Item side is realized in vector multiplicationDeep·IDeepIs the result of the deep learning model. Therefore, the model not only has the generalization capability of a deep learning model, but also has the memory capability based on explicit feature combination and cross, thereby improving the modelThe performance of (c).
6. And (3) reasoning by using a trained model to obtain each User and a vector corresponding to each Item, and searching all the items by using the vector of a certain User to obtain TopK (top match) items.
In the above example, the vectors corresponding to all the features of the User side and the Item side can be accumulated respectively, and the accumulated result is used as the output vector of the FM part of the model; and splicing the output vectors of the deep learning parts of the User side and the Item side and the output vector of the FM part to obtain model output vectors of the User side and the Item side, so that the discretization of the characteristics of the User and the items is completed, the vectors are distributed to the discretized characteristics, and the construction of a fully-connected neural network is helpful for the cumulative summation operation of each discrete characteristic, the splicing operation of the deep learning output vectors and the FM cumulative summation vector, the multiplication of the User vector and the Item vector of a positive sample and a negative sample to calculate a loss function and optimize the model. The method is characterized in that the linear model FM model is added through the two steps of the accumulative summation operation of each discrete feature, the splicing operation of the deep learning output vector and the FM accumulative summation vector, and the deep learning model and the FM model are combined.
Of course, it is understood that other variations of the above detailed flow can be made, and all such variations are intended to fall within the scope of the present invention.
In the embodiment of the invention, discretization and vectorization processing are carried out on the user characteristics to obtain a plurality of user discrete characteristic vectors of the user characteristics; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics; discretizing and vectorizing the project characteristics to obtain a plurality of project discrete characteristic vectors of the project characteristics; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics; performing dot product multiplication on the output total vector of the user characteristic and the output total vector of the project characteristic to obtain a calculation model of the matching degree between the user characteristic and the project characteristic; inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model to obtain the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item; according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item, the recall operation is carried out, so that the accurate calculation of the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item can be realized by establishing a calculation model of the matching degree between the user characteristics and the item characteristics, the calculation of linear discrete feature vectors can be realized by determining linear output vectors, the problem that the recall capability of a recommendation system is limited because the traditional deep learning model can only carry out vector form calculation in the prior art is solved, the explicit second-order crossing of the discrete feature vectors can be realized through the linear output vectors, and the recall accuracy of the recommendation system is improved; meanwhile, by means of linear output vectors, the recommendation system recall has memory capacity brought by feature cross combination based on explicit type, and recommendation system recall efficiency is improved.
The embodiment of the invention also provides a device for recalling the recommendation system, which is described in the following embodiment. Because the principle of the device for solving the problems is similar to the method for recalling the recommendation system, the implementation of the device can be referred to the implementation of the method for recalling the recommendation system, and repeated details are not repeated.
An embodiment of the present invention further provides a recommendation system recall device, so as to improve accuracy and work efficiency of recommendation system recall, as shown in fig. 3, the recommendation system recall device may include:
the user characteristic processing module 01 is used for performing discretization and vectorization processing on the user characteristics to obtain a plurality of user discrete characteristic vectors of the user characteristics; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics;
the project feature processing module 02 is used for performing discretization and vectorization processing on the project features to obtain a plurality of project discrete feature vectors of the project features; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics;
the calculation model modeling module 03 for the matching degree between the user characteristics and the project characteristics is used for multiplying the output total vectors of the user characteristics and the output total vectors of the project characteristics by dot products to obtain a calculation model for the matching degree between the user characteristics and the project characteristics;
the matching degree calculation module 04 is used for inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model to obtain the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item;
and the recommendation system recalling module 05 is used for recalling operation according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item.
In one embodiment, the user characteristic processing module is specifically configured to:
vector accumulation is carried out on a plurality of user discrete feature vectors of the user features according to the following formula to obtain linear output vectors of the user features:
Figure BDA0003266510220000131
wherein, UFMA linear output vector representing a user feature; u shapeiRepresenting the ith user discrete feature vector; m represents the number of user discrete feature vectors.
In one embodiment, the item feature processing module is specifically configured to:
performing vector accumulation on a plurality of item discrete feature vectors of the item features according to the following formula to obtain linear output vectors of the item features:
Figure BDA0003266510220000132
wherein, IFMA linear output vector representing a feature of the item; i isjRepresenting a j-th item discrete feature vector; n represents the number of discrete feature vectors of the item.
In one embodiment, the user characteristic processing module is specifically configured to:
vector splicing is carried out on the deep learning output vector and the linear output vector of the user characteristics according to the following formula to obtain the output total vector of the user characteristics:
Figure BDA0003266510220000141
wherein, UFMA linear output vector representing a user feature; u shapeDeepA deep learning output vector representing a user feature; u represents the output total vector of the user features;
Figure BDA0003266510220000142
representing a vector stitching operation.
In one embodiment, the item feature processing module is specifically configured to:
carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristic according to the following formula to obtain an output total vector of the project characteristic:
Figure BDA0003266510220000143
wherein, IFMA linear output vector representing a feature of the item; i isDeepA deep learning output vector representing a feature of the project; i respectively represents the output total vector of the project characteristics;
Figure BDA0003266510220000144
representing a vector stitching operation.
In an embodiment, the computational model modeling module for matching between the user feature and the item feature is specifically configured to:
and multiplying the output total vector of the user characteristic and the output total vector of the project characteristic by a dot product according to the following formula to obtain a calculation model of the matching degree between the user characteristic and the project characteristic:
Figure BDA0003266510220000145
wherein, the Score represents a calculation model of the matching degree between the user characteristic and the item characteristic; u shapeiAnd IiRespectively representing the ith user discrete feature vector and the jth item discrete feature vector; m and n respectively represent the number of user discrete feature vectors and project discrete feature vectors;
Figure BDA0003266510220000146
wherein, UFMAnd IFMA linear output vector representing user features and a linear output vector representing item features, U, respectivelyDeepAnd IDeepA deep learning output vector representing a user feature and a deep learning output vector representing a project feature, respectively, U and I represent an output total vector of the user feature and an output total vector of the project feature, respectively,
Figure BDA0003266510220000147
representing a vector splicing operation;
Figure BDA0003266510220000148
Figure BDA0003266510220000149
wherein, UiAnd IjRespectively representing the ith user discrete feature vector and the jth item discrete feature vector.
In one embodiment, as shown in fig. 4, further comprising: a model modification module 06 for:
according to the cross entropy loss function, correcting the matching degree calculation model of the user characteristic and the project characteristic to obtain a corrected matching degree calculation model of the user characteristic and the project characteristic;
inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model, including:
and inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the corrected calculation model of the matching degree between the user characteristics and the item characteristics.
Based on the same inventive concept, fig. 5 is a schematic diagram of a computer device provided in the embodiment of the present invention, as shown in fig. 5, the computer device 50 includes a memory 501, a processor 502, and a computer program stored on the memory 503 and capable of running on the processor 502, and the processor 502 implements the method for recalling in the recommendation system when executing the computer program.
Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium storing a computer program for executing the method for recalling a recommendation system.
In the embodiment of the invention, discretization and vectorization processing are carried out on the user characteristics to obtain a plurality of user discrete characteristic vectors of the user characteristics; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics; discretizing and vectorizing the project characteristics to obtain a plurality of project discrete characteristic vectors of the project characteristics; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics; performing dot product multiplication on the output total vector of the user characteristic and the output total vector of the project characteristic to obtain a calculation model of the matching degree between the user characteristic and the project characteristic; inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model to obtain the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item; according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item, the recall operation is carried out, so that the accurate calculation of the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item can be realized by establishing a calculation model of the matching degree between the user characteristics and the item characteristics, the calculation of linear discrete feature vectors can be realized by determining linear output vectors, the problem that the recall capability of a recommendation system is limited because the traditional deep learning model can only carry out vector form calculation in the prior art is solved, the explicit second-order crossing of the discrete feature vectors can be realized through the linear output vectors, and the recall accuracy of the recommendation system is improved; meanwhile, by means of linear output vectors, the recommendation system recall has memory capacity brought by feature cross combination based on explicit type, and recommendation system recall efficiency is improved.
As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The above-mentioned embodiments are intended to illustrate the objects, technical solutions and advantages of the present invention in further detail, and it should be understood that the above-mentioned embodiments are only exemplary embodiments of the present invention, and are not intended to limit the scope of the present invention, and any modifications, equivalent substitutions, improvements and the like made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (16)

1. A method of recommending system recalls, comprising:
discretizing and vectorizing the user features to obtain a plurality of user discrete feature vectors of the user features; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics;
discretizing and vectorizing the project characteristics to obtain a plurality of project discrete characteristic vectors of the project characteristics; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics;
performing dot product multiplication on the output total vector of the user characteristic and the output total vector of the project characteristic to obtain a calculation model of the matching degree between the user characteristic and the project characteristic;
inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model to obtain the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item;
and performing recall operation according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item.
2. The method of claim 1, wherein the vector accumulation is performed on a plurality of user discrete eigenvectors of the user characteristic according to the following formula to obtain a linear output vector of the user characteristic:
Figure FDA0003266510210000011
wherein, UFMA linear output vector representing a user feature; u shapeiRepresenting the ith user discrete feature vector; m represents the number of user discrete feature vectors.
3. The method of claim 1, wherein the vector accumulation is performed on a plurality of item discrete feature vectors for the item feature to obtain a linear output vector for the item feature according to the following equation:
Figure FDA0003266510210000012
wherein, IFMA linear output vector representing a feature of the item; i isjRepresenting a j-th item discrete feature vector; n represents the number of discrete feature vectors of the item.
4. The method of claim 1, wherein the deep learning output vector and the linear output vector of the user feature are vector-spliced to obtain an output total vector of the user feature according to the following formula:
Figure FDA0003266510210000021
wherein, UFMA linear output vector representing a user feature; u shapeDeepA deep learning output vector representing a user feature; u represents the input of user characteristicsObtaining a total vector;
Figure FDA0003266510210000022
representing a vector stitching operation.
5. The method of claim 1, wherein the deep learning output vector and the linear output vector of the project feature are vector-spliced to obtain an output total vector of the project feature according to the following formula:
Figure FDA0003266510210000023
wherein, IFMA linear output vector representing a feature of the item; i isDeepA deep learning output vector representing a feature of the project; i respectively represents the output total vector of the project characteristics;
Figure FDA0003266510210000024
representing a vector stitching operation.
6. The method of claim 1, wherein the output total vectors of the user features and the output total vectors of the item features are multiplied by a dot product according to the following formula to obtain a calculation model of the matching degree between the user features and the item features:
Figure FDA0003266510210000025
wherein, the Score represents a calculation model of the matching degree between the user characteristic and the item characteristic; u shapeiAnd IiRespectively representing the ith user discrete feature vector and the jth item discrete feature vector; m and n respectively represent the number of user discrete feature vectors and project discrete feature vectors;
Figure FDA0003266510210000026
wherein, UFMAnd IFMA linear output vector representing user features and a linear output vector representing item features, U, respectivelyDeepAnd IDeepA deep learning output vector representing a user feature and a deep learning output vector representing a project feature, respectively, U and I represent an output total vector of the user feature and an output total vector of the project feature, respectively,
Figure FDA0003266510210000027
representing a vector splicing operation;
Figure FDA0003266510210000028
Figure FDA0003266510210000029
wherein, UiAnd IjRespectively representing the ith user discrete feature vector and the jth item discrete feature vector.
7. The method of claim 1, further comprising:
according to the cross entropy loss function, correcting the matching degree calculation model of the user characteristic and the project characteristic to obtain a corrected matching degree calculation model of the user characteristic and the project characteristic;
inputting user characteristics of a recall target user and item characteristics of a recall target item into the computational model, including:
and inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the corrected calculation model of the matching degree between the user characteristics and the item characteristics.
8. An apparatus for recommending system recalls, comprising:
the user characteristic processing module is used for carrying out discretization and vectorization processing on the user characteristics to obtain a plurality of user discrete characteristic vectors of the user characteristics; inputting a plurality of user discrete feature vectors of user features into a multilayer fully-connected neural network to obtain deep learning output vectors of the user features; performing vector accumulation on a plurality of user discrete feature vectors of the user features to obtain linear output vectors of the user features; carrying out vector splicing on the deep learning output vector and the linear output vector of the user characteristics to obtain an output total vector of the user characteristics;
the project feature processing module is used for carrying out discretization and vectorization processing on the project features to obtain a plurality of project discrete feature vectors of the project features; inputting a plurality of project discrete feature vectors of the project features into a multilayer fully-connected neural network to obtain deep learning output vectors of the project features; performing vector accumulation on a plurality of project discrete feature vectors of the project features to obtain linear output vectors of the project features; carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristics to obtain an output total vector of the project characteristics;
the calculation model modeling module of the matching degree between the user characteristics and the project characteristics is used for multiplying the output total vector of the user characteristics and the output total vector of the project characteristics by dot product to obtain a calculation model of the matching degree between the user characteristics and the project characteristics;
the matching degree calculation module is used for inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the calculation model to obtain the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item;
and the recommendation system recall module is used for recalling operation according to the matching degree between the user characteristics of the recall target user and the item characteristics of the recall target item.
9. The apparatus of claim 8, wherein the user characteristic processing module is specifically configured to:
vector accumulation is carried out on a plurality of user discrete feature vectors of the user features according to the following formula to obtain linear output vectors of the user features:
Figure FDA0003266510210000031
wherein, UFMA linear output vector representing a user feature; u shapeiRepresenting the ith user discrete feature vector; m represents the number of user discrete feature vectors.
10. The apparatus of claim 8, wherein the item feature processing module is specifically configured to:
performing vector accumulation on a plurality of item discrete feature vectors of the item features according to the following formula to obtain linear output vectors of the item features:
Figure FDA0003266510210000041
wherein, IFMA linear output vector representing a feature of the item; i isjRepresenting a j-th item discrete feature vector; n represents the number of discrete feature vectors of the item.
11. The apparatus of claim 8, wherein the user characteristic processing module is specifically configured to:
vector splicing is carried out on the deep learning output vector and the linear output vector of the user characteristics according to the following formula to obtain the output total vector of the user characteristics:
Figure FDA0003266510210000042
wherein, UFMA linear output vector representing a user feature; u shapeDeepA deep learning output vector representing a user feature; u represents the output total vector of the user features;
Figure FDA0003266510210000043
representing a vector stitching operation.
12. The apparatus of claim 8, wherein the item feature processing module is specifically configured to:
carrying out vector splicing on the deep learning output vector and the linear output vector of the project characteristic according to the following formula to obtain an output total vector of the project characteristic:
Figure FDA0003266510210000044
wherein, IFMA linear output vector representing a feature of the item; i isDeepA deep learning output vector representing a feature of the project; i respectively represents the output total vector of the project characteristics;
Figure FDA0003266510210000045
representing a vector stitching operation.
13. The apparatus of claim 8, wherein the computational model modeling module of the degree of match between the user characteristic and the item characteristic is specifically configured to:
and multiplying the output total vector of the user characteristic and the output total vector of the project characteristic by a dot product according to the following formula to obtain a calculation model of the matching degree between the user characteristic and the project characteristic:
Figure FDA0003266510210000046
wherein, the Score represents a calculation model of the matching degree between the user characteristic and the item characteristic; u shapeiAnd IiRespectively representing the ith user discrete feature vector and the jth item discrete feature vector; m and n respectively represent the number of user discrete feature vectors and project discrete feature vectors;
Figure FDA0003266510210000047
wherein, UFMAnd IFMA linear output vector representing user features and a linear output vector representing item features, U, respectivelyDeepAnd IDeepDepthology separately representing user featuresLearning the output vector and the deep learning output vector of the project characteristic, wherein U and I respectively represent the output total vector of the user characteristic and the output total vector of the project characteristic,
Figure FDA0003266510210000051
representing a vector splicing operation;
Figure FDA0003266510210000052
Figure FDA0003266510210000053
wherein, UiAnd IjRespectively representing the ith user discrete feature vector and the jth item discrete feature vector.
14. The apparatus of claim 8, further comprising: a model modification module to:
according to the cross entropy loss function, correcting the matching degree calculation model of the user characteristic and the project characteristic to obtain a corrected matching degree calculation model of the user characteristic and the project characteristic;
inputting user characteristics of a recall target user and item characteristics of a recall target item into the computational model, including:
and inputting the user characteristics of the recall target user and the item characteristics of the recall target item into the corrected calculation model of the matching degree between the user characteristics and the item characteristics.
15. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any one of claims 1 to 7 when executing the computer program.
16. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program for executing the method of any one of claims 1 to 7.
CN202111088806.7A 2021-09-16 2021-09-16 Recommendation system recall method and device Active CN113742594B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111088806.7A CN113742594B (en) 2021-09-16 2021-09-16 Recommendation system recall method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111088806.7A CN113742594B (en) 2021-09-16 2021-09-16 Recommendation system recall method and device

Publications (2)

Publication Number Publication Date
CN113742594A true CN113742594A (en) 2021-12-03
CN113742594B CN113742594B (en) 2024-02-27

Family

ID=78739403

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111088806.7A Active CN113742594B (en) 2021-09-16 2021-09-16 Recommendation system recall method and device

Country Status (1)

Country Link
CN (1) CN113742594B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110866191A (en) * 2019-11-21 2020-03-06 苏州朗动网络科技有限公司 Recommendation recall method, apparatus and storage medium
CN111062775A (en) * 2019-12-03 2020-04-24 中山大学 A Recall Method for Recommendation System Based on Attention Mechanism
CN111127145A (en) * 2019-12-17 2020-05-08 武汉海云健康科技股份有限公司 Sorting recommendation method and system based on combination of catboost algorithm and deep learning
CN112541130A (en) * 2020-12-07 2021-03-23 东北大学 Deep cross feature fusion based recommendation method and device

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110866191A (en) * 2019-11-21 2020-03-06 苏州朗动网络科技有限公司 Recommendation recall method, apparatus and storage medium
CN111062775A (en) * 2019-12-03 2020-04-24 中山大学 A Recall Method for Recommendation System Based on Attention Mechanism
CN111127145A (en) * 2019-12-17 2020-05-08 武汉海云健康科技股份有限公司 Sorting recommendation method and system based on combination of catboost algorithm and deep learning
CN112541130A (en) * 2020-12-07 2021-03-23 东北大学 Deep cross feature fusion based recommendation method and device

Also Published As

Publication number Publication date
CN113742594B (en) 2024-02-27

Similar Documents

Publication Publication Date Title
Das et al. A group incremental feature selection for classification using rough set theory based genetic algorithm
CN110795527B (en) Candidate entity ordering method, training method and related device
Lyu et al. Resource-constrained neural architecture search on edge devices
CN111127246A (en) An Intelligent Prediction Method of Transmission Line Engineering Cost
CN109614495B (en) Related company mining method combining knowledge graph and text information
CN111160191A (en) Video key frame extraction method and device and storage medium
CN111581545A (en) Method for sorting recalled documents and related equipment
JP7417679B2 (en) Information extraction methods, devices, electronic devices and storage media
CN112905801A (en) Event map-based travel prediction method, system, device and storage medium
CN116719520B (en) Code generation method and device
CN113641819A (en) Debate mining system and method based on multi-task sparse shared learning
CN112380421A (en) Resume searching method and device, electronic equipment and computer storage medium
CN112084307A (en) Data processing method and device, server and computer readable storage medium
CN113590811A (en) Text abstract generation method and device, electronic equipment and storage medium
US12217185B2 (en) Method and apparatus of increasing knowledge based on uncertainty in neural networks
CN111523040A (en) Social contact recommendation method based on heterogeneous information network
Syah et al. Sensitivity of shortest distance search in the ant colony algorithm with varying normalized distance formulas
CN106776782B (en) Semantic similarity obtaining method and device based on artificial intelligence
CN114065769A (en) Training method, device, equipment and medium for emotional reasons to extraction model
CN117829149B (en) Language model hybrid training method and device, electronic equipment and storage medium
CN113742594B (en) Recommendation system recall method and device
CN117112773B (en) Method and device for searching navigable unstructured data based on NLP
CN118553332A (en) Method and device for recommending oxygen consumption of converter steelmaking based on ensemble learning algorithm
CN110083674B (en) Intellectual property information processing method and device
CN111813941A (en) Text classification method, device, device and medium combining RPA and AI

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