CN117033948A - Project recommendation method based on feature interaction information and time tensor decomposition - Google Patents
Project recommendation method based on feature interaction information and time tensor decomposition Download PDFInfo
- Publication number
- CN117033948A CN117033948A CN202311288046.3A CN202311288046A CN117033948A CN 117033948 A CN117033948 A CN 117033948A CN 202311288046 A CN202311288046 A CN 202311288046A CN 117033948 A CN117033948 A CN 117033948A
- Authority
- CN
- China
- Prior art keywords
- layer
- deep
- module
- user
- time
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 230000003993 interaction Effects 0.000 title claims abstract description 53
- 238000000354 decomposition reaction Methods 0.000 title claims abstract description 32
- 238000000034 method Methods 0.000 title claims abstract description 27
- 239000011159 matrix material Substances 0.000 claims abstract description 83
- 239000013598 vector Substances 0.000 claims description 105
- 230000007704 transition Effects 0.000 claims description 26
- 230000006870 function Effects 0.000 claims description 14
- 230000002123 temporal effect Effects 0.000 claims description 10
- 238000013528 artificial neural network Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 2
- 238000011478 gradient descent method Methods 0.000 claims description 2
- 239000000284 extract Substances 0.000 abstract description 4
- 230000007246 mechanism Effects 0.000 abstract description 4
- 238000000605 extraction Methods 0.000 description 4
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 238000005065 mining Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
-
- 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
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- 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
- G06N3/045—Combinations of networks
-
- 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
Abstract
The invention provides a project recommending method based on characteristic interaction information and time tensor decomposition, which is applied to a project recommending model, wherein the project recommending model comprises a characteristic coding layer, a characteristic extracting layer and a learning predicting layer, the characteristic extracting layer comprises a deep TF module and a deep ATF module, and the learning predicting layer comprises a GMF-TF module, a GMF-ATF module and a predicting module; the invention extracts the characteristics of users, projects and time through a multi-layer attention mechanism, acquires rich characteristic representation, introduces generalized matrix decomposition (GMF) into the model to generate second-order characteristic interaction information of the users, the projects, the users, the time and the projects and the time, so that the model can observe potential connection among the users, the projects and the time, and the overall performance of the model is improved.
Description
Technical Field
The invention relates to the technical field of data processing, in particular to a project recommendation method based on characteristic interaction information and time tensor decomposition.
Background
In the recommendation system, matrix factorization (Matrix Factorization, MF) is a representative algorithm for collaborative filtering, which decomposes a user-item matrix into two potential factor matrices to represent users and items, but when the user-item matrix is very sparse, the decomposition can be very unstable, often resulting in low quality local optima. To improve this, the relevant technicians have done much work, but these methods often require a lot of manpower to screen the features and have poor results because the inner product of the matrix decomposition is too linearized, so subsequent researchers have made efforts to improve the matrix decomposition.
On one hand, the related technicians introduce deep learning into matrix decomposition to promote recommendation effects. For example, using a multi-layer perceptron (Multilayer Perceptron, MLP) to extract the user's preference characteristics for items and combine the simple user preferences obtained by dot product; or reallocate the duty cycle to the user's preferences using a multi-layer attention mechanism. However, although these methods extract a relatively rich user item feature by various techniques, they only focus on the overall preference of the user, and do not consider the situation that the user preference changes with time. When scenes of different times are involved, the user interaction with the item cannot reflect the dynamics generated by the time, and recommendation deviation is caused.
On the other hand, the relevant technician tries to extend a two-dimensional matrix representing the user's interactions with the items into a three-dimensional tensor containing time information, and then, the user and the items can be projected into a potential space with time coding using tensor decomposition (Tensor Factorization, TF) techniques, where there are two opposite assumptions, i.e., adjacent points in time are independent or interrelated. However, whether independent or continuous interaction, the method directly inputs the obtained user, item and time feature vector into the multi-layer perceptron to predict the rating, and the problem of information loss is brought to the model to a certain extent due to insufficient mining of potential interaction relation between the user and the item.
Disclosure of Invention
Therefore, the embodiment of the invention provides an item recommendation method based on characteristic interaction information and time tensor decomposition, which aims to solve the problems that the prior art ignores the situation that the user preference changes along with time change and potential interaction relation mining is insufficient.
According to the project recommendation method based on the feature interaction information and the time tensor decomposition, which is provided by the embodiment of the invention, the project recommendation method is applied to a project recommendation model, wherein the project recommendation model comprises a feature coding layer, a feature extraction layer and a learning prediction layer, the feature extraction layer comprises a deep TF module and a deep ATF module, and the learning prediction layer comprises a GMF-TF module, a GMF-ATF module and a prediction module;
the method comprises the following steps:
step 1, performing one-hot coding on historical interaction data of a user at a feature coding layer to obtain one-hot coding of a user number, a project number and a time number, and multiplying the one-hot coding of the user number, the project number and the time number with a potential factor matrix respectively to obtain a user potential feature vector, a project potential feature vector and a time potential feature vector;
step 2, inputting the user potential feature vector, the project potential feature vector and the time potential feature vector into a deep TF module, obtaining user depth features, project depth features and time depth features through multi-layer perceptron operation, and simultaneously inputting the user potential feature vector, the project potential feature vector and the time potential feature vector into a deep ATF module, wherein in the deep ATF module, the output of each layer of neural network is input into the attention network of the next layer, so as to obtain multi-layer attention user feature output of the deep ATF module, multi-layer attention project feature output of the deep ATF module and multi-layer attention time feature output of the deep ATF module;
step 3, inputting the user depth feature, the project depth feature and the time depth feature into a GMF-TF module, outputting the multi-layer attention user feature of the deep ATF module, outputting the multi-layer attention project feature of the deep ATF module and outputting the multi-layer attention time feature of the deep ATF module into the GMF-ATF module, obtaining second-order feature interaction information among the user, the project and the time through generalized matrix decomposition learning, and inputting the second-order feature interaction information into a prediction module to obtain a prediction score value of the user on the project;
and 4, calculating the loss between the predicted score value and the real score value through the objective function, minimizing the loss, optimizing the objective function parameter through back propagation, completing algorithm convergence, further obtaining a final predicted score value, and providing a project recommendation list for a user according to the final predicted score value.
According to the project recommendation method based on the feature interaction information and the time tensor decomposition, the method has the following beneficial effects:
1) According to the invention, time factors are integrated into the model, so that the model can observe the influence of time on user selection, and when a dynamic scene is involved, the interaction between the user and the project can reflect the dynamic property generated by time, so that the scoring of the user on each project at the current time can be predicted more accurately;
2) The invention extracts the characteristics of users, projects and time through a multi-layer attention mechanism, acquires rich characteristic representation, introduces generalized matrix decomposition (GMF) into the model to generate second-order characteristic interaction information of the users, the projects, the users, the time and the projects and the time, so that the model can observe potential connection among the users, the projects and the time, and the overall performance of the model is improved.
Drawings
The foregoing and/or additional aspects and advantages of embodiments of the invention will be apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:
FIG. 1 is a block diagram of the results of an item recommendation model in accordance with an embodiment of the present invention;
FIG. 2 is a flow chart of a project recommendation method based on feature interaction information and time tensor decomposition according to an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. 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.
The embodiment of the invention provides a project recommendation method based on feature interaction information and time tensor decomposition, which is applied to a project recommendation model, referring to fig. 1, wherein the project recommendation model comprises a feature coding layer, a feature extraction layer and a learning prediction layer, the feature extraction layer comprises a deep TF module and a deep ATF module, and the learning prediction layer comprises a GMF-TF module, a GMF-ATF module and a prediction module.
Referring to fig. 2, the method includes steps 1 to 4:
and step 1, performing one-hot coding on historical interaction data of a user at a feature coding layer to obtain one-hot coding of a user number, a project number and a time number, and multiplying the one-hot coding of the user number, the project number and the time number with a potential factor matrix respectively to obtain a user potential feature vector, a project potential feature vector and a time potential feature vector.
When the historical interaction data of the user is transmitted to the project recommendation model, the data is required to be embedded in a feature coding layer. The data is first one-hot encoded, which eliminates the order effects of conventional integer encoding, making each feature equal in computer representation. However, because the vector of the one-hot code is too high-dimensional sparse, computational resources are seriously consumed and model training efficiency is reduced, so that the one-hot code and a potential factor matrix are subjected to matrix multiplication to obtain the feature vector in a dense low-dimensional space.
Specifically, the one-hot codes of user number, project number and time number are respectively combined with the potential factor matrix、、/>Multiplying the three potential feature vectors to obtain the user potential feature vector +.>Item potential feature vector->Temporal latent feature vector->. Wherein (1)>、/>、/>Respectively representing the total number of users, items and time,Lis a potential feature dimension.
And 2, inputting the user potential feature vector, the project potential feature vector and the time potential feature vector into a deep TF module, and obtaining user depth features, project depth features and time depth features through multi-layer perceptron operation, and simultaneously inputting the user potential feature vector, the project potential feature vector and the time potential feature vector into a deep ATF module, wherein in the deep ATF module, the output of each layer of neural network is input into the attention network of the next layer, so as to obtain multi-layer attention user feature output of the deep ATF module, multi-layer attention project feature output of the deep ATF module and multi-layer attention time feature output of the deep ATF module.
In step 2, the user depth feature satisfies the following conditional expression:
；
；
；
；
wherein,user transition characteristics corresponding to the first layer in the user multi-layer perceptron of the deep TF module,/user transition characteristics corresponding to the first layer in the user multi-layer perceptron of the deep TF module>For the first activation function, +.>Weight matrix corresponding to the first layer in the multi-layer perceptron of the user as deep TF module>For the user potential feature vector, +.>Bias vector corresponding to the first layer in the user multi-layer perceptron of deep TF module,/for the user multi-layer perceptron of deep TF module>User transition characteristics corresponding to the n-1 th layer in the user multi-layer perceptron of the deep TF module,/L>Weight matrix corresponding to n-1 th layer in user multi-layer perceptron of deep TF module, < ->N-2 in user multi-layer perceptron as deep TF moduleUser transition feature for layer correspondence->Bias vector corresponding to n-1 th layer in user multi-layer perceptron of deep TF module,/-, and>weight matrix corresponding to the nth layer in the user multi-layer perceptron of deep TF module>Representing the user depth feature, in this embodiment user depth feature +.>Including information on the occupation, hobbies, age, etc. of the user.
In step 2, the project depth feature satisfies the following conditional expression:
；
；
；
；
wherein,item transition characteristics corresponding to the first layer in the item multi-layer perceptron of the deep TF module,/item transition characteristics corresponding to the first layer in the item multi-layer perceptron of the deep TF module>Weight matrix corresponding to the first layer in the item multi-layer perceptron of deep TF module,/I>For item potential feature vector, ++>Bias vector corresponding to the first layer in the item multi-layer perceptron of deep TF module,>item transition characteristics corresponding to the n-1 th layer in the item multi-layer perceptron of the deep TF module,/item transition characteristics>Weight matrix corresponding to n-1 th layer in item multi-layer perceptron of deep TF module, < ->Item transition characteristics corresponding to the n-2 th layer in the item multi-layer perceptron of the deep TF module,/item transition characteristics>Bias vector corresponding to n-1 th layer in item multi-layer perceptron of deep TF module,/->Weight matrix corresponding to the nth layer in the item multi-layer perceptron of deep TF module>Representing project depth features, in this embodiment project depth features +.>Including all category characteristics of the item.
In step 2, the temporal depth profile satisfies the following conditional expression:
；
；
；
；
wherein,time transition characteristics corresponding to the first layer in the time multi-layer perceptron of the deep TF module are +.>Weight matrix corresponding to the first layer in the time multi-layer perceptron of deep TF module>For temporal latent feature vector, ++>Bias vector corresponding to the first layer in the time multi-layer perceptron of deep TF module, ++>Time transition characteristics corresponding to the n-1 th layer in the time multi-layer perceptron of the deep TF module,/L>Weight matrix corresponding to n-1 th layer in time multi-layer perceptron of deep TF module, < ->Time transition characteristics corresponding to the n-2 th layer in the time multi-layer perceptron of the deep TF module,/L>Bias vector corresponding to n-1 th layer in time multi-layer perceptron of deep TF module,/-, and>weight matrix corresponding to the nth layer in the time multi-layer perceptron of deep TF module>Representing temporal depth features, in this embodiment temporal depth features +.>Contains the information of time, year, month, festival, special meaning, etc.
In this embodiment, the deep atf module is different from the deep tf module in that the deep atf module adds an attention layer in the multi-layer perceptron. The input of which is still the output of the feature encoding layer、/>、/>Vector. However, the output of each layer of neural network is not transmitted to the next layer of neural network, but is transmitted to the Attention network (Attention) layer, namely, the MLP-Attention two layers are continuously transmitted in a layered manner, so that the weight of corresponding input information is acquired through the Attention network, and better user, project and time characteristic information is obtained.
Specifically, in step 2, the multi-layer attention user feature output of the deep atf module satisfies the following conditional expression:
；
；
；
；
；
wherein,is the user characteristic output corresponding to the 1 st layer in the user multi-layer perceptron of the deep ATF module,/I>Is the weight matrix corresponding to the 1 st layer in the user multi-layer perceptron of the deep ATF module, and is->Is bias vector corresponding to layer 1 in the user multi-layer perceptron of deep ATF module,/>Is the user characteristic output of the layer 1 attention network in the user multi-layer perceptron of the deep ATF module,/or%>Is thatsoftmaxFunction (F)>Is the weight matrix of the layer 1 attention network in the user multi-layer perceptron of the deep ATF module,/and%>Is the bias vector of the layer 1 attention network in the user multi-layer perceptron of the deep atf module,is the user characteristic output corresponding to the nth layer in the user multi-layer perceptron of the deep ATF module,/I>Is the weight matrix corresponding to the n-th layer in the user multi-layer perceptron of the deep ATF module, and is->Is the user characteristic output of the n-1 layer attention network in the user multi-layer perceptron of the deep ATF module,/for the user>Is the bias vector corresponding to the n-th layer in the user multi-layer perceptron of the deep ATF module,/h>Is a user feature vector extracted through a multi-layer attention network,/->Is the weight matrix of the nth layer attention network in the user multi-layer perceptron of the deep ATF module,/and%>Is the bias vector of the n-th layer of attention network in the user multi-layer perceptron of the deep ATF module,/and>is the user characteristic output corresponding to the n-1 layer in the user multi-layer perceptron of the deep ATF module;
in step 2, the multi-layer attention item feature output of the deep atf module satisfies the following conditional expression:
；
；
；
；
；
wherein,is item characteristic output corresponding to the 1 st layer in the item multi-layer perceptron of the deep ATF module,/I>Is the weight matrix corresponding to the 1 st layer in the item multi-layer perceptron of the deep ATF module,/I>Is the bias vector corresponding to layer 1 in the item multi-layer perceptron of the deep ATF module,/>Item feature output of layer 1 attention network in item multi-layer perceptron of deep ATF module,/I>Is the weight matrix of the layer 1 attention network in the project multi-layer perceptron of the deep ATF module,/and%>Is the bias vector of the layer 1 attention network in the item multi-layer perceptron of the deep ATF module,/for the item multi-layer perceptron>Is the item characteristic output corresponding to the nth layer in the item multi-layer perceptron of the deep ATF module,/I>Is the weight matrix corresponding to the n-th layer in the item multi-layer perceptron of the deep ATF module,/I>Item feature output of the n-1 layer attention network in the item multi-layer perceptron of the deep ATF module,/item feature output of the n-1 layer attention network is added>Is the bias vector corresponding to the n-th layer in the item multi-layer perceptron of the deep ATF module,/I>Is an item extracted through a multi-layer attention networkFeature vector->Is the weight matrix of the nth layer attention network in the project multi-layer perceptron of the deep ATF module,/and%>Is the bias vector of the n-th layer attention network in the item multi-layer perceptron of the deep ATF module,/and>the item characteristic output corresponding to the n-1 th layer in the item multi-layer perceptron of the deep ATF module;
in step 2, the multi-layer attention time feature output of the deep atf module satisfies the following conditional expression:
；
；
；
；
；
wherein,is the time characteristic output corresponding to the 1 st layer in the time multi-layer perceptron of the deep ATF module,/I>Is the weight matrix corresponding to the 1 st layer in the time multi-layer perceptron of the deep ATF module, and is ∈1 st layer>Is the bias vector corresponding to layer 1 in the time multi-layer perceptron of the deep ATF module,/>Is the time characteristic output of the layer 1 attention network in the time multi-layer perceptron of the deep ATF module,/I>Is the weight matrix of the layer 1 attention network in the time multi-layer perceptron of the deep ATF module,/and%>Is the bias vector of the layer 1 attention network in the time multi-layer perceptron of the deep ATF module,/for the layer 1 attention network>Is the corresponding time characteristic output of the nth layer in the time multi-layer perceptron of the deep ATF module,/->Is the weight matrix corresponding to the n-th layer in the time multi-layer perceptron of the deep ATF module,/h>Is the time characteristic output of the n-1 layer attention network in the time multi-layer perceptron of the deep ATF module,/for the time multi-layer perceptron>Is the bias vector corresponding to the n-th layer in the time multi-layer perceptron of the deep ATF module,/->Is a temporal feature vector extracted through a multi-layer attention network,/->Is the weight matrix of the nth layer attention network in the time multi-layer perceptron of the deep ATF module,/and%>Is the bias vector of the n-th layer attention network in the time multi-layer perceptron of the deep ATF module,/and>is the corresponding time characteristic output of the n-1 layer in the time multi-layer perceptron of the deep ATF module.
And 3, inputting the user depth feature, the project depth feature and the time depth feature into a GMF-TF module, outputting the multi-layer attention user feature of the deep ATF module, outputting the multi-layer attention project feature of the deep ATF module and outputting the multi-layer attention time feature of the deep ATF module into the GMF-ATF module, obtaining second-order feature interaction information among the user, the project and the time through generalized matrix decomposition learning, and inputting the second-order feature interaction information into a prediction module to obtain a prediction score value of the user on the project.
In step 3, the user depth feature, the project depth feature and the time depth feature are input to the GMF-TF module, and second-order feature interaction information among the user, the project and the time is obtained through generalized matrix decomposition learning, wherein the expression is as follows:
；
；
；
wherein,for the second order feature interaction vector of the user and the project in the GMF-TF module,/for the user and the project>For the second order feature interaction vector of the user and time in the GMF-TF module,/for the user and time>For the second order feature interaction vector of the item and time in the GMF-TF module,/for the item and time>For the second activation function, +.>Is->Corresponding weight matrix, < >>Is->The corresponding weight matrix is used to determine the weight matrix,is->Corresponding weight matrix, < >>Is->Corresponding bias vector, ">Is->Corresponding bias vector, ">Is->Corresponding bias vector, ">Representing dot product.
In step 3, inputting the multi-layer attention user feature output of the deep ATF module, the multi-layer attention project feature output of the deep ATF module and the multi-layer attention time feature output of the deep ATF module to the GMF-ATF module, and obtaining second-order feature interaction information among the user, the project and the time through generalized matrix decomposition learning, wherein the expression is as follows:
；
；
；
wherein,for the second order feature interaction vector of the user and the item in the GMF-ATF module, +.>For the second order feature interaction vector of the user and time in the GMF-ATF module,/for the user and time>Is the second order feature interaction vector of items and time in the GMF-ATF module, +.>Is->Corresponding weight matrix, < >>Is->Corresponding weight matrix, < >>Is->Corresponding weight matrix, < >>Is->Corresponding bias vector, ">Is->Corresponding bias vector, ">Is->A corresponding bias vector.
In step 3, the predictive score value satisfies the following conditional expression:
；
；
；
；
；
wherein,is indicated at +.>When (4) user->Item->Is a predictive score value for (a); />、/>、/>、/>A weight matrix for the prediction module; />Indicating the number of hidden layers; />、/>、/>、/>For the bias vector of the prediction module, +.>、/>、/>、/>Is an intermediate value.
And 4, calculating the loss between the predicted score value and the real score value through the objective function, minimizing the loss, optimizing the objective function parameter through back propagation, completing algorithm convergence, further obtaining a final predicted score value, and providing a project recommendation list for a user according to the final predicted score value.
In step 4, the expression for minimizing the objective function loss is:
；
wherein,minimizing a function for losses by adopting an Adam gradient descent method; />Is indicated at +.>When the userItem->Is a true score value of (2);U、V、Ta potential factor matrix of user, project, time, respectively,>、/>、/>respectively representing the total number of users, items, and times.
In summary, the project recommendation method based on the characteristic interaction information and the time tensor decomposition provided by the invention has the following beneficial effects:
1) According to the invention, time factors are integrated into the model, so that the model can observe the influence of time on user selection, and when a dynamic scene is involved, the interaction between the user and the project can reflect the dynamic property generated by time, so that the scoring of the user on each project at the current time can be predicted more accurately;
2) The invention extracts the characteristics of users, projects and time through a multi-layer attention mechanism, acquires rich characteristic representation, introduces generalized matrix decomposition (GMF) into the model to generate second-order characteristic interaction information of the users, the projects, the users, the time and the projects and the time, so that the model can observe potential connection among the users, the projects and the time, and the overall performance of the model is improved.
In the description of the present specification, a description referring to terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiments or examples. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
While embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that: many changes, modifications, substitutions and variations may be made to the embodiments without departing from the spirit and principles of the invention, the scope of which is defined by the claims and their equivalents.
Claims (7)
1. The project recommending method based on the feature interaction information and the time tensor decomposition is characterized by being applied to a project recommending model, wherein the project recommending model comprises a feature coding layer, a feature extracting layer and a learning predicting layer, the feature extracting layer comprises a deep TF module and a deep ATF module, and the learning predicting layer comprises a GMF-TF module, a GMF-ATF module and a predicting module;
the method comprises the following steps:
step 1, performing one-hot coding on historical interaction data of a user at a feature coding layer to obtain one-hot coding of a user number, a project number and a time number, and multiplying the one-hot coding of the user number, the project number and the time number with a potential factor matrix respectively to obtain a user potential feature vector, a project potential feature vector and a time potential feature vector;
step 2, inputting the user potential feature vector, the project potential feature vector and the time potential feature vector into a deep TF module, obtaining user depth features, project depth features and time depth features through multi-layer perceptron operation, and simultaneously inputting the user potential feature vector, the project potential feature vector and the time potential feature vector into a deep ATF module, wherein in the deep ATF module, the output of each layer of neural network is input into the attention network of the next layer, so as to obtain multi-layer attention user feature output of the deep ATF module, multi-layer attention project feature output of the deep ATF module and multi-layer attention time feature output of the deep ATF module;
step 3, inputting the user depth feature, the project depth feature and the time depth feature into a GMF-TF module, outputting the multi-layer attention user feature of the deep ATF module, outputting the multi-layer attention project feature of the deep ATF module and outputting the multi-layer attention time feature of the deep ATF module into the GMF-ATF module, obtaining second-order feature interaction information among the user, the project and the time through generalized matrix decomposition learning, and inputting the second-order feature interaction information into a prediction module to obtain a prediction score value of the user on the project;
and 4, calculating the loss between the predicted score value and the real score value through the objective function, minimizing the loss, optimizing the objective function parameter through back propagation, completing algorithm convergence, further obtaining a final predicted score value, and providing a project recommendation list for a user according to the final predicted score value.
2. The item recommendation method based on feature interaction information and time tensor decomposition according to claim 1, wherein in step 2, the user depth feature satisfies the following conditional expression:
；
；
；
；
wherein,user transition characteristics corresponding to the first layer in the user multi-layer perceptron of the deep TF module,/user transition characteristics corresponding to the first layer in the user multi-layer perceptron of the deep TF module>For the first activation function, +.>Weight matrix corresponding to the first layer in the multi-layer perceptron of the user as deep TF module>For the user potential feature vector, +.>Bias vector corresponding to the first layer in the user multi-layer perceptron of deep TF module,/for the user multi-layer perceptron of deep TF module>User transition characteristics corresponding to the n-1 th layer in the user multi-layer perceptron of the deep TF module,/L>Weight matrix corresponding to n-1 th layer in user multi-layer perceptron of deep TF module, < ->User transition characteristics corresponding to the n-2 th layer in the user multi-layer perceptron of the deep TF module,/L>Bias vector corresponding to n-1 th layer in user multi-layer perceptron of deep TF module,/-, and>representing user depth features>The weight matrix corresponding to the nth layer in the multi-layer perceptron of the user of the deep TF module;
in step 2, the project depth feature satisfies the following conditional expression:
；
；
；
；
wherein,item transition characteristics corresponding to the first layer in the item multi-layer perceptron of the deep TF module,/item transition characteristics corresponding to the first layer in the item multi-layer perceptron of the deep TF module>Weight matrix corresponding to the first layer in the item multi-layer perceptron of deep TF module,/I>For item potential feature vector, ++>Bias vector corresponding to the first layer in the item multi-layer perceptron of deep TF module,>item transition characteristics corresponding to the n-1 th layer in the item multi-layer perceptron of the deep TF module,/item transition characteristics>Weight matrix corresponding to n-1 th layer in item multi-layer perceptron of deep TF module, < ->Item transition characteristics corresponding to the n-2 th layer in the item multi-layer perceptron of the deep TF module,/item transition characteristics>Bias vector corresponding to n-1 th layer in item multi-layer perceptron of deep TF module,/->Representing item depth features, < >>The weight matrix corresponding to the nth layer in the project multi-layer perceptron of the deep TF module;
in step 2, the temporal depth profile satisfies the following conditional expression:
；
；
；
；
wherein,time transition characteristics corresponding to the first layer in the time multi-layer perceptron of the deep TF module are +.>Weight matrix corresponding to the first layer in the time multi-layer perceptron of deep TF module>For temporal latent feature vector, ++>Bias vector corresponding to the first layer in the time multi-layer perceptron of deep TF module, ++>Time transition characteristics corresponding to the n-1 th layer in the time multi-layer perceptron of the deep TF module,/L>The weight matrix corresponding to the n-1 layer in the time multi-layer perceptron of the deep TF module,time transition characteristics corresponding to the n-2 th layer in the time multi-layer perceptron of the deep TF module,/L>Bias vector corresponding to n-1 th layer in time multi-layer perceptron of deep TF module,/-, and>representing temporal depth features>The weight matrix corresponding to the nth layer in the time multi-layer perceptron of the deep TF module.
3. The item recommendation method based on feature interaction information and time tensor decomposition according to claim 2, wherein in step 2, the multi-layer attention user feature output of the deep atf module satisfies the following conditional expression:
；
；
；
；
；
wherein,is the user characteristic output corresponding to the 1 st layer in the user multi-layer perceptron of the deep ATF module,/I>User multilayer feel for deep ATF moduleWeight matrix corresponding to layer 1 in knowing machine, < >>Is bias vector corresponding to layer 1 in the user multi-layer perceptron of deep ATF module,/>Is the user characteristic output of the layer 1 attention network in the user multi-layer perceptron of the deep ATF module,/or%>Is thatsoftmaxFunction (F)>Is the weight matrix of the layer 1 attention network in the user multi-layer perceptron of the deep ATF module,/and%>Is the bias vector of the layer 1 attention network in the user multi-layer perceptron of the deep atf module,is the user characteristic output corresponding to the nth layer in the user multi-layer perceptron of the deep ATF module,/I>Is the weight matrix corresponding to the n-th layer in the user multi-layer perceptron of the deep ATF module, and is->Is the user characteristic output of the n-1 layer attention network in the user multi-layer perceptron of the deep ATF module,/for the user>Is the bias vector corresponding to the n-th layer in the user multi-layer perceptron of the deep ATF module,/h>Is a user feature vector extracted through a multi-layer attention network,/->Is the weight matrix of the nth layer attention network in the user multi-layer perceptron of the deep ATF module,/and%>Is the bias vector of the n-th layer of attention network in the user multi-layer perceptron of the deep ATF module,/and>is the user characteristic output corresponding to the n-1 layer in the user multi-layer perceptron of the deep ATF module;
in step 2, the multi-layer attention item feature output of the deep atf module satisfies the following conditional expression:
；
；
；
；
；
wherein,is item characteristic output corresponding to the 1 st layer in the item multi-layer perceptron of the deep ATF module,/I>Is the weight matrix corresponding to the 1 st layer in the item multi-layer perceptron of the deep ATF module,/I>Is the bias vector corresponding to layer 1 in the item multi-layer perceptron of the deep ATF module,/>Item feature output of layer 1 attention network in item multi-layer perceptron of deep ATF module,/I>Is the weight matrix of the layer 1 attention network in the project multi-layer perceptron of the deep ATF module,is the bias vector of the layer 1 attention network in the item multi-layer perceptron of the deep ATF module,/for the item multi-layer perceptron>Is the item characteristic output corresponding to the nth layer in the item multi-layer perceptron of the deep ATF module,/I>Is the weight matrix corresponding to the n-th layer in the item multi-layer perceptron of the deep ATF module,/I>Item feature output of the n-1 layer attention network in the item multi-layer perceptron of the deep ATF module,/item feature output of the n-1 layer attention network is added>Is the bias vector corresponding to the n-th layer in the item multi-layer perceptron of the deep ATF module,/I>Is extracted through a multi-layer attention networkProject feature vector->Is the weight matrix of the nth layer attention network in the project multi-layer perceptron of the deep ATF module,/and%>Is the bias vector of the n-th layer attention network in the item multi-layer perceptron of the deep ATF module,/and>the item characteristic output corresponding to the n-1 th layer in the item multi-layer perceptron of the deep ATF module;
in step 2, the multi-layer attention time feature output of the deep atf module satisfies the following conditional expression:
；
；
；
；
；
wherein,is the time characteristic output corresponding to the 1 st layer in the time multi-layer perceptron of the deep ATF module,/I>Is the weight matrix corresponding to the 1 st layer in the time multi-layer perceptron of the deep ATF module, and is ∈1 st layer>Is the bias vector corresponding to layer 1 in the time multi-layer perceptron of the deep ATF module,/>Is the time characteristic output of the layer 1 attention network in the time multi-layer perceptron of the deep ATF module,/I>Is the weight matrix of the layer 1 attention network in the time multi-layer perceptron of the deep atf module,is the bias vector of the layer 1 attention network in the time multi-layer perceptron of the deep ATF module,/for the layer 1 attention network>Is the corresponding time characteristic output of the nth layer in the time multi-layer perceptron of the deep ATF module,/->Is the weight matrix corresponding to the n-th layer in the time multi-layer perceptron of the deep ATF module,/h>Is the time characteristic output of the n-1 layer attention network in the time multi-layer perceptron of the deep ATF module,/for the time multi-layer perceptron>Is the bias vector corresponding to the n-th layer in the time multi-layer perceptron of the deep ATF module,/->Is a temporal feature vector extracted through a multi-layer attention network,/->Is the weight matrix of the nth layer attention network in the time multi-layer perceptron of the deep ATF module,/and%>Is the bias vector of the n-th layer attention network in the time multi-layer perceptron of the deep ATF module,is the corresponding time characteristic output of the n-1 layer in the time multi-layer perceptron of the deep ATF module.
4. The project recommendation method based on feature interaction information and time tensor decomposition according to claim 3, wherein in step 3, user depth features, project depth features and time depth features are input to a GMF-TF module, and second-order feature interaction information among the user, the project and the time is obtained through generalized matrix decomposition learning, and the expression is as follows:
；
；
；
wherein,for the second order feature interaction vector of the user and the project in the GMF-TF module,/for the user and the project>For the second order feature interaction vector of the user and time in the GMF-TF module,/for the user and time>Is a second-order feature interaction vector of items and time in the GMF-TF module,for the second activation function, +.>Is->Corresponding weight matrix, < >>Is->Corresponding weight matrix, < >>Is->Corresponding weight matrix, < >>Is->Corresponding bias vector, ">Is->The corresponding offset vector is used to determine the offset,is->Corresponding bias vector, ">Representing dot product.
5. The project recommendation method based on feature interaction information and time tensor decomposition according to claim 4, wherein in step 3, the multi-layer attention user feature output of the deep ATF module, the multi-layer attention project feature output of the deep ATF module, and the multi-layer attention time feature output of the deep ATF module are input to the GMF-ATF module, and second-order feature interaction information among the user, the project and the time is obtained through generalized matrix decomposition learning, and the expression is as follows:
；
；
；
wherein,for the second order feature interaction vector of the user and the item in the GMF-ATF module, +.>For the second order feature interaction vector of the user and time in the GMF-ATF module,/for the user and time>Is the second order feature interaction vector of items and time in the GMF-ATF module, +.>Is->Corresponding weight matrix, < >>Is->Corresponding weight matrix, < >>Is in combination withCorresponding weight matrix, < >>Is->Corresponding bias vector, ">Is->Corresponding bias vector, ">Is->A corresponding bias vector.
6. The item recommendation method based on feature interaction information and time tensor decomposition according to claim 5, wherein in step 3, the predictive score value satisfies the following conditional expression:
；
；
；
；
；
wherein,is indicated at +.>When (4) user->Item->Is a predictive score value for (a); />、/>、/>、/>A weight matrix for the prediction module; />Indicating the number of hidden layers; />、/>、/>、/>For the bias vector of the prediction module, +.>、/>、/>、/>Is an intermediate value.
7. The item recommendation method based on feature interaction information and time tensor decomposition according to claim 6, wherein in step 4, the expression for minimizing the objective function loss is:
；
wherein,minimizing a function for losses by adopting an Adam gradient descent method; />Is indicated at +.>When (4) user->Item->Is a true score value of (2);U、V、Ta potential factor matrix of user, project, time, respectively,>、/>、/>respectively representing the total number of users, items, and times.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311288046.3A CN117033948B (en) | 2023-10-08 | 2023-10-08 | Project recommendation method based on feature interaction information and time tensor decomposition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311288046.3A CN117033948B (en) | 2023-10-08 | 2023-10-08 | Project recommendation method based on feature interaction information and time tensor decomposition |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117033948A true CN117033948A (en) | 2023-11-10 |
| CN117033948B CN117033948B (en) | 2024-01-09 |
Family
ID=88641470
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311288046.3A Active CN117033948B (en) | 2023-10-08 | 2023-10-08 | Project recommendation method based on feature interaction information and time tensor decomposition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117033948B (en) |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170132509A1 (en) * | 2015-11-06 | 2017-05-11 | Adobe Systems Incorporated | Item recommendations via deep collaborative filtering |
| EP3179434A1 (en) * | 2015-12-10 | 2017-06-14 | Deutsche Telekom AG | Designing context-aware recommendation systems, based on latent contexts |
| CN112631560A (en) * | 2020-12-29 | 2021-04-09 | 上海海事大学 | Method and terminal for constructing objective function of recommendation model |
| CN113407864A (en) * | 2021-06-21 | 2021-09-17 | 南京邮电大学 | Group recommendation method based on mixed attention network |
| CN114021024A (en) * | 2021-12-09 | 2022-02-08 | 中国科学院重庆绿色智能技术研究院 | Deep reinforcement learning recommendation method fusing user preference prediction |
| US20220058714A1 (en) * | 2020-08-18 | 2022-02-24 | Salesforce.Com, Inc. | Systems and methods for next basket recommendation with dynamic attributes modeling |
| CN115168721A (en) * | 2022-07-15 | 2022-10-11 | 齐鲁工业大学 | User interest recommendation method and system integrating collaborative transformation and temporal perception |
| CN115687757A (en) * | 2022-10-28 | 2023-02-03 | 南宁师范大学 | Recommendation method fusing hierarchical attention and feature interaction and application system thereof |
| CN115809339A (en) * | 2022-07-04 | 2023-03-17 | 中移（杭州）信息技术有限公司 | Cross-domain recommendation method, system, device and storage medium |
| CN116664253A (en) * | 2023-07-28 | 2023-08-29 | 江西财经大学 | Project recommendation method based on generalized matrix decomposition and attention shielding |
-
2023
- 2023-10-08 CN CN202311288046.3A patent/CN117033948B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170132509A1 (en) * | 2015-11-06 | 2017-05-11 | Adobe Systems Incorporated | Item recommendations via deep collaborative filtering |
| EP3179434A1 (en) * | 2015-12-10 | 2017-06-14 | Deutsche Telekom AG | Designing context-aware recommendation systems, based on latent contexts |
| US20220058714A1 (en) * | 2020-08-18 | 2022-02-24 | Salesforce.Com, Inc. | Systems and methods for next basket recommendation with dynamic attributes modeling |
| CN115917577A (en) * | 2020-08-18 | 2023-04-04 | 硕动力公司 | System and method for next basket recommendation using dynamic attribute modeling |
| CN112631560A (en) * | 2020-12-29 | 2021-04-09 | 上海海事大学 | Method and terminal for constructing objective function of recommendation model |
| CN113407864A (en) * | 2021-06-21 | 2021-09-17 | 南京邮电大学 | Group recommendation method based on mixed attention network |
| CN114021024A (en) * | 2021-12-09 | 2022-02-08 | 中国科学院重庆绿色智能技术研究院 | Deep reinforcement learning recommendation method fusing user preference prediction |
| CN115809339A (en) * | 2022-07-04 | 2023-03-17 | 中移（杭州）信息技术有限公司 | Cross-domain recommendation method, system, device and storage medium |
| CN115168721A (en) * | 2022-07-15 | 2022-10-11 | 齐鲁工业大学 | User interest recommendation method and system integrating collaborative transformation and temporal perception |
| CN115687757A (en) * | 2022-10-28 | 2023-02-03 | 南宁师范大学 | Recommendation method fusing hierarchical attention and feature interaction and application system thereof |
| CN116664253A (en) * | 2023-07-28 | 2023-08-29 | 江西财经大学 | Project recommendation method based on generalized matrix decomposition and attention shielding |
Non-Patent Citations (6)
| Title |
|---|
| XING XIANG; ZHIYUA ZHONG; YIBO CHEN; ZUPING ZHANG: "《Document similarity detection based on multi-feature semantic fusion and concept graph》", 《2022 IEEE 2ND INTERNATIONAL CONFERENCE ON ELECTRONIC TECHNOLOGY, COMMUNICATION AND INFORMATION (ICETCI)》 * |
| 冯金慧;陶宏才;: "基于注意力的深度协同在线学习资源推荐模型", 成都信息工程大学学报, no. 02 * |
| 刘慧婷;纪强;刘慧敏;赵鹏;: "基于双层注意力机制的联合深度推荐模型", 华南理工大学学报(自然科学版), no. 06 * |
| 廖国琼；杨乐川: "《注意力感知的群组Next事件推荐策略》", 《计算机科学与探索》 * |
| 胡志斌: "《交互性记忆系统视角下大型建设项目网络结构特征对项目绩效影响机制研究》", 《硕士电子期刊出版信息》 * |
| 贺思源: "《基于自编码器的神经协同过滤推荐算法研究》", 《硕士电子期刊出版信息》 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117033948B (en) | 2024-01-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108648049B (en) | Sequence recommendation method based on user behavior difference modeling | |
| CN111291266A (en) | Artificial intelligence based recommendation method and device, electronic equipment and storage medium | |
| CN114817663B (en) | Service modeling and recommendation method based on class perception graph neural network | |
| CN111241394A (en) | Data processing method and device, computer readable storage medium and electronic equipment | |
| CN113191154B (en) | Semantic analysis method, system and storage medium based on multi-modal graph neural network | |
| CN111611488A (en) | Information recommendation method and device based on artificial intelligence and electronic equipment | |
| CN114580707A (en) | Emotional tendency prediction model, building method and prediction method of multi-feature fusion product | |
| CN112699310A (en) | Cold start cross-domain hybrid recommendation method and system based on deep neural network | |
| Ahamed et al. | A recommender system based on deep neural network and matrix factorization for collaborative filtering | |
| CN116468460A (en) | Consumer finance customer image recognition system and method based on artificial intelligence | |
| CN117033948B (en) | Project recommendation method based on feature interaction information and time tensor decomposition | |
| CN117235264A (en) | Text processing method, apparatus, device and computer readable storage medium | |
| CN117391497A (en) | News manuscript quality subjective and objective scoring consistency evaluation method and system | |
| CN116934380A (en) | E-commerce material supply and demand combined prediction method under abnormal event | |
| CN116340635A (en) | Article recommendation method, model training method, device and equipment | |
| KR102410981B1 (en) | Curation system for recommending contents based on information of user response with ai | |
| CN113095883B (en) | Video payment user prediction method and system based on deep cross attention network | |
| CN111949884B (en) | Multi-mode feature interaction-based depth fusion recommendation method | |
| CN113313535A (en) | Click rate estimation method based on multi-domain partition integrated network | |
| CN113255891B (en) | Method, neural network model and device for processing event characteristics | |
| Yin et al. | PeNet: A feature excitation learning approach to advertisement click-through rate prediction | |
| CN114971784B (en) | Session recommendation method and system based on graph neural network by fusing self-attention mechanism | |
| Feng et al. | Adaptive attention with consumer sentinel for movie box office prediction | |
| CN117786234B (en) | Multimode resource recommendation method based on two-stage comparison learning | |
| CN116150506A (en) | Self-attention sequence recommendation method and system integrating feature comparison learning |
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 |