CN114662652A - Expert recommendation method based on multi-mode information learning - Google Patents

Expert recommendation method based on multi-mode information learning Download PDF

Info

Publication number
CN114662652A
CN114662652A CN202210155520.4A CN202210155520A CN114662652A CN 114662652 A CN114662652 A CN 114662652A CN 202210155520 A CN202210155520 A CN 202210155520A CN 114662652 A CN114662652 A CN 114662652A
Authority
CN
China
Prior art keywords
expert
embedding
information
learning
project
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210155520.4A
Other languages
Chinese (zh)
Inventor
王书海
彭浩
唐翊群
赵晓亮
王辉
胡畅霞
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.)
Shijiazhuang Tiedao University
Original Assignee
Shijiazhuang Tiedao University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shijiazhuang Tiedao University filed Critical Shijiazhuang Tiedao University
Priority to CN202210155520.4A priority Critical patent/CN114662652A/en
Publication of CN114662652A publication Critical patent/CN114662652A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • 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
    • G06F16/951—Indexing; Web crawling techniques
    • 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)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses an expert recommending method based on multi-mode information learning, which comprises the steps of crawling network expert information data and project information data, and sorting expert information data in an existing expert database; constructing an expert review abnormal composition and an expert attribute abnormal composition; the bert model learns the text information of the expert, the graph neural network learns the attribute heterogeneous graph of the expert, a self-attention recommendation model is built to learn the historical review sequence of the expert, and the learned embedded input fusion layer is embedded with the expert information to obtain a pre-training model; extracting expert information embedding by using a pre-training model, coding the item information to obtain item information embedding, inputting the obtained expert attribute representation and the obtained item information representation into a multi-layer perceptron to train the model, and obtaining recommendation scores of the experts and the items. The invention fully integrates abundant semantic and attribute information into the embedding and model parameters of the expert, and improves the accuracy of expert recommendation.

Description

Expert recommendation method based on multi-mode information learning
Technical Field
The invention belongs to the technical field of data processing, and particularly relates to an expert recommendation method based on multi-modal information learning.
Background
With the development of science and technology and the like, theoretical innovation is greatly promoted, the application amount of various innovative projects is greatly increased, and further, the application of science and technology projects is continuously increased. Wherein, a key step, namely recommendation of evaluation experts, exists in the stages of establishing and ending items of scientific research projects. And recommending by the evaluation expert, namely recommending the experts in the related fields to evaluate the project according to the scientific research project document so as to evaluate the actual significance, feasibility and completion quality of the project. This requires that the technology and the field of expertise of the reviewer be matched with the contents of the application to obtain a more accurate review result. Therefore, how to accurately retrieve the selected and recommended review experts that match the project becomes important and critical.
In the scientific research project declaration process, the assessment experts are selected in a manual mode, because the manual understanding of the domain knowledge is limited and has certain subjective tendency, the project declaration quantity is increased, the information quantity of an expert database is huge, and the problems of mechanical word matching, low recall ratio and low precision ratio of retrieval and recommendation exist in the traditional keyword retrieval and recommendation technology, so that the problems that projects distributed by many experts are irrelevant to the research direction of the experts are caused; meanwhile, whether the recommended review expert meets the review standard or not is judged only by the subjective consciousness of the project manager, so that the workload of the manager is huge, and the situation that the expert is not matched with the project is easily caused.
The existing expert recommendation system, such as an algorithm based on project research content and review expert research direction, calculates similarity of text feature vectors by extracting keywords, thereby omitting related other information. However, in practice, the selection of the scientific research project and the review expert often involves other factors, such as personal attributes of the expert. Meanwhile, when a new expert appears, the problem that expert information is incomplete often exists, a model cannot learn the effective representation of the expert, and the expert is difficult to recommend when items needing to be reviewed appear, so that the recommendation accuracy rate is low.
Disclosure of Invention
In order to solve the problems, the invention provides an expert recommendation method based on multi-modal information learning, which utilizes the crawled expert data, project data and the existing expert database data to execute various self-supervised pre-training tasks through multi-task learning, fully integrates rich semantic and attribute information into the embedding and model parameters of experts, and improves the accuracy of expert recommendation.
In order to achieve the purpose, the invention adopts the technical scheme that: an expert recommendation method based on multi-modal information learning comprises the following steps:
s10, crawling network expert information data and project information data, preprocessing the data, and sorting the expert information data in the existing expert database;
s20, constructing an expert historical review project sequence according to the information of the expert acting as a review committee, and constructing an expert attribute abnormal composition according to the expert attribute;
s30, learning the text information of the expert by using a bert model, learning the attribute heteromorphic graph of the expert by using a graph neural network model, building a self-attention recommendation model to learn the historical evaluation project sequence of the expert, and embedding the obtained embedded information into a fusion layer to obtain expert information embedding to obtain a pre-training model;
and S40, extracting expert information embedding by using a pre-training model, simultaneously, coding the project information to obtain the project information embedding through an embedding layer, combining the obtained expert attribute representation and the obtained project information representation, inputting the combined expert attribute representation and the project information representation into a multi-layer perceptron to train the model, and learning to obtain the matching degree score of the expert on the evaluation project.
Further, the step S10 includes:
crawling network expert information data comprising expert text information; crawling project information data in a network; preprocessing the crawled data;
and arranging data in the existing expert database, wherein the data comprises names, positions, research fields, attribute information of working units and review activity information of experts.
Further, the step S20 includes:
establishing a historical evaluation project sequence for the evaluation projects participated by the experts in the time sequence;
and constructing an expert attribute abnormal graph by taking the expert names, positions, research fields and working units as nodes and taking the correlation among the nodes as edges.
Further, in the step S30, the method includes:
inputting expert text information including articles and personal profiles into a bert model, and pre-training the bert model by executing a masking language prediction task to obtain expert text embedding;
on the expert attribute heterogeneous graph, executing a comparison pre-training task, pre-training a graph neural network for learning the expert attribute, and obtaining expert attribute embedding;
inputting the expert historical review sequence into a built self-attention recommendation model, executing a next project prediction task, learning the expert historical review project sequence, and embedding an expert mean (used for basic interest) and a covariance (used for variability of interest);
and fusing the extracted different embeddings to obtain expert information embedding.
Further, inputting expert text information including articles and personal profiles into a bert model, training the bert model by executing a masking language prediction task, and obtaining expert text embedding, comprising the steps of:
s311, converting the characters in the text information into word embedding through an embedding matrix, and adding corresponding segment embedding and position embedding to each character to be used as input of a bert model;
s312, in the masking language prediction task, selecting words related to the expert research field to shield and reconstruct;
s313, the masking language prediction task loss is defined as a cross entropy loss.
Further, a comparison pre-training task is carried out on the expert attribute heterogeneous graph, a graph neural network used for learning the expert attributes is trained, and embedding of the expert attributes is obtained, and the method comprises the following steps:
s321, executing a relation-level pre-training task, and for a given positive example triple, constructing a negative example queue of inconsistent relations and irrelevant nodes for the positive example triple and performing a comparative learning task;
s322, executing a sub-graph pre-training task, generating a metagraph instance on the heterogeneous graph to construct a positive sample, generating a queued negative sample, and distinguishing the positive sample and the negative sample through comparison and learning;
further, inputting the expert history review sequence into a self-attention recommendation model, executing a next project prediction task, learning the expert history review project sequence, and obtaining an expert mean (for basic interest) and a covariance (for variability of interest) embedding, the method comprises the following steps:
s331, a self-attention recommendation model which comprises regularization terms in random embedding, Wasserstein self-attention layer and BPR loss;
s332, in the random embedding layer, by representing the items as a multidimensional elliptic gaussian distribution. The elliptical gaussian distribution is controlled by the mean vector and the covariance vector. For all items, define a mean embedding table Mμ∈R|v|*dSum covariance Embedded Table M∑∈R|v|*d. Introducing separate position embedding P for mean embedding and covariance embedding simultaneouslyμ∈R|v|*dAnd P∑∈R|v|*dThus, mean-inclusive (for basic interest) embedding of sequences is obtained
Figure BDA0003512374660000031
And covariance (for variability of interest) embedding
Figure BDA0003512374660000032
Embedding:
Figure BDA0003512374660000033
Figure BDA0003512374660000041
s333, giving the items Sk and St in the Wasserstein self-attention layer, and respectively embedding the corresponding random embedding into d-dimensional elliptical Gaussian distribution
Figure BDA0003512374660000042
Wherein:
Figure BDA0003512374660000043
Figure BDA0003512374660000044
Figure BDA0003512374660000045
is a trainable matrix;
meanwhile, Wasserstein distance is introduced as attention weight to measure the pairwise relation among items in the sequence, and the linear combination characteristic of Gaussian distribution is adopted to aggregate historical items and obtain sequence representation;
s334, inputting the sequence representation into a feedforward neural network, applying two point-by-point full-connection layers with ELU activation to introduce nonlinearity in learning random embedding, adopting residual connection, layer normalization and dropout layers, and outputting the layers as mean embedding and covariance embedding of the article;
and S335, executing a next project prediction task, wherein the prediction score is the Wasserstein distance between the articles, the distance between the positive sample and the negative sample is increased by using a regularization term, the loss is defined as BRP loss, and a self-attention recommendation model is pre-trained.
Further, the extraction of different embeddings is fused to obtain expert information attributes, including:
embedding the expert text extracted by the pre-training model, embedding the expert review and embedding the expert attribute into an input fusion layer to obtain expert information embedding:
U'＝LN(aWa+wWw+ZμWμ+ZΣWΣ)，
U＝LN(U'+FFN(U'))；
wherein a represents expert attribute embedding, w represents expert text embedding, and ZμMean value embedding, Z, representing an expert review sequenceΣEmbedding covariance representing expert review sequences; wa,Ww,Wμ,W∑For trainable matrices, U is expert information embedding.
Further, in the step S40, the method includes the steps of:
s41, extracting new expert information by using a pre-training model and embedding the new expert information;
s42, encoding the project information, and obtaining the project information embedding after mapping;
s43, fusing the embedding of the new extracted expert information and the embedding of the project, inputting the information into the multi-layer perceptron, and training the multi-layer perceptron according to historical expert review information;
and finally, obtaining recommendation scores of the experts and the items to recommend the experts.
The beneficial effects of the technical scheme are as follows:
the invention crawls network expert information data and project information, preprocesses the data, and sorts the expert information data in the prior expert database. Inputting collected text information of an expert, such as a thesis and the like, into a bert model for learning, constructing an attribute differential graph by using collected attribute characteristics, learning the attribute characteristics by using a graph neural network model, modeling review information of the expert into a historical review sequence, building a self-attention recommendation model for learning the review characteristics, fusing different characteristics to obtain the information characteristics of the expert, and performing training on the model. In the recommendation task, the pre-trained model and parameters are applied to extraction of expert information embedding and fine adjustment, extracted expert information embedding and project information embedding are used for training the multi-layer perceptron, and recommendation scores of experts and projects are learned. According to the invention, various self-supervised pre-training tasks are executed through multi-task learning, abundant semantics, attribute information and review information are fully merged into the embedding and model parameters of experts, and the accuracy of expert recommendation is improved.
The invention builds a self-attention recommendation model for processing historical review sequence information of experts. He includes regularization terms in random embedding, Wasserstein self-attention layer, and BPR loss. The items were modeled as gaussian distributions with random embedding, including mean (for basic interest) and covariance (for variability of interest) embedding. On top of random embedding, the item transformations are measured using distance, which results from metric learning. A novel Wasserstein self-attention layer is proposed that measures attention as a scaled Wasserstein distance between items. A new regularization term is also introduced in the pre-training process, so that the distance between the positive sample item and the negative sample item is considered, the work of considering cooperative transitivity in the SR is considered, random embedding is introduced to measure the inherent basic interest and interest variability in the user behavior, the distance between the positive sample item and the negative sample item is limited through additional regularization, and the recommendation effect is improved.
The expert recommendation method provided by the invention is focused on embedding the learning experts. In order to effectively utilize various structural and non-structural information of experts, potential expert information is mined, expert attributes and expert review information are modeled into an expert attribute graph and an expert historical review sequence, the text information of the experts is input into a bert model, the text representation of the experts is extracted, the attribute graph of the experts is input into a graph neural network model, the attribute representation of the experts is extracted, the expert historical review sequence is input into a built self-attention recommendation model, the expert review representation is learned, the learned representations are fused, the information representation of the experts is obtained, and therefore various information of the experts is fused into the information representation of the experts, more knowledge is obtained, and the recommendation effect is improved.
In order to learn the text information, the attribute information and the review information of the expert, the invention adopts a pre-training model to learn the characteristics of the expert. The model combines a bert model, a graph neural network model, and a self-attention recommendation model. Inputting the text information of an expert into a bert model, inputting the attribute diagram of the expert into a graph neural network model, inputting the historical review sequence of the expert into a self-attention recommendation model, executing various self-supervision pre-training tasks by executing multi-task learning, fully integrating rich semantic and attribute information into the embedding and model parameters of the expert, extracting the expert information by using the trained model, embedding and integrating to obtain the information characteristics of the expert, and improving the recommendation effect.
Drawings
FIG. 1 is a schematic flow chart of an expert recommendation method based on multi-modal information learning according to the present invention;
FIG. 2 is a schematic diagram of a schematic framework of an expert recommendation method based on multi-modal information learning according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a self-attention recommendation model built in an embodiment of the invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described with reference to the accompanying drawings.
In this embodiment, referring to fig. 1 and 2, the present invention provides an expert recommendation method based on multi-modal information learning, including the steps of:
s10, crawling network expert information data and project information data, preprocessing the data, and sorting the expert information data in the existing expert database;
s20, constructing an expert historical review project sequence according to the expert acting as the information of the reviewer, and constructing an expert attribute abnormal composition according to the expert attribute;
s30, learning the text information of the expert by using a bert model, learning the attribute heteromorphic graph of the expert by using a graph neural network model, building a self-attention recommendation model to learn the historical evaluation project sequence of the expert, and embedding the obtained embedded information into a fusion layer to obtain expert information embedding to obtain a pre-training model;
and S40, extracting expert information embedding by using a pre-training model, simultaneously, coding the project information to obtain the project information embedding through an embedding layer, combining the obtained expert attribute representation and the obtained project information representation, inputting the combined expert attribute representation and the project information representation into a multi-layer perceptron to train the model, and learning to obtain the matching degree score of the expert on the evaluation project.
As an optimization scheme of the above embodiment:
the expert information data comprises text information such as articles of experts, the project information data comprises description information such as documents of projects, and the expert database comprises expert attribute information and historical review information. In actual application, due to the fact that data are sparse, experts cannot be well embedded, different patterns are built for the expert attribute information and review information, the expert information is learned by adopting a pre-training mode, the expert information is better extracted to be embedded, and the learned knowledge is used for an existing recommendation system.
By crawling network expert information data, including expert text information; crawling project information data in a network; preprocessing the crawled data;
meanwhile, the data in the existing expert database is sorted, and the data comprises names, positions, research fields, attribute information of working units and review activity information of the experts.
As an optimization scheme of the above embodiment: the construction of the expert attribute abnormal picture and the expert historical review sequence comprises the following steps:
establishing a historical evaluation project sequence for the evaluation projects participated by the experts in the time sequence;
and constructing an expert attribute abnormal graph by taking the expert names, positions, research fields and working units as nodes and taking the correlation among the nodes as edges.
As an optimization scheme of the above embodiment, the bert model and the graph neural network model are used to learn various information of an expert, including:
inputting expert text information including articles and personal profiles into a bert model, and training the bert model by executing a masking language prediction task to obtain expert text embedding;
on the expert attribute heterogeneous graph, a comparison pre-training task is carried out on the expert attribute heterogeneous graph, a graph neural network for learning expert review items is trained, and expert review embedding is obtained;
inputting the expert historical review sequence into a built self-attention recommendation model, executing a next project prediction task, learning the expert historical review project sequence, and embedding an expert mean (used for basic interest) and a covariance (used for variability of interest);
and fusing the extracted different embeddings to obtain the expert information attribute.
The method comprises the following steps of inputting expert text information including articles and personal profiles into a bert model, training the bert model by executing a masking language prediction task, and acquiring expert text embedding, wherein the method comprises the following steps:
s311, converting the characters in the text information into word embedding through an embedding matrix, and adding corresponding segment embedding and position embedding to each character to be used as input of a bert model;
s312, in the masking language prediction task, selecting words related to the expert research field to shield and reconstruct;
s313, the masking language prediction task loss is defined as a cross entropy loss:
Figure BDA0003512374660000081
wherein S isMIs a set of masked positions that are to be masked,
Figure BDA0003512374660000082
and WjAre the predicted word and the original word.
The method comprises the following steps of comparing an expert attribute heterogeneous graph with a pre-training task, training a graph neural network for learning the expert attribute, and acquiring expert attribute embedding, and comprises the following steps:
s321, executing a relation level pre-training task, and regarding a given positive example triple < u, R, v >. epsilon.P in the graphrelFor its structural inconsistency negative example < u, R-W > formed queues and irrelevant nodes < u, x, v-The > load example queue performs a comparative learning task.
S322, executing a subgraph level pre-training task, and constructing a metagraph instance M epsilon M (M) as a set of nodes u according with the metagraph M for a given metagraph M epsilon M and a source node u, wherein I (M) is expressed as a set of all instances of the metagraph M. Based on previous positive samples in the training process, a negative sample is generated by adding the latest positive sample and removing the tail end of the earliest queue, and a comparison training task is performed.
Wherein, inputting the expert history review sequence into the built self-attention recommendation model as shown in fig. 3, executing the next project prediction task, learning the expert history review project sequence, and obtaining the embedding of the expert mean (for basic interest) and covariance (for variability of interest), comprising the steps of:
s331, a self-attention recommendation model which comprises regularization terms in random embedding, Wasserstein self-attention layer and BPR loss;
s332, in the random embedding layer, representing the items as multi-dimensional elliptic Gaussian distribution; the elliptical Gaussian distribution is controlled by a mean vector and a covariance vector; for all items, define a mean embedding table Mμ∈R|v|*dSum covariance Embedded Table M∑∈R|v|*d(ii) a Introducing separate position embedding P for mean embedding and covariance embedding simultaneouslyμ∈R|v|*dAnd P∑∈R|v|*d(ii) a Thus, we get an embedding that includes mean (for basic interest) and covariance (for variability of interest):
Figure BDA0003512374660000083
Figure BDA0003512374660000084
s333, giving the items Sk and St in the Wasserstein self-attention layer, and respectively embedding the corresponding random embedding into d-dimensional elliptical Gaussian distribution
Figure BDA0003512374660000085
Wherein:
Figure BDA0003512374660000091
Figure BDA0003512374660000092
Figure BDA0003512374660000093
is a trainable matrix.
Wasserstein distance is introduced as an attention weight to measure the pairwise relationship between items in the sequence, and the linear combination property of Gaussian distribution is adopted to aggregate historical items and obtain a sequence representation.
The weight A between the item Sk and St is defined as the negative 2-Wasserstein distance W2The measurement is as follows:
Figure BDA0003512374660000094
the output embedding of the items at each position of the sequence is a weighted sum of the embedding from the previous step, where the weight is a normalized attention value
Figure BDA0003512374660000095
Figure BDA0003512374660000096
Each item is represented as a random embedding with mean and covariance, and the items are aggregated using a linear combination property of gaussian distribution, as follows:
Figure BDA0003512374660000097
wherein the content of the first and second substances,
Figure BDA0003512374660000098
w is a trainable matrix.
The output is a random embedding of the sequence:
Figure BDA0003512374660000099
s334, in the feedforward neural network, two point-by-point fully-connected layers with ELU activation are applied to introduce non-linearity in learning random embedding. Residual connection, layer normalization and dropout layers are adopted, and layer output is mean value embedding and covariance embedding of items;
s335, executing next project prediction task, predicting the Wasserstein distance between the items, and using the regularization item
Figure BDA00035123746600000910
The distance between the positive sample and the negative sample is increased, a self-attention recommendation model is pre-trained, and the loss is defined as BRP loss:
Figure BDA00035123746600000911
wherein [ x ]]Max (x, 0) is the standard hinge loss, j+For the true next item, j-Negative examples were chosen for random.
As an optimization scheme of the above embodiment, the extracting different embeddings are fused to obtain expert information attributes, including:
embedding the expert text extracted by the pre-training model, obtaining expert mean (for basic interest) and covariance (for variability of interest) embedding, and embedding expert attributes into the input fusion layer, obtaining expert information embedding:
U'＝LN(aWa+wWw+ZμWμ+ZΣWΣ)，
U＝LN(U'+FFN(U'))；
wherein a represents expert attribute embedding, w represents expert text embedding, and ZμMean value embedding, Z, representing an expert review sequenceΣCovariance embedding of sequences representing expert reviews; wa,Ww,Wμ,W∑Embedding the information of an expert into a trainable matrix U;
and (3) according to the embedding of expert information, predicting the expert score, embedding better representation of the information of the learning expert, wherein the prediction score is r' ═ WU + b, W, b is trainable parameter, and the loss function is defined as:
Figure BDA0003512374660000101
as an optimization scheme of the above embodiment, extracting expert information embedding by using a pre-training model, simultaneously encoding item information to obtain item information embedding, inputting the obtained expert attribute representation and item information representation into a multi-layer perceptron (MLP) to train the model, and obtaining recommendation scores of experts and items, includes the steps of:
s41, extracting new expert information by using a pre-training model and embedding the new expert information;
s42, encoding the project information, and obtaining the project information embedding after mapping;
s43, according to the new expert information embedding and project embedding fusion, input into the multi-layer perceptron, according to historical expert review information training the multi-layer perceptron, the loss function is:
Figure BDA0003512374660000102
and finally, obtaining the recommendation scores of the experts and the items, and recommending the experts.
The foregoing shows and describes the general principles and broad features of the present invention and advantages thereof. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims (8)

1. An expert recommendation method based on multi-modal information learning is characterized by comprising the following steps:
s10, crawling network expert information data and project information data, preprocessing the data, and sorting the expert information data in the existing expert database;
s20, constructing an expert attribute abnormal picture according to the expert attributes, and constructing an expert historical review project sequence according to the information of the expert acting as a review committee;
s30, learning the text information of the expert by using a bert model, learning the attribute heteromorphic graph of the expert by using a graph neural network model, building a self-attention recommendation model to learn the historical evaluation project sequence of the expert, and embedding the obtained embedded information into a fusion layer to obtain expert information embedding to obtain a pre-training model;
and S40, extracting expert information embedding by using a pre-training model, simultaneously, coding the project information to obtain the project information embedding through an embedding layer, combining the obtained expert attribute representation and the obtained project information representation, inputting the combined expert attribute representation and the project information representation into a multi-layer perceptron to train the model, and learning to obtain the matching degree score of the expert on the evaluation project.
2. The expert recommendation method based on multi-modal information learning as claimed in claim 1, wherein in the step S10, it comprises:
crawling network expert information data comprising expert text information; crawling project information data in a network; preprocessing the crawled data;
and arranging data in the existing expert database, wherein the data comprises names, positions, research fields, attribute information of working units and review activity information of experts.
3. The expert recommendation method based on multi-modal information learning as claimed in claim 2, wherein in the step S20, it comprises:
establishing a historical evaluation project sequence for the evaluation projects participated by the experts in the time sequence;
and constructing an expert attribute abnormal graph by taking the expert names, positions, research fields and working units as nodes and taking the correlation among the nodes as edges.
4. The expert recommendation method based on multi-modal information learning as claimed in claim 3, wherein in the step S30, it comprises:
inputting expert text information including articles and personal profiles into a bert model, and pre-training the bert model by executing a masking language prediction task to obtain expert text embedding;
on the expert attribute abnormal graph, executing a comparison pre-training task, pre-training a graph neural network for learning the expert attribute, and obtaining expert attribute embedding;
inputting the expert historical evaluation sequence into a built self-attention recommendation model, executing a next project prediction task, learning the expert historical evaluation project sequence, and obtaining expert mean and covariance embedding;
and fusing the extracted different embeddings to obtain the expert information attribute.
5. The method as claimed in claim 4, wherein the expert text information including articles and personal profiles is inputted into the bert model, and the bert model is trained by executing the masking language prediction task to obtain the expert text embedding, comprising the steps of:
s311, converting the characters in the text information into word embedding through an embedding matrix, and adding corresponding segment embedding and position embedding to each character to be used as input of a bert model;
s312, in the masking language prediction task, selecting words related to the expert research field to shield and reconstruct;
s313, the masking language prediction task loss is defined as a cross entropy loss.
6. The method for recommending experts on review based on multi-modal information learning, as claimed in claim 4, wherein a pre-training task is performed on the heterogeneous graph of expert attributes to train the graph neural network for learning the expert attributes, and obtain the expert attribute embedding, comprising the steps of:
s321, executing a relation-level pre-training task, and performing a comparative learning task for constructing an inconsistent relation and a negative sample queue of an irrelevant node for a given positive sample triple in the abnormal picture;
and S322, executing a sub-graph pre-training task, generating a metagraph instance on the heterogeneous graph to construct a positive sample, generating a queued negative sample, and distinguishing the positive sample from the negative sample through comparison and learning.
The expert review recommendation method based on multi-modal information learning according to claim 4, wherein the expert historical review sequence is input into the built self-attention recommendation model, the next project prediction task is executed, the expert historical review project sequence is learned, and expert mean) and covariance embedding are obtained, comprising the steps of:
s331, a self-attention recommendation model comprising regularization terms in a random embedding layer, a Wasserstein self-attention layer and BPRloss;
s332, in the random embedding layer, representing the items as multi-dimensional elliptic Gaussian distribution; the elliptical Gaussian distribution is controlled by the mean vector and the covariance vector; for all items, define a mean embedding table Mμ∈R|v|*dSum covariance Embedded Table M∑∈R|v|*d(ii) a Introducing separate position embedding P for mean embedding and covariance embedding simultaneouslyμ∈R|v|*dAnd P∑∈R|v|*d(ii) a Thus, a mean value embedding of the sequence is obtained
Figure FDA0003512374650000021
Sum covariance embedding
Figure FDA0003512374650000022
Figure FDA0003512374650000031
Figure FDA0003512374650000032
S333, giving the items Sk and St in the Wasserstein self-attention layer, and respectively obtaining d-dimensional elliptic Gaussian distribution corresponding to random embedding
Figure FDA0003512374650000033
Wherein:
Figure FDA0003512374650000034
Figure FDA0003512374650000035
is a trainable matrix;
meanwhile, Wasserstein distance is introduced as attention weight to measure the pairwise relation among items in the sequence, and the linear combination characteristic of Gaussian distribution is adopted to aggregate historical items and obtain sequence representation;
s334, inputting the sequence representation into a feedforward neural network, applying two point-by-point full-connection layers with ELU activation to introduce nonlinearity in learning random embedding, adopting residual connection, layer normalization and dropout layers, and outputting the layers as mean embedding and covariance embedding of each item;
and S335, executing a next project prediction task, wherein the prediction score is the Wasserstein distance between the articles, the distance between the positive sample and the negative sample is increased by using a regularization term, the loss is defined as BRP loss, and a self-attention recommendation model is pre-trained.
7. The method for recommending review experts based on multi-modal information learning according to claim 4, wherein the extracting different embeddings are fused to obtain expert information attributes, comprising:
embedding the expert text extracted by the pre-training model, embedding the mean value and covariance sequence of the expert and embedding the expert attribute into an input fusion layer to obtain expert information embedding:
U'＝LN(aWa+wWw+ZμWμ+ZΣWΣ)，
U＝LN(U'+FFN(U'))；
wherein a represents expert attribute embedding, w represents expert text embedding, and ZμMean value embedding, Z, representing an expert review sequenceΣEmbedding covariance representing expert review sequences; wa,Ww,Wμ,W∑Embedding U in expert information for a trainable matrix;
and (4) according to the embedding of expert information, predicting the expert score, embedding better representation of the information of the learning expert, and taking the prediction score as r ═ WU + b and W, b as trainable parameters.
8. The expert recommendation method based on multi-modal information learning as claimed in claim 1, wherein in said step S40, comprising the steps of:
s41, extracting new expert information by using a pre-training model and embedding the new expert information;
s42, encoding the project information, and obtaining the project information embedding after mapping;
s43, fusing the embedding of the extracted new expert information and the embedding of the project, inputting the information into the multi-layer perceptron, and training the multi-layer perceptron according to historical expert review information;
and finally, obtaining recommendation scores of the experts and the items to recommend the experts.
CN202210155520.4A 2022-02-21 2022-02-21 Expert recommendation method based on multi-mode information learning Pending CN114662652A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210155520.4A CN114662652A (en) 2022-02-21 2022-02-21 Expert recommendation method based on multi-mode information learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210155520.4A CN114662652A (en) 2022-02-21 2022-02-21 Expert recommendation method based on multi-mode information learning

Publications (1)

Publication Number Publication Date
CN114662652A true CN114662652A (en) 2022-06-24

Family

ID=82027241

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210155520.4A Pending CN114662652A (en) 2022-02-21 2022-02-21 Expert recommendation method based on multi-mode information learning

Country Status (1)

Country Link
CN (1) CN114662652A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115599984A (en) * 2022-09-09 2023-01-13 北京理工大学(Cn) Retrieval method
CN117763238A (en) * 2024-01-09 2024-03-26 南京理工大学 Multi-graph neural network-based academic paper review expert recommendation method

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115599984A (en) * 2022-09-09 2023-01-13 北京理工大学(Cn) Retrieval method
CN115599984B (en) * 2022-09-09 2023-06-09 北京理工大学 Retrieval method
CN117763238A (en) * 2024-01-09 2024-03-26 南京理工大学 Multi-graph neural network-based academic paper review expert recommendation method
CN117763238B (en) * 2024-01-09 2024-05-24 南京理工大学 Multi-graph neural network-based academic paper review expert recommendation method

Similar Documents

Publication Publication Date Title
CN112417306B (en) Method for optimizing performance of recommendation algorithm based on knowledge graph
CN108647226B (en) Hybrid recommendation method based on variational automatic encoder
CN110309195B (en) FWDL (full Width Domain analysis) model based content recommendation method
CN112579778A (en) Aspect-level emotion classification method based on multi-level feature attention
CN114662652A (en) Expert recommendation method based on multi-mode information learning
CN114693397B (en) Attention neural network-based multi-view multi-mode commodity recommendation method
CN114238577B (en) Multi-task learning emotion classification method integrating multi-head attention mechanism
CN113343125B (en) Academic accurate recommendation-oriented heterogeneous scientific research information integration method and system
CN111858940A (en) Multi-head attention-based legal case similarity calculation method and system
CN112989761A (en) Text classification method and device
CN114119058A (en) User portrait model construction method and device and storage medium
CN111582506A (en) Multi-label learning method based on global and local label relation
CN112926325A (en) Chinese character relation extraction construction method based on BERT neural network
CN116975776A (en) Multi-mode data fusion method and device based on tensor and mutual information
CN116976505A (en) Click rate prediction method of decoupling attention network based on information sharing
Castellano et al. Integrating contextual knowledge to visual features for fine art classification
CN115221387A (en) Enterprise information integration method based on deep neural network
CN116127175A (en) Mobile application classification and recommendation method based on multi-modal feature fusion
CN115203529A (en) Deep neural network recommendation model and method based on multi-head self-attention mechanism
Sarang Thinking Data Science: A Data Science Practitioner’s Guide
CN112862569B (en) Product appearance style evaluation method and system based on image and text multi-modal data
CN112100395B (en) Expert cooperation feasibility analysis method
CN114357284A (en) Crowdsourcing task personalized recommendation method and system based on deep learning
CN111782802B (en) Method and system for obtaining commodity corresponding to national economy manufacturing industry based on machine learning
Laaroussi et al. A novel hybrid deep learning approach for tourism demand forecasting

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