A personalized paper recommendation method based on timeliness-assisted task-driven

Abstract

The present invention relates to a personalized paper recommendation method based on time-sensitive auxiliary task driving, comprising the following steps: constructing an academic HIN graph, and using a HIN subgraph extraction function to extract HIN subgraphs within long and short time spans. Adding a fully connected layer to the output of the classical BERT language model for fine-tuning to obtain paper keyword semantic features, aggregating the keyword semantic features through a self-attention mechanism to obtain paper content features, and aggregating the paper content features through a self-attention mechanism to obtain the current user's research direction features. Embedding a GNN network in the HIN subgraphs of the two time spans to capture user reading preference features and paper audience preference features. Performing multi-time-span related feature capture through the cfLSTM predictor in the present invention. Calculating the matching degree between the user's research direction and the paper content and calculating the matching degree between the user's reading preference and the paper audience preference, and linearly fusing the two matching degrees to obtain a recommendation probability value.

Description

Technical Field

[0001] The present invention relates to the field of academic paper recommendation, and in particular to a personalized paper recommendation method driven by time-sensitive auxiliary tasks. Background Art

[0002] With the development of digital libraries and academic paper websites, more and more academic papers are appearing, stored, and disseminated online at an unprecedented rate. In recent years, paper recommendation has become an indispensable tool in academia, and personalized paper recommendation algorithms have also received increasing attention.

[0003] Traditional personalized paper recommendation algorithms typically use collaborative filtering to analyze citation relationships between papers or collaborations between users to recommend relevant papers. However, this approach relies on a single source of information, which can easily lead to cold-start problems. Furthermore, most collaborative filtering methods rely solely on single interaction data for recommendation, neglecting the potential for rich and diverse auxiliary information to improve recommendation performance. To address this issue, a growing number of auxiliary information, such as user attributes, paper content, and paper keywords, has been incorporated, leading to a new recommendation approach: content-based recommendation. Researchers have used various methods to embed various auxiliary data into feature vectors and incorporate them into recommendation algorithms, alleviating the cold-start problem caused by sparse interaction data. However, in reality, relevant auxiliary information comes from diverse sources and forms, and much of it involves privacy concerns. Therefore, determining the weights of various types of information and establishing a unified data representation standard when implementing recommendation strategies presents a new challenge.

[0004] Graph-based recommendation methods have shown great potential for recommendation tasks, but they still have some shortcomings: most existing methods focus only on graph structures and semantic relationships in heterogeneous information networks to obtain user and paper features, and then recommend papers based on these features. However, they ignore whether the recommended results are consistent with the user's current research direction, because a senior researcher generally has a wealth of historical reading data, often covering multiple research directions. Secondly, many existing recommendation methods are based on static data modeling to mine relevant features of users and papers. These methods ignore the dynamic nature of features that change over time, so the final recommendation results may not be what the user is most interested in reading. Summary of the Invention

[0005] To this end, the technical problem to be solved by the present invention is to overcome the problems in the prior art of graph-based paper recommendation methods that ignore the degree of matching between the recommended results and the user's current research direction and the lack of timeliness of the current recommended results.

[0006] To solve the above technical problems, the present invention provides a personalized paper recommendation method based on time-sensitive auxiliary task driving, including:

[0007] Step 1: Construct an academic heterogeneous information network graph, model the dataset as the academic heterogeneous information network graph, which includes four types of nodes: users, papers, publication locations, and keywords, as well as relationships between nodes such as writing relationships, reading relationships, citation relationships, subordination relationships, and publication relationships; through the academic heterogeneous information network subgraph extraction function, from the perspective of a preset time period, obtain the academic heterogeneous information network subgraph under a long time span, and through the extraction function, obtain the academic heterogeneous information network subgraph under a short time span from the perspective of a preset year;

[0008] Step 2: Heterogeneous entity feature representation learning. According to the academic heterogeneous information network graph, first use the classical language model BERT and add a fully connected layer to its output to learn the keyword semantic features, and then aggregate the keyword semantic features in each paper through the self-attention mechanism to obtain the paper content features; to capture the dynamic changes in the user's research direction features, aggregate the paper content features published by the user within the long time span and the short time span through the self-attention mechanism respectively to obtain the user's research direction features under the long time span and the short time span; embed the GNN network in the academic heterogeneous information network subgraphs under the long time span and the short time span to capture the paper audience preference features and user reading preference features under the two time spans of long and short respectively;

[0009] Step 3: Dynamic mining of features under multiple time spans. Based on LSTM, add the user's research direction features, paper audience preference features, and user reading preference features learned in Step 2 to the input of the LSTM to construct a predictor corresponding to the predicted features. By inputting the user's research direction features, audience preference features, and user reading preference features under the long time span and the short time span into the corresponding predictors respectively, obtain the predicted user's research direction features, predicted paper audience preference features, and predicted user reading preference features;

[0010] Step 4: Probability prediction. Dot multiply the paper content features and the user's research direction features predicted in Step 3 to obtain the matching degree between the user's research direction and the paper content, that is, the matching degree under timeliness; dot multiply the predicted paper audience preference features and the predicted user reading preference features in Step 3 to obtain the matching degree between the user's reading preference and the paper audience preference, that is, the preference matching degree score; linearly fuse the timeliness score and the preference matching degree score to obtain the probability that the recommended paper will be read.

[0011] In an embodiment of the present invention, constructing the academic heterogeneous information network subgraph in Step 1 includes:

[0012] Arrange all the years involved in the dataset in ascending order to form a set YL, i.e., YL = [y0, y1, y2, …];

[0013] Use the academic heterogeneous information network subgraph extraction function SubHIN(HIN, Year) to extract the academic heterogeneous information network subgraph at the l-th year, i.e., a short time span, as G l = SubHIN(G, YL[l]), where l ∈ [0, |YL|), and |YL| is the length of the set YL;

[0014] Extract the academic heterogeneous information network subgraph over the past m years, i.e., a long time span, as G sub_l = SubHIN(G, YL[max(l - m, 0):l]), where YL[max(l - m, 0):l], and m ∈ [0, |YL|) is used to control the long time span level.

[0015] In an embodiment of the present invention, to learn the paper content features in step 2, define the set of all keywords involved in the dataset as KL, add a fully connected layer to the output of the classical language model BERT, and output the set KL to the language model BERT to obtain the semantic features of keyword k i of keyword k

[0016] where KL = [k0, k1, k2, …], represents the learnable parameter matrix of the fully connected layer; σ is a non-linear activation function; d BERT = 768 is the output dimension of BERT, and d K is the final embedding dimension of the keyword; is the semantic feature representation of keyword k i ;

[0017] To make the semantic features of keywords output by the classical language model BERT with a fully connected layer more suitable for the paper recommendation scenario, design a pre-training task to train the fully connected layer: construct a keyword semantic training set in the form of triples where When it means that keyword k i and keyword k j are discussed in the same paper, otherwise the keywords do not exist in the same paper; define the pre-training loss function as:

[0018]

[0019]

[0020] where |D K | represents the set D K The number of Represents the predicted keyword k i and keyword k j the correlation between

[0021] According to step 2, the keyword semantic features of the paper can be aggregated by the self-attention mechanism to obtain the content features of the paper, and KL is defined pi For the paper i Keyword set, The characteristics of the paper content are:

[0022]

[0023]

[0024] in For the paper i The number of keywords, SA(*) represents the self-attention mechanism, which is defined as follows:

[0025]

[0026] in is the input vector, They represent the learnable query weight parameter matrix and keyword weight parameter matrix respectively.

[0027] In one embodiment of the present invention, according to step 2, in order to obtain the dynamic changes of the user's research direction, first collect the user u i Papers published in a short time span and papers published over a long period of time The user paper content features in the two time spans are fused through the self-attention mechanism to obtain the user u i User research direction characteristics in a short time span and user u i Characteristics of user research directions over a long period of time Right now

[0028]

[0029]

[0030]

[0031] In one embodiment of the present invention, in order to effectively capture the audience preference characteristics of papers and the changes in the reading preference characteristics of users, two academic heterogeneous information network subgraphs with long time spans and short time spans are embedded into the GNN network to obtain:

[0032]

[0033] where represents the reading preference characteristics of user u i under a short time span, represents the audience preference characteristics of paper p i under a short time span; represents the reading preference characteristics of user u i under a long time span, represents the audience preference characteristics of paper p i under a long time span, and d G is the embedding dimension of the preference characteristics.

[0034] In one embodiment of the present invention, in order to implement the dynamic feature mining under multiple time spans in step 3, the predictor in step 3 is designed as a cfLSTM predictor. The features of the research direction of user u i under the short time span of the current year and the features of the research direction of user u under the long time span of the current year i are input into the cfLSTM predictor, and the research direction features of user u

[0035]

[0036] …

[0037]

[0038] where is the output state of the l-th cfLSTM cell, retains some data features of the l-th cfLSTM cell and is input as the input for calculation into the next cfLSTM cell; and are initialized to 0, and if it is considered that no paper was published in YL[l], then by default, the research direction features of the user under the short time span are the same as those of the previous year, that is If the user has not published a paper within m years in and before YL[l], then by default, the research direction features of the user under the long time span are also the same as those of the previous year, that is If l = 0, then the user research direction features are both 0 in the long time span and the short time span; cfLSTM l The formula is as follows:

[0039]

[0040]

[0041]

[0042] where the operation represents concatenation, ⊙ represents Hadamard product, are learnable parameters, z l , f l , o l are the input gate, forget gate, and output gate of the l-th cfLSTM cell respectively.

[0043] In an embodiment of the present invention, in the cfLSTM predictor, the cosine similarity between the predicted user research direction features for each year and the true user research direction features is used to define the loss function loss2 of the user research direction features:

[0044] A1 = |U| × |YL|,

[0045]

[0046] where |U| represents the number of users.

[0047] In an embodiment of the present invention, it is defined that the predicted value of the reading preference feature of user u i in year YL[l] obtained by the cfLSTM predictor is for paper p j the predicted value of the audience preference feature in year YL[l] is Using and the true user reading preference features the cosine similarity between them is used to define the loss function loss3 of the user reading preference features; using and the true audience preference features of the paper the cosine similarity between them is used to define the loss function loss4 of the audience preference features of the paper:

[0048] A2 = |U| × |YL|, A3 = |P| × |YL|,

[0049]

[0050]

[0051] in

[0052] In one embodiment of the present invention, according to step 4, the user u is predicted i Read the paper in YL[l] j The probability of In the year YL[l], the predicted value of the user research direction feature obtained by the cfLSTM predictor is The predicted value of reading preference characteristics is and paper p j The predicted value of the audience preference feature is The content features of the paper are obtained by aggregating keyword semantic features through the attention mechanism: Define read probability:

[0053]

[0054]

[0055]

[0056] in, The value of represents user u i Research direction and papers j The degree of matching between the contents, that is, the matching degree under the timeliness, The value of represents user u i Reading preferences and paper p j The matching degree between the audience preferences is the preference matching score, α is the weight parameter learned by the model, It is a comprehensive consideration and The final reading prediction value obtained.

[0057] In one embodiment of the present invention, a training task is designed to train the predicted value of the reading probability, and a training set D is constructed in the form of triples, that is, (u i ,r ij ,p j ), where r ij ∈{0,1}, when r ij =0, indicating that user u i Have not read the paper j , when r ij =1, indicating that user u i Read the paper j ;

[0058] The loss function is defined as:

[0059]

[0060] By minimizing the objective function Θ, the papers recommended to the user are made timely, and Θ is defined as follows:

[0061] Θ = loss5 + β(loss4 + loss3 + loss2) + λL2

[0062] Among them, β and λ are hyperparameters set manually, and L2 is a regularization term used to prevent the model from overfitting.

[0063] The above technical solution of the present invention has the following advantages compared with the prior art:

[0064] The present invention proposes a recommendation method driven by a timeliness auxiliary task to solve the timeliness problem of recommendation results and the problem of feature dynamics in paper recommendation, which is called the multi-time-span dynamic perception multi-task recommendation method MDRec. The present invention first designs an auxiliary task, that is, the task of matching the user's research direction with the paper content, and innovatively uses BERT fine-tuning and self-attention mechanism to obtain the paper content features and user research direction features therein. By multiplying the paper content features and user research direction features, the recommendation score under timeliness is obtained to assist in recommendation, overcoming the problem of lack of timeliness in paper recommendation in the prior art. The present invention also proposes a method for mining feature dynamics, and designs a cfLSTM predictor based on LSTM that can perceive feature changes from long time spans and short time spans. The cfLSTM predictor can capture the dynamic changes of any feature embedded based on the heterogeneous information network graph from both long and short time spans, so that the recommendation results obtained by relying on feature embedding are more accurate and more in line with the user's current research direction. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to make the content of the present invention easier to be clearly understood, the following further describes the present invention in detail according to the specific embodiments of the present invention in conjunction with the drawings, where

[0066] Figure 1 is the flowchart of the multi-time-span dynamic perception multi-task recommendation method of the present invention

[0067] Figure 2 is the schematic diagram of the multi-time-span dynamic perception multi-task recommendation method of the present invention

[0068] Figure 3 is the academic heterogeneous information network graph of the present invention

[0069] Figure 4 is Figure 3 the sub-graph of the shown academic heterogeneous information network graph under two time spans in 2002

[0070] Figure 5 It is the detailed diagram of the l-th cfLSTM cell in the cfLSTM predictor of the present invention Specific implementation manners

[0071] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and implement it, but the embodiments given are not intended to limit the present invention.

[0072] Refer to Figure 1 and Figure 2 As shown, the present invention proposes a multi-task recommendation method MDRec with multi-time-span dynamic perception for recommending the most suitable papers for users in their current research directions and of the greatest interest. The method flow chart and method schematic diagram are respectively as shown in Figure 1 、 Figure 2 As shown, and the specific operation steps are as follows:

[0073] Step 1: Refer to Figure 3 As shown, model the data set as an academic heterogeneous information network graph. The HIN graph includes four types of nodes: user User, U, paper Paper, P, venue Venue, V, and keyword Keyword, K, as well as relationships between nodes such as writing relationship UP_W, reading relationship UP_R, citation relationship PP, subordination relationship PK, and publication relationship PV; refer to Figure 4 As shown, through the academic heterogeneous information network subgraph extraction function SubHIN(HIN, Year), obtain the academic HIN subgraph G under a long time span from the perspective of a preset time period. sub_l = SubHIN(G, YL[max(l - m, 0):l]), where YL[max(l - m, 0):l], m ∈ [0, |YL|) is used to control the long time span level. Through the extraction function, obtain the academic HIN subgraph G under a short time span from the perspective of a preset year. l = SubHIN(G, YL[l]), where l ∈ [0, |YL|), and |YL| is the length of the set YL. Figure 4 They are the academic HIN subgraphs under the short time span of 2002 and the long time span of 2002 and the past 1 year.

[0074] Step 2: Heterogeneous entity feature representation learning. According to the academic heterogeneous information network graph, learn the paper content features, user research direction features, user reading preference features, and paper audience preference features.

[0075] First, use the classical language model BERT and add a fully connected layer to its output to learn the semantic features of keywords. Obtain the content features of the paper through the semantic features of keywords, and then acquire the user's research direction features. Define the set KL composed of all keywords involved in the dataset. Add a fully connected layer to the output of the classical language model BERT, and output the set KL to the language model BERT to obtain the semantic feature of keyword k i semantic feature where KL = [k0, k1, k2,...], represents the learnable parameter matrix of the fully connected layer; σ is a non-linear activation function; d BERT = 768 is the output dimension of BERT, and d K is the final embedding dimension of the keyword; is the semantic feature representation of keyword k i .

[0076] To make the semantic features of keywords output by the classical language model BERT after adding the fully connected layer more suitable for the scenario of paper recommendation, design a pre-training task to train the fully connected layer: construct a keyword semantic training set in the form of triples where when it means that keyword k i and keyword k j are discussed in the same paper, otherwise the keywords do not exist in the same paper; define the pre-training loss function as:

[0077]

[0078]

[0079] where |D K | represents the number of the set D K , represents the correlation between the predicted keyword k i and keyword k j .

[0080] Then, aggregate the semantic features of keywords in each paper through the self-attention mechanism to obtain the content features of the paper. Define as the keyword set of paper p i , and the content feature of the paper is:

[0081]

[0082]

[0083] where For paper p i the number of keywords, SA(*) represents the self-attention mechanism, and its definition is as follows:

[0084]

[0085] where is the input vector, respectively represent the learnable query weight parameter matrix and keyword weight parameter matrix.

[0086] To capture the dynamic changes in the characteristics of the user's research direction, the content characteristics of the papers published by the user in the long time span and short time span are aggregated through the self-attention mechanism respectively to obtain the user's research direction characteristics in the long time span and short time span. First, collect the papers published by user u i published in the short time span and the papers published in the long time span Through the self-attention mechanism, the content characteristics of the user's papers in the two time spans are fused respectively to obtain the user's research direction characteristics in the short time span i for user u and the user's research direction characteristics in the long time span for user u i That is Namely

[0087]

[0088]

[0089]

[0090] To obtain the user's reading preference characteristics and the audience preference characteristics of the papers, by embedding the two academic heterogeneous information network subgraphs of the long time span and short time span into the GNN network, the audience preference characteristics and the user's reading preference characteristics in the two time spans of long time and short time are captured respectively:

[0091]

[0092] where represents the reading preference characteristics of user u i in the short time span, represents the audience preference characteristics of paper p i in the short time span; represents the reading preference characteristics of user u i in the long time span, represents the audience preference characteristics of paper p i in the long time span, dG is the embedding dimension of the preference feature.

[0093] Step 3: Mining the dynamics of features under multiple time spans. Based on LSTM, referring to Figure 5 the designed cfLSTM predictor shown in and is used to mine the dynamics of features under multiple time spans. By inputting the user research direction features under the long time span and short time span obtained in Step 2 into the predictor

[0094]

[0095] …

[0096]

[0097] in is the output state of the l-th cfLSTM cell, retains some data features of the l-th cfLSTM cell and is input as the input for calculation in the next cfLSTM cell; and are initialized to 0, and it is considered that no papers are published in year YL[l]. Then, by default, the user research direction features under the short time span are the same as those of the previous year, that is If the user has not published papers within m years in and before year YL[l], then by default, the features of the user's research direction under the long time span are also the same as those of the previous year, that is If l = 0, then the user research direction features under both the long time span and the short time span are 0. For cfLSTM l there is the following definition:

[0098]

[0099]

[0100]

[0101] where the operation represents concatenation, and ⊙ represents the Hadamard product, are learnable parameters, and z l , f l , o l are the input gate, forget gate, and output gate of the l-th cfLSTM cell respectively.

[0102] During the process, the predicted user research direction features for each year and real user research direction characteristics The cosine similarity between them is used to define the loss function loss2 of the user research direction feature:

[0103] A1=|U|×|YL|,

[0104]

[0105] Where |U| represents the number of users.

[0106] Similarly, we can obtain the loss function loss3 of the user reading preference feature and the loss function loss4 of the paper audience preference feature:

[0107] A2=|U|×|YL|, A3=|P|×|YL|,

[0108]

[0109]

[0110] in

[0111] Step 4: Probability prediction, dot product of the paper content features and the user research direction features predicted in step 3 Get the degree of match between the user's research direction and the paper content The matching degree under the instant effect; the dot product is the paper audience preference characteristics predicted in step 3 and predicted user reading preference characteristics Get the degree of match between user reading preferences and paper audience preferences That is, the preference matching score; the probability of the recommended paper being read is obtained by linearly fusing the timeliness score and the preference matching score

[0112]

[0113]

[0114]

[0115] Among them, α is the weight parameter learned by the model, It is a comprehensive consideration and The final reading prediction value obtained.

[0116] For the predicted value of the reading probability, a training task is designed to train it, and a training set D is constructed in the form of triples, namely (u i ,rij , p j ), where r ij ∈ {0, 1}, when r ij = 0, it means that user u i has not read paper p j , when r ij = 1, it means that user u i has read paper p j ;

[0117] Define the loss function as:

[0118]

[0119] By minimizing the objective function Θ, the papers recommended to users are made time-sensitive. Θ is defined as follows:

[0120] Θ = loss5 + β(loss4 + loss3 + Loss2) + λL2

[0121] Among them, β and λ are hyperparameters set manually, and L2 is a regularization term used to prevent the model from overfitting.

[0122] To verify the effectiveness of the present invention, simulation experiments are carried out. The experimental settings are as follows:

[0123] As shown in Table 1, the datasets DBLP and OAG are selected to construct the academic HIN graph. For the dataset DBLP, papers in the "Computer Graphics" field with complete information from 1996 to 2005 are selected, and the reading records of the first nine years are used as the training set, and the reading records of the last year are used as the test set for paper reading recommendation. For the dataset OAG, papers with complete information from 2010 to 2015 are selected, and keywords with less than 10 paper citations and users with less than 5 paper publications are deleted. 20% of the reading records of each user in 2015 are randomly selected as the test set, and the remaining reading data are used as the training set. During the training process, 1 negative sample is randomly generated for each positive sample.

[0124] Table 1 Statistical table of dataset information

[0125]

[0126] In addition, three classic evaluation metrics, F1-score@K, AUC, and NDCG@K, are selected to evaluate the performance of this recommendation method.

[0127] In the experiment, the value of K is taken as 5, indicating that 5 papers are recommended to each user. In the pre-training task of step 2, the final embedding dimension d of the keywords is set in this embodiment K= 32, the training batch size is 1000, the learning rate is 0.002, and the total number of training epochs is 200. In the training of the main task, in this embodiment, the feature dimension d after GNN embedding G is set to 32, and the training batch size is set to 500, and the learning rate is set to 0.0006. For the three features that will change dynamically, namely: the research direction of the user, the reading preference of the user, and the audience preference of the paper, this embodiment uses their most suitable long - term span values for experiments respectively. In addition, in this embodiment, the initial value of the weight parameter α learned by the model is set to 0.5, and β and λ are set to 0.1 and 0.001 respectively in the objective function Θ. Finally, this embodiment sets the total number of training epochs to 50. This embodiment selects the model with the best performance on the training set for testing.

[0128] This embodiment selects several representative baselines HERec, ACKRec, SMIN, DYHATR to compare with the prediction results of this recommended method. In addition, several variants of MDRec are implemented in the process:

[0129] a) Select three widely used graph neural network embedding methods, GraphSage, HAN, and RGCN, to replace GNN (*) respectively to obtain three new methods, MDRec_GS, MDRec_HN, and MDRec_RN, through dot - product embedding, where "GS" represents GraphSage, "HN" represents HAN, and "RN" represents RGCN.

[0130] b) DC: Only recommend based on the timeliness score obtained by matching the paper content with the user's research direction, that is, the parameter weight α learned in the reading probability = 1.

[0131] c) MDRec_GNN_RA: Only recommend based on the recommendation score obtained by matching the paper audience with the user's reading preference, that is, the parameter weight α learned in the reading probability = 0.

[0132] The experimental results and their comparisons involved in this embodiment are shown in Table 2, and it can be obtained that:

[0133] Table 2 Experimental comparison results with different baselines under three metrics

[0134]

[0135] MDRec_RN outperforms all other baselines on two datasets. The improvement is even higher on the full-domain dataset OAG. This validates the effectiveness of MDRec_RN on datasets of any scale. The performance of all MDRec_GNN_RA methods is better than that of the original GNN method, indicating that if the dynamics of features are ignored and node features are directly learned from the entire historical information, the resulting recommendation results are not accurate enough. The performance of all MDRec_GNN_RA methods is better than that of the baseline DYHATR on two datasets, demonstrating that capturing the dynamic changes of features from the perspective of multiple time spans is superior to that of a single time span, indicating the superiority of the proposed method for capturing the dynamic changes of features over multiple time spans. The performance of all MDRec_GNN methods is better than that of MDRec_GNN_RA, which shows that the method of adding auxiliary training tasks is better than the original method, indicating that the auxiliary tasks proposed in this method are effective in improving the recommendation performance.

[0136] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0137] The present application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices produce means for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0138] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0139] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions for implementing the steps of the process Figure 1 one process or multiple processes and / or blocks Figure 1 steps for the functions specified in one block or multiple blocks.

[0140] Obviously, the above embodiments are only examples for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all implementation manners here. And the obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A personalized paper recommendation method based on time-sensitive auxiliary task-driven, characterized in that: A multi-task recommendation method MDRec with dynamic perception of time span is proposed, including the following steps: Step 1: Construct an academic heterogeneous information network graph, model the data set as the academic heterogeneous information network graph, which includes four types of nodes: users, papers, publication locations, and keywords, as well as relationships between nodes such as writing relationships, reading relationships, citation relationships, subordination relationships, and publication relationships; through the academic heterogeneous information network subgraph extraction function, from the perspective of a preset time period, obtain the academic heterogeneous information network subgraph under a long time span, and through the extraction function, obtain the academic heterogeneous information network subgraph under a short time span from the perspective of a preset year; Step 2: Heterogeneous entity feature representation learning. According to the academic heterogeneous information network graph, first use the classical language model BERT and add a fully connected layer to its output to learn keyword semantic features, and then aggregate the keyword semantic features in each paper through the self-attention mechanism to obtain paper content features; to capture the dynamic changes of user research direction features, aggregate the paper content features published by users under long and short time spans through the self-attention mechanism respectively to obtain user research direction features under long and short time spans; embed the GNN network in the academic heterogeneous information network subgraphs under long and short time spans respectively to capture the paper audience preference features and user reading preference features under long and short time spans; Step 3: Dynamic mining of features under multiple time spans. Based on LSTM, in the input of the LSTM, add the user research direction features, paper audience preference features, and user reading preference features learned in Step 2 to construct a predictor corresponding to the predicted features. By inputting the user research direction features, audience preference features, and user reading preference features under long and short time spans into the corresponding predictors respectively, obtain the predicted user research direction features, predicted paper audience preference features, and predicted user reading preference features; Step 4: Probability prediction. Multiply the paper content features and the user research direction features predicted in Step 3 to obtain the matching degree between the user's research direction and the paper content, that is, the matching degree under timeliness; multiply the paper audience preference features predicted in Step 3 and the predicted user reading preference features to obtain the matching degree between the user reading preference and the paper audience preference, that is, the preference matching degree score; linearly fuse the timeliness score and the preference matching degree score to obtain the probability that the recommended paper will be read.

2. The personalized paper recommendation method based on timeliness-assisted task-driven according to claim 1, wherein: Constructing the academic heterogeneous information network subgraph in Step 1 includes: Arrange all the years involved in the data set in ascending order into a set YL, that is, YL = [y0, y1, y2,...]; Use the academic heterogeneous information network subgraph extraction function SubHIN(HIN, Year) to extract the academic heterogeneous information network subgraph at the l-th year, i.e., under a short time span, as G l = SubHIN(G, YL[l]), where l ∈ [0, |YL|), and |YL| is the length of the set YL; Extract the academic heterogeneous information network subgraph over the past m years, i.e., a long time span, as G sub_l = SubHIN(G, YL[max(l - m, 0):l]), where YL[max(l - m, 0):l], m ∈ [0, |YL|) is used to control the long time span level.

3. The personalized paper recommendation method based on a timeliness-assisted task-driven according to claim 2, wherein: To learn the content features of the paper in step 2, define the set KL consisting of all keywords involved in the dataset. Add a fully connected layer to the output of the classical language model BERT, output the set KL to the language model BERT, and obtain the keyword k i semantic features of where KL = [k0, k1, k2, …], represents the learnable parameter matrix of the fully connected layer; σ is a non-linear activation function; d BERT = 768 is the output dimension of BERT, and d K is the final embedding dimension of the keyword; is the semantic feature representation of the keyword k i ; To make the keyword semantic features output by the classical language model BERT after adding a fully connected layer more suitable for the scenarios recommended by the paper, a pre-training task is designed to train the fully connected layer: construct a keyword semantic training set in the form of triples where When it means that keyword k i and keyword k j are discussed in the same paper, otherwise the keywords do not exist in the same paper; define the pre-training loss function as: where |D K | represents the number of elements in set D K , represents the predicted keyword k i and the correlation between keyword k j ; According to the above step 2, the content features of the paper can be obtained by aggregating the semantic features of the keywords of the paper through the self-attention mechanism. Define as the keyword set of paper p i , and as the content features of the paper, that is: Among them is the number of keywords of paper p i , SA(*) represents the self-attention mechanism, and its definition is as follows: wherein is the input vector, respectively represent the learnable query weight parameter matrix and the keyword weight parameter matrix.

4. A personalized paper recommendation method based on a timeliness-assisted task-driven approach according to claim 3, characterized in that: To obtain the dynamic changes in the user's research direction according to step 2, first collect user u i Papers published in a short time span And papers published in a long time span Through the self-attention mechanism, fuse the content features of the user's papers in the two time spans respectively to obtain user u i The user's research direction features in a short time span And user u i The user's research direction features in a long time span That is 5. A personalized paper recommendation method based on a timeliness-assisted task-driven approach according to claim 4, characterized in that: According to Step 2, to effectively capture the changes in paper audience preference features and user reading preference features, embed the academic heterogeneous information network subgraphs under long and short time spans into the GNN network to obtain: Among them represents the reading preference characteristics of user u i under a short time span, represents the audience preference characteristics of paper p i under a short time span; represents the reading preference characteristics of user u i under a long time span, represents the audience preference characteristics of paper p i under a long time span, and d G is the embedding dimension of the preference characteristics.

6. The personalized paper recommendation method based on a timeliness-assisted task-driven according to claim 5, wherein: To achieve the dynamic feature mining under multiple time spans in Step 3, the predictor in Step 3 is designed as a cfLSTM predictor, and user u is input into the cfLSTM predictor i Features of the research direction under the short time span of the current year And features of the research direction under the long time span of the current year It can be predicted that user u i Research direction features in year YL[l + 1] … Among them is the output state of the l-th cfLSTM cell, retains some data features of the l-th cfLSTM cell and is passed as input to the next cfLSTM cell for calculation; and is initialized to 0, and it is considered that no papers were published in year YL[l]. Then, by default, the user's research direction feature in the short time span is the same as that of the previous year, that is If the user has not published papers within m years in and before year YL[l], then by default, the user's research direction feature in the long time span is also the same as that of the previous year, that is If l = 0, then the user's research direction features in both the long time span and the short time span are 0; cfLSTM l The formula is as follows: Among them, the operation represents concatenation, and ⊙ represents the Hadamard product. are learnable parameters, z l , f l , o l are the input gate, forget gate, and output gate of the l-th cfLSTM cell, respectively.

7. A personalized paper recommendation method based on a timeliness-assisted task-driven approach according to claim 6, characterized in that: In the cfLSTM predictor, the cosine similarity between the predicted user research direction features for each year and the true user research direction features is used to define the loss function loss2 of the user research direction features: A1 = |U| × |YL|, where |U| represents the number of users.

8. A personalized paper recommendation method based on a timeliness-assisted task-driven approach according to claim 7, characterized in that: Define the predicted value of the reading preference characteristics of user u obtained by the cfLSTM predictor i in year YL[l] as paper p j the predicted value of the audience preference characteristics in year YL[l] as Use the cosine similarity between the predicted reading preference characteristics of the user and the true reading preference characteristics of the user to define the loss function loss3 of the reading preference characteristics of the user; use the cosine similarity between the predicted audience preference characteristics of the paper and the true audience preference characteristics of the paper to define the loss function loss4 of the audience preference characteristics of the paper: A2 = |U| × |YL|, A3 = |P| × |YL|, Among them 9. The personalized paper recommendation method based on a timeliness-assisted task-driven according to claim 8, characterized in that: According to step 4, predict user u i The probability of reading paper p in year YL[l] is j as follows In year YL[l], the predicted value of the user's research direction feature obtained by the cfLSTM predictor is The predicted value of the reading preference feature is and for paper p j The predicted value of the audience preference feature is The paper content feature obtained by aggregating keyword semantic features through the attention mechanism is Define the reading probability: Among them, the value represents the matching degree between the research direction of user u i and the content of paper p j , that is, the matching degree under timeliness, the value represents the matching degree between the reading preference of user u i and the audience preference of paper p j , that is, the preference matching degree score, and α is the weight parameter learned by the model, is obtained by comprehensively considering and to get the final reading prediction value.

10. A personalized paper recommendation method based on a timeliness-assisted task-driven approach according to claim 9, characterized in that: For the predicted value of the reading probability, design a training task to train it, and construct a training set D in the form of triples, i.e., (u i , r ij , p j ), where r ij ∈{0, 1}. When r ij = 0, it means that user u i has not read paper p j . When r ij = 1, it means that user u i has read paper p j ; Define the loss function as follows: By minimizing the objective function Θ, the papers recommended to users are made time-sensitive, and Θ is defined as follows: Θ = loss5 + β(loss4 + loss3 + loss2) + λL2 where β and λ are hyperparameters set manually, and L2 is a regularization term used to prevent the model from overfitting.

Images