Construction method and application of uncertainty-aware agent structure completion model for sparse academic paper recommendation

Abstract

The application provides a construction method and application of an uncertainty-aware agent structure completion model for sparse academic paper recommendation, learns basic collaborative representation of users and papers based on user-paper interaction graphs; for low-interaction papers, constructs double-agent neighbor representation by using semantic similarity relationship and common author relationship, and performs confidence mixing and gate update according to source confidence characteristics; subsequently, generates candidate paper representation through multiple probability experts, estimates sample-level uncertainty among experts, and controls fusion of agent-enhanced information and original collaborative information by using a trust coefficient; adopts pair-wise ranking, local consistency, cross-view comparison and uncertainty-aware auxiliary loss joint training, and finally outputs Top-N recommendation results. The application supplements local academic association of low-interaction papers without changing the topology of the original interaction graph, and suppresses noise propagation caused by unreliable auxiliary relationships.

Description

Technical Field

[0001] This invention relates to the fields of academic literature data processing, information retrieval, recommendation systems, and graph representation learning, and particularly to a method and application for constructing an uncertainty-aware agent structure completion model for sparse academic paper recommendations, addressing problems such as sparse user-paper interaction graphs, paper cold start, and noisy auxiliary academic relationships. Background Technology

[0002] Academic paper recommendation systems typically construct a user-paper bipartite interaction graph based on implicit feedback such as user reading, saving, downloading, or clicking. They then use collaborative filtering or graph neural networks to obtain user and paper representations, and finally output the Top-N papers based on the matching scores between the user and paper representations. For newly published, long-tail, or low-interaction papers, the training graph has few or even zero observable interactive neighbors, making it difficult to obtain stable collaborative signals during graph propagation.

[0003] Paper similarity and co-authorship relationships can provide supplementary academic connections, but these relationships are not always reliable. Directly injecting representations of similar or co-authored papers into the target paper can easily propagate weak correlations, erroneous associations, or noisy neighbors to the recommended representation. Conventional graph contrastive learning mainly constructs perturbation views on existing interaction graphs, and cannot explicitly recover the missing local structures of low-interaction papers; fixed-weight fusion cannot distinguish the reliability of proxy information from different papers. Therefore, existing solutions have at least the following problems: 1. Papers with low interaction and cold start characteristics lack available collaborative neighbors, indicating insufficient quality; 2. The reliability of auxiliary similarity relationships and co-authorship relationships differs, and direct fusion will produce noise propagation; 3. Fixed fusion or single prediction branches cannot express the uncertainty of proxy representations at the paper level; 4. When training with only ranking loss, there is a lack of local consistency constraints between the completed structure and the final representation, which limits the ability to resist sparsity and noise.

[0004] Therefore, there is an urgent need for an academic paper recommendation method that can supplement reliable local structures for low-interaction papers without modifying the original user-paper interaction graph topology, and suppress noise propagation based on the reliability of auxiliary academic relationships and sample-level uncertainty, so as to improve the paper representation and ranking performance under sparse interaction conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a method and application for constructing an uncertainty-aware proxy structure completion model for sparse academic paper recommendations. Based on preserving the original user-paper interaction graph topology, it constructs proxy neighborhoods for low-interaction papers using semantic similarity relationships and co-author relationships. Furthermore, it suppresses the influence of unreliable auxiliary relationships through source confidence control, gating updates, and multi-expert uncertainty estimation, thereby forming a complete recommendation process consisting of proxy structure completion, reliable fusion, uncertainty suppression, joint optimization, and preference ranking.

[0006] To achieve the above objectives, this technical solution provides a method for constructing an uncertainty-aware proxy structure completion model for sparse academic paper recommendation, including the following steps: S1: Obtain the training dataset, which includes a user-paper bipartite interaction graph, semantic similarity relationships between papers, and co-author relationships. Papers with an interaction degree not exceeding a preset threshold are marked as low-interaction papers. S2: Construct an uncertainty-aware agent structure completion model, including: The basic collaborative representation module is used to process the user-paper bipartite interaction graph to obtain the user basic collaborative representation and the paper basic collaborative representation. A dual-proxy structure completion module is used to process low-interaction papers and obtain semantically similar proxy representations, co-author proxy representations and corresponding source confidence features; The confidence fusion and gating update module is used to perform confidence-driven fusion of two types of proxy representations based on source confidence features to obtain a completed proxy representation, and then combine the completed proxy representation with the paper-based collaborative representation through gating update to obtain a proxy enhancement representation; The multi-expert uncertainty fusion module is used to generate candidate paper representations through multiple probabilistic experts, determine sample-level uncertainty based on expert disagreement, generate a confidence coefficient accordingly, and weight and fuse the candidate paper representations with the basic collaborative representation of the paper to obtain the final paper representation. S3: Train an uncertainty-aware proxy structure completion model using a training dataset containing low-interaction papers. The uncertainty-aware proxy structure completion model uses pairwise ranking loss, local consistency loss, cross-view comparison loss, and multi-expert assistance loss to jointly optimize the model parameters.

[0007] Secondly, this solution provides a paper recommendation method for sparse academic papers, including the following steps: Step 1: Obtain the target user, the candidate paper set, the user-paper bipartite interaction graph constructed based on the observed interactions during the training phase, and the semantic similarity relationships and co-author relationships of the candidate papers; label low-interaction candidate papers based on the interaction degree of the papers in the training interaction graph, including cold-start papers with an interaction degree of 0. Step 2: Input the user-paper bipartite interaction graph, target user, candidate paper and auxiliary relationship obtained in Step 1 into the uncertainty-aware proxy structure completion model constructed and trained according to the method described in the first aspect, and obtain the user-based collaborative representation of the target user and the final paper representation of the candidate paper. Step 3: Calculate the candidate paper preference score based on the user representation and the final paper representation. After removing the papers that the target user has already interacted with during the training phase, output the Top-N recommended paper list according to the preference score.

[0008] Compared with existing technologies, this technical solution has the following characteristics and beneficial effects: 1. This invention completes the representation layer structure of low-interaction papers by using semantically similar proxy neighbors and co-author proxy neighbors, supplementing local academic association information without modifying the original user-paper interaction graph topology.

[0009] 2. This invention uses source confidence-driven hybrid weights and original representation-proxy representation gated updates to adaptively fuse different proxy relationships based on their quality differences, avoiding the direct replacement of the original collaborative representation with auxiliary relationships.

[0010] 3. This invention constructs sample-level uncertainty by constructing divergence in multi-expert candidate representations and utilizes trust gates to control the contribution of proxy enhancement representations, thereby reducing the impact of weak correlations, erroneous associations, or noisy proxy neighbors on the recommended representation.

[0011] 4. This invention enhances the stability and robustness of low-interaction paper representations under sparse and perturbation conditions through cross-view contrastive learning with local consistency constraints and popularity adaptive perturbation. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a paper recommendation method for sparse academic papers according to an embodiment of this application; Figure 2 This is a framework diagram of an uncertainty-aware agent structure completion model according to an embodiment of this application; Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0013] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0014] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0015] Example 1 This solution provides a method for constructing an uncertainty-aware proxy structure completion model for sparse academic paper recommendation, including the following steps: S1: Obtain the training dataset, which includes a user-paper bipartite interaction graph, semantic similarity relationships between papers, and co-author relationships. Papers with an interaction degree not exceeding a preset threshold are marked as low-interaction papers. S2: Construct an uncertainty-aware agent structure completion model, including: The basic collaborative representation module is used to process the user-paper bipartite interaction graph to obtain the user basic collaborative representation and the paper basic collaborative representation. A dual-proxy structure completion module is used to process low-interaction papers and obtain semantically similar proxy representations, co-author proxy representations and corresponding source confidence features; The confidence fusion and gating update module is used to perform confidence-driven fusion of two types of proxy representations based on source confidence features to obtain a completed proxy representation, and then combine the completed proxy representation with the paper-based collaborative representation through gating update to obtain a proxy enhancement representation; The multi-expert uncertainty fusion module is used to generate candidate paper representations through multiple probabilistic experts, determine sample-level uncertainty based on expert disagreement, generate a confidence coefficient accordingly, and weight and fuse the candidate paper representations with the basic collaborative representation of the paper to obtain the final paper representation. S3: Train an uncertainty-aware proxy structure completion model using a training dataset containing low-interaction papers. The uncertainty-aware proxy structure completion model uses pairwise ranking loss, local consistency loss, cross-view comparison loss, and multi-expert assistance loss to jointly optimize the model parameters.

[0016] Regarding step S1 of this plan: This solution acquires user-paper interaction data and constructs a two-part user-paper interaction graph based on the user-paper interaction data. The two-part user-paper interaction graph records the interaction relationship between users and papers.

[0017] In some embodiments, user-paper interaction data includes positive interactions between the user and the paper, such as reading, clicking, saving, downloading, citing, or other pre-defined interactions.

[0018] For example, user-paper interaction data Represented as ,in For user, p represents paper; the user-paper bipartite interaction diagram is represented as follows: , where U is the set of users and P is the set of papers.

[0019] In some embodiments, the semantic similarity relationship between papers can be calculated from the semantic representation of the paper title, abstract, keywords or full text, or obtained from an established paper similarity relationship database; the co-author relationship is determined based on the co-authors, author overlap or author collaboration between the author sets of two papers.

[0020] For each paper, the auxiliary relation input includes at least candidate neighbor identifiers and original relation scores; the source confidence features are calculated by the subsequent dual-agent structure completion module based on the effective neighbor coverage and relation quality.

[0021] Specifically, for each low-interaction paper, a candidate neighbor set excluding itself is formed based on semantic similarity and co-authorship, and its identifier and original score are recorded. The neighbor number limit, threshold filtering, and fallback rules for each source are all executed according to the dual-agent structure completion module described later; the same rule is used in both the training and inference phases.

[0022] The relationship scores of retained candidate neighbors are processed into non-negative aggregate scores and normalized to obtain surrogate neighbor weights. Based on the surrogate neighbor identifiers, the basic collaborative representations of the corresponding papers are read and aggregated according to the surrogate neighbor weights to obtain semantically similar surrogate representations and co-author surrogate representations. The source confidence feature is calculated using relationship scores scaled to [0,1] to generate surrogate source hybrid weights; this quality score differs from the normalized aggregate weights.

[0023] In some embodiments, the interaction degree of each paper is statistically analyzed based on the user-paper bipartite interaction graph, and papers with an interaction degree not greater than a preset threshold are marked as low-interaction papers. In this scheme, low-interaction papers are determined based on the interaction degree of the paper in the user-paper bipartite interaction graph and the preset threshold; that is, papers with an interaction degree lower than the preset threshold are taken as low-interaction papers.

[0024] For example, statistical papers Interaction degree in the user-thesis part two interaction diagram .when At that time, paper p was classified as a low-interaction paper; among which Low-interaction papers are cold-start papers. This is a preset threshold or a preset threshold determined based on historical interaction data statistics.

[0025] In this embodiment, let That is, only the interaction degree is taken. The cold-start paper is treated as a low-interaction paper, and proxy structure completion, confidence fusion, and uncertainty estimation are performed on the low-interaction paper; when configured to process general low-interaction papers, it can... Set to a preset value greater than 0, and satisfy... The same processing is applied to the papers.

[0026] Regarding the basic collaborative representation module of this scheme: The basic collaborative representation module in the uncertainty-aware proxy structure completion model of this scheme performs multi-layer normalized neighbor propagation on the user-paper bipartite interaction graph to obtain the user basic collaborative representation and the paper basic collaborative representation.

[0027] Specifically, for each user, obtain the set of interactive paper neighbors for that user in the user-paper bipartite interaction graph; for each paper, obtain the set of interactive user neighbors for that paper in the user-paper bipartite interaction graph; for each layer of users, update the user collaborative representation by combining the set of interactive user neighbors and the set of interactive paper neighbors; for each layer of papers, update the paper collaborative representation by combining the set of interactive user neighbors and the set of interactive paper neighbors; summarize the user collaborative representations of all layers to obtain the basic user collaborative representation, and summarize the paper collaborative representations of all layers to obtain the basic paper collaborative representation.

[0028] For example, user-based collaboration is represented as follows:

[0029]

[0030] in For user-based collaborative representation, L represents the number of propagation layers. This represents the user collaboration representation at layer l. This represents the user collaboration representation at layer l+1. For users Interactive paper neighbor set, For the thesis The set of interactive user neighbors.

[0031] The basic co-representation of the paper is as follows:

[0032]

[0033] in This is the basic collaborative representation of the paper, where L represents the number of propagation layers. This represents the collaborative representation of the papers at layer l. For the collaborative representation of the papers in layer l+1, For users Interactive paper neighbor set, For the thesis The set of interactive user neighbors.

[0034] It should be noted that the basic collaborative representation module of this scheme performs one or more layers of normalized neighbor propagation without changing the topology of the original user-paper bipartite interaction graph. It also does not change the topology of the original bipartite graph and aggregates the propagation results of each layer to obtain the user basic collaborative representation and the paper basic collaborative representation, providing a benchmark representation for subsequent proxy structure completion.

[0035] like Figure 2 As shown, regarding the dual-proxy structure completion module of this solution: The dual-proxy structure completion module of this scheme is used for low-interaction papers. It selects proxy neighbors from semantic similarity relations and co-author relations respectively, and performs weighted aggregation on the paper basic collaborative representation of the proxy neighbors to obtain semantic similarity proxy representation, co-author proxy representation and corresponding source confidence features.

[0036] Specifically, the dual-proxy structure completion module of this scheme includes a semantic similarity proxy branch and a co-author proxy branch. The semantic similarity proxy branch is for low-interaction papers, selects semantically similar proxy neighbors based on semantic similarity relationships, and performs weighted aggregation on the basic collaborative representations of the papers of the semantically similar proxy neighbors to obtain semantically similar proxy representations and corresponding source confidence features. The co-author proxy branch is for low-interaction papers, selects co-author proxy neighbors based on co-author relationships, and performs weighted aggregation on the basic collaborative representations of the papers of the co-author proxy neighbors to obtain co-author proxy representations and corresponding source confidence features.

[0037] Define a proxy relationship source index r∈{s,a}, where s represents semantically similar sources and a represents sources with co-authorship. Let... Let r be the set of valid proxy neighbors reserved for paper p (when r=s, i.e., the set of semantically similar proxy neighbors). When r=a, it represents the set of co-authors' agent neighbors. , The paper's foundational collaborative representation for the proxy neighbor q, The set represents the original relation score between paper p and its candidate neighbor q under source r. It also serves as the summation domain for the following weight normalization, weighted aggregation, and confidence feature statistics.

[0038] For a low-interaction paper p, candidate neighbors for semantically similar sources and co-author sources are first determined based on the original relation scores. The candidate neighbors that are selected and retained are used as surrogate neighbors for the corresponding sources. The surrogate neighbors selected from the semantically similar relations are defined as semantically similar surrogate neighbors, and the surrogate neighbors selected from the co-author relations are defined as co-author surrogate neighbors.

[0039] Specifically, for each low-interaction paper: the semantic similarity proxy branch retains the top Ks semantically similar candidate neighbors based on their semantic similarity relation scores. Then, it applies a similarity threshold filter to the original relation scores of the candidate neighbors before normalization, and uses the filtered neighbors as semantically similar proxy neighbors to form a set of semantically similar proxy neighbors. When all candidate neighbors are filtered, the candidate neighbor with the highest original relation score is retained as the fallback neighbor to avoid the aggregation weight being zero. The co-author proxy branch retains the top Ka candidate neighbors based on their co-author relation scores as co-author proxy neighbors, forming a co-author proxy neighbor set. Threshold comparisons are always based on the original relation scores before normalization.

[0040] Specifically, for any proxy relation source r∈{s,a}, the original relation scores of the source proxy neighbors of that source are normalized into proxy neighbor weights:

[0041] in To find the weights of the proxy neighbors q of paper p from source r, sum the denominators and iterate through the same set of valid proxy neighbors. ε is a smoothing term to prevent the denominator from being zero. When all semantically similar candidate neighbors are filtered out and only backoff neighbors are retained, If only the fallback neighbor is included, its proxy neighbor weight is 1.

[0042] Semantic similarity proxies are represented as:

[0043] Co-author representation is indicated as follows:

[0044] in, For semantically similar proxy representation, For proxy neighbor weights based on semantic similarity, The paper provides the underlying collaborative representation for semantically similar proxy neighbors q; As a co-author, he / she indicated that For proxy neighbor weights under co-author sources, The underlying collaborative representation of papers for co-authors' proxy neighbors q.

[0045] In addition, each source corresponds to a source confidence feature. Specifically, for a source r∈{s,a}, where s is a semantically similar source and a is a co-author source, the original relation score, after validity filtering (i.e., the neighbor retention process for each source: threshold filtering and backtracking for semantically similar sources, and retention based on score for co-author sources), is scaled to the interval [0,1] (e.g., linear scaling, where the original score itself is within [0,1] can be directly used), denoted as […]. Let Kr be the maximum number of proxy neighbors preset for source r. The source confidence feature is calculated as follows: (The set of valid proxy neighbors reserved for this source is then used to calculate the source confidence feature.)

[0046] like If it is empty, then let = 0.

[0047] This source confidence feature can simultaneously reflect effective neighbor coverage and average relationship quality, and is pre-cached along with the proxy neighbor index and proxy neighbor weight.

[0048] Furthermore, to avoid training the basic collaborative representation of low-interaction papers with inverse perturbation of the proxy neighbors (i.e., neighbor papers used as references), the basic collaborative representation of the aggregated proxy neighbor q is used. At that time, Gradient truncation is performed, meaning that the gradient of the structure completion-related loss is not backpropagated to the representation branch corresponding to the surrogate neighbor. This process focuses the updates of the completion branch on the target low-interaction paper and the parameters of the completion module itself.

[0049] Regarding the confidence fusion and gating update module: The confidence fusion and gating update module of this scheme is used to generate a hybrid weight of proxy sources based on the source confidence features. Based on the hybrid weight of proxy sources, the semantically similar proxy representation and the co-author proxy representation are fused with confidence to obtain the completed proxy representation. Then, the completed proxy representation is combined with the paper-based collaborative representation through gating update to obtain the proxy enhancement representation.

[0050] In some embodiments, source confidence features from semantically similar sources and co-author sources are concatenated to obtain concatenated features. These concatenated features are then input into a mixing function to generate proxy source mixing weights, as shown in the following formula:

[0051] in It is a mixed function. Source confidence features for co-authored sources Source confidence features for semantically similar sources. This is the Sigmoid function, where || represents vector concatenation.

[0052] In some embodiments, the hybrid function may be implemented by a multilayer perceptron comprising a fully connected layer, a ReLU activation layer, and a fully connected layer in sequence.

[0053] In some embodiments, a confidence-driven fusion is performed on the semantically similar proxy representation and the co-author proxy representation based on the proxy source hybrid weight to obtain the complete proxy representation:

[0054] in For proxy source mixed weights, For semantically similar proxy representation, This is a statement made on behalf of co-authors.

[0055] It should be noted that when a proxy representation from one proxy source is unavailable, a proxy representation from another available source is directly used. That is, if a semantically similar proxy representation is unavailable, the co-author proxy representation is directly used as the completed proxy representation; if the co-author proxy representation is unavailable, the semantically similar proxy representation is directly used as the completed proxy representation.

[0056] Subsequently, the original basic collaborative representation and the completed surrogate representation are concatenated to generate a gating vector with the same dimension as the embedding dimension. Then, the surrogate augmentation representation is generated based on the gating vector through gating update, as follows:

[0057] in For gating network parameters, For element-wise multiplication, This is a random deactivation operation. This indicates an enhancement to the agent after the gating update. For the basic collaborative representation of the paper, To complete the proxy representation, This is the gated vector.

[0058] It should be noted that the gating vector Each dimension independently controls the contribution of the original collaborative representation and the proxy structure representation: When the size is larger, more original collaborative semantics are preserved. When the size is smaller, more surrogate structural semantics are introduced; that is, dimensional gating vectors are used to control the contribution ratio of the original collaborative information and the surrogate structural information. Therefore, this step does not directly overwrite the original collaborative representation with the surrogate representation.

[0059] Regarding the multi-expert uncertainty fusion module: The multi-expert uncertainty fusion module of this scheme is used to generate candidate paper representations by taking the paper's basic collaborative representation and side information containing at least source confidence features and paper activity information as input, and using multiple probabilistic experts with mutually independent parameters. The mean variance of the probabilistic experts' outputs on the embedding dimension is used as the sample-level uncertainty, and a trust coefficient is generated based on the sample-level uncertainty. The candidate paper representations and the paper's basic collaborative representations are weighted and fused based on the trust coefficients to obtain the final paper representation.

[0060] In some embodiments, the side information is a concatenated vector of source confidence features, completed proxy representation, and paper activity information, represented as follows:

[0061] in This represents the completed proxy representation after the merging of the two proxies. These are the source confidence features for sources with similar origins and sources with co-authors, respectively. This refers to information on the activity level of papers.

[0062] The formula for calculating paper activity information is as follows:

[0063] in This is a thesis Interaction level in the training interaction graph This is a collection of papers. any paper Training interactivity This represents the maximum interaction degree within the set.

[0064] It should be noted that the proxy enhancement representation obtained from the gating update is not used as input to this module, but is used as the anchor point for subsequent local consistency loss. This module uses the paper-based collaborative representation as the residual update benchmark and the aforementioned side information as conditional input to implement residual correction, so that the expert output always retains the original collaborative semantics.

[0065] Furthermore, multiple probabilistic experts with mutually independent parameters generate a Gaussian distribution of latent variables using side information. During the training phase, reparameterization sampling is performed based on the Gaussian distribution of latent variables. The sampled latent variables are mapped to residual corrections through their decoding network, and the candidate paper representation is obtained using the paper-based collaborative representation as the residual benchmark.

[0066] It should be noted that during the inference phase, the mean path of the Gaussian distribution of the latent variable is used as the latent variable to reduce the ranking fluctuations caused by random sampling. The resampling is based on the mean and log-variance.

[0067] Specifically, multiple probability experts with independent parameters generate the mean and log-variance of a Gaussian distribution of latent variables using lateral information, denoted as...

[0068] in Let be the mean of the Gaussian distribution of the latent variables of the m-th probability expert. Let be the variance of the Gaussian distribution of the latent variables of the m-th probability expert. This is side information.

[0069] Reparameterized sampling is used during the training phase:

[0070]

[0071] The mean path is used in the inference phase:

[0072] in As a latent variable, Let m be a random number, and m be the m-th probability expert.

[0073] Using the basic collaborative representation of the paper as the residual benchmark, and combining it with residual correction, candidate paper representations are generated, as follows:

[0074]

[0075] in For the basic collaborative representation of the paper, For residual correction, As a latent variable, For side information, The candidate experts stated that For the first A decoding network of experts.

[0076] As described above, each probabilistic expert in this scheme uses the original paper-based collaborative representation as the residual update benchmark, so that the expert output retains the original collaborative semantics.

[0077] For low-interaction papers , No. The candidate papers output by the probability experts are denoted as . All The average of the expert outputs is calculated to obtain the average candidate paper representation:

[0078] in, The average number of candidate papers represents The One dimension, Indicates the first The candidate papers output by the experts represent the papers submitted in the first... The values ​​are taken in each dimension. The mean variance of the expert output in each dimension is used as the data for the paper. Sample-level uncertainty:

[0079] in, The embedding dimension is represented in the paper. This represents the number of probabilistic experts with mutually independent parameters. It is calculated during the training phase from the expert outputs of the sampling paths. The inference phase is calculated using expert outputs from the mean path. Both use the same variance calculation rule.

[0080] The confidence coefficient is generated based on sample-level uncertainty and is expressed as follows:

[0081] in Trust coefficient For sample-level uncertainty, The temperature is uncertain.

[0082] Based on the trust coefficient-weighted fusion of candidate paper representation and basic paper collaborative representation, it is represented as follows:

[0083]

[0084] in For the final paper, This represents the average of the candidate papers represented by all probability experts. This serves as the foundational collaborative representation for the paper.

[0085] Furthermore, for papers with low interaction, the basic collaborative representation of the paper is directly taken as the final paper representation, i.e. Therefore, when sample-level uncertainty is high, the confidence coefficient... Reduce the multi-expert average candidate paper representation generated after completing the surrogate representation and source confidence feature. The weighting is reduced, while the basic collaborative representation of the paper is... The weight of the proxy information is increased accordingly, thereby reducing the impact of unreliable proxy information on the final paper representation and preserving more of the original collaborative semantics.

[0086] In one alternative implementation, the proportion of low-interaction papers among all papers is considered. With the preset ratio upper limit Determine the fusion adaptive coefficients The trust coefficient is updated based on the fusion adaptive coefficient. This process limits the overall impact of multi-expert fusion branches when the proportion of low-interaction papers is too high.

[0087] Specifically, when season ;when season and will Limited to between a preset lower limit and 1, where This is the attenuation coefficient.

[0088] In another alternative implementation, based on the trust coefficient Source confidence features combined with agent source Construct a security mask: When and In this case, multi-expert fusion representation is allowed; otherwise, the actual confidence coefficient of the sample is set to zero and the process is regressed to... . These are the preset minimum trust threshold and minimum agent confidence threshold, respectively.

[0089] The loss function of the uncertainty-aware proxy structure completion model in this scheme is a joint of pairwise ranking loss, local consistency loss, cross-view comparison loss, and multi-expert assistance loss, expressed as:

[0090] in For trainable embedding parameters, L2 regularization term; and The weights corresponding to the losses, For pairwise sorting loss, For local consistency alignment loss, For cross-view comparison loss, The loss was mitigated by multiple expert assistance.

[0091] in, The initial embedding squared norm loss without regularization coefficients, Its weight, not enabled by the timer ; , and The corresponding loss weights. This indicates the currently selected pairwise sorting loss. The basic implementation uses unweighted loss, while the optional implementation uses the weighted alternatives given below.

[0092] Local consistency loss is used to constrain the consistency between the agent-enhanced representation and the completed agent representation; cross-view contrast loss is obtained by constructing the original view and the perturbation view from the user representation and the paper representation, wherein the perturbation magnitude on the paper side is adaptively determined based on the paper activity information; multi-expert assistance loss includes at least reconstruction constraints, probability distribution regularization constraints and expert disagreement constraints.

[0093] Regarding the local consistency alignment loss: For low-interaction papers at each training phase The proxy enhancement representation obtained after gating update As the anchor point for local consistency alignment, the completed proxy representation obtained by fusing semantically similar proxy representations and co-author proxy representations will be used. This is used to represent positive samples.

[0094] Extract a local negative sample set from the training paper set. To avoid the proxy neighbors upon which the proxy representation depends being treated as negative samples simultaneously, the local negative sample set does not include the target paper. Semantically similar proxy neighbor set and co-author proxy neighbor set ,Right now:

[0095] Based on surrogate augmentation representation, complete surrogate representation, and local negative sample representation, calculate the normalized similarity of positive sample pairs:

[0096] in, Represents cosine similarity. This indicates the temperature parameter for local consistency alignment. Indicates negative sample papers The basic collaborative representation. The local consistency alignment loss is expressed as:

[0097] The former represents the set of low-interaction anchor point papers that participate in the calculation of local consistency loss. This represents the set of local negative samples sampled for each anchor paper (p). Normalized similarity The denominator is used, while the local consistency loss requires analysis of the corresponding papers at each anchor point. Find the average.

[0098] In one embodiment, the current low-interaction paper is first... Semantic similarity proxy neighbor set set of co-authors' agent neighbors The union of these sets is determined as the set of positive neighbors of the surrogate:

[0099] The papers in the surrogate positive neighbor set are used for weighted aggregation according to the surrogate neighbor weights to construct the surrogate representation. Therefore, it is not considered a negative sample for the local consistency alignment loss.

[0100] For the local consistency alignment loss, the target paper and the surrogate positive neighbor set are excluded from the training paper set to obtain a local negative sample candidate set:

[0101] From the local negative sample candidate set A predetermined number of papers are randomly selected to form a local negative sample set. If the number of candidate papers in the current batch is insufficient, additional samples can be collected from the training paper set according to the same exclusion rules.

[0102] Regarding cross-view comparison loss: Regarding the cross-view contrastive loss, in the current training batch, the user-based collaborative representation of the deduplicated user set and the paper-based collaborative representation of the deduplicated paper set are first obtained based on the same user-paper bipartite interaction graph. The user representation and paper representation without random perturbation are referred to as the user unperturbed representation view and the paper unperturbed representation view, respectively; the unperturbed representation view is not an additionally constructed interaction graph, but rather the representation of nodes in the same user-paper bipartite interaction graph obtained without perturbation.

[0103] Random perturbations are applied to both the user-based collaborative representation and the paper-based collaborative representation to obtain user-perturbed representation views and paper-perturbed representation views. Cross-view contrastive learning is used to constrain the representation of the same user in the unperturbed user representation view and the user-perturbed representation view to be consistent, and to constrain the representation of the same paper in the unperturbed paper representation view and the paper-perturbed representation view to be consistent.

[0104] To improve the stability of user and paper representations under perturbation conditions, original and perturbed views are constructed for the deduplicated user and paper sets in the current training batch, respectively. For any representation vector... First, generate random vectors. Then perform L2 normalization and reconstruct the perturbation view, as follows:

[0105] User node take , ,in To maintain a fixed perturbation amplitude; the paper's nodes are taken as follows: , Its magnitude is based on the same normalized activity. Adaptive determination:

[0106] in, These represent the lower and upper limits of the perturbation, respectively. The lower the paper activity, the larger the perturbation amplitude. Resampling is performed when the random vector norm is 0. The unperturbed and perturbed representations of users and papers are normalized row-wise, and bidirectional InfoNCE constraints are used to constrain the cross-view representations of nodes of the same type; when the row vector norm is 0, a positive smoothing term is used for normalization.

[0107]

[0108]

[0109]

[0110] in, For node type, and Each is obtained by stacking the basic collaborative representation and its perturbation representation of the current batch of deduplicated users in the same order; and These are obtained by stacking the basic collaborative representation and its perturbation representation from the current batch of deduplicated papers. The underlined matrix represents the row-normalized result. This is the similarity matrix between the two views. For temperature comparison.

[0111] Let the number of nodes be ,Label That is, the similarity matrix of the th The positive samples of the row are located in the first row. List. The cross-entropy is calculated by performing softmax on each row of the matrix, then calculating the in-batch mean based on the corresponding label; the two views of the same node are positive samples, and other nodes of the same type are negative samples. Weights are assigned to the user-side and paper-side losses. When the number of nodes for a certain type is less than 2, the contrastive loss for that type is set to zero.

[0112] In the basic implementation, a pairwise ranking loss is constructed using the current training triplet batch B.

[0113] In the implementation with L2 regularization enabled, the initial embeddings of the user, positive papers, and negative papers corresponding to the current batch are calculated as follows:

[0114] In the optional uncertainty reweighting implementation, for papers in the current batch that have already undergone multi-expert fusion, let ,in The lower limit of the weight; papers that did not undergo multi-expert fusion were ordered. .set up To preset the reweighting strength, the triplet weights are: When using this approach, the following weighted loss replaces the basic loss in the overall objective. It should not be added repeatedly to the unweighted item; It degenerates into an unweighted form.

[0115]

[0116] in, This is the set of non-empty training triples in the current batch. This represents the number of triples contained in the set. For users; To the user Positive sample papers selected from the observed interactive papers; For users selected from the training paper set Non-interactive negative sample papers; To train triplet The weights; and users respectively Prediction preference scores for positive and negative sample papers; .

[0117] It should be noted that there is sample-level uncertainty. This measure is used to gauge the degree of disagreement among multiple experts regarding the paper completion result, rather than to assess the learning value of the sample. Larger... This indicates that the candidate papers generated by the experts have low consistency, and the reliability of the surrogate completion results is relatively weak. In this case, appropriately reducing the weight of this sample in the pairwise ranking loss can avoid noise generated by the model forcibly fitting unreliable surrogate relationships. The lower bound of the weights satisfies... Therefore, samples with high uncertainty still retain non-zero ranking gradients. The lower bound of the triplet weights is not less than... High-uncertainty samples are not completely filtered out. Furthermore, this weighting only applies to pairwise ranking losses; high-uncertainty samples still participate in local consistency losses, cross-view comparison losses, and multi-expert assistance losses. Sample-level uncertainty is dynamically calculated during training; as paper representations gradually improve and expert disagreements decrease, their ranking weights increase accordingly.

[0118] Regarding losses assisted by multiple experts: The loss from multi-expert assistance is expressed as:

[0119]

[0120]

[0121]

[0122] in , , The weights corresponding to the losses, To reconstruct constraints, For KL regularization loss, Constraints on expert disagreement. For the basic collaborative representation of the paper, This represents the average of the candidate papers represented by all probability experts. To complete the proxy representation, This is the logarithm of the variance corresponding to the Gaussian distribution. Let be the mean of the Gaussian distribution of the latent variables of m probability experts. The dimension index of the potential representation vector.

[0123] That is, the multi-expert fusion module of this scheme also sets auxiliary constraints. First, the reconstruction constraint makes the average candidate representation close to both the original collaborative representation and the completed surrogate representation. KL regularization sets a standard Gaussian prior for the mean and variance of each probabilistic expert. Expert divergence constraint increases the disorder of expert divergence.

[0124] During the training phase, completion and uncertainty fusion are performed only on low-interaction papers that meet the conditions in the current batch; during the inference phase, all papers in the candidate set that meet the conditions are processed. Furthermore, forward computation is performed on papers with available proxy sources, the model parameters are fixed, and training loss is not calculated. The two stages share relation weights, source confidence rules, and trained module parameters; random sampling and mean paths are switched according to the aforementioned rules. Proxy neighbor indexes, relation weights, source confidence features, and training interaction degrees can be cached, but the proxy aggregation representation updated with model parameters is not permanently fixed as the preprocessed result.

[0125] Training phase completion objects satisfy ,in These papers have positive training interactions, allowing them to be supervised as positive samples for pairwise ranking; each batch still consists of training triples, rather than training the entire base recommendation model using only low-interaction papers. The actual sets entering each auxiliary loss are ranked... and The definition is established.

[0126] In this embodiment, The paper is not used as a positive sample for training triples, nor is it derived from the above. Samples are taken from the negative sample pool; during the inference phase, they obtain surrogate representations based on text semantics, co-authorship, and other auxiliary relationships, and the final paper representation and ranking results are calculated by the trained shared module. When there are no available surrogate sources, the aforementioned basic representation fallback rules are executed.

[0127] In some implementations, the preset threshold for low interactivity, the number of proxy neighbors, the similarity filtering threshold, the number of experts, the uncertainty temperature, and the weights of each loss can be preset according to the data scale and recommendation task; when the metadata of co-authors is unavailable, the same proxy structure completion process can be completed by using only semantically similar proxy sources.

[0128] Example 2 like Figure 1 As shown, this scheme provides a paper recommendation method for sparse academic papers, including the following steps: Step 1: Obtain the set of users to be recommended, the set of candidate papers, the user-paper bipartite interaction graph constructed during the training phase, and the semantic similarity relationships and co-author relationships of the candidate papers. Label low-interaction candidate papers based on their interaction levels in the training user-paper bipartite interaction graph.

[0129] In offline evaluation scenarios, a set of real interaction papers from test users is also obtained; this set of real interaction papers is only used to calculate the recommendation evaluation index after the ranking results are output, and is not used as model input.

[0130] Step 2: Input the interaction graph, target user, candidate paper and auxiliary relationship obtained in Step 1 into the uncertainty-aware proxy structure completion model trained in Example 1 to obtain the user-based collaborative representation of the target user and the final paper representation of the candidate paper. Step 3: Calculate the candidate paper preference score based on the user representation and the final paper representation. After removing the papers that the target user has already interacted with during the training phase, output the Top-N recommended paper list according to the preference score.

[0131] This embodiment uses the model parameters and representation generation method obtained from Example 1. It acquires the target user and candidate paper set, and performs proxy structure completion, confidence-driven fusion, and multi-expert uncertainty estimation on all low-interaction papers in the candidate set to obtain the user-collaborative basic representation of the target user and the final paper representation of the candidate papers. Subsequently, papers that the target user has already interacted with during the training phase are removed from the candidate paper set. The preference score between the user-collaborative basic representation and the final paper representation of the candidate papers is calculated, and the papers are sorted from high to low according to the preference score. The top N papers are then output as the recommendation result. This solution can be applied to personalized paper recommendation scenarios in academic search platforms, institutional knowledge bases, scientific research management systems, literature recommendation systems, and academic knowledge service platforms to generate Top-N recommendation results when interaction data is sparse or candidate paper interaction is insufficient.

[0132] The technical content of Step 1 and Step 2 in this embodiment is the same as that in the first embodiment.

[0133] It should be noted that during the inference phase, the uncertainty-aware agent structure completion model processes all low-interaction papers, and the multi-expert uncertainty fusion module uses the mean path of the Gaussian distribution of latent variables as latent variables to reduce ranking fluctuations caused by random sampling.

[0134] Regarding step 3: Target users Remove papers that the target user has already interacted with during the training phase from the candidate paper set; by Calculate the scores of the candidate papers and sort them in descending order, then output the top score. The recommended papers included [number] papers. For the basic collaborative representation of the target users, This is for the final paper.

[0135] In one alternative implementation, for a low-interaction paper (p), a virtual neighbor score is calculated based on its semantically similar proxy neighbors:

[0136] in, For the thesis A set of semantically similar proxy neighbors. For the thesis Thesis on Agent Neighbors Semantic similarity proxy weights between them For user-based collaborative representation, For the neighbor's paper The basic collaborative representation.

[0137] The virtual neighbor scores in the current batch or candidate paper set are standardized and then transformed using a hyperbolic tangent function to obtain the virtual neighbor score residuals: in, and These are the mean and standard deviation of the virtual neighbor scores in the current batch or candidate paper set, respectively. A positive number set to prevent the denominator from being zero.

[0138] The virtual neighbor score residual is added to the base preference score calculated in Step 3 to obtain the enhanced preference score:

[0139] in,

[0140] The base preference score is calculated in Step 3. For the thesis The corresponding residual weights. For papers already appearing in the training interaction graph, set smaller weights. For cold-start papers with zero interaction, set a larger [interaction level]. For papers with high uncertainty, the uncertainty can be further reduced. .

[0141] When virtual neighbor score residuals are enabled, scores are based on enhanced preference. Sort the candidate papers; if this option is not enabled, let Enhanced preference scores degenerate into basic preference scores .

[0142] Example 3 This embodiment also provides an electronic device, see reference. Figure 3 It includes a memory 402 and a processor 401, the memory 402 storing a computer program, and the processor 401 being configured to run the computer program to perform the steps in any of the above embodiments of the method for constructing an uncertainty-aware agent structure completion model for sparse academic paper recommendations or the recommendation method for sparse academic papers.

[0143] Specifically, the processor 401 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0144] The memory 402 may include a large-capacity memory 402 for data or instructions. The memory 402 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 401.

[0145] The processor 401 reads and executes computer program instructions stored in the memory 402 to implement any of the above embodiments' methods for constructing an uncertainty-aware proxy structure completion model for sparse academic paper recommendations or for recommending sparse academic papers.

[0146] Optionally, the electronic device may further include a transmission device 403 and an input / output device 404, wherein the transmission device 403 is connected to the processor 401 and the input / output device 404 is connected to the processor 401.

[0147] Input / output device 404 is used to input or output information. In this embodiment, the input information may be a user-paper bipartite interaction diagram, etc., and the output information may be the final paper representation, etc.

[0148] Optionally, in this embodiment, the processor 401 can be configured to perform the following steps via a computer program: S1: Obtain the training dataset, which includes a user-paper bipartite interaction graph, semantic similarity relationships between papers, and co-author relationships. Papers with an interaction degree not exceeding a preset threshold are marked as low-interaction papers. S2: Construct an uncertainty-aware agent structure completion model, including: The basic collaborative representation module is used to process the user-paper bipartite interaction graph to obtain the user basic collaborative representation and the paper basic collaborative representation. A dual-proxy structure completion module is used to process low-interaction papers and obtain semantically similar proxy representations, co-author proxy representations and corresponding source confidence features; The confidence fusion and gating update module is used to perform confidence-driven fusion of two types of proxy representations based on source confidence features to obtain a completed proxy representation, and then combine the completed proxy representation with the paper-based collaborative representation through gating update to obtain a proxy enhancement representation; The multi-expert uncertainty fusion module is used to generate candidate paper representations through multiple probabilistic experts, determine sample-level uncertainty based on expert disagreement, generate a confidence coefficient accordingly, and weight and fuse the candidate paper representations with the basic collaborative representation of the paper to obtain the final paper representation. S3: Train an uncertainty-aware proxy structure completion model using a training dataset containing low-interaction papers. The uncertainty-aware proxy structure completion model uses pairwise ranking loss, local consistency loss, cross-view comparison loss, and multi-expert assistance loss to jointly optimize the model parameters.

[0149] or, Step 1: Obtain the target users, the candidate paper set, the user-paper bipartite interaction graph constructed based on the observed interactions during the training phase, and the semantic similarity relationships and co-author relationships of the candidate papers; label low-interaction candidate papers according to the interaction degree of the papers in the training interaction graph; Step 2: Input the user-paper bipartite interaction graph, target user, candidate paper and auxiliary relationship obtained in Step 1 into the trained uncertainty-aware proxy structure completion model to obtain the user-based collaborative representation of the target user and the final paper representation of the candidate paper. Step 3: Calculate the candidate paper preference score based on the user representation and the final paper representation. After removing the papers that the target user has already interacted with, output the Top-N recommended paper list according to the preference score.

[0150] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0151] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0152] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products), including software routines, applets, and / or macros, can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. A computer program product may include one or more computer-executable components configured to perform embodiments when the program is run. One or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted that any block in the logical flow of the figures may represent a program step, or interconnected logical circuitry, blocks and functions, or a combination of program steps and logical circuitry, blocks and functions. The software may be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0153] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0154] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for constructing an uncertainty-aware proxy structure completion model for sparse academic paper recommendation, characterized in that, Includes the following steps: S1: Obtain the training dataset, which includes a user-paper bipartite interaction graph, semantic similarity relationships between papers, and co-author relationships. Papers with an interaction degree not exceeding a preset threshold are marked as low-interaction papers. S2: Construct an uncertainty-aware agent structure completion model, including: The basic collaborative representation module is used to process the user-paper bipartite interaction graph to obtain the user basic collaborative representation and the paper basic collaborative representation. A dual-proxy structure completion module is used to process low-interaction papers and obtain semantically similar proxy representations, co-author proxy representations and corresponding source confidence features; The confidence fusion and gating update module is used to perform confidence-driven fusion of two types of proxy representations based on source confidence features to obtain a completed proxy representation, and then combine the completed proxy representation with the paper-based collaborative representation through gating update to obtain a proxy enhancement representation; The multi-expert uncertainty fusion module is used to generate candidate paper representations through multiple probabilistic experts, determine sample-level uncertainty based on expert disagreement, generate a confidence coefficient accordingly, and weight and fuse the candidate paper representations with the basic collaborative representation of the paper to obtain the final paper representation. S3: Train an uncertainty-aware proxy structure completion model using a training dataset containing low-interaction papers. The uncertainty-aware proxy structure completion model uses pairwise ranking loss, local consistency loss, cross-view comparison loss, and multi-expert assistance loss to jointly optimize the model parameters.

2. The method for constructing an uncertainty-aware proxy structure completion model for sparse academic paper recommendation according to claim 1, characterized in that, For each user, obtain the set of interactive paper neighbors of that user in the user-paper interaction graph; for each paper, obtain the set of interactive user neighbors of that paper in the user-paper interaction graph. For each layer of users, the user collaborative representation of the current layer is obtained by combining the set of interacting user neighbors and the set of interacting paper neighbors; For each layer of papers, the collaborative representation of the current layer is obtained by combining the set of interactive user neighbors and the set of interactive paper neighbors; the user basic collaborative representation is obtained by summing the user collaborative representations of all layers, and the paper basic collaborative representation is obtained by summing the paper collaborative representations of all layers.

3. The method for constructing an uncertainty-aware proxy structure completion model for sparse academic paper recommendation according to claim 1, characterized in that, The dual-proxy structure completion module includes a semantic similarity proxy branch and a co-author proxy branch. The semantic similarity proxy branch is designed for low-interaction papers. It selects semantic similarity proxy neighbors based on semantic similarity relationships and performs weighted aggregation on the paper's basic collaborative representation of the semantic similarity proxy neighbors to obtain the semantic similarity proxy representation and the corresponding source confidence features. The co-author proxy branch targets low-interaction papers, selects co-author proxy neighbors based on co-author relationships, and performs weighted aggregation on the basic collaborative representations of the co-author proxy neighbors to obtain co-author proxy representations and corresponding source confidence features.

4. The method for constructing an uncertainty-aware proxy structure completion model for sparse academic paper recommendation according to claim 1, characterized in that, The source confidence features of semantic similarity sources and co-author sources are concatenated to obtain concatenated features. The concatenated features are input into a mixing function to generate surrogate source mixing weights. Based on the surrogate source mixing weights, the semantic similarity surrogate representation and the co-author surrogate representation are fused with confidence to obtain the completed surrogate representation. The original paper-based collaborative representation and the completed surrogate representation are concatenated to generate a gating vector with the same dimension as the embedding dimension. Then, the gating vector is updated to generate the surrogate augmentation representation.

5. The method for constructing an uncertainty-aware proxy structure completion model for sparse academic paper recommendation according to claim 1, characterized in that, The multi-expert uncertainty fusion module is used to generate candidate paper representations by multiple probabilistic experts, taking the paper's basic collaborative representation and side information as input. The side information is a concatenated vector of source confidence features, completed surrogate representation, and paper activity information.

6. The method for constructing an uncertainty-aware proxy structure completion model for sparse academic paper recommendation according to claim 1, characterized in that, Multiple probabilistic experts with independent parameters generate a Gaussian distribution of latent variables using side information. During the training phase, reparameterization sampling is performed based on the Gaussian distribution of latent variables. The sampled latent variables are mapped to residual corrections through their decoding network, and the candidate paper representation is obtained by using the paper-based collaborative representation as the residual benchmark.

7. The method for constructing an uncertainty-aware proxy structure completion model for sparse academic paper recommendation according to claim 1, characterized in that, Local consistency loss is used to constrain the consistency between the agent-enhanced representation and the completed agent representation; cross-view contrast loss is obtained by constructing the original view and the perturbation view from the user representation and the paper representation, wherein the perturbation magnitude on the paper side is adaptively determined based on the paper activity information; multi-expert assistance loss includes at least reconstruction constraints, probability distribution regularization constraints and expert disagreement constraints.

8. A paper recommendation method for sparse academic papers, characterized in that, Includes the following steps: Step 1: Obtain the target user, the candidate paper set, the user-paper bipartite interaction graph constructed based on the observed interactions during the training phase, and the semantic similarity relationships and co-author relationships of the candidate papers; label low-interaction candidate papers based on the interaction degree of the papers in the training interaction graph, including cold-start papers with an interaction degree of 0. Step 2: Input the user-paper bipartite interaction graph, target user, candidate paper and auxiliary relationship obtained in Step 1 into the uncertainty-aware proxy structure completion model constructed and trained according to any one of claims 1 to 7, to obtain the user-based collaborative representation of the target user and the final paper representation of the candidate paper; Step 3: Calculate the candidate paper preference score based on the user representation and the final paper representation. After removing the papers that the target user has already interacted with during the training phase, output the Top-N recommended paper list according to the preference score.

9. The paper recommendation method for sparse academic papers according to claim 8, characterized in that, During the inference phase, the uncertainty-aware agent structure completion model processes all low-interaction papers, and the multi-expert uncertainty fusion module uses the mean path of the Gaussian distribution of latent variables as latent variables.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to execute the method for constructing an uncertainty-aware agent structure completion model for sparse academic paper recommendation as described in any one of claims 1 to 7, or the paper recommendation method for sparse academic papers as described in any one of claims 8 to 9.

Images