Debiasing sequence recommendation method and system based on long short-term interest separation
By separating user interest representation into long-term and short-term, and utilizing multi-user interest interaction graphs and contrastive learning methods, the problem of popularity bias in existing technologies is solved, achieving more accurate user interest prediction and recommendation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- QILU UNIVERSITY OF TECHNOLOGY (SHANDONG ACADEMY OF SCIENCES)
- Filing Date
- 2024-02-06
- Publication Date
- 2026-07-21
AI Technical Summary
Existing sequence recommendation models struggle to effectively distinguish between long-term and short-term conformity when dealing with user interest prediction, resulting in popularity bias in the recommendation results and ignoring the user's true intent.
By separating user interest representations into long-term and short-term representations, and extracting conformity factors using multi-user interest interaction graphs and graph convolutional networks, and combining contrastive learning methods to perform contrastive learning and adaptive fusion of long-term and short-term interest representations respectively, the influence of popularity bias is mitigated.
It enables more accurate capture of users' true intentions, improves the accuracy and fairness of sequence recommendation, and effectively alleviates the popularity bias problem.
Smart Images
Figure CN117951382B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of sequence recommendation technology, specifically relating to a debiased sequence recommendation method and system based on the separation of long and short-term interests. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] The goal of personalized sequence recommendation is to predict the next item a user will be interested in. However, when fitting a recommendation model to reconstruct observed user behavior data, the training model often becomes biased towards recommending popular items, resulting in unfair recommendation results.
[0004] Early research primarily relied on Markov chain models, which mainly captured local information within sequences but lacked the ability to capture long-term dependencies. To address this issue, some studies used recurrent neural networks (RNNs) to model latent interests in user interaction sequences, with GRU4Rec being a prime example. Later, some research shifted to using convolutional neural networks (CNNs) for sequence modeling. CNNs can learn local features from specific regions and understand the relationships between different regions, effectively capturing joint-level collective dependencies often overlooked by other models. Caser is a representative work in this field, capturing general user preferences and sequence patterns simultaneously at the joint and point levels through convolutional operations. Furthermore, graph neural network (GNN)-based models such as SR-GNN, GCSAN, and SURGE were proposed for sequence recommendation tasks. Because GNNs can capture global connections, these models also achieved performance improvements. Later, the introduction of Transformers inspired researchers to develop attention-based sequence recommendation models like SASRec and BERT4Rec, which better capture dependencies between user interaction items and achieve superior performance. Many current sequence recommendation models incorporate self-supervised signals through contrastive learning frameworks to enhance the learning of user and item representations. For example, CL4SRec introduces sequence data augmentation into the contrastive learning task to obtain more robust sequence representations, effectively alleviating the data sparsity problem. However, data augmentation may disrupt the internal associations of the original sequence, resulting in low-quality positive sample pairs. Therefore, DuoRec uses a Dropout-based model-level augmentation contrastive learning method to achieve better semantic preservation. On the other hand, to capture different user intentions in temporal interaction behaviors, ICLRec performs clustering and contrastive learning on user intentions, enhancing sequence recommendation by improving the representation of user interests.
[0005] Over the past few years, many researchers have proposed sequence recommendation methods based on deep neural networks, including attention-based and GNN-based methods, which can capture the implicit user interests in sequences. However, these methods often suffer from data sparsity, making it difficult to learn sufficiently good interest representations for recommendations. Contrastive learning can effectively address this problem by introducing self-supervised signals to enhance the representation learning of users and items. For example, various data augmentations on the sequence are used to force consistency between enhanced views for embedding contrast, and contrastive learning is applied by identifying semantic alignment between sequences or recommended items. While these contrastive learning methods have achieved some success in alleviating the data sparsity problem, they have not solved another important issue affecting recommendation performance: popularity bias. Therefore, existing technologies have also attempted to incorporate conformity signals into contrastive learning for sequence recommendation tasks to mitigate popularity bias and improve sequence recommendation performance. Although these methods have improved recommendation performance to some extent by introducing conformity factors to alleviate this problem, they still lack a comprehensive consideration of user conformity, ignoring the differences in users' conformity levels in the long and short term, making it difficult to accurately model users' true intentions. Summary of the Invention
[0006] To address the aforementioned issues, this invention proposes a debiased sequence recommendation method and system based on the separation of long-term and short-term interests. This invention separates the long-term and short-term representations of user interests and compares them with the embedded representations of multi-user interest interaction graphs. During the comparative learning process, long-term conformity factors and short-term conformity factors are introduced to assist in debiasing, giving the sequence recommendation a strong debiasing capability, so that the interest representations learned by the model can better express the user's true intentions.
[0007] According to some embodiments, the first aspect of the present invention provides a bias-free sequence recommendation method based on long-short-term interest separation, employing the following technical solution:
[0008] Bias-free sequence recommendation methods based on long-term and short-term interest separation include:
[0009] A multi-user interest interaction graph is constructed based on user interaction sequences to obtain a global user interest representation.
[0010] Construct long-term and short-term interest representations based on user interaction sequences;
[0011] Based on long-term and short-term interest representations, comparative learning is performed using conformity factors to obtain user long-term and short-term interest representations.
[0012] The user's long-term interest representation and short-term interest representation are adaptively fused to form a user fused interest representation, which is used for the prediction task of sequence recommendation.
[0013] Furthermore, the step of constructing a multi-user interest interaction graph based on user interaction sequences to obtain a global user interest representation includes:
[0014] Construct a multi-user interest interaction graph by using adjacent item pairs in each user sequence as edges of the sequence graph;
[0015] By using graph convolutional networks to encode items in a multi-user interest interaction graph, a global user interest representation is obtained.
[0016] Furthermore, long-term interest representations are constructed based on user interaction sequences, including:
[0017] The user interaction sequence is mapped to a long-term user embedding, and the long-term user embedding is concatenated with the positional embedding of the user interaction item to obtain the long-term interest embedding.
[0018] The long-term interest is encoded using an encoder to obtain a long-term interest representation.
[0019] Furthermore, short-term interest representations are constructed based on user interaction sequences, including:
[0020] Short-term interaction sequences are selected based on user interaction sequences;
[0021] Short-term interaction sequences are mapped to short-term user embeddings, and the short-term user embeddings and the positional embeddings of user interaction items are concatenated to obtain short-term interest embeddings.
[0022] Short-term interest is embedded in the encoder to obtain a short-term interest representation.
[0023] Furthermore, the step of obtaining user long-term interest representation and user short-term interest representation by comparative learning using conformity factors based on long-term interest representation and short-term interest representation includes:
[0024] By using a contrastive learning loss function, the long-term interest representation is compared with the long-term conformity factor control to obtain the user's long-term interest representation.
[0025] By using a contrastive learning loss function to compare and learn the short-term interest representation with the short-term conformity factor control, a short-term interest representation of the user is obtained.
[0026] Furthermore, the extraction process of the long-term conformity factor is specifically as follows:
[0027] Based on the multi-user interest interaction graph, the degree centrality algorithm is used to calculate the degree centrality of each item node.
[0028] Normalize the degree centrality of each user's project node;
[0029] The long-term conformity factor is obtained by aggregating the normalized degree centrality of all project nodes for each user.
[0030] Furthermore, the extraction process of the short-term conformity factor is specifically as follows:
[0031] Based on the multi-user interest interaction graph, the degree centrality algorithm is used to calculate the degree centrality of each item node.
[0032] Normalize the degree centrality of each user's project node;
[0033] The short-term conformity factor is obtained by aggregating the normalized degree centrality of each user's short-term interaction item node.
[0034] According to some embodiments, the second aspect of the present invention provides a bias-free sequence recommendation system based on long-short-term interest separation, employing the following technical solution:
[0035] Bias-free sequence recommendation systems based on long-short-term interest separation include:
[0036] The global interest representation extraction module is configured to construct a multi-user interest interaction graph based on user interaction sequences to obtain a global user interest representation.
[0037] The long-term and short-term interest representation extraction module is configured to construct long-term interest representations and short-term interest representations respectively based on the user interaction sequence;
[0038] The long-term and short-term interest representation comparison learning module is configured to perform comparative learning based on long-term and short-term interest representations, using conformity factors to obtain user long-term and short-term interest representations respectively.
[0039] The sequence recommendation module is configured to adaptively fuse user long-term interest representations and user short-term interest representations to form a user fused interest representation, which is used for the prediction task of sequence recommendation.
[0040] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium.
[0041] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the bias-removing sequence recommendation method based on long-short-term interest separation as described in the first aspect above.
[0042] According to some embodiments, a fourth aspect of the present invention provides a computer device.
[0043] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the bias-removed sequence recommendation method based on long-short-term interest separation as described in the first aspect above.
[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0045] This invention comprehensively considers both long-term and short-term conformity in contrastive learning-based sequence recommendation, enabling the learning of interest representations that better express users' true intentions. By decomposing user interest representations into long-term and short-term representations, it captures the dynamic changes in user interests more precisely. To comprehensively consider the interaction relationships among multiple user interests, a multi-user interest interaction graph is introduced, and a graph convolutional network is used to obtain the conformity embeddings of user interests on this graph, while simultaneously extracting long-term and short-term conformity factors. Subsequently, contrastive learning is used to effectively compare the long-term and short-term interest representations with the conformity representations. During the contrastive learning process, long-term and short-term conformity factors are introduced to effectively mitigate the impact of popularity bias in contrastive learning-based sequence recommendation systems. Attached Figure Description
[0046] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0047] Figure 1 This is a flowchart of the bias-free sequence recommendation method based on long-short-term interest separation in this embodiment of the invention;
[0048] Figure 2 This is a diagram illustrating the architecture of the comparative learning sequence recommendation method for separating long and short-term interests according to an embodiment of the present invention. Detailed Implementation
[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0050] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0051] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0052] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0053] Example 1
[0054] like Figure 1 As shown, this embodiment provides a debiased sequence recommendation method based on long-short-term interest separation. This embodiment uses the application of this method to a server as an example for illustration. It is understood that this method can also be applied to terminals, and can also be applied to systems including terminals, servers, and other components, and implemented through interaction between the terminal and the server. The server can be an independent physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communication, middleware services, domain name services, CDN security services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein. In this embodiment, the method includes the following steps:
[0055] A multi-user interest interaction graph is constructed based on user interaction sequences to obtain a global user interest representation.
[0056] Construct long-term and short-term interest representations based on user interaction sequences;
[0057] Based on long-term and short-term interest representations, comparative learning is performed using conformity factors to obtain user long-term and short-term interest representations.
[0058] The user's long-term interest representation and short-term interest representation are adaptively fused to form a user fused interest representation, which is used for the prediction task of sequence recommendation.
[0059] Specifically, this embodiment first constructs a multi-user interest interaction graph using global sequences and obtains a global user interest representation using GCN, which contains rich cross-user semantic information. Second, long-term and short-term interests are modeled separately to obtain long-term and short-term interest representations, which are then compared and learned with the global user interest representation. Third, conformity factors are used to model users' conformity psychology. Specifically, short-term conformity factors extracted from the global sequence graph are incorporated into short-term contrastive learning, which helps to dynamically control knowledge transfer between the global representation and short-term interest representation across sequences, enhancing the short-term interest representation. Long-term conformity factors extracted from the global sequence graph are then incorporated into long-term contrastive learning to better model users' long-term interests. Finally, the short-term and long-term representations are adaptively fused to form the user's final interest representation, used for sequence recommendation prediction tasks.
[0060] The specific implementation steps of this embodiment are as follows:
[0061] The model of this invention demonstrates the proposed contrastive learning sequence recommendation method LSDRec based on long-short-term interest separation, whose architecture is as follows: Figure 2 As shown. The framework consists of three main components: (1) an interest encoding module, which includes a long-term interest encoder and a short-term interest encoder, used to model users’ long-term and short-term interests, respectively; (2) a multi-user interaction interest graph module, used to model the interests of multiple users, which we regard as user conformity modeling, and extract conformity factors with reasonable interaction levels from it; (3) a contrastive learning module, used to integrate the conformity factors in (2) into our embedding contrastive paradigm to assist in knowledge transfer between graph modeling representation and sequence modeling representation.
[0062] Multi-user interactive interest graph
[0063] Multi-user interaction interest graph construction, wherein the construction of a multi-user interest interaction graph based on user interaction sequences to obtain a global user interest representation includes:
[0064] Construct a multi-user interest interaction graph by using adjacent item pairs in each user sequence as edges of the sequence graph;
[0065] By using graph convolutional networks to encode items in a multi-user interest interaction graph, a global user interest representation is obtained.
[0066] To capture the transition relationships between projects from a multi-user sequence, each sequence S u Adjacent item pairs in the sequence are used as edges in the sequence transformation graph G. Given the item sequences of all users... The adjacency matrix of item relevance in graph G The construction method can be expressed by the following formula:
[0067]
[0068]
[0069] in, S represents the interaction sequence of user u. u The connection built, i a and i b Let a and b represent two distinct items in the interaction sequence, where a and b are their position indices. Finally, the adjacency matrix of all users u is calculated. The summation yields the final multi-user sequence adjacency matrix A. G .
[0070] Understandably, the construction of the interest graph is the first step, while the subsequent GCN encoding and Long Short-Term Encoding are performed in parallel.
[0071] The graph encoder based on GCN is used by GCN to construct a conformity representation for encoding the global graph. This representation is an embedding used for comparison and learning with the representation produced by the subsequent long short-term encoder.
[0072] Redundant transformation and activation operations during message transmission on graph G have been removed, as shown below:
[0073]
[0074] Where l represents the number of layers in the GCN, Indicates the neighbors of i. Let be the embedding representation of i, i′∈I at layer 0, and d represent the embedding size in the latent space. A residual connection is added in the last layer to mitigate the vanishing gradient problem. Finally, we sum the representations of all hidden layers to obtain the final embedding for item i. We view this as a representation of user conformity embedding, i.e. Using this GCN-based encoder, the transitional relationships between items in a multi-user interaction interest graph can be captured, and the resulting item embeddings contain rich cross-user semantic information.
[0075] Long-term user interest encoding involves constructing long-term interest representations based on user interaction sequences, including:
[0076] The user interaction sequence is mapped to a long-term user embedding, and the long-term user embedding is concatenated with the positional embedding of the user interaction item to obtain the long-term interest embedding.
[0077] The long-term interest is encoded using an encoder to obtain a long-term interest representation.
[0078] Long-term interests reflect the overall state of user preferences and are less influenced by recent interactions. In other words, long-term interests can be inferred from the entire historical interaction sequence. Therefore, in the long-term interest encoder, we model the user's long-term interests using the entire interaction information of user u, and then consider the interaction sequence S. u Mapping to Embedding Use P u Let represent the location embedding of the user's interactive items. The concatenation of these two embeddings serves as the input to the user's long-term interest encoder. The overall embedding matrix can be represented as follows: We encode user interests using a Transformer layer, embedding the representation. Input a long-term interest encoder and apply a multi-head self-attention module to compute user representations:
[0079]
[0080]
[0081]
[0082] Among them, W l ∈R d×d It is a learnable projection matrix, projection matrix W i Q W i K , Q represents the query, K represents the key, V represents the value, and h is the number of attention heads.
[0083] To enhance the model's expressive power, nonlinearity is injected into the generation process, and a feedforward neural network (FFN) is used within the interest encoder for representation transformation, defined as:
[0084]
[0085] FFN(x)=GELU(xW (1) +b (1) W (2) +b (2)
[0086] in, It represents the end user's long-term interest; GELU is the activation function; W (1) W (2) ,b (1) ,b (2) These are learnable weight parameters.
[0087] Short-term interest encoding, which constructs short-term interest representations based on user interaction sequences, includes:
[0088] Short-term interaction sequences are selected based on user interaction sequences;
[0089] Short-term interaction sequences are mapped to short-term user embeddings, and the short-term user embeddings and the positional embeddings of user interaction items are concatenated to obtain short-term interest embeddings.
[0090] Short-term interest is embedded in the encoder to obtain a short-term interest representation.
[0091] As users interact with recommended items, their short-term interests change. For example, users may develop new interests after interacting with items. Simultaneously, users may gradually lose some interests. In other words, short-term interests are time-dependent variables, indicating that a user's short-term interests are significantly influenced by recent interactions. Therefore, we use a short-term interest encoder based on the short-term interaction sequence of user u containing k items. Model users' short-term interests and map them to embeddings Embed it with position encoding P u Concatenation is used as input to the short-term interest encoder: Similar to long-term interest encoders, we use a Transformer layer to encode user interests, embedding the user's short-term historical interaction sequences into the representation. As input to the multi-head attention layer:
[0092]
[0093]
[0094]
[0095] in, S is a short-term interest representation of the end user. s It is a hidden representation that has passed through a multi-head attention layer (MH), W s ∈R d×d It is a learnable projection matrix, and FFN is a feedforward neural network consistent with the long-term interest encoder.
[0096] The herd factor extractor, specifically the extraction process of the long-term herd factor, is as follows:
[0097] Based on the multi-user interest interaction graph, the degree centrality algorithm is used to calculate the degree centrality of each item node.
[0098] Normalize the degree centrality of each user's project node;
[0099] The long-term conformity factor is obtained by aggregating the normalized degree centrality of all project nodes for each user.
[0100] The extraction process of the short-term herd mentality factor is as follows:
[0101] Based on the multi-user interest interaction graph, the degree centrality algorithm is used to calculate the degree centrality of each item node.
[0102] Normalize the degree centrality of each user's project node;
[0103] The short-term conformity factor is obtained by aggregating the normalized degree centrality of each user's short-term interaction item node.
[0104] In a graph network, the higher the degree of a node, the higher its degree centrality, which also means that the node is more closely related to other nodes in the graph. Based on the multi-user interaction interest graph established earlier, the degree centrality algorithm is used to calculate the degree centrality of each project node to reflect the popularity of different projects.
[0105]
[0106] Among them, C D (v i The degree centrality of node i is represented by ), and N represents the number of project nodes. This is used to calculate the number of direct connections between a node and the other N-1 nodes. To better represent the popularity of the project, the initial degree center calculation value is further normalized using Min-Max:
[0107] D i =MinMax(log C) D (v i )),
[0108] MinMax(x) = (x - min) / (max - min)
[0109] Among them, D i The degree centrality of the i-th node is normalized. Each node corresponds to an interaction item. Finally, the conformity factor ω is obtained by aggregating the k interaction items of each user, which is used to represent the degree of conformity of the user.
[0110]
[0111] For calculating short-term conformity, k is taken as the number of items the user interacts with in a short period of time, thus obtaining the short-term conformity factor. For calculating long-term conformity, k is taken as the total number of user interactions in the sequence, n, to obtain the long-term conformity factor.
[0112] The contrastive learning module, based on long-term and short-term interest representations, uses conformity factors to perform contrastive learning to obtain user long-term and short-term interest representations, including:
[0113] By using a contrastive learning loss function, the long-term interest representation is compared with the long-term conformity factor control to obtain the user's long-term interest representation.
[0114] By using a contrastive learning loss function to compare and learn the short-term interest representation with the short-term conformity factor control, a short-term interest representation of the user is obtained.
[0115] Short-term interest contrastive learning. To better mitigate popularity bias during contrastive learning, we introduce a conformity factor in the contrastive learning of short-term sequence interests to adaptively control the degree of knowledge transfer among different users. Specifically, we perform contrastive learning between short-term interest embeddings and GCN-encoded conformity embeddings, using the short-term conformity factor extracted by the conformity signal extractor. This is applied to the contrastive learning loss function.
[0116]
[0117] exist We employ InfoNCE for embedding comparison. By incorporating the conformity factor ω learned in our conformity signal extractor, we allow for the representation of short-term interests. and the conformity representation e encoded by GCN U Adaptive knowledge transfer,
[0118] Long-term interest contrastive learning. To better model users' long-term intrinsic interests, we also conduct contrastive learning between long-term interest embeddings and conformity embeddings encoded by GCN. During the contrastive learning process, we use long-term conformity factors. Controlling the degree of contrastive learning allows for an appropriate amount of knowledge transfer, thereby enhancing long-term interest embedding representations.
[0119]
[0120] We incorporate both long-term and short-term interest contrastive learning losses into our contrastive learning paradigm to provide self-supervised signals for model training, enabling our model to obtain interest representations that more effectively represent the user's current true intent.
[0121] Model training and prediction
[0122] User sequence S u The last item serves as the label for model optimization during training. To obtain the interest embedding that best matches the user's current intent, we aggregate the user's long-term and short-term interests into the final interest representation e using learnable attention weights.u :
[0123]
[0124]
[0125] To supplement the recommendation loss with an enhanced self-supervised learning task within a multi-task training framework. We will jointly optimize the objectives Defined as:
[0126]
[0127] Here, λ1 and λ2 are hyperparameters used to balance task-specific losses. It is the cross-entropy loss function in sequence recommendation.
[0128] This embodiment proposes a novel sequence recommendation method: a debiased sequence recommendation method based on contrastive learning with long-term and short-term interest separation. This method decomposes user interest representations into long-term and short-term representations to capture the dynamic changes in user interests more precisely. To comprehensively consider the interaction relationships of multiple user interests, we introduce a multi-user interest interaction graph and obtain the conformity embedding of user interests on this graph through a graph convolutional network, while simultaneously extracting long-term and short-term conformity factors. Subsequently, we effectively compare the long-term and short-term interest representations with the conformity representations through contrastive learning. During the contrastive learning process, we introduce long-term and short-term conformity factors to effectively mitigate the impact of popularity bias in the contrastive learning-based sequence recommendation system. Through a series of detailed experimental verifications, we clearly demonstrate the effectiveness of this invention and obtain more significant results compared to other baseline methods.
[0129] Example 2
[0130] This embodiment provides a bias-free sequence recommendation system based on long-short-term interest separation, including:
[0131] The global interest representation extraction module is configured to construct a multi-user interest interaction graph based on user interaction sequences to obtain a global user interest representation.
[0132] The long-term and short-term interest representation extraction module is configured to construct long-term interest representations and short-term interest representations respectively based on the user interaction sequence;
[0133] The long-term and short-term interest representation comparison learning module is configured to perform comparative learning based on long-term and short-term interest representations, using conformity factors to obtain user long-term and short-term interest representations respectively.
[0134] The sequence recommendation module is configured to adaptively fuse user long-term interest representations and user short-term interest representations to form a user fused interest representation, which is used for the prediction task of sequence recommendation.
[0135] The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1 above. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.
[0136] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0137] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.
[0138] Example 3
[0139] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the bias-removing sequence recommendation method based on long-short-term interest separation as described in Embodiment 1 above.
[0140] Example 4
[0141] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the bias-removing sequence recommendation method based on long-short-term interest separation as described in Embodiment 1 above.
[0142] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0143] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0144] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0145] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0146] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.
[0147] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A bias-free sequence recommendation method based on long-short-term interest separation, characterized in that, include: A multi-user interest interaction graph is constructed based on user interaction sequences to obtain a global user interest representation. Construct long-term and short-term interest representations based on user interaction sequences; Based on long-term and short-term interest representations, comparative learning is performed using conformity factors to obtain user long-term and short-term interest representations, including: The long-term interest representation is obtained by comparing it with the long-term conformity factor control using a contrastive learning loss function; the extraction process of the long-term conformity factor is as follows: Based on the multi-user interest interaction graph, the degree centrality algorithm is used to calculate the degree centrality of each item node. Normalize the degree centrality of each user's project node; The long-term conformity factor is obtained by aggregating the normalized degree centrality of all project nodes for each user. The user's short-term interest representation is obtained by comparing the short-term interest representation with the short-term conformity factor control using a contrastive learning loss function; the extraction process of the short-term conformity factor is as follows: Based on the multi-user interest interaction graph, the degree centrality algorithm is used to calculate the degree centrality of each item node. Normalize the degree centrality of each user's project node; The short-term conformity factor is obtained by aggregating the normalized degree centrality of each user's short-term interaction item node. The user's long-term interest representation and short-term interest representation are adaptively fused to form a user fused interest representation, which is used for the prediction task of sequence recommendation.
2. The bias-free sequence recommendation method based on long-short-term interest separation as described in claim 1, characterized in that, The construction of a multi-user interest interaction graph based on user interaction sequences to obtain a global user interest representation includes: Construct a multi-user interest interaction graph by using adjacent item pairs in each user sequence as edges of the sequence graph; By using graph convolutional networks to encode items in a multi-user interest interaction graph, a global user interest representation is obtained.
3. The bias-free sequence recommendation method based on long-short-term interest separation as described in claim 1, characterized in that, Construct long-term interest representations based on user interaction sequences, including: The user interaction sequence is mapped to a long-term user embedding, and the long-term user embedding is concatenated with the positional embedding of the user interaction item to obtain the long-term interest embedding. The long-term interest is encoded using an encoder to obtain a long-term interest representation.
4. The bias-free sequence recommendation method based on long-short-term interest separation as described in claim 1, characterized in that, Short-term interest representations are constructed based on user interaction sequences, including: Short-term interaction sequences are selected based on user interaction sequences; Short-term interaction sequences are mapped to short-term user embeddings, and the short-term user embeddings and the positional embeddings of user interaction items are concatenated to obtain short-term interest embeddings. Short-term interest is embedded in the encoder to obtain a short-term interest representation.
5. A bias-free sequence recommendation system based on long-short-term interest separation, characterized in that, include: The global interest representation extraction module is configured to construct a multi-user interest interaction graph based on user interaction sequences to obtain a global user interest representation. The long-term and short-term interest representation extraction module is configured to construct long-term interest representations and short-term interest representations respectively based on the user interaction sequence; The long-term and short-term interest representation contrastive learning module is configured to perform contrastive learning based on long-term and short-term interest representations using conformity factors to obtain user long-term and short-term interest representations, including: The long-term interest representation is obtained by comparing it with the long-term conformity factor control using a contrastive learning loss function; the extraction process of the long-term conformity factor is as follows: Based on the multi-user interest interaction graph, the degree centrality algorithm is used to calculate the degree centrality of each item node. Normalize the degree centrality of each user's project node; The long-term conformity factor is obtained by aggregating the normalized degree centrality of all project nodes for each user. The user's short-term interest representation is obtained by comparing the short-term interest representation with the short-term conformity factor control using a contrastive learning loss function; the extraction process of the short-term conformity factor is as follows: Based on the multi-user interest interaction graph, the degree centrality algorithm is used to calculate the degree centrality of each item node. Normalize the degree centrality of each user's project node; The short-term conformity factor is obtained by aggregating the normalized degree centrality of each user's short-term interaction item node. The sequence recommendation module is configured to adaptively fuse user long-term interest representations and user short-term interest representations to form a user fused interest representation, which is used for the prediction task of sequence recommendation.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the biased sequence recommendation method based on long-short-term interest separation as described in any one of claims 1-4.
7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the bias-removing sequence recommendation method based on long-short-term interest separation as described in any one of claims 1-4.