Cross-domain knowledge migration cold start recommendation method based on dynamic intention perception

Abstract

The invention relates to a cross-domain knowledge migration cold start recommendation method based on dynamic intention perception. A system model of the method comprises a decoupling feature extractor based on graph convolution, an intention bridging network, a self-adaptive knowledge fusion mechanism and a recommendation generation unit. The method comprises the following steps: firstly, decomposing expressions of a user and an article into a plurality of intention subspaces through a decoupling feature extractor; then, establishing a soft mapping relation between intention subspaces of a source domain and a target domain by using an intention bridging network to realize intention alignment of fine granularity; the migration degree of source domain knowledge is dynamically adjusted through an adaptive knowledge fusion mechanism, and differentiated migration strategies are adopted for different users and articles; and finally, stable learning and smooth knowledge migration of the intention mapping relation are ensured by adopting a three-stage progressive training strategy of a recommendation generation unit. The problems of non-correspondence of intention semantics, negative migration and the like in traditional cross-domain recommendation are effectively solved, the recommendation effect is remarkably improved, and the method is particularly suitable for a cold start scene with sparse target domain data.

Description

Technical Field

[0001] The present invention belongs to the technical field of material performance testing equipment, and specifically relates to a cross-domain knowledge migration cold start recommendation method based on dynamic intent perception, which is suitable for solving the problems of knowledge migration and cold start recommendation between different fields. Background Art

[0002] The rapid development of internet technology and e-commerce platforms has led to increasingly diverse online consumer behaviors, with cross-platform and cross-domain shopping becoming the norm. Recommender systems, as information filtering tools, play a vital role in helping users discover content of interest. However, with the diversification of user behavior, traditional single-domain recommendation systems face significant challenges, especially in emerging domains or cold-start scenarios, where data sparsity often makes it difficult to guarantee recommendation quality.

[0003] Cross-domain recommendation systems provide an effective approach to addressing these challenges by building knowledge bridges between different domains. Existing cross-domain recommendation methods primarily include those based on shared representations, transfer learning, and adversarial domain adaptation. While these methods alleviate data sparsity to some extent, they still have significant shortcomings. First, most existing methods transfer user and item representations as a whole, ignoring the diversity of user interests and the potential correspondence between intents across domains. Users often exhibit significant differences in intent across domains. For example, a user in the electronics domain may prioritize performance, configuration, and brand reputation, while focusing on style, design, and styling in the clothing domain. These domain-specific user intents not only influence the user's decision-making process but also determine the effective boundaries of knowledge transfer between domains. Secondly, with the introduction of disentangled representation learning methods, recommender systems can decompose user interests into multiple independent intent factors, more precisely capturing users' diverse preferences. However, this also introduces new challenges in cross-domain knowledge transfer: disentangled intent factors in different domains may exhibit semantic mismatches. For example, the first intent factor in the source domain might represent "brand preference," while the first intent factor in the target domain might represent "price sensitivity." Simply mapping these factors one-to-one can lead to semantic confusion and even a breakdown of the intent structure. Furthermore, different samples differ in their need for source domain knowledge. For popular items with ample data in the target domain, over-reliance on source domain knowledge can actually interfere with the effective representations learned. Cold-start items, on the other hand, require even greater support from source domain knowledge.

[0004] Based on the above analysis, the current technical field urgently needs a cross-domain recommendation method that can achieve accurate knowledge transfer while maintaining the semantic consistency of decoupled representation, so as to effectively solve the problems of intent semantic mismatch and negative transfer existing in existing technologies. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a cross-domain knowledge transfer cold start recommendation method based on dynamic intent perception, aiming to solve the problem of intent space inconsistency in cross-domain recommendation, especially the problem of intent semantic mismatch when using decoupled representation learning method.

[0006] The present invention solves the technical problem by the following technical solutions:

[0007] A cross-domain knowledge transfer cold-start recommendation method based on dynamic intent perception is proposed. The recommendation system model adopted by the method includes a decoupled feature extractor based on graph convolution, an intent bridging network, and an adaptive knowledge fusion mechanism. The steps of the method are as follows:

[0008] S1. Decompose the user and item representation spaces in the source and target domains into multiple intent subspaces using a graph convolution-based decoupled feature extractor. Each intent subspace captures the interaction pattern under a specific intent.

[0009] S2. Establish an intent bridging network to automatically discover the intent correspondence between different domains through a learnable intent mapping matrix;

[0010] S3. Design an adaptive knowledge fusion mechanism to dynamically adjust the degree of source domain knowledge transfer according to the characteristics of the target domain samples;

[0011] S4. A three-stage progressive training strategy is adopted, including a warm-up stage, a joint optimization stage, and a fine-tuning stage, to ensure that the model can stably and effectively learn cross-domain intent mapping relationships and recommendation tasks.

[0012] Preferably, the S1 is specifically:

[0013] (1) The graph convolution-based decoupled feature extractor evenly divides the d-dimensional embedding vectors of user u and item i into K sub-vectors, each of which represents a representation of a potential intent;

[0014] (2) Through intent-aware graph convolution operations, the interaction patterns are learned separately in each intent subspace, and the representations of each layer are fused using residual connections to retain structural information at different levels.

[0015] Preferably, the S2 is specifically:

[0016] (1) The intention bridging network establishes an intention mapping matrix M, where each element M ij represents the association strength between the i-th intention in the source domain and the j-th intention in the target domain;

[0017] (2) Normalize the mapping matrix column-wise to ensure that each target domain intent is a weighted combination of different source domain intents, and the sum of the weights is 1;

[0018] (3) Based on the normalized mapping matrix, the intent representation in the source domain is mapped to the intent space of the target domain;

[0019] (4) The intent mapping relationship is optimized through mapping similarity loss, mapping sparsity loss and mapping coverage loss. The mapping similarity loss encourages the mapped source domain representation and the target domain representation to have high similarity in the corresponding intent subspace; the mapping sparsity loss enables each source domain intent to be mainly mapped to a few target domain intents; the mapping coverage loss ensures that each intent in the target domain can obtain sufficient knowledge from the source domain.

[0020] Preferably, the S3 is specifically:

[0021] (1) Construct a context vector containing the complete feature information of the current interaction pair;

[0022] (2) Calculate the gate value through the multi-layer perceptron structure to control the influence of source domain knowledge on the final representation;

[0023] (3) Based on the gated value, the source domain mapped representation is weightedly fused with the target domain original representation to obtain the final representation.

[0024] Preferably, the S4 is specifically:

[0025] (1) The warm-up phase focuses on training the intent bridging network while temporarily freezing the parameters of the source and target domain models;

[0026] (2) The joint optimization phase simultaneously optimizes the parameters of the intent bridging network and the target domain model, coordinating knowledge transfer and target domain task learning;

[0027] (3) The fine-tuning phase fixes the learned intent mapping relationship and focuses on optimizing the performance of the target domain model.

[0028] The advantages and beneficial effects of the present invention are:

[0029] This invention achieves more precise knowledge transfer by preserving the semantic structure of decoupled representations through an innovative intent-bridging mechanism. Furthermore, its adaptive knowledge fusion mechanism adjusts the degree of knowledge transfer based on the characteristics of target domain samples, ensuring both accuracy and effectiveness. This invention is particularly well-suited for cold-start scenarios where target domain data is sparse, significantly improving recommendation performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Schematic diagram of a cross-domain recommendation system model based on dynamic intent perception according to the present invention;

[0031] Figure 2 A schematic diagram of the structure of a bridge network intended by the present invention;

[0032] Figure 3 This is a flow chart of the three-stage progressive training strategy of the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the scope of protection of the present invention cannot be limited thereto.

[0034] A cross-domain knowledge transfer cold start recommendation method based on dynamic intent perception is innovative in that the cross-domain recommendation system consists of two different domains: the source domain and the target domain. The source domain has rich user-item interaction data, while the target domain may face data sparsity or cold start problems. The goal of this invention is to use the knowledge of the source domain to improve the recommendation performance of the target domain, especially for cold start scenarios. Figure 1 As shown, the recommendation system model adopted by the method includes a decoupled feature extractor based on graph convolution, an intention bridging network and an adaptive knowledge fusion mechanism; the steps of the method are:

[0035] The decoupled feature extractor based on graph convolution decomposes the user and item representation space into multiple intent subspaces, each of which captures interaction patterns under a specific intent. Specifically, for each user u and item i, its d-dimensional embedding vector is evenly divided into K subvectors, each representing a representation of a potential intent. Through intent-aware graph convolution operations, interaction patterns are learned separately in each intent subspace, enabling the model to more nuancedly understand user preferences across different intent dimensions.

[0036] In the lth layer of graph convolution, the representation update of node v in the kth intention subspace is achieved by aggregating its neighbor information:

[0037] in: represents the neighbor set of node $v$; W k is the transformation matrix of the kth intention subspace; is the weight coefficient of the edge (u, v) in the kth intent subspace. The weight coefficient calculated by the attention mechanism can adaptively adjust the importance of different neighbor nodes, enabling the model to better capture intent-specific interaction patterns.

[0038] In order to avoid the over-smoothing problem caused by deep graph convolution and retain structural information at different levels, this paper adopts residual connections to fuse the representations of each layer, so that the model can simultaneously utilize shallow local structural features and deep global semantic information:

[0039]

[0040] Where: L represents the total number of graph convolutional layers.

[0041] The structure of the intent bridge network is as follows Figure 2 As shown in Figure 2, the network establishes a mapping relationship between the source domain and target domain intent subspaces through a learnable intent mapping matrix, which enables the model to automatically discover the intent correspondence between different domains instead of enforcing a simple one-to-one correspondence, thus avoiding knowledge misleading caused by semantic mismatch. The core of the intent bridging network is an intent mapping matrix. Where K is the number of intents, each element of the matrix M ij represents the association strength between the i-th intent in the source domain and the j-th intent in the target domain. The matrix is ​​normalized column-wise to ensure that each target domain intent is a weighted combination of different source domain intents, and the sum of the weights is 1:

[0042] Based on the normalized mapping matrix, the intent representation in the source domain is mapped to the intent space of the target domain: The jth intent representation in the target domain is a weighted combination of all intent representations in the source domain, with the weights being the corresponding columns of the mapping matrix Decide.

[0043] In order to learn effective intent mapping relationships, the present invention designs an intent alignment loss function composed of mapping similarity loss, mapping sparsity loss, and mapping coverage loss. Mapping similarity loss encourages the mapped source domain representation and target domain representation to have high similarity in the corresponding intent subspace; mapping sparsity loss maximizes the negative entropy of the mapping probability distribution, so that each source domain intent is mainly mapped to a few semantically related target domain intents, avoiding unnecessary cross-influence; mapping coverage loss ensures that each intent in the target domain can obtain sufficient knowledge support from the source domain. Through the joint action of the three constraints, the intent mapping can be made both accurate and comprehensive, effectively preserving the semantic structure of the decoupled representation.

[0044] The adaptive knowledge fusion mechanism generates a gating value based on the characteristics of the target domain samples to control the influence of the source domain knowledge on the final representation. This allows the model to adopt differentiated migration strategies for different samples, relying more on the source domain knowledge in areas with sparse target domain data, while retaining more of the target domain's own characteristics in areas with abundant data. The mechanism first constructs a context vector containing the complete feature information of the current interaction pair: c ui =[e u,t ;e i,t Based on this context vector, a gate value is calculated through the multi-layer perceptron structure: g ui =σ(W2(ReLU(W1c ui +b1))+b2).

[0045] Based on the gate value, the source domain mapped representation is weightedly fused with the target domain original representation to obtain the final representation: When the gating value is close to 0, it means that the model judges that the source domain knowledge may not be applicable to the current sample or may have potential interference, so it retains more of the original features of the target domain; when the gating value is close to 1, it means that the source domain knowledge is very valuable to the current sample and should be fully utilized.

[0046] Figure 3 This paper demonstrates the three-stage progressive training strategy employed by the present invention, consisting of a warm-up phase, a joint optimization phase, and a fine-tuning phase. This progressive training approach enables the model to focus on different learning objectives at each stage, avoiding interference during the optimization process and ensuring the coordinated development of intent mapping relationships and recommendation tasks.

[0047] The warm-up phase focuses on training the intent bridging network, temporarily freezing the parameters of the source and target domain models. Feature representations are extracted using pre-trained models from both the source and target domains, and then the intent bridging network is trained to establish a reasonable intent mapping relationship. By minimizing the intent alignment loss, the model can identify which intents can be effectively transferred and which intents are too different to be suitable for transfer, thereby reducing the risk of inappropriate knowledge transfer.

[0048] The joint optimization phase simultaneously optimizes the parameters of the intent bridging network and the target domain model, coordinating knowledge transfer and target domain task learning. By simultaneously optimizing the recommendation task loss and the intent alignment loss, the model ensures that knowledge transfer is actually helpful for the target domain task, naturally suppressing ineffective or harmful knowledge transfer. The parameters of the source domain model remain fixed and are not updated, ensuring the stability of the source domain knowledge.

[0049] The fine-tuning phase fixes the learned intent mappings and focuses on optimizing the target domain model performance. This phase allows the model to further adjust parameters specific to the target domain while retaining valuable transferred knowledge, mitigating performance degradation that can occur due to over-transfer. By reducing the constraints on intent alignment, the model can focus more on adapting to the characteristics of the target domain.

[0050] After completing the three-stage training, the model is ready for prediction. Given a user u and item i in the target domain, the model first uses a decoupled feature extractor to obtain their representations in various intent subspaces. Then, through an intent bridging network and an adaptive knowledge fusion mechanism, it transfers relevant knowledge from the source domain to the target domain and generates the final user and item representations. Finally, by calculating the similarity (e.g., inner product) between the user and item representations, it generates a prediction score for use in recommendation decisions.

[0051] A specific example application scenario of the method of the present invention is cross-category recommendation in an e-commerce platform. For example, a user has a rich history of interaction in the field of electronic products, but is a new user in the field of household goods. Through the method of the present invention, the system can identify the various intentions expressed by users in the field of electronic products (such as brand preference, price sensitivity, functional requirements, etc.), and map these intentions to the corresponding intention dimensions in the field of household goods through the intention bridging network. For quality intentions that users may be highly concerned about, the system will give a higher migration weight; for domain-specific intentions (such as the technical specifications of electronic products), it may give a lower migration weight. Through this fine-grained selective knowledge transfer, the system can provide the user with more personalized and accurate recommendations in the field of household goods.

[0052] Through the above steps, the present invention can achieve more accurate and efficient cross-domain knowledge transfer while maintaining the integrity of the intent decoupling structure, especially when the target domain data is sparse, significantly improving the recommendation performance.

[0053] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A cross-domain knowledge transfer cold start recommendation method based on dynamic intent perception, characterized by: The recommendation system model adopted by the method includes a decoupled feature extractor based on graph convolution, an intent bridging network and an adaptive knowledge fusion mechanism; the steps of the method are: S1. Decompose the user and item representation spaces in the source and target domains into multiple intent subspaces using a graph convolution-based decoupled feature extractor. Each intent subspace captures the interaction pattern under a specific intent. S2. Establish an intent bridging network to automatically discover the intent correspondence between different domains through a learnable intent mapping matrix; S3. Design an adaptive knowledge fusion mechanism to dynamically adjust the degree of source domain knowledge transfer according to the characteristics of the target domain samples; S4. A three-stage progressive training strategy is adopted, including a warm-up stage, a joint optimization stage, and a fine-tuning stage, to ensure that the model can stably and effectively learn cross-domain intent mapping relationships and recommendation tasks.

2. The cross-domain knowledge transfer cold start recommendation method based on dynamic intent perception according to claim 1 is characterized by: The S1 is specifically: (1) The graph convolution-based decoupled feature extractor evenly divides the d-dimensional embedding vectors of user u and item i into K sub-vectors, each of which represents a representation of a potential intent; (2) Through intent-aware graph convolution operations, the interaction patterns are learned separately in each intent subspace, and the representations of each layer are fused using residual connections to retain structural information at different levels.

3. The cross-domain knowledge transfer cold start recommendation method based on dynamic intent perception according to claim 1 is characterized by: The S2 is specifically: (1) The intention bridging network establishes an intention mapping matrix M, where each element M ij represents the association strength between the i-th intention in the source domain and the j-th intention in the target domain; (2) Normalize the mapping matrix column-wise to ensure that each target domain intent is a weighted combination of different source domain intents, and the sum of the weights is 1; (3) Based on the normalized mapping matrix, the intent representation in the source domain is mapped to the intent space of the target domain; (4) The intent mapping relationship is optimized through mapping similarity loss, mapping sparsity loss and mapping coverage loss. The mapping similarity loss encourages the mapped source domain representation and the target domain representation to have high similarity in the corresponding intent subspace; the mapping sparsity loss enables each source domain intent to be mainly mapped to a few target domain intents; the mapping coverage loss ensures that each intent in the target domain can obtain sufficient knowledge from the source domain.

4. The cross-domain knowledge transfer cold start recommendation method based on dynamic intent perception according to claim 1 is characterized by: The S3 is specifically: (1) Construct a context vector containing the complete feature information of the current interaction pair; (2) Calculate the gate value through the multi-layer perceptron structure to control the influence of source domain knowledge on the final representation; (3) Based on the gated value, the source domain mapped representation is weightedly fused with the target domain original representation to obtain the final representation.

5. The cross-domain knowledge transfer cold start recommendation method based on dynamic intent perception according to claim 1 is characterized by: The S4 is specifically: (1) The warm-up phase focuses on training the intent bridging network while temporarily freezing the parameters of the source and target domain models; (2) The joint optimization phase simultaneously optimizes the parameters of the intent bridging network and the target domain model, coordinating knowledge transfer and target domain task learning; (3) The fine-tuning phase fixes the learned intent mapping relationship and focuses on optimizing the performance of the target domain model.

Images