Cold start recommendation method and system based on dynamic adaptive fusion of multi-modal features

Abstract

This invention discloses a cold-start recommendation method and system based on dynamic adaptive fusion of multimodal features. First, a multimodal semantic graph is constructed to extract features. Second, cold-start items are identified by calculating the deviation between the number of item interactions and the average number of item interactions. Third, a neighborhood set of cold-start items is extracted, and its average neighborhood similarity in the multimodal semantic graph is calculated. Next, base weights are calculated based on the number of item interactions, and a dynamic index is calculated by combining the average neighborhood similarity to determine the final weights. Then, the enhanced interaction matrix is ​​symmetrically normalized, and multi-signal extraction is performed on the symmetrically normalized enhanced interaction matrix. The signals are then fused to obtain the final preference probability. Finally, a Top-K recommendation list is generated based on the final preference probability as the final recommendation result. This invention can achieve high-precision, low-latency, and highly reliable personalized recommendations in cold-start scenarios, and is suitable for large-scale real-time recommendation platforms such as e-commerce and content distribution.

Description

Technical Field

[0001] This invention relates to the field of recommendation system technology, specifically to a cold-start recommendation method and system based on dynamic adaptive fusion of multimodal features. Background Technology

[0002] With the rapid development of artificial intelligence technology, large-scale pre-trained models have long been capable of extracting high-semantic-density feature embeddings from raw text and images. For example, BERT for text understanding and VisionTransformers for image recognition have become core technological supports for the development of Multimodal Recommender Systems (MRSs). In practical applications such as e-commerce and content distribution, users' demand for personalized recommendations continues to rise. However, traditional recommendation systems have a fatal flaw: they rely too heavily on user-item interaction data. Once they encounter a "cold start" scenario—that is, when new users have no historical behavior records and new items have no accumulated interaction data—performance drops drastically. New items have no chance to reach potential users, and new users are lost because they cannot receive suitable recommendations, becoming a key bottleneck restricting platforms from improving user experience and maintaining a healthy ecosystem.

[0003] However, multimodal recommendation systems provide a core approach to the cold start problem. By integrating non-interactive modal information such as text (e.g., product function descriptions, content titles) and visual information (e.g., product images, video covers), even without any interactive data, they can infer the semantic attributes and potential user preferences of items based on their content features, thereby providing a basis for recommendations for cold start entities (new users, new items).

[0004] However, in actual cold start scenarios, multimodal recommendation systems still cannot avoid the difficult dilemma of reconciling "accuracy-efficiency-reliability", which includes the following aspects:

[0005] The mainstream approach is based on Graph Convolutional Networks (GCNs). While these models can indeed improve recommendation accuracy by modeling higher-order relationships between users and items, and between items themselves, their computational cost is exorbitantly high, making them unsuitable for the real-time requirements of cold-start scenarios. These models require building multimodal item similarity graphs and user-item interaction graphs before using multiple rounds of graph convolutional layers for message passing and parameter iteration. In large-scale data scenarios like e-commerce platforms with millions of items, training time can easily reach hours. The inference phase is even more problematic; as new items are dynamically added, the graph structure needs to be readjusted and parameters optimized, making it difficult to quickly integrate into the recommendation process. For example, during major e-commerce promotions, thousands of new products are added daily. The inefficiency of GCN models means that these cold-start products cannot be recommended in a timely manner, meaning that the technological advantages often fail to translate into actual business value.

[0006] Another type of training-independent method, such as Singular Value Decomposition (SVD) and Graph Filtering (GF), avoids parameter iteration by relying on deterministic matrix operations or graph signal processing, compressing training time to minutes or even seconds, which meets the efficiency requirements of cold start scenarios. However, the fixed weights used in feature aggregation cannot distinguish the differences in the "reliability" of multimodal content. In cold start scenarios, the quality of content features of new items is inherently uneven; for example, product images may be blurry, functional descriptions may be ambiguous, and text summaries may lack key information. Fixed weights will treat these low-quality content features as valid signals and include them in the recommendation process, which will actually exacerbate the problem of "false recommendations." For example, pushing visually similar but functionally unrelated products to users will not only fail to improve recommendation accuracy but will also reduce user satisfaction.

[0007] More challenging is the inherent contradiction between modal heterogeneity and fusion efficiency, which further limits the performance of multimodal recommendation systems in cold-start scenarios. Text, visual data, and user interaction data originate from different semantic spaces: text carries logical semantic relationships, visual data reflects intuitive feature attributes, and interaction data embodies dynamic user behavior preferences. Moreover, these three types of data differ significantly in data structure (text is a sequence, visual data is a matrix, and interaction data is a sparse graph) and statistical distribution. Existing models either simply concatenate multimodal features, easily leading to semantic misalignment, or design complex cross-modal fusion mechanisms, such as multi-layer attention networks and contrastive learning modules. While these can alleviate the heterogeneity problem, they further increase the computational burden, resulting in a vicious cycle of "increased accuracy → decreased efficiency." Cold-start scenarios require both "high accuracy" and "fast response," and this cycle clearly cannot balance these two core requirements.

[0008] Furthermore, most existing technologies fail to consider the reliability assessment of multimodal content in cold start scenarios, resulting in insufficient robustness of recommendation systems. The content features of cold start items often lack verification through human review or user feedback, leading to a high proportion of low-quality content. However, existing models lack a content quality screening mechanism for cold start items, blindly fusing all content features and interaction signals. This not only fails to effectively compensate for the sparsity of interaction data but also introduces a large amount of noise, exacerbating the cold start recommendation performance. For example, if a newly launched garment has blurred visual features due to the shooting angle, existing models will perform similarity matching with the visual features of other garments and ultimately recommend it to users who prefer a clearer style. This degrades the user experience and causes a loss of conversion rates.

[0009] In summary, existing multimodal recommendation systems cannot simultaneously meet the three core requirements of "high accuracy", "high computational efficiency" and "high content reliability" in cold start scenarios. Therefore, there is an urgent need for an innovative solution that can overcome these technical bottlenecks to adapt to the performance and real-time requirements of cold start recommendation in practical applications. Summary of the Invention

[0010] This invention addresses the shortcomings of existing multimodal recommendation systems in achieving a balance between accuracy, efficiency, and reliability in cold-start scenarios by proposing a cold-start recommendation method and system based on dynamic adaptive fusion of multimodal features.

[0011] In a first aspect, the present invention provides a cold-start recommendation method based on dynamic adaptive fusion of multimodal features, comprising the following steps:

[0012] Multimodal semantic graph construction: The visual feature matrix and the text feature matrix are normalized, and the multimodal similarity is calculated based on the normalized features. The multimodal similarity matrix is ​​obtained by weighted fusion. Then, a sparse weighted adjacency matrix is ​​constructed using a dual filtering strategy and symmetric normalization is performed on it.

[0013] Cold start item identification: Based on the user-item interaction matrix, cold start items are identified by calculating the deviation between the number of item interactions and the average number of item interactions in the dataset;

[0014] Content reliability assessment: Extract the neighborhood set of the cold-start item from the sparse weighted adjacency matrix, and calculate its average neighborhood similarity in the multimodal semantic graph as a content reliability measure;

[0015] Dynamic fusion weight calculation: The base weight is calculated based on the number of item interactions, and a dynamic index is calculated by combining the average similarity of the neighborhood. The base weight is amplified by the index to suppress the fusion of low-quality content and to determine the final weight of the item.

[0016] Graph fusion: Construct a diagonal matrix based on the final weights of the items, propagate the content signal through sparse matrix multiplication, and then weightedly fuse the original interaction signal and the content signal to obtain an enhanced interaction matrix;

[0017] Interaction matrix normalization: Based on the enhanced interaction matrix, construct an angle matrix, add a numerical stability constant, and then perform symmetric normalization on the enhanced interaction matrix;

[0018] Multi-signal extraction: Perform training-free singular value decomposition on the symmetrically normalized enhanced interaction matrix to calculate the user's personalized preference signal for items, and extract high-order collaborative signals through single graph propagation;

[0019] Signal fusion and scoring: The personalized preference signal and the higher-order collaborative signal are normalized, and after being weighted and fused according to the signal fusion weight, the fusion result is compressed to obtain the user's final preference probability for the item;

[0020] Recommendation list generation: A Top-K recommendation list is generated based on the final preference probability as the final recommendation result.

[0021] Secondly, the present invention provides a cold-start recommendation system based on dynamic adaptive fusion of multimodal features, comprising:

[0022] The multimodal semantic graph construction module is used to normalize the visual feature matrix and the text feature matrix, and calculate the multimodal similarity based on the normalized features. The multimodal similarity matrix is ​​obtained by weighted fusion. Then, a sparse weighted adjacency matrix is ​​constructed using a dual filtering strategy and symmetric normalization is performed on it.

[0023] The cold start item recognition module is used to identify cold start items based on the user-item interaction matrix by calculating the deviation between the number of item interactions and the average number of item interactions in the dataset.

[0024] The content reliability assessment module is used to extract the neighborhood set of the cold start item from the sparse weighted adjacency matrix and calculate its neighborhood average similarity in the multimodal semantic graph as a content reliability metric.

[0025] The dynamic fusion weight calculation module is used to calculate the base weight based on the number of item interactions, calculate the dynamic index by combining the neighborhood average similarity, suppress the fusion of low-quality content by amplifying the base weight through the index, and determine the final weight of the item.

[0026] The graph fusion module is used to construct a diagonal matrix based on the final weights of the items, propagate the content signal through sparse matrix multiplication, and then weightedly fuse the original interaction signal and the content signal to obtain an enhanced interaction matrix.

[0027] The interaction matrix normalization module is used to construct an angle matrix based on the enhanced interaction matrix, add a numerical stability constant, and then perform symmetric normalization on the enhanced interaction matrix.

[0028] The multi-signal extraction module is used to perform training-free singular value decomposition on the symmetric normalized enhanced interaction matrix, calculate the user's personalized preference signal for items, and extract high-order cooperative signals through single graph propagation.

[0029] The signal fusion and scoring module is used to normalize the personalized preference signal and the higher-order collaborative signal, and after weighted fusion according to the signal fusion weight, compress the fusion result to obtain the user's final preference probability for the item.

[0030] The recommendation list generation module is used to generate a Top-K recommendation list as the final recommendation result based on the final preference probability.

[0031] The beneficial effects of this invention are:

[0032] In scenarios where new user / item interaction data is sparse and multimodal features (visual / textual) are heterogeneous, this invention not only ensures that the recommendation results match the user's potential preferences and the semantic attributes of the items, but also fully considers the need for rapid integration of new users / items, thus achieving more accurate and efficient personalized recommendations.

[0033] Meanwhile, this invention introduces a content reliability assessment and dynamic graph fusion strategy to effectively filter low-quality multimodal features. While reducing the risk of noise interference, it minimizes recommendation bias caused by modality misalignment or data sparsity, thereby improving recommendation quality while ensuring the efficiency and fairness of the system. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0035] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention. Detailed Implementation

[0036] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0037] This invention provides a cold-start recommendation method based on dynamic adaptive fusion of multimodal features, such as... Figure 1 As shown, the core logic and data flow of the three stages of multimodal semantic graph construction, adaptive graph alignment, and recommendation generation are illustrated, specifically including the following steps:

[0038] S1: Multimodal Semantic Graph Construction

[0039] Furthermore, the multimodal semantic graph construction aims to generate high-quality sparse semantic graphs to provide reliable content signals, including the following steps:

[0040] First, the visual feature matrix With text feature matrix Perform L2 normalization separately: Formula To ensure fairness in similarity calculation, feature scale differences are eliminated; then, multimodal similarity is calculated based on normalized features, where the visual similarity matrix is:

[0041]

[0042] Text similarity matrix:

[0043]

[0044] Then, the multimodal similarity matrix is ​​obtained by weighted fusion with a modal fusion weight of λ=0.1:

[0045]

[0046] Subsequently, a sparse weighted adjacency matrix A is constructed using a dual filtering strategy of "K-nearest neighbor selection + similarity threshold". The rule is to retain only the k=5 nearest neighbors of item i with similarity. 0.75 connection (otherwise) ), reducing matrix complexity from Down to Finally, to mitigate the nodal degree bias, A is symmetrically normalized. D is a diagonal matrix of A. This ensures that the information dissemination influence of all nodes is balanced, and cold-start items can effectively aggregate information from neighboring regions.

[0047] S2: Cold Start Item Recognition

[0048] Furthermore, the cold start item identification is based on the sparsity of interaction data to define cold start items, providing targets for subsequent targeted fusion, including the following steps:

[0049] First, the number of interactions for each item i is calculated based on the user-item interaction matrix R. Next, calculate the average number of item interactions across the datasets (16.82 for the Baby dataset and 16.15 for the Sports dataset), and set the cold start threshold θ to "average number of interactions ± 2" (θ = 17 for the Baby dataset and θ = 16 for the Sports dataset); finally, determine the cold start item: if... If it is a cold start item, it requires subsequent content signal enhancement. This means that popular items rely solely on interaction signals, avoiding the introduction of noise through content signals.

[0050] S3: Content Reliability Assessment

[0051] Furthermore, the content reliability assessment quantifies the cold-start item content quality based on "neighborhood consistency" to avoid low-quality content contaminating recommendations, and includes the following steps:

[0052] First, for each cold-start item i, extract its neighborhood set from the sparse weighted adjacency matrix A. Then calculate the average similarity of the item's neighborhood. ,like If a valid neighborhood exists, then ,like 0, an isolated node, has no valid neighborhood. , Furthermore, the closer the value is to 1, the higher the content quality; finally, all cold start items... Stored as a reliability score vector s, it is used for subsequent dynamic fusion weight calculation.

[0053] S4: Dynamic Fusion Weight Calculation

[0054] Furthermore, the dynamic fusion weight calculation combines the cold start degree of an item with the reliability of its content to calculate dynamic weights, achieving "targeted enhancement of cold start and avoidance of noise from popular items," including the following steps:

[0055] First, the base weights are calculated using an inverse power law function based on the number of item interactions. (p=0.1 controls the decay rate). The smaller the value (the higher the cold start capability), the better. The larger the value, the more content enhancements the cold-start item receives; then, based on the average similarity of the item's neighborhood... Calculate the dynamic index (Optional, preset minimum threshold) Preset maximum threshold ). The smaller the value (the less reliable the content), the less reliable it is. The larger the base weight, the more it is used to suppress the fusion of low-quality content through exponential amplification of the base weight decay; finally, the final weight is determined. Apply only to cold start items Non-cold start items This ensures that high-quality content signals for cold-start items are utilized and that popular items have no additional noise.

[0056] S5: Graph Fusion

[0057] Furthermore, the graph fusion accurately fuses semantic graph content signals and interaction graph behavior signals to generate an enhanced interaction matrix, including the following steps:

[0058] First, based on the final weight Construct two diagonal matrices: the fusion weight matrix (Used for content signal weighting), original signal matrix (Used for weighting the original interaction signal) (m is the total number of items); then the content signal is propagated through sparse matrix multiplication, the formula is: ( (This is the normalized semantic graph adjacency matrix). Finally, the original interaction signals and content signals are weighted and fused to obtain the enhanced interaction matrix. + This matrix retains reliable interaction signals for popular items while supplementing high-quality content signals for cold-start items, thus resolving the contradiction between "behavioral sparsity" and "content noise".

[0059] S6: Interaction Matrix Normalization

[0060] Furthermore, the interaction matrix normalization aims to suppress popularity bias and provide a fair preference representation for signal extraction, including the following steps:

[0061] Firstly, based on the enhanced interaction matrix Constructing the angle matrix: User-defined angle matrix (Total number of user interactions), item pair angle matrix (Total number of interactions with item i); plus a numerical stability constant. (To avoid division by zero), for Perform symmetric normalization, the formula is as follows After normalization It expresses "relative preference intensity" rather than "absolute interaction frequency" to ensure that cold-start items are evaluated on the same scale as popular items.

[0062] S7: Multi-signal extraction

[0063] Furthermore, the multi-signal extraction aims to extract complementary preference signals from individual and group dimensions, including the following steps:

[0064] First, a training-free singular value decomposition (SVD) is used to extract personalized preference signals. The decomposition yields the user feature matrix U, the singular value matrix S, and the item feature matrix V (latent dimension k=64), which are then processed using the formula... Calculate user u's direct personalized preference for item i; then extract higher-order collaborative signals through one-time graph propagation, first calculating the collaborative similarity between users. Multiply by the left again To achieve preference propagation, the formula is: , It reflects implicit connections at the group level and is especially suitable for cold-start users without direct interaction.

[0065] S8: Signal Fusion and Scoring

[0066] Furthermore, the signal fusion and scoring aims to fuse the two signals to generate the user's final preference probability for the item, including the following steps:

[0067] First of all and Min-Max normalization is performed separately to compress the values ​​to the [0,1] interval to eliminate scale differences; then, the values ​​are weighted and fused according to a signal fusion weight γ=0.5 (balancing individual and group preferences), as shown in the formula. Finally, the fusion result is compressed to the (0,1) interval using the Sigmoid function to obtain the final preference probability. The closer the probability value is to 1, the higher the user's preference level and the higher the recommendation priority.

[0068] S9: Recommendation List Generation

[0069] Furthermore, the generation of the recommendation list, based on the final preference probability, generates a Top-K recommendation list, including the following steps:

[0070] First, for each test user u, filter out items they have already interacted with in the training / validation set to avoid duplicate recommendations; then, sort the remaining items according to... Sort the items in descending order and select the top K items (in this example, K=10 / 20, which conforms to the conventional settings of recommendation systems); finally, use the Top-K items as the final recommendation result to complete one recommendation process.

[0071] Therefore, in this embodiment of the invention, when a new item is listed, the system first acquires the visual and textual features of the item, calculates multimodal similarity after normalization, constructs a sparse semantic graph through K-nearest neighbor filtering and a similarity threshold, and performs symmetric normalization to balance node influence. New users / items do not need to participate in model iteration training; they only need to quickly extract potential factors from the sparse interaction matrix through training-free SVD. Subsequently, the system identifies cold-start items, evaluates their content reliability based on neighborhood consistency, calculates dynamic fusion weights based on the number of interactions, and weights and fuses the semantic graph and user-item interaction graph to generate an enhanced interaction matrix. After normalizing the enhanced matrix, the recommendation module fuses the personalized preference signal extracted by SVD with the high-order collaborative signal obtained by graph propagation, selects the optimal recommendation list, and feeds it back to the user. After a user interacts with an item, the system updates the content reliability score based on the interaction result, providing a basis for subsequent weight adjustments for cold-start recommendations.

[0072] Furthermore, the embodiments of the present invention will demonstrate the implementation results:

[0073] This invention is validated based on publicly available benchmark datasets, and the data distribution is shown in Table 1. This invention selects the Amazon Baby and Sports datasets from an e-commerce scenario, covering the typical cold start and data sparsity characteristics of multimodal recommendation. Item visual features are extracted using a pre-trained ResNet-50 dataset, and text features are extracted using a pre-trained BERT-base dataset. Cold start users are defined as users with fewer than 5 interactions in the training set, to simulate the scarcity of new user data in real-world scenarios.

[0074] Table 1 Data Parameter Distribution

[0075]

[0076] Further, the implementation results are shown in Tables 2, 3, and 4. Table 2 in this embodiment shows the impact of different sensitivity parameters p on the NDCG@20 of cold-start users. The data in Table 2 shows that p has a significant regulatory effect on the system's cold-start performance. Within the range of p ∈ [0.08, 0.12], the system maintains a high NDCG@20 value of 0.0415-0.0428 with good stability, confirming that this range can effectively balance the content signal fusion intensity of cold-start items. The performance is optimal when p = 0.1, with NDCG@20 reaching 0.0428. When p < 0.08, the base weight attenuation is insufficient, leading to excessive fusion of low-quality content signals, and the performance drops to 0.0392. When p > 0.12, the attenuation is too strong, suppressing the enhancement effect of the content signal on cold-start items, and the performance drops to 0.0385. This indicates that parameter p needs to be controlled within a reasonable range to balance reliability and enhancement effect.

[0077] Table 2. Effects of different sensitivity parameters p on cold-start users NDCG@20

[0078]

[0079] Table 3 in this embodiment shows the impact of different modal fusion coefficients λ on the overall recommendation Recall@10. Experimental data show that the value of λ directly affects the synergistic effect of multimodal signals. When λ∈[0.1,0.3], the system Recall@10 maintains a high level of 0.061-0.0635 (Baby dataset), with the best performance when λ=0.1 (Baby: 0.0635, Sports: 0.0747) – at this ratio, the semantic association of text features and the intuitive attributes of visual features effectively complement each other. When λ<0.1, the weight of text features is too high, which easily introduces ambiguous semantic noise, and the performance drops to 0.058; when λ>0.3, the weight of visual features is too high, leading to false recommendations of "visual similarity but functional irrelevance", and the performance drops to 0.057, indicating that λ needs to be within the range of [0.1,0.3] to balance the contributions of the two modalities.

[0080] Table 3. Impact of different modal fusion coefficients λ on overall recommendation Recall@10

[0081]

[0082] Table 4 in this embodiment shows a comparison between this invention and other mainstream recommendation schemes. BPR is a classic matrix factorization method, LightGCN is a lightweight GCN model, VBPR is an early multimodal model, LATTICE and SLMRec are current methods in the multimodal recommendation field, this invention-V1 is a simplified version retaining only collaborative signals, and this invention-V2 is a simplified version retaining only content signals. Experiments compared the performance of the seven schemes on Recall@10 and NDCG@10 for cold-start users (interaction count less than or equal to 5): the static methods BPR and VBPR performed the worst (Recall@10≤0.045), reflecting their inability to adapt to cold-start scenarios; while GCN-like models LightGCN and LATTICE showed improvement, their cold-start performance remained lagging due to computational efficiency limitations (Recall@10≤0.055); the schemes in this embodiment performed best, with the basic version achieving Recall@10 of 0.0662 and NDCG@10 of 0.0362, representing improvements of 56.13% and 54.70% respectively compared to the strongest baseline LATTICE. Data shows that the "cooperative-content" dual-signal fusion strategy of the present invention is significantly better than the single-signal scheme. Even the simplified versions (V1, V2) are better than the traditional baseline, achieving a good balance between accuracy and cold start adaptability.

[0083] Table 4 Comparison of cold start user performance with other recommendation schemes (Baby dataset)

[0084]

[0085] Based on the same concept as the above method embodiments, this invention also proposes a cold-start recommendation system based on dynamic adaptive fusion of multimodal features, including:

[0086] The multimodal semantic graph construction module is used to normalize the visual feature matrix and the text feature matrix, and calculate the multimodal similarity based on the normalized features. The multimodal similarity matrix is ​​obtained by weighted fusion. Then, a sparse weighted adjacency matrix is ​​constructed using a dual filtering strategy and symmetric normalization is performed on it.

[0087] The cold start item recognition module is used to identify cold start items based on the user-item interaction matrix by calculating the deviation between the number of item interactions and the average number of item interactions in the dataset.

[0088] The content reliability assessment module is used to extract the neighborhood set of the cold start item from the sparse weighted adjacency matrix and calculate its neighborhood average similarity in the multimodal semantic graph as a content reliability metric.

[0089] The dynamic fusion weight calculation module is used to calculate the base weight based on the number of item interactions, calculate the dynamic index by combining the neighborhood average similarity, suppress the fusion of low-quality content by amplifying the base weight through the index, and determine the final weight of the item.

[0090] The graph fusion module is used to construct a diagonal matrix based on the final weights of the items, propagate the content signal through sparse matrix multiplication, and then weightedly fuse the original interaction signal and the content signal to obtain an enhanced interaction matrix.

[0091] The interaction matrix normalization module is used to construct an angle matrix based on the enhanced interaction matrix, add a numerical stability constant, and then perform symmetric normalization on the enhanced interaction matrix.

[0092] The multi-signal extraction module is used to perform training-free singular value decomposition on the symmetric normalized enhanced interaction matrix, calculate the user's personalized preference signal for items, and extract high-order cooperative signals through single graph propagation.

[0093] The signal fusion and scoring module is used to normalize the personalized preference signal and the higher-order collaborative signal, and after weighted fusion according to the signal fusion weight, the fusion result is compressed by the Sigmoid function to obtain the user's final preference probability for the item.

[0094] The recommendation list generation module is used to generate a Top-K recommendation list as the final recommendation result based on the final preference probability.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A cold-start recommendation method based on dynamic adaptive fusion of multimodal features, characterized in that, Includes the following steps: Multimodal semantic graph construction: The visual feature matrix and the text feature matrix are normalized, and the multimodal similarity is calculated based on the normalized features. The multimodal similarity matrix is ​​obtained by weighted fusion. Then, a sparse weighted adjacency matrix is ​​constructed using a dual filtering strategy and symmetric normalization is performed on it. Cold start item identification: Based on the user-item interaction matrix, cold start items are identified by calculating the deviation between the number of item interactions and the average number of item interactions in the dataset; Content reliability assessment: Extract the neighborhood set of the cold-start item from the sparse weighted adjacency matrix, and calculate its average neighborhood similarity in the multimodal semantic graph as a content reliability measure; Dynamic fusion weight calculation: The base weight is calculated based on the number of item interactions, and a dynamic index is calculated by combining the average similarity of the neighborhood. The base weight is amplified by the index to suppress the fusion of low-quality content and to determine the final weight of the item. Graph fusion: Construct a diagonal matrix based on the final weights of the items, propagate the content signal through sparse matrix multiplication, and then weightedly fuse the original interaction signal and the content signal to obtain an enhanced interaction matrix; Interaction matrix normalization: Based on the enhanced interaction matrix, construct an angle matrix, add a numerical stability constant, and then perform symmetric normalization on the enhanced interaction matrix; Multi-signal extraction: Perform training-free singular value decomposition on the symmetrically normalized enhanced interaction matrix to calculate the user's personalized preference signal for items, and extract high-order collaborative signals through single graph propagation; Signal fusion and scoring: The personalized preference signal and the higher-order collaborative signal are normalized, and after being weighted and fused according to the signal fusion weight, the fusion result is compressed to obtain the user's final preference probability for the item; Recommendation list generation: A Top-K recommendation list is generated based on the final preference probability as the final recommendation result.

2. The method according to claim 1, characterized in that, The dual filtering strategy includes K-nearest neighbor filtering and similarity thresholding. By retaining connections between items that are close to each other and have a similarity greater than the threshold, a sparse weighted adjacency matrix is ​​constructed.

3. The method according to claim 1 or 2, characterized in that, If the number of interactions with an item is lower than the cold start threshold, it is identified as a cold start item. When the number of interactions with an item is not less than the cold start threshold, it is determined to be a popular item, relying solely on interaction signals to avoid introducing noise from content signals.

4. The method according to claim 3, characterized in that, The base weight is calculated using an inverse power law based on the number of item interactions. The fewer the number of item interactions, the larger the base weight, ensuring that cold-start items receive more content enhancements.

5. The method according to claim 4, characterized in that, The dynamic index calculation process is as follows: Obtain the average similarity of the item's neighborhood, the preset minimum threshold, and the preset maximum threshold; Calculate the threshold range, which is the difference between the maximum threshold and the minimum threshold; Calculate the content unreliability, where the content unreliability is the difference between 1 and the average similarity of the item's neighborhood; Multiplying the threshold range by the content unreliability yields the dynamic adjustment amount; The dynamic index is obtained by adding the dynamic adjustment amount to the minimum threshold.

6. The method according to claim 1, characterized in that, The diagonal matrix includes a fused weight matrix and an original signal matrix; The fusion weight matrix is ​​used for content signal weighting; The original signal matrix is ​​used for weighting the original interactive signals.

7. The method according to claim 1 or 6, characterized in that, The training-free singular value decomposition is performed on the symmetrically normalized enhanced interaction matrix to obtain the user feature matrix, singular value matrix and item feature matrix, and then the user's direct personalized preference for items is calculated.

8. The method according to claim 7, characterized in that, The extraction of high-order cooperative signals via single-graph propagation includes: Calculate the collaboration similarity between users; Multiply by the normalized enhanced interaction matrix on the left to achieve preference propagation.

9. The method according to claim 1, characterized in that, Based on the final preference probability, a Top-K recommendation list is generated as the final recommendation result, and the following steps are performed: First, for each test user, filter out the items they have already interacted with in the training / validation set to avoid duplicate recommendations; Secondly, sort the remaining items in descending order of their final preference probability and select the top K items; Finally, the Top-K items are used as the final recommendation results to complete a recommendation process.

10. A cold-start recommendation system based on dynamic adaptive fusion of multimodal features, characterized in that, include: The multimodal semantic graph construction module is used to normalize the visual feature matrix and the text feature matrix, and calculate the multimodal similarity based on the normalized features. The multimodal similarity matrix is ​​obtained by weighted fusion. Then, a sparse weighted adjacency matrix is ​​constructed using a dual filtering strategy and symmetric normalization is performed on it. The cold start item recognition module is used to identify cold start items based on the user-item interaction matrix by calculating the deviation between the number of item interactions and the average number of item interactions in the dataset. The content reliability assessment module is used to extract the neighborhood set of the cold start item from the sparse weighted adjacency matrix and calculate its neighborhood average similarity in the multimodal semantic graph as a content reliability metric. The dynamic fusion weight calculation module is used to calculate the base weight based on the number of item interactions, calculate the dynamic index by combining the neighborhood average similarity, suppress the fusion of low-quality content by amplifying the base weight through the index, and determine the final weight of the item. The graph fusion module is used to construct a diagonal matrix based on the final weights of the items, propagate the content signal through sparse matrix multiplication, and then weightedly fuse the original interaction signal and the content signal to obtain an enhanced interaction matrix. The interaction matrix normalization module is used to construct an angle matrix based on the enhanced interaction matrix, add a numerical stability constant, and then perform symmetric normalization on the enhanced interaction matrix. The multi-signal extraction module is used to perform training-free singular value decomposition on the symmetric normalized enhanced interaction matrix, calculate the user's personalized preference signal for items, and extract high-order cooperative signals through single graph propagation. The signal fusion and scoring module is used to normalize the personalized preference signal and the higher-order collaborative signal, and after weighted fusion according to the signal fusion weight, compress the fusion result to obtain the user's final preference probability for the item. The recommendation list generation module is used to generate a Top-K recommendation list as the final recommendation result based on the final preference probability.

Images