Cross-domain AI knowledge aggregation method based on collaborative filtering

Abstract

The invention discloses a cross-domain AI knowledge aggregation method based on collaborative filtering. The method comprises the steps that S1, multi-source heterogeneous AI knowledge data and user behavior data are collected and preprocessed; s2, constructing a double-tower cross-domain embedded network, and outputting a cross-domain semantic fusion sequence; s3, modeling and analyzing user preferences through the improved Bi-GRU network, and generating user behavior preference vectors; s4, performing semantic diffusion and neighborhood reasoning on the cold start user, and complementing interest features; s5, adopting a double-tower recall structure and an XGBoost model to sort and generate a cross-domain recommendation list; s6, constructing a context rule base to execute context adaptability judgment, and generating a matched knowledge aggregation recommendation list; and S7, performing incremental learning according to user feedback information, and dynamically updating the double-tower cross-domain embedded network and the improved Bi-GRU network. According to the method, the knowledge matching precision, the cold start adaptability and the scene adaptation capability of cross-domain recommendation content are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence knowledge management and recommendation technology, and in particular to a cross-domain AI knowledge aggregation method based on collaborative filtering. Background Technology

[0002] With the rapid development of artificial intelligence technology and the increasing abundance of knowledge resources across industries, the organization, aggregation, and intelligent recommendation of cross-domain AI knowledge have gradually become important research directions in knowledge management. Currently, AI knowledge exhibits characteristics such as multi-source heterogeneity, wide distribution, and frequent updates, encompassing various forms including structured knowledge graphs, semi-structured API documents, and unstructured scientific research corpora, and is widely distributed across different platforms such as model platforms, open-source communities, and scientific research databases. However, existing knowledge aggregation methods are mostly limited to single domains, single modalities, or static knowledge bases, making it difficult to achieve unified modeling and semantic alignment of heterogeneous knowledge. When facing cross-domain user behavior modeling and multi-scenario knowledge adaptation and recommendation, the following problems exist:

[0003] The semantic systems of knowledge across different domains differ significantly, and the contextual labels are inconsistent, making it impossible for traditional collaborative filtering methods to directly uncover users' cross-domain interest migration patterns, which severely restricts the generalization ability of knowledge recommendation. When faced with cold-start users or new knowledge units, existing methods lack effective completion mechanisms, which easily leads to problems such as recommendation sparsity and response failure. In complex application scenarios, recommended content often ignores the differences in the context environment when users use the content, resulting in a significant deviation between recommended knowledge and actual needs, affecting user experience and recommendation accuracy.

[0004] Therefore, how to provide a cross-domain AI knowledge aggregation method based on collaborative filtering is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a cross-domain AI knowledge aggregation method based on collaborative filtering. This invention fully integrates key technologies such as dual-tower semantic embedding networks, improved Bi-GRU behavior modeling, cold-start interest completion, dual-tower recall ranking, and context rule discrimination. It describes in detail the entire process of achieving unified knowledge modeling, user preference construction, and scenario-adaptive recommendation in a multi-source heterogeneous knowledge environment. It has the advantages of high knowledge matching accuracy, strong applicability of recommended content, and good cold-start robustness.

[0006] A cross-domain AI knowledge aggregation method based on collaborative filtering according to an embodiment of the present invention includes the following steps: S1. Collect multi-source heterogeneous AI knowledge data and user behavior data, and preprocess them to obtain a structured knowledge feature set and a user-knowledge interaction matrix; S2. Construct a dual-tower cross-domain embedding network consisting of a shared semantic tower and a domain-specific tower, perform cross-domain semantic alignment on the structured knowledge feature set, and output a cross-domain semantic fusion sequence. S3. Using an improved Bi-GRU network, user preference modeling and analysis are performed on the user-knowledge interaction matrix to generate user behavior preference vectors; S4. For cold-start users, interest features are completed based on semantic diffusion and neighborhood reasoning to obtain interest completion vectors, which are then used as auxiliary inputs to the improved Bi-GRU network. S5. A dual-tower recall structure is used to perform semantic mapping between cross-domain semantic fusion sequences and user behavior preference vectors, and non-linear scoring and ranking are performed through the XGBoost model to generate a cross-domain recommendation list. S6. Construct a context rule base, perform context adaptability judgment and filtering on the cross-domain recommendation list, and generate a knowledge aggregation recommendation list that matches the user's usage scenario; S7. Based on user feedback, perform incremental learning and dynamically update the dual-tower cross-domain embedded network and the improved Bi-GRU network.

[0007] Optionally, step S1 specifically includes: S11. The multi-source heterogeneous AI knowledge data includes structured knowledge graph data, semi-structured API interface documents, and unstructured scientific research text corpus. S12, The user behavior data includes the user's search records, citation history, access frequency and rating feedback on various knowledge domain platforms; S13, The preprocessing includes: The structured knowledge graph data is subjected to graph parsing, entity deduplication, and triple normalization operations to generate a standardized knowledge graph tensor. Perform field recognition, key-value extraction, and semantic tag mapping operations on the semi-structured API interface document data to generate an interface structure mapping table. Unstructured scientific research text corpora are subjected to word segmentation, named entity recognition, and relation extraction operations to generate text entity semantic feature vectors; Standardized knowledge graph tensors, interface structure mapping tables, and text entity semantic feature vectors are formatted and semantically aligned to form a structured knowledge feature set; Perform user identifier normalization, behavior sequence sorting, and interaction intensity filtering operations on user behavior data to generate a user-knowledge interaction matrix.

[0008] Optionally, step S2 specifically includes: S21. The dual-tower cross-domain embedding network includes a shared semantic tower, a domain-specific tower, a shared semantic mapping layer, and a gated residual fusion module. S22. The shared semantic tower uses a pre-trained SciBERT model as the encoder: The WordPiece word segmenter, built on SCIVOCAB, is used to transform each knowledge unit in the structured knowledge feature set into a token sequence. The identifier [CLS] is inserted at the beginning of the token sequence and the identifier [SEP] is inserted at the end to form the input index sequence. The input index sequence is mapped to word embeddings, position embeddings, and paragraph embeddings respectively, and then combined to form an embedding representation sequence; The embedded representation sequence is encoded with contextual semantic embedding through a 12-layer Transformer encoder, and the output is a shared semantic embedding vector. The shared semantic embedding vectors corresponding to all knowledge units in the structured knowledge feature set are combined to form a shared semantic embedding sequence; S23. The domain-specific tower performs domain modeling processing on the shared semantic embedding sequence, specifically including: For each knowledge unit to which the knowledge domain belongs, an independent Adapter module is introduced under the corresponding domain. The Adapter module includes a dimension reduction layer, a non-linear activation layer and a dimension increase layer connected in sequence. The shared semantic embedding vectors are transformed using the corresponding knowledge domain's Adapter module to generate domain-specific embedding vectors. All domain-specific embedding vectors are then combined to form a domain-specific embedding sequence. S24. The shared semantic mapping layer performs semantic space alignment processing on the shared semantic embedding sequence and the domain-specific embedding sequence through linear mapping, and inputs them into the gated residual fusion module; S25. The gated residual fusion module adopts a dual-gate structure, introducing independent fusion weights for the shared semantic embedding vector and the domain-specific embedding vector respectively, and outputting a cross-domain semantic fusion vector. The fusion weights are dynamically calculated through the Softmax function, and all cross-domain semantic fusion vectors are combined to form a cross-domain semantic fusion sequence.

[0009] Optionally, step S3 specifically includes: S31. The improved Bi-GRU network includes a behavior decomposition module, a behavior encoding module, a domain control module, an attention pooling module, and a preference fusion module; S32. The behavior decomposition module decomposes the user-knowledge interaction matrix according to the order of behavior occurrence and behavior type, generating click behavior sequence, rating behavior sequence and reference behavior sequence; S33. The behavior encoding module includes three sets of parallel Bi-GRU sub-networks. The click behavior sequence, rating behavior sequence and reference behavior sequence are concatenated with the corresponding behavior type embedding vectors and input into the corresponding Bi-GRU sub-networks respectively to perform bidirectional state encoding to obtain the click behavior hidden state, rating behavior hidden state and reference behavior hidden state at each time step. S34. The update gate structure in the Bi-GRU sub-network introduces behavior type embedding vectors for gating adjustment; S35. The domain control module introduces domain label embedding vectors to semantically correct the hidden states of click behavior, rating behavior, and reference behavior at each time step, and obtains the corrected hidden states of behavior, rating behavior, and reference behavior. The domain label embedding vector is generated by selecting the knowledge domain label to which the user behavior sequence belongs from a preset domain label set and transforming it through an embedding mapping function; S36. The attention pooling module performs attention weighted aggregation operations on the corrected click behavior hidden state sequence, the corrected rating behavior hidden state sequence, and the corrected reference behavior hidden state sequence to obtain the click behavior preference vector, the rating behavior preference vector, and the reference behavior preference vector, respectively. S37. The preference fusion module performs a weighted fusion operation on the click behavior preference vector, rating behavior preference vector and reference behavior preference vector to generate a user behavior preference vector.

[0010] Optionally, step S4 specifically includes: S41, The cold start user is a user whose user behavior data does not meet the set behavior quantity threshold; S42. Construct a knowledge semantic graph based on a structured knowledge feature set. The nodes of the knowledge semantic graph are knowledge units, and the edges are similarity scores between nodes. The semantic similarity scores are obtained by calculating the cosine similarity between the cross-domain semantic fusion vectors of the knowledge units. S43. Select the knowledge units that have been interacted with by the cold start user as the initial seed nodes, set the initial interest vector, update the interest feature vector iteratively through semantic diffusion on the knowledge semantic graph, and select the K knowledge units with the highest similarity ranking to form the interest neighborhood set. S44. Calculate the statistical distribution of the domain labels of each knowledge unit in the interest neighborhood set, construct the neighborhood label weight vector, and concatenate and fuse the interest feature vector with the neighborhood label weight vector to generate the interest completion vector. S45. The interest completion vector is used as an auxiliary input to the improved Bi-GRU network, and is concatenated with the click behavior sequence, rating behavior sequence and reference behavior sequence respectively, and then input into the corresponding Bi-GRU sub-network to participate in bidirectional state encoding.

[0011] Optionally, step S5 specifically includes: S51. The dual-tower recall structure includes a user recall sub-tower and a knowledge recall sub-tower. The user recall sub-tower uses a two-layer perceptron network to perform semantic compression mapping on the user behavior preference vector to generate a user semantic vector. The knowledge recall sub-tower uses a two-layer perceptron network that is symmetrical to the user recall sub-tower structure to perform semantic compression mapping on the cross-domain semantic fusion vector in the cross-domain semantic fusion sequence to generate a corresponding knowledge semantic vector. All knowledge semantic vectors are combined to form a knowledge semantic set. S52. Perform vector similarity calculation between the user semantic vector and each knowledge semantic vector in the knowledge semantic set using cosine similarity to obtain a similarity score; S53. Sort the knowledge semantic set according to similarity scores, select the top-M knowledge semantic vectors, and extract the corresponding knowledge units according to the index positions of the top-M knowledge semantic vectors in the structured knowledge feature set to form a preliminary recall recommendation list. S54. Concatenate the Top-M knowledge semantic vectors with the user semantic vector and the corresponding context feature vector at the feature level to generate M candidate interaction feature vectors. S55. Using the XGBoost model, input the M candidate interaction feature vectors into the XGBoost model to perform non-linear scoring, obtain the ranking score of each candidate interaction feature vector, and re-evaluate the knowledge units in the preliminary recall recommendation list based on the ranking score to generate a cross-domain recommendation list.

[0012] Optionally, step S6 specifically includes: S61. The context rule base includes a set number of context discrimination rules, and each context discrimination rule consists of a context condition item and an adaptation constraint item; The contextual conditions include the user's current usage scenario tag, terminal type tag, task urgency tag, time period tag, and geographical location tag. The adaptation constraints include the domain tag of the knowledge unit, presentation format, recommended content length, and terminal display adaptability. S62. Obtain the context feature vector corresponding to each knowledge unit in the cross-domain recommendation list, and perform Boolean discrimination operation on the context feature vector with each context discrimination rule in the context rule base in turn: if the context feature vector satisfies all context discrimination rules, then determine that the current knowledge unit matches the user's usage scenario and retain the current knowledge unit; otherwise, remove the current knowledge unit. S63. Compile the retained knowledge units into a knowledge aggregation recommendation list.

[0013] Optionally, step S7 specifically includes: S71. Collect user behavior feedback data during actual use, and supplement and update the user-knowledge interaction matrix in chronological order; S72. Based on the updated user-knowledge interaction matrix, reconstruct the click behavior sequence, rating behavior sequence, and reference behavior sequence, and input them into the improved Bi-GRU network to perform user preference update modeling and generate an incrementally optimized user behavior preference vector. S73. Based on the newly added knowledge units involved in the user feedback data, expand the structured knowledge feature set and input it into the dual-tower cross-domain embedding network to perform shared semantic embedding and domain-specific embedding processing, and update the cross-domain semantic fusion vector. S74. Calculate the predicted interaction rating based on the updated user behavior preference vector and cross-domain semantic fusion vector, minimize the mean square error between the predicted interaction rating and the actual rating in user feedback behavior, jointly optimize the dual-tower cross-domain embedding network and the improved Bi-GRU network, and regenerate the knowledge aggregation recommendation list.

[0014] The beneficial effects of this invention are: First, by constructing a dual-tower cross-domain embedding network, the collaborative modeling capabilities of the shared semantic tower and the domain-specific tower are utilized to achieve unified representation and semantic alignment of structured knowledge feature sets across different knowledge domains, thereby improving the expression quality and domain generalization ability of cross-domain semantic fusion sequences.

[0015] Secondly, an improved Bi-GRU network was designed to model and analyze user preferences in the user-knowledge interaction matrix. By using behavior decomposition, behavior encoding, and domain regulation modules, combined with attention pooling and preference fusion mechanisms, the differential encoding and unified modeling of multiple types of behaviors such as clicks, ratings, and references were achieved, which effectively improved the semantic integrity and personalized expression of user behavior preference vectors.

[0016] Furthermore, semantic diffusion and neighborhood inference are introduced to construct interest completion vectors for cold-start users, which makes up for the recommendation defects of traditional collaborative filtering methods in scenarios with insufficient user behavior; the suitability of the recommendation list is judged by combining the context rule base to ensure the acceptability and task matching of the recommended content in the use scenario.

[0017] In summary, this invention achieves unified modeling, accurate recommendation, and scenario adaptation of cross-domain AI knowledge, improving the recommendation accuracy, coverage, and user satisfaction of knowledge aggregation in heterogeneous environments, and has good versatility and expansion value. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a schematic diagram of a cross-domain AI knowledge aggregation method based on collaborative filtering proposed in this invention; Figure 2 This is a schematic diagram of the dual-tower cross-domain embedded network structure in this invention; Figure 3 This is a flowchart of cross-domain knowledge preference modeling and cold-start interest completion in this invention; Figure 4 This is a schematic diagram of the dual-tower recall and XGBoost sorting structure in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-4 A cross-domain AI knowledge aggregation method based on collaborative filtering includes the following steps: S1. Collect multi-source heterogeneous AI knowledge data and user behavior data, and preprocess them to obtain a structured knowledge feature set and a user-knowledge interaction matrix; S2. Construct a dual-tower cross-domain embedding network consisting of a shared semantic tower and a domain-specific tower, perform cross-domain semantic alignment on the structured knowledge feature set, and output a cross-domain semantic fusion sequence. S3. Using an improved Bi-GRU network, user preference modeling and analysis are performed on the user-knowledge interaction matrix to generate user behavior preference vectors; S4. For cold-start users, interest features are completed based on semantic diffusion and neighborhood reasoning to obtain interest completion vectors, which are then used as auxiliary inputs to the improved Bi-GRU network. S5. A dual-tower recall structure is used to perform semantic mapping between cross-domain semantic fusion sequences and user behavior preference vectors, and non-linear scoring and ranking are performed through the XGBoost model to generate a cross-domain recommendation list. S6. Construct a context rule base, perform context adaptability judgment and filtering on the cross-domain recommendation list, and generate a knowledge aggregation recommendation list that matches the user's usage scenario; S7. Based on user feedback, perform incremental learning and dynamically update the dual-tower cross-domain embedded network and the improved Bi-GRU network.

[0022] In this embodiment, step S1 specifically includes: S11. The multi-source heterogeneous AI knowledge data includes structured knowledge graph data, semi-structured API interface documents, and unstructured scientific research text corpus. S12, The user behavior data includes the user's search records, citation history, access frequency and rating feedback on various knowledge domain platforms; S13, The preprocessing includes: The structured knowledge graph data is subjected to graph parsing, entity deduplication, and triple normalization operations to generate a standardized knowledge graph tensor. Perform field recognition, key-value extraction, and semantic tag mapping operations on the semi-structured API interface document data to generate an interface structure mapping table. Unstructured scientific research text corpora are subjected to word segmentation, named entity recognition, and relation extraction operations to generate text entity semantic feature vectors; Standardized knowledge graph tensors, interface structure mapping tables, and text entity semantic feature vectors are formatted and semantically aligned to form a structured knowledge feature set; Perform user identifier normalization, behavior sequence sorting, and interaction intensity filtering operations on user behavior data to generate a user-knowledge interaction matrix.

[0023] In this embodiment, step S2 specifically includes: S21. The dual-tower cross-domain embedding network includes a shared semantic tower, a domain-specific tower, a shared semantic mapping layer, and a gated residual fusion module. S22. The shared semantic tower uses a pre-trained SciBERT model as the encoder: The WordPiece word segmenter, built on SCIVOCAB, is used to transform each knowledge unit in the structured knowledge feature set into a token sequence. The identifier [CLS] is inserted at the beginning of the token sequence and the identifier [SEP] is inserted at the end to form the input index sequence. The input index sequence is mapped to word embeddings, position embeddings, and paragraph embeddings respectively, and then combined to form an embedding representation sequence; The embedded representation sequence is encoded with contextual semantic embedding through a 12-layer Transformer encoder, and the output is a shared semantic embedding vector. The shared semantic embedding vectors corresponding to all knowledge units in the structured knowledge feature set are combined to form a shared semantic embedding sequence; S23. The domain-specific tower performs domain modeling processing on the shared semantic embedding sequence, specifically including: For each knowledge unit to which the knowledge domain belongs, an independent Adapter module is introduced under the corresponding domain. The Adapter module includes a dimension reduction layer, a non-linear activation layer and a dimension increase layer connected in sequence. The shared semantic embedding vectors are transformed using the corresponding knowledge domain's Adapter module to generate domain-specific embedding vectors: ; in, Indicates the first Domain-specific embedding vectors of knowledge units Indicates the first Shared semantic embedding vectors of knowledge units For the first The knowledge domain label to which each knowledge unit belongs. Represents the parameters of the dimensionality reduction layer. Indicates the number of dimensional levels. The non-linear activation function combines all domain-specific embedding vectors to form a domain-specific embedding sequence. S24. The shared semantic mapping layer performs semantic space alignment processing on the shared semantic embedding sequence and the domain-specific embedding sequence through linear mapping, and inputs them into the gated residual fusion module; S25. The gated residual fusion module adopts a dual-gate structure, introducing independent fusion weights for the shared semantic embedding vector and the domain-specific embedding vector respectively, and outputting a cross-domain semantic fusion vector. The fusion weights are dynamically calculated through the Softmax function, and all cross-domain semantic fusion vectors are combined to form a cross-domain semantic fusion sequence.

[0024] The dual-tower cross-domain embedding network constructed in this invention integrates two types of structures: a shared semantic tower and a domain-specific tower, achieving deep semantic modeling of cross-domain AI knowledge. The shared semantic tower uses the pre-trained language model SciBERT as the encoder, which can fully mine the common semantic features between different knowledge units, enhancing the consistency and context awareness of semantic representation. The domain-specific tower, by introducing an independent Adapter module for each knowledge domain, further injects domain-specific features on top of the shared semantics, effectively improving the discriminativeness and professionalism of the representation. Simultaneously, the shared semantic mapping layer and the gated residual fusion module support the alignment and dynamic fusion of cross-tower features in the semantic space, avoiding semantic conflicts and redundant interference. The proposed dual-tower cross-domain embedding network improves the accuracy, transferability, and semantic integrity of cross-domain knowledge representation.

[0025] In this embodiment, step S3 specifically includes: S31. The improved Bi-GRU network includes a behavior decomposition module, a behavior encoding module, a domain control module, an attention pooling module, and a preference fusion module; S32. The behavior decomposition module decomposes the user-knowledge interaction matrix according to the order of behavior occurrence and behavior type, generating click behavior sequence, rating behavior sequence and reference behavior sequence; S33. The behavior encoding module includes three sets of parallel Bi-GRU sub-networks. The click behavior sequence, rating behavior sequence and reference behavior sequence are concatenated with the corresponding behavior type embedding vectors and input into the corresponding Bi-GRU sub-networks respectively to perform bidirectional state encoding to obtain the click behavior hidden state, rating behavior hidden state and reference behavior hidden state at each time step. S34. The update gate structure in the Bi-GRU sub-network introduces behavior type embedding vectors for gating adjustment: ; in, This indicates updating the input vector of the gate. This indicates updating the output vector of the gate. Indicates time step Behavior type embedding vector, This indicates the hidden state of the previous time step. This is the weight matrix. This represents the sigmoid activation function; S35. The domain control module introduces domain label embedding vectors to semantically correct the hidden states of click behavior, rating behavior, and reference behavior at each time step, and obtains the corrected hidden states of behavior, rating behavior, and reference behavior. The domain label embedding vector is generated by selecting the knowledge domain label to which the user behavior sequence belongs from a preset domain label set and transforming it through an embedding mapping function; S36. The attention pooling module performs attention weighted aggregation operations on the corrected click behavior hidden state sequence, the corrected rating behavior hidden state sequence, and the corrected reference behavior hidden state sequence to obtain the click behavior preference vector, the rating behavior preference vector, and the reference behavior preference vector, respectively. S37. The preference fusion module performs a weighted fusion operation on the click behavior preference vector, rating behavior preference vector and reference behavior preference vector to generate a user behavior preference vector.

[0026] In this invention, an improved Bi-GRU network is designed to perform refined modeling of user click, rating, and citation behavior sequences across different knowledge domains. The behavior encoding module introduces a behavior type-aware update gate structure, enabling the differentiation of the impact of different behavior types on hidden state updates and enhancing the accuracy of temporal modeling. The domain regulation module achieves semantic calibration based on domain label embedding, effectively introducing domain knowledge constraints. The attention pooling module focuses on key behavior nodes, improving the interpretability of preference representations. The preference fusion module integrates features along the behavior dimension, ultimately generating a high-quality, generalizable user behavior preference vector. This improved Bi-GRU network strengthens the expressive power of user interests in cross-domain environments, significantly enhancing the multi-dimensional behavior modeling and semantic awareness capabilities of this invention.

[0027] In this embodiment, step S4 specifically includes: S41, The cold start user is a user whose user behavior data does not meet the set behavior quantity threshold; S42. Construct a knowledge semantic graph based on a structured knowledge feature set. The nodes of the knowledge semantic graph are knowledge units, and the edges are similarity scores between nodes. The semantic similarity scores are obtained by calculating the cosine similarity between the cross-domain semantic fusion vectors of the knowledge units. S43. Select the knowledge units that have been interacted with by the cold start user as the initial seed nodes, set the initial interest vector, update the interest feature vector iteratively through semantic diffusion on the knowledge semantic graph, and select the K knowledge units with the highest similarity ranking to form the interest neighborhood set. S44. Calculate the statistical distribution of the domain labels of each knowledge unit in the interest neighborhood set, construct the neighborhood label weight vector, and concatenate and fuse the interest feature vector with the neighborhood label weight vector to generate the interest completion vector. S45. The interest completion vector is used as an auxiliary input to the improved Bi-GRU network, and is concatenated with the click behavior sequence, rating behavior sequence and reference behavior sequence respectively, and then input into the corresponding Bi-GRU sub-network to participate in bidirectional state encoding.

[0028] This invention constructs a knowledge semantic graph and, based on the known interactive behaviors of cold-start users, effectively simulates the interest propagation path of users in the semantic space, thereby uncovering potential interest hotspots. Simultaneously, it utilizes domain label statistics to fuse neighborhood label information, further enhancing the semantic completeness and domain adaptability of interest completion. The generated interest completion vectors not only serve as auxiliary input in the improved Bi-GRU network modeling but also significantly improve the personalized understanding and recommendation accuracy for users with insufficient behavior, thus effectively alleviating the modeling blind spots and insufficient cold user coverage issues present in cold-start scenarios.

[0029] In this embodiment, step S5 specifically includes: S51. The dual-tower recall structure includes a user recall sub-tower and a knowledge recall sub-tower. The user recall sub-tower uses a two-layer perceptron network to perform semantic compression mapping on the user behavior preference vector to generate a user semantic vector. The knowledge recall sub-tower uses a two-layer perceptron network that is symmetrical to the user recall sub-tower structure to perform semantic compression mapping on the cross-domain semantic fusion vector in the cross-domain semantic fusion sequence to generate a corresponding knowledge semantic vector. All knowledge semantic vectors are combined to form a knowledge semantic set. S52. Perform vector similarity calculation between the user semantic vector and each knowledge semantic vector in the knowledge semantic set using cosine similarity to obtain a similarity score; S53. Sort the knowledge semantic set according to similarity scores, select the top-M knowledge semantic vectors, and extract the corresponding knowledge units according to the index positions of the top-M knowledge semantic vectors in the structured knowledge feature set to form a preliminary recall recommendation list. S54. The Top-M knowledge semantic vectors are concatenated with the user semantic vector and the corresponding context feature vector at the feature level to generate M candidate interaction feature vectors. The context feature vectors are formed by concatenating the user's current usage scenario label, terminal type label, task urgency label, time period label and geographical location label after being converted into dense vectors by an embedding mapping function. S55. Using the XGBoost model, input the M candidate interaction feature vectors into the XGBoost model to perform non-linear scoring, obtain the ranking score of each candidate interaction feature vector, and re-evaluate the knowledge units in the preliminary recall recommendation list based on the ranking score to generate a cross-domain recommendation list.

[0030] This invention uses a dual-tower structure to perform semantic compression mapping between user behavior preference vectors and cross-domain semantic fusion vectors, enabling efficient matching between users and knowledge units in a unified semantic space. After initial recall of Top-M candidate knowledge units based on cosine similarity, contextual features are introduced and user semantic vectors are concatenated. The XGBoost model is then used for non-linear scoring and ranking, which improves the ranking accuracy, contextual adaptability, and user acceptance of knowledge recommendation results while ensuring the coverage of recommendations.

[0031] In this embodiment, step S6 specifically includes: S61. The context rule base includes a set number of context discrimination rules, and each context discrimination rule consists of a context condition item and an adaptation constraint item; The contextual conditions include the user's current usage scenario tag, terminal type tag, task urgency tag, time period tag, and geographical location tag. The adaptation constraints include the domain tag of the knowledge unit, presentation format, recommended content length, and terminal display adaptability. S62. Obtain the context feature vector corresponding to each knowledge unit in the cross-domain recommendation list, and perform Boolean discrimination operation on the context feature vector with each context discrimination rule in the context rule base in turn: if the context feature vector satisfies all context discrimination rules, then determine that the current knowledge unit matches the user's usage scenario and retain the current knowledge unit; otherwise, remove the current knowledge unit. S63. Compile the retained knowledge units into a knowledge aggregation recommendation list.

[0032] For example, if a user is currently in a "desktop - low urgency task - morning time" scenario, their contextual feature vector includes a terminal type label of "desktop", a task urgency label of "low", and a time period label of "morning". The contextual condition items of a judgment rule in the contextual rule base are set as follows: "terminal type = desktop", "task urgency = low", and "time period = morning". The adaptation constraints are set as follows: the domain label of the knowledge unit is "AI basic theory", the recommended content length is within 1000 words, and the presentation format is "PDF".

[0033] Boolean discrimination was performed on two knowledge units in the cross-domain recommendation list: Knowledge unit A, with the domain label "AI basic theory," a content length of 924 characters, and a "PDF format" presentation, fully met the discrimination rule and was therefore judged as a "match" and retained. Knowledge unit B, with the domain label "visual model optimization," a content length of 1842 characters, and a "Markdown webpage" presentation, failed to meet all contextual conditions and adaptation constraints due to a mismatched domain label, excessive content length, and mismatched presentation format, and was therefore judged as a "mismatch" and removed. Boolean discrimination effectively filters out knowledge units that fit the current user scenario, improving the relevance and practicality of the recommendation results.

[0034] In this invention, by setting a context rule base and combining contextual features such as the user's current usage scenario tags, terminal type, task urgency, time period, and geographical location, the knowledge units in the cross-domain recommendation list are judged and filtered to ensure that the final recommendation results are highly consistent with the user's environment and task requirements. This method significantly improves the relevance and practicality of knowledge aggregation recommendation results, avoids interference from incompatible content, and enhances the adaptability and user satisfaction of this invention under multiple scenarios and multiple terminal conditions.

[0035] In this embodiment, step S7 specifically includes: S71. Collect user behavior feedback data during actual use, and supplement and update the user-knowledge interaction matrix in chronological order; S72. Based on the updated user-knowledge interaction matrix, reconstruct the click behavior sequence, rating behavior sequence, and reference behavior sequence, and input them into the improved Bi-GRU network to perform user preference update modeling and generate an incrementally optimized user behavior preference vector. S73. Based on the newly added knowledge units involved in the user feedback data, expand the structured knowledge feature set and input it into the dual-tower cross-domain embedding network to perform shared semantic embedding and domain-specific embedding processing, and update the cross-domain semantic fusion vector. S74. Calculate the predicted interaction rating based on the updated user behavior preference vector and cross-domain semantic fusion vector, minimize the mean square error between the predicted interaction rating and the actual rating in user feedback behavior, jointly optimize the dual-tower cross-domain embedding network and the improved Bi-GRU network, and regenerate the knowledge aggregation recommendation list.

[0036] Example 1: To verify the feasibility of this invention in practice, it was applied to the knowledge service module of an artificial intelligence open platform. This platform aggregates AI technology resources from multiple fields, including natural language processing, computer vision, knowledge graphs, and autonomous driving. It contains 380,000 structured model entries, 120,000 API interface documents, over 30 million words of scientific research paper corpus, and usage behavior data from nearly 200,000 active users, covering multiple application scenarios such as scientific research, industry, and education.

[0037] Traditional knowledge recommendation methods often face challenges when performing cross-domain knowledge recommendation tasks. These challenges include irrelevant recommendations, lack of response from new users, and recommendations that are out of context for the user. This is particularly true on this platform, where newly registered users account for over 22% of users, while monthly new knowledge units account for 18%. The cold start problem and the issue of heterogeneous knowledge integration are particularly prominent, severely hindering knowledge aggregation efficiency and user satisfaction.

[0038] After applying the method of this invention, the platform parses structured knowledge graphs, semi-structured API documents, and unstructured scientific research corpora into unified structured knowledge feature sets. A dual-tower cross-domain embedding network is used to uniformly model and semantically align the knowledge feature sets. The SciBERT model is employed to capture the contextual semantic information of knowledge units. By setting independent Adapter modules for each domain to preserve domain characteristics, knowledge semantic fusion under a shared representation space is achieved. In the user modeling stage, the platform uses an improved Bi-GRU network to process the user-knowledge interaction matrix, temporally encoding user clicks, ratings, and citations, and introducing knowledge domain labels in conjunction with a domain regulation mechanism to achieve refined modeling of user behavior. For cold-start user modeling, the platform constructs an interest completion mechanism based on semantic diffusion for users with insufficient behavior, effectively inferring their potential interest vectors. Finally, the recommendation list is semantically matched and optimized for scenario adaptation through a dual-tower recall structure and a context rule base.

[0039] To more comprehensively verify the actual effect of the method of this invention, the following three representative recommendation methods were selected as comparison methods: Traditional collaborative filtering recommendation method was used as the basic comparison method. This method calculates the recommendation list based on the similarity of user behavior, but lacks semantic understanding and cross-domain fusion capabilities, and cannot effectively handle heterogeneous knowledge data and the cold start problem; BERT+MLP recommendation method was used as the semantic modeling comparison method. This method extracts the semantic representation of knowledge units through BERT and uses MLP to match and model user behavior with knowledge content; Graph neural network recommendation method was used as the structural modeling comparison method. This method uses graph neural networks to perform feature propagation and interest modeling on the user-knowledge interaction graph, and can characterize the structured relationships between users, but its capabilities in multimodal knowledge fusion and context adaptation are limited, and it suffers from high inference latency. Specific comparison data are shown in Table 1.

[0040] Table 1. Performance Comparison of the Invention Method and Comparative Methods in Platform Recommendation Systems

[0041] The data above shows that the method of this invention outperforms the comparative methods in all indicators. In terms of recommendation accuracy, the method of this invention achieves 0.726, an improvement of approximately 24.5% compared to traditional collaborative filtering methods, and improvements of 11.9% and 8.0% compared to BERT+MLP and graph neural network recommendation methods, respectively, demonstrating superior knowledge matching capabilities. Regarding the average click-through rate, the method of this invention achieves 0.188, significantly higher than other methods, indicating that the recommended content is more attractive and user interaction is more active. Addressing the cold start problem, the cold start user recall rate of the method of this invention is 0.541, far exceeding the 0.296 of traditional collaborative filtering methods, and improvements of 28.0% and 20.8% compared to BERT+MLP and graph neural network recommendation methods, respectively, fully demonstrating that the interest completion and semantic diffusion of this invention effectively alleviate the cold start bottleneck. In terms of recommended content matching rate, the method of this invention achieves 0.803, higher than the comparative methods, indicating that the method of this invention has a strong advantage in knowledge semantic alignment and recommended content fit.

[0042] While slightly faster than traditional collaborative filtering in terms of recommendation response time, it still significantly outperforms BERT+MLP and graph neural network methods, balancing semantic depth and response efficiency with acceptable real-time performance. In terms of user experience, this method achieved a user satisfaction score of 4.62, the highest among all methods, indicating that the relevance and practicality of the recommended content are widely recognized by users. Regarding recommendation coverage, this method reached 0.742, higher than all compared methods, demonstrating strong domain generalization ability and comprehensive recommendation coverage.

[0043] This invention has good practicality and promotional value in real platform environments, and has significant advantages in practical systems with high knowledge heterogeneity, large cold start ratios, and complex usage scenarios.

[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A cross-domain AI knowledge aggregation method based on collaborative filtering, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous AI knowledge data and user behavior data, and preprocess them to obtain a structured knowledge feature set and a user-knowledge interaction matrix; S2. Construct a dual-tower cross-domain embedding network consisting of a shared semantic tower and a domain-specific tower, perform cross-domain semantic alignment on the structured knowledge feature set, and output a cross-domain semantic fusion sequence. S3. Using an improved Bi-GRU network, user preference modeling and analysis are performed on the user-knowledge interaction matrix to generate user behavior preference vectors; S4. For cold-start users, interest features are completed based on semantic diffusion and neighborhood reasoning to obtain interest completion vectors, which are then used as auxiliary inputs to the improved Bi-GRU network. S5. A dual-tower recall structure is used to perform semantic mapping between cross-domain semantic fusion sequences and user behavior preference vectors, and non-linear scoring and ranking are performed through the XGBoost model to generate a cross-domain recommendation list. S6. Construct a context rule base, perform context adaptability judgment and filtering on the cross-domain recommendation list, and generate a knowledge aggregation recommendation list that matches the user's usage scenario; S7. Based on user feedback, perform incremental learning and dynamically update the dual-tower cross-domain embedded network and the improved Bi-GRU network.

2. The cross-domain AI knowledge aggregation method based on collaborative filtering according to claim 1, characterized in that, Step S1 specifically includes: S11. The multi-source heterogeneous AI knowledge data includes structured knowledge graph data, semi-structured API interface documents, and unstructured scientific research text corpus. S12, The user behavior data includes the user's search records, citation history, access frequency and rating feedback on various knowledge domain platforms; S13, The preprocessing includes: The structured knowledge graph data is subjected to graph parsing, entity deduplication, and triple normalization operations to generate a standardized knowledge graph tensor. Perform field recognition, key-value extraction, and semantic tag mapping operations on the semi-structured API interface document data to generate an interface structure mapping table. Unstructured scientific research text corpora are subjected to word segmentation, named entity recognition, and relation extraction operations to generate text entity semantic feature vectors; Standardized knowledge graph tensors, interface structure mapping tables, and text entity semantic feature vectors are formatted and semantically aligned to form a structured knowledge feature set; Perform user identifier normalization, behavior sequence sorting, and interaction intensity filtering operations on user behavior data to generate a user-knowledge interaction matrix.

3. The cross-domain AI knowledge aggregation method based on collaborative filtering according to claim 1, characterized in that, Step S2 specifically includes: S21. The dual-tower cross-domain embedding network includes a shared semantic tower, a domain-specific tower, a shared semantic mapping layer, and a gated residual fusion module. S22. The shared semantic tower uses a pre-trained SciBERT model as the encoder: The WordPiece tokenizer, built on SCIVOCAB, is used to transform each knowledge unit in the structured knowledge feature set into a token sequence. The identifier [CLS] is inserted at the beginning of the token sequence and the identifier [SEP] is inserted at the end to form the input index sequence. The input index sequence is mapped to word embeddings, position embeddings, and paragraph embeddings respectively, and then combined to form an embedding representation sequence; The embedded representation sequence is encoded with contextual semantic embedding through a 12-layer Transformer encoder, and the output is a shared semantic embedding vector. The shared semantic embedding vectors corresponding to all knowledge units in the structured knowledge feature set are combined to form a shared semantic embedding sequence; S23. The domain-specific tower performs domain modeling processing on the shared semantic embedding sequence, specifically including: For each knowledge unit to which the knowledge domain belongs, an independent Adapter module is introduced under the corresponding domain. The Adapter module includes a dimension reduction layer, a non-linear activation layer and a dimension increase layer connected in sequence. The shared semantic embedding vectors are transformed using the corresponding knowledge domain's Adapter module to generate domain-specific embedding vectors. All domain-specific embedding vectors are then combined to form a domain-specific embedding sequence. S24. The shared semantic mapping layer performs semantic space alignment processing on the shared semantic embedding sequence and the domain-specific embedding sequence through linear mapping, and inputs them into the gated residual fusion module; S25. The gated residual fusion module adopts a dual-gate structure, introducing independent fusion weights for the shared semantic embedding vector and the domain-specific embedding vector respectively, and outputting a cross-domain semantic fusion vector. The fusion weights are dynamically calculated through the Softmax function, and all cross-domain semantic fusion vectors are combined to form a cross-domain semantic fusion sequence.

4. The cross-domain AI knowledge aggregation method based on collaborative filtering according to claim 1, characterized in that, Step S3 specifically includes: S31. The improved Bi-GRU network includes a behavior decomposition module, a behavior encoding module, a domain control module, an attention pooling module, and a preference fusion module; S32. The behavior decomposition module decomposes the user-knowledge interaction matrix according to the order of behavior occurrence and behavior type, generating click behavior sequence, rating behavior sequence and reference behavior sequence; S33. The behavior encoding module includes three sets of parallel Bi-GRU sub-networks. The click behavior sequence, rating behavior sequence and reference behavior sequence are concatenated with the corresponding behavior type embedding vectors and input into the corresponding Bi-GRU sub-networks respectively to perform bidirectional state encoding to obtain the click behavior hidden state, rating behavior hidden state and reference behavior hidden state at each time step. S34. The update gate structure in the Bi-GRU sub-network introduces behavior type embedding vectors for gating adjustment; S35. The domain control module introduces domain label embedding vectors to semantically correct the hidden states of click behavior, rating behavior, and reference behavior at each time step, and obtains the corrected hidden states of behavior, rating behavior, and reference behavior. The domain label embedding vector is generated by selecting the knowledge domain label to which the user behavior sequence belongs from a preset domain label set and transforming it through an embedding mapping function; S36. The attention pooling module performs attention weighted aggregation operations on the corrected click behavior hidden state sequence, the corrected rating behavior hidden state sequence, and the corrected reference behavior hidden state sequence to obtain the click behavior preference vector, the rating behavior preference vector, and the reference behavior preference vector, respectively. S37. The preference fusion module performs a weighted fusion operation on the click behavior preference vector, rating behavior preference vector and reference behavior preference vector to generate a user behavior preference vector.

5. The cross-domain AI knowledge aggregation method based on collaborative filtering according to claim 1, characterized in that, Step S4 specifically includes: S41, The cold start user is a user whose user behavior data does not meet the set behavior quantity threshold; S42. Construct a knowledge semantic graph based on a structured knowledge feature set. The nodes of the knowledge semantic graph are knowledge units, and the edges are similarity scores between nodes. The semantic similarity scores are obtained by calculating the cosine similarity between the cross-domain semantic fusion vectors of the knowledge units. S43. Select the knowledge units that have been interacted with by the cold start user as the initial seed nodes, set the initial interest vector, update the interest feature vector iteratively through semantic diffusion on the knowledge semantic graph, and select the K knowledge units with the highest similarity ranking to form the interest neighborhood set. S44. Calculate the statistical distribution of the domain labels of each knowledge unit in the interest neighborhood set, construct the neighborhood label weight vector, and concatenate and fuse the interest feature vector with the neighborhood label weight vector to generate the interest completion vector. S45. The interest completion vector is used as an auxiliary input to the improved Bi-GRU network, and is concatenated with the click behavior sequence, rating behavior sequence and reference behavior sequence respectively, and then input into the corresponding Bi-GRU sub-network to participate in bidirectional state encoding.

6. The cross-domain AI knowledge aggregation method based on collaborative filtering according to claim 1, characterized in that, Step S5 specifically includes: S51. The dual-tower recall structure includes a user recall sub-tower and a knowledge recall sub-tower. The user recall sub-tower uses a two-layer perceptron network to perform semantic compression mapping on the user behavior preference vector to generate a user semantic vector. The knowledge recall sub-tower uses a two-layer perceptron network that is symmetrical to the user recall sub-tower structure to perform semantic compression mapping on the cross-domain semantic fusion vector in the cross-domain semantic fusion sequence to generate a corresponding knowledge semantic vector. All knowledge semantic vectors are combined to form a knowledge semantic set. S52. Perform vector similarity calculation between the user semantic vector and each knowledge semantic vector in the knowledge semantic set using cosine similarity to obtain a similarity score; S53. Sort the knowledge semantic set according to similarity scores, select the top-M knowledge semantic vectors, and extract the corresponding knowledge units according to the index positions of the top-M knowledge semantic vectors in the structured knowledge feature set to form a preliminary recall recommendation list. S54. Concatenate the Top-M knowledge semantic vectors with the user semantic vector and the corresponding context feature vector at the feature level to generate M candidate interaction feature vectors. S55. Using the XGBoost model, input the M candidate interaction feature vectors into the XGBoost model to perform non-linear scoring, obtain the ranking score of each candidate interaction feature vector, and re-task the knowledge units in the preliminary recall recommendation list based on the ranking score to generate a cross-domain recommendation list.

7. The cross-domain AI knowledge aggregation method based on collaborative filtering according to claim 1, characterized in that, Step S6 specifically includes: S61. The context rule base includes a set number of context discrimination rules, and each context discrimination rule consists of a context condition item and an adaptation constraint item; The contextual conditions include the user's current usage scenario tag, terminal type tag, task urgency tag, time period tag, and geographical location tag. The adaptation constraints include the domain tag of the knowledge unit, presentation format, recommended content length, and terminal display adaptability. S62. Obtain the context feature vector corresponding to each knowledge unit in the cross-domain recommendation list, and perform Boolean discrimination operation on the context feature vector with each context discrimination rule in the context rule base in turn: if the context feature vector satisfies all context discrimination rules, then determine that the current knowledge unit matches the user's usage scenario and retain the current knowledge unit; otherwise, remove the current knowledge unit. S63. Compile the retained knowledge units into a knowledge aggregation recommendation list.

8. The cross-domain AI knowledge aggregation method based on collaborative filtering according to claim 1, characterized in that, Step S7 specifically includes: S71. Collect user behavior feedback data during actual use, and supplement and update the user-knowledge interaction matrix in chronological order; S72. Based on the updated user-knowledge interaction matrix, reconstruct the click behavior sequence, rating behavior sequence, and reference behavior sequence, and input them into the improved Bi-GRU network to perform user preference update modeling and generate an incrementally optimized user behavior preference vector. S73. Based on the newly added knowledge units involved in the user feedback data, expand the structured knowledge feature set and input it into the dual-tower cross-domain embedding network to perform shared semantic embedding and domain-specific embedding processing, and update the cross-domain semantic fusion vector. S74. Calculate the predicted interaction rating based on the updated user behavior preference vector and cross-domain semantic fusion vector, minimize the mean square error between the predicted interaction rating and the actual rating in user feedback behavior, jointly optimize the dual-tower cross-domain embedding network and the improved Bi-GRU network, and regenerate the knowledge aggregation recommendation list.

Images