Personalized intelligent recommendation method and system based on large language model

Abstract

The embodiment of the application discloses a kind of personalized intelligent recommendation method and system based on large language model, belong to the technical field of electric digital data processing, including: construct heterogeneous bipartite graph, the semantic feature vector of original text is extracted by first large language model, generate weighted adjacency matrix representation user-text logical relationship, and output explanation logical rule set by second language model, finally utilize multimodal splicing scheme to calculate the joint probability distribution of user-content-text and implement corresponding intelligent recommendation.The embodiment of the application solves the problem that user-content-text representation is fragmented in the prior art by multiple large language models, the generated explanation logical rule set not only fits user real behavior trajectory, but also has explainability, using the joint probability distribution of user-content-text, significantly improves the accuracy and transparency of personalized recommendation.

Description

Technical Field

[0001] This application relates to a personalized intelligent recommendation method and system based on a large language model, belonging to the technical field of electronic digital data processing. Background Technology

[0002] With the exponential growth of internet information, users often struggle to quickly find information that truly matches their interests and needs when faced with massive amounts of content. Personalized recommendation systems have emerged to address this, analyzing users' historical behavioral data to predict their potential preferences and thus achieve precise recommendations. Existing recommendation methods can be broadly categorized into three types: collaborative filtering-based methods, content-based methods, and hybrid recommendation methods. While these three types of existing intelligent recommendation methods can capture common user traits to some extent, they suffer from limitations such as imprecise matching, susceptibility to the quality of feature engineering, and shortcomings in heterogeneous data fusion, dynamic semantic understanding, and interpretability.

[0003] In recent years, Large Language Models (LLMs) have demonstrated outstanding performance in natural language understanding and generation tasks, and their powerful semantic modeling capabilities have provided a new technical path for personalized recommendations. Existing research has attempted to introduce LLMs into recommendation systems, such as guiding models to generate recommendation reasons through Prompt Engineering, or using LLMs to encode text content to enrich item representations. However, most existing large language models remain at a shallow semantic matching level, failing to fully utilize the potential of LLMs in complex relationship modeling and logical reasoning. They also generally lack an understanding of the heterogeneity of user behavior, making it difficult to effectively integrate heterogeneous information such as users, content, interactions, and text. This results in a trade-off between accuracy, diversity, and interpretability in recommendation results, which no longer meets people's requirements and urgently needs improvement. Summary of the Invention

[0004] The purpose of this application is to provide a personalized intelligent recommendation method and system based on a large language model. By constructing a unified framework to represent various heterogeneous data and realize the fusion of heterogeneous data, implicit behavioral logic is transformed into explicit natural language rules through an interpretable mechanism, realizing transparent presentation of "user-text" and accurate intelligent recommendation, thus solving the shortcomings of the existing technology.

[0005] The embodiments of this application provide the following solutions:

[0006] According to one aspect of the embodiments of this application, a personalized intelligent recommendation method based on a large language model is provided for implementing a personalized intelligent recommendation system based on a large language model, comprising: acquiring user behavior data; constructing a heterogeneous bipartite graph based on the user behavior data, the heterogeneous bipartite graph including user nodes, content nodes, interaction relationships, and original text; inputting the original text into a first large language model and extracting semantic feature vectors, binding the semantic feature vectors with content nodes to form content enhancement nodes; traversing the interaction relationships with the content enhancement nodes as constraints, obtaining the intersection between content nodes and interaction relationships in the interaction relationships, constructing weighted edges between user nodes and content enhancement nodes based on the intersection, generating a weighted adjacency matrix through the weights of the weighted edges, the weighted adjacency matrix being used to represent the "user-text" logical relationship; obtaining corresponding confidence association information based on user behavior sequences and combined with the weighted adjacency matrix and generating natural language description prompt information, inputting it into a second language model and outputting an explanation logic rule set;

[0007] The set of interpretation logic rules is multimodal concatenated, and the joint probability distribution of user-content-text is calculated based on the multimodal concatenation result. The corresponding intelligent recommendation is then performed based on the probability value of the joint probability distribution.

[0008] According to at least one specific embodiment of the present application, the user node is used to represent the user entity and its attributes, including user ID and user basic profile; the content node is used to represent the interaction object; the interaction relationship is used to represent a single interaction record between the user and the content; and the original text is specifically unstructured text.

[0009] According to at least one specific embodiment of the present application, before inputting the original text into the first large language model and extracting semantic feature vectors, the original text is subjected to structured processing. The structured processing includes deduplication, word segmentation, stop word filtering, part-of-speech tagging, and entity extraction. Specifically: the original text is segmented and filtered for stop words, and then part-of-speech tagging and entity extraction are performed to obtain a preprocessing sequence. The preliminary entity table in the preprocessing sequence is vectorized and encoded to generate at least three types of vectors: semantic vector, pinyin vector, and part-of-speech vector. The semantic vector is fused with the pinyin vector and the part-of-speech vector to obtain a fused pinyin vector and a fused part-of-speech vector. Based on cosine similarity, the fused pinyin vector and the fused part-of-speech vector are spatially aligned under a contrastive learning framework. Head entities are extracted using the semantic vectors. The head entities are added to the fused pinyin vector, the fused part-of-speech vector, and the semantic vector to form a joint representation value. The joint representation value is fed into a tail entity and relation extraction network to output triple results.

[0010] According to at least one specific implementation of the embodiments of this application, in the process of spatially aligning the fused pinyin vector and the fused part-of-speech vector based on cosine similarity within a contrastive learning framework, the process further includes: synchronously inputting the pinyin vector and the part-of-speech vector into the contrastive learning framework; firstly, calling the cosine similarity function within the framework to calculate the distance metric between the two vectors; and then performing backpropagation based on the gradient signal of the progressive contrastive loss function to drive the spatial alignment module to sequentially apply linear mapping and normalization operations to the pinyin vector and the part-of-speech vector until the two vectors are aligned in a unified metric space.

[0011] According to at least one specific implementation of the embodiments of this application, in the process of binding the semantic feature vector with the content node to form a content enhancement node, the content enhancement node specifically involves: packaging the semantic feature vector, pinyin feature vector and part-of-speech feature vector obtained after the original text has been processed by the first large language model, and enhancing the original node with additional semantic information based on the extracted entity tags to form the content enhancement node.

[0012] According to at least one specific embodiment of the present application, in the process of generating and constructing a weighted adjacency matrix, a user behavior sequence is obtained and a confidence algorithm is run. The confidence algorithm includes a behavior-weight joint tensor, and the parameters in the behavior-weight joint tensor include at least a set of behavior types. The confidence is calculated based on the behavior-weight joint tensor, and the confidence calculation result is compared with a preset threshold. Joint tensor entries that retain preset conditions are retained, and a corresponding high-confidence association set is generated. The high-confidence association set is subjected to natural language transformation according to a preset natural language template to generate a description set, and the description set is input into a second language model. The second language model outputs a corresponding set of explanation logic rules.

[0013] According to at least one specific embodiment of the present application, the set of interpretation logic rules is interpretable symbolic logic, and the joint probability distribution of user-content-text is calculated based on the interpretable symbolic logic through a multimodal neural network.

[0014] According to another aspect of the embodiments of this application, a personalized intelligent recommendation system based on a large language model is provided, used to implement the personalized intelligent recommendation method based on the large language model, comprising: a heterogeneous bipartite graph generation module, which acquires user behavior data and constructs a heterogeneous bipartite graph based on the user behavior data, the heterogeneous bipartite graph including user nodes, content nodes, interaction relationships, and original text; a content enhancement node generation module, which inputs the original text into a first large language model and extracts semantic feature vectors, and binds the semantic feature vectors with content nodes to form content enhancement nodes; and a weighted adjacency matrix generation module, which traverses the interaction relationships with the content enhancement nodes as constraints to obtain content in the interaction relationships. The system identifies the intersection between nodes and interaction relationships. Based on this intersection, weighted edges are constructed between user nodes and content enhancement nodes. A weighted adjacency matrix is ​​generated using the weights of these edges, representing the "user-text" logical relationship. An explanation logic rule set output module, based on user behavior sequences and the weighted adjacency matrix, obtains corresponding confidence-related information and generates natural language description prompts. This information is input into a second language model and outputs an explanation logic rule set. A multimodal concatenation module performs multimodal concatenation on the explanation logic rule set, calculates the joint probability distribution of user-content-text based on the concatenation results, and performs corresponding intelligent recommendations based on the probability values ​​of the joint probability distribution.

[0015] According to another aspect of the embodiments of this application, an electronic device is provided, including: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method.

[0016] According to other aspects of embodiments of this application, a computer-readable storage medium is provided that stores a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the steps of the method.

[0017] The embodiments of this application have the following advantages compared with the prior art:

[0018] This application embodiment achieves semantic enhancement by constructing a heterogeneous bipartite graph, and then calculates the joint probability using an interpretation logic rule set, unifying and integrating four types of heterogeneous data: user, content, interaction, and text. Based on the first major language model, text is converted into semantic features, and these features are bound to content nodes to form content enhancement nodes. Then, using the content enhancement nodes as constraints, the interaction relationships are traversed to calculate the high-order weighted edges between user nodes and content enhancement nodes, constructing a "user-text" weighted adjacency matrix. The user behavior sequence and the weighted adjacency matrix are then fed into the second language model to generate an interpretation logic rule set (an interpretable natural language logic rule set). Finally, the interpretation logic rule set is multimodally concatenated to calculate the joint probability distribution of user-content-text, and intelligent recommendation is completed accordingly.

[0019] This application's embodiments utilize content-enhanced nodes formed by the first language model to solve the problem of fragmented user-content-text representation in existing technologies. By traversing interaction relationships using content-enhanced nodes, semantic features and user interaction frequency jointly determine weights. While retaining sufficient information in the original text, it provides a data foundation for subsequent probabilistic statistical processing. User behavior sequences and weighted adjacency matrices are input into the second language model, generating an interpretation logic rule set that not only matches the user's actual behavior trajectory but also possesses interpretability, avoiding the black-box processing problem in existing technologies. Finally, the natural language rules are multimodally concatenated and directly used to estimate the joint probability distribution of user-content-text, significantly improving the accuracy and transparency of personalized recommendations. Attached Figure Description

[0020] To more clearly illustrate the specific implementation methods of the embodiments of this application or the technical solutions in the prior art, the drawings used in the description of the specific implementation methods or the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a personalized intelligent recommendation method based on a large language model, as described in an embodiment of this application.

[0022] Figure 2 This is a flowchart of the optimization technical solutions provided in steps S11 to S13.

[0023] Figure 3 This is a flowchart of the optimization technical solution provided in steps S121 to S122.

[0024] Figure 4 This is a flowchart of the optimization technical solutions provided in steps S31 to S33.

[0025] Figure 5 This is an architecture diagram of a personalized intelligent recommendation system based on a large language model, according to an embodiment of this application.

[0026] Figure 6 This is a schematic diagram of the electronic device. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the embodiments of this application.

[0028] Definitions:

[0029] Heterogeneous bipartite graph: Define a bipartite graph G=(U∪C,E), where U and C are two sets of nodes, and E is the set of edges. If U and C are further subdivided into nodes with different attributes or modalities, it is called "heterogeneous". In the embodiments of this application, U is the set of user nodes UserSet, C is the set of content nodes ContentSet, and the edges E are composed of the set of interaction relationships InteractionSet. At the same time, each content node is also attached with the original text set TextSet. Therefore, the heterogeneous bipartite graph of this application embodiment simultaneously carries four types of objects. The heterogeneous bipartite graph can map four heterogeneous objects—users, content, interactions, and text—to the same graph, so that any subsequent algorithm can utilize all the information at once, avoiding the information loss caused by splicing multi-source data.

[0030] Large Language Models (LLMs): Natural Language Processing models with a parameter scale of ≥ one billion, based on the Transformer architecture, and pre-trained on a large scale without supervision, possessing context understanding and generation capabilities. In this embodiment, the first large language model can be an M3E-based large language model, specifically designed for extracting semantic feature vectors. The second language model can be a general-purpose large model (such as GPT-4 / ChatGLM), generating interpretable rules based on a weighted adjacency matrix and behavioral sequences. The first large language model elevates the text to a semantic space, while the second large language model translates the graph structure and temporal behavior into natural language rules, achieving a dual-pathway fusion of semantics and structure.

[0031] Content-enhanced nodes: High-dimensional representation nodes formed by superimposing semantic features on the original content nodes.

[0032] In this embodiment, a portion of the original text in the original text dataset TextSet is encoded using a first language model to obtain a semantic vector v_text, which is then bound to the corresponding content node c_i in the content object dataset ContentSet to form a new node. This new node retains both discrete symbolic attributes and carries continuous semantic information, allowing graph convolution or attention mechanisms to operate directly in the semantic space, significantly improving representation accuracy.

[0033] Weighted adjacency matrix: Define a square matrix W, where the elements w_ij in the matrix W are used to quantify the strength of the relationship between nodes i and j.

[0034] In this embodiment, the matrix rows of the square matrix W correspond to UserSet, the matrix of the square matrix W corresponds to the content enhancement node set (EnhancedContentSet), and the weight w(u,c)=sim(V_user,V_text). sim(V_user,V_text) is the similarity measurement function between the user's historical behavior vector V_user and the content text semantic vector V_text, that is, the similarity between the user's historical behavior vector and the content semantic vector.

[0035] The weighted adjacency matrix quantifies both "statistical frequency" and "semantic matching degree" in a unified way, providing interpretable weights for subsequent calculation of the joint probability distribution.

[0036] Multimodal concatenation: The operation of mapping information from different modalities (text, vector, graph, rule) into the same representation space.

[0037] In this embodiment, the set of interpretation logic rules (text modality) output by the second language model is vector-concatenated with the weighted adjacency matrix (graph modality) and the user behavior sequence (temporal modality) to form a unified multimodal tensor. Multimodal concatenation can break down modal barriers, utilize the joint probability distribution model, and simultaneously leverage the prior knowledge of symbolic rules and the statistical signals of numerical features to improve generalization and interpretability.

[0038] like Figure 1 The personalized intelligent recommendation method based on a large language model, shown below, is used to implement a personalized intelligent recommendation system based on a large language model, including:

[0039] Step S1: Obtain user behavior data and construct a heterogeneous bipartite graph based on the user behavior data. The heterogeneous bipartite graph includes user nodes, content nodes, interaction relationships, and original text. In this step, user behavior data refers to the recordable and quantifiable operational traces generated by users during their use of various platforms, such as: user identification (e.g., user ID, device ID, or login account), content identification (e.g., content ID, article title, video VID, product SKU), interaction type (click, browse, favorite, like, comment, share, purchase, dwell time), or contextual time sequence (e.g., timestamp of interaction behavior, session order, geographical location, terminal type). The role of contextual time sequence is to provide a timeline basis for constructing the heterogeneous bipartite graph. User behavior data can be saved in JSON format, for example: {user_id: "U123", content_id: "C456", action: "like", timestamp: "2025-05-11T14:32:00Z", dwell_time: 23.5s}.

[0040] Step S2: Input the original text into the first language model and extract the semantic feature vector. Bind the semantic feature vector to the content node to form the content enhancement node.

[0041] Step S3: Traverse the interaction relationship with the content enhancement node as the constraint condition, obtain the intersection between the content node and the interaction relationship, construct the weighted edge between the user node and the content enhancement node based on the intersection, and generate the weighted adjacency matrix through the weight of the weighted edge. The weighted adjacency matrix is ​​used to represent the "user-text" logical relationship.

[0042] Steps S2 and S3 involve the concept of content enhancement nodes, which include weighted edges and a weight calculation formula. This formula provides the computational basis for subsequent weighted adjacency matrices and multimodal concatenation. In steps S2 and S3, the weighted edge can be defined using the concept of a set: e(u,c) = (u,c,w), where u∈UserSet (user node), c∈EnhancedContentSet (content enhancement node), and w=w(u,c) is the real-valued weight, quantifying the association strength between user u and content c.

[0043] The weight calculation formula is: w(u,c)=sim(V_user,V_text)×f(interaction). Based on cosine or dot product similarity, it measures the semantic matching degree between the user's historical interest vector V_user and the content semantic vector V_text. f(interaction) is a monotonic function of the interaction relationship (such as the number of interactions, time decay, and interaction type weight) to retain statistical information. Finally, the historical interest vector, content semantic vector, and monotonic function are multiplied and incorporated into the weight calculation formula.

[0044] Step S4: Based on the user behavior sequence and combined with the weighted adjacency matrix, obtain the corresponding confidence association information and generate natural language description prompts. Input these prompts into the second language model and output an explanation logic rule set. For example, the second language model could be ChatGLM-Pro (with hundreds of billions of parameters, capable of generating sentences based on Chinese semantic understanding and rules).

[0045] Step S5: Perform multimodal concatenation on the interpretation logic rule set, calculate the joint probability distribution of user-content-text based on the multimodal concatenation result, and make corresponding intelligent recommendations based on the probability values ​​of the joint probability distribution.

[0046] The technical solution provided in steps S1 to S5 uses a heterogeneous bipartite graph to unify heterogeneous data. It utilizes a first language model to upscale the text into semantic vectors, binds content nodes to form content-enhanced nodes, and calculates weighted edges between users and content using these enhanced nodes as anchors. This generates a weighted adjacency matrix, which is then input into a second language model along with user behavior sequences. Finally, it outputs an interpretable logic rule set (interpretable natural language rules) and performs multimodal concatenation to estimate the joint probability distribution of the corresponding "user-content-text," achieving accurate intelligent recommendation. Because it achieves unified processing of heterogeneous data, semantic enhancement is implemented on top of the structured language in steps S1 and S2, providing a structurally consistent input signal with rich semantic information for subsequent calculations. Then, semantic-statistical analysis is implemented in steps S2 and S3, using weighting formulas to semantically complete user frequency data and provide a data foundation for subsequent probability statistical calculations. Finally, an interpretable interpretable logic rule set is generated through the encoding rules of the weighted adjacency matrix and multimodal tensors, and the joint probability distribution is calculated. Accurate personalized recommendation is achieved through the probability model. In summary, steps S1 to S5 deeply integrate the semantic splitting capability of the large language model with the behavioral analysis of traditional recommendation systems, breaking through the semantic bottleneck of collaborative filtering and avoiding the data sparsity problem of pure semantic recommendation. In terms of technical effect, it achieves the unity of accurate recommendation and semantic interpretability.

[0047] For example, one patent application scenario of the technical solution in steps S1 to S5 is as follows: During the content distribution process of a short video platform, the average duration of user swiping and staying is less than 2 seconds. After applying the technical solution in steps S1 to S5, the four types of objects "user UID-video VID-interaction behavior (like / complete playback / comment)-video title / subtitle / background music lyrics" are included in the heterogeneous bipartite graph through a heterogeneous bipartite graph. The first language model is called to encode the subtitles and background music lyrics into semantic vectors and bind them to the VID node to form content enhancement nodes. Based on the "content enhancement nodes" as constraints, the interactions such as likes, complete playback, and comments are traversed to generate a weighted adjacency matrix. The weights reflect both the similarity between "the user's historical preferences and video semantics" and the intensity of interaction between the user and the platform.

[0048] Example of a user behavior sequence: A user first watches a tech review video and then likes a drone aerial video. This user behavior sequence (watching the video and then liking it) is fed into a second language model along with a weighted adjacency matrix. The model outputs rules in real time, determining that the user prefers tech reviews from an aerial perspective. After multimodal concatenation, the model estimates the joint probability distribution of the user, video, and text. Based on the value of the joint probability distribution, the model decides to recommend high frame rate drone race clips to the user, thereby increasing the user's average time spent on the short video platform.

[0049] The main steps from S1 to S5 can be summarized as follows: Heterogeneous bipartite graph construction: Users, content, interactions, and text are uniformly modeled as a heterogeneous bipartite graph. Graph neural networks (GNNs) are applied to the recommendation system. Then, semantic feature vectors (such as pinyin, parts of speech, and entity tags) of the text are extracted through a large language model (LLM) and bound to content nodes. This enriches node representation, improves recommendation accuracy, and enhances semantic features. Then, weights are calculated by combining user behavior frequency and semantic similarity (such as cosine similarity) to generate a weighted adjacency matrix. Semantics are filtered and fused. Finally, natural language rules are generated through a second language model, and a joint probability distribution is calculated by combining multimodal data. This balances the two factors of recommendation accuracy and interpretability, achieving the goal of precise push notification.

[0050] Preferably, in step S1, the heterogeneous bipartite graph includes user nodes, content nodes, interaction relationships, and raw text. User nodes represent user entities and their attributes, including user ID and basic user profile. Content nodes represent interactive objects (such as articles, videos, and products). Interaction relationships represent single interaction records between users and content (directed / undirected edges between users and content, such as clicks, favorites, and comments). Each interaction record should include a timestamp and weight value. The raw text is specifically unstructured text, such as titles, body text, user comments, etc. The heterogeneous bipartite graph unifies the modeling of user behavior data (UID-VID interaction), user attributes (UID profile), content attributes (VID metadata), and raw text (titles / comments).

[0051] User node: Indexed by UID, it is bound to the user's static profile (such as age and gender).

[0052] Content nodes: Indexed by VID, associated with content structure attributes (such as category and tag).

[0053] Interaction relationships: Directed / undirected edges between users and content, with timestamps and initial weights (e.g., click = 0.5, favorite = 1.0).

[0054] Raw text: Unstructured data (such as the video title "AI Technology Overview") is mounted to the corresponding content node.

[0055] The heterogeneous bipartite graph after unified modeling solves the data silo problem, breaks the segmented storage state of user behavior, content attributes, and text data, realizes a unified topological expression of multi-source data, and supports complex queries. Cross-modal data chains such as "user A-video B-video title-user C's comment" can be directly obtained through traversal. User behavior time series subgraphs can be constructed through timestamps (e.g., a user clicks on video VID1 at time t1 and collects video VID2 at time t2), providing a data foundation for subsequent time decay calculation (f(interaction) in step S3).

[0056] like Figure 2 As shown, preferably, before inputting the original text into the first large language model and extracting semantic feature vectors in step S1, the original text undergoes structured processing. This structured processing includes deduplication, word segmentation, stop word filtering, part-of-speech tagging, and entity extraction, and also includes the following steps:

[0057] Step S11: The original text is segmented and filtered for stop words, and then part-of-speech tagging and entity extraction are performed to obtain a preprocessed sequence. The preliminary entity table in the preprocessed sequence is vectorized and encoded to generate at least three types of vectors: semantic vector, pinyin vector, and part-of-speech vector.

[0058] Step S12: The semantic vector is fused with the pinyin vector and the memory vector to obtain the fused pinyin vector and the fused part-of-speech vector. Based on cosine similarity, the fused pinyin vector and the fused part-of-speech vector are spatially aligned under a contrastive learning framework (for example, contrastive learning can be used as the contrastive learning framework).

[0059] Step S13: Extract head entities using semantic vectors. Add the head entity to the fused pinyin vector, fused part-of-speech vector, and semantic vector to form a joint representation value. Feed this joint representation value into a tail entity and relation extraction network, outputting a triplet result. In this step, the tail entity can be based on a linear layer with two pointers (start / end). The input is the joint representation value, and the output is the probability distribution of the tail entity's start and end positions in the text. The tail entity's role is to lock the "object" boundary. The relation extraction network can be a multilayer perceptron (MLP) with the hidden states of the head and tail entities as input, outputting the softmax probability of the candidate relation category (i.e., a normalized probability output, transforming any real vector into a probability distribution with a sum of 1). Based on this, the relation extraction network can determine the semantic relation label between "subject-object".

[0060] For example, in some implementations, the tail entity and relation extraction networks can share the same joint representation encoder to ensure feature consistency. Tail entities are fed into the relation extraction network with start and end indices as hard constraints, while the relation probabilities generated by the relation sub-network, in turn, guide the tail entity pointers to adjust boundaries via the loss function gradient, achieving a "boundary-relationship" closed-loop optimization. Based on the tail entity and relation extraction networks, the overall loss entropy of the original text can be determined (overall loss entropy = tail entity cross entropy + relation classification cross entropy), enabling end-to-end backpropagation. This allows the localization of tail entities and relation determination to converge simultaneously under a unified objective, ultimately outputting a complete triplet (head entity, relation, tail entity).

[0061] In the optimized technical solutions provided in steps S11 to S13, a joint representation value is constructed by fusing vectors and contrastive learning. Semantic, pinyin, and part-of-speech vectors are generated synchronously for the preprocessed text, preserving the multimodal features of the language. A contrastive learning framework is used to achieve cross-modal spatial mapping between pinyin and part-of-speech vectors, solving the compatibility problem of heterogeneous features. Finally, a joint representation is constructed: based on head entity-guided triple extraction, achieving the fusion of structured knowledge and deep semantics. Deep semantic fusion improves the accuracy of semantic precision. Part-of-speech tagging (such as verb / noun recognition) constrains the entity extraction boundary, avoiding erroneous entity segmentation (such as "Apple phone" being mistakenly split into "Apple + phone"). It achieves the complementary effect of multi-vector fusion: pinyin vectors capture homophones and near-synonyms (such as "algorithm" and "calculate"), while part-of-speech vectors strengthen grammatical structures (such as verb-object relationships), complementing the semantic vectors.

[0062] In the optimization scheme provided in steps S11 to S13, the entity participates in both vector generation (G) and entity boundary correction (D). A contrastive learning framework is used to simultaneously optimize the pinyin / part-of-speech vectors to adapt to the semantic space (H). Finally, the extraction range is constrained by the head entity (e.g., "movie"), using "director-lead actor" rather than irrelevant attributes. The joint representation value improves the entity recognition success rate by more than ten percentage points, enhances the domain adaptability of semantic understanding, improves the generalization ability of contrastive learning, and makes the triple network transferable. The tail entity extraction network can be reused in new domains; only the head entity type needs to be adjusted (e.g., in the medical field, for videos or articles, the keyword "medicine" can be set as the head entity), which can multiply the information processing capacity of the subsequent first and second language models. The distinction and relationship between head entities and tail entities can be understood through relational triples (h, r, t). Head entity h is the subject object described by relation r, equivalent to the subject in a natural language sentence. Tail entity t is the target object that head entity h points to through relation r, i.e., the receiver or attribute value of relation r, equivalent to the object in a natural language sentence. Head and tail entities constitute semantic units through relation r; the head entity provides the subject, and the tail entity provides the object or attribute value. Together, they completely express a relational triple.

[0063] like Figure 3 As shown, for example, in step S12, the process of spatially aligning the fused pinyin vector and the fused part-of-speech vector based on cosine similarity within a contrastive learning framework further includes:

[0064] Step S121: Input the pinyin vector and part-of-speech vector into the contrastive learning framework simultaneously. Within the framework, the cosine similarity function is first called to calculate the distance metric between the two vectors.

[0065] Step S122 involves performing backpropagation based on the gradient signal of the progressive contrastive loss function. This drives the spatial alignment module to sequentially apply linear mapping and normalization operations to the pinyin vector and part-of-speech vector until the two vectors are aligned within a unified metric space. In step S122, the unified metric space refers to mapping the pinyin vector and part-of-speech vector to a common-dimensional Euclidean space. The distance between any two vectors in this Euclidean space is calculated using the same metric function (cosine distance or Euclidean distance), ensuring that features of different modalities can be compared on the same scale. The aligned state refers to the state within this Euclidean space where the distance between the pinyin vector and part-of-speech vector of the same class is minimized (intra-class compactness), the distance between samples of different classes is maximized (inter-class separation), and their distribution centers coincide, satisfying the convergence condition of the progressive contrastive loss.

[0066] The optimization techniques provided in steps S121 and S122 are essentially a progressive cross-modal vector alignment method. This method achieves spatial unification of features from heterogeneous data through dynamic gradient-guided spatial mapping. Step S121 establishes preliminary spatial associations through cosine similarity calculation, while step S122 achieves nonlinear calibration of the vector space through progressive contrastive loss-driven linear transformation. The similarity calculation in step S121 provides an initial alignment benchmark for step S122, avoiding mapping offsets caused by random initialization. Step S121 can improve the initial alignment accuracy of pinyin and part-of-speech vectors by more than 30%. Building upon step S121, step S122 addresses the premature convergence problem that may occur during contrastive learning by dynamically adjusting the gradient signal strength (e.g., emphasizing directional alignment initially and distance compression later). The distance metric in step S121 provides prior data for step S122, while the incremental optimization in step S122 can, in turn, correct any metric biases that may exist in step S121. The synergistic cooperation between steps S121 and S122 forms a closed-loop feedback, which is equivalent to constructing a joint distribution space of sound-form-meaning, thus solving the problem of the separation between NLP pinyin input method and grammatical analysis in the existing technology.

[0067] Preferably, in step S2, during the process of binding semantic feature vectors with content nodes to form content enhancement nodes, the content enhancement node specifically involves: packaging the semantic feature vectors, pinyin feature vectors, and part-of-speech feature vectors obtained after the original text has been processed by the first language model, and enhancing the original nodes with additional semantic information based on the extracted entity tags to form content enhancement nodes.

[0068] For example, the following steps can also be used to form a content-enhanced node: Content-enhanced node = original node + multi-channel semantic packet, and the content-enhanced node is constructed according to the following steps:

[0069] Step S123: Receive the original content node and extract static attributes such as node ID, title, and publication time.

[0070] Step S124: Perform word segmentation, stop word filtering, and entity extraction on the original text to obtain entity label table E.

[0071] Step S125: Feed the cleaned text into the first language model and output the semantic feature vector V_sem, the pinyin feature vector V_py, and the part-of-speech feature vector V_pos simultaneously.

[0072] Step S126: Concatenate the semantic feature vector V_sem, the pinyin feature vector V_py, the part-of-speech feature vector V_pos, and the entity label table E into a multi-channel tensor F_enhance according to a fixed dimension.

[0073] Step S127: Write the multi-channel tensor F_enhance into the node attribute field and bind it with the node ID to form an indexable and computable content-enhanced node.

[0074] In the optimization solutions provided in steps S123 to S127, through the coordinated efforts of different steps, a single node can carry indexable and computable multidimensional semantics, avoiding repeated text parsing. Multi-channel tensors are cascaded and connected to nodes, reducing the feature reading latency when traversing heterogeneous bipartite graphs, shortening the response time for intelligent recommendation or retrieval, improving the user experience, and aligning heterogeneous data such as semantics, pinyin, parts of speech, and entities within the same node. This enhances the downstream model's ability to model the joint probability distribution of user-content-text, laying a data foundation for providing more accurate intelligent recommendation results.

[0075] like Figure 4 As shown, preferably, in step S3, the process of generating the weighted adjacency matrix further includes:

[0076] Step S31: Obtain the user behavior sequence and run the confidence algorithm. The confidence algorithm includes a behavior-weight joint tensor. The parameters in the behavior-weight joint tensor include at least a set of behavior types.

[0077] Step S32: Calculate the confidence level based on the behavior-weight joint tensor, compare the confidence level calculation result with the preset threshold, retain the joint tensor entries that retain the preset conditions, and generate the corresponding high-confidence association set.

[0078] Step S33: Perform natural language transformation on the high-confidence association set according to the preset natural language template to generate a description set, and input the description set into the second language model. The second language model then outputs the corresponding set of interpretation logic rules. For example, in this step, the set of interpretation logic rules is interpretable symbolic logic, and the joint probability distribution of user-content-text is calculated based on the interpretable symbolic logic using a multimodal neural network.

[0079] The optimization solutions provided in steps S31 to S33 achieve transparency and accuracy in recommendation decisions through the synergy of confidence filtering and symbolic logic generation. The optimization solutions model the behavior-weight joint, bind user behavior types (such as clicks and favorites) with dynamic weight parameters, construct a tensor structure that can quantify the value of behavior, perform confidence filtering based on threshold driving, remove low-confidence associations through preset thresholds, retain high-value user-content interactions, and achieve interpretable transformation, converting high-confidence associations into natural language descriptions, and then generating interpretable symbolic logic rules through a language model, and finally calculating the joint probability distribution.

[0080] In steps S31 and S32, during collaborative filtering, multi-dimensional behavioral features (e.g., "favorites = 0.8, clicks = 0.3") are integrated using a behavior-weight joint tensor to quantify the contribution differences of different behaviors. Noise in the interaction (e.g., accidental clicks) is filtered out through threshold comparison, resulting in a high-confidence set (e.g., "User A - Video B: Favorites + Long Views"). Symbolic logic (e.g., "IF user's historical favorites of technology videos THEN recommendation weight + 0.2") directly constrains the joint probability distribution, avoiding the black-box random bias of traditional models. In step S33, natural language templates can be used as a bridge to transform tensor entries (e.g., "User A - Video B: confidence 0.9") into natural language descriptions (e.g., "User A's preference strength for video B is 90%)." A second language model achieves a dual interpretable transformation of "data-language-rules," realizing dynamic adaptive optimization, dynamic quantification of behavioral value, and a balance between interpretability and recommendation accuracy. In a test on an e-commerce platform, the optimization solution from steps S31 to S33 increased the recommendation conversion rate by about 20%, while user satisfaction with the recommendation reasons reached 92%.

[0081] Preferably, in step S5, the interpretation logic rule set is multimodal concatenated, and the multimodal concatenation result is used to calculate the joint probability distribution of user-content-text. Based on the probability values ​​of the joint probability distribution, corresponding intelligent recommendations are made. This method is suitable for the following application scenarios:

[0082] Application Scenario 1: The following scenario data exists in the technology channel of a short video platform:

[0083] User U1: Historical Behavior Sequence = "Browsing Tech Videos - Liking Drone Reviews - Completing Aerial Photography Tutorials". Content C1: Video Title: "XX Drone's First Flight Experience".

[0084] Original text T1: 120 characters in total, including video subtitles and bullet comments, containing the keywords "4K60fps" and "folding paddle".

[0085] Calculation results: Joint probability distribution: P(U1,C1,t1)=0.83;

[0086] Triggering rule: "IF user preference for 'aerial photography + technology' and text containing 'XX drone manufacturer' THEN recommendation weight +0.35" Recommendation action: The system inserts C1 in the second position of "Recommended for You" with an explanation: "Because you like aerial photography and technology content, and this video is highly relevant to your interests, please watch it."

[0087] Application Scenario 2: Data from the finance section of a news app:

[0088] User U2: Behavior sequence = "Read macroeconomic articles - bookmark X Bank's interest rate hike report".

[0089] Content C2: Article "X Bank's September Interest Rate Decision Preview".

[0090] Original text T2: 800 words of main text, entity extraction results in "X Bank, CPI, interest rate hike".

[0091] Calculation result: P(U2,C2,t2)=0.91;

[0092] Recommended action: News app push notification pop-up "You may be interested in: Bank interest rate hike analysis";

[0093] Post-event analysis: User click-through rate increased to 17.4% (compared to a baseline of 8.1%).

[0094] Application Scenario 3: Scenario data from the outdoor sports section of an e-commerce platform:

[0095] User U3: Behavior sequence = "Search for tents - Add to cart - Browse camping guides".

[0096] Content C3: Product "Lightweight Two-Person Tent".

[0097] The original text T3 contains product descriptions and user reviews, which contain the high-frequency words "lightweight" and "rainproof". However, "lightweight" has a special meaning in other fields, so it is more appropriate to change "lightweight" to "lightweight" in this application scenario.

[0098] Calculation result: P(U3,C3,T3)=0.75 (background data);

[0099] Recommended action: The system displays C3 in the 3rd position of the 1st row of "You May Like" and generates a natural language explanation: It matches your recent camping gear needs.

[0100] In the three application scenarios mentioned above, the three types of heterogeneous data—user behavior logs, content metadata, and raw text—are uniformly modeled into a heterogeneous bipartite graph. A large language model is used to upscale the raw text into semantic feature packages, which are then bound to content nodes to form "content-enhanced nodes." Subsequently, using these content-enhanced nodes as anchors, user behavior sequences and weighted adjacency matrices are combined to drive a second language model to generate interpretable symbolic rules. Finally, the symbolic rules, weight matrices, and behavior sequences are multimodally concatenated to estimate the joint probability of the "user-content-text" ternary model and rank and recommend the data. Analysis of these three application scenarios shows that this embodiment provides computable semantics through content-enhanced nodes, ensures cross-modal consistency through spatial alignment, filters noise, and ultimately binds users, content, and text to the same decision plane by calculating the joint probability distribution value, achieving "accurate + interpretable" personalized recommendations.

[0101] For the method steps disclosed in the above embodiments, the method steps are described as a series of actions for the purpose of simplicity. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily necessary for the embodiments of this application.

[0102] Any flowchart or other description of a process or method can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a specific logical function or process. Furthermore, the scope of preferred embodiments of this application includes additional implementations in which functions may be performed and implemented not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, or in accordance with program structures such as loops, branches, etc., to execute computer instructions and implement corresponding functions. This is something that those skilled in the art would readily understand when implementing embodiments of this application.

[0103] like Figure 5 As shown in the embodiments of this application, the company also discloses a personalized intelligent recommendation system based on a large language model, used to implement the aforementioned personalized intelligent recommendation method based on a large language model, including:

[0104] The heterogeneous bipartite graph generation module acquires user behavior data and constructs a heterogeneous bipartite graph based on the user behavior data. The heterogeneous bipartite graph includes user nodes, content nodes, interaction relationships, and original text.

[0105] The content enhancement node generation module inputs the original text into the first language model and extracts semantic feature vectors, then binds the semantic feature vectors to content nodes to form content enhancement nodes;

[0106] The weighted adjacency matrix generation module traverses the interaction relationship with the content enhancement node as a constraint, obtains the intersection between the content node and the interaction relationship, constructs a weighted edge between the user node and the content enhancement node based on the intersection, and generates a weighted adjacency matrix through the weight of the weighted edge. The weighted adjacency matrix is ​​used to represent the "user-text" logical relationship.

[0107] The explanation logic rule set output module, based on the user behavior sequence and combined with the weighted adjacency matrix, obtains the corresponding confidence association information and generates natural language description prompts, inputs them into the second language model, and outputs the explanation logic rule set.

[0108] The multimodal splicing module performs multimodal splicing on the interpretation logic rule set, calculates the joint probability distribution of user-content-text based on the multimodal splicing result, and performs corresponding intelligent recommendations based on the distribution of the joint probability.

[0109] The implementation methods of the system described above are merely illustrative. For example, the various functional modules, units, or subsystems within the system may or may not be physically separate, or they may or may not be physical units; that is, they may be located in the same place or distributed across multiple different systems and their subsystems or modules. Those skilled in the art can select some or all of the functional modules, units, or subsystems to achieve the objectives of the embodiments of this application according to actual needs. Those skilled in the art can understand and implement the above-described situations without any creative effort.

[0110] like Figure 6 As shown, this application embodiment, in addition to providing a personalized intelligent recommendation method and system based on a large language model, also provides a corresponding electronic device and storage medium:

[0111] An electronic device includes: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of a personalized intelligent recommendation method based on a large language model.

[0112] In some embodiments, the electronic device is specifically a computer device, including a non-volatile storage medium and a system bus, wherein the non-volatile storage medium includes an operating system, computer programs and a database, and the system bus is connected to internal memory, a network interface and a processor.

[0113] A computer-readable storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform steps of a personalized intelligent recommendation method based on a large language model.

[0114] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0115] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the specification of the embodiments of this application.

[0116] In the description of the embodiments of this application, the reference to terms such as "an embodiment," "example," "specific example," etc., means that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the embodiments of this application. In the description of the embodiments of this application, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0117] In the embodiments of this application, the functional modules can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. If implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various specific embodiments of the embodiments of this application. The aforementioned storage medium includes: USB flash drives, mobile hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0118] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this disclosure are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state disk (SSD)).

[0119] In this application, the term "exemplary" is used to mean "used as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes are not described in detail to avoid obscuring the description of this application with unnecessary detail. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of this application, and are not intended to limit them. Although the embodiments of this application have 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, and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the specific implementation methods of the embodiments of this application.

Claims

1. A personalized intelligent recommendation method based on a large language model, used to implement a personalized intelligent recommendation system based on a large language model, characterized in that, include: Obtain user behavior data, and construct a heterogeneous bipartite graph based on the user behavior data. The heterogeneous bipartite graph includes user nodes, content nodes, interaction relationships, and original text. The original text is input into the first language model and semantic feature vectors are extracted. The semantic feature vectors are then bound to content nodes to form content enhancement nodes. In the process of binding the semantic feature vectors to content nodes to form content enhancement nodes, the content enhancement nodes are specifically formed by: packaging the semantic feature vectors, pinyin feature vectors, and part-of-speech feature vectors obtained after the original text has been processed by the first language model, and enhancing the original nodes with additional semantic information based on the extracted entity tags to form the content enhancement nodes. Using the content enhancement node as a constraint, traverse the interaction relationship to obtain the intersection between the content node and the interaction relationship. Construct a weighted edge between the user node and the content enhancement node based on the intersection. Generate a weighted adjacency matrix through the weight of the weighted edge. The weighted adjacency matrix is ​​used to represent the "user-text" logical relationship. Based on the user behavior sequence and the weighted adjacency matrix, the corresponding confidence association information is obtained and natural language description prompt information is generated. This information is then input into the second language model and an explanation logic rule set is output. The set of interpretation logic rules is multimodal concatenated, and the joint probability distribution of user-content-text is calculated based on the multimodal concatenation result. The corresponding intelligent recommendation is then performed based on the probability value of the joint probability distribution.

2. The personalized intelligent recommendation method based on a large language model according to claim 1, characterized in that, The user node is used to represent the user entity and its attributes, including user ID and user basic profile. The content node is used to represent the interaction object. The interaction relationship is used to represent a single interaction record between the user and the content. The original text is specifically unstructured text.

3. The personalized intelligent recommendation method based on a large language model according to claim 1, characterized in that, Before inputting the original text into the first language model and extracting semantic feature vectors, the original text undergoes structured processing. This structured processing includes deduplication, word segmentation, stop word filtering, part-of-speech tagging, and entity extraction, wherein: The original text is segmented and filtered for stop words, and then part-of-speech tagging and entity extraction are performed to obtain a preprocessed sequence. The preliminary entity table in the preprocessed sequence is vectorized and encoded to generate at least three types of vectors: semantic vector, pinyin vector, and part-of-speech vector. The semantic vector is fused with the pinyin vector and part-of-speech vector to obtain a fused pinyin vector and a fused part-of-speech vector. Based on cosine similarity, the fused pinyin vector and the fused part-of-speech vector are spatially aligned under the contrastive learning framework. The head entity is extracted using semantic vectors. The head entity is then added to the fused pinyin vector, fused part-of-speech vector, and semantic vector to form a joint representation value. The joint representation value is then fed into a tail entity and relation extraction network to output triple results.

4. The personalized intelligent recommendation method based on a large language model according to claim 3, characterized in that, In the process of spatially aligning the fused pinyin vectors and fused part-of-speech vectors within a contrastive learning framework based on cosine similarity, the process further includes: The pinyin vector and part-of-speech vector are simultaneously input into the contrast learning framework. Within this framework, the cosine similarity function is first called to calculate the distance metric between the two vectors. Backpropagation is performed based on the gradient signal of the progressive contrastive loss function, driving the spatial alignment module to apply linear mapping and normalization operations to the pinyin vector and part-of-speech vector in sequence until the two vectors are aligned in a unified metric space.

5. The personalized intelligent recommendation method based on a large language model according to claim 1, characterized in that, During the process of generating and constructing the weighted adjacency matrix, user behavior sequences are obtained and a confidence algorithm is run. The confidence algorithm includes a behavior-weight joint tensor, and the parameters in the behavior-weight joint tensor include at least a set of behavior types. The confidence level is calculated based on the behavior-weight joint tensor. The confidence level calculation result is compared with a preset threshold. Joint tensor entries that retain the preset conditions are retained, and a corresponding high-confidence association set is generated. The high-confidence association set is subjected to natural language transformation according to a preset natural language template to generate a description set. The description set is then input into a second language model, which outputs the corresponding set of explanation logic rules.

6. The personalized intelligent recommendation method based on a large language model according to claim 5, characterized in that, The set of interpretation logic rules is interpretable symbolic logic, and the joint probability distribution of user-content-text is calculated based on the interpretable symbolic logic through a multimodal neural network.

7. A personalized intelligent recommendation system based on a large language model, characterized in that, The personalized intelligent recommendation method based on a large language model, as described in any one of claims 1 to 6, comprises: The heterogeneous bipartite graph generation module acquires user behavior data and constructs a heterogeneous bipartite graph based on the user behavior data. The heterogeneous bipartite graph includes user nodes, content nodes, interaction relationships, and original text. The content enhancement node generation module inputs the original text into the first language model and extracts semantic feature vectors. It then binds the semantic feature vectors to content nodes to form content enhancement nodes. Specifically, the content enhancement node involves packaging the semantic feature vectors, pinyin feature vectors, and part-of-speech feature vectors obtained after processing the original text through the first language model. Based on the extracted entity tags, additional semantic information is used to enhance the features of the original nodes, thus forming the content enhancement node. The weighted adjacency matrix generation module traverses the interaction relationship with the content enhancement node as a constraint, obtains the intersection between the content node and the interaction relationship, constructs a weighted edge between the user node and the content enhancement node based on the intersection, and generates a weighted adjacency matrix through the weight of the weighted edge. The weighted adjacency matrix is ​​used to represent the "user-text" logical relationship. The explanation logic rule set output module, based on the user behavior sequence and combined with the weighted adjacency matrix, obtains the corresponding confidence association information and generates natural language description prompt information, inputs it into the second language model and outputs the explanation logic rule set; The multimodal splicing module performs multimodal splicing on the interpretation logic rule set, calculates the joint probability distribution of user-content-text based on the multimodal splicing result, and performs corresponding intelligent recommendations based on the probability values ​​of the joint probability distribution.

8. An electronic device, characterized in that, include: The system includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus; the memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the steps of the method according to any one of claims 1 to 6.

Images