Personalized recommendation method and system based on cooperation of large language model and domain model

Abstract

The invention discloses a personalized recommendation method and system based on cooperation of a large language model and a domain model, and relates to the technical field of information recommendation. According to the method, under a target recommendation scene, unstructured data is processed through a large language model to obtain semantic pattern features, meanwhile, quantifiable operation records of a user are analyzed through a specified domain model, behavior pattern features are output, and synchronous extraction of unstructured semantic information and structured behavior information is achieved; bidirectional information supplement and knowledge transfer are carried out on the two types of features, a collaborative optimization feature mapping set is constructed, and unified conversion and synchronous scheduling of cross-modal features are realized; in combination with real-time interaction information reflecting the current intention of the user and scene demand changes, an initial recommendation list is generated through dual-model collaborative reasoning, and whether a personalized recommendation result is output or not is judged after dynamic sorting, so that accurate adaptation between the real-time demand of the cross-scene user and personalized recommendation is realized, the recommendation timeliness is improved, and the user experience is improved. And thus, the rapid adaptability of personalized recommendation is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information recommendation, and in particular to a personalized recommendation method and system based on cooperation of a large language model and a domain model. BACKGROUND

[0002] With the continuous expansion of the magnitude of digital content and services, the demand for users to efficiently obtain target content from massive information is increasingly urgent, and personalized recommendation has become the core link connecting users and information. At the same time, large language model (LLM, Large Language Model) technology has developed rapidly, with powerful semantic understanding, natural language interaction and cross-domain knowledge integration capabilities, providing a new paradigm for deep mining of user potential intent and optimizing recommendation interaction experience. While the domain model has accumulated professional data processing experience, scenario-based rules and recommendation effect stability advantages in specific industries in long-term practice, and the synergy of the two has become a key driving force to promote the upgrading of personalized recommendation technology to be more accurate and more scenario-based.

[0003] Existing personalized recommendation technology mainly realizes application through three types of paths. The first type relies on collaborative filtering and matrix decomposition technology, based on user historical behavior data to mine the association between users and specified targets, complete basic interest matching, and is suitable for early e-commerce and content platform scenarios. The second type uses deep learning models such as neural collaborative filtering (NCF, Neural Collaborative Filtering), integrates user portraits, specified target attributes and scenario features, and enhances feature interaction through neural networks to improve recommendation accuracy. The third type begins to integrate large language model capabilities, and some solutions use LLM to generate user interest text and specified target knowledge labels as supplementary features, or use LLM to optimize the natural language explanation of recommendation results, and preliminarily explore the application mode of the combination of the two.

[0004] For example, the patent number CN118551123B Chinese invention patent discloses a movie personalized recommendation method and system fusing a large language model and a knowledge graph, which includes: collecting a data set containing movie detailed information and preprocessing; constructing a knowledge graph based on the preprocessed data set; based on the knowledge graph, fine-tuning the large language model with prompt words, and constructing a fusion scoring model combined with a traditional recommendation algorithm; training the fusion scoring model, and using the trained fusion scoring model for scoring prediction and personalized recommendation.

[0005] For example, the publication number: CN120492590A Chinese invention patent application discloses a personalized content recommendation method, system and electronic equipment based on a large model, which comprises: in response to receiving a content query instruction, using at least one large language model to generate a question text corresponding to the content query instruction; according to the similarity between each question text and the text block in the content library, determine the corresponding text block; based on the first prompt word generated according to each question text and the corresponding text block, using at least one large language model to generate a recommendation result corresponding to the question text; output the recommendation result.

[0006] Under cross-scene recommendation, it is necessary to realize seamless connection of user's needs in different scenes and consistency of preference continuation. However, the features output by the large language model are usually non-structured vectors, and the existing technology usually does not fully consider the unified feature conversion and synchronous scheduling mechanism, resulting in time sequence misalignment of unstructured semantic features and structured behavior features of the domain model in the transmission and analysis link, further exacerbating the asynchronous problem that the large language model has output the recommendation candidate when the behavior calculation of the domain model is not completed, so that the timeliness of cross-modal feature collaboration is reduced, and there is a problem that the personalized recommendation adaptability under cross-scene recommendation is not high. SUMMARY

[0007] In order to solve the technical problem that the large-scale multi-source data has low computing efficiency in the dynamic expansion process in the prior art, the embodiments of the present application provide a personalized recommendation method and system based on the cooperation of large language model and domain model. The technical scheme is as follows: On the one hand, a personalized recommendation method based on the cooperation of large language model and domain model is provided, which comprises: step one, in the target recommendation scene, using a large language model to perform semantic understanding on the acquired unstructured data to output semantic pattern features, and using a specified domain model to perform collaborative analysis on the acquired user behavior data to output behavior pattern features, the user behavior data representing quantifiable operation records of the user in the target recommendation scene; step two, supplementing the semantic pattern features output by the large language model and the behavior pattern features output by the specified domain model in both directions to form a collaborative optimized feature mapping set; step three, combining the user real-time interaction information and the formed feature mapping set, generating an initial recommendation list through the collaborative reasoning of the large language model and the specified domain model, and determining whether to output a personalized recommendation result after dynamic sorting, the user real-time interaction information being used to reflect the current operation state and the change of the user's scene demand, and the personalized recommendation result being used to support fast adaptation under cross-modal scenes.

[0008] In another aspect, a personalized recommendation system based on large language model and field model collaboration is provided, which comprises: a double model mode feature output module, configured to, in a target recommendation scenario, utilize a large language model to perform semantic understanding on acquired unstructured data to output semantic mode features, and simultaneously utilize a specified field model to perform collaborative analysis on acquired user behavior data to output behavior mode features; A bidirectional information supplementing and mapping set construction module is configured to perform bidirectional information supplementing on the semantic mode features output by the large language model and the behavior mode features output by the specified field model to form a collaborative optimization feature mapping set; and an initial recommendation list generation and determination module is configured to combine real-time interaction information of a user and the formed feature mapping set to generate an initial recommendation list through collaborative reasoning of the large language model and the specified field model, and determine whether to output a personalized recommendation result after dynamic sorting.

[0009] The technical scheme provided by the embodiments of the present application has at least the following beneficial effects: 1. In the target recommendation scenario, the present application processes data in parallel through double models: the large language model is utilized to perform semantic understanding on unstructured data to output semantic mode features, and the specified field model is utilized to analyze user quantifiable operation records to output behavior mode features. This step realizes synchronous extraction of two types of features, avoids the one-sidedness of a single model focusing on semantics or behavior, eliminates information bias at the data source for subsequent cross-modal collaboration, performs bidirectional information supplementing and knowledge transfer on the semantic mode features and the behavior mode features, and constructs a collaborative optimization feature mapping set. Through a unified feature conversion and scheduling mechanism, this process solves the time sequence misalignment problem of transmission and analysis of the two types of features in the prior art, and fundamentally alleviates the low efficiency of cross-modal feature collaboration. In combination with real-time interaction information reflecting the current operation state of the user and changes in scene demand, the initial recommendation list is generated through double model collaborative reasoning, and whether to output a personalized recommendation result is determined after dynamic sorting. This step seamlessly connects the user's cross-scene demand, consistently continues the preference, and effectively improves the low cross-scene recommendation adaptability problem in the prior art.

[0010] 2、By first judging the semantic content of unstructured data, if it exceeds the preset semantic content, the sub-feature is refined to the preset number according to the semantic granularity, if it does not exceed the preset semantic content, the sorted coupled semantic vector is obtained, and then the semantic mode feature is generated and cross-text verification is performed to ensure stable matching of semantic requirements; then the behavior mode feature is output: the specified behavior frequency is calculated from the time sequence dimension to obtain the time sequence difference parameter, the behavior trend change is judged to adjust the time sequence weight or directly output according to the rule, and the new user refers to the behavior of the same tag group. This process solves the problem of unstructured semantic feature analysis through dynamic adjustment of semantic granularity and cross-text verification, optimizes the output of structured behavior features with the help of time sequence difference parameters and weight control, constructs a unified feature conversion and synchronous scheduling mechanism, avoids the time sequence misalignment of semantic and behavior feature transmission and analysis, reduces the asynchronous problem of model output, improves the efficiency of cross-modal feature coordination, realizes seamless connection of user demand and consistency of preference in cross-scene recommendation, and improves the adaptability of personalized recommendation.

[0011] 3、By first standardizing the semantic and behavior mode features in the formation of the collaborative optimized feature mapping set, different dimension features are converted to the same target space through the feature mapping matrix to reduce the dimension difference; then the semantic-behavior feature correlation is calculated, first Fourier transform is performed to generate frequency domain intensity value to capture time sequence correlation, then binary coding and cosine similarity are used to form discrete matching degree index, and finally the geometric mean is used to balance the influence of the two to avoid single index interference; then the feature weight is adjusted according to the deviation value of the correlation and the midpoint of the preset range, if the deviation exceeds the upper limit, the semantic mode feature weight is lowered and the behavior mode feature weight is raised, if the deviation value is negative and the opposite number exceeds the midpoint, the semantic mode feature weight is raised and the behavior mode feature weight is lowered, and if the deviation value is within the allowed range, the weight is maintained, and finally the feature mapping set is obtained by alignment matching. This process solves the dimension difference problem of unstructured and structured features through unified dimension conversion, ensures the accuracy of correlation calculation through geometric mean, realizes feature coordination through dynamic weight adjustment, constructs a unified feature conversion mechanism, avoids transmission and analysis time sequence misalignment and model output asynchrony, improves the efficiency of cross-modal feature coordination, helps seamless connection of user demand and preference continuity in cross-scene recommendation, and improves the adaptability of personalized recommendation.

[0012] 4. By first calling operation data from the user real-time interaction log and converting it into a real-time feature vector when generating the initial recommendation list, matching the correlation degree score with the feature mapping set, and comparing the candidate recommendation table generated based on the large language model with the selected list generated based on the specified field model, the specified target overlap rate is obtained; if the correlation degree score exceeds the preset correlation degree score and the specified target overlap rate is lower than the preset specified target overlap rate, it indicates that the user's current demand and historical matching degree are high but the list difference is large, and differential cross verification is required to complete the non-overlapping target and perform adaptability verification; if the correlation degree score does not exceed the preset and the overlap rate is not lower than the preset, the initial recommendation list containing ID, score and recommendation reason is directly generated based on the overlapping target; when determining whether to output the personalized recommendation result, the recommendation priority of the target in the initial list and the specified target proportion are used as the basis, and if the target meets the requirements, the personalized recommendation result is outputted and sorted according to the priority; if the target does not meet the requirements, the optimization or manual verification is prompted. The process relies on the feature mapping set to realize real-time and historical feature matching, solves the coordination problem of unstructured semantic and structured behavior characteristics through differential verification and adaptability verification, avoids time sequence misplacement and model asynchrony, improves cross-modal coordination efficiency, helps seamless connection of user demand and preference continuation in cross-scenarios, and improves the adaptability of personalized recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0014] Figure 1 The flowchart of the personalized recommendation method based on the cooperation of the large language model and the field model provided by the embodiments of the present application; Figure 2 The output flowchart of the behavior mode feature provided by the embodiments of the present application; Figure 3 The generation flowchart of the initial recommendation list provided by the embodiments of the present application; Figure 4 The structural schematic diagram of the personalized recommendation system based on the cooperation of the large language model and the field model provided by the embodiments of the present application. DETAILED DESCRIPTION

[0015] The technical solutions in the present application will be described below with reference to the drawings.

[0016] In the embodiments of the present application, the words such as "exemplary", "for example", etc. are used to represent an example, illustration or description. Any embodiment or design scheme described as "exemplary" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the word "exemplary" is intended to present the concept in a specific manner. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be one of the two.

[0017] To make the technical problems, technical solutions and advantages to be solved by the present application clearer, the following will be described in detail in combination with the drawings and specific embodiments.

[0018] The embodiments of the present application provide a personalized recommendation method based on large language model and field model cooperation, as shown in Figure 1 The flowchart of the personalized recommendation method based on large language model and field model cooperation, the processing flow of the method can include the following steps: Step one, in the target recommendation scene, the obtained unstructured data is understood by the large language model to output semantic pattern features, and the obtained user behavior data is analyzed by the specified field model to output behavior pattern features, realizing the synchronous extraction of unstructured semantic information and structured behavior information; unstructured data represents information that cannot be stored and parsed in a fixed format, usually including user comments, consultation texts, browsing content descriptions, scene dialogue records, user behavior data represents quantifiable operation records of users in the target recommendation scene, usually including quantifiable operation records of users in the target recommendation scene, usually including click frequency, stay time, collection / cancellation operation, purchase / browsing history times.

[0019] Step two, the semantic pattern features output by the large language model and the behavior pattern features output by the specified field model are supplemented in both directions to form a feature mapping set optimized in cooperation, realizing the unified conversion and synchronous scheduling of cross-modal features; the semantic pattern features are used to capture the potential semantic needs of users, such as understanding user needs and scenes through text, the behavior pattern features are used to quantify the explicit operation rules of users, such as reflecting actual user needs through behavior frequency and timing changes, and the feature mapping set is used to integrate general semantic knowledge and field behavior rules, providing a unified feature benchmark for subsequent real-time feature matching and recommendation list generation.

[0020] Step three involves combining real-time user interaction information and the resulting feature mapping set to generate an initial recommendation list through collaborative reasoning between a large language model and a specified domain model. After dynamic sorting, a decision is made on whether to output personalized recommendation results. This achieves seamless integration of user needs across scenarios and accurate adaptation of personalized recommendations, improving the timeliness and adaptability of recommendations. Real-time user interaction information reflects the user's current operational status and changes in scenario-based needs. Personalized recommendation results support rapid adaptation across modal scenarios. The initial recommendation list carries the core information of candidate recommendation targets, typically including the specified target ID, matching score, and recommendation reason.

[0021] Specifically, the large language model is built using the Transformer-BERT architecture: First, it is pre-trained using massive amounts of general text data (web pages, books, dialogue corpora) to capture semantic, grammatical, and common-sense associations through self-supervised learning (such as masked language modeling tasks). Then, it is fine-tuned using unstructured data from recommendation scenarios (user comments, consultation texts) to adapt to the needs of contextualized semantic understanding, ultimately resulting in a model that can output structured semantic pattern features. Its input is unstructured data from the target recommendation scenario, and its output is semantic pattern features such as user demand vectors and contextual semantic labels.

[0022] The specific domain model is constructed using a temporal collaborative filtering model: It uses historical user behavior data (clicks, purchase records) as training samples, learns user behavior patterns and temporal changes through the model, and incorporates recommendation scenario business rules (such as product category weights and user segmentation rules) to optimize model parameters, forming a targeted domain model. Its input is user behavior data (click frequency, dwell time), and its output is user behavior trend vectors, preference weight values, and other behavioral pattern features.

[0023] The two types of models have a clear division of labor. By inputting different data sources, they output semantic and behavioral pattern features respectively, which provides a foundation for subsequent feature fusion and recommendation list generation, solves the problem of asynchronous cross-modal features, and improves recommendation adaptability.

[0024] In a specific implementation, taking the online course recommendation scenario as an example: unstructured data includes user questions in the course consultation area (e.g., looking for a Python data analysis course suitable for beginners, preferably with practical examples) and course evaluation text (e.g., this course is too fast-paced, I hope there will be more explanations of basic concepts); user behavior data includes the user's course browsing history in the past 15 days (clicked on Python courses 8 times and data analysis courses 6 times), video viewing time (watched a single episode of a basic Python course for more than 40 minutes), and favorites (favorited 2 courses with practical tags).

[0025] First, unstructured data is processed using a large language model based on Transformer-BERT, outputting semantic pattern features such as Python needs for beginners, emphasis on practical examples, and preference for slow-paced explanations. Then, a temporal collaborative filtering domain model is used to analyze behavioral data, outputting behavioral pattern features such as a Python course preference weight of 0.8 and behavioral trend vectors for practical courses. Subsequently, both types of features are dimensionality-standardized (mapped to 512 dimensions), and weights are adjusted based on relevance to form a feature mapping set. When a user searches for Python practical courses in real time, the real-time features are matched against the mapping set, and the two models collaboratively generate an initial list containing course IDs, matching scores (e.g., a matching score of 91 for a Python practical course for beginners), and recommendation reasons. Personalized recommendations are then output after sorting.

[0026] It solves the problem of the disconnect between users' semantic needs and behavioral data in online education scenarios. The recommendations not only match the learning preferences expressed by users in the text, but also conform to actual browsing and operation habits. The connection between needs across learning stages (such as from basic to advanced) is smoother, which effectively improves the accuracy of course recommendations and user learning conversion rate.

[0027] Furthermore, the specific output process of semantic pattern features is as follows: First, the semantic content volume in the unstructured data is obtained. If the semantic content volume is greater than a preset semantic content volume, the semantic granularity ratio is calculated based on the obtained semantic content volume deviation, using the semantic granularity ratio formula. This ratio is the ratio of the semantic content volume deviation to the preset semantic content volume. The semantic content is then progressively divided into dimensions corresponding to the semantic content volume, decomposing the semantic content into a preset number (usually 10-20) of semantic sub-features. If the semantic content volume is not greater than the preset semantic content volume, the semantic vectors corresponding to the semantic content in the unstructured data are sorted by similarity. Each sorted semantic vector is assigned a corresponding weight coefficient and coupled, i.e., each semantic vector is multiplied by its corresponding weight coefficient, and all results are summed to obtain the semantic feature result. The weight coefficient is usually set by the pre-defined personnel based on the business objectives of historical recommendation scenarios and historical semantic matching effects, used to quantify the importance of each semantic vector. The vector sequence of the semantic feature result is obtained, and semantic pattern features are generated sequentially from high to low semantic association strength. Cross-text semantic consistency verification is then performed to improve the matching degree between the semantic pattern features and user needs.

[0028] The semantic granularity ratio is a quantitative representation of the difference between the preset semantic content amount and the acquired semantic content amount, used to reflect the fineness of semantic parsing of unstructured data. The preset semantic content amount is set by analyzing the average effective semantic length of historical unstructured data and combining it with the accuracy requirements of semantic parsing in the recommendation scenario to ensure that it is adapted to the needs of the scenario. The semantic content amount deviation represents the difference between the acquired semantic content amount and the preset semantic content amount. The vector sequence represents the ordered vector set formed by the semantic feature results according to the preset semantic logic. The semantic association strength represents the degree of matching between each ordered vector and the user's current needs. The semantic association strength is obtained by calculating the cosine similarity between each ordered vector and the user's current needs vector.

[0029] like Figure 2 The flowchart for outputting behavioral pattern features, as shown, follows this logic: First, obtain the temporal difference parameters. Then, determine if their absolute values ​​are not greater than a preset difference threshold. If so, directly output the behavioral pattern feature. If not, increase the temporal weight corresponding to the deviation of the temporal difference parameters and check if the temporal weight exceeds a preset range. If it does, pause the feature calculation process; if it does not, continue monitoring the temporal difference parameters. Furthermore, for special cases such as new users, behavioral pattern features can also be output, but these are based on the behavioral distribution patterns of user groups with the same feature label, ensuring the comprehensiveness and adaptability of the process.

[0030] It is important to understand that the specific output process of behavioral pattern features is as follows: From the time-series dimension of user behavior data, the frequency of specified behaviors within two adjacent statistical periods is retrieved. The difference between the frequency of specified behaviors in the later period and the frequency of specified behaviors in the previous period is defined as the time-series difference parameter, which is used to quantify the changing trend of user behavior over time. If the absolute value of the time-series difference parameter is greater than the preset difference threshold, it is determined that the behavior trend of the corresponding user has changed significantly, and a time-series weight corresponding to the deviation of the time-series difference parameter is added to the specified domain model. At the same time, it is monitored whether the time-series weight exceeds the preset time-series weight range. The preset difference threshold is set by analyzing the frequency of user's historical behaviors and calculating the upper limit of the frequency difference of 90%. The preset time-series weight range is set by referring to the optimal weight range of the same scenario domain model. The optimal weight range is usually 0.3-0.8.

[0031] If the temporal weight exceeds the preset temporal weight range, the feature calculation process based on that weight is paused to prevent excessive interference with the behavioral feature output process. If the temporal weight is within the preset temporal weight range, monitoring continues until the absolute value of the temporal difference parameter is not greater than the preset difference threshold, at which point the behavioral pattern feature is output. If the absolute value of the temporal difference parameter is not greater than the preset difference threshold, the corresponding user's behavioral trend is determined to be stable, and behavioral pattern features are output based solely on the corresponding user's historical behavioral patterns. If it is a new user, behavioral pattern features are output based on the behavioral distribution patterns of user groups with the same feature label. The same feature label represents a set of labels consistent with the new user's current scenario and behavioral patterns.

[0032] The cross-text semantic consistency verification process involves: acquiring semantic fragments from semantic pattern features, generating semantic vectors corresponding to each semantic fragment using a large language model, and calculating the cosine similarity of each semantic vector; if the cosine similarity is higher than a preset cosine similarity, the semantic features are considered consistent, and the semantic pattern features are output to ensure output stability in cross-text scenarios. The preset cosine similarity is represented by the sum and average of historical cosine similarities of historical cross-text semantic vectors; otherwise, the process goes back to the semantic pattern feature output stage, re-decomposes the semantic content into a preset number of semantic sub-features, and after a preset number of decompositions, if the re-acquired cosine similarity is still not higher than the preset cosine similarity, the preset personnel are prompted to perform semantic calibration until the semantic features are considered consistent.

[0033] In this embodiment, when outputting semantic pattern features, the processing method is dynamically adjusted according to the amount of semantic content. When the amount of semantic content exceeds the preset amount, semantic sub-features are split proportionally. When the amount is insufficient, semantic vectors are weighted and coupled. Semantic accuracy is also ensured through cross-text consistency verification, effectively improving the parsing accuracy of unstructured data and avoiding semantic bias caused by a single processing method. When outputting behavioral pattern features, the temporal difference parameters are combined with the preset difference threshold to dynamically adjust the temporal weights. New users are referenced to the behavior of groups with the same tag to accurately capture changes in user behavior trends. Both types of feature outputs rely on historical data and scene adaptation to set parameters, reducing the temporal misalignment between unstructured semantic and structured behavioral features, alleviating the asynchronous problem of the two models, laying the foundation for subsequent cross-modal feature collaboration, and significantly improving the adaptability of cross-scene personalized recommendations.

[0034] Furthermore, a collaboratively optimized feature mapping set is formed, specifically including: performing dimensionality standardization on the output semantic pattern features and behavioral pattern features to transform the two types of features to a specified dimensional space through a feature mapping matrix, such as mapping the 1024-dimensional vector of semantic features and the 256-dimensional vector of behavioral features to a 512-dimensional target space, reducing the dimensionality difference caused by different feature sources; based on the semantic-behavioral feature correlation degree in the current recommendation scenario, calculating the deviation value between the current semantic-behavioral feature correlation degree and the midpoint of the preset feature correlation degree range, where the preset feature correlation degree range is obtained by considering the historical semantic-behavioral feature correlation degrees of the target recommendation scenario over the past 3-6 months. Minimum and maximum values ​​are used to determine the feature correlation range after removing extreme outliers (such as data below the 1st percentile or above the 99th percentile). The minimum value of the remaining data is used as the lower limit and the maximum value is used as the upper limit. The midpoint of the preset feature correlation range represents the median of the historical semantic-behavioral feature correlation within the preset feature correlation range for this scenario. If the deviation value is within the allowable range, the current semantic and behavioral pattern feature weights are kept unchanged. The two types of features are mapped and matched according to the dimension corresponding to the current weight through a feature alignment algorithm to eliminate dimensional deviation. The dimension vector after mapping and matching is statistically analyzed to obtain a feature mapping set that integrates general semantic knowledge and domain behavioral rules.

[0035] A positive deviation value indicates that if the deviation value is not less than the upper limit of the allowable deviation range, the coupling redundancy is determined to be too high. Based on the first deviation value, the semantic pattern feature weights are adjusted downwards by a certain percentage, while the behavioral pattern feature weights are adjusted upwards by the same percentage. The sum of the semantic pattern feature weights and the behavioral pattern feature weights is 1. The first deviation value represents the ratio of the difference between the obtained deviation value and the upper limit of the allowable deviation range to the upper limit of the allowable deviation range. If the deviation value is less than the upper limit of the allowable deviation range, the current semantic pattern feature weights and behavioral pattern feature weights remain unchanged. The upper limit of the allowable deviation range is determined by statistically analyzing the maximum deviation value without redundancy risk in historical feature coupling, adding a 10%-15% tolerance interval, and finally determining the upper limit to ensure that coupling redundancy is only determined when the deviation significantly exceeds the safe range. The lower limit of the allowable deviation range is usually set to 0.

[0036] The deviation value is negative: If the opposite of the deviation value is not less than the midpoint of the allowable deviation value range, the coupling depth is determined to be insufficient. The semantic pattern feature weights are adjusted upward based on the second deviation value, while the behavioral pattern feature weights are adjusted downward by the same proportion. The second deviation value represents the ratio of the difference between the midpoint of the allowable deviation value range and the obtained deviation value to the midpoint of the allowable deviation value range. If the opposite of the deviation value is less than the midpoint of the allowable deviation value range, the current semantic pattern feature weights and behavioral pattern feature weights remain unchanged. Both the first and second deviation values ​​are used to quantify the degree of feature coupling deviation from the expected state, and the numerical range is limited to [0,1] to ensure that the weight adjustment ratio is controllable. The midpoint of the allowable deviation value range is usually taken as 50% of the upper limit of the allowable deviation value range, which reflects the balanced expectation of feature coupling and can be compared with the deviation value (positive / negative) to accurately determine whether the coupling is redundant or insufficient in depth.

[0037] The semantic-behavioral feature correlation is obtained as follows: Fourier transform is performed on the semantic pattern features and behavioral pattern features after dimensionality standardization to convert the two types of features from the time domain to the frequency domain, presenting the periodic pattern of the two types of features changing over time. Then, the amplitude and phase of each frequency point in the spectrum of the two types of features are extracted. For each frequency point, the Pearson correlation coefficient of the amplitude of the semantic and behavioral pattern features (quantifying the consistency of the amplitude change trend, with a value of [-1,1], the closer to 1, the higher the amplitude overlap) and the mean of the phase difference between the two are calculated (measuring phase coordination, the smaller the mean, the more synchronized the phase). Then, the Pearson correlation coefficient (forwardly mapped to [0,1]) and the mean phase difference (backwardly mapped to [0,1], the smaller the difference, the larger the mapped value) are normalized and then weighted and summed according to a 4:6 ratio (the degree of amplitude overlap is more critical to the periodic correlation) to obtain a comprehensive quantized frequency domain intensity value, which is set to [0,1], in order to capture the correlation pattern of the above two types of features over time. The higher the value, the closer the correlation between the two types of features in the periodic trend.

[0038] Binary encoding is used to convert the median of continuous values ​​of two types of features into a threshold. Continuous values ​​greater than the threshold are mapped to 1, and continuous values ​​less than or equal to the threshold are mapped to 0, thus transforming both types of features into equal-length 0-1 binary sequences. The two binary sequences are then treated as vectors, and the cosine similarity between the sequences is calculated, which is the vector dot product divided by the product of the magnitudes of the two vectors. This forms a discrete matching index used to quantify the degree of matching between discretized features. For example, if the semantic feature sequence is [1,0,1,1] and the behavioral feature sequence is [1,0,0,1], the cosine similarity formula yields a value that reflects the directional consistency between the two features in the discrete dimension. The discrete matching index represents the degree of directional consistency between the two types of features in the vector space after discretization; a higher value indicates a stronger matching between the two types of features in the discrete feature dimension.

[0039] Since the influence of frequency domain intensity (reflecting periodic correlation) and discrete matching degree index (reflecting discrete dimensional matching) on ​​the correlation degree needs to be balanced, a geometric mean calculation is used for fusion: the values ​​of the two indices are multiplied, and then the square root of the product is taken to obtain the semantic-behavioral feature correlation degree. If the frequency domain intensity value is 0.8 and the discrete matching degree index is 0.9, the correlation degree is... =0.847, this result can avoid the excessive dominance of a single indicator (such as an excessively high frequency domain intensity value) on the overall correlation, and more objectively reflect the degree of synergy between the two types of features.

[0040] In this embodiment, dimensional standardization eliminates the dimensional differences between semantic and behavioral features, avoiding fusion barriers caused by different sources and laying the foundation for subsequent collaboration. The semantic-behavioral feature correlation calculation integrates the periodicity of the frequency domain and the matching of discrete dimensions, and uses geometric averaging to balance the influence of the index, ensuring that the correlation evaluation is objective and accurate. For cases where the deviation value is positive, the semantic weight is lowered and the behavioral weight is raised when the upper limit is exceeded, which can effectively reduce coupling redundancy and avoid the excessive dominance of semantic features. When the deviation value is negative and the absolute value exceeds the midpoint, the semantic weight is raised and the behavioral weight is lowered, which can compensate for insufficient coupling depth and strengthen the collaboration between the two types of features. At the same time, the weight adjustment is based on the quantized first and second deviation values ​​and is limited to the range [0,1], ensuring that the adjustment is controllable and fits the scenario expectations. The final feature mapping set integrates general semantic knowledge and domain behavioral rules, which not only alleviates the problem of cross-modal feature temporal misalignment, but also provides high-quality feature support for subsequent recommendations, significantly improving the adaptability and accuracy of cross-scenario recommendations.

[0041] like Figure 3 The flowchart shown illustrates the initial recommendation list generation process. Its design logic is as follows: First, obtain the relevance score and the specified target overlap rate. By comparing the relevance score with the corresponding preset value, determine whether to proceed directly to the next step. If the relevance score is not greater than the corresponding preset value, then based on the comparison of the specified target overlap rate with the preset value, decide whether to perform differentiated cross-validation and suitability verification. If the specified target overlap rate exceeds the corresponding preset value, prompt for manual review. Only after verification or meeting the conditions will the initial recommendation list be generated. This logic ensures that recommendations are based on reasonable relevance and overlap rates, balancing automation and manual intervention to guarantee recommendation quality.

[0042] Further understanding is needed regarding the specific steps for generating the initial recommendation list: Retrieving current operation data from real-time user interaction logs and converting it into a real-time feature vector; then performing feature matching calculations with the existing feature mapping set to obtain the correlation score between real-time features and historical mapping features; comparing the candidate recommendation list generated based on a large language model with the candidate list generated based on a specified domain model to obtain the specified target overlap rate; the correlation score, obtained through the Euclidean distance between real-time features and historical mapping features, quantifies the closeness of the match between them, with higher scores indicating a higher degree of fit between current needs and historical features; if the correlation score is greater than a preset correlation score, and the specified target overlap rate is less than a preset overlap rate, it indicates a high degree of match between the user's current needs and historical features, but significant differences between the two recommendation lists, requiring differentiated cross-validation to supplement the corresponding information in the two lists. While identifying non-overlapping target information, the system improves the matching degree between non-overlapping targets and users' real-time needs. Simultaneously, it performs adaptive verification of differentiated cross-validation results. The preset relevance score is determined by statistically analyzing the mean Euclidean distance of users' historical features during the recommendation process, mapping it back to the base score, and adding a 5%-8% tolerance range. If the relevance score is not greater than the preset relevance score, and the overlap rate of the specified targets is not less than the preset overlap rate set by preset personnel, it indicates that the user's current needs have a low matching degree with historical features, but the consistency between the two types of recommendation lists is high. An initial recommendation list containing the specified target ID, matching score, and recommendation reason is directly generated based on the overlapping target information. If the above conditions are not met, it indicates that the matching degree between the user's current needs and historical features, and the consistency between the two types of recommendation lists, are in an intermediate state. Preset personnel are then prompted to intervene and manually review the information to avoid fluctuations in recommendation accuracy due to the intermediate state.

[0043] The adaptation verification of the differentiated cross-validation results is specifically as follows: The specified targets corresponding to the non-overlapping specified target information in the candidate recommendation table are designated as semantic-priority targets, and the specified targets corresponding to the non-overlapping specified target information in the candidate list are designated as behavior-priority targets. Behavioral pattern features in the feature mapping set are called to obtain the length of the longest common subsequence between the feature pattern features and the corresponding user's historical behavior sequence. Combined with the total length of the historical behavior sequence, the length ratio is obtained, i.e., the ratio of the longest common subsequence length to the total length of the historical behavior sequence. Simultaneously, the user behavior weight coefficients in the feature mapping set are combined to generate the behavioral fit of the semantic-priority targets through weighted summation. Simultaneously, the semantic pattern features in the feature mapping set are called to calculate the cosine similarity between the attribute labels of the behavior-priority targets and the semantic vector of the user's current operation state, thus obtaining the semantic fit of the behavior-priority targets.

[0044] Only targets with behavioral fit no less than a preset behavioral fit and semantic fit no less than a preset semantic fit, corresponding to semantic priority targets and behavioral priority targets, are retained. Targets that do not meet the conditions (i.e., behavioral fit less than a preset behavioral fit, and semantic fit less than a preset semantic fit, corresponding to semantic priority targets and behavioral priority targets) are removed, and the reasons for removal are synchronized to the designated personnel. The preset behavioral fit is determined by analyzing the behavioral fit data of behavioral priority targets clicked / converted in the historical recommendations of the target scenario, taking the average value and lowering it by 5%-8% to ensure coverage of most effective behavioral matching targets. The preset semantic fit is determined by statistically analyzing the average semantic fit of semantic priority targets recognized by historical users, plus a 5%-10% error tolerance range, which avoids missing high-quality targets and reduces interference from low-fit targets.

[0045] The determination of whether to output personalized recommendation results is as follows: First, obtain the recommendation priority and percentage of each specified target in the initial recommendation list. Recommendation priority includes allowing and disallowing recommendations. If the recommendation priority is allowing recommendations and the percentage of specified targets is not less than a preset percentage, then personalized recommendation results are output. This means using the generated initial recommendation list and sorting the corresponding initial recommendation lists in descending order of recommendation priority. The preset percentage is typically set to 60%-70% of the total number of targets in the initial recommendation list. This percentage is determined by considering user browsing habits in the target recommendation scenario (e.g., the average number of recommended targets a user focuses on per visit) and the required recommendation accuracy. This ensures that the output personalized recommendation list contains enough high-quality targets (allowed recommendations) while avoiding limited user choices due to insufficient quantity. If the recommendation priority is disallowed recommendations and the percentage of specified targets is less than the preset percentage, then the initial recommendation list needs optimization, and a recommendation anomaly warning is issued. Otherwise, it indicates that neither the output nor optimization conditions are met, and a manual verification warning is issued.

[0046] Specifically, the target overlap rate represents the number of identical specified targets in the candidate recommendation table generated by the large language model and the candidate list generated by the specified domain model. It is used to quantify the consistency of the recommendation results from both models. Both the candidate recommendation table and the candidate list output targets from two dimensions: semantic understanding and behavioral analysis, based on user recommendation needs. The overlap rate reflects the consistency of the recommendations from both models. Large language models perform semantic understanding on real-time user interaction information and unstructured content (such as user reviews, product descriptions, and content tags), and combine this with their built-in contextual association and intent reasoning capabilities to generate a set of recommendation candidates that highly matches the user's semantic needs, i.e., a candidate recommendation table. Domain-specific models focus on in-depth analysis of user behavior data, such as user purchase history, browsing time, and add-to-cart behavior in the e-commerce field, and user collection, likes, and forwarding sequences in the content field. Through reasoning on this behavior data, they generate a candidate list based on behavioral trajectories.

[0047] In this embodiment, a dual-indicator approach of correlation score and overlap rate is used for judgment. A correlation threshold is set based on historical data, and manual review is triggered in intermediate states. This avoids misjudgment based on a single indicator and reduces accuracy fluctuations. Differentiated cross-validation completes the list of non-overlapping targets, and adaptability verification filters for behavioral / semantic fit, eliminating low-quality targets to ensure that recommended targets align with user needs. Result judgment is based on allowable recommendation priority and a 60%-70% proportion of recommended targets, satisfying user selection needs while improving adaptability accuracy. When targets are not met, optimization prompts or warnings are issued to adapt to recommendation requirements in different scenarios. The entire process relies on historical data to set thresholds and record the reasons for elimination, achieving standardized recommendations while facilitating manual traceability and adjustment. This effectively alleviates the problems of poor demand coordination and insufficient feature collaboration across scenarios, improving the reliability of personalized recommendations and user satisfaction.

[0048] This invention provides a personalized recommendation system based on the collaboration of a large language model and a domain model, such as... Figure 4 The diagram shows the structure of a personalized recommendation system based on the collaboration of a large language model and a domain model. This system can include: a dual-model pattern feature output module, used to perform semantic understanding on acquired unstructured data in a target recommendation scenario using a large language model to output semantic pattern features, and simultaneously using a specified domain model to collaboratively analyze acquired user behavior data to output behavioral pattern features; a bidirectional information supplementation and mapping set construction module, used to supplement the semantic pattern features output by the large language model and the behavioral pattern features output by the specified domain model with bidirectional information to form a collaboratively optimized feature mapping set; and an initial recommendation list generation and judgment module, used to combine real-time user interaction information and the formed feature mapping set to generate an initial recommendation list through collaborative reasoning between the large language model and the specified domain model, and then dynamically sort it to determine whether to output a personalized recommendation result.

[0049] In this embodiment, the dual-model pattern feature output module parses unstructured data using a large language model to output semantic pattern features, and analyzes user behavior data using a domain model to output behavioral pattern features. The two models perform their respective functions while cooperating, avoiding the limitations of a single model in processing data and providing an accurate feature foundation for subsequent recommendations. The bidirectional information supplementation and mapping set construction module supplements the two types of features bidirectionally, constructing a collaboratively optimized feature mapping set to eliminate information discrepancies between semantic and behavioral features, making the features more comprehensive and aligned with user needs. The initial recommendation list generation and judgment module combines real-time user interaction information and the feature mapping set, generates an initial recommendation list through collaborative reasoning of the two models, and dynamically sorts and judges the output results. This not only quickly responds to real-time user preferences but also improves recommendation accuracy, reduces invalid recommendations, and significantly optimizes user experience and recommendation conversion efficiency.

[0050] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, 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., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0051] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0052] In various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0053] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0054] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A personalized recommendation method based on the collaboration of a large language model and a domain model, characterized in that, The method includes: Step 1: In the target recommendation scenario, use a large language model to perform semantic understanding on the acquired unstructured data to output semantic pattern features. At the same time, use a specified domain model to perform collaborative analysis on the acquired user behavior data to output behavior pattern features. The user behavior data represents the quantifiable operation records of users in the target recommendation scenario. Step 2: The semantic pattern features output by the large language model and the behavioral pattern features output by the specified domain model are supplemented by two-way information to form a collaboratively optimized feature mapping set. Step 3: Combining real-time user interaction information and the resulting feature mapping set, an initial recommendation list is generated through collaborative reasoning between a large language model and a specified domain model. After dynamic sorting, it is determined whether to output personalized recommendation results. The real-time user interaction information is used to reflect the user's current operation status and changes in scenario-based needs. The personalized recommendation results are used to support rapid adaptation in cross-modal scenarios.

2. The personalized recommendation method based on the collaboration of a large language model and a domain model as described in claim 1, characterized in that, The specific output process of the semantic pattern features is as follows: The semantic content volume in the unstructured data is obtained. If the semantic content volume is greater than the preset semantic content volume, the semantic granularity ratio is calculated based on the deviation of the obtained semantic content volume and the semantic granularity ratio formula. The semantic content is then decomposed into a preset number of semantic sub-features by progressively splitting the semantic content volume into the dimensions of the semantic content. The semantic granularity ratio is represented by the result of quantifying the difference between the preset semantic content amount and the acquired semantic content amount and the preset semantic content amount, and is used to reflect the fineness of the semantic parsing of unstructured data. If the amount of semantic content is not greater than the preset amount of semantic content, the semantic vectors of the corresponding semantic content in the unstructured data are sorted by similarity, and each sorted semantic vector is assigned a corresponding weight coefficient and coupled to obtain the semantic feature result. Obtain the vector sequence of semantic feature results, and generate semantic pattern features in descending order of semantic association strength. Perform cross-text semantic consistency verification to improve the matching degree between semantic pattern features and user needs. The vector sequence represents an ordered set of vectors formed by semantic feature results according to preset semantic logic, and the semantic association strength represents the degree of matching between each ordered vector and the user's current needs.

3. The personalized recommendation method based on the collaboration of a large language model and a domain model as described in claim 2, characterized in that, The cross-text semantic consistency verification is specifically performed as follows: Semantic fragments are obtained from semantic pattern features, and semantic vectors corresponding to each semantic fragment are generated through a large language model. The cosine similarity of each semantic vector is then calculated. If the cosine similarity is higher than the preset cosine similarity, the semantic features are determined to be consistent, and the semantic pattern features are output to ensure the output stability in cross-text scenarios. Conversely, if the semantic content is not fully decomposed, the process returns to the semantic pattern feature output stage, where the semantic content is re-decomposed into a preset number of semantic sub-features. After a preset number of decompositions, if the re-obtained cosine similarity is still not higher than the preset cosine similarity, the preset personnel are prompted to perform semantic calibration until the semantic features are determined to be consistent.

4. The personalized recommendation method based on the collaboration of a large language model and a domain model as described in claim 1, characterized in that, The specific output process of the behavioral pattern features is as follows: From the time-series dimension of user behavior data, the frequency of a specified behavior within two adjacent statistical periods is retrieved. The difference between the frequency of a specified behavior in the later period and the frequency of a specified behavior in the previous period is defined as the time-series difference parameter, which is used to quantify the changing trend of user behavior over time. If the absolute value of the temporal difference parameter is greater than the preset difference threshold, then a temporal weight corresponding to the deviation of the temporal difference parameter is added to the specified domain model, and the temporal weight is monitored to see if it exceeds the preset temporal weight range. If the temporal weight exceeds the preset temporal weight range, the feature calculation process based on that weight will be paused to prevent excessive interference with the behavioral feature output process. If the temporal weight is within the preset temporal weight range, monitoring will continue until the absolute value of the temporal difference parameter is not greater than the preset difference threshold, and then the behavioral pattern feature will be output. If the absolute value of the temporal difference parameter is not greater than the preset difference threshold, then the behavioral pattern feature is output based on the historical behavioral pattern of the corresponding user. If the user is new, behavioral pattern features are output based on the behavioral distribution patterns of user groups with the same feature tags. The same feature tags represent a set of tags that are consistent with the current scenario and behavioral patterns of the new user.

5. The personalized recommendation method based on the collaboration of a large language model and a domain model as described in claim 1, characterized in that, The formation of the collaboratively optimized feature mapping set specifically includes: The first step is to perform dimensionality standardization on the output semantic pattern features and behavioral pattern features, so as to transform the two types of features to a specified dimensional space through the feature mapping matrix, thereby reducing the dimensionality difference caused by different feature sources. The second step is to calculate the deviation between the current semantic-behavioral feature correlation degree and the midpoint of the preset feature correlation degree range, based on the semantic-behavioral feature correlation degree in the current recommendation scenario. The third step is to make the deviation value positive: if the deviation value is not less than the upper limit of the allowable range of the deviation value, it is determined that the coupling redundancy is too high, and the reduction ratio of the semantic pattern feature weight is obtained based on the first deviation value. At the same time, the behavioral pattern feature weight is increased by the same ratio. The sum of the semantic pattern feature weight and the behavioral pattern feature weight is 1. If the deviation value is less than the upper limit of the allowable deviation range, the current semantic pattern feature weights and behavioral pattern feature weights remain unchanged. Fourth step, the deviation value is negative: if the opposite of the deviation value is not less than the midpoint of the allowable range of the deviation value, it is determined that the coupling depth is insufficient, and the adjustment ratio of the semantic pattern feature weight is obtained based on the second deviation value, while the behavioral pattern feature weight is adjusted down by the same ratio. If the negative of the deviation value is less than the midpoint of the allowable range of the deviation value, then the current semantic pattern feature weights and behavioral pattern feature weights remain unchanged. Both the first deviation value and the second deviation value are used to quantify the degree to which the feature coupling degree deviates from the expected state; Fifth, if the deviation value is within the allowable range, the current semantic and behavioral pattern feature weights are kept unchanged, and the two types of features are aligned and matched according to the dimensions corresponding to the current weights to obtain a feature mapping set that integrates general semantic knowledge and domain behavioral rules.

6. The personalized recommendation method based on the collaboration of a large language model and a domain model as described in claim 5, characterized in that, The semantic-behavioral feature correlation degree is obtained as follows: Fourier transform is performed on the semantic pattern features and behavioral pattern features after dimensionality standardization to convert the two types of features from the time domain to the frequency domain. At the same time, frequency domain intensity values ​​are generated by analyzing the corresponding spectral distribution state to capture the correlation between the two types of features and time. The frequency domain intensity value represents the degree of overlap and phase coordination of the amplitudes of the two types of features at each frequency component in the frequency domain space; Binary encoding is used to map the continuous values ​​of two types of features to a 0-1 sequence. The cosine similarity between the sequences is calculated using the binary sequences of the two types of features to form a discrete matching index for quantifying the degree of matching of discrete features. The discrete matching degree index represents the degree of directional consistency between two types of features in the vector space after discretization. The obtained frequency domain intensity value and discrete matching degree index are geometrically averaged to obtain the semantic-behavioral feature correlation degree. The geometric average operation is used to balance the influence of frequency domain intensity value and discrete matching degree index on the correlation degree result.

7. The personalized recommendation method based on the collaboration of a large language model and a domain model as described in claim 1, characterized in that, The specific steps for generating the initial recommendation list are as follows: The system retrieves current operation data from real-time user interaction logs, converts it into real-time feature vectors, and performs feature matching calculations with the established feature mapping set to obtain the correlation score between real-time features and historical mapping features. By comparing the candidate recommendation table generated based on the large language model with the candidate list generated based on the specified domain model, the system obtains the overlap rate of the specified target. The correlation score is obtained by the Euclidean distance between real-time features and historical mapping features, and is used to quantify the degree of matching between real-time features and historical mapping features. If the relevance score is greater than the preset relevance score and the overlap rate of the specified target is less than the preset overlap rate, then differential cross-validation is performed to supplement the information of the specified targets that do not overlap in the two lists, improve the matching degree between the specified targets that do not overlap and the user's real-time needs, and verify the adaptability of the differential cross-validation results. If the relevance score is not greater than the preset relevance score and the overlap rate of the specified target is not less than the preset overlap rate, then an initial recommendation list containing the specified target ID, matching score and recommendation reason will be generated based on the overlapping specified target information. If the above conditions are not met, a pre-selected person will be prompted to intervene for manual review.

8. The personalized recommendation method based on the collaboration of a large language model and a domain model as described in claim 7, characterized in that, The adaptation verification of the differentiated cross-validation results specifically includes: The specified targets corresponding to the specified target information that do not overlap in the candidate recommendation table are recorded as semantic priority targets, and the specified targets corresponding to the specified target information that do not overlap in the candidate list are recorded as behavior priority targets. Call the behavioral pattern features in the feature mapping set, obtain the length of the longest common subsequence between the feature and the corresponding user's historical behavior sequence, and combine it with the total length of the historical behavior sequence to obtain the length ratio. At the same time, combine the user behavior weight coefficient and generate the behavioral fit of the semantic priority target by weighted summation. Synchronously call the semantic pattern features in the feature mapping set, calculate the cosine similarity between the attribute labels of the behavior priority target and the semantic vector of the user's current operation state, and obtain the semantic fit of the behavior priority target. Only retain the semantic and behavioral priority targets that have a behavioral fit of no less than the preset behavioral fit and a semantic fit of no less than the preset semantic fit. For the specified targets that do not meet the conditions, remove them and synchronize the recorded removal reasons to the preset personnel.

9. The personalized recommendation method based on the collaboration of a large language model and a domain model as described in claim 8, characterized in that, The determination of whether to output a personalized recommendation result specifically involves: Obtain the recommendation priority and percentage of each specified target in the initial recommendation list, wherein the recommendation priority includes allowed recommendation results and disallowed recommendation results; If the recommendation priority is to allow the recommendation results, and the specified target proportion is not less than the preset proportion, then the personalized recommendation results will be output, and the corresponding initial recommendation list will be sorted in descending order of recommendation priority. If the recommendation priority is to disallow recommended results, and the specified target percentage is less than the preset percentage, then it is determined that the initial recommendation list needs to be optimized, and a recommendation error message will be issued. Otherwise, it indicates that neither the output conditions nor the optimization conditions are met, and manual verification and early warning will be performed.

10. A personalized recommendation system based on the collaboration of a large language model and a domain model, employing the personalized recommendation method based on the collaboration of a large language model and a domain model as described in any one of claims 1-9, characterized in that, include: The dual-model pattern feature output module is used in target recommendation scenarios to perform semantic understanding on the acquired unstructured data using a large language model to output semantic pattern features, and at the same time to perform collaborative analysis on the acquired user behavior data using a specified domain model to output behavioral pattern features. The bidirectional information supplementation and mapping set construction module is used to supplement the semantic pattern features output by the large language model and the behavioral pattern features output by the specified domain model with bidirectional information to form a collaboratively optimized feature mapping set. The initial recommendation list generation and determination module combines real-time user interaction information and the resulting feature mapping set to generate an initial recommendation list through collaborative reasoning between a large language model and a specified domain model. After dynamic sorting, it determines whether to output personalized recommendation results.

Images