A user multi-interest recommendation method and system based on a large model and frequency domain decomposition

Abstract

This application relates to the field of large-scale model applications, and more specifically to a user multi-interest recommendation method and system based on a large-scale model and frequency domain decomposition. The method includes: obtaining the objective attribute features of a user's historical interaction items and candidate items based on the objective attributes of the items; selecting several objective interest sequences from the historical interaction items to form candidate items based on the objective attribute features; processing user comment information on candidate items and historical interaction items using a large language model to select several subjective interest sequences from the historical interaction items to constitute candidate items; sequentially performing temporal encoding, frequency domain decomposition, and inverse Fourier transform on the sequences to obtain temporal representations; fusing the temporal representations corresponding to the sequences to obtain a comprehensive interest representation; and performing preference prediction based on the comprehensive interest representation and the objective attribute features of the candidate items to output the user's item recommendation results. This application aims to improve the matching degree between the accuracy and personalized experience of the recommendation results.

Description

Technical Field

[0001] This invention relates to the field of large model applications, and more specifically, to a user multi-interest recommendation method and system based on large models and frequency domain decomposition. Background Technology

[0002] In today's digital service ecosystem, recommender systems serve as the core hub connecting massive amounts of information with user needs, and their performance directly determines the quality of user experience and the platform's commercial value. To accurately capture users' personalized preferences, existing technologies typically rely on data mining of historical user interaction sequences, such as clicks, browsing, and purchases, using sequence modeling methods to depict the evolution of user interests. To more comprehensively describe the complex interest structures of users across different preference dimensions, multi-interest modeling methods have emerged. These methods construct multiple interest vectors, attempting to characterize users' diverse preferences from a parallel perspective.

[0003] While the methods described above improve the accuracy of recommendations to some extent, they still have significant limitations in deeply exploring users' decision-making psychology and satisfaction-driving mechanisms.

[0004] First, existing technologies have limited capabilities in integrating user subjective experience feedback. Current multi-interest modeling methods are mainly based on interaction behavior data. Even when some solutions incorporate user comment text, they typically only extract general semantics and struggle to delve into specific attributes or aspects. This coarse-grained analysis approach cannot accurately capture differences in user decision-making at a finer-grained level, resulting in insufficient ability to characterize the deeper mechanisms of user interest evolution.

[0005] Secondly, existing methods are insufficient in filtering key information when processing user historical behavior sequences and are easily interfered with by noisy interactions. User historical behavior is often disorganized, containing a large number of records with low relevance to the current recommendation goal. Traditional modeling methods still struggle to effectively identify key subsequences that truly reflect the user's core needs, resulting in interest representations being interfered with by irrelevant behaviors, affecting the accuracy of modeling and the interpretability of results.

[0006] Finally, existing technologies typically model user behavior on a uniform timeline, which can easily lead to the model over-reliance on historical high-frequency behaviors and may also make it susceptible to short-term noise. This limits the system's ability to express the multi-level dynamic evolution of user interests, thereby affecting the accuracy and personalization of recommendation results. Summary of the Invention

[0007] This invention provides a user multi-interest recommendation method and system based on a large model and frequency domain decomposition, which improves the matching degree of recommendation results in terms of accuracy and personalized experience.

[0008] According to a first aspect of this application, a user multi-interest recommendation method based on a large model and frequency domain decomposition is provided, the method comprising:

[0009] Based on preset objective attributes of the project, obtain the objective attribute characteristics of the user's historical interaction projects;

[0010] Based on the objective attributes of the project, obtain the objective attribute characteristics of the candidate project;

[0011] Based on the objective attribute characteristics of the candidate items and the objective attribute characteristics of the historical interaction items, a number of historical interaction items are selected from the historical interaction items to form an objective interest sequence of the candidate items;

[0012] The user comment information of the candidate items and the user comment information of the historical interaction items are processed by a large language model, and several items are selected from the historical interaction items to form the subjective interest sequence of the candidate items;

[0013] The objective interest sequence and the subjective interest sequence are sequentially subjected to time-series coding, frequency domain decomposition, and inverse Fourier transform to obtain their corresponding time-domain representations.

[0014] The temporal representations corresponding to the objective interest sequence and the subjective interest sequence are fused to obtain the corresponding comprehensive interest representation;

[0015] Based on the comprehensive interest representation and the objective attribute features of the candidate items, preference prediction is performed, and the item recommendation results for the user are output.

[0016] Understandably, this application effectively addresses the limited ability of existing technologies to integrate interest modeling and experience feedback by constructing a dual-sequence architecture of objective and subjective interest sequences. In this process, it utilizes a large language model to deeply mine semantic information in user comments, capturing users' subjective feelings before and after decision-making, thus overcoming the deficiency of relying solely on objective interaction behavior to reflect users' true psychological perceptions. Simultaneously, a frequency domain decomposition strategy is introduced to separate and fuse user interest signals across multiple time scales, explicitly characterizing interest patterns at different time scales. This results in a structured, dynamic, and interpretable representation of user interests, ultimately generating user recommendation results that significantly improve the accuracy and personalization of the recommendation outcomes.

[0017] Preferably, the step of obtaining the objective attribute characteristics of a user's historical interaction items based on preset objective attributes of the items includes:

[0018] Preset the objective attribute categories of the project, and preset the trainable embedding lookup table corresponding to each objective attribute category of the project;

[0019] Obtain the objective attributes of the historical interaction items, and based on the objective attributes of the historical interaction items, find the corresponding low-dimensional dense vectors from the corresponding embedding lookup tables respectively.

[0020] The low-dimensional dense vectors of the historical interaction items are sequentially concatenated, dimension aligned, and dimension mapped to obtain the first objective attribute representation vector.

[0021] Several objective interest perspectives are preset, and each objective interest perspective corresponds to a learnable linear projection matrix;

[0022] Based on the linear projection matrix corresponding to the objective interest perspective, the first objective attribute representation vector is feature-mapped to obtain the corresponding initial item representation;

[0023] The initial item representations of the historical interaction items under the corresponding objective interest perspective are sorted according to the interaction time between the corresponding historical interaction items and the user to obtain the initial item representation sequence. The initial item representation sequence is then input into a temporal coding network for processing, and the temporally enhanced item representation sequence is output.

[0024] The time-enhanced item representation sequence of the historical interaction item under all objective interest perspectives is used as the objective attribute feature of the historical interaction item, wherein each time-enhanced item representation sequence includes the item representation corresponding to each historical interaction item under each objective interest perspective.

[0025] Preferably, obtaining the objective attribute features of candidate projects based on the objective attributes of the projects includes:

[0026] Obtain the objective attributes of the candidate projects;

[0027] Based on the objective attributes of the candidate projects, the corresponding low-dimensional dense vectors are retrieved from the corresponding embedding lookup tables.

[0028] The low-dimensional dense vectors of the candidate projects are sequentially concatenated, dimension aligned, and dimension mapped to obtain the second objective attribute representation vector.

[0029] Based on the linear projection matrix corresponding to the objective interest perspective, the second objective attribute representation vector is feature-mapped to obtain the corresponding item representation;

[0030] The project representation of the candidate project under all objective interest perspectives is taken as the objective attribute feature of the candidate project.

[0031] Understandably, this application, through the synergy of multi-view projection and temporal modeling, not only decouples the objective attributes of the project from multiple dimensions, but also captures the dynamic evolution of user interests, thereby improving the fineness and timeliness of feature representation and providing a more accurate and interpretable user interest modeling foundation for subsequent personalized recommendations.

[0032] Preferably, the step of selecting a plurality of historical interaction items from the historical interaction items to form an objective interest sequence of the candidate items based on the objective attribute characteristics of the candidate items and the objective attribute characteristics of the historical interaction items includes:

[0033] For each objective interest perspective, based on the corresponding project representation, the objective similarity between the candidate project and the historical interaction project is obtained;

[0034] For each objective interest perspective, based on the objective similarity, several historical interaction items are selected from the historical interaction items to form the objective interest sequence of the candidate items under the corresponding objective interest perspective.

[0035] Understandably, this application constructs objective interest sequences based on objective similarity, ensuring the relevance of objective interest sequences in terms of objective attributes, and enhancing the balance between content relevance and user experience satisfaction in the recommendation results of this application.

[0036] Preferably, the step of processing user comment information of the candidate items and user comment information of the historical interaction items through a large language model, and selecting several items from the historical interaction items to form a subjective interest sequence of the candidate items, includes:

[0037] The preset prompts are combined with the user comments of the candidate items and the user comments of the historical interaction items to construct input information, which is then input into the large language model to obtain the corresponding scores, pre-event sentiment features and post-event sentiment features.

[0038] The subjective features of the historical interaction items and the subjective features of the candidate items are obtained based on the ratings, pre-event sentiment features, and post-event sentiment features corresponding to the historical interaction items and the candidate items, respectively.

[0039] Based on the subjective characteristics of the historical interaction items and the subjective characteristics of the candidate items, several items are selected from the historical interaction items to form the subjective interest sequence of the candidate items.

[0040] Understandably, this application utilizes a large language model to deeply analyze user comment information, explicitly extracting pre-event and post-event emotional features. This effectively overcomes the limitations of existing technologies that rely solely on objective interaction behavior. It accurately captures the psychological gap between users' expectations before decision-making and their actual perceptions after interaction, transforming difficult-to-quantify subjective experiences into calculable structured features. This approach more realistically reflects users' decision-making psychology and satisfaction-driven mechanisms, thereby enhancing the semantic richness and interpretability of user interest representations.

[0041] Preferably, obtaining the subjective features of the historical interaction items and the subjective features of the candidate items based on the ratings, pre-event sentiment features, and post-event sentiment features of the historical interaction items and the candidate items respectively includes:

[0042] Based on the historical interaction items, the ratings corresponding to the candidate items, and the post-event sentiment features, respectively, the post-event sentiment scores of the historical interaction items and the candidate items are obtained.

[0043] Based on the historical interaction items, the corresponding ratings, pre-event sentiment features, and post-event sentiment features, respectively, obtain the experience expectation gap value of the historical interaction items and the experience expectation gap value of the candidate items;

[0044] Subjective features of the historical interaction items are constructed based on their post-event sentiment scores and experience expectation gaps, and subjective features of the candidate items are constructed based on their post-event sentiment scores and experience expectation gaps.

[0045] Understandably, this application constructs subjective features by combining the experience expectation gap value and post-event emotional score, achieving a comprehensive quantification of users' psychological state. Specifically, the experience expectation gap value accurately captures the psychological gap between users before and after decision-making, while the post-event emotional score objectively reflects the final satisfaction and genuine experience feedback after user interaction. The fusion of these two factors to construct subjective features can more accurately filter historical interaction items with a high degree of consistency in emotional experience, significantly enhancing the robustness of subjective interest sequences, and thus improving the matching degree and interpretability of recommendation results to users' potential psychological needs.

[0046] Preferably, the step of processing user comment information of the candidate items and user comment information of the historical interaction items through a large language model, and selecting several items from the historical interaction items to form a subjective interest sequence of the candidate items, includes:

[0047] The preset prompts are combined with the user comments of the candidate items and the user comments of the historical interaction items to construct input information, which is then input into the large language model to obtain the preliminary evaluation categories corresponding to the candidate items and the historical interaction items, as well as the scores, pre-event sentiment features and post-event sentiment features corresponding to the preliminary evaluation categories.

[0048] The preliminary evaluation categories of the historical interaction items and the preliminary evaluation categories of the candidate items are clustered and merged based on semantic similarity to obtain standardized evaluation categories;

[0049] Based on the scores, pre-event sentiment features, and post-event sentiment features corresponding to each standardized evaluation category of the historical interaction project, the subjective features of each standardized evaluation category of the historical interaction project are obtained; and based on the scores, pre-event sentiment features, and post-event sentiment features corresponding to each standardized evaluation category of the candidate project, the subjective features of each standardized evaluation category of the candidate project are obtained.

[0050] Based on the subjective characteristics of each standardized evaluation category of the historical interaction projects and the subjective characteristics of each standardized evaluation category of the candidate projects, several projects are selected from the historical interaction projects to form the subjective interest sequence of the candidate projects.

[0051] Understandably, this application uses a large language model to perform aspect-level analysis of user review information, explicitly extracting the expectation-perception gap signal, and transforming users' subjective experiences into computable structured features. This overcomes the limitations of existing technologies that rely solely on objective behavior, and more realistically reflects users' decision-making psychology and satisfaction-driven mechanisms, thereby enhancing the semantic richness and interpretability of user interest representations. Simultaneously, through a semantic similarity-based evaluation category clustering and merging mechanism, fragmented user reviews are transformed into standardized evaluation categories. This captures users' preferences on specific attributes from the attribute experience level, making the construction of subjective interest sequences more accurate and effectively matching the potential needs of candidate items in specific attribute dimensions, further improving the accuracy and personalization of recommendations.

[0052] Preferably, obtaining the subjective characteristics of each standardized evaluation category of the historical interaction project based on the rating, pre-event sentiment characteristics, and post-event sentiment characteristics corresponding to each standardized evaluation category of the historical interaction project includes:

[0053] Based on the scores and post-event sentiment features corresponding to each standardized evaluation category of the historical interaction project, the post-event sentiment score of each standardized evaluation category of the historical interaction project is obtained.

[0054] Based on the scores, pre-event emotional features, and post-event emotional features corresponding to each standardized evaluation category of the historical interaction project, obtain the experience expectation gap value for each standardized evaluation category of the historical interaction project.

[0055] Based on the ex-post emotional score and experience expectation gap for each standardized evaluation category of the historical interaction project, subjective features for each standardized evaluation category of the historical interaction project are constructed.

[0056] And / or, based on the scores, pre-event sentiment features, and post-event sentiment features corresponding to each standardized evaluation category of the candidate projects, obtain the subjective features of each standardized evaluation category of the candidate projects, including:

[0057] Based on the scores and post-event sentiment features corresponding to each standardized evaluation category of the candidate projects, the post-event sentiment score for each standardized evaluation category of the candidate projects is obtained.

[0058] Based on the scores, pre-event emotional features, and post-event emotional features corresponding to each standardized evaluation category of the candidate projects, obtain the experience expectation gap value for each standardized evaluation category of the candidate projects.

[0059] Based on the ex-post affective score and experience expectation gap for each standardized evaluation category of the candidate projects, subjective features for each standardized evaluation category of the candidate projects are constructed.

[0060] Understandably, this application calculates the experience expectation gap value under standardized evaluation categories, which can achieve extremely high-precision decoupling of user interests, thereby constructing corresponding subjective features. This allows for the flexible use of interactive items with similar attributes in history when generating subjective interest sequences, based on the specific attribute weaknesses or strengths of candidate items. This effectively overcomes the lack of experience difference modeling in the attribute dimension of existing technologies, enabling recommendation methods to accurately avoid or satisfy specific user pain points, and significantly improving the robustness of recommendations in complex scenarios.

[0061] Preferably, based on the subjective characteristics of the historical interaction items and the subjective characteristics of the candidate items, several items are selected from the historical interaction items to form a subjective interest sequence of the candidate items, including:

[0062] Based on a preset matching function, the subjective matching degree between the subjective features of the candidate items and the subjective features of each historical interaction item is calculated.

[0063] Based on the subjective matching degree, several items are selected from the historical interaction item sequence to form the subjective interest sequence of the candidate items.

[0064] Understandably, this application effectively addresses the technical problems of insufficient sequential information filtering capabilities and susceptibility to noisy interactions by filtering historical interaction items based on subjective matching. Because it aligns candidate items with historical interaction items at the emotional and experiential levels based on their subjective characteristics, it selects only historical interaction items that are highly correlated in subjective experience to form a subjective interest sequence. This ensures that subsequent modeling processes focus on key sequences that truly reflect the user's current interest intentions, thereby improving the purity of interest representation.

[0065] Preferably, the step of sequentially performing time-domain encoding, frequency-domain decomposition, and inverse Fourier transform on the objective interest sequence and the subjective interest sequence to obtain the corresponding time-domain representation includes:

[0066] The objective interest sequence and the subjective interest sequence are respectively time-series encoded to obtain corresponding time-series encoded signals;

[0067] The time-coded signal is decomposed in the frequency domain to obtain frequency domain components of at least two preset frequency bands;

[0068] The frequency domain components are subjected to inverse Fourier transform to obtain the corresponding time domain representation.

[0069] Understandably, this application, by introducing a frequency domain decomposition strategy, achieves explicit separation and structured modeling of users' interests across multiple time scales, effectively overcoming the limitations of traditional single-time-domain modeling in balancing long-term and short-term dependencies. Specifically, the interest sequence is decomposed into multiple frequency band components in the frequency domain, accurately corresponding to long-term stable preferences, periodic interest patterns, and short-term dynamic behaviors. This allows the application to more meticulously depict the multi-level structure and evolutionary patterns of user interests, effectively avoiding mutual interference between features at different time scales, and significantly improving the application's adaptability to dynamic changes in user interests and its modeling accuracy.

[0070] According to a second aspect of this application, a user multi-interest recommendation system based on a large model and frequency domain decomposition is provided, the system comprising:

[0071] The first feature acquisition module is used to acquire the objective attribute features of the user's historical interaction items based on preset objective attributes of the items;

[0072] The second feature acquisition module is used to acquire the objective attribute features of the candidate projects based on the objective attributes of the projects.

[0073] An objective interest sequence acquisition module is used to select several historical interaction items from the historical interaction items based on the objective attribute characteristics of the candidate items and the objective attribute characteristics of the historical interaction items, thereby forming an objective interest sequence of the candidate items;

[0074] The subjective interest sequence acquisition module is used to process the user comment information of the candidate items and the user comment information of the historical interaction items through a large language model, and select several items from the historical interaction items to form the subjective interest sequence of the candidate items;

[0075] The time-domain representation acquisition module is used to sequentially encode, decompose in the frequency domain, and perform inverse Fourier transform on the objective interest sequence and the subjective interest sequence to obtain the corresponding time-domain representations.

[0076] The comprehensive interest representation acquisition module is used to fuse the time-domain representations corresponding to the objective interest sequence and the subjective interest sequence respectively to obtain the corresponding comprehensive interest representation;

[0077] The recommendation module is used to predict preferences based on the comprehensive interest representation and the objective attribute features of the candidate items, and output the item recommendation results for the user.

[0078] According to a third aspect of this application, an electronic device is provided, comprising:

[0079] Memory, used to store one or more computer programs;

[0080] A processor, when the one or more computer programs are executed by the processor, implements the user multi-interest recommendation method based on large model and frequency domain decomposition as described in the first aspect above.

[0081] According to a fourth aspect of this application, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the user multi-interest recommendation method based on large model and frequency domain decomposition described in the first aspect.

[0082] Based on any of the above aspects, the user multi-interest recommendation method, system, electronic device, and storage medium based on large model and frequency domain decomposition provided in the embodiments of this application have the following beneficial effects:

[0083] This application uses a large language model to perform aspect-level analysis of user comment information, explicitly extracting the expectation-perception gap signal, and transforming the user's subjective experience into computable structured features. This effectively overcomes the limitations of existing technologies that rely solely on objective behavior or overall sentiment analysis, and can more realistically reflect the user's decision-making psychology and satisfaction-driven mechanisms. This significantly enhances the semantic richness and interpretability of user interest representation, realizes the perception and fusion of fine-grained experience expectation gaps, and improves the psychological credibility of interest modeling.

[0084] This application proposes a dual-path parallel screening mechanism based on objective attribute similarity and subjective experience matching. This mechanism identifies historical interaction item sequences that are highly relevant to the current candidate items from different perspectives, effectively mitigating the interference of long-tail behavior and noisy interactions on interest modeling. It achieves complementary fusion of subjective and objective signals, improves the accuracy and robustness of sequence information screening, and thus effectively reduces the negative impact of noisy interactions on the recommendation results.

[0085] This application introduces a frequency domain decomposition strategy, which decomposes the interest sequence into multiple frequency band components in the frequency domain, corresponding to long-term stable preferences, periodic interest patterns, and short-term dynamic behaviors, respectively. This breaks through the limitations of traditional single time domain modeling, and can more meticulously depict the multi-level structure and evolution of user interests, significantly improving the adaptability to dynamic changes in user interests.

[0086] This application, through the fusion of multi-perspective and multi-scale interest representations, can simultaneously consider users' long-term preferences, cyclical habits, and immediate intentions, generating a more comprehensive and three-dimensional user profile. Experiments on public datasets show that this application outperforms existing baseline models in terms of accuracy, recall, and normalized depreciation cumulative gain (NDCG), especially in long-tail item recommendation and cross-domain recommendation scenarios, significantly improving the accuracy, personalization, and diversity of recommendation results.

[0087] This application does not rely on domain-specific knowledge and can be applied to various recommendation scenarios such as e-commerce, content, and services. Its modular design supports flexible replacement or expansion of various components, has high engineering feasibility and system adaptability, strong method versatility, and is easy to expand and deploy, providing a reusable technical framework for recommendation systems in different fields. Attached Figure Description

[0088] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0089] Figure 1 This diagram illustrates an application scenario of the user multi-interest recommendation method based on a large model and frequency domain decomposition provided in this embodiment.

[0090] Figure 2 The flowchart shows the user multi-interest recommendation method based on large model and frequency domain decomposition provided in this embodiment.

[0091] Figure 3 This embodiment provides a schematic diagram of the sub-steps of step S100.

[0092] Figure 4This embodiment provides a schematic diagram of the sub-steps of step S300.

[0093] Figure 5 This embodiment provides a functional module diagram of a user multi-interest recommendation system based on a large model and frequency domain decomposition.

[0094] Figure 6 This embodiment provides a schematic diagram of the electronic device. Detailed Implementation

[0095] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this application. To better illustrate the following embodiments, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product; it is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0096] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0097] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0098] The specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0099] An exemplary diagram illustrates an application scenario of a user multi-interest recommendation method based on a large model and frequency domain decomposition, provided in an embodiment of this application. Figure 1 As shown, the application scenario includes at least a server 100 and a terminal 200 that can communicate with the server 100.

[0100] Understandably, the server 100 can be an independent electronic device or a cluster of multiple electronic devices; the terminal 200 can be a smartphone terminal, personal computer, tablet computer, vehicle terminal, etc., but is not limited to these.

[0101] In one possible implementation, server 100 and terminal 200 can respectively execute the user multi-interest recommendation method based on large model and frequency domain decomposition provided in the embodiments of this application. Alternatively, the user multi-interest recommendation method based on large model and frequency domain decomposition provided in the embodiments of this application can be partially executed in server 100 and partially executed in terminal 200.

[0102] like Figure 2 As shown, this embodiment provides a user multi-interest recommendation method based on a large model and frequency domain decomposition. The method can be further divided into the following steps:

[0103] S100: Based on preset objective attributes of the project, obtain the objective attribute characteristics of the user's historical interaction projects;

[0104] In the specific implementation process, items that have been interacted with by the user are obtained in advance as the historical interaction items. Based on this, a historical interaction item sequence can be constructed. The historical interaction item sequence can be composed of a triple (user ID, item ID, timestamp), where the timestamp represents the exact moment when the user interacts with the item. Furthermore, objective attributes of the items can be preset and associated with the item ID.

[0105] Specifically, such as Figure 3 As shown, the specific steps of step S100 may include:

[0106] S110. Preset the objective attribute categories of the project, and preset the trainable embedding lookup table corresponding to each objective attribute category of the project;

[0107] Specifically, the objective attribute categories of the project are pre-defined, such as brand, primary category, price range, etc., and an independent trainable embedding lookup table is pre-defined for each objective attribute category of the project. Each trainable embedding lookup table maps the specific discrete values ​​under each objective attribute category of the project into a low-dimensional dense vector.

[0108] S120. Obtain the objective attributes of the historical interaction items;

[0109] Specifically, based on the objective attribute category of the project, the objective attributes of each of the historical interaction projects are obtained.

[0110] S130. Based on the objective attributes of the historical interaction items, find the corresponding low-dimensional dense vectors from the corresponding embedding lookup tables.

[0111] S140. Concatenate the low-dimensional dense vectors of the historical interaction items to obtain the corresponding first concatenated vector;

[0112] Specifically, the low-dimensional dense vectors of the objective attributes of each item in the same historical interaction project are sequentially concatenated to obtain a first concatenated vector with a higher dimension.

[0113] S150. Perform dimension alignment and dimension mapping on the first concatenated vector to obtain the first objective attribute representation vector.

[0114] Specifically, each of the first concatenated vectors is transformed through a linear layer (weight matrix + bias) to achieve dimension alignment: the transformed first concatenated vectors are mapped to a unified target dimension to obtain the first objective attribute representation vector corresponding to the historical interaction item.

[0115] S160. Several objective interest perspectives are preset, and each objective interest perspective corresponds to a learnable linear projection matrix.

[0116] S170. Based on the linear projection matrix corresponding to the objective interest perspective, perform feature mapping on the first objective attribute representation vector to obtain the corresponding initial item representation.

[0117] S180. The initial item representations of the historical interaction items under the corresponding objective interest perspective are sorted according to the interaction time between the corresponding historical interaction items and the user to obtain an initial item representation sequence. The initial item representation sequence is input into a temporal coding network for processing, and a temporally enhanced item representation sequence is output.

[0118] Specifically, in the time-enhanced item representation sequence, each of the historical interaction items has a time-enhanced item representation.

[0119] For example, the temporal coding network is implemented using a Transformer Encoder structure, which consists of a multi-head self-attention layer and a feedforward neural network, to capture long-term dependencies and attribute association patterns in user behavior sequences.

[0120] S190. The time-enhanced item representation sequence of the historical interaction items under all objective interest perspectives is taken as the objective attribute feature of the historical interaction items.

[0121] S200. Based on the objective attributes of the project, obtain the objective attribute features of the candidate project;

[0122] Specifically, the specific steps of step S200 may include:

[0123] S210. Obtain the objective attributes of the candidate projects;

[0124] Specifically, the objective attributes of each candidate project are obtained based on the objective attribute category of the project.

[0125] S220. Based on the objective attributes of the candidate projects, find the corresponding low-dimensional dense vectors from the corresponding embedding lookup tables.

[0126] Specifically, the embedded lookup table can be obtained based on the above-described step S110.

[0127] S230. Concatenate the low-dimensional dense vectors of the candidate items to obtain the corresponding second concatenated vector;

[0128] Specifically, the low-dimensional dense vectors of the objective attributes of each candidate project are sequentially concatenated to obtain a second concatenated vector with a higher dimension.

[0129] S240. Perform dimension alignment and dimension mapping on the second concatenated vector to obtain the second objective attribute representation vector.

[0130] Specifically, each of the second concatenated vectors is transformed through a linear layer (weight matrix + bias) to achieve dimension alignment: the transformed second concatenated vectors are mapped to a unified target dimension to obtain the second objective attribute representation vector corresponding to the candidate item;

[0131] S250. Based on the linear projection matrix corresponding to the objective interest perspective, perform feature mapping on the second objective attribute representation vector to obtain the corresponding item representation.

[0132] Specifically, the linear projection matrix is ​​obtained based on the above step S160.

[0133] S260. The project representation of the candidate project under all objective interest perspectives shall be taken as the objective attribute feature of the candidate project.

[0134] S300. Based on the objective attribute characteristics of the candidate items and the objective attribute characteristics of the historical interaction items, select several historical interaction items from the historical interaction items to form an objective interest sequence of the candidate items.

[0135] Specifically, such as Figure 4 As shown, the specific steps of step S300 include:

[0136] S310. For each objective interest perspective, based on the corresponding item representation, obtain the objective similarity between the candidate item and the historical interaction item;

[0137] Specifically, for each objective interest perspective, the candidate item has a corresponding item representation, and each historical interaction item has a corresponding item representation (i.e., the time-enhanced item representation obtained in step S180). The cosine similarity between the item representation of the candidate item and the item representation of the historical interaction item is calculated as the objective similarity between the candidate item and the historical interaction item under the corresponding objective interest perspective.

[0138] Furthermore, considering the overall credibility and popularity of the project, a behavioral intensity weight is introduced to calculate the objective similarity. This behavioral intensity weight can be obtained from historical average ratings. Number of historical ratings The historical average rating is constructed using a Bayesian averaging approach; the historical average rating is obtained by calculating the historical average rating of all users for the historical interaction item, and the number of historical ratings is obtained by calculating the total number of historical ratings of all users for the historical interaction item. Specifically, the behavior intensity weight... w rating (i t ) The calculation formula is as follows:

[0139] w rating (i t )=

[0140] in, i t This represents the historical interaction items, and C represents the adjustment smoothing constant, which controls the sensitivity of the quantity influence.

[0141] Furthermore, item co-occurrence weights can be introduced to calculate the objective similarity. These weights can be obtained by statistically analyzing the normalized co-occurrence scores of historical interaction items and candidate items in user behavior. The item co-occurrence weights are used to emphasize item pairs that frequently co-occur in actual interactions. Specifically, the item co-occurrence weights... w co (i t ,i ) It can be obtained based on the following formula:

[0142] w co (i t ,i )=

[0143] in, Indicates candidate projects. Indicates simultaneous interaction with historical projects and candidate projects The number of users who have interacted with the user; Indicates interaction with historical projects or candidate projects The total number of users who have interacted at least once; w co ( i t ,i ) This represents the co-occurrence weight of the items, used to indicate the correlation between two items in real user behavior patterns.

[0144] Based on the above, the final objective similarity between the candidate item and the historical interaction item from the perspective of objective interest is obtained according to the cosine similarity, the behavior intensity weight, and the item co-occurrence weight. The specific calculation is as follows:

[0145] score (k) (i t ,i )=sim (k) (i t ,i ) w rating (i t ) w co (i t ,i )

[0146] in, Representing an objective perspective of interest ; score (k) (i t ,i ) Representing an objective perspective of interest Below, historical interactive projects With candidate projects The final objective similarity between them; sim (k) (i t ,i ) Representing an objective perspective of interest Below, historical interactive projects With candidate projects Cosine similarity between them.

[0147] S320. For each objective interest perspective, based on the objective similarity, select several historical interaction items from the historical interaction items to form the objective interest sequence of the candidate items under the corresponding objective interest perspective.

[0148] For example, if there are K objective interest perspectives, then the candidate items have K objective interest sequences.

[0149] S400. Process the user comment information of the candidate items and the user comment information of the historical interaction items through a large language model, and select several items from the historical interaction items to form the subjective interest sequence of the candidate items;

[0150] In practical implementation, the Large Language Model (LLM) refers to a neural network model trained on massive amounts of text data, possessing powerful language generation, reasoning, and context understanding capabilities. Before use, the LLM can be pre-trained. During the training phase, the LLM undergoes domain adaptation using a large amount of user comment information as corpus, enabling fine-tuning to specifically enhance its user comment information analysis capabilities. The fine-tuned LLM can then analyze the input user comment information.

[0151] Specifically, when acquiring the user and their historical interaction items, user comment information on the historical interaction items can also be acquired. Similarly, based on candidate items, user comment information on the candidate items can be acquired.

[0152] In practical applications, the user comment information is input as text, one by one, into the large language model. By obtaining the contextual semantic representation of the user comment information, the analysis results are obtained. Based on the analysis results, several items are selected from the user's historical interaction items as an objective interest sequence of candidate items.

[0153] In one specific embodiment, step S400 may include the following steps:

[0154] A410. Construct input information by combining the preset prompt information with the user comment information of the candidate items and the user comment information of the historical interaction items, and input the input information into the large language model to obtain the corresponding scores, pre-event sentiment features and post-event sentiment features respectively.

[0155] In the specific implementation process, the prompt information refers to the structured input text constructed for the large language model, which is used to guide the large language model to perform structured parsing of user comment information. This allows the large language model to analyze the content described by the user comment information and obtain analysis results. Based on the analysis results, the rating, pre-event sentiment features, and post-event sentiment features corresponding to the user comment information can be obtained.

[0156] For example, a prompt message can guide a large language model to perform structured parsing of user review information and output a tuple containing multiple fields. The tuple may include, but is not limited to: specific product / service names appearing in the user review information, whether there is a predictable prior expectation, prior sentiment, post-event sentiment, and objectivity rating. The existence of a predictable prior expectation can be represented by 0 to indicate that there is no predictable prior expectation and 1 to indicate that there is a predictable prior expectation. The prior sentiment outputs a specific value under the condition that there is a predictable prior expectation, which can be represented by (prior sentiment polarity value [-1,1] × prior sentiment intensity value [0,1]). The post-event sentiment can be represented by (post-event sentiment polarity value [-1,1] × post-event sentiment intensity value [0,1]). The objectivity rating can be directly output as a rating value.

[0157] A420. Obtain the subjective features of the historical interaction items and the subjective features of the candidate items based on the ratings, pre-event sentiment features, and post-event sentiment features corresponding to the historical interaction items and the candidate items, respectively.

[0158] For example, the corresponding score can be the objective score in the tuple, the pre-emotional feature can be the product of the pre-emotional polarity value and the pre-emotional intensity value in the pre-emotional feature, and the post-emotional feature can be the product of the post-emotional polarity value and the post-emotional intensity value in the post-emotional feature.

[0159] Specifically, step A420 may include the following steps:

[0160] A421. Based on the ratings and post-event sentiment features corresponding to the historical interaction items, obtain the post-event sentiment scores of the historical interaction items; and based on the ratings and post-event sentiment features corresponding to the candidate items, obtain the post-event sentiment scores of the candidate items.

[0161] Specifically, for each historical interaction item and each candidate item, there is at least one user comment. Each user comment is input into the large language model and will yield a corresponding score and post-event sentiment features. The post-event sentiment score of the historical interaction item can be obtained based on the scores and post-event sentiment features of all user comment information corresponding to each historical interaction item. Similarly, the post-event sentiment score of the candidate item can be obtained based on the scores and post-event sentiment features of all user comment information corresponding to each candidate item.

[0162] Specifically, the formula for calculating the post-event sentiment score can be as follows:

[0163]

[0164] in, This refers to the project. It can be a candidate item or a historical interaction item; Indicates project Post-event emotional score For user comment information The rating; Indicates user comment information The corresponding post-event emotional polarity value, Indicates user comment information The corresponding post-event emotional intensity value; Indicates project A collection of user comments.

[0165] A422. Based on the ratings, pre-event sentiment features, and post-event sentiment features corresponding to the historical interaction items and the candidate items, respectively, obtain the experience expectation gap value of the historical interaction items and the experience expectation gap value of the candidate items;

[0166] Specifically, the formula for calculating the experience expectation gap value can be as follows:

[0167] =

[0168] in, Indicates project The gap between expected and unexpected experiences This indicates the existence of predictable, anticipated user comment information. For user comment information The rating, Indicates user comment information The corresponding post-event emotional polarity value, Indicates user comment information The corresponding post-event emotional intensity value, Indicates user comment information The corresponding pre-event emotional polarity value, Indicates user comment information The corresponding pre-event emotional intensity value; Indicates project All corresponding user comment information The set. Understandably, if the project... If there are no predictable pre-expected expectations for any of the user comments, then the experience expectation gap value can be calculated solely based on the post-expected sentiment score, i.e., solely based on... and This can be calculated.

[0169] Specifically, for each historical interaction item and each candidate item, there is at least one user review. Each user review is input into the large language model and will yield a corresponding rating and post-event sentiment features. The pre-event sentiment features only have specific values ​​when there are predictable pre-event expectations. Therefore, user review information with predictable pre-event expectations can be filtered out. Then, based on the rating, pre-event sentiment features, and post-event sentiment features of the user review information filtered for each historical interaction item, the experience expectation gap value of that historical interaction item is obtained. The experience expectation gap value of each candidate item is obtained in the same way.

[0170] A423. Based on the post-event emotional score and experience expectation gap value of the historical interaction items, construct the subjective features of the historical interaction items; and based on the post-event emotional score and experience expectation gap value of the candidate items, construct the subjective features of the candidate items.

[0171] Specifically, the post-event emotional score and the experience expectation gap value can be combined to form a vector as the subjective feature, i.e., the subjective feature. .

[0172] A430. Based on the subjective characteristics of the historical interaction items and the subjective characteristics of the candidate items, select several items from the historical interaction items to form the subjective interest sequence of the candidate items.

[0173] Specifically, an experience matching function can be preset to calculate the subjective matching degree between the subjective features of the candidate items and the subjective features of each historical interaction item. Based on the subjective matching degree, several historical interaction items with the highest subjective matching degree are selected from the historical interaction items to form the subjective interest sequence.

[0174] Example, subjective matching degree ( , ) The calculation formula is as follows:

[0175] ( , ) =

[0176] in, The experience matching function is used to characterize the consistency of the experience similarity or gap patterns between two items. Weights can be learned or set based on statistical information; For historical interactive projects Post-event emotional score; For historical interactive projects The gap between expected and unexpected experiences; Candidate projects Post-event emotional score Candidate projects The gap between expected and unexpected experiences.

[0177] Specifically, the study characterizes the evolution mechanism of user interests from the perspective of subjective experience feedback. By using the gap between post-event emotional scores and experience expectations, the study selects historical interaction items that are most relevant to the current candidate items in terms of experience, and constructs a subjective interest sequence that reflects the user's true satisfaction to make up for the psychological differences ignored by relying solely on objective behavior and attribute modeling.

[0178] Specifically, based on the subjective matching degree, several candidate items with the highest subjective matching degree are selected from the historical interaction items. The most relevant historical interaction items at the experiential level are ranked to form a subjective interest sequence for the candidate items. Specifically, the ranking can be based on the most recent interaction time between the historical interaction items and the user. The subjective interest sequence depicts the user's historical decision-making and feedback patterns that are most similar to the current candidate items in terms of subjective experience.

[0179] In another specific embodiment, step S400 may include the following steps:

[0180] B410. Construct input information by combining the preset prompt information with the user comment information of the candidate items and the user comment information of the historical interaction items, and input the input information into the large language model to obtain the preliminary evaluation categories corresponding to the candidate items and the historical interaction items, as well as the scores, pre-event sentiment features and post-event sentiment features corresponding to the preliminary evaluation categories.

[0181] For example, after the large language model performs structured parsing of user review information using prompts, the output tuples may include not only the specific product / service names appearing in the user review information, but also preliminary evaluation categories. Each preliminary evaluation category corresponds to: whether there is a predictable pre-purchase expectation, pre-purchase sentiment, post-purchase sentiment, and objectivity rating. The preliminary evaluation categories can be extracted from the user review information. For example, the taste of a product can be used as a preliminary evaluation category. Then, the analysis of the product taste as a preliminary evaluation category is performed to determine whether the user has pre-purchase expectations for the product taste in the user review information corresponding to this preliminary evaluation category. If so, the value of "whether there is a predictable pre-purchase expectation" is 1; otherwise, it is 0. Similarly, the pre-purchase sentiment, post-purchase sentiment, and objectivity rating corresponding to the product taste are output.

[0182] B420. The preliminary evaluation categories of the historical interaction items and the preliminary evaluation categories of the candidate items are clustered and merged based on semantic similarity to obtain standardized evaluation categories;

[0183] Specifically, the purpose of step B420 is to merge the preliminary evaluation categories from different user comment information that have different expressions but similar semantics, so as to provide a unified semantic framework for subsequent quantitative aggregation.

[0184] In practice, the preliminary evaluation categories of the candidate items and the historical interaction items can be submitted to a large language model, which then performs clustering and merging based on semantic similarity to obtain standardized evaluation categories. In the specific implementation process, the large language model performing the preliminary evaluation category aggregation can be the same as the large language model in S400, or a separate large language model can be constructed.

[0185] For example, if three preliminary evaluation categories, namely "coffee temperature", "drink hot or cold", and "latte hot or cold", are obtained from all the corresponding user comments for historical interaction items and candidate items, then these three preliminary evaluation categories are clustered and merged based on semantic similarity to obtain the standardized evaluation category "temperature".

[0186] B430. Based on the scores, pre-event sentiment features, and post-event sentiment features corresponding to each standardized evaluation category of the historical interaction project, obtain the subjective features of each standardized evaluation category of the historical interaction project; and based on the scores, pre-event sentiment features, and post-event sentiment features corresponding to each standardized evaluation category of the candidate project, obtain the subjective features of each standardized evaluation category of the candidate project.

[0187] Specifically, the steps in step B430 may include:

[0188] B431. Based on the scores and post-event sentiment features corresponding to each standardized evaluation category of the historical interaction project, obtain the post-event sentiment score of the historical interaction project under the standardized evaluation category; and based on the scores and post-event sentiment features corresponding to each standardized evaluation category of the candidate project, obtain the post-event sentiment score of the candidate project under the standardized evaluation category.

[0189] Specifically, the formula for calculating the ex-post sentiment score in standardized evaluation category c can be as follows:

[0190]

[0191] in, Indicates project The post-event sentiment score corresponding to standardized evaluation category c. For user comment information The score corresponding to standardized evaluation category c; Indicates user comment information Corresponding to the post-event sentiment polarity value of standardized evaluation category c, Indicates user comment information The post-event emotional intensity value corresponding to standardized evaluation category c.

[0192] B432. Based on the scores, pre-event emotional features, and post-event emotional features corresponding to each standardized evaluation category of the historical interaction project, obtain the experience expectation gap value for each standardized evaluation category of the historical interaction project; and based on the scores, pre-event emotional features, and post-event emotional features corresponding to each standardized evaluation category of the candidate project, obtain the experience expectation gap value for each standardized evaluation category of the candidate project.

[0193] Specifically, the formula for calculating the experience expectation gap value in standardized evaluation category c can be as follows:

[0194] =

[0195] in, Indicates project The experience expectation gap value under standardized evaluation category c. Indicates user comment information Corresponding to the post-event sentiment polarity value of standardized evaluation category c, Indicates user comment information Corresponding to the post-event affective intensity value of standardized evaluation category c, Indicates user comment information Corresponding to the pre-evaluation polarity value of standardized evaluation category c, Indicates user comment information Corresponding to the pre-emotional intensity value of standardized evaluation category c, For user comment information The score corresponding to standardized evaluation category c; Indicates project There exists a set of user comment information that can be predicted in advance.

[0196] Specifically, for each project (historical interaction project or candidate project), each of its standardized evaluation categories may correspond to several preliminary evaluation categories, and each preliminary evaluation category corresponds to a score, pre-event sentiment feature and post-event sentiment feature. Then, for the score, pre-event sentiment feature and post-event sentiment feature corresponding to each project under each standardized evaluation category, the scores, pre-event sentiment features and post-event sentiment features of all preliminary evaluation categories corresponding to that standardized evaluation category are weighted and aggregated to obtain the results.

[0197] B433. Based on the ex-post sentiment score and experience expectation gap value of each standardized evaluation category of the historical interaction project, construct the subjective features of each standardized evaluation category of the historical interaction project; and based on the ex-post sentiment score and experience expectation gap value of each standardized evaluation category of the candidate project, construct the subjective features of each standardized evaluation category of the candidate project.

[0198] Specifically, the post-event affective scores and experience expectation gap values ​​corresponding to each standardized evaluation category can be combined to form a vector as the subjective feature, i.e., the item. Subjective characteristics .

[0199] B440. Based on the subjective characteristics of each standardized evaluation category of the historical interaction project and the subjective characteristics of each standardized evaluation category of the candidate project, select several historical interaction projects from the historical interaction projects to form the subjective interest sequence of the candidate projects.

[0200] Specifically, an experience matching function can be preset to calculate the subjective matching degree between the subjective features corresponding to each candidate item and the subjective features corresponding to each historical interaction item for the same standardized evaluation category. Based on the subjective matching degree, several historical interaction items with the highest subjective matching degree are selected from the historical interaction items to form the subjective interest sequence.

[0201] Example, subjective matching degree ( ,i ) The calculation formula can also be as follows:

[0202] ( , ) =

[0203] Where C represents the set of standardized evaluation categories; The weights for standardizing evaluation category c can be learned or set based on statistical information; Represents historical interactive items The post-event sentiment score corresponding to standardized evaluation category c. Indicates candidate projects The post-event sentiment score corresponding to standardized evaluation category c. Represents historical interactive items The experience expectation gap value corresponding to standardized evaluation category c. Indicates candidate projects The experience expectation gap value corresponding to standard evaluation category c.

[0204] S500. The objective interest sequence and the subjective interest sequence are sequentially subjected to time-series coding, frequency domain decomposition, and inverse Fourier transform to obtain the corresponding time-domain representations.

[0205] Specifically, the specific steps of step S500 may include:

[0206] S510. Perform time-series coding on the objective interest sequence and the subjective interest sequence respectively to obtain the corresponding time-series coded signals;

[0207] Specifically, if there are K objective interest perspectives, then the candidate project has K objective interest sequences, plus one subjective interest sequence, resulting in a total of K+1 interest sequences.

[0208] Specifically, the objective interest sequence and the subjective interest sequence are respectively mapped into continuous time signals through a preset second time-series coding network, which serve as time-series coding signals. H u (k) Specifically, it can mean:

[0209] H u (k) =Encoder(S u (k) )

[0210] in, S u (k)The k-th interest sequence can be either an objective interest sequence or a subjective interest sequence; the Encoder represents the second temporal coding network. For example, the second temporal coding network can adopt a bidirectional gated recurrent unit (GRU) or other temporal modeling structure to characterize the temporal dependencies within the sequence.

[0211] S520. Perform frequency domain decomposition on the time-series coded signal to obtain frequency domain components of at least two preset frequency bands;

[0212] Specifically, step S520 includes the following steps:

[0213] S521. Perform a Fast Fourier Transform (FFT) on the time-series coded signal to obtain the frequency domain signal, that is, map the sequence from the time domain to the frequency domain space:

[0214] F u (k) =FFT(H u (k) )

[0215] in, F u (k) This represents the frequency domain signal corresponding to the k-th interest sequence; in the frequency domain representation, different frequency components correspond to the changing patterns of user interests at different time scales.

[0216] S522. The frequency domain signal is filtered by at least two filters of different frequency bands to obtain frequency domain components of at least two preset frequency bands.

[0217] In practice, a set of learnable frequency domain filters is introduced to separate the spectrum of the frequency domain signal in order to model the interest at different time scales.

[0218] For example, when the frequency bands are set to three, filtering is achieved through a low-pass filter, a band-pass filter, and a high-pass filter, respectively:

[0219]

[0220] Here, the Hadamard product (⊙) represents digit-wise multiplication. , , These represent low-pass, band-pass, and high-pass filters, respectively; the filter parameters are automatically learned through end-to-end training, eliminating the need for manual frequency boundary setting. F u (k ,l) ,F u (k,m) , F u (k,h) This represents the frequency domain components obtained by passing the frequency domain signal corresponding to the k-th interest sequence through a low-pass filter, a band-pass filter, and a high-pass filter, respectively.

[0221] S530. Perform an inverse Fourier transform on the frequency domain components to obtain the corresponding time domain representation.

[0222] Performing an inverse Fourier transform (IFFT) on each frequency domain component reconstructs it back into a time domain representation. This yields time domain representations corresponding to different time scales for both subjective and objective interest sequences.

[0223] If each interest sequence has three multi-scale time-domain representations (low frequency, mid frequency, and high frequency), then K+1 interest sequences will have a total of 3×(K+1) time-domain representations.

[0224] This approach uses frequency domain decomposition to break down both objective and subjective interest sequences into interest components at multiple time scales: low-frequency components characterize relatively stable long-term user preferences; mid-frequency components describe recurring interest patterns within a certain time range; and high-frequency components capture short-term interest fluctuations generated in recent user interactions. This method effectively alleviates the problem of interference between long-term and short-term signals in single-time-domain modeling, providing structured, multi-scale interest inputs for subsequent interest fusion and recommendation prediction.

[0225] S600. The temporal representations corresponding to the objective interest sequence and the subjective interest sequence are fused to obtain the corresponding comprehensive interest representation;

[0226] Specifically, to integrate information from each interest sequence at different time scales, learnable scale weights are introduced to fuse all temporal representations corresponding to the same interest sequence, resulting in a comprehensive interest representation for the same interest sequence.

[0227] For example, the time-domain representations of low-frequency, mid-frequency, and high-frequency corresponding to the same interest subsequence are fused. This fusion process can be achieved by introducing an attention mechanism to assign learnable weights to the time-domain representations at different time scales and then summing them up to obtain the comprehensive interest representation of the interest sequence.

[0228] Subsequently, at the cross-subsequence level, the comprehensive interest representation of all subsequences is input into the attention fusion network. By learning the importance weights of different subsequences in the current recommendation decision, the network is weighted and aggregated to form the final overall user interest representation.

[0229] S700. Based on the comprehensive interest representation and the objective attribute features of the candidate items, perform preference prediction and output the item recommendation results for the user.

[0230] Specifically, step S700 includes the following steps:

[0231] S710. The comprehensive interest representations corresponding to the objective interest sequence and the subjective interest sequence are weighted and fused to obtain the fused interest representation;

[0232] Specifically, to integrate interest information from objective and subjective interest sequences, a sub-sequence-level fusion mechanism is introduced to perform weighted aggregation on the comprehensive interest representation:

[0233]

[0234] These are learnable parameters; This represents the k-th interest sequence, which can be either an objective interest sequence or a subjective interest sequence. This represents the comprehensive interest representation of the k-th interest sequence. This represents the fused interest representation. Indicates user .

[0235] S720. Based on the fused interest representation and the objective attribute features of the candidate items, calculate the user's preference score for the candidate items using vector matching. :

[0236]

[0237] in, Indicates candidate projects Its objective attributes and characteristics.

[0238] S730. Sort the candidate items according to the preference scores to obtain a candidate item sequence, and extract and output the user's item recommendation results based on the candidate item sequence.

[0239] Specifically, each candidate item receives a corresponding preference score. The candidate items can then be sorted according to the magnitude of the preference scores to obtain a candidate item sequence. A preset number of candidate items with the highest preference scores are then selected from the candidate item sequence as the user's item recommendation results and recommended to the user.

[0240] like Figure 5 As shown in the illustration, this application also provides a user multi-interest recommendation system based on a large model and frequency domain decomposition. Optionally, the system includes:

[0241] The first feature acquisition module 811 is used to acquire the objective attribute features of the user's historical interaction items based on preset objective attributes of the items;

[0242] In this embodiment, the first feature acquisition module 811 can be used to perform... Figure 2 For a detailed description of the first feature acquisition module 811, please refer to the description of step S100 shown.

[0243] The second feature acquisition module 812 is used to acquire the objective attribute features of the candidate projects based on the objective attributes of the projects.

[0244] In this embodiment, the second feature acquisition module 812 can be used to perform... Figure 2 For a detailed description of step S200 shown, and the second feature acquisition module 812, please refer to the description of step S200.

[0245] The objective interest sequence acquisition module 813 is used to select several historical interaction items from the historical interaction items based on the objective attribute characteristics of the candidate items and the objective attribute characteristics of the historical interaction items, to form the objective interest sequence of the candidate items;

[0246] In this embodiment, the objective interest sequence acquisition module 813 can be used to perform... Figure 2 For a detailed description of the objective interest sequence acquisition module 813, see step S300 shown below. For a detailed description of step S300, please refer to the description of step S300.

[0247] The subjective interest sequence acquisition module 814 is used to process the user comment information of the candidate items and the user comment information of the historical interaction items through a large language model, and select several items from the historical interaction items to form the subjective interest sequence of the candidate items;

[0248] In this embodiment, the subjective interest sequence acquisition module 814 can be used to perform... Figure 2 For a detailed description of the subjective interest sequence acquisition module 814, see step S400 shown below. For a detailed description of step S400, please refer to the description of step S400.

[0249] The time-domain representation acquisition module 815 is used to sequentially encode, decompose in the frequency domain, and perform inverse Fourier transform on the objective interest sequence and the subjective interest sequence to obtain the corresponding time-domain representations.

[0250] In this embodiment, the time-domain representation acquisition module 815 can be used to perform... Figure 2 For a detailed description of the time-domain representation acquisition module 815, please refer to the description of step S500 shown.

[0251] The comprehensive interest representation acquisition module 816 is used to fuse the time-domain representations corresponding to the objective interest sequence and the subjective interest sequence respectively to obtain the corresponding comprehensive interest representation;

[0252] In this embodiment, the comprehensive interest representation acquisition module 816 can be used to perform... Figure 2 For a detailed description of the integrated interest representation acquisition module 816, please refer to the description of step S600 shown.

[0253] The recommendation module 817 is used to predict preferences based on the comprehensive interest representation and the objective attribute features of the candidate items, and output the item recommendation results for the user.

[0254] In this embodiment, the recommendation module 817 can be used to perform... Figure 2 For a detailed description of the recommendation module 817, please refer to the description of step S700 shown.

[0255] This application also provides an electronic device, the structure of which is as follows: Figure 6 As shown, the electronic device includes a memory 611, a processor 612, a communication module 613, and an input / output interface 614, etc. Optionally, the memory 611, the processor 612, the communication module 613, and the input / output interface 614 can be connected and communicate with each other through a bus 615.

[0256] The memory 611 is used to store one or more computer programs and to transfer the code of the computer programs to the processor 612; when the one or more computer programs are executed by the processor 612, a user multi-interest recommendation method based on a large model and frequency domain decomposition in this application embodiment is implemented.

[0257] Optionally, the electronic device can be connected to a network via communication module 613 to communicate with other devices, such as terminals or servers, to achieve data interaction. The electronic device can be various forms of digital computers, exemplarily such as desktop computers, servers, workbenches, mainframes, or other types of computers. The electronic device can also be various forms of mobile terminals, exemplarily such as smartphones, tablets, wearable devices (such as helmets, glasses, watches, etc.), and other similar mobile terminals.

[0258] Optionally, the electronic device can connect to required input / output devices, such as a keyboard or display device, via the input / output interface 614. The electronic device itself may have a display device, and other display devices can also be connected externally via the input / output interface 614. Optionally, a storage device, such as a hard disk, can also be connected via the input / output interface 614 to store data from the electronic device, read data from the storage device, or store data from the storage device in the memory 611. It is understood that the input / output interface 614 can be a wired interface or a wireless interface. Depending on the actual application scenario, the device connected to the input / output interface 614 can be a component of the electronic device or an external device connected to the electronic device when needed.

[0259] Optionally, the memory 611 may be a volatile memory and / or a non-volatile memory. The volatile memory may be a random access memory, etc., and the non-volatile memory may be a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, or a flash memory, etc.

[0260] Optionally, the computer program stored in the memory 611 can be divided into one or more modules, which are stored in the memory 611 and executed by the processor 612 to perform the method provided in this embodiment. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the electronic device.

[0261] Optionally, processor 612 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 612 include, but are not limited to, central processing units, graphics processing units, digital signal processors, various special-purpose artificial intelligence computing chips, various processors running machine learning model algorithms, and can also be any suitable controller, microcontroller, processor, etc. Processor 612 executes the various methods and processes of this embodiment, exemplarily, such as a user multi-interest recommendation method based on large model and frequency domain decomposition according to an embodiment of this application.

[0262] Optionally, the bus 615 may include a path for transmitting information. Depending on its function, the bus 615 may be divided into an address bus, a data bus, a control bus, etc.

[0263] In an optional implementation, this application embodiment also provides a computer storage medium storing a computer program thereon, which, when executed by a computer, enables the computer to perform the methods described in the above method embodiments. Part or all of the computer program can be loaded and / or installed on the memory 611 of an electronic device. When the computer program is executed by the processor 612, it can perform one or more steps of a user multi-interest recommendation method based on a large model and frequency domain decomposition according to an embodiment of this application.

[0264] Optionally, the computer-readable storage medium may be a random access memory, a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, etc.

[0265] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A user multi-interest recommendation method based on a large model and frequency domain decomposition, characterized in that, The method includes: Preset the objective attribute categories of the project, and preset the trainable embedding lookup table corresponding to each objective attribute category of the project; Obtain the objective attributes of historical interaction items, and based on the objective attributes of the historical interaction items, find the corresponding low-dimensional dense vectors from the corresponding embedding lookup tables respectively. The low-dimensional dense vectors of the historical interaction items are sequentially concatenated, dimension aligned, and dimension mapped to obtain the first objective attribute representation vector. Several objective interest perspectives are preset, and each objective interest perspective corresponds to a learnable linear projection matrix; Based on the linear projection matrix corresponding to the objective interest perspective, the first objective attribute representation vector is feature-mapped to obtain the corresponding initial item representation; The initial item representations of the historical interaction items under the corresponding objective interest perspective are sorted according to the interaction time between the corresponding historical interaction items and the user to obtain the initial item representation sequence. The initial item representation sequence is input into a temporal coding network for processing, and the temporally enhanced item representation sequence is output. The time-enhanced item representation sequence of the historical interaction item under all objective interest perspectives is used as the objective attribute feature of the historical interaction item. Each time-enhanced item representation sequence includes the item representation corresponding to each historical interaction item under each objective interest perspective. Based on the objective attributes of the project, obtain the objective attribute characteristics of the candidate project; Based on the objective attribute characteristics of the candidate projects and the objective attribute characteristics of the historical interaction projects, a number of historical interaction projects are selected from the historical interaction projects to form an objective interest sequence of the candidate projects; The user comment information of the candidate items and the user comment information of the historical interaction items are processed by a large language model, and several items are selected from the historical interaction items to form the subjective interest sequence of the candidate items; The objective interest sequence and the subjective interest sequence are sequentially subjected to time-series coding, frequency domain decomposition, and inverse Fourier transform to obtain their corresponding time-domain representations. The temporal representations corresponding to the objective interest sequence and the subjective interest sequence are fused to obtain the corresponding comprehensive interest representation; Based on the comprehensive interest representation and the objective attribute features of the candidate items, preference prediction is performed, and the item recommendation results for the user are output. The process of processing user comment information of the candidate items and user comment information of the historical interaction items through a large language model, and selecting several items from the historical interaction items to form a subjective interest sequence of the candidate items, includes: constructing input information by combining preset prompt information with user comment information of the candidate items and user comment information of the historical interaction items, inputting the input information into the large language model, and obtaining corresponding ratings, pre-event sentiment features and post-event sentiment features respectively; Based on the historical interaction items, the ratings corresponding to the candidate items, and the post-event sentiment features, respectively, the post-event sentiment scores of the historical interaction items and the candidate items are obtained. Based on the historical interaction items, the corresponding ratings, pre-event sentiment features, and post-event sentiment features, respectively, obtain the experience expectation gap value of the historical interaction items and the experience expectation gap value of the candidate items; Based on the post-event emotional scores and experience expectation gap values ​​of the historical interaction items, subjective features of the historical interaction items are constructed, and subjective features of the candidate items are constructed based on the post-event emotional scores and experience expectation gap values ​​of the candidate items. According to the subjective features of the historical interaction items and the subjective features of the candidate items, several items are selected from the historical interaction items to form the subjective interest sequence of the candidate items.

2. The method of claim 1, wherein, The step of obtaining the objective attribute features of candidate projects based on the objective attributes of the projects includes: Obtain the objective attributes of the candidate projects; Based on the objective attributes of the candidate projects, the corresponding low-dimensional dense vectors are retrieved from the corresponding embedding lookup tables. The low-dimensional dense vectors of the candidate projects are sequentially concatenated, dimension aligned, and dimension mapped to obtain the second objective attribute representation vector. Based on the linear projection matrix corresponding to the objective interest perspective, the second objective attribute representation vector is feature-mapped to obtain the corresponding item representation; The project representation of the candidate project under all objective interest perspectives is taken as the objective attribute feature of the candidate project.

3. The method of claim 2, wherein, The step of selecting several historical interaction items from the historical interaction items to form an objective interest sequence of the candidate items based on the objective attribute characteristics of the candidate items and the objective attribute characteristics of the historical interaction items includes: For each objective interest perspective, based on the corresponding project representation, the objective similarity between the candidate project and the historical interaction project is obtained; For each objective interest perspective, based on the objective similarity, several historical interaction items are selected from the historical interaction items to form the objective interest sequence of the candidate items under the corresponding objective interest perspective.

4. The method according to any one of claims 1 to 3, characterized in that, The process involves using a large language model to process user comment information on the candidate items and user comment information on historical interaction items, selecting several items from the historical interaction items to form a subjective interest sequence for the candidate items, or including: The preset prompts are combined with the user comments of the candidate items and the user comments of the historical interaction items to construct input information, which is then input into the large language model to obtain the preliminary evaluation categories corresponding to the candidate items and the historical interaction items, as well as the scores, pre-event sentiment features and post-event sentiment features corresponding to the preliminary evaluation categories. The preliminary evaluation categories of the historical interaction items and the preliminary evaluation categories of the candidate items are clustered and merged based on semantic similarity to obtain standardized evaluation categories; Based on the scores, pre-event sentiment features, and post-event sentiment features corresponding to each standardized evaluation category of the historical interaction project, the subjective features of each standardized evaluation category of the historical interaction project are obtained; and based on the scores, pre-event sentiment features, and post-event sentiment features corresponding to each standardized evaluation category of the candidate project, the subjective features of each standardized evaluation category of the candidate project are obtained. Based on the subjective characteristics of each standardized evaluation category of the historical interaction projects and the subjective characteristics of each standardized evaluation category of the candidate projects, several projects are selected from the historical interaction projects to form the subjective interest sequence of the candidate projects.

5. The method of claim 4, wherein, The step of obtaining the subjective characteristics of each standardized evaluation category of the historical interaction project based on the scores, pre-event sentiment characteristics, and post-event sentiment characteristics corresponding to each standardized evaluation category of the historical interaction project includes: Based on the scores and post-event sentiment features corresponding to each standardized evaluation category of the historical interaction project, the post-event sentiment score of each standardized evaluation category of the historical interaction project is obtained. Based on the scores, pre-event emotional features, and post-event emotional features corresponding to each standardized evaluation category of the historical interaction project, obtain the experience expectation gap value for each standardized evaluation category of the historical interaction project. Based on the ex-post emotional score and experience expectation gap for each standardized evaluation category of the historical interaction project, subjective features for each standardized evaluation category of the historical interaction project are constructed. And / or, based on the scores, pre-event sentiment features, and post-event sentiment features corresponding to each standardized evaluation category of the candidate projects, obtain the subjective features of each standardized evaluation category of the candidate projects, including: Based on the scores and post-event sentiment features corresponding to each standardized evaluation category of the candidate projects, the post-event sentiment score for each standardized evaluation category of the candidate projects is obtained. Based on the scores, pre-event emotional features, and post-event emotional features corresponding to each standardized evaluation category of the candidate projects, obtain the experience expectation gap value for each standardized evaluation category of the candidate projects. Based on the ex-post affective score and experience expectation gap for each standardized evaluation category of the candidate projects, subjective features for each standardized evaluation category of the candidate projects are constructed.

6. The method of claim 1, wherein, The step of selecting several items from the historical interaction items to form a subjective interest sequence of the candidate items based on the subjective characteristics of the historical interaction items and the subjective characteristics of the candidate items includes: Based on a preset matching function, the subjective matching degree between the subjective features of the candidate items and the subjective features of each historical interaction item is calculated. Based on the subjective matching degree, several items are selected from the historical interaction item sequence to form the subjective interest sequence of the candidate items.

7. The method according to any one of claims 1 to 3, characterized in that, The step of sequentially performing time-domain encoding, frequency-domain decomposition, and inverse Fourier transform on the objective interest sequence and the subjective interest sequence to obtain the corresponding time-domain representation includes: The objective interest sequence and the subjective interest sequence are respectively time-series encoded to obtain corresponding time-series encoded signals; The time-coded signal is decomposed in the frequency domain to obtain frequency domain components of at least two preset frequency bands; The frequency domain components are subjected to inverse Fourier transform to obtain the corresponding time domain representation.

8. A user multi-interest recommendation system based on a large model and frequency domain decomposition, characterized in that, The system is used to implement the method according to any one of claims 1-7, the system comprising: The first feature acquisition module is used to acquire the objective attribute features of the user's historical interaction items based on preset objective attributes of the items; The second feature acquisition module is used to acquire the objective attribute features of the candidate projects based on the objective attributes of the projects. An objective interest sequence acquisition module is used to select several historical interaction items from the historical interaction items based on the objective attribute characteristics of the candidate items and the objective attribute characteristics of the historical interaction items, thereby forming an objective interest sequence of the candidate items; The subjective interest sequence acquisition module is used to process the user comment information of the candidate items and the user comment information of the historical interaction items through a large language model, and select several items from the historical interaction items to form the subjective interest sequence of the candidate items; The time-domain representation acquisition module is used to sequentially encode, decompose in the frequency domain, and perform inverse Fourier transform on the objective interest sequence and the subjective interest sequence to obtain the corresponding time-domain representations. The comprehensive interest representation acquisition module is used to fuse the time-domain representations corresponding to the objective interest sequence and the subjective interest sequence respectively to obtain the corresponding comprehensive interest representation; The recommendation module is used to predict preferences based on the comprehensive interest representation and the objective attribute features of the candidate items, and output the item recommendation results for the user.

Images