A trusted recommendation model based on a two-dimension constraint large language model

Abstract

The application provides a trusted recommendation model based on a double-dimension constraint large language model, and comprises the following steps: S1, a negative description module and a similar behavior generation module: the negative description module and the similar behavior generation module provide a multi-dimensional balanced perspective for subsequent reasoning; S2, a knowledge consistency verification module: the knowledge consistency verification module makes the features of the input model highly reliable through a lightweight fact consistency detection process; S3, a thinking chain reasoning module: the thinking chain reasoning part performs task decomposition, and the complete reasoning chain from user preference analysis to final score generation; S4, a fusion module: the fusion module adjusts the adaptive weight, fuses the statistical results of the traditional recommendation channel and the reasoning results of the large language model channel, and generates the final recommendation result, can solve the illusion and deviation problem of the large language model, and enhance the recommendation performance by integrating the deep reasoning large language model.

Description

Technical Field

[0001] This invention relates to the technical field of recommender systems, and in particular to the technical field of a reliable recommender model based on a two-dimensional constrained large language model. Background Technology

[0002] Recommender systems have become a core component of modern digital platforms, playing an irreplaceable role in e-commerce, social media, online streaming, and other fields. However, traditional recommendation models have limitations when dealing with complex semantic understanding.

[0003] Chinese patent application number CN202211284644.9 discloses a method for a sequence recommendation model, including obtaining historical behavior sequences of related object recommendation business; generating original interest representations based on historical objects contained in the historical behavior sequences; generating original interest representations based on each historical object; calculating object similarity between each historical object and a benchmark preference object based on the attention influence parameters of each historical object in at least one dimension; determining target historical objects in the historical behavior sequences based on object similarity, and generating counterfactual interest representations based on target historical objects; generating optimized parameter pairs based on the original interest representations, the benchmark preference representations corresponding to the benchmark preference objects, and the counterfactual interest representations, and training a sequence recommendation model for related object recommendation business based on the optimized parameter pairs. However, this method still has shortcomings, namely, it does not explicitly model users' long-term and short-term preferences, cannot accurately generate personalized recommendations, does not fully utilize the application of item relationships in sequence recommendation, and its recommendation performance is not ideal when facing cold start problems or sparse data.

[0004] Chinese patent application number CN201810714988.6 discloses a method, apparatus, storage medium, and electronic device for recommending insurance information. This method obtains the identifier of a target user and, based on that identifier, retrieves the target user's insurance family information from a pre-constructed graph database. Based on this insurance family information, it recommends insurance information to objects within the insurance family. This achieves the identification of insurance families based on the constructed graph database, and the identification of the entire insurance family based on any object within the family, increasing the convenience and efficiency of insurance information recommendation. Furthermore, recommending insurance information to objects within an insurance family based on its information enables efficient, accurate, and personalized insurance service recommendations. However, this method still has shortcomings: the application of big data and recommendation technologies is not deep enough, the model's generalization ability is weak, the accuracy of the recommendation results is insufficient, and the depth, granularity, and objectivity of the analysis of user needs are inadequate, failing to deeply explore users' potential protection needs.

[0005] Chinese Patent Application No. CN201410093793.6 discloses an online teaching resource recommendation system and method. The system includes: a data construction module, which constructs teacher behavior data, teacher model data, course model data, and resource model data; an offline data processing module, which initializes and adjusts the course model data and resource model data, infers teacher identity using teacher behavior data, calculates the correlation between resources based on teacher behavior data, calculates the similarity between resources based on resource model data, and calculates the correlation between courses and resources based on resource model data and course model data; and an online recommendation module, which recommends resources online using the correlation between resources, the similarity between resources, the correlation between courses and resources, and dynamic descriptions of teachers, and also recommends resources based on teachers' feedback. The system provides feedback and recommends resource tags, and transmits teacher behavior data to the teacher behavior data in the data construction module through UI interaction. It employs a content- and tag-based recommendation method. New content can have initial tags extracted based on its features, and then the tags are filtered and expanded using teacher behavior records, reducing the inaccuracy of initial keyword extraction in the content recommendation method. This method avoids the impact of cold starts on the recommendation system. However, this method still has shortcomings. It cannot achieve smooth release of new content. Once initial tags are extracted for new content, the system will push this resource to all user interfaces involving those tags. If the initial tags are inaccurate or the content quality is problematic, it may maximize the negative impact of the content, which is detrimental to risk control.

[0006] Large language models, with their powerful contextual modeling and natural language understanding capabilities, have opened up new avenues for overcoming the aforementioned bottlenecks. Applying large language models to recommender systems involves designing prompts to stimulate their deep reasoning capabilities, generating personalized recommendation lists, enhancing semantic information using large language models, employing a knowledge-aware recommendation framework enhanced by large language models, designing a semantic user intent modeling module, aggregating multiple intent signals into a comprehensive and semantically enhanced intent embedding using large language models, analyzing user interaction sequences using masked language modeling techniques to predict future user behavior, optimizing graph-based recommendation denoising processes using large language models, and demonstrating the application potential of large language models in recommender systems through semantic knowledge-enhanced representation learning. However, directly applying large language models to rating prediction tasks faces the following challenges: First, large language models suffer from the illusion of fact problem: the model may generate seemingly reasonable but factually inaccurate answers; second, large language models have an inherent systematic rating bias: due to the bias in the distribution of positive and negative training samples, the ratings generated by large language models tend to concentrate in the high-score range, making it difficult to guarantee the reliability of the prediction results. Summary of the Invention

[0007] The purpose of this invention is to solve the problems in the prior art and propose a reliable recommendation model based on a two-dimensional constrained large language model, which can solve the illusion and bias problems of large language models and enhance recommendation performance by integrating a deep reasoning large language model.

[0008] To achieve the above objectives, this invention proposes a reliable recommendation model based on a two-dimensional constrained large language model, comprising the following steps:

[0009] S1. Negative description module and similar behavior generation module: The negative description module and similar behavior generation module provide a multi-dimensional balanced perspective for subsequent reasoning;

[0010] S2. Knowledge Consistency Verification Module: The knowledge consistency verification module uses a lightweight fact consistency detection process to ensure that the features of the input model have high reliability.

[0011] S3, Thinking Chain Reasoning Module: The thinking chain reasoning part decomposes the task, from user preference analysis to the generation of the final score, forming a complete reasoning chain;

[0012] S4. Fusion Module: The fusion module integrates the statistical results of the traditional recommendation channel with the inference results of the large language model channel through adaptive weight adjustment to generate the final recommendation result.

[0013] Preferably, in step S1, the negative description module constructs a negative description of the project: it constructs a negative description from real user data, using user information with low ratings for the current project and negative information carried by the project itself as information sources; it uses the Deepseek-Distil-32B model, taking the project's negative information and low-rated user information as input, to generate the description.

[0014] Preferably, in step S1, the similar behavior generation module constructs user similar behavior: it constructs a collaborative filtering signal extraction based on a predefined type similarity matrix, identifies user behavior patterns with similar preferences by utilizing the semantic relevance of item types, defines a set covering multiple main types, defines a semantic similarity matrix by calculating the cosine similarity between their semantic vector representations, sets type sets for the current item and the user's historical items respectively, uses a maximum matching strategy for similarity calculation, selects the top few historical items most similar to the current item type based on the calculated similarity, and integrates the user's corresponding interaction behavior and item metadata information to form an input representation of a collaborative filtering signal.

[0015] Preferably, in step S2, the knowledge consistency verification module performs knowledge consistency verification: a multi-level verification framework based on semantic similarity is designed; for the target item, feature vectors of its key attributes are extracted from the knowledge base, the corresponding features generated by the large language model are converted into vectors, and the similarity between the two is calculated. If the similarity value exceeds a set threshold, it is considered a successful match; if the match fails, the content in the knowledge base is referenced; firstly, the text generated by the original large language model and the standard knowledge base text are preprocessed, including text cleaning, word segmentation and standardization. Then, the corresponding vector representation is obtained through TF-IDF vectorization. The similarity calculation uses cosine similarity measurement, and finally reliable knowledge is generated for use as input for subsequent large language models and traditional models.

[0016] Preferably, in step S3, the thought chain reasoning module performs structured thought chain reasoning: the large language model performs structured thought chain reasoning based on enhanced context; it clearly lists the potential positive and negative factors of the item, and makes a comprehensive judgment based on this, decomposing the scoring task into multiple reasoning steps, each with a clear goal and output: in the preference analysis stage, the model infers the user's interest preference pattern based on the user's historical behavior, and at the same time combines the negative description of the item to determine the user's final preference; after determining the preference, the model then uses the scoring range specified by the corresponding preference, the basic characteristics of the item, and the scoring distribution of similar types of items to determine the scoring interval and give the final score.

[0017] Preferably, in step S3, during the structured thinking chain reasoning process, the evaluation criteria are constrained, and the model considers both positive and negative factors during the evaluation process, making a comprehensive trade-off, and finally generating results in a structured form.

[0018] Preferably, in step S4, the fusion module extracts the score of the large language model from the structured output result in step S3, and obtains the target result required for the score prediction task by directional parsing of the content of the corresponding score field in the structured data. The final prediction result is then linearly fused with the output of a traditional recommendation model that is independent of the model to generate the final prediction.

[0019] Preferably, in step S3, the thought chain reasoning module performs context-guided prompt word engineering: constructing a multi-step thought chain reasoning process, guiding the model to conduct progressive deep thinking; the thought chain structure is designed with five distinct stages, each stage building upon the reasoning result of the previous stage, forming a complete logical loop:

[0020] (1) Given role and task: The prompt words set a clear professional background and task goal for the model; the model is positioned as a recommendation system with professional knowledge in the project domain, and makes a comprehensive judgment based on the provided multi-dimensional information;

[0021] (2) Given user information: Construct a comprehensive and balanced contextual information; enable the large language model to simultaneously access positive and negative evidence, individual history and group behavior, and collaborative signals as the data basis for its reasoning;

[0022] (3) Determine user preferences: The prompt words guide the model to deeply analyze the user's historical behavior data. From this data, the model identifies the user's interests, aesthetic preferences and rating habits, forming a comprehensive understanding of the user's characteristics;

[0023] (4) Combine collaborative signals to determine the score: guide the model to deeply associate user preferences with project characteristics; the model comprehensively considers the project attributes after knowledge base verification, the behavior patterns of similar users, and the advantages and disadvantages of the project itself. At the same time, the model adopts dialectical thinking to balance positive and negative factors and avoids excessive optimism or pessimism.

[0024] (5) Structured output: The complex reasoning process is transformed into a clear and readable result through format specifications, and the prompt words clearly specify the specific format and requirements of the output.

[0025] The beneficial effects of this invention are as follows: This invention improves the reliability of large language models from both input and reasoning perspectives. First, it utilizes an external knowledge base for knowledge verification, suppressing generative illusions at the source. Next, it employs context-guided thought chains for reasoning and balances positive and negative knowledge before final scoring, effectively neutralizing biases. Furthermore, it enhances recommendation performance by integrating a deep reasoning model. Firstly, to prevent illusions from arising at the source, this invention proposes a dual-constraint recommendation framework of context guidance and knowledge verification, abbreviated as TrustRec. This framework performs knowledge verification, utilizing an external knowledge base to fact-check key information. Secondly, to ensure... This framework effectively neutralizes inherent biases and provides clear explanations. During the reasoning phase, it employs a context-guided thought process, using verified knowledge and collaborative signals as the basis for reasoning. It also forces the model to perform structured analysis that includes both positive and negative trade-offs. This framework can fully leverage the strengths of two types of models: traditional recommendation models excel at understanding ID features and collaborative information, while large language-based models stand out in text information understanding and reasoning. Through the synergy of knowledge verification and context-guided reasoning, it generates credible recommendations, uses a knowledge base to combat generative illusions, and uses a multi-stage reasoning process with positive and negative trade-offs to neutralize biases. This provides a solid and reliable implementation path for a collaborative paradigm.

[0026] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0027] Figure 1This is a framework diagram of the TrustRec model, a trusted recommendation model based on a two-dimensional constrained large language model, according to the present invention.

[0028] Figure 2 This is a structured thought chain prompt template diagram for a reliable recommendation model based on a two-dimensional constrained large language model, as proposed in this invention.

[0029] Figure 3 This is a graph showing the impact of different numbers of neighbors on the recommendation performance of a reliable recommendation model based on a two-dimensional constrained large language model according to the present invention.

[0030] Figure 4 This is a graph showing the impact of different thresholds on the recommendation performance of a reliable recommendation model based on a two-dimensional constrained large language model according to the present invention.

[0031] Figure 5 This is a diagram showing the impact of different fusion weights on recommendation performance of a trusted recommendation model based on a two-dimensional constrained large language model according to the present invention. Detailed Implementation

[0032] This invention proposes a trustworthy recommendation model based on a dual-dimensional constrained large language model, including the overall architecture and implementation details of a dual-constraint model (TrustRec) that combines contextual guidance and knowledge verification. This framework consists of five core modules, such as... Figure 1 As shown, firstly, the negative description module and the similar behavior generation module provide a multi-dimensional balanced perspective for subsequent reasoning. Secondly, the knowledge consistency verification module ensures the high reliability of the features of the input model through a lightweight fact consistency detection process. Next, the thought chain reasoning part decomposes the task to realize a complete reasoning chain from user preference analysis to final score generation. Finally, the fusion module effectively integrates the statistical results of the traditional recommendation channel with the reasoning results of the large language model channel through adaptive weight adjustment to generate the final recommendation result.

[0033] 1.1 Task Definition

[0034] The core of the rating prediction task is to predict the strength of a target user's preference for a specific item. This task can be formally defined as: given a set of users and movie collection Learn a mapping function This makes it possible for any user And movies ,have ,in It is the predicted score, and its value range is... ;

[0035] This invention is an improvement on the Lamar baseline model, which relies on an integration strategy of traditional models and large language models. The large language model recommendation module uses a diverse set of cue words to generate predictions and obtains the results through an aggregation function. This prediction is then independent of the output of a traditional recommendation model. Perform linear fusion to generate the final prediction. This method has the following limitations: First, the content generated by the large language model contains factual illusions, which impairs the reliability of reasoning; second, the large language model exhibits systematic bias, with its predictions tending to be high-scoring and lacking a rigorous reasoning process, resulting in insufficient prediction ability for low-scoring movies; third, traditional recommendation models have difficulty effectively utilizing rich semantic information. Experiments show that directly introducing rich semantic features into traditional models does not bring significant performance improvement, reflecting their inherent limitations in deep semantic understanding and fusion.

[0036] 1.2 Context Construction for Knowledge Enhancement

[0037] In the context building stage of knowledge enhancement, this invention provides comprehensive, accurate and balanced input information for subsequent large language model inference through the system integration of the following three key components, and provides reliable knowledge for traditional models;

[0038] 1.2.1 Construction of Negative Descriptions for the Project

[0039] To overcome the biases of large language models, the strategy of this invention is to construct negative descriptions from real user data. The information sources include two aspects: firstly, user information indicating low ratings for the current project; and secondly, negative information inherent in the project itself (for example, in the case of a movie project, negative information might include a dragging plot and a chaotic structure). Specifically, the Deepseek-Distil-32B model, with its greater knowledge base, is used as input to generate descriptions, taking the project's negative information and low-rated user information as input.

[0040] ;

[0041] in Indicates model parameters, For project information, Constructing negative descriptions for user information with low interaction scores can not only identify obvious negative features, but also discover significant negative information that is ignored in positive descriptions, providing an important basis for subsequent objective reasoning.

[0042] 1.2.2 Constructing User Similar Behaviors

[0043] When constructing user similarity behavior contexts, this invention builds a collaborative filtering signal extraction method based on a predefined type similarity matrix. The core idea of ​​this method is to use the semantic relevance of item types to identify user behavior patterns with similar preferences, that is, to find the nearest neighbor items corresponding to the current item. This invention defines a scope covering... Collections of major types A semantic similarity matrix is ​​defined by calculating the cosine similarity between their semantic vector representations. For the current project With user history projects Let their type sets be respectively and The similarity calculation uses a maximum matching strategy, which can effectively handle situations where the number of types is asymmetrical.

[0044] ;

[0045] Based on the calculated similarity, the top K historical neighbors most similar to the current project type are selected. These are then integrated with user interaction behavior and project metadata information to form the input representation of a collaborative filtering signal. :

[0046] ;

[0047] in, Indicates project information, This design, which represents user interaction information, provides the model with personalized collaborative filtering signals;

[0048] 1.2.3 Knowledge Consistency Verification

[0049] To alleviate the problem of factual illusions generated by large language models, this invention designs a lightweight verification framework based on semantic similarity. For a target item, feature vectors of its key attributes are extracted from the knowledge base, and the corresponding features generated by the large language model are converted into vectors. Similarity is calculated between the two; a successful match is achieved if the similarity value exceeds a set threshold. If the match fails, content from the knowledge base is referenced. Taking a movie dataset as an example, this framework is specifically optimized for four key semantic fields: director, screenwriter, actors, and keywords. For each semantic field... First, the text generated by the original large language model and the text from the standard knowledge base are preprocessed, including text cleaning, word segmentation, and standardization. Then, the corresponding vector representations are obtained through TF-IDF vectorization. The similarity calculation uses the widely used cosine similarity metric.

[0050] ;

[0051] in It is the original text vector representation. It is a standard knowledge base text vector representation.

[0052] The final calculation formula is as follows:

[0053] ;

[0054] in, It is a threshold that ultimately generates reliable knowledge. The keyword features were used as input for subsequent large language models and traditional models. In experiments where keyword features were introduced into traditional recommendation models, it was found that most evaluation metrics showed performance degradation, with only a few showing performance improvement. The experimental results show that the underlying architecture of the model determines its ability to process high-dimensional sparse semantic features.

[0055] Through the system integration of the above three key components, comprehensive, accurate, and balanced input information is provided for large language model inference. This process can be formally represented as:

[0056] ;

[0057] in, This represents the final constructed enhanced context. This indicates an information concatenation operation, where... This represents the context of a negative description of the project, derived from the process of constructing a negative description of the project as described in section 1.2.1. This represents the context of similar user behaviors, derived from the user similar behavior construction module in section 1.2.2. This represents a rigorously validated factual knowledge context, derived from the knowledge consistency verification process described in 1.2.3.

[0058] 1.3 Context-guided large language model recommendation

[0059] To alleviate the scoring bias of large language models and optimize their reasoning process, this invention proposes a context-guided objective thinking chain framework. The core of this framework is to construct a balanced cognitive foundation, thereby guiding large language models to make more neutral and reliable inferences.

[0060] 1.3.1 Structured Thinking Chain Reasoning

[0061] This stage requires large language models to be based on enhanced context. This invention employs structured reasoning, which requires clearly identifying the potential positive and negative factors of a film. Based on this, a comprehensive judgment is made. The rating task is broken down into multiple reasoning steps, each with a clear objective and output: In the preference analysis phase, the model infers the user's interest preference pattern based on their historical behavior, while simultaneously determining the user's final preference by considering the negative descriptions of items. The preference result is a binary "like" or "dislike." After determining the preference, the model then uses the rating range defined by the corresponding preference, the basic characteristics of the item, and the rating distribution of similar items to determine a reasonable rating interval and provide the final rating. Furthermore, this invention constrains the evaluation criteria, requiring the model to consider both positive and negative factors during the evaluation process, conducting a comprehensive trade-off, and ultimately generating results in a structured form. The reasoning process is as follows:

[0062] ;

[0063] in It is a reasoning text containing a complete analysis of both sides. These are the prompt words designed in this invention. Based on this, this invention further extracts the score of the large language model from the above structured output results. By analyzing the content of the corresponding rating fields in the structured data, the target result required for the rating prediction task can be obtained:

[0064] ;

[0065] The final prediction result is independent of the output of a traditional recommendation model. Perform linear fusion to generate the final prediction. ,in, These are the corresponding fusion weights;

[0066] 1.3.2 Context-guided prompt word engineering

[0067] The core of this invention lies in constructing a multi-step thought chain reasoning process. By guiding the model to engage in progressive, in-depth thinking, it ensures the logical rigor and reliability of the scoring and prediction process. The thought chain structure is designed with five distinct stages, each built upon the reasoning results of the previous stage, forming a complete logical loop, such as... Figure 2 As shown;

[0068] 1) Given the role and task

[0069] The prompts set a clear professional background and task objectives for the model. The model is positioned as a recommendation system with professional knowledge in the film field, which needs to make comprehensive judgments based on the provided multi-dimensional information. This stage not only establishes the basic task framework of the model, but more importantly, it establishes a professional thinking paradigm and analytical perspective for the subsequent reasoning process.

[0070] 2) Given user information

[0071] To guide the large language model in performing accurate and controllable rating prediction tasks, a comprehensive and balanced contextual information was constructed for it, enabling the large language model to simultaneously access positive and negative evidence, individual history and group behavior, and collaborative signals, thus laying a solid data foundation for its reasoning.

[0072] 3) Determine user preferences

[0073] At this stage, the prompt-guided model deeply analyzes the user's historical behavioral data, including viewing records, rating patterns, and genre preferences. The model needs to identify the user's interests, aesthetic preferences, and rating habits from this data to form a comprehensive understanding of the user's characteristics. This analysis provides a key cognitive foundation for subsequent personalized recommendations, ensuring that the recommendation results are consistent with the user's preferences. Finally, it tells the user whether they like the current movie, with the preference result being "like" or "dislike".

[0074] 4) Determine the score by combining the cooperative signal.

[0075] At this stage, the model is guided to deeply associate user preferences with project characteristics. The model needs to comprehensively consider the movie attributes verified by the knowledge base, the behavior patterns of similar users, and the advantages and disadvantages of the project itself. In this process, the prompt words particularly emphasize the application of dialectical thinking, requiring the model to balance positive and negative factors and avoid biases of excessive optimism or pessimism. This multi-dimensional balancing ensures the comprehensiveness and objectivity of the scoring decision.

[0076] 5) Structured output

[0077] The structured output stage transforms the complex reasoning process into clear and readable results through strict format specifications. The prompts clearly define the specific format and requirements of the output, ensuring that the scoring results generated by the model not only meet professional standards but also facilitate subsequent processing. This design not only improves the operability of the results but also enhances the interpretability of the entire reasoning process.

[0078] This complete set of prompt words ensures that the large language model can maintain a professional level of judgment, execute the system's reasoning process, and produce standardized results in recommendation tasks. Each prompt word component has been carefully debugged and repeatedly verified, maximizing functionality while maintaining simplicity, and providing a solid technical guarantee for the reliable operation of the recommendation system.

[0079] 2. Experiment

[0080] This invention aims to address the following research questions by conducting experiments on real-world datasets to evaluate the performance of various methods in score prediction:

[0081] RQ1: Will using consistency verification improve the overall predictive ability of the model?

[0082] RQ2: Can negative descriptive information make the model's predictions more objective and improve the model's ability to predict low-scoring cases?

[0083] RQ3: Which works better, a single, well-developed cue word or multiple cue words with different focuses?

[0084] RQ4: How do key hyperparameters affect model performance?

[0085] 2.1 Experimental Setup

[0086] 2.1.1 Dataset

[0087] This invention conducts comprehensive experiments on two widely used datasets: one is the MovieLens-100K (ML-100K) dataset, widely used for evaluating recommendation algorithms in movie rating prediction scenarios; the other is Kaggle-Movie, an extended MovieLens dataset released on the Kaggle platform. Following the method of the Lamar authors, this invention uses the last interaction as the test set, the second-to-last interaction as the validation set, and the rest as the training set. The features of each dataset are summarized in Table 1:

[0088] Table 1: Dataset Data

[0089]

[0090] 2.1.2 Baseline Model

[0091] This invention integrates TrustRec with the rating prediction model commonly used in recommendation systems, compares it with the Lamar method, and uses the Lamar model as the backbone model of the large language model.

[0092] DeepFM employs a parallel integrated architecture of factorization machine (FM) and deep neural network (DNN) with a shared input embedding layer, enabling efficient learning of low-order and high-order feature interactions.

[0093] DCN: It uses cross-networks to improve the wide layer part of the deep wide model, thereby learning high-order feature interactions more efficiently.

[0094] xDeepFM: Adds a Compressed Interaction Network (CIN) that runs in parallel with a deep network, while retaining the second-order interaction capability of the factorization machine, thus enabling more refined explicit learning of higher-order feature interactions;

[0095] AutoInt: By leveraging a multi-head self-attention mechanism, it adaptively assigns weights to interactions between different features, thereby automatically constructing higher-order features;

[0096] AFN: Based on the factorization mechanism, an attention network is introduced to dynamically assign contribution weights to different feature interactions, thereby more accurately strengthening the learning of key high-order feature interactions.

[0097] DIFM: It adopts a dual-branch structure to separate linear features and interactive features for modeling, and combines deep networks with interactive perception mechanisms to capture high-order feature interactions in complex scenes more efficiently.

[0098] DCNMix: Based on the DCN-V2 cross network, it incorporates the ideas of low-rank matrix factorization and hybrid expert network, compresses redundant parameters and improves the representation ability of cross layers, thereby more efficiently balancing high-order feature interaction learning and model inference efficiency.

[0099] 2.1.3 Evaluation Indicators

[0100] This invention uses two commonly used metrics: Root Mean Square Error (RMSE) is calculated by first squaring the prediction error of each user-item rating, then averaging all squared errors, and finally taking the square root. This is better at capturing extreme prediction errors that may seriously affect user experience and can indicate that the model has a problem of "recommendation results deviating from core user preferences." The smaller the RMSE value, the smaller the deviation between the model's prediction results and the true values, and the higher the prediction accuracy. Mean Absolute Error (MAE) is calculated by taking the absolute value of each rating prediction error and then averaging all absolute errors. In recommendation systems, this more objectively reflects the model's prediction accuracy for most user-item interactions. The smaller the MAE, the more accurately the model captures the preferences of most users, and the higher the average reliability of the recommendation results.

[0101] 2.1.4 Experimental Setup

[0102] To ensure fairness in the comparison, the optimal hyperparameter settings reported in the original papers of each benchmark model were adopted, and the hyperparameters of all benchmark models were fine-tuned through grid search. For each user, a leave-one-out strategy was used: the item the user interacted with last was used for testing, the item interacted with second to last was used for validation, and the remaining items were used for model training. The results were verified through experiments. The value is 5. =0.78, =0.3;

[0103] 2.2 Performance Comparison

[0104] To comprehensively evaluate the effectiveness of TrustRec in recommendation tasks, the model of this invention was compared with seven deep recommendation models and variants using existing LLM enhancement methods (Lamar) on the ML-100K and Kaggle-Movie datasets. The experimental results are summarized in Table 2, from which the following key results are obtained. The best results in the table are marked in bold.

[0105] Table 2: Results after using different methods

[0106]

[0107] As shown in Table 2, on all backbone networks and two datasets, the method of this invention consistently outperforms the baseline model and the Lamar comparison model in terms of root mean square error (RMSE) and mean absolute error (MAE). This comprehensive validation fully demonstrates the effectiveness and robustness of the model of this invention in improving the accuracy of score prediction.

[0108] Compared to traditional recommendation methods, both Lamar and Trust-Rec demonstrate improved performance. This reflects the limitations of conventional models in capturing deep semantic relationships between users and products, further highlighting the significant value of integrating the semantic understanding capabilities of large language models with traditional recommendation mechanisms. However, Lamar's performance fluctuates or even degrades in specific scenarios, indicating that simply introducing content or features generated by LLM is insufficient for continuous improvement. The key lies in achieving deep knowledge fusion between LLM and the recommendation structure. Trust-Rec effectively corrects the systematic scoring bias in LLM through a unique knowledge constraint mechanism, while injecting factual semantic knowledge into the traditional recommendation model. This design allows the model to leverage the deep semantic understanding capabilities of large language models and enhance its discriminative effectiveness based on collaborative behavior, thereby improving recommendation performance.

[0109] 2.3 Ablation Experiment

[0110] 2.3.1 Module Contribution (RQ1, RQ2)

[0111] Ablation experiments were conducted on different modules of the TrustRec framework to explore the performance impact of each module on the overall recommendation task; the ablation results are shown in Table 3, with the best results highlighted in bold.

[0112] Table 3: Ablation experiments on the ML-100K dataset, with DeepFM as the backbone model.

[0113]

[0114] Experimental results show that the thought chain module contributes most significantly to improving recommendation performance. Removing this module leads to a sharp drop in model performance metrics. This result confirms that refined instructions can fully unleash the deep reasoning capabilities of large language models and play a crucial role in improving recommendation quality. The knowledge consistency verification module ranks second, and its absence leads to a significant performance degradation, indicating that knowledge constraints can effectively suppress factual illusions and improve the accuracy of model recommendations. In contrast, the similar behavior module and the negative description module both bring moderate performance improvements. The former enhances recommendation personalization by integrating similar behaviors, while the latter optimizes the completeness of preference learning by balancing positive and negative feedback. Both exhibit strong robustness on different datasets.

[0115] Overall, removing any component leads to a performance degradation, and the degradation is proportional to the criticality of that component in solving the core challenges of trusted recommendation. The complete TrustRec framework consistently outperforms all ablation variants, proving that each component overcomes different technical difficulties, and its integration forms a comprehensive solution that surpasses the sum of the contributions of each module.

[0116] 2.3.2 Prompt Strategy (RQ3)

[0117] To verify the impact of prompt words on model performance, this invention systematically compared the effects of different prompt word strategies. Paired t-tests were used to compare the performance differences between single-prompt and multi-prompt methods in the rating prediction task. The experiment was based on 5 sets of paired data, with a statistical significance level set at α=0.05. Effect sizes were calculated using Hedges' g correction. As shown in Table 4, single structured prompt words significantly outperformed the multi-prompt ensemble strategy in all indicators. Specifically, for the MAE indicator, the average value of the single-prompt method was 0.6949, while the average value of the multi-prompt method was 0.7891. The paired t-test showed that this difference was statistically significant (t(4)=-36.19, p<0.001, Hedges' g=-12.95). For the RMSE indicator, the average value of the single-prompt method was 0.9001, while the average value of the multi-prompt method was 1.0044. The paired t-test also showed a significant difference (t(4)=-42.49, p<0.001, Hedges' g=-12.95). (g=-15.20), all five sets of experimental data consistently show that the single-prompt word method performs better, and the statistical results support the advantage of single structured prompt words in terms of stability;

[0118] Table 4 Summary of paired t-test results (n=5)

[0119]

[0120] The analysis results show that while the multi-cue word strategy can increase the diversity of output, it also introduces the risk of decision inconsistency. In contrast, a single but deeply structured cue word guides the reasoning through a clear reasoning path, ensuring the coherence and reliability of logical reasoning. Compared with increasing the number of cue words, optimizing the quality and structure of a single cue word is more effective in improving performance.

[0121] 2.3.3 Impact of key hyperparameters on model performance (RQ4)

[0122] In this invention, three hyperparameters have a significant impact on recommendation performance: (1) the number of neighbors hyperparameter (2) Hyperparameter threshold ;

[0123] (3) Fusion weight hyperparameters This invention uses the recommended model DeepFM as the basic model for parameter sensitivity analysis, proving... The value is 5. =0.78, =0.3;

[0124] 1) Number of neighbors

[0125] During model optimization, the optimal parameter for the number of recommended neighbors is determined by adjusting the Top-Number parameter. K The value of a neighbor number is a key factor affecting the performance of a recommendation system. This can be achieved by setting different values. The values ​​were tested, and the results were as follows: Figure 3 As shown, overall, appropriately increase Values ​​help cover more user interests and improve recommendation accuracy; however, Values ​​that are too large or too small will both lead to performance degradation, such as... Figure 3 As shown, when =1 or When the value is 7, the model performs worse than others in both mean absolute error (MAE) and root mean square error (RMSE). =5 indicates that a list that is too long will introduce noise, while a list that is too short will not be able to fully learn user preference features;

[0126] 2) Hyperparameter threshold

[0127] A systematic experiment was conducted to evaluate the impact of different threshold settings on model performance. The experiment covered a threshold range of 0.6-0.9, with a step size of 0.02. Each threshold was applied to a uniform model architecture and dataset to ensure the comparability and reliability of the results. The experimental results show that when the threshold is set to 0.78, the model achieves the best performance in both metrics. However, thresholds that are too low or too high lead to a decline in model performance: when the threshold is below 0.78, both error metrics generally increase, indicating that a threshold that is too small may introduce noise; when the threshold is increased to 0.8 or above, although the model reduces the error to some extent, it loses too much generated information due to over-reliance on the true labels. Specific results are as follows: Figure 4 As shown;

[0128] 3) Fusion weight hyperparameters

[0129] This invention systematically evaluates different fusion weights. The experimental results show that the fusion weights affect the overall performance of the model. The value of is crucial to the recommendation performance. Experiments covered a range from 0.2 to 0.4, all applied to a unified model architecture and dataset to ensure the comparability and reliability of the results. Specific results are shown in the figure. Figure 5 As shown, when If the value is too small, the model relies too heavily on the results of traditional recommendation models, and the knowledge enhancement and semantic understanding capabilities of large language models are not fully utilized; conversely, when the value is too small, the model becomes overly reliant on the results of traditional recommendation models. When the value is too large, exceeding 0.35, the large language model dominates the fusion decision excessively. This may introduce inherent uncertainty and weaken the accurate signals captured by traditional models from user behavior data. A comprehensive analysis of the experimental results shows that when... When the weight is set to 0.3, the model achieves the best performance balance on the test set. Under this weight configuration, the accurate recommendation capability of the traditional recommendation model and the reasoning capability of the large language model form the most effective synergy.

[0130] This invention aims to systematically solve the illusion and bias problems of large language models, thereby providing a solid and reliable implementation path for the collaborative paradigm. It not only verifies the effectiveness of this path, but also further discovers that: First, traditional recommendation models generally suffer from structural bottlenecks when facing high-dimensional sparse semantics, and only a few models with specific interaction architectures can benefit to a limited extent. This fundamentally establishes the irreplaceable position of large language models in the semantic understanding process. Second, the reliability of large language model components is the cornerstone of the entire system, and a logically rigorous single prompt word is the key design for achieving reliable reasoning.

[0131] The above embodiments are illustrative of the present invention and are not intended to limit the present invention. Any simple modifications to the present invention are within the scope of protection of the present invention.

Claims

1. A reliable recommendation model based on a two-dimensional constrained large language model, characterized in that: Includes the following steps: S1. Negative description module and similar behavior generation module: The negative description module and similar behavior generation module provide a multi-dimensional balanced perspective for subsequent reasoning; S2. Knowledge Consistency Verification Module: The knowledge consistency verification module uses a lightweight fact consistency detection process to ensure that the features of the input model have high reliability. S3, Thinking Chain Reasoning Module: The thinking chain reasoning part decomposes the task, from user preference analysis to the generation of the final score, forming a complete reasoning chain; S4. Fusion Module: The fusion module integrates the statistical results of the traditional recommendation channel with the inference results of the large language model channel through adaptive weight adjustment to generate the final recommendation result. In step S3, the thought chain reasoning module performs context-guided prompt word engineering: constructing a multi-step thought chain reasoning process, guiding the model to conduct progressive deep thinking; the thought chain structure is designed with five distinct stages, each stage building upon the reasoning result of the previous stage, forming a complete logical loop: (1) Given role and task: The prompt words set a clear professional background and task goal for the model; the model is positioned as a recommendation system with professional knowledge in the project domain, and makes a comprehensive judgment based on the provided multi-dimensional information; (2) Given user information: Construct a comprehensive and balanced contextual information; This enables large language models to simultaneously access positive and negative evidence, individual history and group behavior, and collaborative signals as the data foundation for their reasoning; (3) Determine user preferences: The prompt words guide the model to deeply analyze the user's historical behavior data. From this data, the model identifies the user's interests, aesthetic preferences and rating habits, forming a comprehensive understanding of the user's characteristics; (4) Combine collaborative signals to determine the score: guide the model to deeply associate user preferences with project characteristics; the model comprehensively considers the project attributes after knowledge base verification, the behavior patterns of similar users, and the advantages and disadvantages of the project itself. At the same time, the model adopts dialectical thinking to balance positive and negative factors and avoids excessive optimism or pessimism. (5) Structured output: The complex reasoning process is transformed into a clear and readable result through format specifications, and the prompt words clearly specify the specific format and requirements of the output.

2. The reliable recommendation model based on a two-dimensional constrained large language model as described in claim 1, characterized in that: In step S1, the negative description module constructs negative descriptions of the project: it constructs negative descriptions from real user data, using user information with low ratings of the current project and negative information carried by the project itself as information sources; it uses a model with a larger knowledge base and richer content, taking the project's negative information and low-rated user information as inputs, to generate descriptions.

3. The reliable recommendation model based on a two-dimensional constrained large language model as described in claim 1, characterized in that: In step S1, the similar behavior generation module constructs user similar behavior: it constructs a collaborative filtering signal extraction based on a predefined type similarity matrix, and uses the semantic relevance of item types to identify user behavior patterns with similar preferences; it defines a set covering multiple main types, and defines a semantic similarity matrix by calculating the cosine similarity between their semantic vector representations; for the current item and the user's historical items, it sets their type sets respectively, and the similarity calculation adopts the maximum matching strategy; based on the calculated similarity, it selects the top few historical items that are most similar to the current item type, and integrates the user's corresponding interaction behavior and item metadata information to form an input representation of a collaborative filtering signal.

4. A reliable recommendation model based on a two-dimensional constrained large language model as described in claim 1, characterized in that: In step S2, the knowledge consistency verification module performs knowledge consistency verification: a multi-level verification framework based on semantic similarity is designed; for the target item, feature vectors of its key attributes are extracted from the knowledge base, and the corresponding features generated by the large language model are converted into vectors. The similarity between the two is calculated. If the similarity value exceeds a set threshold, it is considered a successful match. If the match fails, the content in the knowledge base is referenced; firstly, the text generated by the original large language model and the standard knowledge base text are preprocessed, including text cleaning, word segmentation and standardization. Then, the corresponding vector representation is obtained through TF-IDF vectorization. The similarity calculation uses cosine similarity measurement. Finally, reliable knowledge is generated and used as input for subsequent large language models and traditional models.

5. A reliable recommendation model based on a two-dimensional constrained large language model as described in claim 1, characterized in that: In step S3, the thought chain reasoning module performs structured thought chain reasoning: the large language model, based on enhanced context, performs structured thought chain reasoning; it clearly lists the potential positive and negative factors of the item, and makes a comprehensive judgment based on this, decomposing the scoring task into multiple reasoning steps, each with a clear goal and output: in the preference analysis stage, the model infers the user's interest preference pattern based on the user's historical behavior, and at the same time combines the negative description of the item to determine the user's final preference; after determining the preference, the model then uses the scoring range specified by the corresponding preference, the basic characteristics of the item, and the scoring distribution of similar types of items to determine the scoring interval and give the final score.

6. A reliable recommendation model based on a two-dimensional constrained large language model as described in claim 5, characterized in that: In step S3, during the structured thinking chain reasoning process, the thinking chain reasoning module constrains the evaluation criteria. The model considers both positive and negative factors during the evaluation process, making a comprehensive trade-off, and finally generates the results in a structured form.

7. A reliable recommendation model based on a two-dimensional constrained large language model as described in claim 1, characterized in that: In step S4, the fusion module extracts the score of the large language model from the structured output result in step S3. By parseting the content of the corresponding score field in the structured data in a targeted manner, the target result required for the score prediction task is finally obtained. The final prediction result is linearly fused with the output of a traditional recommendation model that is independent of the model to generate the final prediction.

Images