Recommendation method for enhancing semantics and interest perception by using large language model

Abstract

The invention discloses a recommendation method for enhancing semantics and interest perception by using a large language model. The method comprises the following steps: firstly, performing semantic modeling on unstructured text information such as user comments, article description and the like by utilizing the powerful capability of a large language model in semantic comprehension and user preference modeling aspects, so as to improve the deep perception capability of a recommendation system on user interests and article semantic attributes; then, through a semantic feature alignment and discretization strategy, the problem that continuous semantic representation generated by a large language model is incompatible with features of a traditional recommendation system in the aspect of an expression structure is solved; finally, unified modeling of semantic information and traditional recommendation signals is achieved through a recommendation integration mechanism, and recommendation performance and model interpretability are improved.

Description

Technical Field

[0001] The present invention relates to a recommendation method for enhancing semantics and interest perception by utilizing a large language model, and belongs to the field of artificial intelligence and recommendation technology. Background Art

[0002] With the development of e-commerce and content platforms, recommendation systems have been widely studied as a core module for improving user experience and platform revenue. Traditional recommendation methods rely primarily on collaborative filtering, content feature matching, and deep learning models, but their performance is limited when faced with cold start issues, data sparsity, and complex expressions of user interests.

[0003] In recent years, large language models (LLMs) have demonstrated remarkable semantic modeling capabilities in natural language processing, providing a new opportunity to introduce semantic reasoning into recommendation systems. However, directly using the semantic representations generated by LLMs for recommendation modeling faces the following challenges: LLM outputs are typically high-dimensional semantic embeddings, which differ significantly from the embedding dimensions of ID features in recommendation systems; recommendation systems typically employ a single epoch for rapid training, while LLM representations require long-term optimization cycles; embeddings are difficult to interpret and integrate: the continuous nature of semantic embeddings is incompatible with the primarily discretized features of recommendation systems; and interest modeling lacks granularity: it is difficult to capture the matching distribution of users' diverse interests. Therefore, an innovative approach that tightly integrates the advantages of large language models with the structure of recommendation systems is urgently needed to address these issues and improve recommendation effectiveness. Summary of the Invention

[0004] To address the problem that traditional item recommendation methods lack sufficient semantic information modeling, which limits the accuracy of recommendations, this paper proposes a recommendation method that uses a large language model to enhance semantic and interest perception. This invention is achieved through the following technical solutions:

[0005] Step 1: Semantic and Interest Perception

[0006] This step leverages the language understanding and reasoning capabilities of the Large Language Model (LLM) to extract semantically deep user and item representations from unstructured text. This step is a crucial perception module in the overall recommendation process. First, the system designs context-aware prompts for all item text attributes on the platform, including titles, descriptions, user reviews, and technical specifications, and feeds them into the LLM. Using natural language generation, the system generates a semantic profile for the item, capturing its core selling points, implicit attributes, and target demographic. Subsequently, to enhance the model's understanding of user preferences, this step further uses reverse reasoning prompts to guide the LLM in answering questions like "Why did the user purchase this item?" or "What user interest tags might this item attract?", thereby constructing reasoning features for the item. This "user-motivation-driven" semantic completion significantly improves the comprehensiveness of item semantic representations. For user-side modeling, the system collects a list of historically interacted items, concatenates the corresponding item semantic profiles into a semantic sequence, and feeds the LLM with this sequence. This automatically generates a user profile text, refining their long-term interests, value orientations, and preferred themes. Finally, the semantic text is further converted into a neural network-readable embedding representation, constructing user semantic embeddings and interest vectors, laying the semantic foundation for subsequent feature alignment and modeling. This module overcomes the limitation of traditional ID recommendation that cannot utilize unstructured information, enabling the recommendation system to perceive and model user interests and item semantics in a fine-grained manner.

[0007] Step 2: Semantic Embedding Alignment

[0008] This step mainly solves the structural difference problem of "difficult to directly integrate" the high-dimensional semantic embeddings generated by the large language model in the recommendation system. Since the vector dimension generated by LLM is relatively high (such as 1024 or 2048), and the embedding dimension commonly used in the recommendation system is much lower than this (such as 64 to 128 dimensions), direct splicing will lead to unstable model training, overfitting or redundant computing resources. Therefore, this step first uses a trainable multi-layer perceptron (MLP) dimensionality reduction network to map the semantic embeddings of items and users to the target recommendation space respectively to achieve dimensional alignment. On this basis, the present invention introduces a pre-training mechanism, and designs a CTR click-through rate prediction task as a self-supervised training target for the semantic vector after dimensionality reduction, and uses the BPR loss function to perform multiple rounds of optimization training on the semantic module to alleviate the training delay problem of semantic embedding. In addition, considering that the recommendation system relies on contextual interaction information, this step introduces the DIN (Deep Interest Network) model to dynamically match the user's semantic history sequence with the current candidate item, and enhances the contextual semantic modeling capability of the target item through the attention mechanism. The final output of aligned semantic embedding not only has semantic reasoning capabilities, but can also seamlessly integrate with the traditional recommendation model structure, greatly improving the expressiveness and robustness of the recommendation system in cold start and long-tail scenarios.

[0009] Step 3: Discrete feature construction

[0010] This step aims to convert semantic interest representation from a "continuous representation" to a "discrete feature" that is more suitable for industrial recommendation systems, thereby improving the interpretability, fusionability, and computational efficiency of semantic information. Specifically, based on the user interest representation obtained in step 1 and the item inference features after dimensionality reduction in step 2, the similarity between them is calculated through vector dot product to obtain a set of matching score sequences. Next, the system sets a series of thresholds (such as dividing the [-1,1] interval into 10 intervals) and divides all matching values ​​into corresponding buckets to form a discrete distribution vector (Bucket Histogram). This distribution vector can be regarded as a "similarity response map" of the user's interest to the item. To further enhance the model's perception ability, this step also calculates the high-order statistical features of the distribution vector, including mean, variance, skewness, and kurtosis, which are used to characterize the concentration, directionality, and extreme preferences of the interest distribution. Ultimately, bucket counts are concatenated with statistical features to form a unified discrete semantic interest vector. This not only preserves semantic matching information but can also be efficiently integrated with traditional discrete ID features into model training. This module avoids the training burden of directly introducing high-dimensional text embeddings while improving the recommendation system's ability to characterize the distribution of user behavior. It serves as a key link in the fusion of semantics and structure.

[0011] Step 4: Recommended Integration

[0012] After semantic embedding alignment and discrete feature construction, this step integrates various feature information into a model to predict user preference ratings for items. Specifically, the system first concatenates the user's semantic embedding, discrete interest vector, and high-order statistics with their underlying ID features (such as gender, age, region, and spending power) into a unified vector. Similarly, the item's semantic embedding is fused with basic attribute features (such as item ID, category, and brand). The user and item vectors are then concatenated into a joint representation and fed into a multi-layer perceptron (MLP) model. This model typically consists of several layers of nonlinear mapping, supplemented by normalization (LayerNorm) and activation functions (ReLU) to fully capture the nonlinear relationships and high-order interaction features between users and items. The final output layer uses a sigmoid function to map the model's predicted values ​​to probabilities between 0 and 1, representing the user's preference rating for clicking or purchasing the item. This step not only integrates multi-source semantic and structural information but also exhibits excellent scalability and industrial deployment efficiency. It is suitable for large-scale online recommendation systems, significantly improving recommendation accuracy and user experience.

[0013] The present invention takes into account the depth of semantic perception, fusion of structural features, fine-grained interest modeling and adaptability of computing efficiency, and can be widely used in scenarios such as e-commerce, content platforms, and advertising recommendations.

[0014] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method for enhancing semantics and interest perception using a large language model is implemented.

[0015] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the recommendation method for enhancing semantics and interest perception using a large language model.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] (1) Compared with traditional recommendation systems that only rely on sparse ID information or shallow text feature extraction methods, the present invention uses a large language model to perform inference modeling on unstructured texts such as user comments and item descriptions, comprehensively extracting item semantic portraits and user interest semantic expressions, and significantly improving the modeling capabilities of the recommendation system in cold start, long-tail items and user interest migration scenarios.

[0018] (2) The present invention proposes a semantic embedding alignment mechanism, which uses a multi-layer perceptron to map high-dimensional semantic embeddings to the target dimension of the recommendation model, and optimizes the semantic representation through the CTR pre-training task to solve the problem in the existing technology that LLM output and the recommendation system are difficult to directly integrate, thereby improving the compatibility and training stability of the model.

[0019] (3) To address the problem that semantic interest representation is difficult to interpret and utilize in industrial recommendation systems, this paper innovatively uses a bucket statistical method to discretize the similarity scores, and further extracts statistics such as mean, variance, skewness and kurtosis to construct a discrete interest vector that is both interpretable and expressive, thereby optimizing the recommendation feature design method. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a diagram of the overall framework of the KG-RL method involved in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to deepen the understanding of the present invention, the present invention is further illustrated in conjunction with specific examples. Figure 1 , a recommendation method that uses a large language model to enhance semantic and interest perception. The specific implementation steps are as follows:

[0022] Step 1: Semantic and interest perception, including the following sub-steps:

[0023] Sub-step 1-1: Item semantic information extraction is first performed by To analyze, we first use LLM to generate a description of the item as an item portrait. The formula is as follows:

[0024]

[0025] It is the text form of the image of the item, and LLM_Prompt1 is the prompt word for constructing the image of the item.

[0026] Thus, we introduce the external knowledge of LLM, add semantic information, and obtain the object image. Then, we use LLM to convert the object image from text to a vector that the model can recognize. The formula is as follows:

[0027]

[0028] Thus, the semantic vector of the item is obtained Convert textual object descriptions into corresponding embeddings.

[0029] Sub-steps 1-2: Modeling item inference features. Then, this module lets LLM simulate users. The specific formula is as follows:

[0030] F i =LLM_Prompt3(v i )

[0031] The F here i is the potential feature obtained by using LLM external knowledge to reason about the item, LLM_Prompt3 is the prompt word for the corresponding feature reasoning, v i Is the name of the item.

[0032] Thus, LLM is used to infer the feature F i Converted to the corresponding embedding, the specific formula is as follows:

[0033]

[0034] Finally, the description of the item and the features of the reasoning are converted into embeddings processed by the neural network

[0035] Sub-steps 1-3: User semantic profile generation based on item description This module describes the items that the user interacts with After merging and processing by the large language model, the formula is as follows:

[0036]

[0037] In this way, LLM can reason and summarize the description of the user interaction items and obtain the user portrait Pj As the semantic information on the user side, the user portrait P is then converted into j Convert to the corresponding embedding P_E j LLM_Prompt2 is a prompt word for inferring the user's interaction history items.

[0038] Sub-steps 1-4, user interest modeling. In the recommendation system, user interests can be modeled by which items they interact with, and the features of the items themselves can be used as proxy representations of user interests. This module embeds the inference features of the items that users interact with into the For aggregation, the formula is as follows:

[0039]

[0040] In this way, we can get the representation of users’ diverse interests. j User interests are characterized by similarity statistical bucketing in the following module, and direct concatenation does not lose information.

[0041] Step 2: Semantic embedding alignment, which includes the following sub-steps:

[0042] Sub-step 2-1, semantic embedding dimensionality reduction: The semantic embedding dimension obtained by LLM conversion is much larger than the dimension of the first version of recommendation features. This module attempts to set the conversion layer with different hyperparameters to reduce the dimension of the semantic embedding of the item. The formula is as follows:

[0043]

[0044] This is how we get the embedding of the item after dimensionality reduction In actual modeling, we experimented with various hyperparameters for the transformation layer's dimensions to evaluate their impact on recommendation effectiveness and training efficiency. Appropriate dimensionality reduction not only effectively mitigates the interference of high-dimensional semantic features on recommendation model training but also improves the final model's fusion performance and generalization capabilities.

[0045] Sub-step 2-2, pre-training and alignment of semantic embedding: Based on the importance analysis of the semantic alignment module, the semantic alignment module is introduced to decouple the optimization of semantic embedding from the optimization of traditional recommendation features. The goal of the semantic alignment module is to pre-train parameters related to semantic representation over multiple periods to ensure their convergence. By using the CTR prediction task as the recommendation pre-training task, the loss function uses the bpr loss to achieve pre-training of semantic embedding. This module uses the DIN model to process the semantic representation of the target item and the historical interaction item, and outputs P_E DIN , the formula is as follows:

[0046]

[0047] Characterization It is fed into MLP for the final prediction, and the formula is as follows:

[0048]

[0049] This gives the predicted probability This module continues to optimize the cross entropy loss L between the predicted click probability and the binary click label. Align , the calculation formula is as follows:

[0050]

[0051] By pre-training semantic embedding, the alignment of semantic embedding and ID embedding is achieved.

[0052] Step 3: Discrete feature construction, which includes the following sub-steps:

[0053] Sub-step 3-1, semantic similarity bucket modeling: first calculate the inference feature v of the target candidate c and the current user interest representation Interest j The similarity is, so the calculation formula is as follows:

[0054]

[0055] By performing dot product between the inference features of the candidate items and the representation of the user’s interests, we can obtain the matching degree vector s between the candidate items and the user’s interests. i After calculating the similarity score, the module defines N-1 thresholds to separate s i The score range of [-1.0, 1.0] for each dimension of the vector is divided into N buckets. In each bucket, the number of similarity scores falling into the bucket is counted as follows:

[0056]

[0057] Where U(·) is the indicator function, when s i Falls into the kth bucket b k-1 and b k When h k Represents the similarity score count of the k-th bucket. Thus, an N-dimensional vector H is obtained as the user bucket interest feature, as shown in the formula:

[0058] H j =[h1,h2,…,h N ]

[0059] The bucket interest distribution H jIt is used to construct new auxiliary features and enhance the information representation capabilities of recommendation systems. Bucket statistics can effectively reduce the high-dimensionality of text embeddings in recommendation systems, avoiding the increased computational complexity and decreased model performance associated with directly concatenating text embeddings.

[0060] Sub-step 3-2, high-order statistical feature extraction and enhancement: In order to further enhance the expression ability of interest features, after obtaining the discretized bucket interest distribution H j Finally, this module introduces statistical feature calculation operations to extract high-order statistical information of interest distribution and construct richer feature expressions. Specifically, for the bucket interest distribution H j The following four statistical features are calculated using the following formulas:

[0061]

[0062] where μ j , Skew j and Kurt j The kurtosis and skewness represent the mean, variance, skewness, and kurtosis of bucketed interest values, respectively. The mean reflects the user's overall interest in a candidate item, the variance measures the dispersion of interest, and helps determine the concentration of user preferences. The skewness reveals the direction of the interest distribution and is used to identify personalized or exploratory needs. The kurtosis describes the sharpness of the interest distribution, making it easier to capture extreme preferences or generalization tendencies. These statistical features not only improve feature interpretability but also enhance the adaptability of recommendation models to different interest structures, providing more stable and effective auxiliary information for downstream prediction tasks.

[0063] Sub-step 3-3, interest enhancement feature vector construction: these four statistical features are combined with the bucket interest distribution H j Perform splicing to obtain enhanced interest representation vector The input features used to construct the recommendation model are as follows:

[0064]

[0065] This design captures the overall shape of user interests from a distributional perspective, avoiding the curse of dimensionality associated with directly using high-dimensional text embeddings while also enhancing the model's discriminative and generalizable capabilities in interest modeling. Recommendation systems primarily rely on discrete features (such as click counts and categories) rather than continuous features. Discrete features generated by bucketing statistics can be better integrated with traditional recommendation features, improving model compatibility. Furthermore, by performing bucketing on features, this method adapts to the distribution of diverse user interests and improves recommendation accuracy.

[0066] Step 4: Recommend integration, which includes the following sub-steps:

[0067] Sub-step 4-1, Multi-source Feature Normalization: Receive the user interest discrete features H_j^final, user semantic embedding vector P_E_j, item semantic embedding P_E_i, and ID embedding features from traditional recommendation methods, namely user basic features F_j and item basic features F_i. These features include a user's historical preference description, semantic information, and structured basic attributes.

[0068] Normalization is performed on various features. The operations include: applying LayerNorm normalization to the user interest feature H_j^final; performing LayerNorm on the semantic embeddings P_E_j and P_E_i respectively; and performing BatchNorm on the basic features F_j and F_i to maintain the stability of the input features during training.

[0069] Sub-step 4-2, joint feature splicing and modeling: splice the normalized feature vectors to construct the user-side feature vector u = concat(H_j^final, P_E_j, F_j) and the item-side feature vector v = concat(P_E_i, F_i), and further splice them into the overall input feature vector z = concat(u, v).

[0070] The joint feature vector z is input into the multi-layer perceptron (MLP) and nonlinear mapping learning is performed in sequence. The layer-by-layer processing logic is as follows: for the first L-1 layers, ReLU activation and LayerNorm normalization are applied: z = ReLU(LayerNorm(W_l*z+b_l)); for the last layer, only linear transformation is performed: z = W_L*z+b_L; finally, the user's preference score for the item y_ij = σ(z) is calculated through the Sigmoid activation function, where y_ij∈[0,1].

[0071] Sub-step 4-3, model training and loss optimization: Use the cross-entropy loss function (Cross-Entropy Loss) to measure and optimize the error between the predicted result y_ij and the actual interaction label y. The loss function is defined as follows: Loss = -y*log(σ(y_ij))-(1-y)*log(1-σ(y_ij)).

[0072] The MLP network weight W_l and bias b_l are gradient updated through the back-propagation algorithm to optimize the recommendation model performance.

[0073] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions fall within the scope of protection of the claims of the present invention.

Claims

1. A recommendation method for enhancing semantics and interest perception using a large language model, characterized in that: The method comprises the following steps: Step 1: The semantic encoder encodes the semantics and interest perception of the source text; Step 2: Use pre-training and dimension conversion to align the semantic encoding with the traditional recommendation encoding; Step 3: Use interest statistics bucketing to construct discrete features; Step 4: Use the aligned semantic encoding and interest discrete features for integrated recommendation.

2. The recommendation method for enhancing semantics and interest perception using a large language model according to claim 1, characterized in that: Step 1 specifically includes the following sub-steps: Sub-step 1-1: Item semantic information extraction is first performed by To analyze, we first use LLM to generate a description of the item as an item portrait. The formula is as follows: is the text form of the item portrait, and LLM_Prompt1 is the prompt word for constructing the item portrait, thereby introducing the external knowledge of LLM, adding semantic information, and obtaining the item portrait. Then, the item portrait is converted from text to a vector that the model can recognize using LLM. The formula is as follows: Thus, the semantic vector of the item is obtained Convert the textual image of an item into its corresponding embedding. Sub-steps 1-2: Modeling item inference features. Then, this module lets LLM simulate users. The specific formula is as follows: F i =LLM_Prompt3(v i ) The F here i is the potential feature obtained by using LLM external knowledge to reason about the item, LLM_Prompt3 is the prompt word for the corresponding feature reasoning, v i is the name of the item, Thus, LLM is used to infer the feature F i Converted to the corresponding embedding, the specific formula is as follows: Finally, the description of the item and the features of the reasoning are converted into embeddings processed by the neural network Sub-steps 1-3: User semantic profile generation based on item description This module describes the items that the user interacts with After merging and processing by the large language model, the formula is as follows: In this way, LLM can reason and summarize the description of the user interaction items and obtain the user portrait P j As the semantic information on the user side, the user portrait P is then converted into j Convert to the corresponding embedding P_E j ,LLM_Prompt2 is the prompt word for reasoning about the user interaction history items, Sub-steps 1-4, user interest modeling, in the recommendation system, this module embeds the inference features of the items that the user interacts with into For aggregation, the formula is as follows: In this way, we can get the representation of users’ diverse interests. j ,User interests are characterized by similarity statistical buckets in the following module, and direct ,joining will not lose information.

3. The recommendation method for enhancing semantics and interest perception using a large language model according to claim 2, characterized in that: Step 2 specifically includes the following sub-steps: Sub-step 2-1, semantic embedding dimensionality reduction: The semantic embedding dimension obtained by LLM conversion is much larger than the dimension of the first version of recommendation features. This module attempts to set the conversion layer with different hyperparameters to reduce the dimension of the semantic embedding of the item. The formula is as follows: This is how we get the embedding of the item after dimensionality reduction Sub-step 2-2, pre-training and alignment of semantic embedding: Based on the importance analysis of the semantic alignment module, the semantic alignment module is introduced to decouple the optimization of semantic embedding from the optimization of traditional recommendation features. The goal of the semantic alignment module is to pre-train parameters related to semantic representation in multiple periods to ensure its convergence. By using the CTR prediction task as the recommendation pre-training task, the loss function uses the bpr loss to achieve pre-training of semantic embedding. The DIN model is used to process the semantic representation of the target item and the historical interaction item, and the output is P_E DIN , the formula is as follows: Characterization It is fed into MLP for the final prediction, and the formula is as follows: This gives the predicted probability This module continues to optimize the cross entropy loss L between the predicted click probability and the binary click label. Align , the calculation formula is as follows: By pre-training semantic embedding, the alignment of semantic embedding and ID embedding is achieved.

4. The recommendation method for enhancing semantics and interest perception using a large language model according to claim 3, characterized in that: Step 3 specifically includes the following sub-steps: Sub-step 3-1, semantic similarity bucket modeling: first calculate the inference feature v of the target candidate c and the current user interest representation Interest j The similarity is calculated as follows: By performing dot product between the inference features of the candidate items and the representation of the user’s interests, we can obtain the matching degree vector s between the candidate items and the user’s interests. i After calculating the similarity score, the module defines N-1 thresholds to separate s i The score range of [-1.0, 1.0] for each dimension of the vector is divided into N buckets. In each bucket, the number of similarity scores falling into the bucket is counted as follows: Where I(·) is the indicator function, when s i Falls into the kth bucket b k-1 and b k When the count increases by 1, h k Represents the similarity score count of the k-th bucket, thereby obtaining an N-dimensional vector H as the user bucket interest feature, as shown in the formula: H j =[h1,h2,…,h N ] The bucket interest vector H j It is used to construct new auxiliary features and enhance the information representation ability of the recommendation system. The bucket statistical method can effectively reduce the high-dimensional impact of text embedding in the recommendation system and avoid the problems of increased computational complexity and decreased model performance caused by directly concatenating text embeddings. Sub-step 3-2, high-order statistical feature extraction and enhancement: In order to further enhance the expression ability of interest features, after obtaining the discretized bucket interest vector H j Finally, this module introduces statistical feature calculation operations to extract high-order statistical information of interest distribution and construct richer feature expressions. Specifically, the bucket interest vector H j The following four statistical features are calculated using the following formulas: where μ j , Skew j and Kurt j Respectively represent the mean (Mean), variance (Variance) skewness (Skewness) and kurtosis (Kurtosis) of the bucket interest distribution, where, Sub-step 3-3, interest enhancement feature vector construction: these four statistical features are combined with the bucket interest vector H j Perform splicing to obtain enhanced interest representation vector The input features used to construct the recommendation model are as follows:

5. The recommendation method for enhancing semantics and interest perception using a large language model according to claim 4, characterized in that: Step 4 specifically includes the following sub-steps: Sub-step 4-1, multi-source feature normalization processing: Receive the user interest discrete features H_j^final, user semantic embedding vector P_E_j, item semantic embedding P_E_i generated by the above modules, and the ID embedding features in traditional recommendation methods, namely user basic features F_j and item basic features F_i. The above features include user historical preference descriptions, semantic information, and structured basic attributes. Normalization is performed on various features. The operations include: applying LayerNorm normalization to the user interest feature H_j^final; performing LayerNorm on the semantic embeddings P_E_j and P_E_i respectively; and performing BatchNorm on the basic features F_j and F_i to maintain the stability of the input features during training. Sub-step 4-2, joint feature concatenation and modeling: concatenate the normalized feature vectors to construct the user-side feature vector u = concat(H_j^final, P_E_j, F_j) and the item-side feature vector v = concat(P_E_i, F_i), and further concatenate them into the overall input feature vector z = concat(u, v). The joint feature vector z is input into the multi-layer perceptron (MLP) and nonlinear mapping learning is performed in sequence. The layer-by-layer processing logic is as follows: For the first L-1 layers, ReLU activation and LayerNorm normalization are applied: z = ReLU(LayerNorm(W_l*z+b_l)); for the last layer, only linear transformation is performed: z = W_L*z+b_L; finally, the user's preference score for the item y_ij = σ(z) is calculated through the Sigmoid activation function, where y_ij∈[0,1], Sub-step 4-3, model training and loss optimization: Use the cross-entropy loss function (Cross-Entropy Loss) to measure and optimize the error between the predicted result y_ij and the actual interaction label y. The loss function is defined as follows: Loss = -y*log(σ(y_ij))-(1-y)*log(1-σ(y_ij)). The MLP network weight W_l and bias b_l are gradient updated through the back-propagation algorithm to optimize the recommendation model performance.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the recommendation method for enhancing semantics and interest perception by using a large language model is implemented as described in any one of claims 1 to 5.

7. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, the recommendation method for enhancing semantics and interest perception by using a large language model as described in any one of claims 1 to 5 is implemented.

Images