A causal dynamic semantic alignment method for LLM enhanced recommendation system

Abstract

This invention discloses a causal dynamic semantic alignment method for LLM-enhanced recommender systems, comprising: step one, initial user profile construction based on confounding factor awareness; step two, a dual causal decoupling mechanism; step three, iterative profile correction and semantic alignment training; and step four, lightweight deployment with zero online LLM overhead. Within the interdisciplinary field of artificial intelligence and recommender systems, this invention aims to address the problems existing in the prior art by providing a method for constructing and using user profiles that can dynamically update, remove bias, and semantically align with the recommendation target.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and recommendation systems, specifically to a causal dynamic semantic alignment method for LLM-enhanced recommendation systems. Background Technology

[0002] In recent years, Large Language Models (LLMs) have made significant progress in natural language understanding and generation, and have been introduced into recommender systems to enhance user intent modeling. Existing methods typically transform user historical behavior into textual profiles, which are then encoded into vectors and used as auxiliary input to the recommender model, thereby improving the interpretability and semantic richness of recommendations. However, current LLM-enhanced recommender systems still face the following prominent problems in practical applications: First, there's the issue of static user profiles becoming fixed. Current mainstream paradigms typically employ a "generate once, use throughout" strategy: in the early stages of training, the system calls an LLM (Local Level Model) once to generate a fixed text profile for each user, encodes it into an embedding vector, and then freezes it, never updating it again. This design ignores the dynamic learning characteristics of recommendation systems—as the model continuously optimizes on interactive data, its understanding of user preferences gradually becomes more refined and even corrected. However, static user profiles cannot perceive this evolutionary process, leading to a gradual disconnect between semantic representations and the preference representations learned internally by the recommendation model, resulting in a "profile-model cognitive misalignment."

[0003] Second, there is the issue of systemic bias contamination. LLM relies heavily on users' historical interaction records when generating user profiles. However, these records are themselves contaminated by the exposure mechanisms of the recommendation system—popular items receive more clicks due to high exposure, while less popular but high-quality items are ignored due to a lack of display opportunities. This hybrid effect, comprised of popularity bias, exposure bias, and selection bias, makes the profiles generated by LLM inherently carry systemic bias. This bias is further amplified by the recommendation model, forming a vicious cycle of "bias → profile contamination → recommendation bias → more bias," exacerbating the Matthew effect and harming the interests of long-tail content creators.

[0004] Third, there is the mismatch between the semantic space and the recommendation space. LLM is pre-trained on a large-scale corpus, and its understanding of "similarity" mainly comes from lexical co-occurrence and contextual semantics, while recommendation models learn "behavioral similarity" from user-item interaction graphs. Although these two types of similarity overlap, they are fundamentally different. If the semantic embeddings generated by LLM are directly injected into the recommendation model without an explicit alignment mechanism, the two will "talk to themselves," leading to mutual interference and even conflict between semantic and behavioral signals.

[0005] In summary, current LLM-enhanced recommendation systems have limitations in terms of dynamic adaptability, causal fairness, spatial alignment, and engineering feasibility. To address these limitations, this invention proposes a causal dynamic semantic alignment method, aiming to achieve self-evolution, self-purification, and self-alignment of user profiles through iterative correction, causal decoupling, and semantic alignment. Summary of the Invention

[0006] This invention aims to solve the problems existing in the prior art and provide a method for constructing and using user profiles that can be dynamically updated, remove biases, and be semantically aligned with the recommendation target.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a causal dynamic semantic alignment method for LLM-enhanced recommender systems, comprising the following steps: Step 1: Constructing the initial user profile based on confounding factor awareness: A debiased suggestion template is constructed, explicitly requiring the large language model to ignore external interference factors such as popularity, exposure frequency, brand, or price when generating user profiles, focusing only on intrinsic content attributes such as theme, style, and sentiment. Subsequently, the generated natural language profiles are encoded into high-dimensional vectors through a pre-trained embedding model, and then reduced to a predetermined dimension using principal component analysis (PCA) to obtain the initial user profile vector. At the same time, three confounding factor vectors—popularity, activity, and category entropy—are calculated based on users' historical interaction data as supervision signals.

[0008] Step Two: Dual Causal Decoupling Mechanism: Through the synergy of adversarial training and invariant risk minimization (IRM), false information related to confounding factors is removed from the initial profile, extracting a pure causal representation. Specifically, this includes: 1. Adversarial debiasing: Construct a causal network (multilayer perceptron MLP) to extract causal representations from the profile, and introduce a discriminator network to try to predict confounding factors from the causal representations; through min-max adversarial optimization, force the output causal representations to be independent of the confounding factor statistics; 2. Environment Invariance Constraint: Users are divided into three "environments"—high, medium, and low—based on popularity. Based on the Invariant Risk Minimization (IRM) principle, the causal representation is required to have consistent gradient directions for the recommendation task in different environments, thereby improving the model's generalization ability across environments. Through the synergy of adversarial training and IRM, it is ensured that the causal representation only encodes semantic information related to real user preferences.

[0009] Step 3, Iterative Image Correction and Semantic Alignment Training: A multi-round offline iterative framework is adopted, with the following sub-steps executed in each round: Training and Alignment: The current causal representation is jointly trained with the recommendation model, and semantic alignment loss is introduced. KL divergence is used to align the semantic similarity distribution generated by LLM with the recommendation score distribution output by the recommendation model, so that the recommendation results are close to the true interests expressed semantically. Profile Refinement: Based on the embedding of the current recommendation model, retrieve Top-K items, construct reverse prompts emphasizing "de-popularity", and call LLM again to generate more accurate profile text; Fusion Update: The newly generated profile text is input into the causal network to extract new causal representations, and the exponential moving average (EMA) method is used to smoothly update the profile, thereby realizing the dynamic evolution and semantic refinement of the profile.

[0010] Through multiple iterations, the semantic granularity of user profiles is refined round by round, and the correlation with confounding factors such as popularity is continuously reduced.

[0011] Step 4: Lightweight deployment with zero online LLM overhead: After offline iterative optimization, all LLM-related components are discarded, retaining only the trained recommendation model and its user and item embedding vectors. During online inference, only embedding lookup and inner product calculation are required, without calling LLM, achieving high concurrency and low latency industrial-grade deployment.

[0012] Preferably, in step one, the specific calculation method for the confounding factor vector is as follows: Definition 1: User profile vector. A user profile is encoded using natural language descriptions generated by a large language model, denoted as , where is the profile dimension. Item profiles are defined similarly.

[0013] Definition 2: Confounding Factors. Confounding factors refer to variables that are unrelated to actual user preferences but have a spurious impact on observed interaction behavior, denoted as a vector. .

[0014] include: Popularity:

[0015] Activity level:

[0016] Category entropy:

[0017] in This is a collection of items that user u has interacted with historically. This represents the percentage of interactions for category k.

[0018] A bias-removing prompt template is constructed, explicitly requiring the large language model to ignore external interference factors such as popularity, exposure frequency, brand, or price when generating user profiles, focusing only on intrinsic content attributes such as theme, style, and sentiment. The generated natural language profiles are then encoded into high-dimensional vectors using a pre-trained embedding model (e.g., text-embedding-ada-002), and principal component analysis (PCA) is used to reduce the dimensionality to 128 dimensions, yielding the initial user profile vector. ; Calculate a three-dimensional confounding factor vector containing popularity, activity, and category entropy based on user historical interactions. This provides key monitoring signals for subsequent causal decoupling.

[0019] Preferably, in step two, the dual causal decoupling mechanism specifically refers to: Definition 3. Causal representation: Causal representation refers to features extracted from an image that are statistically independent of confounding factors, denoted as... Through causal networks (Multilayer Perceptron (MLP)) generate.

[0020] Definition 4 Discriminator Network: Construct a discriminator to attempt to predict confounding factors from causal representations. .

[0021] Adversarial debiasing: Introducing a discriminator D to attempt to predict confounding factors from causal representations. , and causal network Then, through min-max adversarial optimization, the discriminator maximization stage: the discriminator parameters are fixed, and the discriminator is trained to make the prediction more accurate; the causal network minimization stage: the discriminator parameters are fixed, and the causal network is trained to "deceive" the discriminator. The two stages are executed alternately, forcing the output causal representation to be... and Statistical irrelevant.

[0022]

[0023] Definition 5, the Invariant Risk Minimization (IRM) principle, states that true causality should remain stable under different environments, while spurious correlations will collapse as the environment changes.

[0024] Environment Invariance Constraint (IRM): Users are divided into three "environments" of high, medium and low popularity. The causal representation is required to have consistent gradient directions for the recommendation task in different environments, thereby improving the model's generalization ability across environments.

[0025] The two work together to ensure Only semantic information related to actual preferences is encoded.

[0026] Preferably, the semantic alignment training in step three specifically includes: Phase 1 training: training the current causal representation (... It is jointly trained with recommendation models (such as LightGCN) and the semantic similarity distribution and recommendation score distribution are aligned using KL divergence. Definition 6: Recommendation model loss. For example, LightGCN uses BPR loss for training. For each user u, if they click on item i but not item j, then the model should satisfy... ( If user u's predicted score for item i is given, then the recommendation loss is: .

[0027] Definition 7 Semantic alignment utilizes KL divergence to align two probability distributions, with the semantic similarity distribution generated by LLM being the key difference. First, the cosine similarity between the user and each item is calculated. Then, the softmax function is used to transform these similarities into a probability distribution. The recommendation score distribution output by the recommendation model is then calculated. Recommendation models (such as LightGCN) generate embedding vectors for user u and each item i. The user's predicted score for an item is the inner product, which is also transformed into a probability distribution using softmax. The semantic alignment loss is then...

[0028] Phase 2 Revision: Based on the current recommended embedding ( , Retrieve Top-K items, construct biased inverse suggestions (e.g., "The recommended model is closest to these low-to-medium popularity but high-rated items, please adjust the profile accordingly"), and call LLM to generate more accurate profile text; Phase 3 fusion: The new profile is fed back into the causal network and updated smoothly using an exponential moving average (EMA). To prevent optimization oscillations.

[0029] This mechanism enables user profiles to "self-evolve" as the recommendation learning process progresses, with semantic granularity being refined round by round.

[0030] .

[0031] Preferably, step four specifically involves: after completing T iterations, completely discarding all LLM-related components (including prompt templates, causal networks, discriminators, and embedding encoders), retaining only the finally trained lightweight recommendation model and its user / item embeddings. , During online inference, the system only needs to perform table lookup and inner product calculation: No LLM calls are required.

[0032] The causal dynamic semantic alignment method for LLM-enhanced recommender systems proposed in this invention achieves the following beneficial effects: 1. Through multiple rounds of training and correction iterations, the user profiles generated by LLM can be refined synchronously as the recommendation model deepens its understanding of user preferences, completely solving the cognitive misalignment problem caused by the solidification of static profiles and improving the timeliness and effectiveness of semantic information.

[0033] 2. By combining bias-removing prompt templates with a dual causal decoupling mechanism, this system intervenes for the first time from both the language generation source and representation learning level, effectively removing confounding factors such as popularity and exposure bias from the user profile, breaking the bias amplification cycle, and significantly improving the fairness of the recommendation system and its ability to cover long-tail content.

[0034] 3. By designing a semantic alignment loss based on KL divergence, the learning direction of the recommendation model is explicitly constrained, so that its output distribution is consistent with the deep semantic similarity distribution expressed by LLM. This solves the problem of mismatch between semantic space and recommendation space, and enables externally introduced semantic knowledge to effectively guide the recommendation goal.

[0035] 4. It adopts an architecture of offline iterative optimization and online lightweight inference. After training is completed, the computationally expensive LLM component is completely eliminated. The online service only relies on efficient inner product operations, achieving zero additional LLM inference cost, while ensuring low latency and high throughput, and is easy to integrate into existing recommendation system infrastructure. Attached Figure Description

[0036] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall process of the method of the present invention. Detailed Implementation

[0037] The technical inventions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0038] It should be noted that the terms “front,” “back,” “left,” “right,” “up,” and “down” used in the following description refer to the directions shown in the attached diagram, while the terms “inside” and “outside” refer to the directions toward or away from the geometric center of a specific component, respectively.

[0039] This invention provides a causal dynamic semantic alignment method for LLM-enhanced recommender systems, the overall process of which is as follows: Figure 1 As shown below, the invention will be described in detail with a specific embodiment of a book recommendation scenario: Example: This embodiment is implemented on an online reading platform, aiming to provide users with personalized book recommendations. The platform possesses user interaction data such as historical clicks, reading activity, and ratings.

[0040] Taking user "Xiao Li" as an example, his historical interaction records include popular science fiction works such as "The Three-Body Problem", "The Wandering Earth", "Dune" and "Foundation".

[0041] Step 1: Initialization—Building the Semantic Starting Point Based on users' historical interaction behavior (such as the types of restaurants or cuisines visited), structured cue words are constructed and a Large Language Model (LLM) is invoked to generate a user interest profile in natural language form. Subsequently, a text embedding model (such as text-embedding-ada-002) is used to encode this profile into a vector, and the dimensionality is reduced to 128 dimensions as the initial semantic representation. Simultaneously, three confounding factors—popularity, activity, and category entropy—are explicitly calculated to characterize structural biases in the data. This step not only solves the cold-start problem of traditional collaborative filtering in sparse scenarios but also provides an interpretable semantic foundation and supervision signals for subsequent causal debiasing.

[0042] Taking the user "Xiao Li" on the reading platform as an example, his historical activity only includes popular science fiction works such as *The Three-Body Problem*, *The Wandering Earth*, *Dune*, and *Foundation*. If relying solely on collaborative filtering, the system would simply categorize him as a "popular science fiction enthusiast." However, by constructing debiased prompts (such as "ignore the popularity of the work, focus only on the theme and philosophical speculation"), the LLM-generated profile is: "Preference for hard science fiction works with a grand cosmic view, strong scientific logic support, and integration of civilizational reflection and existentialist themes." After embedding and dimensionality reduction, the text forms an initial semantic vector. At the same time, the system calculates its popularity as high as 0.92 and its category entropy as low as 0.10, accurately quantifying the structural bias in the data.

[0043] Step 2: First Iteration – Closed-Loop Feedback and Causal Intervention First, the LightGCN recommendation model is trained using the initial causal decoupling profile as input, and semantic alignment loss is introduced to constrain its learning direction, making its recommendation results closer to the true interests expressed by LLM. Then, low-popularity recommendations are selected from the Top-K output, and reverse prompts emphasizing "de-popularity" are constructed to guide LLM to dynamically correct user profiles, generate more accurate and de-biased natural language descriptions, and re-encode them. Finally, the new profile is fed back into the causal decoupling module to filter out information related to confounding factors, and is updated smoothly through exponential moving average (EMA) to form the causal interest vector after the first round of optimization, which significantly reduces the correlation between profile and popularity.

[0044] In "Xiao Li's" case, the recommended list after the first round of training not only included popular sequels like *The Three-Body Problem 2*, but also high-scoring works with moderate popularity such as *Solaris* and *Story of Your Life*. The system automatically filtered these "high-value long-tail items" and constructed a reverse prompt: "The current recommendations involve topics such as 'language relativity' and 'nonlinear time.' Please reassess whether the user is being overshadowed by popular IPs?" Based on this, LLM output a new profile: "Those truly interested in philosophical science fiction that explores language, time perception, and the limitations of human cognition through extraterrestrial contact." After causal decoupling and EMA updates, the user indicated a reduced correlation with popularity, achieving effective de-biasing for the first time.

[0045] Step 3: Second and Third Rounds of Iteration and Interest Refinement The closed-loop process of "adversarial debiasing → training the recommendation model → generating recommendations → LLM dynamic profile correction → EMA smooth update" is repeatedly executed, continuously refining the expression of user interests in multiple rounds of interaction. Each iteration makes the semantics of the profile more focused (e.g., from "likes Sichuan cuisine" to "prefers the spicy and numbing layers of stir-fried Sichuan dishes"), while further weakening the correlation with confounding factors such as popularity. After 2-3 rounds, the user profile is almost completely decoupled from exposure bias, and the correlation drops to near zero, thereby effectively improving the ability to discover long-tail, niche but semantically matching items.

[0046] In the second round, "Xiao Li's" profile was further refined to "a science fiction reader who prefers to explore the nature of consciousness and the ethics of non-carbon-based intelligence," and in the third round, it stabilized as "a philosophical science fiction reader who focuses on the triangular relationship between language, time, and consciousness." Simultaneously, the proportion of underrated gems in the recommendation list increased, while the correlation between the profile and popularity decreased significantly, almost completely decoupling. At this point, the system could reliably recommend highly relevant but previously undiscovered underrated works such as *Blindness* and *Breathing*, significantly improving its long-tail discovery capabilities and user satisfaction.

[0047] Step 4: Deployment – ​​Industrial-Friendly Design After offline iterative optimization, only the finally converged user and item embedding vectors are saved, and complex components such as LLM calls, hint engineering, and causal networks are completely removed. In the online inference stage, only efficient inner product calculation and Top-K sorting are required, without the need for any external APIs or large models. This design achieves zero LLM inference cost, is fully compatible with existing recommendation system architectures (such as FAISS or Redis vector retrieval), and has good engineering feasibility.

[0048] For "Xiao Li," after three rounds of offline iterations, the system only retains Xiao Ming and the book vectors. When he refreshes the recommendation page in the app, the system directly returns accurate results such as "Solaris" and "The Story of Your Life" through inner product calculation, without any LLM calls or external dependencies.

[0049] Experimental verification: Comparative experiments on public datasets show that the recommendation system using the method of this invention significantly improves core recommendation accuracy metrics (approximately 8%-12%) compared to traditional static LLM profile enhancement methods. It also improves long-tail item coverage by approximately 15% and reduces online inference latency by more than 90%, fully verifying the comprehensive advantages of this invention in terms of effectiveness, fairness, and efficiency.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, material, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, material, or apparatus.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A causal dynamic semantic alignment method for LLM-enhanced recommender systems, characterized in that, Includes the following steps: Step 1: Initial User Profile Construction Based on Confounding Factor Perception: Construct a debiased prompt template, explicitly requiring the large language model to ignore external interference factors such as popularity, exposure frequency, brand, or price when generating user profiles, focusing only on intrinsic content attributes such as theme, style, and sentiment. The generated natural language profile is then encoded into a high-dimensional vector using a pre-trained embedding model, and principal component analysis (PCA) is used to reduce its dimensionality to a predetermined dimension, obtaining the initial user profile vector. Simultaneously, three types of confounding factor vectors—popularity, activity, and category entropy—are calculated based on user historical interaction data as supervision signals. Step 2: Dual Causal Decoupling Mechanism: Through the synergy of adversarial training and invariant risk minimization (IRM), false information related to confounding factors is removed from the initial profile, extracting a pure causal representation. Specifically, this includes:

1. Adversarial Debiasing: Constructing a causal network (Multilayer Perceptron MLP) to extract causal representations from the profile and introducing discriminant... The network attempts to predict confounding factors from the causal representation; through min-max adversarial optimization, the output causal representation is forced to be independent of the confounding factor statistics; 2. Environment invariance constraint: users are divided into three groups of "environments" of high, medium and low popularity. Based on the principle of invariant risk minimization (IRM), the gradient direction of the causal representation for the recommendation task is required to be consistent in different environments, thereby improving the model's generalization ability across environments; through the synergy of adversarial training and IRM, it is ensured that the causal representation only encodes semantic information related to real user preferences; Step 3, iterative profile correction and semantic alignment training: a multi-round offline iterative framework is adopted, and the following sub-steps are executed in each round:

1. Training and alignment: the current causal representation and the recommendation model are jointly trained, and semantic alignment loss is introduced. KL divergence is used to align the semantic similarity distribution generated by LLM with the recommendation score distribution output by the recommendation model, so that the recommendation result is close to the real interest of semantic expression; 2. Profile Correction: Based on the embedding of the current recommendation model, retrieve Top-K items, construct reverse prompts emphasizing "de-popularity", and call LLM again to generate more accurate profile text; 3. Fusion Update: Input the newly generated profile text into the causal network to extract new causal representations and perform smooth updates to achieve dynamic evolution of user profiles; through multiple iterations, the semantic granularity of user profiles is refined round by round, and the correlation with confounding factors such as popularity is continuously reduced; Step 4, Lightweight Deployment with Zero Online LLM Overhead: After completing offline iterative optimization, discard all LLM-related components and retain only the trained recommendation model and its user and item embedding vectors; during online inference, only embedding table lookup and inner product calculation are required, without calling LLM, achieving high concurrency and low latency industrial-grade deployment.

2. The causal dynamic semantic alignment method for LLM-enhanced recommender systems according to claim 1, characterized in that: In step one: the specific calculation method for the confounding factor vector is as follows: Definition 1 User profile vector: The user profile is encoded by natural language description generated by a large language model, denoted as , where is the profile dimension; Item portraits are similarly defined as; Definition 2: Confounding factor, denoted as a vector. ; include: Popularity: Activity level: Category entropy: ,in This is a collection of items that user u has interacted with historically. The percentage of interactions for category k; A bias-free suggestion template is constructed, focusing only on intrinsic content attributes such as theme, style, and sentiment. The generated natural language profile is then encoded into a high-dimensional vector using a pre-trained embedding model, and principal component analysis (PCA) is used to reduce its dimensionality to 128 dimensions, yielding the initial user profile vector. ; A three-dimensional confounding factor vector, comprising popularity, activity, and category entropy, is calculated based on user historical interaction data. This provides a monitoring signal for causal decoupling.

3. The causal dynamic semantic alignment method for LLM-enhanced recommender systems according to claim 1, characterized in that: In step two, the dual causal decoupling mechanism is specifically defined as follows: Definition 3: Causal representation refers to the feature extracted from the image that is statistically independent of confounding factors, denoted as... Through causal networks (Multilayer Perceptron (MLP)) Generate; Define 4 Discriminator Network, construct a discriminator, and attempt to predict confounding factors from causal representations. Adversarial debiasing: A discriminator D is introduced to attempt to predict confounding factors from causal representations. causal network Through min-max adversarial optimization, the discriminator maximization phase involves fixing the causal network parameters and training the discriminator to make accurate predictions. The causal network minimization phase involves fixing the discriminator parameters and training the causal network to "deceive" the discriminator. These two phases are executed alternately, forcing the output causal representation to be... and Statistically irrelevant; Definition 5, the Invariant Risk Minimization (IRM) principle, states that true causal relationships should remain stable across different environments, while spurious correlations will collapse with environmental changes. Environment Invariance Constraint (IRM) categorizes users into high / medium / low "environments" based on popularity, requiring that the gradient direction of the causal representation for the recommendation task remains consistent across different environments. This improves the model's generalization ability across environments. The two work synergistically to ensure... Only semantic information related to actual preferences is encoded.

4. The causal dynamic semantic alignment method for LLM-enhanced recommender systems according to claim 1, characterized in that: The semantic alignment training described in step three specifically involves: Phase 1 training: The current causal representation ( Jointly trained with recommendation models (such as LightGCN), the semantic similarity distribution and recommendation score distribution are aligned using KL divergence; Definition 6: Recommendation model loss. For example, LightGCN uses BPR loss for training. For each user u, if they click on item i but not item j, then the model should satisfy... ( If user u's predicted score for item i is given, then the recommendation loss is: ; Definition 7 Semantic alignment utilizes KL divergence to align two probability distributions, with the semantic similarity distribution generated by LLM being the key difference. First, the cosine similarity between the user and each item is calculated. Then, the similarity is transformed into a probability distribution using the softmax function. The recommendation score distribution output by the recommendation model is then used. Recommendation models (such as LightGCN) generate embedding vectors for user u and each item i. The user's predicted score for an item is the inner product, which is also transformed into a probability distribution through softmax. The semantic alignment loss is then... ; Phase 2 Revision: Based on the current recommended embedding ( , Retrieve Top-K items, construct debiased inverse hints, and call LLM to generate more accurate profile text; Phase 3 fusion: Input the new profile back into the causal network and use exponential moving average (EMA) for smoothing updates. To prevent optimization oscillations, this mechanism enables user profiles to "self-evolve" during the recommendation learning process, with semantic granularity being refined round by round. 。 5. The causal dynamic semantic alignment method for LLM-enhanced recommender systems according to claim 1, characterized in that: Step four specifically involves: after completing T iterations, completely discarding all LLM-related components, retaining only the finally trained lightweight recommendation model and its user / item embeddings. , During online inference, the system only needs to perform table lookups and inner product calculations. .

6. The causal dynamic semantic alignment method for LLM-enhanced recommender systems according to claim 1, characterized in that: The recommendation model is a recommendation model based on embedding representation.

7. The causal dynamic semantic alignment method for LLM-enhanced recommender systems according to claim 1, characterized in that: The bias-reduction prompt template constructed in step one and the reverse prompt constructed in step three are configured to guide the large language model to focus on the intrinsic content attributes of user preferences, while ignoring the popularity, exposure frequency and commercial attributes of items.

8. The causal dynamic semantic alignment method for LLM-enhanced recommender systems according to claim 1, characterized in that: In step three, the exponential moving average method is used to smoothly update the new causal representation extracted in each iteration.

Images