Multi-scene user controllable personalized recommendation method and system based on super network adaptation

Abstract

The invention provides a multi-scene user controllable personalized recommendation method and system based on super network adaptation. The method comprises the following steps: acquiring a natural language query input by a user; mapping the natural language query into a control condition text based on a large language model; encoding the control condition text into a condition embedding vector by using a pre-trained text encoder; inputting the condition embedding vector into a super network, and generating an adapter parameter matrix for at least one dense layer in the basic recommendation model by the super network; adding the adapter parameter matrix and an original weight matrix of a corresponding dense layer in the basic recommendation model to obtain an adapted recommendation model; and using the adapted recommendation model to predict preference scores of the user for the articles, and generating a recommendation list based on the preference scores. According to the technical scheme, the super network is introduced to dynamically generate the lightweight adapter, so that the limitation of a traditional recommendation system in the aspects of user controllability, scene adaptability and semantic comprehension is effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of recommendation systems, specifically a multi-scene user controllable personalized recommendation method and system based on super network adaptation. BACKGROUND

[0002] As a key information filtering technology, recommendation systems have been widely applied in e-commerce, social media, content platforms, etc. Its core task is to actively recommend information (such as goods, videos, information, etc.) that may be of interest to users by analyzing their historical behavior data, interest preferences, etc.

[0003] Traditional recommendation systems, such as collaborative filtering, matrix decomposition or deep learning-based models, mainly rely on historical interaction data between users and items to build representations, and then match high-relevance recommendation results. However, this kind of method has the following significant shortcomings when dealing with complex and changing user needs in the real world: Dynamicity of user interest is difficult to capture: User interest preferences are not constant, but change dynamically with time, place, situation, and even mood. Traditional models trained based on static historical data are difficult to capture and respond to users' immediate and situational needs in real time.

[0004] Lack of transparency and controllability in the recommendation process: The recommendation results generated by the system are often like a "black box", and users do not know why they see certain recommendations, and cannot directly intervene in the recommendation logic to meet their specific or temporary needs. This weakens users' trust and autonomy.

[0005] In order to solve the above problems, the academic and industrial circles have begun to explore user-controllable recommendation systems. This kind of system allows users to directly express their recommendation preferences through natural language queries, operation interfaces, etc., thereby affecting the final recommendation list.

[0006] The closest prior art to the present scheme is a user-controllable recommendation system based on pre-defined profiles and categories, which pre-constructs a structured user profile (such as age, gender, occupation, etc. Demographic characteristics) and item label system (such as movie genre, music genre, product category, etc.). In the recommendation process, users are allowed to check, filter or adjust the weight of these pre-defined labels through a graphical interface. For example, users can request "increase the proportion of science fiction movies" or filter "animation suitable for children to watch". After receiving these instructions, the system modifies the internal ranking algorithm or uses post-adjustment strategies to rearrange the initial recommendation list, so that the result is more in line with the user's explicit control intent.

[0007] Technical shortcomings: Although this kind of method gives users some control to some extent, it still has the following fundamental limitations: Lacking semantic understanding capabilities, this method relies heavily on a predefined, structured tagging system. However, in reality, item categorization and attribute labeling are often incomplete, failing to cover diverse user semantic preferences. For example, for natural language queries containing complex, implicit semantics, such as "a heartwarming movie suitable for the whole family to watch on the weekend," the system cannot accurately understand the true intent behind it because it cannot directly map the query to any predefined tag.

[0008] Poor scenario adaptability and limitation to a single dimension: Most existing methods are designed for a single recommendation scenario (such as movies or music), and the control dimension is limited to the content attributes of items. However, users' real needs are often complex needs that cross scenarios and multiple dimensions, which may simultaneously involve content themes ("inspirational"), time context ("suitable for watching alone at night"), social preferences, etc. Existing solutions cannot model such complex, cross-scenario semantic preferences.

[0009] Insufficient generalization ability makes it difficult to handle novel queries: User query combinations are open and novel. For query combinations that have not appeared in the training data (such as "classical music podcasts suitable for listening to on the commute"), the existing system struggles to generate reasonable control strategies due to a lack of deep understanding of query semantics and generalization ability, resulting in a significant decline in recommendation quality.

[0010] Lack of online dynamic adaptability: Most existing methods are applied to offline recommendation scenarios, that is, generating one-time recommendation results based on users' static control options, which is difficult to adapt to the dynamic changes in user preferences and intentions in online scenarios.

[0011] Therefore, how to design a system that can accurately understand the complex semantics behind users' natural language queries, adapt to the complex needs of multiple scenarios, and have good generalization ability to achieve truly intelligent, flexible, and user-controllable recommendations is a technical challenge that urgently needs to be solved in the current technology field. Summary of the Invention

[0012] The problem addressed by this invention is how to achieve controllable recommendations that accurately understand the complex semantics behind users' natural language queries.

[0013] To address the aforementioned issues, this invention provides a multi-scenario user-controllable personalized recommendation method, recommendation system, electronic device, and storage medium based on hyper-network adaptation.

[0014] In a first aspect, the present invention provides a multi-scenario user-controllable personalized recommendation method based on hypernetwork adaptation, including: Obtain the natural language query input by the user; The natural language query is mapped to control condition text based on a large language model; The control condition text is encoded into a conditional embedding vector using a pre-trained text encoder; The conditional embedding vector is input into the hypernetwork, which generates an adapter parameter matrix for at least one dense layer in the base recommendation model. The adapter parameter matrix is ​​added to the original weight matrix of the corresponding dense layer in the basic recommendation model to obtain the adapted recommendation model. Using the adapted recommendation model, the user's preference score for items is predicted, and a recommendation list is generated based on the preference score.

[0015] Optionally, the control condition text is mapped from a predefined multi-scenario control intent corpus, which contains topic-based, time-based, and social-based control intents.

[0016] Optionally, the supernetwork includes a parameter generation block, which is a multilayer perceptron network, used to map the projected conditional embedding vector into a vector and reshape the vector into the adapter parameter matrix.

[0017] Optionally, during the training phase, the last layer of the supernetwork is initialized to zero so that the adapter parameter matrix is ​​a zero matrix at the start of training.

[0018] Optionally, the model involved in the method is trained through a semantically aware optimization objective, the loss function of which includes: The first optimization term is used to ensure that, under specific control conditions, the user's predicted score for positive sample items is higher than the predicted score for negative sample items. The second optimization term is used to introduce a semantic matching signal by comparing the predicted scores of the same negative sample item under different control conditions.

[0019] Optionally, the loss function is as follows: ; In the formula, This refers to items that the user has interacted with. This indicates the specific control conditions associated with the item. This refers to negative sample items sampled from a set of items that the user has not interacted with. Indicates under control conditions Next user For positive sample items The predicted score Indicates under control conditions Next user For negative sample items The predicted score From the set of control conditions Different control conditions for sampling It is the sigmoid function. Used to balance the contribution of the semantic contrast term in the total loss formula.

[0020] Optionally, the hypernetwork generates a corresponding adapter parameter matrix for each dense layer of the base recommendation model.

[0021] Secondly, embodiments of the present invention provide a multi-scenario user-controllable personalized recommendation system based on hypernetwork adaptation, comprising: The query receiving module is used to obtain natural language queries input by the user; The intent extraction module is used to map the natural language query into control condition text based on a large language model; A conditional embedding module is used to encode the control condition text into a conditional embedding vector using a pre-trained text encoder; An adapter generation module is used to input the conditional embedding vector into a hypernetwork, and the hypernetwork generates an adapter parameter matrix for at least one dense layer in the base recommendation model. The model adjustment module is used to add the adapter parameter matrix to the original weight matrix of the corresponding dense layer in the basic recommendation model to obtain the adapted recommendation model. The recommendation generation module is used to predict the user's preference score for items using the adapted recommendation model, and generate a recommendation list based on the preference score.

[0022] Thirdly, embodiments of the present invention provide an electronic device, including a processor, a communication interface, a memory, and a bus, wherein the processor, the communication interface, and the memory communicate with each other through the bus, and the processor can call logical instructions in the memory to execute the steps of the method provided in the first aspect.

[0023] Fourthly, embodiments of the present invention provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the multi-scenario user-controllable personalized recommendation method based on hypernetwork adaptation as described in the first aspect.

[0024] The beneficial effects of the multi-scenario user-controllable personalized recommendation method based on hypernetwork adaptation of the present invention are as follows: This invention effectively addresses the limitations of traditional recommender systems in terms of user controllability, scenario adaptability, and semantic understanding by introducing a hypernetwork-based dynamic generation of lightweight adapters. First, the system accurately understands the complex semantic intent in users' natural language queries, transforming abstract requirements into precise adjustments to the recommender model through a pre-trained language model and hypernetwork adaptation mechanism. Second, employing a parameter-efficient adapter structure, it maintains the personalization capabilities of the base model while requiring only a few additional parameters to flexibly adapt to various recommendation scenarios, significantly improving the system's practicality and scalability. Finally, through a semantically aware contrastive learning objective and a two-stage inference framework, the system exhibits powerful zero-shot and combinatorial generalization capabilities, maintaining stable and excellent recommendation performance even when faced with novel queries or complex conditions not encountered during training. Attached Figure Description

[0025] Figure 1 This is a flowchart of a multi-scenario user-controllable personalized recommendation method based on hypernetwork adaptation in an embodiment of the present invention; Figure 2 This is a schematic diagram of the HGA framework in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hypernetic network generation adapter parameters in an embodiment of the present invention; Figure 4 This is a performance comparison chart of the Yelp dataset HGA and its variants over 50 training epochs in an embodiment of the present invention. Figure 5 The figure shows a performance comparison of the KuaiRec dataset HGA and its variants over 50 training epochs in this embodiment of the invention; Figure 6 This is a performance comparison chart for zero-sample generalization in this embodiment of the invention; Figure 7 This is a performance comparison chart under combined generalization in the embodiments of the present invention; Figure 8 This is a structural block diagram of a multi-scenario user-controllable personalized recommendation system based on hypernetwork adaptation, as described in an embodiment of the present invention. Figure 9 This is a structural block diagram of the electronic device in an embodiment of the present invention. Detailed Implementation

[0026] To better understand the purpose, technical solution, and advantages of this application, the application is described and explained below in conjunction with the accompanying drawings and embodiments.

[0027] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by one of ordinary skill in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.

[0028] To better understand the technical solution of this invention, it is first briefly described that this invention proposes a novel multi-scenario user-controllable recommendation framework based on Hypernetwork-Generated Adapters (HGA). This framework utilizes hypernetwork technology to dynamically generate corresponding lightweight adapter parameters for different recommendation control intentions. By modulating the parameters of key layers of the basic recommendation model, it introduces scenario-specific control signals while preserving the model's inherent personalized capabilities, thereby achieving flexible multi-scenario user-controllable recommendations. The overall structure of the HGA framework is as follows: Figure 2 As shown, the HGA framework will be described together with the method proposed in this invention.

[0029] like Figure 1 As shown in the figure, a multi-scenario user-controllable personalized recommendation method based on hypernetwork adaptation is provided in this embodiment of the invention, including: S1. Obtain the natural language query input by the user; S2. Based on the large language model, map the natural language query into control condition text; The control condition text is mapped from a predefined multi-scenario control intent corpus, which contains topic-based, time-based, and social-based control intents.

[0030] Specifically, the HGA framework mainly includes the following key components: Conditional embedding: Each control condition is embedded using a pre-trained text encoder. Mapped to a fixed-dimensional embedding vector In this way, the rich semantics of scene description are effectively captured; ; in, A powerful pre-trained text encoding model (such as BERT) projects text conditions c into a dense semantic vector space. By introducing a powerful language model as a query encoder, HGA can mine the rich semantic information contained in user queries, overcoming the limitations of existing methods based on literal matching.

[0031] To accommodate the diversity of user control intentions, HGA constructs an offline corpus containing multi-scenario control intentions based on domain knowledge during the model training phase. This corpus covers three typical categories of controllable intentions: topic-based, time-based, and social, characterizing users' complex semantic control needs from the perspectives of content items themselves, user behavior patterns, and user social preferences. During the inference phase, HGA utilizes instruction tuning to map users' immediate query requests to this offline, multi-scenario semantic space, enhancing the model's ability to perceive and adapt to control signals of different granularities.

[0032] S3. Encode the control condition text into a conditional embedding vector using a pre-trained text encoder; S4. Input the conditional embedding vector into the hypernetwork, and generate an adapter parameter matrix for at least one dense layer in the basic recommendation model from the hypernetwork. The supernetwork includes a parameter generation block, which is a multilayer perceptron network, used to map the projected conditional embedding vector into a vector and reshape the vector into the adapter parameter matrix.

[0033] During the training phase, the last layer of the supernetwork is initialized to zero so that the adapter parameter matrix is ​​a zero matrix at the start of training.

[0034] The hypernetwork generates a corresponding adapter parameter matrix for each dense layer of the basic recommendation model.

[0035] Furthermore, combined Figure 3 As shown, the HGA framework includes adapter generation based on hypernetworks: given conditional embedding First, the projection vector is obtained through the projection layer. : Use a pre-trained text encoder to process each control condition Mapped to a fixed-dimensional embedding vector In this way, the rich semantics of scene descriptions can be effectively captured.

[0036] ; in The parameter is Multilayer perceptron networks (MLPs) project conditional embeddings into a low-dimensional space.

[0037] Then, for each dense layer of the base model This invention uses a separate hypernetwork. To generate the corresponding adapter parameters : ; ; Each supernetwork From parameters Parameterization, instantiated as an MLP (called a parameter generation block), will Mapping to vector , and They represent the first The output and input dimensions of each dense layer. The generated vector. Then it was reshaped into a matrix. As the basic recommendation model The adapter parameters for each dense layer.

[0038] To enhance training stability, each supernetwork The last layer is initialized to zero, thus The initial adapter parameters are unaffected. This initialization ensures model stability during training and allows for fine-tuning by gradually incorporating control conditions.

[0039] S5. Add the adapter parameter matrix to the original weight matrix of the corresponding dense layer in the basic recommendation model to obtain the adapted recommendation model; S6. Using the adapted recommendation model, predict the user's preference score for items, and generate a recommendation list based on the preference score.

[0040] The model involved in the method is trained through a semantically aware optimization objective, the loss function of which includes: The first optimization term is used to ensure that, under specific control conditions, the user's predicted score for positive sample items is higher than the predicted score for negative sample items. The second optimization term is used to introduce a semantic matching signal by comparing the predicted scores of the same negative sample item under different control conditions.

[0041] Furthermore, the HGA framework also includes adapted preference prediction: adapted dense layer weights. The calculation is performed by adding the generated adapter parameters to the original weights: ; Then, the basic recommender Predict conditional preference scores using adapted weights : ; in, This represents the parameters of the neural network excluding the dense layers.

[0042] It's worth noting that the hypernetwork only generates adapter parameters for the deep, dense layers of the recommendation model, not the embedding layers (which typically contain most of the model capacity). This focuses semantic adaptation on the key layers responsible for learning high-order feature interactions from user and item representations. HGA achieves efficient and flexible adaptation to various scenarios by introducing a small number of additional parameters.

[0043] The key advantage of introducing the adapter structure lies in its ability to maximize the personalized representation capabilities of the original recommendation model while minimizing the introduction of additional parameters. The adapter mechanism preserves the structure and parameters of the model's backbone network, instead fine-tuning them with a small number of generated additional parameters to dynamically activate parts related to control intent in different scenarios. This parameter-efficient and modular design allows HGA to flexibly support user-controllable needs across multiple scenarios within a unified framework without sacrificing recommendation performance.

[0044] To effectively learn adapter parameters, this invention proposes a novel semantic-aware optimization objective. Given user-item interactions... Sample multiple relevant conditions based on item metadata and interaction context. It covers three semantic levels: category level (item side), time level (behavior side), and social level (user side).

[0045] The loss function is defined as follows: ; In the formula, This refers to items that the user has interacted with. This indicates the specific control conditions associated with the item. This refers to negative sample items sampled from a set of items that the user has not interacted with. Indicates under control conditions Next user For positive sample items The predicted score Indicates under control conditions Next user For negative sample items The predicted score From the set of control conditions Different control conditions for sampling It is the sigmoid function. Used to balance the contribution of the semantic contrast term in the total loss formula.

[0046] The loss function above comprises two terms: the first optimizes personalized ranking under specific control conditions, resulting in higher prediction scores for positive samples compared to negative samples; the second, by comparing the relevance of items under different conditions, introduces direct semantic matching signals, enhancing the model's scene recognition capability. Combining these two optimization objectives, HGA can balance personalized recommendation with the satisfaction of control conditions, improving the overall recommendation performance.

[0047] A key challenge in practical recommendations is leveraging the set of control conditions. This allows for semantic adaptation learned from previous learning while handling unconstrained user queries. To bridge this gap, a two-stage inference process is proposed: Query intent extraction: A large language model (LLM) with instruction tuning is used to extract the intent of a specific scenario from the user query q and map it to a set of control conditions. Corresponding conditional text in : ; in Indicates identification query The process of extracting the intent of key semantic aspects expressed in the text.

[0048] Adaptive recommendation for specific scenarios: extracted conditional text This is then processed by HGA to generate personalized recommendations adapted to specific semantic scenarios. This is done using the same text encoder as during training. right The algorithm encodes the network and uses the trained supernetwork to generate corresponding adapter parameters to achieve preference prediction.

[0049] As shown in Tables 1 and 2, the effectiveness of this invention has been verified on two real-world datasets, Yelp and KuaiRec, through extensive experiments. Specific conclusions are as follows: Table 1 Overall performance comparison on the Yelp dataset.

[0050] Table 2 Overall performance comparison on the KuaiRec dataset

[0051] The performance of HGA was evaluated on two real-world datasets, Yelp and KuaiRec, covering different contexts such as time, social, and topical scenarios. Table 1 shows the performance comparison of HGA and baseline methods on the two datasets.

[0052] HGA achieved state-of-the-art performance across all metrics related to user-level personalization and item-level control effectiveness, demonstrating its ability to capture user semantic needs and recommend relevant items while adjusting the recommendation distribution of the base model to meet user requirements. HGA's performance improvement on KuaiRec was particularly significant, likely due to its dense user-item interactions and concentrated category information. Furthermore, HGA maintained high recommendation diversity on Yelp while ensuring strong control effectiveness, highlighting its ability to maintain item diversity while satisfying user semantic needs.

[0053] To verify the effectiveness of the HGA critical design, the following variations were considered: (a) Without adapter: Traditional hypernetwork method, where the generated parameters directly replace the original model parameters.

[0054] (b) No zero initialization: The last layer of the HGA supernetwork uses non-zero initialization.

[0055] (c) Using random semantics: Controls the semantic text to be randomly sampled from C, independent of the training items.

[0056] Figure 4 and Figure 5 This paper presents a performance comparison of HGA and its variants over 50 training epochs on the Yelp and KuaiRec datasets. For the NFM and KGAT backbone models, HGA consistently outperforms all variants across different datasets, demonstrating its superior performance. Variants (a) and (b) are suboptimal compared to HGA, while variant (c) fails to learn an effective control pattern and frequently triggers early stopping. Direct parameter substitution in variant (a) may perturb the original parameter space distribution, impairing personalized recommendation capabilities. Non-zero initialization in variant (b) causes adapter parameters to introduce significant interference to the original model in the early stages of training, affecting the model's learning and convergence speed. Furthermore, the random semantics in variant (c) cannot be integrated with... The data is correlated, which makes it difficult to learn discriminative distributions and results in weak semantic awareness.

[0057] This invention evaluates the generalization ability of HGA from two aspects: zero-shot generalization and combinatorial generalization.

[0058] Zero-shot generalization: This invention evaluates the zero-shot generalization capability of HGA by comparing its performance in two settings: (1) Seen: The semantic text of the query mapping appears in the training control text during inference; (2) Unseen: The semantic text of the mapping does not exist in the training control text. Figure 6 Results are shown on the Yelp and KuaiRec datasets.

[0059] HGA maintains competitive performance even in zero-shot settings, while U-NDCG shows only a slight decrease. HGA consistently outperforms UCI-M in both settings because UCI-M relies on matching to predefined control labels for tuning and lacks zero-shot capability. HGA's ability to learn controllability in the semantic vector space is validated, enabling it to handle a wide range of queries within the contextual semantic space.

[0060] Combinatorial Generalization: This scheme investigates the combinatorial generalization capability of HGA by evaluating its performance under multi-conditional control, where the target item needs to meet multiple criteria. The number of control conditions is varied from 2 (dual) to 3 (triple). Note that during training, each The sample contains only a single control condition, so the composite semantic vector is not seen in the training set. Figure 7 Results are shown on the Yelp and KuaiRec datasets.

[0061] HGA significantly outperforms Pretrain in both dual and triple control settings, especially on KuaiRec, demonstrating its strong combinatorial generalization ability. Although a direct comparison of dual and triple settings is not possible due to the different distribution of target items, HGA's consistent advantage over Pretrain validates HyperAdapter's ability to generalize complex semantics in a predefined semantic space, even without exposure to complex conditions during training. Combined with its zero-shot generalization ability, this suggests that HGA has the potential to learn generalizable semantic awareness in a context space, adapting to diverse semantic needs.

[0062] The following uses an online video recommendation platform as an example to illustrate the implementation of the present invention.

[0063] System environment: The system is deployed on a cloud server and includes a data storage module, a model training module, and an online inference service module.

[0064] Data storage module: Stores user information (ID, viewing history), video information (ID, title, tags, description), and user interaction logs with videos.

[0065] Model training module: responsible for offline training of the recommendation model proposed in this invention.

[0066] Online reasoning service module: responsible for receiving requests from front-end users and returning recommendation results in real time.

[0067] Implementation steps: Model preparation and training; Base Recommendation Model Selection: Choose a mature deep learning recommendation model as the base model, such as the Neural Collaborative Filtering (NCF) model. This model includes a user embedding layer, an item embedding layer, and multiple densely connected layers for learning higher-order interaction features.

[0068] Multi-scenario Corpus Construction: Based on the characteristics of the video platform, a multi-scenario control intent corpus is manually constructed. For example: Thematic categories: "Food making videos", "Travel Vlogs", "Classic science fiction movie explanations".

[0069] Time-related categories: "Suitable for watching while eating", "Relax during lunch break", "Recommended anime to catch up on on weekends".

[0070] Social / Atmosphere Category: "Recently popular on Bilibili", "Suitable for watching with family", "Tear-jerking and touching".

[0071] Model Training: The entire HGA framework is trained end-to-end using the platform's historical interaction data and the constructed corpus. After training, the trained base model parameters, text encoder, all hypernetworks, and large language model intent extractor are deployed to the online inference service module.

[0072] Online recommendations and user interaction; User input: Users can enter natural language queries in the search box or the dedicated "Smart Recommendation" dialog box on the video platform. For example, a user might enter: "Looking for some short, soothing videos that are suitable for watching alone quietly at night."

[0073] Intent Extraction: After receiving the query, the online inference service module first calls the Large Language Model (LLM) to extract intent. The LLM analyzes the query and maps it to the most relevant semantic tags in the corpus. For example, it extracts two control condition texts: "watching at night" and "healing".

[0074] Adapter generation and recommendation: The service module converts the texts "watching at night" and "healing" into semantic vectors using a text encoder.

[0075] These two semantic vectors are fed separately (or in combination) into the deployed hypernetworks to generate two sets (or a fused set) of adapter parameter matrices for each dense layer of the basic recommendation model.

[0076] The generated adapter parameters are added to the original weights of the base model to obtain a temporary "customized" recommendation model that simultaneously satisfies the conditions of "watching at night" and "healing".

[0077] Using this customized model, a recommendation score is calculated for all videos in the candidate video library for the user.

[0078] Results returned: The top-K videos with the highest scores (e.g., the top 10) will be presented to the user as a recommendation list. The videos in this list will be biased towards "healing" content, and the sorting may take into account some characteristics related to the "nighttime viewing" scenario (such as moderate length, quiet atmosphere, etc.).

[0079] Through the above implementation methods, the present invention successfully converts users' natural language commands into precise control over the internal logic of the recommendation model in real time, realizing highly personalized, scenario-based, and controllable recommendation services, and greatly improving the user experience.

[0080] This invention also provides a multi-scenario user-controllable personalized recommendation system based on hypernetwork adaptation. This system is used to implement the above-described method embodiments, and details already described will not be repeated. The terms "module," "unit," and "subunit" used below refer to combinations of software and / or hardware that perform predetermined functions. Although the system described in the following embodiments is preferably implemented in software, hardware implementation or a combination of software and hardware is also possible and contemplated.

[0081] like Figure 8 As shown, Figure 8 This is a structural block diagram of the multi-scenario user-controllable personalized recommendation system based on hypernetwork adaptation in this invention. The system includes: The query receiving module 101 is used to obtain natural language queries input by the user; The intent extraction module 102 is used to map the natural language query into control condition text based on a large language model; Conditional embedding module 103 is used to encode the control condition text into a conditional embedding vector using a pre-trained text encoder; The adapter generation module 104 is used to input the conditional embedding vector into the super network, and the super network generates an adapter parameter matrix for at least one dense layer in the basic recommendation model. The model adjustment module 105 is used to add the adapter parameter matrix to the original weight matrix of the corresponding dense layer in the basic recommendation model to obtain the adapted recommendation model. The recommendation generation module 106 is used to predict the user's preference score for items using the adapted recommendation model, and generate a recommendation list based on the preference score.

[0082] like Figure 9 As shown in the figure, an electronic device provided by an embodiment of the present invention includes: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute the following method: Obtain the natural language query input by the user; The natural language query is mapped to control condition text based on a large language model; The control condition text is encoded into a conditional embedding vector using a pre-trained text encoder; The conditional embedding vector is input into the hypernetwork, which generates an adapter parameter matrix for at least one dense layer in the base recommendation model. The adapter parameter matrix is ​​added to the original weight matrix of the corresponding dense layer in the basic recommendation model to obtain the adapted recommendation model. Using the adapted recommendation model, the user's preference score for items is predicted, and a recommendation list is generated based on the preference score.

[0083] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0084] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0085] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-scenario user-controllable personalized recommendation method based on hypernetwork adaptation, characterized in that, include: Obtain the natural language query input by the user; The natural language query is mapped to control condition text based on a large language model; The control condition text is encoded into a conditional embedding vector using a pre-trained text encoder; The conditional embedding vector is input into the hypernetwork, which generates an adapter parameter matrix for at least one dense layer in the base recommendation model. The adapter parameter matrix is ​​added to the original weight matrix of the corresponding dense layer in the basic recommendation model to obtain the adapted recommendation model. Using the adapted recommendation model, the user's preference score for items is predicted, and a recommendation list is generated based on the preference score.

2. The method according to claim 1, characterized in that, The control condition text is mapped from a predefined multi-scenario control intent corpus, which contains topic-based, time-based, and social-based control intents.

3. The method according to claim 1, characterized in that, The supernetwork includes a parameter generation block, which is a multilayer perceptron network, used to map the projected conditional embedding vector into a vector and reshape the vector into the adapter parameter matrix.

4. The method according to claim 1, characterized in that, During the training phase, the last layer of the supernetwork is initialized to zero so that the adapter parameter matrix is ​​a zero matrix at the start of training.

5. The method according to claim 1, characterized in that, The model involved in the method is trained through a semantically aware optimization objective, the loss function of which includes: The first optimization term is used to ensure that, under specific control conditions, the user's predicted score for positive sample items is higher than the predicted score for negative sample items. The second optimization term is used to introduce a semantic matching signal by comparing the predicted scores of the same negative sample item under different control conditions.

6. The method according to claim 5, characterized in that, The loss function is as follows: ； In the formula, This refers to items that the user has interacted with. This indicates the specific control conditions associated with the item. This refers to negative sample items sampled from a set of items that the user has not interacted with. Indicates under control conditions Next user For positive sample items The predicted score Indicates under control conditions Next user For negative sample items The predicted score From the set of control conditions Different control conditions for sampling It is the sigmoid function. Used to balance the contribution of the semantic contrast term in the total loss formula.

7. The method according to claim 1, characterized in that, The hypernetwork generates a corresponding adapter parameter matrix for each dense layer of the basic recommendation model.

8. A multi-scenario user-controllable personalized recommendation system based on hypernetwork adaptation, characterized in that, include: The query receiving module is used to obtain natural language queries input by the user; The intent extraction module is used to map the natural language query into control condition text based on a large language model; A conditional embedding module is used to encode the control condition text into a conditional embedding vector using a pre-trained text encoder; An adapter generation module is used to input the conditional embedding vector into a hypernetwork, and the hypernetwork generates an adapter parameter matrix for at least one dense layer in the base recommendation model. The model adjustment module is used to add the adapter parameter matrix to the original weight matrix of the corresponding dense layer in the basic recommendation model to obtain the adapted recommendation model. The recommendation generation module is used to predict the user's preference score for items using the adapted recommendation model, and generate a recommendation list based on the preference score.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a bus. The processor, the communication interface, and the memory communicate with each other through the bus. The processor can call logical instructions in the memory to execute the multi-scenario user-controllable personalized recommendation method based on hyper-network adaptation as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the multi-scenario user-controllable personalized recommendation method based on hypernetwork adaptation as described in any one of claims 1 to 7.

Images