Environment diffusion method for user-oriented recommendation fairness
By introducing a time-conditional discriminator and a course learning approach, a calibration dataset was constructed and a diffusion model was trained. This solved the unfairness problem of disadvantaged users caused by sparse user interactions in the recommender system, and improved the performance and fairness of the recommendation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2026-05-07
- Publication Date
- 2026-07-03
AI Technical Summary
Existing recommendation systems ignore the continuous nature of user activity when dealing with sparse user interactions, leading to unfairness within disadvantaged user groups. Furthermore, existing methods struggle to effectively improve recommendation performance and fairness.
A time-conditional discriminator is introduced to construct a calibration dataset. By tracking the forward diffusion noise, high-fidelity representations and degraded representations are distinguished. A course learning based on the degree of degrade is used to train the diffusion model, recover the distribution of user interests, and generate a list of recommended items.
It improves the overall utility and fairness of the recommendation system, and solves the performance limitations caused by the collapse of interest distribution and the unfairness within disadvantaged user groups.
Smart Images
Figure CN122332871A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of computer technology, specifically relating to a research on an environmental diffusion method for user recommendation fairness. Background Technology
[0002] Recommender systems, as fundamental tools for information delivery, aim to predict user interests and maximize the overall utility of recommendations by utilizing user interaction data. However, this global optimization strategy often neglects user-oriented fairness. Specifically, sparsely interacting users cannot provide sufficient signals for learning, leading to degradation in their representations. Consequently, the model performs poorly when serving long-tail users, resulting in unfairness.
[0003] In related technologies, user groups are typically simply divided into a binary advantageous group and an disadvantageous group to reduce the difference in recommendation metrics between the two groups. However, in implementing this application, the inventors discovered that this method ignores the continuous nature of user activity and the resulting representation degradation. Specifically, in terms of evaluation, this masks the unfairness within the large disadvantageous user group; methodologically, it forcibly aligns the degraded representations of disadvantageous users with a limited number of advantageous prototypes, leading to a collapse of interest distribution, limiting the improvement of recommendation performance, and making it difficult to handle intra-group unfairness. Summary of the Invention
[0004] The purpose of this application is to provide a research on an environmental diffusion method for user recommendation fairness, which can solve the recommendation performance limitations caused by the collapse of interest distribution in related technologies, as well as the problem of difficulty in handling unfairness within a large disadvantaged user group.
[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide an environmental diffusion method for user recommendation fairness, the method comprising: A calibration dataset is constructed based on a time-conditional discriminator configured to distinguish the evolutionary differences between high-fidelity representations and degraded representations by tracking increasing forward diffusion noise. The calibration dataset includes the representation state of each user under different degrees of degradation. The diffusion model is trained using the calibration dataset, and the sampling range of the calibration dataset is dynamically expanded in order of increasing degradation level during the model training process. The high-fidelity interest prototypes of the target users are determined using the diffusion model after training, and an item recommendation list is generated based on the high-fidelity interest prototypes.
[0006] Secondly, embodiments of this application provide an environmental diffusion device for recommending fairness to users, the device comprising: A dataset construction module is used to construct a calibration dataset based on a time-conditional discriminator configured to distinguish the evolutionary differences between high-fidelity representations and degraded representations by tracking increasing forward diffusion noise. The calibration dataset includes the representation state of each user under different degrees of degradation. The model training module is used to train a diffusion model using the calibration dataset, and dynamically expands the sampling range of the calibration dataset in order of increasing degradation level during the model training process; The user recommendation module is used to determine the high-fidelity interest prototypes of target users using the diffusion model after training, and to generate a list of recommended items based on the high-fidelity interest prototypes.
[0007] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the environmental diffusion method for user-oriented recommendation fairness as described in the first aspect.
[0008] Fourthly, embodiments of this application provide a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the user-oriented recommendation fairness environment diffusion method described in the first aspect.
[0009] Fifthly, embodiments of this application provide a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the user-oriented recommendation fairness environment diffusion method as described in the first aspect.
[0010] In this embodiment, a temporal conditional discriminator is introduced to distinguish the evolutionary differences between high-fidelity representations and degraded representations by tracking the continuously increasing forward diffusion noise. This takes into account the continuous representation degradation caused by the sparsity of user interactions and accurately measures the degree of degradation of each user representation, thereby constructing a calibration dataset covering all degrees of degradation. Furthermore, an environmental diffusion training process based on the degree of degradation is introduced to learn the distribution of user interests. This enriches the learned distribution using the full dataset without sacrificing its fidelity, enabling the model to adaptively recover degraded user representations, thereby effectively improving overall utility and fairness. Thus, this application can effectively solve the problems of recommendation performance limitations caused by the collapse of interest distribution in related technologies, as well as the difficulty in handling the unfairness within a large disadvantaged user group. Attached Figure Description
[0011] Figure 1A flowchart illustrating the implementation of an environmental diffusion method for ensuring fairness in user recommendations, provided in this application embodiment; Figure 2 A schematic diagram of an environment diffusion framework based on course scheduling provided in an embodiment of this application; Figure 3 A schematic diagram of an environmental diffusion device for recommending fairness to users, provided as an embodiment of this application; Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0013] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.
[0014] First, to facilitate understanding of the technical solutions provided in this application, the main technical concepts involved in the embodiments of this application will be briefly explained below.
[0015] Diffusion model: It is a generative artificial intelligence model based on deep learning. In the forward process, noise is gradually added to the data until it is completely randomized. In the backward process, the neural network (i.e., the denoising network) is trained to learn to gradually remove noise, thereby reconstructing or generating new data samples from random noise.
[0016] In related technologies, there are currently three main solutions to the problem of fairness for users in recommender systems: Re-ranking strategy: As a model-agnostic post-processing step, this method limits the utility gap between the dominant and disadvantaged groups by adjusting the recommendation list.
[0017] Split-Brow Bar Optimization: This method intervenes directly during training by optimizing the worst-case distribution by assigning higher gradients to disadvantaged users, thereby offsetting the dominance of advantageous users in minimizing empirical risk.
[0018] Knowledge transfer method: This method addresses the problem of scarce collaborative signals by aligning the representations of disadvantaged users with the representations of advantageous users whose interaction patterns are similar, thereby transferring high-fidelity user interest knowledge to the disadvantaged group.
[0019] Furthermore, in terms of evaluation systems, traditional methods mostly rely on coarse-grained binary indicators, that is, dividing users into non-overlapping advantageous and disadvantageous groups based on activity levels, and measuring fairness by calculating the average performance difference between these two groups.
[0020] Based on the above analysis, current recommendation systems mostly rely on simple binary partitioning, forcibly dividing users into advantageous and disadvantaged groups. This approach ignores the continuous nature of user interaction frequency. Specifically, at the evaluation level, this masks the significant intra-group unfairness caused by varying degrees of representation degradation within the large disadvantaged group. At the methodological level, re-ranking and re-weighting methods struggle to fundamentally recover the collaborative information lost due to data sparsity, limiting the overall improvement in recommendation utility. Knowledge transfer methods fail to fully utilize the residual signals retained in degraded representations, forcibly aligning a large number of disadvantaged user representations with a limited number of advantageous prototypes. This causes the learned interest distribution to collapse towards the preferences of a minority group, restricting the improvement of recommendation performance. Furthermore, because it treats all disadvantaged users equally, it also struggles to effectively address intra-group unfairness.
[0021] To address the aforementioned issues, this application provides a research on an environmental diffusion method for user recommendation fairness. It introduces an environmental diffusion framework based on course scheduling (which can be called FairDiff), uses environmental diffusion to learn the distribution of user interests, and enriches the learned distribution with the full set of data without sacrificing its fidelity. This enables the model to adaptively recover degraded user representations, thereby effectively improving overall utility and fairness.
[0022] The following, in conjunction with the accompanying drawings, provides a detailed description of a user-oriented recommendation fairness environmental diffusion method provided by the embodiments of this application, through specific implementations and application scenarios.
[0023] See Figure 1 The diagram shown is an implementation flowchart of an environmental diffusion method for user recommendation fairness provided in this application embodiment. The method may include the following steps: Step S101: Construct a calibration dataset based on a time-conditional discriminator, wherein the time-conditional discriminator is configured to distinguish the evolutionary differences between high-fidelity representations and degraded representations by tracking an increasing spread forward noise, and the calibration dataset includes the representation state of each user under different degrees of degradation.
[0024] In this embodiment of the application, unlike related technologies that simply and crudely divide the user group into "advantageous" and "disadvantageous" groups, this application considers the continuous representation degradation characteristics caused by the sparsity of user interaction.
[0025] Specifically, refer to Figure 2 The diagram illustrates an environment diffusion framework based on course scheduling. To accurately measure the degree of degradation for each user representation, this application proposes an adaptive degradation calibration module, which utilizes a time-conditional discriminator. The system tracks the increasing forward diffusion noise and uses this to distinguish between high-fidelity representations (also known as high-fidelity representation states, i.e., user representations that truly, accurately, and completely reflect the user's original interests and behavioral characteristics) and degraded representations (also known as degraded representation states, i.e., user representations that deviate from their true interests and suffer quality degradation due to external intervention or data defects). The time-conditional discriminator is designed for input at time step t. Includes: representations by the original users (i.e., the initial diffusion state) Following a predefined forward noise scheduling strategy The continuous perturbation state obtained after injecting Gaussian noise.
[0026] It should be noted that as the time step t increases, Gaussian noise gradually becomes dominant, and the distributions of the noisy degradation representation and the high-fidelity representation will tend to be consistent. Therefore, a time condition discriminator can be used to find the smallest time step that makes the noisy degradation representation and the high-fidelity representation statistically indistinguishable, so as to accurately measure the degree of degradation.
[0027] With the premise of measuring the degree of degradation based on the time conditional discriminator, the representation state of different users under different degrees of degradation (such as high-fidelity representation state and perturbation representation state obtained after adding noise) can be collected to construct the calibration dataset, so that the diffusion model can be trained to learn the global high-fidelity user interest distribution.
[0028] Step S102: Train the diffusion model using the calibration dataset, and dynamically expand the sampling range of the calibration dataset in order of increasing degradation level during the model training process.
[0029] In the embodiments of this application, such as Figure 2 As shown, in order to avoid directly adding degraded samples with large perturbations to the early training and thus destroying the convergence stability, this application proposes an environment diffusion model training module based on the degree of degradation to guide course learning. It dynamically expands the sampling range of the calibration dataset according to the order of degradation degree from low to high by adopting a strategy from easy to difficult, so that the model can first use high-fidelity representations to establish a stable basic distribution, and then gradually integrate high-noise degraded representations.
[0030] Step S103: Use the diffusion model after training to determine the high-fidelity interest prototype of the target user, and generate an item recommendation list based on the high-fidelity interest prototype.
[0031] In practical implementation, after the diffusion model is trained, it can be used to repair the degraded representations of the target user to recover the high-fidelity interest prototype. Subsequently, the high-fidelity interest prototype is used... Embedding with pre-trained items Through the scoring function Calculate the predicted preference score Then, a list of recommended items is generated based on the score.
[0032] Therefore, in response to the shortcomings of knowledge transfer-based fairness recommendation methods in related technologies, which ignore the continuous differences in the degree of user representation degradation and forcibly align the representations of disadvantaged users with those of advantageous groups, resulting in distribution collapse and masking the unfairness within the group, this application proposes an environment diffusion framework based on course scheduling. This framework can be flexibly adapted to various traditional and multimodal recommendation backbone networks. Specifically, it uses a time-conditional discriminator to accurately measure the degree of degradation of each user representation to achieve fine-grained fairness discrimination. Furthermore, it designs an adaptive representation repair method based on environment diffusion to compensate for the aforementioned shortcomings of related technologies. This breaks the binary limitation and alleviates the long-tail unfairness phenomenon at the distribution level, thereby significantly enhancing the fairness of the system while improving the overall performance of the recommendation system.
[0033] As can be seen from the above technical solutions, this application introduces a temporal conditional discriminator to distinguish the evolutionary differences between high-fidelity representations and degraded representations by tracking the continuously increasing forward diffusion noise. This considers the continuous representation degradation caused by the sparsity of user interactions and accurately measures the degree of degradation of each user representation, thereby constructing a calibration dataset covering all degrees of degradation. Furthermore, it introduces an environmental diffusion training process based on the degree of degradation to learn the user interest distribution, enriching the learned distribution with full-scale data without sacrificing its fidelity. This allows the model to adaptively recover degraded user representations, effectively improving overall utility and fairness. Therefore, this application effectively solves the recommendation performance limitations caused by the collapse of interest distribution in related technologies, as well as the difficulty in handling the unfairness within a large disadvantaged user group.
[0034] In some embodiments, constructing the calibration dataset based on the time-conditional discriminator includes: For each degraded user's representation, N independent discriminator inference trials are performed on the degraded user's representation by sampling different Gaussian noise in each independent forward process, generating a set of calibration pairs. In each discriminator inference trial, the minimum time step that makes the noisy degraded representation and the high-fidelity representation statistically indistinguishable is determined based on the time-conditional discriminator, and the minimum time step and the perturbation representation state under the minimum time step are used as calibration pairs. The minimum time step is used to measure the original degree of degraded representation. The calibration dataset is obtained by integrating the calibration pair set with the high-fidelity representation states associated with the dominant user group.
[0035] In this embodiment of the application, in order to eliminate the random bias caused by a single noise addition process, a multi-experiment data augmentation strategy is further designed, and a specific minimum time step is used as a fine-grained fairness index to solve the deficiency of related technologies in lacking explicit modeling of the degree of degradation, so as to facilitate the subsequent reasonable guidance of the model to learn the distribution of interest based on the different degradation levels of users.
[0036] Specifically, for each downgraded user's representation (i.e., the original representation of the inferior user to be noisyned), different Gaussian noise is sampled in each independent forward pass, and then the process is performed. (Its value is greater than 1) independent discriminator reasoning trials; where, in the th In this experiment, based on the time-conditional discriminator, the minimum time step that makes the noisy degraded representation and the high-fidelity representation statistically indistinguishable is obtained (for example, finding the minimum time step that makes the output probability of the time-conditional discriminator stable around 0.5). And obtain the perturbation representation state (i.e., the degradation representation) at that time step. .
[0037] Based on the inference information obtained under different noise levels through the above process, a stable and robust set of calibration pairs can be generated for each downgraded user sample. and compare it with the advantageous user groups in By integrating the high-fidelity representations of the states at each time step, an enhanced training set covering all degrees of degradation can be constructed as the calibration dataset.
[0038] Optionally, the minimum time step is determined by the following formula:
[0039] in, Indicates the minimum time step. The larger the value The original degree of degradation (i.e. The more severe the degradation of the original representation of the corresponding downgraded user before adding noise, the more severe the degradation. inf represents the infmum calculation operation. This represents a time-conditional discriminator, where t represents the time step. This represents the perturbation characterization state at time t. This indicates the preset tolerance threshold.
[0040] In some embodiments, the optimization objective used during model training is represented as follows:
[0041] in, Indicates the diffusion loss due to environmental conditions. Indicates the expected computational operation. This indicates the calibrated downgraded characterization observation state (i.e., the reference starting point). and Representing time steps and The predefined noise intensity coefficient is below. Indicates will Further noise increase state, This represents the prediction results of the denoising network in the diffusion model. This represents the semantic guidance conditions in the input of the diffusion model.
[0042] In this embodiment of the application, conditional environmental diffusion loss is designed. The optimization objective is to ensure that the model can fully absorb the residual signals from all users.
[0043] Specifically, this optimization objective introduces a two-stage noise intensity square ratio (i.e. The parameter adjustment term allows the model to legally operate directly with incomplete observations. For the purpose of supervision, it is completely equivalent to the clean distribution that approximates the real one in terms of mathematical expectation; this breaks through the limitation of relying only on data from a few advantageous users, and can safely absorb the residual features of a large number of disadvantaged users into the optimization, which not only enriches the expression of the global interest manifold, but also strictly prevents low-quality noise from polluting the high-fidelity distribution.
[0044] Optionally, the semantic guidance conditions are determined through the following steps: Based on users' preference scores for historically interacted items, normalized attention weights are calculated. The historical item embeddings are weighted and aggregated according to the normalized attention weights to obtain the interactive anchoring semantic guidance conditions as the semantic guidance conditions.
[0045] In this embodiment of the application, during the training phase (and application phase), the input to the diffusion model may include the perturbation representation state in the calibration dataset. Further noise-increasing perturbation state (Time step) ), time step and semantic guidance conditions The model output is the predicted expectation representation. .
[0046] To avoid large disturbances (such as large...) Adding downgraded samples directly to early training disrupts convergence stability. A time-step-guided learning strategy can be used, progressing from easy to difficult. As a calibration time step, the sampling range is dynamically expanded in order of increasing calibration time step, so that the model can first use high-fidelity representation to establish a stable basic distribution, and then gradually integrate high-noise degradation representation.
[0047] Meanwhile, to prevent the restored representation from losing its personalized features, this application calculates normalized attention weights based on users' preference scores for historical interactive items, and then performs weighted aggregation of historical item embeddings, thereby constructing interactive anchoring semantic guidance conditions as semantic guidance conditions. .
[0048] In some embodiments, determining the high-fidelity interest prototype of the target user using the completed diffusion model includes: Based on the time condition discriminator, the perturbation representation of the target user at the target time step is calculated to obtain the calibrated degraded representation observation state. The target time step is: the minimum time step that makes the noisy degraded representation and the high-fidelity representation of the target user statistically indistinguishable. Using the diffusion model, the degraded representation observation state is taken as the starting point of the inverse denoising trajectory. Through iterative execution of deterministic sampling and state updates, the high-fidelity interest prototype of the target user is obtained.
[0049] In the embodiments of this application, such as Figure 2 As shown, a high-fidelity prototype recovery module based on adaptive condition guidance is proposed to be used to repair the degradation features of the target user after the diffusion model training is completed.
[0050] Specifically, unlike traditional diffusion models that start with pure Gaussian noise, this module employs Calibrate Reverse Initialization (CRI) technology, directly applying the calibrated downgraded representation of the observed state (i.e., the target user at the target time step). Perturbation characterization The calculation method can refer to the calibration pair generation process in the training phase) as the starting point of the model's inverse denoising trajectory; thus, the effective semantic information of the skeleton network can be preserved, and it can act as an adaptive filter: high-activity users (i.e., dominant users, who usually have low...) Users with low activity levels (i.e., disadvantaged users, who typically have high...) can retain their characteristics with very few steps. The signal loss is then compensated for through deep generative reconstruction.
[0051] Subsequently, the model performs deterministic sampling using the Probability Flow ODE and iteratively updates the state through Euler discretization. This allows us to obtain a high-fidelity interest prototype of the target user (which reflects the user's true preferences).
[0052] Optionally, in each sampling step, the guiding noise can be calculated using classifier-free guidance (CFG) technology. That is, the guiding noise used in the deterministic sampling process of each iteration can be determined by the following formula:
[0053] in, Indicates guiding noise. This represents the user representation to be denoised in the current iteration (the diffusion model will use this guiding noise as a basis to iteratively update the rules in the next iteration). Recursion to Until the high-fidelity prototype of interest is finally restored. ), Indicates the guiding ratio. This represents the denoising network in the diffusion model. This represents the semantic guidance conditions in the input of the diffusion model. This indicates a null condition marker in the diffusion model input.
[0054] It should be noted that the environmental diffusion method for recommending fairness to users provided in this application can be executed by an environmental diffusion device for recommending fairness to users, or by a control module within that device for executing the environmental diffusion method for loading fairness to users. This application embodiment uses the execution of the environmental diffusion method for loading fairness to users by an environmental diffusion device for recommending fairness to users as an example to illustrate the environmental diffusion method for recommending fairness to users provided in this application embodiment.
[0055] This application provides an environmental diffusion device for recommending fairness to users, such as... Figure 3 As shown, the device includes: A dataset construction module is used to construct a calibration dataset based on a time-conditional discriminator configured to distinguish the evolutionary differences between high-fidelity representations and degraded representations by tracking increasing forward diffusion noise. The calibration dataset includes the representation state of each user under different degrees of degradation. The model training module is used to train a diffusion model using the calibration dataset, and dynamically expands the sampling range of the calibration dataset in order of increasing degradation level during the model training process; The user recommendation module is used to determine the high-fidelity interest prototypes of target users using the diffusion model after training, and to generate a list of recommended items based on the high-fidelity interest prototypes.
[0056] Optionally, the dataset construction module is also used to perform the following steps: For each degraded user's representation, N independent discriminator inference trials are performed on the degraded user's representation by sampling different Gaussian noise in each independent forward process, generating a set of calibration pairs. In each discriminator inference trial, the minimum time step that makes the noisy degraded representation and the high-fidelity representation statistically indistinguishable is determined based on the time-conditional discriminator, and the minimum time step and the perturbation representation state under the minimum time step are used as calibration pairs. The minimum time step is used to measure the original degree of degraded representation. The calibration dataset is obtained by integrating the calibration pair set with the high-fidelity representation states associated with the dominant user group.
[0057] Optionally, the minimum time step is determined by the following formula:
[0058] in, Indicates the minimum time step, and inf denotes the infmum calculation operation. This represents a time-conditional discriminator, where t represents the time step. This represents the perturbation characterization state at time t. This indicates the preset tolerance threshold.
[0059] Optionally, the optimization objective used during model training can be represented as follows:
[0060] in, Indicates the diffusion loss due to environmental conditions. Indicates the expected computational operation. This indicates the degraded characterization of the observation state after calibration. and Representing time steps and The predefined noise intensity coefficient is below. Indicates will Further noise increase state, This represents the prediction results of the denoising network in the diffusion model. This represents the semantic guidance conditions in the input of the diffusion model.
[0061] Optionally, the model training module is further configured to perform the following steps: Based on users' preference scores for historically interacted items, normalized attention weights are calculated. The historical item embeddings are weighted and aggregated according to the normalized attention weights to obtain the interactive anchoring semantic guidance conditions as the semantic guidance conditions.
[0062] Optionally, the user recommendation module is further configured to perform the following steps: Based on the time condition discriminator, the perturbation representation of the target user at the target time step is calculated to obtain the calibrated degraded representation observation state. The target time step is: the minimum time step that makes the noisy degraded representation and the high-fidelity representation of the target user statistically indistinguishable. Using the diffusion model, the degraded representation observation state is taken as the starting point of the inverse denoising trajectory. Through iterative execution of deterministic sampling and state updates, the high-fidelity interest prototype of the target user is obtained.
[0063] Optionally, the guiding noise used in the deterministic sampling process for each iteration is determined by the following formula:
[0064] in, Indicates guiding noise. This represents the user representation to be denoised in the current iteration round. Indicates the guiding ratio. This represents the denoising network in the diffusion model. This represents the semantic guidance conditions in the input of the diffusion model. This indicates a null condition marker in the diffusion model input.
[0065] As can be seen from the above technical solutions, this application introduces a temporal conditional discriminator to distinguish the evolutionary differences between high-fidelity representations and degraded representations by tracking the continuously increasing forward diffusion noise. This considers the continuous representation degradation caused by the sparsity of user interactions and accurately measures the degree of degradation of each user representation, thereby constructing a calibration dataset covering all degrees of degradation. Furthermore, it introduces an environmental diffusion training process based on the degree of degradation to learn the user interest distribution, enriching the learned distribution with full-scale data without sacrificing its fidelity. This allows the model to adaptively recover degraded user representations, effectively improving overall utility and fairness. Therefore, this application effectively solves the recommendation performance limitations caused by the collapse of interest distribution in related technologies, as well as the difficulty in handling the unfairness within a large disadvantaged user group.
[0066] The environmental diffusion device for recommending fairness to users in this application embodiment can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, in-vehicle electronic devices, wearable devices, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network-attached storage (NAS), PCs, televisions (TVs), ATMs, or self-service machines, etc. This application embodiment does not impose specific limitations.
[0067] The environmental diffusion device for recommending fairness to users in this application embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit it.
[0068] The environmental diffusion device for recommending fairness to users provided in this application embodiment can achieve... Figure 1 The various processes implemented in the example of the user-oriented recommendation fairness environment diffusion method shown will not be described again here to avoid repetition.
[0069] Optionally, this application embodiment also provides an electronic device, including a processor 110, a memory 109, and a program or instructions stored in the memory 109 and executable on the processor 110. When the program or instructions are executed by the processor 110, they implement the various processes of the above-described user-oriented recommendation fairness environment diffusion method embodiment and achieve the same technical effect. To avoid repetition, they will not be described again here.
[0070] It should be noted that the electronic devices in the embodiments of this application include the mobile electronic devices and non-mobile electronic devices described above.
[0071] Figure 4 A schematic diagram of the hardware structure of an electronic device to implement an embodiment of this application.
[0072] The electronic device 100 includes, but is not limited to, components such as: radio frequency unit 101, network module 102, audio output unit 103, input unit 104, sensor 105, display unit 106, user input unit 107, interface unit 108, memory 109, and processor 110.
[0073] Those skilled in the art will understand that the electronic device 100 may also include a power supply (such as a battery) for supplying power to various components. The power supply may be logically connected to the processor 110 through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system. Figure 4 The electronic device structure shown does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements, which will not be elaborated here.
[0074] The processor 110 is used to perform the following steps: A calibration dataset is constructed based on a time-conditional discriminator configured to distinguish the evolutionary differences between high-fidelity representations and degraded representations by tracking increasing forward diffusion noise. The calibration dataset includes the representation state of each user under different degrees of degradation. The diffusion model is trained using the calibration dataset, and the sampling range of the calibration dataset is dynamically expanded in order of increasing degradation level during the model training process. The high-fidelity interest prototypes of the target users are determined using the diffusion model after training, and an item recommendation list is generated based on the high-fidelity interest prototypes.
[0075] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiment of the environment diffusion method for user-oriented recommendation fairness and achieve the same technical effect. To avoid repetition, they will not be described again here.
[0076] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.
[0077] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above-described embodiment of the environment diffusion method for user recommendation fairness, and can achieve the same technical effect. To avoid repetition, it will not be described again here.
[0078] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.
[0079] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.
[0080] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, 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 is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.
[0081] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.
Claims
1. An environmental diffusion method for user recommendation fairness, characterized in that, The method includes: A calibration dataset is constructed based on a time-conditional discriminator configured to distinguish the evolutionary differences between high-fidelity representations and degraded representations by tracking increasing forward diffusion noise. The calibration dataset includes the representation state of each user under different degrees of degradation. The diffusion model is trained using the calibration dataset, and the sampling range of the calibration dataset is dynamically expanded in order of increasing degradation level during the model training process. The high-fidelity interest prototypes of the target users are determined using the diffusion model after training, and an item recommendation list is generated based on the high-fidelity interest prototypes.
2. The method according to claim 1, characterized in that, The calibration dataset, constructed based on the time-conditional discriminator, includes: For each degraded user's representation, N independent discriminator inference trials are performed on the degraded user's representation by sampling different Gaussian noise in each independent forward process, generating a set of calibration pairs. In each discriminator inference trial, the minimum time step that makes the noisy degraded representation and the high-fidelity representation statistically indistinguishable is determined based on the time-conditional discriminator, and the minimum time step and the perturbation representation state under the minimum time step are used as calibration pairs. The minimum time step is used to measure the original degree of degraded representation. The calibration dataset is obtained by integrating the calibration pair set with the high-fidelity representation states associated with the dominant user group.
3. The method according to claim 2, characterized in that, The minimum time step is determined by the following formula: in, Indicates the minimum time step, and inf denotes the infmum calculation operation. This represents a time-conditional discriminator, where t represents the time step. This represents the perturbation characterization state at time t. This indicates the preset tolerance threshold.
4. The method according to claim 1, characterized in that, The optimization objective used during model training is represented as follows: in, Indicates the diffusion loss due to environmental conditions. Indicates the expected computational operation. This indicates the degraded characterization of the observation status after calibration. and Representing time steps and The predefined noise intensity coefficient is below. Indicates will Further noise increase state, This represents the prediction results of the denoising network in the diffusion model. This represents the semantic guidance conditions in the input of the diffusion model.
5. The method according to claim 4, characterized in that, The semantic guidance conditions are determined through the following steps: Based on users' preference scores for historically interacted items, normalized attention weights are calculated. The historical item embeddings are weighted and aggregated according to the normalized attention weights to obtain the interactive anchoring semantic guidance conditions as the semantic guidance conditions.
6. The method according to any one of claims 1-5, characterized in that, The process of determining the high-fidelity interest prototypes of target users using the completed diffusion model includes: Based on the time condition discriminator, the perturbation representation of the target user at the target time step is calculated to obtain the calibrated degraded representation observation state. The target time step is: the minimum time step that makes the noisy degraded representation and the high-fidelity representation of the target user statistically indistinguishable. Using the diffusion model, the degraded representation observation state is taken as the starting point of the inverse denoising trajectory. Through iterative execution of deterministic sampling and state updates, the high-fidelity interest prototype of the target user is obtained.
7. The method according to claim 6, characterized in that, The guiding noise used in the deterministic sampling process for each iteration is determined by the following formula: in, Indicates guiding noise. This represents the user representation to be denoised in the current iteration round. Indicates the guiding ratio. This represents the denoising network in the diffusion model. This represents the semantic guidance conditions in the input of the diffusion model. This indicates a null condition flag in the diffusion model input.
8. An environmental diffusion device for recommending fairness to users, characterized in that, The device includes: A dataset construction module is used to construct a calibration dataset based on a time-conditional discriminator configured to distinguish the evolutionary differences between high-fidelity representations and degraded representations by tracking increasing forward diffusion noise. The calibration dataset includes the representation state of each user under different degrees of degradation. The model training module is used to train a diffusion model using the calibration dataset, and dynamically expands the sampling range of the calibration dataset in order of increasing degradation level during the model training process; The user recommendation module is used to determine the high-fidelity interest prototypes of target users using the diffusion model after training, and to generate a list of recommended items based on the high-fidelity interest prototypes.
9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the environmental diffusion method for user-oriented recommendation fairness as described in any one of claims 1-7.
10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the environmental diffusion method for user-oriented recommendation fairness as described in any one of claims 1-7.