Dual-exposure mechanism-aware sequential recommendation method facing user acceptance decision prediction
Patent Information
- Application Number
- CN202611133014.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-29
- Publication Date
- 2026-10-27
AI Technical Summary
[0007]本发明的目的在于提供一种面向用户接受决策预测的双曝光机制感知序列推荐方法，以解决现有序列推荐模型在观测曝光反馈与干预曝光反馈共同存在时，容易混淆不同曝光机制下的反馈规律、产生表示干扰，以及干预曝光样本较少时预测准确性不足的问题
（1）通过融合用户—物品交互图中的长期协同信息和历史交互序列中的动态兴趣信息，使用户偏好表示同时包含长期结构偏好与动态序列兴趣，提高用户偏好建模的完整性。
Smart Images
Figure CN122986176A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of recommender systems, user behavior modeling, and machine learning, and in particular to a dual-exposure mechanism-based perceptual sequence recommendation method for predicting user acceptance decisions. Background Technology
[0002] As the amount of online content, goods, and services continues to grow, recommendation systems typically predict items a user is likely to accept from a pool of candidate items based on their historical interaction records, and generate recommendations accordingly. Existing sequential recommendation methods generally organize user behavior according to the order of interaction, extracting dynamic interests through recurrent neural networks, self-attention networks, or other sequence coding structures; some methods also aggregate collaborative relationships based on user-item interaction graphs to obtain long-term user preference representations.
[0003] In practical recommendation systems, user feedback is generated not only by user preferences but also by factors such as platform exposure strategies, item popularity, user self-selection behavior, and historical recommendation paths. Interaction logs generated during natural recommendation are typically filtered by existing ranking strategies and represent feedback formed under observed exposure conditions. Feedback generated by random exposure, exploratory exposure, or selection of target items by randomization rules is less affected by direct filtering by existing ranking strategies compared to natural exposure feedback and can be considered as interventional exposure feedback, distinct from observed exposure mechanisms.
[0004] Observational exposure feedback and interventional exposure feedback have different data generation mechanisms. Observational exposure logs are usually more numerous and retain relatively obvious local behavioral continuity; interventional exposure feedback can reflect users' acceptance of target items, but in real-world systems, it is often less abundant. If the two types of feedback are directly mixed and trained using the same prediction branch, the model is prone to introducing local feedback dependencies and strategy selection features from the observational exposure scenario into the interventional exposure prediction, causing representational interference between the training signals of the two types of feedback. If only a limited number of interventional exposure feedback samples are used to train the model, problems such as insufficient representation of user preferences, large training fluctuations, and unstable predictions are likely to occur.
[0005] Furthermore, existing methods typically use the feedback states in historical sequences as a unified sequence feature, rarely setting up separate decision-making processes for different exposure mechanisms, and lack a mechanism that can transform the local feedback continuity in the observation exposure log into intervention exposure learning auxiliary information without replacing the true label of the intervention feedback with the observation feedback.
[0006] Therefore, it is necessary to propose a sequence recommendation method based on dual exposure mechanism perception, which distinguishes between observation exposure decision and intervention exposure decision based on user preference representation, and improves the accuracy of user acceptance decision prediction under intervention exposure conditions through auxiliary constraints and staged training methods suitable for intervention exposure learning. Summary of the Invention
[0007] The purpose of this invention is to provide a dual-exposure mechanism-based perceived sequence recommendation method for user acceptance decision prediction, in order to solve the problems of existing sequence recommendation models, which are prone to confusion of feedback patterns under different exposure mechanisms and generation of representation interference when observation exposure feedback and intervention exposure feedback coexist, as well as insufficient prediction accuracy when the number of intervention exposure samples is small.
[0008] To achieve the above objectives, this invention provides a dual-exposure mechanism-based perceptual sequence recommendation method for user acceptance decision prediction, comprising the following steps: S1: Obtain dual-exposure training samples. The training samples include historical interaction sequences, target items, user feedback tags, and exposure mechanism identifiers. Based on the exposure mechanism identifier, they are divided into observation exposure samples and intervention exposure samples. This includes: obtaining the target user's historical interaction sequence before the target time, as well as the target item, user feedback tags, and exposure mechanism identifier at the target time; a value of 1 for the feedback status and user feedback tag indicates that user acceptance behavior has been observed, and a value of 0 indicates that acceptance behavior has not been observed; the exposure mechanism identifier is used to distinguish between observation exposure samples generated under natural exposure conditions and intervention exposure samples generated by random exposure, exploratory exposure, or target item selection by randomization rules.
[0009] S2: Generate User Preference Representation. A user preference representation is generated based on the target user's historical interaction information. Historical interaction information includes user-item interaction relationships and historical interaction sequences. A structural preference representation is generated based on the user-item interaction graph, and a sequence interest representation is generated based on the historical interaction sequences. The user-item interaction graph is constructed from historical interaction records during the training phase. The structural preference representation is obtained by performing multi-level neighborhood aggregation on user nodes and item nodes. The sequence interest representation is obtained by combining historical item representations, feedback state representations, and position representations, and processing the combined sequence using a sequence encoder. The structural preference representation and the sequence interest representation are then fused to obtain the user preference representation.
[0010] S3: Constructing Dual-Exposure Decision Branches. The target item is represented and mapped to obtain its representation. Based on the user preference representation and the target item representation, observation and intervention exposure decision branches are constructed respectively. The observation exposure decision branch further uses the user preference representation, the target item representation, and the previous time-step feedback state as explicit conditional inputs, while the intervention exposure decision branch does not use the previous time-step feedback state as an independent explicit conditional input. The user preference representation is generated through a gating fusion mechanism. The gating fusion mechanism calculates gating weights based on the structural preference representation and the sequence interest representation, and then uses these gating weights to adaptively fuse the two types of representations.
[0011] S4: Construct a recursive consistency reference target. Under the two conditions of "not accepted" and "accepted" feedback states in the previous time step, obtain two conditional prediction probabilities for the current target item using the observation exposure decision branch. Then, weight the two conditional prediction probabilities using the user acceptance probability and its complementary probability output by the observation exposure decision branch for the same user sequence in the previous time step, to obtain the recursive consistency reference target. This target is used as an auxiliary constraint for training the intervention exposure samples. The recursive consistency reference target is only used to provide auxiliary constraints for the intervention exposure decision branch; the actual feedback labels of the intervention exposure samples still serve as the primary training basis.
[0012] S5: Perform observational feedback training and interventional feedback adaptation. During the observational feedback training phase, the model parameters used to generate user preference representations and the parameters of the observational exposure decision branch are updated using observed exposure samples. During the interventional feedback adaptation phase, the parameters of the interventional exposure decision branch are updated using the real feedback labels of the interventional exposure samples, and auxiliary constraints are provided using a recursive consistency reference objective.
[0013] S6: Generate user acceptance decision prediction results. Calculate the user acceptance probability under the intervention exposure condition using the trained intervention exposure decision branch, and generate user acceptance decision prediction results accordingly. After observation feedback training is completed, keep the model parameters used to generate user preference representations and the parameters of the observation exposure decision branch unchanged, and construct a recursive consistency reference target using the observation exposure decision branch; update the parameters of the intervention exposure decision branch during the intervention feedback adaptation phase.
[0014] Compared with the prior art, the present invention has at least the following beneficial effects: (1) By integrating long-term collaborative information in the user-item interaction graph and dynamic interest information in the historical interaction sequence, the user preference representation can simultaneously include long-term structural preferences and dynamic sequence interests, thereby improving the completeness of user preference modeling.
[0015] (2) By constructing observation exposure decision branches and intervention exposure decision branches, and only explicitly introducing the feedback state of the previous moment in the observation exposure decision branch, the direct impact of local feedback dependence in natural exposure scenarios on intervention exposure prediction is reduced.
[0016] (3) By using recursive consistency reference targets to provide historical state perception auxiliary constraints for intervention exposure decision branches, and adopting a phased training method of observation feedback training and intervention feedback adaptation, the interference caused by simultaneous optimization of feedback generated by different exposure mechanisms is reduced, thereby improving the accuracy of user acceptance decision prediction in dual exposure recommendation scenarios. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating the overall process of a dual-exposure mechanism-based perceptual sequence recommendation method for predicting user acceptance decisions, as described in this invention. Detailed Implementation
[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are used to illustrate the technical solutions of the present invention, and not to limit the scope of protection of the present invention. Where there is no conflict, the technical features of each embodiment can be combined with each other. Equivalent transformations or substitutions that can be directly obtained by those skilled in the art based on the description herein should also be included within the technical concept of the present invention.
[0019] This implementation method first defines the overall structure and data.
[0020] like Figure 1 As shown, the dual-exposure mechanism-based sequence recommendation method in this embodiment sequentially includes acquiring dual-exposure training samples, generating user preference representations, constructing dual-exposure decision branches, constructing recursive consistency reference targets, training observation feedback and adapting intervention feedback, and generating user acceptance decision prediction results. The user preference representation is obtained by fusing structural preference representations and sequence interest representations, and serves as the common input to both the observation exposure decision branch and the intervention exposure decision branch.
[0021] Let the set of users be U, and the set of items be... I For target user u∈ U The historical interaction sequence before the target time t is represented as: in, These are the historical items that the target user encountered before the target time t; The feedback states for the corresponding historical interactions are listed in order. A feedback state value of 1 indicates that the user has accepted the corresponding item, while a feedback state value of 0 indicates that no acceptance behavior has been observed.
[0022] The target item at time t is denoted as User acceptance of decision label is recorded as In different application scenarios, acceptance behavior can be determined based on clicks, purchases, favorites, completion of playback, reaching a preset viewing time, rating reaching a preset threshold, or other behaviors that can represent the user's willingness to accept. Correspondingly, no observed acceptance behavior can be determined by records of no clicks, no purchases, failure to reach preset behavior thresholds, or other records that do not meet the acceptance judgment conditions.
[0023] To differentiate the generation mechanisms of feedback samples, an exposure mechanism identifier is set for each target sample:
[0024] When the exposure mechanism identifier is set to "obs", the target sample is an observed exposure sample generated under natural exposure conditions; when the exposure mechanism identifier is set to "int", the target sample is an intervention exposure sample generated by random exposure, exploratory exposure, or target item selection by randomization rules. The exposure mechanism identifier describes the generation mechanism of the target feedback sample and is not a fixed attribute of the target item itself.
[0025] In one data construction method, interaction records are sorted according to the interaction timestamps of the same user or the original record order, and a historical interaction sequence is constructed using records preceding the target interaction. For each target sample, the target user, the historical interaction sequence before the target time, the target item, user feedback tags, and exposure mechanism identifiers are obtained. To avoid leakage of target information, the historical interaction sequence does not include interaction records at or after the target time.
[0026] The historical interaction sequence is constructed from records of interactions that occurred to the target user before the target time. The exposure mechanism for the current target sample is determined by... Individually identified, observation exposure samples and intervention exposure samples participate in the corresponding training phases according to this identifier.
[0027] In step S1, double exposure training samples are obtained.
[0028] Obtain user identifiers, item identifiers, interaction sequence, feedback status, and exposure source information from recommendation system logs, user behavior databases, or offline training datasets. For each target interaction, extract historical interaction records prior to the target time to form... and obtain the target item. and user acceptance of decision labels .
[0029] The exposure mechanism identifier is determined based on the exposure type or sample source recorded in the recommendation system logs or offline training dataset, and the training samples are divided into the observation exposure sample set and the intervention exposure sample set according to the exposure mechanism identifier.
[0030] In one implementation, the observed exposure samples come from user feedback records generated under natural exposure conditions; the intervened exposure samples come from random exposure records, or from random survey records of the item to be fed back, selected by randomization rules. For the same item, it may correspond to different exposure mechanism identifiers at different times, with different users, or under different exposure conditions. Therefore, the exposure mechanism identifier is used to describe the generation mechanism of the target feedback sample and is not fixedly bound to the item identifier.
[0031] In step S2, a user preference representation is generated.
[0032] First, generate a structural preference representation.
[0033] Construct a user-item interaction graph using historical interaction records from the training phase:
[0034] Where U is the set of user nodes. For a set of item nodes, This is the set of interaction edges between user nodes and item nodes. When a user... With items When an interaction that satisfies the graph construction conditions occurs during the training phase, at the user node... With item nodes Establish interaction edges between .
[0035] The conditions for constructing the graph can be any valid interaction, the occurrence of an accepted action, or the fulfillment of a preset feedback threshold. Preferably, only training data is used to construct the user-item interaction graph, and the target interactions of the verification or prediction phases are not added to the graph to avoid subsequent target information from entering the structural preference representation.
[0036] Initialize low-dimensional representations for user nodes and item nodes respectively. Let the first... The layer user node representation and item node representation are respectively and Then, an update method based on normalized neighborhood aggregation is:
[0037] in, For user nodes The set of adjacent item nodes, For item nodes The set of adjacent user nodes.
[0038] go through After layer neighborhood aggregation, the user node representations of the target user in each layer are weighted or averaged to obtain the structural preference representation:
[0039] Structural preference representation This is used to represent the long-term collaborative preferences of target users in the user-item interaction graph. The neighborhood aggregation method described above is a preferred implementation; in other implementations, other graph encoding methods capable of generating target user structural preference representations based on the user-item interaction graph can also be used. Subsequently, a sequence interest representation is generated.
[0040] For the first in the historical interaction sequence Each historical location will house historical artifacts. Mapped to item representation , will provide feedback status Mapped to feedback state representation and obtain position Position representation The three types of representations combine to form a sequence input:
[0041] The input representations corresponding to each historical position are arranged in interactive order to obtain the sequence input matrix:
[0042] Input the sequence input matrix into the sequence encoder:
[0043] in, For sequence encoders, This represents the sequential interest of the target user at the target time.
[0044] In a preferred embodiment, the sequence encoder employs a self-attention coding structure to capture temporal dependencies in the historical interaction sequence. The output of the last historical position of the sequence encoder is used as... , to represent the dynamic interest state before the target time.
[0045] When the length of the historical sequence exceeds the preset maximum length, the preset number of interactions closest to the target time can be retained; when the length of the historical sequence is less than the preset length, padding can be used to fill the gaps, and a mask can be used to prevent the padding positions from participating in effective feature aggregation.
[0046] Then, the structural preference representation and the sequence interest representation are gated and fused.
[0047] Structural preference representation With sequence interest representation These describe long-term collaborative preferences and dynamic interest states, respectively. To adaptively determine the contribution of these two types of information based on the current user and historical states, gating weights are calculated:
[0048] Where [·;·] represents a concatenation operation, and These are the learnable parameters of the gated network. This is the Sigmoid activation function.
[0049] Based on the fusion of structural preference representation and sequential interest representation using gating weights:
[0050] Where ⊙ represents element-wise multiplication, and LayerNorm(·) represents the layer normalization function. This represents user preferences.
[0051] To improve the consistency between structural preference representation and sequence interest representation in the representation space, graph order representation alignment constraints can be introduced during the observation feedback training phase, with the following loss function:
[0052] Uniformity constraints can also be introduced, with the following loss function. Graph order representation alignment constraints are used to reduce the deviation between the structural preference representation and the sequence interest representation of the same user, while uniformity constraints are used to reduce the risk of excessive clustering of user preference representations of different users; the above two types of constraints are optional implementations.
[0053]
[0054] In step S3, a dual-exposure decision branch is constructed.
[0055] First, the target item Perform representation mapping to obtain the target item representation:
[0056] in, This is the item representation mapping function. The target item representation can share parameters with the initial item representation generated during the structural preference representation generation process, or it can use independent item embedding parameters.
[0057] The status feedback from the previous moment Mapped to feedback state representation:
[0058] Where g(·) is the feedback state embedding function.
[0059] The observation and exposure decision branch represents user preferences. Target item representation and feedback status representation The images are stitched together to obtain the observed exposure prediction score:
[0060] The user acceptance probability under the observed exposure conditions was obtained through Sigmoid transformation:
[0061] Intervention exposure decision branch represents user preferences and target item representation By stitching the results together, we obtain the intervention exposure prediction score:
[0062] And the user acceptance probability under the intervention exposure condition is obtained through Sigmoid transformation:
[0063] in, To observe the exposure decision function, This is the decision function for intervening in exposure.
[0064] The observation exposure decision branch explicitly utilizes the previous moment's feedback state to characterize the relationship between user feedback and local historical feedback states in a natural exposure scenario. The intervention exposure decision branch still utilizes user preference representations generated from the complete historical sequence, but does not treat the previous moment's feedback state as an independent explicit conditional input, thereby reducing the direct dependence of intervention exposure prediction on a single local feedback state.
[0065] In step S4, a recursive consistency reference target is constructed.
[0066] In the process of intervention feedback adaptation, in order to utilize the local feedback continuity learned from the observed exposure logs, and at the same time avoid directly using the observed exposure prediction results as the labels of the intervention exposure samples, this implementation constructs a recursive consistency reference target.
[0067] Given the user preference representation and target item representation of the current target sample, the feedback state representations representing the unaccepted state and the accepted state are respectively input into the observation exposure decision branch to obtain two conditional prediction probabilities at the current time:
[0068] in, This represents the probability that a user will accept the feedback at the current moment when the feedback status at the previous moment was "not accepted". This represents the probability that a user would accept the feedback at the current moment if the feedback state at the previous moment was "accept". These two states are only used to calculate the conditional prediction results and do not change the actual feedback states already recorded in the historical sequence.
[0069] The observation-exposure decision branch is used to predict the previous time-series samples of the same user sequence, and the soft history state weight is denoted as... The weights of the soft history states are determined according to the following formula:
[0070] When the same user sequence has a prediction result from the previous time step... Take the user acceptance probability at the previous moment from the output of the observation exposure decision branch. When there is no prediction result from the previous time step, a neutral weight of 0.5 is used. The probability of the conditional prediction corresponding to the accepting state is used as the weight, and its complementary probability is used as the weight of the probability of the conditional prediction corresponding to the unaccepting state.
[0071] This leads to the recursive consistency reference target:
[0072] The recursive consistency reference target reflects the recursive combination relationship between the prediction results of the two conditions at the current moment and the weights of the soft historical state. This reference target is only used to provide auxiliary constraints on the intervention exposure decision branch and is not used as the true feedback label for the intervention exposure sample. The true feedback label is still used as the main training basis.
[0073] Let the number of batches of the intervention exposure samples be 1. The recursive consistency constraint loss is:
[0074] in, This indicates that the batch of samples exposed during the intervention was affected. This indicates the probability of user acceptance of the output of the intervention exposure decision branch. This represents the corresponding recursive consistency reference target.
[0075] In step S5, observational feedback training and interventional feedback adaptation are performed.
[0076] This implementation method employs two stages for model training: observational feedback training and interventional feedback adaptation, in order to reduce mutual interference between training signals corresponding to different exposure mechanisms from the initial stage.
[0077] During the observation feedback training phase, the following processing is performed.
[0078] Select training batches from the observed exposure sample set. The model parameters used to generate user preference representations and the parameters of the observation-based exposure decision branch are jointly updated using observed exposure samples. The observation feedback prediction loss is:
[0079] When using graph order to represent alignment and uniformity constraints, the optimization objective during the observation feedback training phase is:
[0080] in, and These represent the weighting coefficients for alignment and uniformity constraints, respectively. These weighting coefficients can be determined using validation data, cross-validation, or preset parameters.
[0081] The observation and feedback training phase can employ stochastic gradient descent, Adam, AdamW, or other gradient-based parameter optimization algorithms. Training termination conditions can include reaching a preset number of iterations, verifying that the loss no longer decreases within the preset number of iterations, or satisfying other convergence criteria.
[0082] During the intervention feedback adaptation phase, the following procedures are performed.
[0083] After completing the observation feedback training, training batches are selected from the intervention exposure sample set. The parameters of the intervention exposure decision branch are updated using the real feedback labels of the intervention exposure samples. The intervention feedback prediction loss is:
[0084] Using the recursive consistency constraint loss as an auxiliary regularization term, we obtain the optimization objective for the intervention feedback adaptation stage:
[0085] in, This represents the weighting coefficient for the recursive consistency constraint loss. The intervention feedback prediction loss is calculated based on the actual feedback labels, while the recursive consistency constraint loss only provides auxiliary constraints on the predicted probability of intervention exposure.
[0086] In a preferred embodiment, after the observation feedback training is completed, the model parameters used to generate the user preference representation and the parameters of the observation exposure decision branch are not updated, and only the parameters of the intervention exposure decision branch are updated during the intervention feedback adaptation phase. Using this approach, both the model parameters used to generate the user preference representation and the observation exposure decision function used to construct the recursive consistency reference target remain stable in the second phase.
[0087] In step S6, the prediction result of the user's acceptance decision is generated.
[0088] After the model is trained, it is then used to target users. At the target time Targeted Items The prediction task generates user preference representations based on historical interaction sequences and user-item interaction graphs prior to the target time. And perform a representation mapping on the target item to obtain the target item representation. .
[0089] The user preference representation and the target item representation are input into the trained intervention exposure decision branch to obtain the intervention exposure prediction score, and the user acceptance probability is calculated by Sigmoid transformation:
[0090] For a single target item, the user's acceptance probability can be compared with a preset decision threshold. If the user's acceptance probability is greater than or equal to the preset decision threshold, an acceptance prediction is generated; if the user's acceptance probability is less than the preset decision threshold, a non-acceptance prediction is generated. The preset decision threshold can be 0.5, or it can be set according to business objectives, validation data, or different error costs.
[0091] For recommendation tasks involving multiple candidate items, the probability of user acceptance for each candidate item under the intervention exposure condition is calculated separately. These items are then sorted from highest to lowest acceptance probability, and one or more of the top-ranked candidate items are selected to generate a recommendation list. User acceptance probability can also be combined with item availability, display constraints, diversity constraints, or other recommendation strategy rules to generate the final recommendation result.
[0092] The above embodiments illustrate the technical solution of the present invention, but the present invention is not limited to the specific structures and parameters described above. For example, the structural preference representation generation method, the sequence interest representation generation method, the gating fusion structure, the decision function, the loss function, and the parameter update method can all be equivalently replaced or adjusted while maintaining the distinction between the dual exposure mechanisms, the recursive consistency auxiliary constraints, and the prediction relationship of the intervention exposure acceptance probability.
Claims
1. A dual-exposure mechanism-based perceptual sequence recommendation method for user acceptance decision prediction, characterized in that: Includes the following steps: S1: Obtain dual-exposure training samples: The training samples include historical interaction sequences, target items, user feedback tags, and exposure mechanism identifiers; Training samples are categorized into observation exposure samples and intervention exposure samples based on the exposure mechanism identifier; S2: Generate user preference representation: Generate user preference representation based on the target user's historical interaction information; S3: Construct dual exposure decision branches: Map the target item to obtain the target item representation, and construct the observation exposure decision branch and the intervention exposure decision branch based on the user preference representation and the target item representation, respectively; The observation exposure decision branch uses the feedback state of the previous moment as an explicit conditional input, while the intervention exposure decision branch does not use the feedback state of the previous moment as an independent explicit conditional input. S4: Construct a recursive consistency reference target: Under the two conditions of non-acceptance and acceptance in the previous time step, the observation exposure decision branch is used to obtain two conditional prediction probabilities of the current target item. The observation exposure decision branch is used to weight the user acceptance probability and its complementary probability output by the previous time step of the same user sequence to obtain a recursive consistency reference target, which is then used as an auxiliary constraint for intervening in the training of exposure samples. S5: Perform observation feedback training and intervention feedback adaptation: In the observation feedback training phase, update the model parameters used to generate user preference representations and the parameters of the observation exposure decision branch using the observation exposure samples; in the intervention feedback adaptation phase, update the parameters of the intervention exposure decision branch using the real feedback labels of the intervention exposure samples, and provide auxiliary constraints using the recursive consistency reference target. S6: Generate user acceptance decision prediction results: Calculate the user acceptance probability using the trained intervention exposure decision branch, and generate user acceptance decision prediction results based on the user acceptance probability.
2. The dual-exposure mechanism-based perceptual sequence recommendation method for user acceptance decision prediction according to claim 1, characterized in that, S1 includes: acquiring the target user's historical interaction sequence before the target time, as well as the target item, user feedback tag, and exposure mechanism identifier at the target time; when the feedback status and user feedback tag are 1, it indicates that the user's acceptance behavior has been observed, and when they are 0, it indicates that no acceptance behavior has been observed; the exposure mechanism identifier is used to distinguish between observed exposure samples generated under natural exposure conditions and intervention exposure samples generated by random exposure, exploratory exposure, or target item selection by randomization rules.
3. The dual-exposure mechanism-based perceptual sequence recommendation method for user acceptance decision prediction according to claim 1, characterized in that, S2 includes: constructing a user-item interaction graph using historical interaction records from the training phase, and performing neighborhood aggregation on user nodes and item nodes in the graph to obtain a structural preference representation; inputting the historical item representation, feedback state representation, and position representation from the historical interaction sequence into a sequence encoder to obtain a sequence interest representation; and fusing the structural preference representation and the sequence interest representation to obtain a user preference representation.
4. The dual-exposure mechanism-based perceptual sequence recommendation method for user acceptance decision prediction according to claim 3, characterized in that, The user preference representation is obtained through a gating fusion mechanism; the gating fusion mechanism calculates gating weights based on the structural preference representation and the sequential interest representation, and uses the gating weights to perform weighted fusion of the structural preference representation and the sequential interest representation.
5. The dual-exposure mechanism-based perceptual sequence recommendation method for user acceptance decision prediction according to claim 1, characterized in that, S3 includes: mapping the previous feedback state to a feedback state representation; inputting the user preference representation, target item representation, and feedback state representation into the observation exposure decision branch to obtain the user acceptance probability under the observation exposure condition; and inputting the user preference representation and target item representation into the intervention exposure decision branch to obtain the user acceptance probability under the intervention exposure condition.
6. The dual-exposure mechanism-based perceptual sequence recommendation method for user acceptance decision prediction according to claim 5, characterized in that, S4 includes: inputting the feedback state representations representing the non-acceptance state and the acceptance state, together with the user preference representation and the target item representation at the current moment, into the observation and exposure decision branch to obtain the conditional prediction probability corresponding to the non-acceptance state and the conditional prediction probability corresponding to the acceptance state; using the user acceptance probability output by the observation and exposure decision branch for the previous moment's sample of the same user sequence as the weight of the conditional prediction probability corresponding to the acceptance state, and using the complementary probability of the user acceptance probability as the weight of the conditional prediction probability corresponding to the non-acceptance state, weighting the two conditional prediction probabilities to obtain the recursive consistency reference target; when the prediction result of the previous moment's sample of the same user sequence cannot be obtained, using 0.5 as a neutral weight to weight the two conditional prediction probabilities.
7. The dual-exposure mechanism-based perceptual sequence recommendation method for user acceptance decision prediction according to claim 6, characterized in that, The recursive consistency reference target is only used to provide auxiliary constraints on the intervention exposure decision branch and is not used as the true feedback label of the intervention exposure sample. The recursive consistency constraint loss is constructed based on the difference between the user acceptance probability output by the intervention exposure decision branch and the recursive consistency reference target.
8. The dual-exposure mechanism-based perceptual sequence recommendation method for user acceptance decision prediction according to claim 1, characterized in that, S5 includes: optimizing the observation feedback prediction loss based on the true feedback labels of the observed exposure samples to update the model parameters used to generate user preference representations and the parameters of the observation exposure decision branch; optimizing the intervention feedback prediction loss based on the true feedback labels of the intervention exposure samples, and updating the parameters of the intervention exposure decision branch in combination with the recursive consistency constraint loss.
9. The dual-exposure mechanism-based perceptual sequence recommendation method for user acceptance decision prediction according to claim 8, characterized in that, After the observation feedback training is completed, the model parameters used to generate user preference representations and the parameters of the observation exposure decision branch are kept unchanged, and a recursive consistency reference target is constructed using the observation exposure decision branch; the parameters of the intervention exposure decision branch are updated during the intervention feedback adaptation phase.
10. The dual-exposure mechanism-based perceptual sequence recommendation method for user acceptance decision prediction according to claim 1, characterized in that, S6 includes: generating a user preference representation based on the target user's historical interaction sequence before the target time and the user-item interaction graph, and obtaining a target item representation; inputting the user preference representation and the target item representation into the trained intervention exposure decision branch to obtain the user acceptance probability; generating a user acceptance decision prediction result based on the user acceptance probability, or ranking multiple candidate items.