Stability-induced weight disturbance method for streaming recommendation and electronic equipment

Abstract

The invention discloses a stability-induced weight disturbance method for streaming recommendation and electronic equipment, and the method comprises the steps: screening representative samples from historical interaction data according to the model influence information of interaction samples on a recommendation model and the graph structure information of the interaction samples, and constructing an experience buffer pool; defining a stability induced loss function based on local worst case parameter disturbance applied to current model parameters of the recommendation model; the stability induced loss function is used for constraining prediction distribution consistency of the recommendation model on samples in the experience buffer pool under parameter disturbance; and combining the stability induced loss function with a recommendation loss function to obtain a comprehensive objective function of the recommendation model in streaming recommendation, and optimizing parameters of the recommendation model by minimizing the comprehensive objective function. According to the method, a mechanism of sample representativeness representation and prediction stability enhancement can be taken into consideration in streaming recommendation, so that catastrophic forgetting is effectively inhibited.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and recommendation system technology, and in particular to a stability-induced weight perturbation method and electronic device for streaming recommendation. Background Technology

[0002] Recommender systems aim to provide personalized suggestions based on users' historical interactions. Since user interests and trends evolve over time, capturing these dynamic changes is crucial for providing accurate recommendations. Streaming recommendation systems have emerged to address this need and have received widespread attention from academia and industry in recent years. Unlike traditional offline recommendation models, streaming recommendation models continuously learn from a constant stream of incoming data, enabling them to capture users' latest interests and preferences. However, streaming recommendation systems face a fundamental challenge: "catastrophic forgetting." When learning new interaction data, recommendation models often overwrite knowledge learned from historical data, leading to performance degradation on older data reflecting long-term user preferences.

[0003] To alleviate this challenge, researchers have proposed various continuous learning techniques, such as regularization-based methods, parameter amplification-based methods, and experience replay-based methods. Among these, experience replay-based methods have attracted widespread attention because they can effectively preserve historical knowledge by directly retraining a small number of historical samples (i.e., the experience buffer). In this field, the Graph Experience Replay (GER) method stands out as an advanced technique. It further utilizes graph convolutional neural networks (GCNs) to model the complex relationships between users and items and preserves structured information by replaying representative subgraphs.

[0004] While graph experience replay methods mitigate catastrophic forgetting to some extent, they still suffer from two major limitations when dealing with streaming recommendations. First, existing sample selection strategies often overlook the synergy between the predictive dynamics of the recommendation model and the graph structure: existing solutions typically rely either on recommendation model-based metrics (such as prediction error or influence functions) or graph structure-based metrics (such as node centrality or diversity). Since the evolving preferences captured by the recommendation model are intertwined with the underlying graph topology reflecting user behavior patterns, this separation leads to insufficient representativeness of samples in the buffer pool. Second, existing graph experience replay methods neglect the stability of recommendation model predictions to small parameter perturbations: in streaming settings, recommendation model parameters undergo local drift to adapt to new data. However, if the recommendation model's predictions for replay data are not stable within a certain neighborhood of the parameter space, then even with replay, parameter updates for new data can easily disrupt the recommendation model's memory of historical knowledge, thus exacerbating forgetting.

[0005] Therefore, designing a mechanism that can simultaneously consider sample representativeness and prediction stability enhancement in streaming recommendation to effectively suppress catastrophic forgetting has become a pressing technical problem in this field. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a stability-induced weight perturbation method and electronic device for streaming recommendation, which can adapt to local parameter drift caused by new interactive data while maintaining the stability of predictions for replay data.

[0007] To achieve the above objectives, this invention provides a stability-induced weight perturbation method for streaming recommendation, comprising the following steps: Based on the influence information of interaction samples on the recommendation model and their graph structure information, representative samples are selected from historical interaction data and an experience buffer pool is constructed. Based on the local worst-case parameter perturbation applied to the current model parameters of the recommendation model, a stability-induced loss function is defined; the stability-induced loss function is used to constrain the consistency of the prediction distribution of the samples in the experience buffer pool by the recommendation model under parameter perturbation. The stability-induced loss function is combined with the recommendation loss function to obtain the comprehensive objective function of the recommendation model in streaming recommendation, and the parameters of the recommendation model are optimized by minimizing the comprehensive objective function.

[0008] Furthermore, the step of selecting representative samples from historical interaction data and constructing an experience buffer pool based on the model influence information and graph structure information of the interaction samples on the recommendation model also includes: Define an influence scoring function to quantify the importance of each interaction sample; The influence scoring function takes the optimal model parameters of the current snapshot and the graph structure information of the interaction samples as input, and the experience buffer pool as output; wherein, the optimal model parameters of the current snapshot are obtained by minimizing the recommendation loss function on the current training set.

[0009] Furthermore, the model influence information is quantified by the model influence score of the recommendation model through interaction samples; the model influence score is calculated based on the alignment between the loss gradient of the interaction sample and the aggregated gradient of historical data, and is used to measure the strength of the influence of the interaction sample on the optimization of the recommendation model.

[0010] Furthermore, the graph structure information is quantified through a structural influence score; the structural influence score is calculated based on the degree of the user nodes and item nodes associated with the interaction sample, and the lower the node degree, the higher the structural influence score of the interaction sample.

[0011] Furthermore, it also includes: The model influence score and the structural influence score are normalized respectively; The normalized model influence score and structural influence score are weighted and averaged to obtain the fusion score. All interaction samples are sorted in descending order based on the fusion score, and the top-ranked interaction samples are selected according to a preset number to construct the experience buffer pool.

[0012] Furthermore, the stability-induced loss function is defined as the maximum KL divergence of the predicted distribution on the empirical buffer pool under the worst-case perturbation within the neighborhood of the current model parameters of the recommendation model; the magnitude of the perturbation is constrained by a preset neighborhood radius.

[0013] Furthermore, the stability-induced loss function constrains the knowledge forgetting of the recommendation model in streaming recommendation within an explicit upper bound; the explicit upper bound of knowledge forgetting is jointly determined by the stability-induced loss function and the size of the experience buffer pool, and the upper bound of knowledge forgetting decreases as the stability-induced loss function decreases or the experience buffer pool increases.

[0014] Furthermore, the recommendation loss function employs Bayesian personalized ranking loss to capture users' short-term preferences.

[0015] Furthermore, in the streaming recommendation scenario, the training set of each time snapshot is composed of the data in the experience buffer pool that was built at the end of the previous time snapshot and the newly arrived interaction data in the current time snapshot; the size of the experience buffer pool is constrained by a preset replay ratio, which represents the proportion of historical interaction data selected into the experience buffer pool.

[0016] To achieve the above objectives, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute the computer program stored in the memory to implement the stability-induced weight perturbation method for streaming recommendation as described above.

[0017] The stability-induced weight perturbation method for streaming recommendation provided by this invention selects representative samples and constructs an experience buffer pool by integrating model influence information and graph structure information to retain key historical interaction data; at the same time, it adopts a stability-induced loss function based on local worst-case parameter perturbation to enhance the stability of the recommendation model's prediction of experience buffer pool data.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a three-dimensional structural schematic diagram of a stability-induced weight perturbation method for streaming recommendation according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the stability-induced weight perturbation graph empirical replay framework according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the impact of three key hyperparameters according to an embodiment of the present invention on model performance on an Amazon dataset; Figure 4 This is a schematic diagram comparing the average training time of each snapshot window for different models according to an embodiment of the present invention; Figure 5 This is a schematic diagram of an electronic device structure according to an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the invention. It should be understood that the accompanying drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the invention.

[0022] The term "comprising" and its variations as used in this invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; and the term "some embodiments" means "at least some embodiments".

[0023] It should be noted that the concepts of "first" and "second" may be mentioned in this invention only to distinguish different devices, components or parts, and are not used to limit the order of the functions performed by these devices, components or parts or their interdependence.

[0024] It should be noted that the terms "one" and "multiple" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless explicitly stated otherwise in the context, they should be understood as "one or more". "Multiple" should be understood as two or more.

[0025] The STAB4Srec described in this invention is a stability-inducing weight-perturbed graph experience replay framework for streaming recommendation. The core idea of ​​STAB4SRec is to enhance the predictive stability of the recommendation model on this historical data by replaying representative historical data and inducing smoothness in model parameters, thereby suppressing knowledge forgetting caused by parameter drift. Specifically, this invention first constructs a compact experience buffer by integrating model influence information and graph structure information. By combining model prediction dynamics and graph topological features, the most representative interactions for maintaining long-term interest can be selected. Subsequently, to enhance the predictive stability of the recommendation model, this invention designs a stability-inducing loss function. This loss function is based on weight perturbations in the local worst case, and by minimizing the KL divergence between perturbed and unperturbed predictions in the buffer, it forces the recommendation model to maintain consistent outputs on historical knowledge within the neighborhood of parameter changes. Furthermore, this invention theoretically proves that the proposed stability-inducing loss provides a tight upper bound for the measure of knowledge forgetting in streaming recommendation, providing a theoretical basis for the effectiveness of the method.

[0026] In this invention, the model influence information refers to the model influence score of the interactive sample on the recommendation model, and the graph structure information refers to the structural information of the interactive sample in the graph structure. Definitions of other terms will be given below.

[0027] Example 1 Figure 1 The flowchart below shows a stability-induced weight perturbation method for streaming recommendation according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the recommended model architecture according to an embodiment of the present invention. The following will be combined with... Figure 1 and Figure 2 The specific embodiments of the present invention will be described in further detail below.

[0028] To clearly describe the scenarios to which this invention is applicable, this invention first provides the following formal definitions of streaming recommendation and graph experience replay: Definition 1, Streaming Recommendation: In real-world recommendation systems, user interactions with items arrive as a continuous data stream, denoted as... This data stream Divided into a series of snapshots ordered by time. These snapshots are spaced at fixed time intervals. In each snapshot... At any given moment, the recommendation model is in the current snapshot. Newly arriving interactive data The parameters of its recommendation model are updated, while utilizing data from earlier snapshot sequences. The recommendation performance is evaluated over the entire timeline based on the knowledge accumulated in the process.

[0029] Definition 2, Graph Experience Replay: For each snapshot At any time, build the training set .in, Indicates in snapshot At the end, the already maintained experience buffer pool is generated from the previous incremental dataset. The selected representative samples constitute the dataset (i.e., representative samples selected from historical data). In the context of streaming recommendation, graph experience replay can be formally represented as:

[0030] in, Represents the overall loss function; buffer pool capacity constraint. This indicates the replay ratio of historical interaction data (the proportion of historical interaction data selected into the experience buffer pool). This represents the recommendation loss function (e.g., BPR loss (Bayesian Personalized Ranking Loss)). Indicates the current snapshot The parameters of the recommendation model that are being optimized at any given time (e.g., user / item embedding vectors). This indicates the size of the experience buffer pool; Indicates as of snapshot time The amount of all interactive data at any given moment.

[0031] Experience buffer pool The quality of the data directly determines the overall performance of incremental recommendations. To address this, this invention designs a sampling strategy that integrates model influence information and graph structure information to prioritize the retention of interaction samples that have the greatest impact on the update of the recommendation model, namely, key and representative interaction behaviors.

[0032] like Figure 2 As shown in the figure express A constantly maintained experience buffer. The recommendation model proposed in this invention consists of two components: a model-structure fusion sampling component and a stability-induced weight perturbation component. The model-structure fusion sampling component constructs an experience buffer by combining model influence information and graph structure information to retain key long-term interaction samples, while the stability-induced weight perturbation component enhances the stability of the recommendation model's prediction of replay data under local parameter perturbations by introducing stability-induced regularization. These two components work synergistically to improve the stability of streaming recommendation against catastrophic failures.

[0033] First, in step 101, the influence scoring function is defined.

[0034] In embodiments of this invention, "influence" is characterized as a combined manifestation of changes in the recommendation model's state and graph structure, which can significantly perturb recommendation performance. To select representative samples from historical interaction data and construct an experience buffer pool based on the model influence information and graph structure information of the interaction samples, an influence scoring function is first defined to quantify the importance of each interaction sample. This influence scoring function takes the optimal model parameters of the current snapshot and the graph structure information of the interaction samples as input, and the experience buffer pool as output. The optimal model parameters of the current snapshot are obtained by minimizing the recommendation loss function on the current training set.

[0035] Specifically, the importance of interaction samples is quantified as their combined impact on changes in model state and graph structure, and an impact scoring function is defined. The following is used to guide empirical sampling:

[0036]

[0037] in, Indicates the size of the experience buffer pool. Represents graph structure information, Let represent the optimal recommendation model parameters at time snapshot t. These parameters are used to characterize the model's influence on interaction samples, where:

[0038] in, The independent variable represents the value at which the minimum value is obtained. Indicates the parameters of the recommendation model The search space Representative from the current training set The first sampled A number of user-item interaction samples. Therefore, the goal of formula (2) is to continuously find representative interaction samples, thereby further improving the recommendation performance of future timestamps.

[0039] In step 102, the model influence score of the interaction sample is calculated.

[0040] In embodiments of the present invention, model influence information is quantified through the model influence score of the recommendation model based on interaction samples. The model influence score is calculated based on the alignment between the loss gradient of the interaction sample and the aggregated gradient of historical data, and is used to measure the strength of the influence of the interaction sample on the optimization of the recommendation model.

[0041] From a model perspective, an interaction sample is considered influential if its associated loss gradient causes a significant change in the parameters of the learned recommendation model. Therefore, an interaction sample is defined as such by quantifying its model influence by measuring the alignment between its gradient and the aggregated gradients of historical data. The model influence scores are as follows:

[0042] in, Measured interaction samples The strength of the impact on the optimization of the recommendation model, while gradient alignment represents the consistency with historical recommendation model knowledge.

[0043] In step 103, the structural influence score of the interaction sample is calculated.

[0044] In embodiments of the present invention, graph structure information is quantified through a structural influence score. The structural influence score is calculated based on the degree of the user nodes and item nodes associated with the interaction sample; the lower the node degree, the higher the structural influence score of the interaction sample.

[0045] From a structural perspective, graph topology also implies the importance of interaction; therefore, graph topology can be used to characterize the impact of interactions. Interaction samples related to users or items with low node degrees are considered to have higher structural information content. Based on this, an interaction sample is defined. The structural influence scores are as follows:

[0046] in, Representation and interaction samples Related user and item nodes, Indicates the degree of a node.

[0047] In step 104, the model-structure fusion score of the interactive samples is calculated.

[0048] In embodiments of the present invention, in order to make full use of comprehensive graph data, a weighted average method is used to combine the model influence score in the interaction importance score. and structural influence score Since these two types of scores have different dimensions, they need to be normalized before combination. The normalization function expression is as follows:

[0049] in, Represents the normalization function. , These represent the minimum and maximum value functions, respectively. Score based on influence.

[0050] Then, the fusion score is obtained by weighted averaging. The normalized model influence score is weighted by the structural influence score, as shown in formula (7):

[0051] in, These are hyperparameters used to adjust the sampling focus. Higher values ​​indicate a higher level of sensitivity. The value indicates that the model influence of interactive samples is emphasized more during the fusion sampling process.

[0052] In an embodiment of the present invention, after obtaining the fusion score, all interactive samples are sorted in descending order of fusion score, and the top samples are selected. Top- Interactive samples are used to build an experience buffer pool. This experience buffer, serving as a compact and information-rich summary of historical interactions, is used for replay during incremental optimization to promote stable updates of recommendation model parameters and effectively mitigate the catastrophic forgetting problem in streaming recommendation. This also provides the necessary foundation for another component (stability-induced weight perturbation): based on the construction of a high-quality experience buffer, this invention further introduces a stability-induced loss function to enhance the stability of the recommendation model's predictions of experience buffer data under local parameter perturbations, thereby more effectively suppressing knowledge forgetting.

[0053] In embodiments of the present invention, after constructing a compact experience buffer pool by combining the influence and structural information driven by the recommendation model in the above steps, the present invention further focuses on how to effectively utilize this experience buffer pool to enhance the stability of the recommendation model when adapting to new data reflecting short-term user preferences, thereby better mitigating the catastrophic forgetting of long-term user interests. Specifically, the present invention introduces a stability-induced loss function and combines it with the recommendation objective to improve the stability of the recommendation model in subsequent incremental updates.

[0054] In step 105, knowledge forgetting is formally measured.

[0055] In an embodiment of the present invention, a construction-based experience buffer pool is proposed. The knowledge metric encapsulates previously learned stable long-term user preferences. Accordingly, the degree of knowledge forgetting is defined as the KL divergence between the predicted distributions of the current and future recommendation models, evaluated on an experience buffer:

[0056] in, and These represent the current and future recommendation model parameters, respectively. This represents the probability distribution predicted by the recommendation model for buffered interaction samples (on the empirical buffer pool); The colon (") serves as a separator, indicating the range for KL divergence calculation. This represents a sample in the buffer pool, where Indicates the user's historical interaction behavior; This indicates the corresponding target project.

[0057] loss function The aim is to ensure that predictions for buffered interactive samples remain consistent in the future. Intuitively, It measures the degree of change in the recommendation model's predicted values ​​for replayed historical interaction samples after future parameter updates. The larger the value, the more severe the prediction bias of the recommendation model in the experience buffer pool, resulting in more severe knowledge forgetting; conversely, a smaller value means that the historical knowledge encoded in the buffer pool is effectively preserved.

[0058] In step 106, a stability-induced loss function based on local worst-case parameter perturbation is defined.

[0059] In embodiments of the present invention, in At any given moment, the parameters of the future recommendation model It is unknown whether this invention uses the worst-case perturbation within the neighborhood of the current recommendation model parameters to approximate the future recommendation model parameters. , that is, use Alternative ,in ( (where the space is real numbers), thereby reducing the forgetting phenomenon in recommendation models without depending on the future direction of model parameter updates. Based on this, this invention proposes a stability-induced loss function, denoted as... Its definition is as follows:

[0060] in, It is a stability-induced loss function. It is an arbitrarily given positive number used to control the scale of the perturbation neighborhood. In practical applications, The value can be set according to the model size and parameter update range; generally, 10 is selected. -3 ~10 -1 This allows for better performance. In this embodiment, the parameters... Set to 0.01. Indicates from arrive The parameter update increment is limited to the radius. Inside the ball.

[0061] Intuitively, the STAB loss function promotes consistency in the predicted distribution of replay samples within the local parameter neighborhood of the recommendation model. This grants subsequent recommendation model updates greater flexibility and robustness in the parameter space, thereby reducing the likelihood of future forgetting in streaming recommendations. This stability-aware optimization helps prevent key long-term user preferences from being overridden by short-term fluctuations, thus achieving more reliable knowledge retention over time.

[0062] In step 107, the stability-induced loss function and the recommendation loss function are combined to obtain the comprehensive objective function of the recommendation model.

[0063] In embodiments of the present invention, the STAB loss function is combined with the aforementioned recommendation loss function. Seamlessly integrated to reduce forgetting and improve the overall performance of the recommendation model. Specifically, for a specific recommendation objective in a given streaming recommendation, such as BPR or its variants, It can be added as a regularization term to obtain the comprehensive objective function of the recommendation model in this streaming recommendation. The expression is as follows:

[0064] in, This represents the balance coefficient, used to balance plasticity and stability. Generally, better performance can be achieved in the range of [0, 2]. Capture short-term user preferences, and Constrain prediction drift on replay data to mitigate catastrophic forgetting in streaming recommendations.

[0065] By minimizing this comprehensive objective function, the recommendation model can capture short-term user preferences while preventing long-term knowledge from being overwritten, thus adapting well to evolving user behavior.

[0066] In the embodiments of the present invention, it is further theoretically proven that the stability-induced loss (STAB4SRec) proposed in this invention provides an explicit upper bound for knowledge forgetting in streaming recommendation, thereby ensuring the stability of the method of the present invention.

[0067] Theorem 1 (Stability-Induced Forgetting Bound): In the streaming recommendation setting, the proposed STAB4SRec provides an explicit upper bound for the knowledge forgetting metric under bounded parameter perturbations.

[0068] Proof: Assume the recommendation model is in Constant updates mean that the recommendation model's impact scoring function It is optimal. (In snapshot) The knowledge forgetting is quantified into two consecutive timestamps during the training process. and The change in model error between, i.e. Its expression is as shown in formula (11):

[0069] here, It is an indicator function. The model represents the distribution of data in the buffer pool B. t The mathematical expectation of the degree of knowledge forgetting. The variation in error can be determined by the upper bound of the shift between the prediction distributions, which is quantified by the total variation (TV) distance, as in Equation (12):

[0070] According to Pinsker's inequality, the TV distance allows for an upper bound based on the KL divergence, as shown in Equation (13):

[0071] Subsequently, after averaging the empirical buffer pool, we apply the Cauchy-Schwarz inequality to combine the square roots, as shown in formula (14):

[0072] Due to arrive The actual incremental update is located within the radius considered by the stability objective. Within the set of disturbances, i.e. Therefore, the resulting KL drift is correspondingly worst-case. Stability Term (The upper bound constraint defined in formula (9) is as shown in formula (15):

[0073] As shown in Equation (15), knowledge forgetting is governed by two factors: it decreases with increasing experience pool size and increases with increasing proposed stability-induced loss. Therefore, for a fixed pool size, explicitly optimizing the stability-induced loss provides an effective mechanism for mitigating forgetting, consistent with the theoretical intuition of this invention. Furthermore, the quality of the experience pool is guaranteed by shaping the drift evaluated on the pool during incremental updates and the associated stability term loss, thus implicitly influencing knowledge forgetting. Therefore, under bounded parameter perturbations, knowledge forgetting caused by incremental updates has a closed upper bound controlled by the stability term, thus establishing a formal stability guarantee for the proposed recommendation model in streaming recommendation scenarios.

[0074] To verify the effectiveness of this invention, extensive experiments were conducted on three public datasets (Amazon, Yelp, and Taobao), and comparisons were made with various baseline methods. The experimental setup, evaluation metrics, and results are as follows.

[0075] First, the experimental setup: Regarding the datasets, this invention uses the following three public datasets covering different real-world streaming recommendation scenarios: (1) Amazon: This dataset contains 93,471 ratings from 9,826 users for 16,172 products in the CD and vinyl record category. The time span is 2,064 days, from January 1, 2018 to August 27, 2023. Interactions with ratings higher than 2 are retained as positive feedback, and users or items with fewer than 10 interactions are filtered out.

[0076] (2) Yelp: This dataset contains 942,395 reviews from 40,863 users for 25,338 businesses on the local services platform. The time span is 1,826 days, from December 13, 2014 to December 13, 2019. Interactions with a rating higher than 3 are retained as positive feedback, and users or items with fewer than 10 interactions are filtered out.

[0077] (3) Taobao: This dataset contains 749,438 interactions from 8,844 users on 39,103 products in the online shopping system. The time span is 30 days, from November 18, 2014 to December 18, 2014. Users or items with fewer than 10 interactions were removed.

[0078] Regarding evaluation, to simulate a streaming recommendation scenario, the interactions in each dataset were arranged chronologically. 60% of the data was allocated as a pre-training set. The remaining 40% was considered incremental data for dividing the interactive data. The incremental portion was further evenly divided into... Each consecutive stage generates a series of data blocks. In terms of evaluation, in ( After training the recommendation model, use the subsequent data blocks. The first half is validated, and the remaining half is tested. Use Recall@ and NDCG@ (in To evaluate the recommendation model and ultimately report all The average result across several incremental data blocks. For fair comparison, LightGCN was used as the base recommendation model in all incremental methods.

[0079] Regarding the baseline method, this experiment included eight competing methods, which can be roughly categorized into the following three representative groups: Group 1, Traditional Baseline Method: Fine-Tune only applies to incremental data segments. The above fine-tunes the previously trained recommendation model; RETRAIN requires refactoring the entire dataset with each data update. Conduct a thorough retraining.

[0080] The second group consists of knowledge distillation methods: LWC-KD, LWC-KD-PIW, and PISA. These methods utilize knowledge distillation to preserve historical knowledge and constrain the updated recommendation model's predictions to maintain a close resemblance to the predictions of the previous recommendation model.

[0081] The third group consists of experience replay methods: UNIFORM is a classic experience replay baseline that replays historical interactions by uniformly sampling them during incremental updates; INVDEG samples historical interactions inversely proportional to historical popularity; and INFER is a method that samples from the perspective of model influence.

[0082] To ensure a fair comparison, the same experience buffer pool size is maintained.

[0083] Subsequently, the recommendation performance is compared: This invention reports the overall experimental results of different methods, as shown in Table 1, where R and N represent Recall and NDCG, respectively. These results clearly demonstrate that the proposed STAB4SRec consistently achieves the best performance on all three datasets, outperforming existing knowledge distillation methods, experience replay methods, and traditional baseline methods. These experimental results indicate that by jointly utilizing model influence and graph structure information, and explicitly enhancing the predictive stability of the recommendation model on replay data, recommendation performance can be effectively improved.

[0084] Table 1. Average recommendation performance of STAB4SRec and eight competing methods over time.

[0085] As shown in Table 1, the best results are shown in bold, and the second-best results are shown underlined. “Improv.” indicates relative improvement relative to the strongest baseline. All results are reported as mean ± standard deviation of five independent runs and presented as a percentage.

[0086] Observations and analysis of the experiments: First, the traditional FINE-TUNE method exhibits poor performance because it relies solely on interactions within the current data window, overemphasizing plasticity at the expense of stability, leading to severe catastrophic forgetting. Interestingly, the RETRAIN method performs even worse than fine-tuning. This phenomenon can be attributed to the highly dynamic nature of user preferences in real-world scenarios and the substantial noise present in historical data. Indiscriminately training on all historical interactions may cause the model to overfit to this noisy information, thus reducing recommendation performance. These observations highlight the necessity of an effective sampling strategy—selectively retaining key interaction data—to enable the model to remain robust to noise.

[0087] Secondly, knowledge distillation-based methods (such as LWC-KD, LWC-KD-PIW, and PISA) have shown improved results, emphasizing the importance of preserving historical knowledge during incremental updates. These methods mitigate catastrophic forgetting to some extent by transferring key knowledge from training models in previous time windows. However, existing knowledge distillation paradigms primarily rely on indirect consistency constraints and do not pay sufficient attention to the selective preservation of key knowledge, which limits their effectiveness in dynamic incremental learning scenarios.

[0088] Finally, the experience-based replay approach also demonstrated competitive performance, consistently outperforming the Fine-Tune and Retrain baselines. Their advantage over Fine-Tune indicates that revisiting historical data enhances model stability and effectively mitigates catastrophic forgetting. Furthermore, their superior performance compared to the Retrain strategy suggests that indiscriminately using all historical data is not optimal, and selectively sampling representative interactions is crucial to preventing the model from being overwhelmed by noisy data. Meanwhile, the significant performance differences between different replay methods highlight the importance of sampling strategies, as the quality and representativeness of replay samples have a significant impact on recommendation performance. Despite these improvements, all existing replay-based methods still fall short of the STAB4SRec proposed in this invention. This observation implies that simply improving sample quality is insufficient to completely solve the catastrophic forgetting problem, and explicitly enhancing the stability of model predictions on replay data is crucial for achieving robust performance in streaming recommendation scenarios.

[0089] Furthermore, to demonstrate the role of each module in the recommendation model, three sets of ablation experiments were conducted, corresponding to three different variants of the recommendation model. The results of these experiments are summarized in Table 2.

[0090] Table 2 Ablation experimental results of STAB4SRec on three datasets

[0091] Regarding w / o Model (defined in Equation (4)): This experiment analyzes the impact of removing model influence scores during streaming recommendation sampling. Ablation experiments show that the sampling method with model influence awareness is consistently superior to the corresponding simplified version, indicating that model influence information can effectively improve sampling quality and significantly enhance recommendation performance in streaming recommendation scenarios.

[0092] Regarding w / o Structure (defined in Equation (5)): This experiment investigated the impact of removing graph structure information during streaming recommendation sampling. Ablation results showed that the structure-guided sampling strategy can bring stable and consistent performance improvement, indicating that graph structure information helps improve sample quality and enhances recommendation performance in streaming environments.

[0093] Regarding w / o Stab Loss (defined in Equation (9)): This experiment investigated the impact of removing the stability-induced loss function proposed in this invention during the replay stage of streaming recommendation. Ablation experiments showed that the performance of the recommendation model significantly decreased after removing the stability-induced loss, indicating that enhancing the predictive stability of the recommendation model on replay data is crucial for mitigating catastrophic forgetting during incremental updates.

[0094] In embodiments of the present invention, hyperparameter sensitivity analysis was also performed: key hyperparameters were investigated. , and Impact on model performance. Figure 3 The experimental results on the Amazon dataset are presented. Specifically: 1) Used to control the playback ratio. With As the number of users increases, the performance of the recommendation model continues to improve, indicating that experience replay can enhance the stability of the recommendation model and effectively alleviate the catastrophic forgetting problem in streaming recommendation scenarios. 2) Used to balance model influence information and graph structure information during the sampling process. When When a suitable value is chosen, the recommendation model achieves optimal performance, indicating that a reasonable combination of these two types of information is crucial for selecting representative replay samples. 3) This determines the strength of the stability regularization term. With... As the value increases, the performance of the recommendation model improves initially, but when... When the value is too large, the performance actually decreases, reflecting the trade-off between stability and flexibility in streaming recommendation scenarios.

[0095] Figure 4 This is a schematic diagram comparing the average training time of different models for each snapshot window according to embodiments of the present invention, as shown in the figure. Figure 4 As shown, FINE-TUNE has the lowest training cost, but its performance is limited due to the lack of retention of historical knowledge; while RETRAIN exhibits a significantly higher training cost, making it impractical for streaming recommendation. Knowledge distillation-based methods further increase computational overhead by introducing additional constraints, especially on large-scale datasets. Although STAB4SRec incurs slightly higher computational cost than simple replay-based methods, it consistently delivers superior recommendation performance, achieving a favorable balance between effectiveness and efficiency in streaming recommendation.

[0096] In summary, the stability-induced weight perturbation method for streaming recommendation provided in this invention is a novel stability-induced weight perturbation graph experience replay method, aiming to further alleviate the catastrophic forgetting problem in streaming recommendation. First, this method constructs a compact experience buffer by combining model influence information and graph structure information to store influential historical interaction data. Then, to enhance the stability of the recommendation model's predictions for the replay buffer under local parameter drift, this invention introduces a stability-induced loss function based on local worst-case parameter perturbation. This loss function reduces the KL divergence between the recommendation model's predictions for the experience buffer under perturbation and unperturbed settings. Together, these two methods form a complementary framework for streaming recommendation. Theoretically, the stability-induced loss proposed in this invention derives an explicit upper bound for the forgetting metric in streaming recommendation. Finally, this invention provides extensive experimental results and analyses on three widely used benchmark datasets, demonstrating that the method consistently outperforms other state-of-the-art methods in streaming recommendation tasks.

[0097] Example 2 In embodiments of the present invention, an electronic device is also provided. Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention, such as... Figure 5 As shown, the electronic device of the present invention includes a processor 501 and a memory 502, wherein, The memory 502 stores a computer program, which, when read and executed by the processor 501, performs the steps described above in the embodiment of the stability-induced weight perturbation method for streaming recommendation.

[0098] It will be understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A stability-induced weight perturbation method for streaming recommendation, characterized in that, Includes the following steps: Based on the influence information of interaction samples on the recommendation model and their graph structure information, representative samples are selected from historical interaction data and an experience buffer pool is constructed. Based on the local worst-case parameter perturbation applied to the current model parameters of the recommendation model, a stability-induced loss function is defined; the stability-induced loss function is used to constrain the consistency of the prediction distribution of the samples in the experience buffer pool by the recommendation model under parameter perturbation. The stability-induced loss function is combined with the recommendation loss function to obtain the comprehensive objective function of the recommendation model in streaming recommendation, and the parameters of the recommendation model are optimized by minimizing the comprehensive objective function.

2. The stability-induced weight perturbation method for streaming recommendation according to claim 1, characterized in that, The step of selecting representative samples from historical interaction data and constructing an experience buffer pool based on the model influence information and graph structure information of the interaction samples on the recommendation model further includes: Define an influence scoring function to quantify the importance of each interaction sample; The influence scoring function takes the optimal model parameters of the current snapshot and the graph structure information of the interaction samples as input, and the experience buffer pool as output; wherein, the optimal model parameters of the current snapshot are obtained by minimizing the recommendation loss function on the current training set.

3. The stability-induced weight perturbation method for streaming recommendation according to claim 2, characterized in that, The model influence information is quantified by the model influence score of the recommendation model through interactive samples; the model influence score is calculated based on the alignment between the loss gradient of the interactive sample and the aggregated gradient of historical data, and is used to measure the strength of the influence of the interactive sample on the optimization of the recommendation model.

4. The stability-induced weight perturbation method for streaming recommendation according to claim 3, characterized in that, The graph structure information is quantified by a structural influence score; the structural influence score is calculated based on the degree of the user nodes and item nodes associated with the interaction sample. The lower the node degree, the higher the structural influence score of the interaction sample.

5. The stability-induced weight perturbation method for streaming recommendation according to claim 4, characterized in that, Also includes: The model influence score and the structural influence score are normalized respectively; The normalized model influence score and structural influence score are weighted and averaged to obtain the fusion score. All interaction samples are sorted in descending order based on the fusion score, and the top-ranked interaction samples are selected according to a preset number to construct the experience buffer pool.

6. The stability-induced weight perturbation method for streaming recommendation according to claim 1, characterized in that, The stability-induced loss function is defined as the maximum KL divergence of the predicted distribution on the empirical buffer pool under the worst-case perturbation in the neighborhood of the current model parameters of the recommended model; the magnitude of the perturbation is constrained by a preset neighborhood radius.

7. The stability-induced weight perturbation method for streaming recommendation according to claim 6, characterized in that, The stability-induced loss function constrains the knowledge forgetting of the recommendation model in streaming recommendation within an explicit upper bound. The explicit upper bound of knowledge forgetting is jointly determined by the stability-induced loss function and the size of the experience buffer pool. As the stability-induced loss function decreases or the experience buffer pool increases, the upper bound of knowledge forgetting decreases.

8. The stability-induced weight perturbation method for streaming recommendation according to claim 1, characterized in that, The recommendation loss function employs Bayesian personalized ranking loss to capture users' short-term preferences.

9. The stability-induced weight perturbation method for streaming recommendation according to claim 1, characterized in that, In streaming recommendation scenarios, the training set for each time snapshot consists of data from the experience buffer pool that was built at the end of the previous time snapshot and newly arrived interaction data in the current time snapshot. The size of the experience buffer pool is constrained by a preset replay ratio, which represents the proportion of historical interaction data selected into the experience buffer pool.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor is configured to execute the computer program stored in the memory to implement the stability-induced weight perturbation method for streaming recommendation as described in any one of claims 1-9.

Images