A data enhancement method for multi-behavior sequence recommendation

Abstract

The application discloses a kind of data enhancement methods of multi-behavior sequence recommendation, comprising: according to the original sequence of the interaction of user and article is divided into several sub-sequences according to behavior type;Determine the probability of each item in each sub-sequence being sampled based on the sampling strategy of position;According to the determined probability of each item being sampled, the data of several sub-sequences is sequentially disturbed, and the behavior perception data enhancement operation of redundancy reduction and pair exchange is carried out, to obtain the sequence recommendation data after enhancement.The application proposes a general behavior perception data enhancement framework designed for multi-behavior sequence recommendation, generates new samples consistent with user shopping patterns while retaining the information of the original sequence as much as possible through three behavior perception data enhancement operations.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, and in particular to a data augmentation method for multi-behavioral sequence recommendation. Background Technology

[0002] Recommender systems have become an indispensable part of various online platforms, such as e-commerce, streaming services, and social media. They can greatly alleviate information overload and help users find items that match their interests. As user preferences dynamically change over time, effectively utilizing the sequence of user interactions with items is becoming increasingly important. Sequential Recommendation (SR), which aims to predict the next most likely item a user will interact with based on their historical interaction sequences, has become a research hotspot in recent years. However, most existing sequential recommendation models typically focus only on single behavior types, neglecting valuable information between heterogeneous behaviors. This limitation leads to insufficient learning of user preferences by the models, thus prompting research into Multi-Behavior Sequential Recommendation (MBSR).

[0003] Multi-behavior sequence recommendation aims to leverage the dependencies between different user behavior types to enhance the understanding of various sequence patterns, thereby better inferring users' dynamic preferences for target behaviors. With the flourishing development of deep neural networks, models based on recurrent neural networks (RNNs), graph neural networks (GNNs), and Transformers have become mainstream research directions. However, despite the encouraging results achieved by these methods, they still face the challenge of the inherent sparsity problem of real-world data. Due to the lack of comprehensive user behavior information, models often struggle to accurately capture complex cooperative signals, thus hindering model training. To address this issue, data augmentation methods have been proposed and have attracted considerable attention from researchers.

[0004] Among existing multi-behavior sequence recommendation models, research on MBSR is relatively limited and mainly focuses on deep learning. The RLBL model divides heterogeneous sequences into multiple time windows, combining the advantages of RNN and LBL (log-bilinear) networks to capture users' short- and long-term preferences, and introduces behavior-specific transition matrices to model different behavior types. The RIB model encodes each item and behavior in the user's historical interaction sequence into a corresponding embedding vector, captures sequence information through a GRU network, and uses an attention layer to capture the different influences of different behavior types. The BINN model proposes a novel context-aware LSTM network that can simultaneously memorize item and behavior information from heterogeneous sequences. The BAR model designs a behavior attention layer and a task-specific layer to learn the relationship between the target behavior and items under different behavior types, and can convert single-behavior sequence recommendation (SBSR) models into MBSR models. The DyMus model divides the entire sequence into several behavior-specific subsequences and learns sequence representations through dynamic routing.

[0005] In addition, another important research direction is the use of graph neural networks (GNNs) to capture complex behavioral transition relationships between items. For example, the MKM-SR model uses a gated GNN (GGNN) and a GRU layer to learn item features and behavioral representations at each time step, then concatenates them and uses a soft attention mechanism to obtain the final sequence representation. The MGNN-SPred model constructs a multi-relationship item graph (MRIG) to capture behavioral transitions between purchases and clicks, in order to better learn user intent. The GPG4HSR model captures behavioral transitions through a global graph (GG) and introduces a user-specific personalized graph (PG) to mine user personalized interests. Furthermore, it designs a global personalization fusion layer (GPF) to integrate global and personalized information.

[0006] With the significant success of contrastive learning in recommender systems, a series of models utilizing contrastive learning for data augmentation have emerged. The CL4SRec model proposes three augmentation operations: cropping, masking, and reordering. It treats augmented sequences constructed from the same sequence as positive sample pairs and employs a Transformer-based model to encode user representations. The CoSeRec model, building upon CL4SRec, introduces two information-rich augmentation operations: replacement and insertion of relevant items, thereby improving the quality of augmented sequences. The DuoRec model constructs positive and negative samples using both unsupervised and supervised methods. It generates semantically similar but feature-different positive samples using different dropout masks and assumes that sequences with the same target item contain similar user preferences, treating sequences with the same target item as positive sample pairs. The MoCo4SRec model was the first to integrate a momentum updating mechanism into sequence recommendation to enhance the learning of positive and negative samples. However, most of these models are based on the Transformer architecture and are not suitable for short sequences. Another recent research approach relies on pre-trained models to extend original sequences or generate new sequences. The L2Aug model categorizes users into transient and core users, using core users to assist in modeling transient users. The ASRep and BiCAT models extend sequences by generating items preceding them, enhancing the sequence while preserving original information to the greatest extent possible. The GraphDA model first pre-trains a graph-based model to obtain useful user and item representations, then uses a top-k sampling strategy to augment the user-item matrix. In addition, the RSS model considers every item in the sequence to be a potential target item. It employs a recency-based sequence sampling strategy to reconstruct training samples, a completely model-independent method.

[0007] Multi-behavior sequence recommendation (MBSR) models have achieved significant success by modeling the sequentiality and heterogeneity of users' historical interaction sequences to better learn multifaceted user intentions. Despite their effectiveness, the performance of these methods can be limited by data sparsity. Current data augmentation methods in recommender systems primarily focus on modeling single behavior types, neglecting the biases of different behavior types in expressing user preferences. During sample augmentation, such data augmentation methods are prone to introducing excessive interference or noise, thus misleading the recommendation of the next item.

[0008] Therefore, existing technologies still need improvement. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a data augmentation method for multi-behavior sequence recommendation, in order to address the shortcomings of existing technologies and solve the problem that existing multi-behavior sequence recommendation methods cannot achieve data augmentation.

[0010] The technical solution adopted by this invention to solve the technical problem is as follows:

[0011] In a first aspect, the present invention provides a data augmentation method for multi-behavior sequence recommendation, comprising:

[0012] The original sequence of user-item interaction is divided into several sub-sequences based on the behavior type;

[0013] Location-based sampling strategies determine the probability that each item in each subsequence will be sampled;

[0014] Based on the determined probability of each item being sampled, the data of several subsequences are subjected to behavior-aware data enhancement operations such as order perturbation, redundancy reduction, and pairwise swapping to obtain enhanced sequence recommendation data.

[0015] In one implementation, the process of dividing the original sequence of user-item interaction into several sub-sequences based on behavior type includes:

[0016] Each purchase and the click behavior between two consecutive purchases are set as the user's short-term preference within a specific period;

[0017] The original sequence of user-item interactions is divided into several sub-sequences based on the short-term preferences.

[0018] In one implementation, the location-based sampling strategy determines the probability that each item in each subsequence will be sampled, including:

[0019] Based on the order of the items' positions, each item in each subsequence is assigned a corresponding score. Then, at the k-th position, the score g(k) is represented as follows:

[0020] g(k)=λ k

[0021] Calculate the probability f(k) of sampling for each item based on the corresponding score assigned to it:

[0022]

[0023] Where 0≤λ≤1 represents the weighting coefficient.

[0024] In one implementation, the behavior-aware data enhancement operation of perturbing the order of data from several sub-sequences includes:

[0025] Calculate the sampling probability f for each subsequences ;

[0026] Based on the sampling probability f of each subsequence s Extract a specific subsequence from a set of subsequences. s op ;

[0027] Calculate subsequence s op The sampling probability f of each clicked item v ;

[0028] Based on the sampling probability f v The value is used to perform a sequence perturbation operation, resulting in a new subsequence after the sequence perturbation:

[0029]

[0030] Where m and n represent the original sequences respectively. The number of subsequences in s op The length of c represents s op The number of items clicked;

[0031] By perturbing the order of α percent of the original subsequence, the enhanced complete sequence is obtained. Where 0≤α≤1, This is the floor function.

[0032] In one implementation, redundancy reduction and behavior-aware data augmentation operations are performed on several sub-sequences of data, including:

[0033] Select the subsequences whose redundancy needs to be reduced. s RR and for the subsequence s RR Sampling is performed to determine the subsequence. s RR The clicked items to be deleted in the list;

[0034] Delete the subsequence s RR The items to be deleted in the sequence are used to obtain a new sequence with reduced redundancy:

[0035]

[0036]

[0037] use Controlling redundancy reduces the number of repeated operations, resulting in enhanced sequences. Where 0 ≤ β ≤ 1.

[0038] In one implementation, a behavior-aware data augmentation operation involving pairwise swapping of data from several subsequences includes:

[0039] From the original sequence S u Randomly select a source subsequence And obtain its index idx source ;

[0040] Calculate the position-aware sampling probability p(·) of other subsequences:

[0041]

[0042]

[0043] Where p(k) is the sampling probability of the k-th subsequence; 0≤γ≤1 is a parameter that controls the probability distribution;

[0044] Determine the target subsequence based on the position of the source subsequence.

[0045] The source subsequence and the target subsequence are paired and swapped to obtain the enhanced sequence.

[0046]

[0047]

[0048] In one implementation, the method further includes:

[0049] Based on the enhanced sequence recommendation data, diverse training samples are constructed, and these diverse training samples are input into the multi-behavior sequence recommendation model to obtain the trained and optimized multi-behavior sequence recommendation model.

[0050] Secondly, the present invention provides a data augmentation apparatus for multi-behavior sequence recommendation, comprising:

[0051] The sequence segmentation module is used to divide the original sequence of user-item interaction into several sub-sequences based on the behavior type;

[0052] The sampling probability calculation module is used to determine the probability of each item being sampled in each subsequence based on a location-based sampling strategy.

[0053] The behavior-aware data enhancement module is used to perform behavior-aware data enhancement operations such as order perturbation, redundancy reduction, and pairwise swapping on the data of several sub-sequences based on the determined probability of each item being sampled, so as to obtain enhanced sequence recommendation data.

[0054] Thirdly, the present invention provides a terminal, comprising: a processor and a memory, the memory storing a data augmentation program for multi-line sequence recommendation, the data augmentation program for multi-line sequence recommendation being executed by the processor to implement the operation of the data augmentation method for multi-line sequence recommendation as described in the first aspect.

[0055] Fourthly, the present invention also provides a medium, which is a computer-readable storage medium, storing a data augmentation program for multi-row sequence recommendation, which, when executed by a processor, is used to implement the operation of the data augmentation method for multi-row sequence recommendation as described in the first aspect.

[0056] The present invention, by employing the above technical solution, has the following effects:

[0057] This invention divides the original sequence of user-item interactions into several sub-sequences based on behavior type, and determines the probability of each item being sampled in each sub-sequence using a location-based sampling strategy. Based on the determined sampling probability of each item, the data of several sub-sequences undergoes behavior-aware data augmentation operations such as order perturbation, redundancy reduction, and pairwise swapping, thereby obtaining augmented sequence recommendation data. This invention proposes a general behavior-aware data augmentation framework designed for multi-behavior sequence recommendations. Through three behavior-aware data augmentation operations, it generates new samples that conform to the user's shopping patterns while preserving as much information as possible from the original sequence. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0059] Figure 1 This is a flowchart of the data augmentation method for multi-behavior sequence recommendation in this invention.

[0060] Figure 2 This is the overall architecture diagram of data augmentation for multi-behavior sequence recommendation in this invention; Figure 2 (a) is a schematic diagram of the subsequence partitioning; Figure 2 Figure (b) shows a schematic diagram of three data augmentation operations.

[0061] Figure 3 These are the experimental results of the data augmentation operations (sequence perturbation and redundancy reduction) in this invention on different multi-behavior sequence recommendation models.

[0062] Figure 4This is a schematic diagram illustrating the impact of data sparsity on data augmentation for multi-behavioral sequence recommendation in this invention.

[0063] Figure 5 This is a schematic diagram of the data augmentation algorithm for multi-row sequence recommendation in this invention.

[0064] Figure 6 This is a functional schematic diagram of the terminal in one implementation of the present invention.

[0065] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0067] Exemplary methods

[0068] Existing Multi-Behavior Sequence Recommendation (MBSR) models have achieved significant success by modeling the sequentiality and heterogeneity of users' historical interaction sequences to better learn multifaceted user intentions. Despite their effectiveness, the performance of these methods can be limited by data sparsity. Current data augmentation methods in recommender systems primarily focus on modeling single behavior types, neglecting the biases of different behavior types in expressing user preferences. During sample augmentation, such data augmentation methods are prone to introducing excessive interference or noise, thus misleading the recommendation of the next item.

[0069] To address the above technical problems, this invention provides a data augmentation method for multi-behavior sequence recommendation. This method divides the original sequence of user-item interactions into several sub-sequences based on behavior type, and determines the probability of sampling each item in each sub-sequence using a location-based sampling strategy. Based on the determined sampling probability of each item, behavior-aware data augmentation operations such as order perturbation, redundancy reduction, and pairwise swapping are performed on the data of several sub-sequences to obtain augmented sequence recommendation data. This invention proposes a general behavior-aware data augmentation framework designed for multi-behavior sequence recommendation. Through three behavior-aware data augmentation operations, it generates new samples that match the user's shopping patterns while preserving as much information as possible from the original sequence.

[0070] like Figure 1 As shown, this embodiment of the invention provides a data augmentation method for multi-behavior sequence recommendation, including the following steps:

[0071] Step S100: Divide the original sequence of user-item interaction into several sub-sequences according to the behavior type.

[0072] In this embodiment, it is noted that there is currently no data augmentation method specifically designed for multi-behavior sequence recommendation models, and the challenges remain due to short sequences. Therefore, this embodiment proposes a general behavior-aware data augmentation framework (such as...) specifically designed for multi-behavior sequence recommendation models. Figure 2 As shown in the figure, it is called the Multi-Behavior Data Augmentation for Sequential Recommendation (MBASR).

[0073] The data augmentation method for multi-behavior sequence recommendation in this embodiment focuses only on the data level and is a model-agnostic and non-intrusive approach. From the perspective of multi-behavior sequences, this embodiment designs three behavior-aware data augmentation operations. These operations aim to generate new samples that match the user's shopping patterns while preserving as much information as possible from the original sequence.

[0074] Specifically, this embodiment designs three behavior-aware data augmentation operations to construct rich training samples. Each augmentation operation considers the correlation between behaviors and conforms to users' shopping patterns. Simultaneously, this embodiment introduces a location-based sampling strategy, which can effectively reduce the perturbation of the original data caused by the augmentation operations. It is worth noting that this embodiment is entirely data-driven and can be flexibly embedded into different downstream multi-behavior sequence recommendation models; therefore, the entire framework is universal. Extensive experiments on three real-world datasets demonstrate the effectiveness of this embodiment and showcase its applicability to various mainstream multi-behavior sequence recommendation models.

[0075] In this embodiment, the specific problem definition is given first:

[0076] set up and Let represent the sets of users and items, respectively. and These represent the number of users and the number of items, respectively. This example assumes the existence of... Types of behavior, and using This indicates that each user... Associated with a chronologically ordered sequence, each sequence consisting of multiple (item, action) pairs, denoted as... in, Indicates user u's behavior The l-th item to be interacted with.

[0077] Given a multi-action interaction sequence for all users User set Item collection and the set of behavior types Multi-behavior sequence recommendation (MBSR) aims to predict the item that user u is most likely to buy at time step L+1. This problem can be formalized as:

[0078]

[0079] Based on the above definition, the original sequence of user-item interaction is divided into several sub-sequences according to the behavior type, so as to realize the sequence division of behavior perception.

[0080] Specifically, in one implementation of this embodiment, step S100 includes the following steps:

[0081] Step S101: Set each purchase behavior and the click behavior between two consecutive purchase behaviors as the user's short-term preference within a specific period;

[0082] Step S102: Divide the original sequence of user-item interaction into several sub-sequences according to the short-term preferences.

[0083] In practical applications, users typically click on or compare multiple related items before purchasing a particular item. Therefore, the click behavior between each purchase and between two consecutive purchases can be viewed as short-term user preferences within a specific period. Based on this perspective, this embodiment presents the original sequence of user-item interactions. It is divided into several subsequences to better understand users' purchasing behavior and short-term interests.

[0084] like Figure 2 As shown, in Figure 2 In (a), this embodiment describes the sub-sequence division process in detail, where dark circles represent purchase actions and light circles represent click actions. Clearly, Figure 2 The complete user-item interaction sequence in (a) is divided into three sub-sequences. Each sub-sequence begins with a click and ends with a purchase.

[0085] For simplicity, the superscript u and action b for each item u are omitted in this embodiment. Thus, This can be further expressed as follows:

[0086]

[0087] Where s represents a subsequence consisting of multiple items, and m represents The total number of neutron sequences.

[0088] like Figure 1 As shown, in one implementation of this invention, the data augmentation method for multi-behavior sequence recommendation further includes the following steps:

[0089] Step S200: The location-based sampling strategy determines the probability of each item being sampled in each subsequence.

[0090] In this embodiment, after the original sequence is divided into behavior-aware sequences, multiple sub-sequences are obtained; for all sub-sequences, a location-based sampling strategy determines the probability of each item being sampled in each sub-sequence.

[0091] Specifically, in one implementation of this embodiment, step S200 includes the following steps:

[0092] Step S201: Assign a corresponding score to each item in each subsequence based on the order of their positions;

[0093] Step S202: Calculate the probability f(k) of sampling based on the corresponding score assigned to each item.

[0094] In this embodiment, it is assumed that items closer to the current location are inherently more aligned with the user's current preferences and will have a greater impact on downstream tasks. Therefore, the sampling strategy should depend on the order of item positions, meaning that items closer to the current location have a lower probability of being sampled. This embodiment uses g(·) to represent the scoring function, which assigns higher scores to items with lower positional importance. At the k-th position, g(k) can be expressed as follows:

[0095] g(k)=λ k (3)

[0096] This embodiment obtains the sampling probability f(k) using the following method:

[0097]

[0098] Here, 0 ≤ λ ≤ 1 represents the weighting coefficient, which can adjust the importance of different positions, thus affecting the sampling probability. When λ → 0, the items at the beginning of the sequence are more likely to be sampled. When λ → 1, the probability of each item in the sequence being sampled tends to be consistent.

[0099] like Figure 1 As shown, in one implementation of this invention, the data augmentation method for multi-behavior sequence recommendation further includes the following steps:

[0100] Step S300: Based on the determined probability of each item being sampled, perform behavior-aware data enhancement operations such as order perturbation, redundancy reduction, and pairwise exchange on the data of several sub-sequences to obtain enhanced sequence recommendation data.

[0101] In this embodiment, after calculating the probability of each item being sampled in each subsequence, behavior-aware data augmentation can be performed on the data of all subsequences to obtain augmented sequence recommendation data. The augmented sequence recommendation data can be used for the training and optimization process of the multi-behavior sequence recommendation model, thereby enabling the multi-behavior sequence recommendation model to improve the prediction probability of a user purchasing an item in the next moment.

[0102] This embodiment proposes three data augmentation methods in the data augmentation stage: subseq order perturbation (OP), subseq redundancy reduction (RR), and pairwise subseq swapping (PS). Subsequence order perturbation and subsequence redundancy reduction both occur within the subsequence; this embodiment refers to these as intra-subseq augmentation. Pairwise subsequence swapping occurs between subsequences; this embodiment refers to it as inter-subseq augmentation. Figure 2 As shown, in this embodiment... Figure 2 (b) provides examples and demonstrations of these three data augmentation operations, and finally, in Algorithm 1 (such as...) Figure 5 The complete learning algorithm is summarized in (as shown).

[0103] Specifically, in one implementation of this embodiment, the behavior-aware data enhancement operation of perturbing the order of data from several sub-sequences includes the following steps:

[0104] Step S311: Calculate the sampling probability f of each subsequence. s ;

[0105] Step S312, based on the sampling probability f of each subsequence s Extract a specific subsequence from a set of subsequences. s op ;

[0106] Step S313, calculate the subsequence s op The sampling probability f of each clicked item v ;

[0107] Step S314, based on the sampling probability fv The value is used to perform a sequence perturbation operation, resulting in a new subsequence after the sequence perturbation;

[0108] Step S315: Perturb the order of α percent of the original subsequence to obtain the enhanced complete sequence.

[0109] In this embodiment, research on user shopping habits revealed that users often click on a series of similar items before deciding to purchase a particular item. Due to the similarity of the items, these clicks do not strictly adhere to a specific order, and the correlation between consecutively clicked items is not particularly strong. Based on this observation, this embodiment proposes injecting controllable order perturbations into the clicked items in a sub-sequence to introduce diverse sequence patterns.

[0110] Specifically, the original sequence is first calculated according to formulas (3) and (4). The sampling probability f of each subsequence s Then use the calculated probability f s from Extracting a specific subsequence from s op To maintain simplicity of operation, in s op In this embodiment, formulas (3) and (4) are used again to determine the sampling probability f of each clicked item. v Finally, according to f v The value is used to perform the order perturbation operation.

[0111] It is worth noting that, due to f s and f v The calculation follows the viewpoint mentioned above, namely that items closer to the current item have a lower probability of being sampled. Therefore, when... s op When items are disturbed, this embodiment aims to maintain their original relative positions as much as possible, rather than randomly shuffling them. The new sequence can be represented as:

[0112]

[0113]

[0114] Where m and n represent respectively The number of subsequences in s op The length of c is represented by s op Click on the quantity of the item. Then, repeat the above steps. This involves perturbing the original subsequence by a factor of α, where 0 ≤ α ≤ 1. It is a floor function. Ultimately, the enhanced complete sequence can be obtained.

[0115] Specifically, in one implementation of this embodiment, a behavior-aware data enhancement operation to reduce redundancy in several sub-sequences includes the following steps:

[0116] Step S321: Select the subsequences whose redundancy needs to be reduced. s RR and for the subsequence s RR Sampling is performed to determine the subsequence. s RR The clicked items to be deleted in the list;

[0117] Step S322, delete the sub-sequence s RR The items to be deleted in the sequence are used to obtain a new sequence with reduced redundancy;

[0118] Step S323, using Controlling redundancy reduces the number of repeated operations, resulting in enhanced sequences.

[0119] In this embodiment, similarity or redundancy will inevitably occur when a user clicks on items within a short period. For example, a user may click on several similar T-shirts before ultimately purchasing one. Given this phenomenon, it is believed that reducing redundancy can enhance user representation. On the one hand, this preserves user preferences as much as possible. On the other hand, intentionally reducing redundancy allows the model to capture skip-level sequence patterns, thereby enhancing the model's learning ability. Therefore, this embodiment chooses a relatively simple operation: deleting specific clicked items. By randomly deleting some similar or redundant items, this embodiment can simulate the inherent randomness and diversity in the user decision-making process. Furthermore, this method helps prevent the model from relying too heavily on historical user behavior data; this change will facilitate the model's better adaptation to potential changes in user preferences and the evolution of new interests.

[0120] Specifically, firstly, the subsequences are processed according to formulas (3) and (4). s RR Sampling is performed, and then, to further mitigate the information loss caused by deleting items, this embodiment again uses formulas (3) and (4) to sample the clicked items to be deleted. Experience shows that purchased items contain a wealth of valuable information, therefore, the original information of purchased items should be preserved as much as possible. Therefore, this embodiment only deletes clicked items before purchase, and the new sequence can be represented as:

[0121]

[0122]

[0123] Similar to the sequential perturbation operation, this embodiment uses To control the number of times the above operation is repeated, in order to obtain an enhanced sequence. Where 0 ≤ β ≤ 1.

[0124] Specifically, in one implementation of this embodiment, the behavior-aware data enhancement operation of pairwise swapping of data from several sub-sequences includes the following steps:

[0125] Step S331, from the original sequence Randomly select a source subsequence And obtain its index idx source ;

[0126] Step S332: Calculate the position-aware sampling probability p(·) of other subsequences;

[0127] Step S333: Determine the target subsequence based on the position of the source subsequence.

[0128] Step S334: Perform pairwise swapping on the source subsequence and the target subsequence to obtain the enhanced sequence.

[0129] In this embodiment, based on the assumption above that subsequences characterize users' interests at a specific time, this embodiment treats them as the smallest unit, allowing for moderate adjustments to the order between subsequences. Furthermore, to strike a balance between data sample diversity and model performance, this embodiment chooses to perform the swap operation only on a pair of subsequences.

[0130] First, from the sequence Randomly select a source subsequence And obtain its index idx source To minimize noise amplification during the swapping operation, this embodiment designs a sampling probability p(·) based on the position of the source subsequence. Specifically, the subsequence closer to the source subsequence is sampled as the target subsequence. The probability is relatively high. For the k-th subsequence, its sampling probability is denoted as p(k), which can be formally expressed as:

[0131]

[0132]

[0133] Here, 0 ≤ γ ≤ 1 is a parameter controlling the probability distribution. This relative position sampling strategy ensures that positions closer to the source subsequence have higher priority when sampling the target subsequence. Given the original sequence Once the source and target subsequences are determined, this embodiment enhances the sequence by swapping these two subsequences. The final enhanced sequence is obtained

[0134]

[0135]

[0136] Based on the three behavior-aware data enhancement operations described above, the enhanced multi-behavior sequence recommendation data can be obtained.

[0137] In one implementation of this invention, the data augmentation method for multi-behavior sequence recommendation further includes the following steps:

[0138] Step S400: Construct diverse training samples based on the enhanced sequence recommendation data, and input the diverse training samples into the multi-behavior sequence recommendation model to obtain the trained and optimized multi-behavior sequence recommendation model.

[0139] In this embodiment, data augmentation operations can be used to construct diverse training samples, enabling the model to better learn users' dynamic preferences. After obtaining the augmented sequences... Then, it is input into a multi-behavior sequence recommendation (MBSR) model to obtain the user representation h. u The specific formalization is as follows:

[0140]

[0141] in, It is a multi-behavior sequence recommendation (MBSR) model designed to effectively capture complex behavioral dependencies by modeling heterogeneous behaviors in order to better depict dynamic user interests.

[0142] In addition to constructing diverse training samples, the enhanced sequence recommendation data can also be used for prediction and model optimization; specifically, after obtaining the sequence representation and item embeddings, for each candidate item... A user's preference score can be calculated in the following ways:

[0143]

[0144] in, It is item v iThe embedding is then performed. Next, the softmax function is applied to normalize the preference scores, yielding the probability that user u will purchase item v in the next time step. i Predicted probability:

[0145]

[0146] This example uses three datasets from the real world:

[0147] (i)Tmall: A public e-commerce dataset released in the 2015 IJCAI competition, containing six months of user shopping logs.

[0148] (ii)UB: Another e-commerce dataset containing various behaviors released in the 2016 IJCAI competition.

[0149] (iii) JD: A dataset released in the 2019 competition. For these datasets, this embodiment only retains click and purchase records and performs the following preprocessing:

[0150] i) Discard cold start items that have fewer than 10 records in UB and fewer than 20 records in Tmall and JD;

[0151] ii) Discard cold-start users who have fewer than 10 records in Tmall and fewer than 5 records in the other two datasets;

[0152] iii) Sort each dataset by timestamp and keep only the first (user, item, behavior) triplet that is repeated in the sequence;

[0153] iv) Following the LOO (Leave-One-Out) validation method, for each dataset, this embodiment uses the last purchase as the test set, the penultimate purchase as the validation set, and the rest as the training set;

[0154] v) To simulate data sparsity, this embodiment uses a subset of items at equal intervals from each user's interaction sequence for training. Specifically, this embodiment processes Tmall and UB at intervals of 7 and 5 respectively, while JD is not sparsified.

[0155] To evaluate the model's recommendation performance, this embodiment uses two common top-k metrics: hit rate (HR) and normalized depreciation cumulative gain (NDCG). HR@k refers to the proportion of the next item that actually appears among the top k recommended items, emphasizing the model's accuracy. NDCG@k is a location-aware metric that emphasizes item ranking, meaning that items ranked higher are more important.

[0156] To demonstrate the effectiveness of the MBASR proposed in this embodiment, it is integrated into a series of representative models, including some multi-behavior sequence recommendation models based on RNN, attention mechanisms, and GNN.

[0157] This embodiment considers the following six models:

[0158] (i)RLBL: An RNN-based model that combines RNN and LBL (log-bilinear) to capture long-term and short-term preferences.

[0159] (ii)RIB: An RNN-based model that uses the concatenation of item embeddings and behavior embeddings as input to a GRU layer, and then employs an attention layer to distinguish the effects of different behavior types.

[0160] (iii) BINN: An RNN-based model that designs a contextual long short-term memory (CLSTM) network to integrate the user’s history and current preferences.

[0161] (iv) GRUBAR: An extended model of BAR, which uses GRU4Rec as its backbone model. In BAR, a behavior attention layer is used to encode the relationship between the target behavior and items with different behaviors.

[0162] (v)SASBAR: An extended model of BAR that uses SASRec as its backbone model. SASRec is an attention-based model that uses multi-head attention to learn sequence representations.

[0163] (vi)GPG4HSR: A GNN-based model that constructs a global graph to capture transitions between different behavior types and leverages personalized graphs to enhance sequence representations with contextual information.

[0164] Furthermore, this embodiment compares the performance of MBASR after its enhancement with that of the highly advanced data augmentation method RSS, because RSS, like this embodiment, is a data-oriented augmentation method that can ensure the fairness of the experiment. RSS uses a recency-based sampling strategy to generate new training samples, so that every item in the sequence can be selected as the target item.

[0165] The main experimental results are reported in Tables 1 and 2 of this embodiment.

[0166] In Table 1, this embodiment applies the pairwise exchange (SS) data augmentation operation to MBASR and integrates it into various multi-behavior sequence recommendation models. The best results are marked in bold, and "Imprv." indicates the improvement of MBASR over the original model on HR@10. Observing Table 1, it can be seen that the MBASR proposed in this embodiment achieves significant improvements over various mainstream MBSR models on all datasets, demonstrating the superiority and applicability of MBASR.

[0167] Specifically, in terms of average performance across the three datasets, MBASR achieved performance improvements of 25.17%, 4.55%, 5.5%, 7.62%, 10.20%, and 4.18% on the HR@10 metric, respectively, across RLBL, RIB, BINN, GRUBAR, SASBAR, and GPG4HSR. On one hand, this embodiment views the historical interaction sequences between users and items from a more granular perspective, constructing simple yet information-rich samples through specific behavior-aware data augmentation operations. This helps the model capture additional sequence patterns and enhances its robustness and generalization ability. On the other hand, this embodiment mitigates the impact of noise on model performance through a location-based sampling strategy. The combination of these two factors ultimately leads to excellent performance.

[0168] Table 1 shows the experimental results on different MBSR models using (w / ) or not using (w / o) MBASR on three datasets.

[0169]

[0170]

[0171] To demonstrate the superiority of MBASR, this embodiment compares it with the state-of-the-art data augmentation model RSS. Specifically, this embodiment uses SASRec and SASBAR as base models to evaluate the performance improvements brought by RSS and the proposed MBASR from the perspectives of single-behavior sequence recommendation and multi-behavior sequence recommendation, respectively. The results are shown in Table 2, with the best results highlighted in bold and the second-best results underlined. The following key conclusions can be drawn from Table 2:

[0172] First, MBASR outperforms RSS on almost all datasets and shows significant performance improvements on representative SBSR and MBSR models. These results further demonstrate the superiority and applicability of MBASR. Compared to the original model, MBASR has the advantage of learning richer sequence representations from newly constructed samples, which is a key factor in its significant performance improvement.

[0173] Secondly, it can be observed that SASRec, which integrates RSS, achieves significant performance improvements on both the UB and JD datasets. However, when RSS is combined with the MBSR model SASBAR, it performs poorly in most cases. These findings indicate that RSS is ineffective in MBSR scenarios because it does not consider the impact of different behavior types on user preferences. This further highlights the advantages of the MBASR proposed in this embodiment in addressing the sparsity problem of multi-behavior sequences.

[0174] Table 2 compares the performance of the data augmentation model RSS with that of MBASR in this embodiment. MBASR in this embodiment uses pairwise swapping as the data augmentation method.

[0175]

[0176] To more comprehensively verify the contributions of different data augmentation operations, this embodiment conducted a set of more detailed experiments. For example... Figure 3 As shown, this example demonstrates the performance of two data augmentation operations (i.e., order perturbation and redundancy reduction) combined with various MBSR models on three datasets. As... Figure 3 As shown, the experimental results are similar to the trends observed in Table 1. On each dataset, each data augmentation method proposed in this embodiment delivers varying degrees of performance improvement across all downstream tasks. Taking Tmall as an example, the sequential perturbation operation (OP) achieves performance improvements of 21.01%, 7.04%, 11.07%, 10.57%, 9.33%, and 7.49% on the HR@10 metric for RLBL, RIB, BINN, GRUBAR, SASBAR, and GPGHSR, respectively. Correspondingly, for the redundancy reduction operation (RR), it achieves performance improvements of 24.37%, 2.05%, 9.69%, 2.48%, 4.26%, and 5.87% on the HR@10 metric for RLBL, RIB, BINN, GRUBAR, SASBAR, and GPGHSR, respectively. This further demonstrates the universality of MBASR and validates the rationality of the design of each data augmentation operation. It is worth noting that each enhancement method designed in this embodiment conforms to real shopping patterns, providing strong evidence for its feasibility in practical applications.

[0177] To investigate the impact of data sparsity on MBASR, this embodiment intentionally introduces sparsity into the UB dataset. Items are drawn at different intervals (1, 3, 5) within each user sequence. This process produces three sparsified datasets, which are denoted in this embodiment as UB(1), UB(3), and UB(5), representing progressively increasing sparsity. Figure 4As shown, this embodiment demonstrates the HR@10 results under three different data augmentation operations when MBASR uses SASBAR as the backbone model on these three datasets.

[0178] The results were consistent with expectations: the performance of HR@10 gradually deteriorated as data sparsity increased. Meanwhile, MBASR's ability to improve the original model continued to rise. This indicates that MBASR demonstrates a stronger ability to handle increasingly sparse data.

[0179] This embodiment achieves the following technical effects through the above technical solution:

[0180] This embodiment divides the original sequence of user-item interactions into several sub-sequences based on behavior type, and determines the probability of each item being sampled in each sub-sequence using a location-based sampling strategy. Based on the determined sampling probability of each item, behavior-aware data augmentation operations such as order perturbation, redundancy reduction, and pairwise swapping are applied to the data of several sub-sequences to obtain augmented sequence recommendation data. This embodiment proposes a general behavior-aware data augmentation framework designed for multi-behavior sequence recommendations. Through three behavior-aware data augmentation operations, it generates new samples that conform to the user's shopping patterns while preserving as much information as possible from the original sequence.

[0181] Exemplary device

[0182] Based on the above embodiments, the present invention also provides a data augmentation device for multi-behavior sequence recommendation, comprising:

[0183] The sequence segmentation module is used to divide the original sequence of user-item interaction into several sub-sequences based on the behavior type;

[0184] The sampling probability calculation module is used to determine the probability of each item being sampled in each subsequence based on a location-based sampling strategy.

[0185] The behavior-aware data augmentation module is used to perform behavior-aware data augmentation operations such as order perturbation, redundancy reduction, and pairwise swapping on the data of several sub-sequences based on the determined probability of each item being sampled, so as to obtain the augmented sequence recommendation data.

[0186] This embodiment achieves the following technical effects through the above technical solution:

[0187] This embodiment divides the original sequence of user-item interactions into several sub-sequences based on behavior type, and determines the probability of each item being sampled in each sub-sequence using a location-based sampling strategy. Based on the determined sampling probability of each item, behavior-aware data augmentation operations such as order perturbation, redundancy reduction, and pairwise swapping are applied to the data of several sub-sequences to obtain augmented sequence recommendation data. This embodiment proposes a general behavior-aware data augmentation framework designed for multi-behavior sequence recommendations. Through three behavior-aware data augmentation operations, it generates new samples that conform to the user's shopping patterns while preserving as much information as possible from the original sequence.

[0188] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 6 As shown.

[0189] The terminal includes: a processor, a memory, an interface, a display screen, and a communication module connected via a system bus; wherein, the processor of the terminal provides computing and control capabilities; the memory of the terminal includes a storage medium and internal memory; the storage medium stores the operating system and computer programs; the internal memory provides an environment for the operation of the operating system and computer programs in the storage medium; the interface is used to connect to external devices; the display screen is used to display relevant information; and the communication module is used to communicate with a cloud server or other devices.

[0190] When executed by a processor, this computer program is used to implement a data augmentation method for multi-beginner sequence recommendations.

[0191] It will be understood by those skilled in the art that Figure 6 The schematic diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0192] In one embodiment, a terminal is provided, comprising: a processor and a memory, the memory storing a data augmentation program for multi-line sequence recommendation, the data augmentation program for multi-line sequence recommendation being executed by the processor to implement the operations of the data augmentation method for multi-line sequence recommendation as described above.

[0193] In one embodiment, a storage medium is provided, wherein the storage medium stores a data augmentation program for multi-row sequence recommendation, which, when executed by a processor, is used to implement the operations of the data augmentation method for multi-row sequence recommendation as described above.

[0194] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, database, or other media used in the embodiments provided by this invention can include both non-volatile and volatile memory.

[0195] In summary, this invention provides a data augmentation method for multi-behavior sequence recommendation. The method includes: dividing the original sequence of user-item interaction into several sub-sequences according to behavior type; determining the probability of each item being sampled in each sub-sequence using a location-based sampling strategy; and performing behavior-aware data augmentation operations such as order perturbation, redundancy reduction, and pairwise swapping on the data of the several sub-sequences according to the determined probability of each item being sampled, to obtain augmented sequence recommendation data. This invention proposes a general behavior-aware data augmentation framework designed for multi-behavior sequence recommendation. Through three behavior-aware data augmentation operations, it generates new samples that conform to the user's shopping pattern while preserving as much information as possible from the original sequence.

[0196] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A data augmentation method for multi-behavior sequence recommendation, characterized in that, include: The original sequence of user-item interaction is divided into several sub-sequences based on the behavior type; Location-based sampling strategies determine the probability that each item in each subsequence will be sampled; Based on the determined probability of each item being sampled, the data of several subsequences are subjected to behavior-aware data enhancement operations such as order perturbation, redundancy reduction, and pairwise exchange to obtain enhanced sequence recommendation data; The original sequence of user-item interaction is divided into several sub-sequences based on behavior type, including: Each purchase and the click behavior between two consecutive purchases are set as the user's short-term preference within a specific period; The original sequence of user-item interactions is divided into several sub-sequences based on the short-term preferences. The location-based sampling strategy determines the probability of each item being sampled in each subsequence, including: Based on the order of the items' positions, each item in each subsequence is assigned a corresponding score. Then, in the [number]th [sequence]... In each position, the score It is expressed as follows: The probability of being sampled is calculated based on the corresponding score assigned to each item. : in, Indicates the weighting coefficient; The behavior-aware data augmentation operation that perturbs the order of data from several sub-sequences includes: Calculate the sampling probability of each subsequence ; Based on the sampling probability of each subsequence Extract a specific subsequence from a set of subsequences. ; Calculate subsequence Sampling probability of each clicked item ; Based on sampling probability The value is used to perform a sequence perturbation operation, resulting in a new subsequence after the sequence perturbation: in, and Representing the original sequence The number of subsequences in Length, express The number of items clicked; For the original subsequence of percent The subsequences are perturbed sequentially to obtain the enhanced complete sequence. ;in, , This is the floor function; Behavior-aware data augmentation operations involving pairwise swapping of several subsequences of data include: From the original sequence Randomly select a source subsequence And obtain its index. ; Calculate the position-aware sampling probability of other subsequences : in, For the first The sampling probability of each subsequence; It is a parameter that controls the probability distribution; Determine the target subsequence based on the position of the source subsequence. ; The source subsequence and the target subsequence are paired and swapped to obtain the enhanced sequence. : 。 2. The data augmentation method for multi-behavior sequence recommendation according to claim 1, characterized in that, Redundancy reduction and behavior-aware data augmentation operations are performed on several subsequences of data, including: Select the subsequences whose redundancy needs to be reduced. and for the subsequence Sampling is performed to determine the subsequence. The clicked items to be deleted in the list; Delete the subsequence The items to be deleted in the sequence are used to obtain a new sequence with reduced redundancy: use Controlling redundancy reduces the number of repeated operations, resulting in enhanced sequences. ;in, .

3. The data augmentation method for multi-behavior sequence recommendation according to claim 1, characterized in that, The method further includes: Based on the enhanced sequence recommendation data, diverse training samples are constructed, and these diverse training samples are input into the multi-behavior sequence recommendation model to obtain the trained and optimized multi-behavior sequence recommendation model.

4. A data augmentation apparatus for multi-behavior sequence recommendation, used to implement the data augmentation method for multi-behavior sequence recommendation as described in any one of claims 1-3, characterized in that, include: The sequence segmentation module is used to divide the original sequence of user-item interaction into several sub-sequences based on the behavior type; The sampling probability calculation module is used to determine the probability of each item being sampled in each subsequence based on a location-based sampling strategy. The behavior-aware data enhancement module is used to perform behavior-aware data enhancement operations such as order perturbation, redundancy reduction, and pairwise swapping on the data of several sub-sequences based on the determined probability of each item being sampled, so as to obtain enhanced sequence recommendation data.

5. A terminal, characterized in that, include: The processor and memory, the memory storing a data augmentation program for multi-row sequence recommendation, which, when executed by the processor, is used to implement the operation of the data augmentation method for multi-row sequence recommendation as described in any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a data augmentation program for multi-row sequence recommendation, which, when executed by a processor, is used to implement the operation of the data augmentation method for multi-row sequence recommendation as described in any one of claims 1-3.

Images