A sequence recommendation method and system based on sememe generalization diffusion model

Abstract

The present invention relates to the technical field of sequence recommendation, and provides a sequence recommendation method and system based on a semantic generalization diffusion model, including: obtaining a user's behavior sequence; extracting the user's prior semantic behavior from an item-semetic dependency matrix based on the behavior sequence; statistically analyzing the occurrence frequency of each semantic behavior based on the prior semantic behavior to obtain a statistical sequence, and combining the time interval reweighting strategy to obtain the user's original interest distribution for the semantics; performing a reverse conversion on the user's original interest distribution for the semantics to obtain a generalized interest distribution; and predicting and obtaining a recommended item based on the generalized interest distribution. The space and time complexity of DM is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of sequence recommendation, and in particular relates to a sequence recommendation method and system based on a semantic primitive generalization diffusion model. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Sequential Recommendation (SR) usually infers users' dynamic interests by encoding their sequential behaviors. In the evolution of SR technology, early sequential recommendation methods generally infer users' short-term preferences through Markov chains.

[0004] Inspired by the outstanding performance of diffusion models (DM) in fields such as image generation, semantic segmentation, and machine translation, some researchers have tried to incorporate DM into sequence recommendation. PDRec (Plug-in diffusion model for sequential recommendation, published in the journal AAAIConference on Artificial Intelligence in 2024) proposed three plug-in modules to fully utilize the diffusion preference on item indexes to improve existing sequence modeling algorithms. Since the DM in PDRec is pre-trained based on item indexes, the time complexity of PDRec will be limited by the number of item corpora, thus encountering scalability challenges. Secondly, the diffusion based on item indexes causes PDRec to rely on overlapping users to retrain DM from scratch in CDSR tasks. Finally, the DM based on item indexes is challenging in fully utilizing the association information in the semantic hierarchy of items, which may lead to suboptimal performance results.

[0005] Existing DM-based recommendation algorithms usually deploy DM on discrete item indexes or continuous embeddings of items to inject uncertainty in recommendations. Therefore, the increase in the number of items leads to a doubling of its time and space complexity, making it difficult to deploy in real-world million-level recommendation systems. In addition, these DM-based recommendation algorithms rely only on item indexes, which shows that they lack scalability and cannot fully utilize the semantic information naturally present in the items. Whenever encountering a new dataset or experimental environment, existing algorithms find it difficult to generalize using the semantic information widely present in the corpus, but have to start training from scratch, which leads to a repetitive process that requires a lot of computing power and time consumption. Summary of the invention

[0006] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a sequence recommendation method and system based on the semantic primitive generalization diffusion model, which extracts the user's prior semantic primitive behavior from the project-semene primitive dependency matrix, so that DM can be transferred from the project index domain to the semantic primitive domain, reducing the space and time complexity of DM and improving the sequence recommendation speed.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A first aspect of the present invention provides a sequence recommendation method based on a semantic generalization diffusion model, comprising:

[0009] Get the user's behavior sequence;

[0010] Based on the behavior sequence, extracting the user's prior semantic behavior from the item-semene dependency matrix;

[0011] Based on the prior semantic primitive behavior, the occurrence frequency of each semantic primitive is counted to obtain the statistical sequence, and combined with the time interval reweighting strategy, the original interest distribution of users on the semantic primitive is obtained;

[0012] For the user's original interest distribution for the sememe, reverse conversion is performed to obtain a generalized interest distribution;

[0013] Based on the generalized interest distribution, recommendation items are predicted.

[0014] Furthermore, the step of predicting the recommended items includes: based on the generalized interest distribution, obtaining multiple interest prompt representations through multiple generators, and incorporating the interest prompt representations into the self-attention function to predict the recommended items.

[0015] Furthermore, the time interval re-weighting strategy includes:

[0016] Based on the statistical sequence and in combination with the timestamp sequence at the project level, time interval weights are generated for all sememes;

[0017] The time interval weight is assigned to each sememe, and then the weights of the same sememe are combined to obtain the user's original interest distribution for the sememe.

[0018] Furthermore, the reverse conversion reversely and gradually removes noise in the user's original interest distribution for the sememe through a learnable transition kernel.

[0019] Furthermore, the back-transformation approximates the posterior distribution by minimizing the weighted mean squared error loss.

[0020] Furthermore, the generator is constructed by a two-layer fully connected network with an activation function.

[0021] A second aspect of the present invention provides a sequence recommendation system based on a semantic generalization diffusion model, comprising:

[0022] A data acquisition module is configured to: acquire a user's behavior sequence;

[0023] A priori semantic primitive behavior extraction module is configured to: extract the user's prior semantic primitive behavior from the item-semene primitive dependency matrix based on the behavior sequence;

[0024] The original interest distribution calculation module is configured to: based on the prior semantic primitive behavior, count the occurrence frequency of each semantic primitive to obtain a statistical sequence, and combine the time interval reweighting strategy to obtain the user's original interest distribution for the semantic primitive;

[0025] A reverse conversion module is configured to: perform reverse conversion on the user's original interest distribution of the sememe to obtain a generalized interest distribution;

[0026] A prediction module is configured to predict a recommendation item based on the generalized interest distribution.

[0027] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the sequence recommendation method based on the semantic primitive generalization diffusion model as described above.

[0028] The fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the sequence recommendation method based on the semantic primitive generalization diffusion model as described above are implemented.

[0029] A fifth aspect of the present invention provides a computer program product, which is a computer program product containing a computer program. When a processor executes the computer program, the steps in the sequence recommendation method based on the semantic primitive generalization diffusion model as described above are implemented.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The present invention extracts the user's prior semantic primitive behavior from the project-semene dependency matrix, so that DM can be transferred from the project index domain to the semantic primitive domain; the semantic primitive concept is universal to multiple domains and represents the user's minimum interest unit. Each project in multiple fields can be identified as a combination of multiple semantic primitives. Diffusion at the semantic primitive level can achieve generalization based on semantic information, thereby eliminating the need for training from scratch; the space and time complexity of DM is reduced, so that the diffusion model based on semantic primitives has practical scalability.

[0032] The present invention does not introduce additional noise, but regards the user's original interest distribution for sememes as an inherent noisy sememe distribution, and directly performs a reverse transformation on it to obtain a generalized interest distribution, which not only retains more personalized information to enhance the accuracy of DM, but also ensures that the reasoning process does not start from a completely disordered state, thereby making the reasoning process more robust.

[0033] The present invention converts the user's generalized interest distribution on discrete semantic primitives into multiple continuous prompt representations, which can alleviate the deviation between semantics and behavioral intentions, and guide sequential modeling to understand personalized behavior patterns, thereby enriching user representation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0035] Figure 1 It is a flowchart of a sequence recommendation method based on a semantic generalization diffusion model according to the first embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to enable those skilled in the art to better understand the technical solutions in this application, the following will further illustrate the technical solutions for achieving the invention purpose of this application through several specific embodiments. It should be noted that the technical solutions claimed for protection in this application include but are not limited to the following embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of this application.

[0037] Terminology explanation:

[0038] Sememe: The most basic and smallest unit of meaning that is not easily divisible; for example, the meaning of the word "boy" can be expressed by the combination of the three sememes "human", "male" and "child", and the meaning of "girl" can be expressed by the combination of the three sememes "human", "female" and "child".

[0039] Embodiment 1

[0040] This embodiment provides a sequence recommendation method based on a semantic generalization diffusion model.

[0041] This embodiment provides a sequence recommendation method based on the semantic primitive generalization diffusion model, and proposes an effective and universal semantic primitive-based generalization diffusion framework (SeeDRec, referred to as the semantic primitive generalization diffusion model) for sequential recommendation tasks, which makes full use of the low complexity, cross-domain transferability and semantic relevance of the semantic primitives in the data set to simplify the existing diffusion-based sequence recommendation algorithm. Specifically, inspired by the concept of Sememe in the field of NLP (Natural Language Processing), SeeDRec first defined a similar concept of "recommendation sememe" in the recommendation field to represent the user's smallest interest unit, thereby elevating the diffusion target from the project index level to the sememe level; based on this, SeeDRec proposed a sememe-to-interest diffusion model (S2IDM), which can accurately capture the generalized interest distribution learned by users from local interest evolution and global interest generalization while maintaining low computational cost; subsequently, SeeDRec also proposed an interest-aware prompt enhancement (IPE) strategy, which converts the discrete user interest distribution learned by S2IDM into a continuous prompt representation through multiple prompt generators, thereby better guiding the personalized behavior modeling of each user.

[0042] Different from the complex network structure and random sequence enhancement methods specifically proposed in previous sequence recommendation algorithms, the SeeDRec proposed in this embodiment cleverly utilizes the currently popular diffusion model and incorporates common user dynamic interests into sequence pattern modeling.

[0043] In contrast, SeeDRec proposed in this embodiment creatively proposes S2IDM, which mines the generalized interest distribution of each user by modeling the minimum interest unit on discrete semantics. Therefore, even if a single-layer MLP (Multilayer Perceptron) is used as the reverse approximator, the time complexity of the DM side can be reduced from Simplified to here, represents the time complexity of all other parts of DM except the approximator. The emergence of S2IDM successfully solved the scalability problem that has long plagued PDRec. In addition, SeeDRec also demonstrated the ability to fully utilize semantic association relationships and cross-domain transferability under the item-semite hierarchy. This effectively solves the cross-domain generalization and semantic association problems inherent in PDRec.

[0044] By comparing with existing algorithms, it is found that the SeeDRec proposed in this embodiment is effective and deployable in real-world large-scale real data sets of millions of levels, which is basically consistent with the motivation for designing this algorithm, that is, to help accurately model user preferences through the simplicity, cross-domain transferability and semantic relevance of the diffusion model deployed on the concept of semantic primitives. The experimental results prove that SeeDRec can significantly improve the consistency of existing SOTA sequence recommendation algorithms, and also prove the universality of SeeDRec in data sets without explicit item-category semantic associations, on different types of sequence recommendation algorithms, and on different recommendation tasks.

[0045] Therefore, this embodiment can correspond to specific application scenarios, that is, the proposed universal framework SeeDRec can first construct item-semene hierarchical associations from the existing category information in the WeChat Kanyi platform, and then pre-train the recommendation algorithm based on the diffusion model based on the timing and frequency relationship of the sememes that each user has interacted with. Finally, through the prompt learning paradigm, the generalized user interest distribution learned by the diffusion model is converted into a continuous preference representation for downstream sequence recommendation tasks, thereby improving the recommendation performance of users under the video tab in WeChat Kanyi.

[0046] This example takes SR as an example to explore the DM enhanced recommendation task in various recommendation scenarios. First, define the user The behavior sequence is in is the jth interaction item of user u, and p represents the length of the user's historical behavior. In this embodiment, SASRec is used as a sequence encoder to predict target items that the user may be interested in.

[0047] The model-independent generalized diffusion model SeeDRec based on sememes for sequential recommendation proposed in this embodiment explores the atomic multi-interest units of each user through the diffusion model deployed on the sememes, thereby achieving various recommendation tasks. Figure 1As shown in the figure, the proposed SeeDRec consists of two main parts, including the semantic-to-interest diffusion model (S2IDM) and the interest-aware prompt enhancement strategy (IPE). Specifically, SeeDRec first proposes S2IDM, which creatively performs diffusion at the semantic level rather than the item level used in traditional works, with the aim of modeling the user's local minimum interest unit from the perspective of global interest diffusion. In order to direct the attention of sequential modeling to the personalized interests of users, SeeDRec designs an IPE strategy to transform the diffused generalized interest distribution into a continuous prompt representation through multiple personalized prompt generators. It can accurately extract interest-centric information from the user's historical behavior sequence, thereby achieving accurate user dynamic interest modeling. It is worth noting that SeeDRec is a universal framework that is task-independent and baseline-model-independent, so it can be easily migrated to cross-domain sequential recommendation tasks.

[0048] (1) Diffusion model from semantic origin to interest.

[0049] The overall structure of the semantic source-to-interest diffusion model S2IDM based on the DDPM (Denoising Diffusion Probabilistic Models) framework is as follows: Figure 1 As shown in the left part. Unlike DiffRec (Diffusion recommender model, published in the International Conference on Research on Development in Information Retrieval (SIGIR) in 2023) and PDRec (Plug-In Diffusion Model for Sequential Recommendation, published in The 38th Annual AAAI Conference on Artificial Intelligence (AAAI) in 2024), which directly adopt diffusion on the index of items, the proposed S2IDM creatively deploys DM at the semantic level and optimizes it accordingly according to the mutual correlation between semantics, the overlap of semantics in items, and the drift of multiple interests in behavior sequences.

[0050] In the forward process, S2IDM gradually injects Gaussian noise in a discrete manner to destroy the original semantic preference distribution of each user. Specifically, given the user's behavior sequence, S2IDM first extracts the prior semantic behavior, reweights it by the frequency of occurrence of each semantic and the time interval between adjacent semantics, obtains the original interest distribution, and finally destroys it in a discrete manner.

[0051] Diffusion based on semantic origin. Given the behavior sequence of user u in is the jth interactive item of user u (the jth short video that can be watched by user u), Represents a collection of interactive items (a collection of short videos), Represents the number of interactive items in the set of interactive items; first, from the item-semene dependency matrix Obtain the prior semantic behavior of user u As initial input in Indicates project i p The kth sememe of (e.g., short video 1 corresponds to sememe 3, sememe 5, and sememe 8, short video 2 corresponds to sememe 1, sememe 2, and sememe 5, short video 3 corresponds to sememe 1, sememe 4, sememe 8, and so on, then the prior seme behavior sequence can be expressed as {semme 3, sememe 5, sememe 8; sememe 1, sememe 2, sememe 5; sememe 1, sememe 4, sememe 8; …}). For a short video, a sememe can be funny, educational, or parenting, etc. p represents the pth behavior in the behavior sequence, Representing a set of semantic primitives This allows DM to move from the item index domain to the semantic domain, resulting in three advantages: (1) As defined in [1], it reduces the spatial and temporal complexity of DM, thus making the semantic diffusion model based on semantic primitives practically scalable; (2) due to its basic definition as the minimum interest unit, the inherent multi-domain sharing property of semantic primitives endows it with transferability in the cross-domain sequence recommendation (CDSR) task; (3) it introduces the semantic association relationship inherent in semantic primitives, thus providing comprehensive support for DM.

[0052] Frequency-time-aware reweighting. Given the common one-to-many association between items and sememes, highly related sememes tend to appear repeatedly in user behavior sequences. Therefore, indiscriminately using the reweighting strategy based on sequential relationships or time interval information in T-DiffRec and TI-DiffRec may obscure the user's true interests. To this end, we firstly weight the user behavior sequence based on the prior sememe behavior. The frequency of occurrence of each sememe in the statistic sequence s0 is simply counted to obtain the statistical sequence s0. The size of the statistical sequence s0 is 1×|S|, that is, the length is consistent with the number of sememes. Each element in the statistical sequence s0 represents the number of times each sememe has been interacted in the prior sememe behavior. Then, the time interval reweighting strategy (frequency-time-aware reweighting strategy) in TI-DiffRec is used to perform a reweighting based on the timestamp sequence at the project level. (According to the actual user interaction, for example, the interaction time of the first interactive project is 2022.5.6, the interaction time of the second interactive project is 2023.5.7, ...) is the interaction time of project i k All the sememes Generate time interval weights here is the kth interaction item i in s0 k Finally, these weights are assigned to each sememe, and then the weights of the same sememe are combined to calculate the user's original interest distribution s′0 for the sememe. Specifically, after calculating the time interval weights of all semesmes in each interactive project, the weight values ​​of the same sememes in all interactive projects are combined (summed up) to obtain the user's original interest distribution for the sememe.

[0053] Discrete Noise. Discrete Noise is one of the core stages of the diffusion model, which injects uncertainty into the user's original interest distribution. Different from the existing DM-based recommender, S2IDM uses forward transformation to Gradually destroy the user's original interest distribution for the sememe, and obtain the perturbed sememe distribution s′ t , where β t ∈(0,1) represents the scale of the Gaussian noise added at step t, s′ t represents the user's interest distribution at step t, represents the overall conversion algorithm, and I represents the standard Gaussian distribution. It is worth noting that S2IDM follows the method commonly used in many related studies, generating the noise scale β added in each step through a linear noise scheduling method. t Using the reparameterization technique and the inherent additivity of independent Gaussian distributions, we can obtain in is the added Gaussian noise, α t =1-β t , α t′ From β t Parameter obtained by expansion. t represents the number of diffusion steps. As t→+∞, s′ t Gradually approaching the standard Gaussian distribution.

[0054] Diffusion process. The task of the reverse process in the diffusion model is to gradually remove the added Gaussian noise, thereby obtaining the original distribution s′ from the disturbance. t The user's generalized interest distribution s′0 is recovered from . The related reverse transformation can be defined as The mean μ θ (s′ t ,t) and variance ∑ θ (s′ t ,t) can be modeled using a deep neural network. This is because the relatively small Gaussian noise can ensure that the transformation kernel p θ (s′ t-1 ∣s′ t ) follows a Gaussian distribution. The iterative process can be described as follows:

[0055]

[0056] Optimization strategy. The training goal of S2IDM is to force the reverse transformation p θ (s′ t-1 ∣s′ t ) is close to the posterior distribution q(s′) t-1 ∣s′ t ,s′0). This is done by minimizing the Kullback-Leibler (KL) divergence D KL (q(s′ t-1 ∣s′ t ,s′0)||p θ (s′ t-1 ∣s′ t )) is achieved. Through derivation simplification and importance sampling techniques, the above KL divergence can be rewritten as a weighted mean square error (MSE) loss This allows us to focus on the more difficult denoising task and reduce unnecessary noise:

[0057]

[0058] Among them, x θ (.) is a deep neural network constructed by multiple layers of MLP; represents the sampling probability at each step t, which follows And dynamically update according to the top ten values ​​of each loss item; Express expectations, That is The square of represents the difference between the signal-to-noise ratio (SNR) at step size t and the signal-to-noise ratio (SNR) at step size t-1, x θ (.) is a deep neural network.

[0059] Optimization strategy. The purpose of the reverse process is to recover the user's interaction from the standard Gaussian distribution through a step-by-step denoising transformation process. Specifically, given a Gaussian distribution vector x′ T In the case of θ (x′ t-1 ∣x′ t ) reversely and step by step removes the noise and restores the original interaction. The reverse transition phase can be defined as:

[0060]

[0061] Among them, μ θ (x′ t ,t) and ∑ θ (x′ t ,t) is parameterized by a deep neural network (DNN), and θ represents the model parameters. The complex interactive generation process can be modeled through this step-by-step iterative reverse process.

[0062] Diffusion reasoning. Intuitively, there is inherent occasional noise information in user behaviors collected from the real world, and user generalization interest mining at the semantic level will inevitably introduce a certain degree of bias. Therefore, instead of introducing additional noise, the original semantic distribution s′0 is regarded as an inherently noisy semantic distribution Directly reverse it, that is Get the generalized interest distribution This not only retains more personalized information to enhance the accuracy of DM, but also ensures that the reasoning process does not start from a completely disordered state, making the reasoning process more robust, which just fits the inherent nature of the recommendation task.

[0063] (2) Enhanced interest perception cues.

[0064] Intuitively, there is an insurmountable difference between the diffusion-based user interest distribution and the user's temporal preferences, which motivates the exploration of how to incorporate the generalized interest distribution into sequential recommendation. Moreover, the generalized interest distribution learned from S2IDM (a) is correlated with global user interests and can provide attention cues associated with the sequential recommendation task; (b) it realizes multi-domain associations at the semantic primitive level, which can exploit its inherent semantic associations to discover subtle user preferences and obtain more information. To this end, following the prompt learning paradigm, an interest-aware prompt enhancement (IPE) strategy is proposed to convert the user's generalized interest distribution on discrete semantic primitives into multiple continuous prompt representations. This can alleviate the deviation between semantics and behavioral intentions, and guide sequential modeling to understand personalized behavior patterns, thereby enriching user representations.

[0065] Specifically, given the generalized interest distribution First, through multiple personalized prompt generators f θ (.) It is projected into the same feature space. The generator is constructed by a two-layer fully connected network with a Sigmoid activation function. After the projection, multiple interest prompt representations M = {m1,m2,…,m k}, rewrite the input matrix as in m i Indicates prompt representation, e p represents the embedded representation of the project index (obtained by the model self-learning), p p Embedded representation of the location of the behavior (learned by the model itself); represents the generalized interest distribution; p, k, and d represent the number of user behaviors, the number of prompt representations, and the size of the dimension, respectively. The interest prompt representation M is used to construct the input matrix The pre-trained sequence encoder (FFN) is sent to learn through the self-attention function. Here, the length of the multi-interest prompt is specified as 3 to simplify the training and parameter adjustment process. In the pre-trained sequence model, the generated interest prompt is incorporated into the self-attention function to complete the diffusion of user generalized interest injection, and the recommended items (behavior of the next time step) are predicted. In addition to improving the diffusion target in S2IDM, the low computational complexity of SeeDRec is also reflected in many aspects. The first is the generalized interest distribution used in IPE It only needs to be generated once through the inference process of S2IDM before model training, which makes the time complexity in online service asymptotically similar to its sequence encoder. Second, the deployment of IPE only requires the additional training of a limited number of prompt generators, which also alleviates the space complexity of existing DM-based recommenders. It should be noted that the proposed IPE is a general strategy and does not require unique designs for different sequence modeling algorithms in SR. Therefore, IPE can make full use of possible advances in sequence modeling in the future, thereby extending the life cycle of SeeDRec.

[0066] According to the classic SR method, the binary cross entropy (BCE) loss is selected for each user-item (u,i) pair in the training set As optimization goal:

[0067]

[0068] in refers to the user representation u and the item embedding representation e i The predicted probability between u,i =1 and y u,i = 0 represents positive and negative samples respectively. User representation u and item embedding e iIt is a 64-dimensional feature, which is self-learned. Each feature is represented by 64 double-precision numbers. User representation is obtained by model self-learning. In sequence recommendation, user representation refers to the final representation of the user behavior sequence learned by the sequence encoder (such as SASRec), and has nothing to do with the specific user id (user number).

[0069] This example implements SeeDRec using two traditional sequence recommendation algorithms, SASRec and CL4SRec, and a state-of-the-art diffusion model-based sequence recommendation algorithm, PDRec, as baseline algorithms, and compares them with the individual implementations of these algorithms, thereby comprehensively analyzing the effectiveness and universality of the proposed SeeDRec framework.

[0070] Table 1. Experimental results of sequence recommendation

[0071]

[0072] In the performance comparison, three classic evaluation indicators are selected to evaluate the effectiveness of the proposed SeeDRec, including normalized discounted cumulative gain NDCG@k(N@k), hit rate Hit Rate@k(HR@k) and area under the receiver operating characteristic curve AUC, where k is selected from [1, 5, 10]. 999 negative samples are randomly selected for each positive sample in the test set, and the best results of each method with the same baseline sequence recommendation model are shown in bold in Table 1. The following conclusions are observed:

[0073] (1) Compared with the baseline sequence model, SeeDRec achieves significant improvements on all metrics and datasets with a significance level of p>0.05. This shows that SeeDRec can accurately mine the user's generalized interest distribution through customized multi-intention cue representation based on different types of basic SR models and successfully integrate it into downstream SR tasks.

[0074] (2) In the horizontal comparison of different datasets, it is found that SeeDRec is more advantageous on relatively dense datasets. This is consistent with its basic assumption that S2IDM can better capture the generalized interest distribution of users by diffusing the smallest interest unit on datasets with rich user behaviors.

[0075] (3) SeeDRec, which uses PDRec, the most advanced diffusion model-based sequential recommendation algorithm, as its baseline model, achieves the best results on most datasets. This finding confirms that SeeDRec has a different diffusion model utilization mechanism from PDRec, and also shows that it can continue to play a role based on more advanced DM-based recommendation algorithms in the future.

[0076] The generalized diffusion model based on semantics for sequential recommendation proposed in the present invention includes a prompt perception strategy for modeling user preferences at the semantics level through a diffusion model and migrating the user's generalized interest distribution to the downstream sequential recommendation task, which is mainly designed based on the real interactive relationship between users and projects in the sequential recommendation scenario. Similarly, the proposed SeeDRec can also be migrated to the cross-domain recommendation scenario to demonstrate its superiority and universality. Alternative solutions may exist in different recommendation tasks other than sequential recommendation and cross-domain recommendation, or utilize the category information existing in the data set itself but intentionally ignore the concept of "semite". Therefore, the S2IDM deployed at the semantics level proposed in this embodiment may be modified into a diffusion process specially designed based on the semantic labels or semantic associations of the project itself, thereby realizing the generalized diffusion model based on semantics proposed in the present invention from another perspective.

[0077] Embodiment 2

[0078] This embodiment provides a sequence recommendation system based on a semantic generalization diffusion model, which specifically includes:

[0079] A data acquisition module is configured to: acquire a user's behavior sequence;

[0080] A priori semantic primitive behavior extraction module is configured to: extract the user's prior semantic primitive behavior from the item-semene primitive dependency matrix based on the behavior sequence;

[0081] The original interest distribution calculation module is configured to: based on the prior semantic primitive behavior, count the occurrence frequency of each semantic primitive to obtain a statistical sequence, and combine the time interval reweighting strategy to obtain the user's original interest distribution for the semantic primitive;

[0082] A reverse conversion module is configured to: perform reverse conversion on the user's original interest distribution of the sememe to obtain a generalized interest distribution;

[0083] A prediction module is configured to predict a recommendation item based on the generalized interest distribution.

[0084] It should be noted here that each module in this embodiment corresponds to each step in Example 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0085] Embodiment 3

[0086] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the sequence recommendation method based on the semantic primitive generalization diffusion model as described in the first embodiment above are implemented.

[0087] Embodiment 4

[0088] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the sequence recommendation method based on the semantic primitive generalization diffusion model described in the first embodiment are implemented.

[0089] Embodiment 5

[0090] This embodiment provides a computer program product, which is a computer program product containing a computer program. When a processor executes the computer program, the steps in the sequence recommendation method based on the semantic primitive generalization diffusion model described in the first embodiment are implemented.

[0091] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0092] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0093] The above is only a preferred embodiment of the present application and does not constitute any form of obstruction to the present application. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present application shall fall within the protection scope of the present application.

Claims

1. A sequence recommendation method based on the semantic generalization diffusion model, characterized in that: include: Get the user's behavior sequence; Based on the behavior sequence, extracting the user's prior semantic behavior from the item-semene dependency matrix; Specifically: From the project-semite dependency matrix Obtain the prior semantic behavior of user u As initial input in Indicates project i p The kth sememe of p represents the pth behavior in the behavior sequence, Representing a set of semantic primitives The total length of , which makes the diffusion model transfer from the item index domain to the semantic original domain; Based on the prior semantic primitive behavior, the occurrence frequency of each semantic primitive is counted to obtain the statistical sequence, and combined with the time interval reweighting strategy, the original interest distribution of users on the semantic primitive is obtained; For the user's original interest distribution for the sememe, reverse conversion is performed to obtain a generalized interest distribution; Based on the generalized interest distribution, predict recommended items; The step of predicting the recommended items includes: obtaining multiple interest prompt representations through multiple generators based on the generalized interest distribution, incorporating the interest prompt representations into a self-attention function to predict the recommended items; Specifically, after obtaining the projection, multiple interest prompt representations M = {m1, m2, ..., m k }, rewrite the input matrix as in m i Indicates prompt representation, e p represents the item index embedding representation, p p Embedded representations representing the location of actions; represents the generalized interest distribution; p, k, d represent the number of user behaviors, the number of prompt representations, and the size of the dimension respectively; the interest prompt representation M is used to construct the input matrix It is fed into the pre-trained sequence encoder for learning through the self-attention function.

2. The sequence recommendation method based on the semantic generalization diffusion model according to claim 1, characterized in that: The time interval re-weighting strategy includes: Based on the statistical sequence and in combination with the timestamp sequence at the project level, time interval weights are generated for all sememes; The time interval weight is assigned to each sememe, and then the weights of the same sememe are combined to obtain the user's original interest distribution for the sememe.

3. The sequence recommendation method based on the semantic generalization diffusion model according to claim 1, characterized in that: The reverse transformation reversely and gradually removes noise in the user's original interest distribution for sememes through a learnable transition kernel.

4. The sequence recommendation method based on the semantic generalization diffusion model according to claim 1, characterized in that: The back-transformation approximates the posterior distribution by minimizing the weighted mean squared error loss.

5. The sequence recommendation method based on the semantic generalization diffusion model according to claim 1, characterized in that: The generator is constructed by a two-layer fully connected network with an activation function.

6. A sequence recommendation system based on the semantic generalization diffusion model, characterized in that: include: A data acquisition module is configured to: acquire a user's behavior sequence; A priori semantic primitive behavior extraction module is configured to: extract the user's prior semantic primitive behavior from the item-semene primitive dependency matrix based on the behavior sequence; Specifically: From the project-semite dependency matrix Obtain the prior semantic behavior of user u As initial input in Indicates project i p The kth sememe of p represents the pth behavior in the behavior sequence, Representing a set of semantic primitives The total length of , which makes the diffusion model transfer from the item index domain to the semantic original domain; The original interest distribution calculation module is configured to: based on the prior semantic primitive behavior, count the occurrence frequency of each semantic primitive to obtain a statistical sequence, and combine the time interval reweighting strategy to obtain the user's original interest distribution for the semantic primitive; A reverse conversion module is configured to: perform reverse conversion on the user's original interest distribution of the sememe to obtain a generalized interest distribution; A prediction module, configured to: predict a recommendation item based on the generalized interest distribution; The step of predicting the recommended items includes: obtaining multiple interest prompt representations through multiple generators based on the generalized interest distribution, incorporating the interest prompt representations into a self-attention function to predict the recommended items; Specifically, after obtaining the projection, multiple interest prompt representations M = {m1, M2, ..., m k }, rewrite the input matrix as in m i Indicates prompt representation, e p represents the item index embedding representation, p p Embedded representations representing the location of actions; represents the generalized interest distribution; p, k, d represent the number of user behaviors, the number of prompt representations, and the size of the dimension respectively; the interest prompt representation M is used to construct the input matrix It is fed into the pre-trained sequence encoder for learning through the self-attention function.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of a sequence recommendation method based on a semantic generalization diffusion model as described in any one of claims 1 to 5 are implemented.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the sequence recommendation method based on the semantic generalization diffusion model according to any one of claims 1 to 5 are implemented.

9. A computer program product, which is a computer program product containing a computer program, characterized in that When the processor executes the computer program, the steps in the sequence recommendation method based on the sememe generalization diffusion model as described in any one of claims 1 to 5 are implemented.

Images