Recommendation generation method, system and equipment based on potential space semantic enhancement and conditional diffusion modeling and storage medium

Abstract

The invention discloses a recommendation generation method, system and device based on potential space semantic enhancement and conditional diffusion modeling and a storage medium. The method comprises the steps of firstly collecting user interaction data; extracting articles from the interactive data, embedding and dividing article categories, and compressing an interactive vector into a low-dimensional vector by using variational self-coding; secondly, generating latent vector representation, and performing semantic enhancement and regularization processing on sparse user latent vectors in the forward process of the conditional diffusion model; guiding a reverse denoising process of the conditional diffusion model through a Transform encoder and a cross attention mechanism, and generating an original latent vector representation; dividing the latent vector representation according to article categories, and predicting an interaction vector between a user and an article; and finally, carrying out training reasoning on the conditional diffusion model to generate Top-K recommendation items. According to the method, sparsity and noise influence caused by a high-dimensional interaction matrix are reduced, and personalized article representation conforming to user preferences is generated.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data mining and recommendation system, in particular to a recommendation generation method, system, device and storage medium based on latent space semantic enhancement and conditional diffusion modeling. BACKGROUND

[0002] With the continuous popularization of Internet technology applications, e-commerce, entertainment and social media platforms have generated a large amount of user behavior data. As an important means to solve information overload, recommendation systems have played a key role in various application scenarios. Past recommendation methods mostly rely on traditional techniques such as collaborative filtering, matrix factorization, deep neural networks, etc., to predict user preferences by mining the interaction between users and items. However, these methods have certain limitations in modeling the complex interest distribution of users and capturing the characteristics of long-tail items, which can easily lead to single recommendation results and decreased accuracy.

[0003] In recent years, generative models have made significant progress in computer vision, natural language processing and other fields. Diffusion models (DM) have shown better generation quality than traditional generative adversarial networks (GAN) and variational autoencoders (VAE) due to their strong data distribution modeling capabilities. In the recommendation task, the goal of the recommendation model is similar to that of the diffusion model, because the recommendation model essentially infers the probability of future interactions based on user interaction history containing noise (false positives, false negatives). Therefore, exploring diffusion models for recommendations has great potential to more accurately simulate complex interactions through powerful representation capabilities.

[0004] Existing diffusion recommendation models based on user-item interaction matrices mostly directly diffuse and inverse diffuse in the high-dimensional user-item interaction space, or use user behavior history and context as conditions to control the generation process. Although these methods can achieve personalized recommendations to some extent, they have high computational overhead and convergence difficulties due to the high dimensionality, sparsity and noise of the user interaction matrix. In addition, the quality of the latent vectors after sparse coding is poor, and the recommendation results are also easily affected by abnormal interactions, affecting the stability and quality of the recommendation results. At the same time, in the diffusion model, if only traditional reverse denoising processes are used to gradually obtain the predicted original vector, there are many challenges, such as how to effectively recover the personalized information of the original input while eliminating most of the noise. SUMMARY

[0005] To solve the problems raised in the background art, the present application proposes a recommendation generation method, system, device and storage medium based on latent space semantic enhancement and conditional diffusion modeling.

[0006] In a first aspect, the present application provides a recommendation generation method based on latent space semantic enhancement and conditional diffusion modeling, comprising the following steps:

[0007] Collecting multi-source interaction data of users in a service system, sorting and screening according to timestamps, to construct interaction vectors between users and items and obtain user interaction sequences;

[0008] Using a LightGCN model to obtain embedding representations of each item from an item set, classifying items through K-means clustering, and simultaneously compressing the interaction vectors between the users and the items into low-dimensional vectors using a variational autoencoder;

[0009] Generating latent vector representations of users by connecting the low-dimensional vectors, performing semantic enhancement and regularization processing on sparse user latent vectors, and perturbing historical interaction information of the user latent vectors in the low-dimensional latent space during the forward process of the conditional diffusion model;

[0010] Encoding the user interaction sequences through a Transformer historical interaction sequence encoder, dynamically guiding the reverse denoising of the conditional diffusion model through a cross-attention mechanism, and gradually generating original latent vector representations from data contaminated by Gaussian noise;

[0011] Dividing the generated original latent vector representations according to item categories, and predicting interaction vectors between users and items through decoders corresponding to the categories;

[0012] Training and reasoning the conditional diffusion model, constructing its total loss, and generating Top-K recommended items according to item rankings in the reasoning phase.

[0013] In a second aspect, the present application provides a recommendation generation system based on latent space semantic enhancement and conditional diffusion modeling, comprising:

[0014] A collection module for collecting multi-source interaction data of users in a service system, sorting and screening according to timestamps, to construct interaction vectors between users and items and obtain user interaction sequences;

[0015] An encoder module for using a LightGCN model to obtain embedding representations of each item from an item set, classifying items through K-means clustering, and simultaneously compressing the interaction vectors between the users and the items into low-dimensional vectors using a variational autoencoder;

[0016] A forward process module for generating latent vector representations of users by connecting the low-dimensional vectors, performing semantic enhancement and regularization processing on sparse user latent vectors, and perturbing historical interaction information of the user latent vectors in the low-dimensional latent space during the forward process of the conditional diffusion model;

[0017] The reverse process module is used to encode the user interaction sequence through the Transformer historical interaction sequence encoder, and then dynamically guide the conditional diffusion model to reverse denoising through the cross-attention mechanism, so as to gradually generate the original latent vector representation from the data contaminated by Gaussian noise.

[0018] The decoder module is used to divide the generated original latent vector representations into item categories, and predict the interaction vectors between users and items through the decoder of the corresponding category;

[0019] The training and inference module is used to train and infer the conditional diffusion model, construct its total loss, and generate Top-K recommended items based on the item ranking during the inference phase.

[0020] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the aforementioned recommendation generation method based on latent spatial semantic enhancement and conditional diffusion modeling.

[0021] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps in the aforementioned recommendation generation method based on latent spatial semantic enhancement and conditional diffusion modeling.

[0022] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the aforementioned recommendation generation method based on latent spatial semantic enhancement and conditional diffusion modeling.

[0023] The beneficial effects of this invention are as follows: By introducing item clustering and multiple encoders, this invention performs diffusion modeling of user interaction representations in a latent low-dimensional space, effectively reducing the sparsity and noise effects caused by high-dimensional interaction matrices. This overcomes the problems of high computational cost and convergence difficulties caused by directly modeling in the original interaction space in existing recommendation models. Simultaneously, by combining historical interaction sequences encoded by multi-layer Transformer as conditional vectors and introducing a cross-attention mechanism in the diffusion denoising stage, the generation process is dynamically guided by user history in the latent space, thereby generating more personalized item representations that better match user preferences. This invention can provide more accurate, diverse, and efficient personalized recommendation services in real-world large-scale recommendation scenarios, and has significant application value and potential for widespread adoption. Attached Figure Description

[0024] 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 these drawings without creative effort.

[0025] Figure 1 A schematic diagram of the system architecture for a recommendation generation method based on latent spatial semantic enhancement and conditional diffusion modeling provided by the present invention;

[0026] Figure 2 This is a schematic diagram of the Sparse User Latent Vector Enhancement Module (SLEM) model provided by the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please refer to Figure 1 This invention provides a recommendation generation method based on latent spatial semantic enhancement and conditional diffusion modeling, comprising the following steps:

[0029] S1: Collect user interaction data in the service system, and construct the data structure and content required for this recommendation method by dividing and building it.

[0030] Optionally, user interaction data within the service system can be collected. This data can originate from user interaction data such as book purchases, movie viewings, and restaurant reviews. First, all interactions are sorted chronologically by timestamp, and user interactions with ratings less than 4 are removed. This constructs the interaction vector between users and items. For a given user... Its collection of items The history of interaction, namely ,in =1 or 0 indicates the user Is it related to the item? The interaction has occurred. After obtaining the user interaction vector, the user's... The sequence of the last n interaction items ,in This represents the n most recent real interaction records of a user arranged in chronological order. This interaction sequence is encoded and used to guide the reverse process of diffusion, and to characterize the user's short-term interest dynamics and immediate preference changes.

[0031] S2: Use the K-means clustering method to cluster each item into different categories based on the item representation, and obtain the user's representation on different categories of items based on the user interaction vector. Then, use different variational autoencoders to compress the user representation on different categories into low-dimensional vectors.

[0032] Optionally, the item set can be obtained in advance using the LightGCN model. Each item is represented by an embedding, and then the item set is clustered into groups using the K-means clustering method. Categories Then, for these categories, user interaction vectors are... Divided into Each part, namely ,in Indicates user exist Interaction on the screen.

[0033] Then use a by A parameterized variational autoencoder will convert each Compressed into a low-dimensional vector The encoder predicts and As a variational distribution The mean and variance are described below:

[0034]

[0035] in, Indicates a multivariate Gaussian distribution. This indicates that the encoder network responds to the input. The output latent mean vector, It is the encoder network for the input The output latent variance vector, This represents the set of learnable parameters for the encoder of the corresponding category.

[0036] Optionally, S2 uses clustering to help identify the latent category structure in the interactive data. The conditional diffusion model propagates in the latent space after clustering, which can better capture the category information in the data and ensure that the category structure and more personalized information about the user are preserved during the generation process. At the same time, in the low-dimensional latent space, the training and generation processes can achieve parallel computation of different categories, breaking the full connection between multiple encoders to save parameters.

[0037] S3: As Figure 2As shown, by connecting low-dimensional vectors of different categories, the latent vector representation of the user in the latent space is obtained. The latent vectors are semantically enhanced and regularized to improve the sparse user representation. A forward diffusion process is performed on the user latent vectors in the low-dimensional latent space to disrupt the user's historical interactions in the data.

[0038] Optionally, via series Obtain the latent vector for the user in the latent space To address the issues of unstable latent representation quality and degraded diffusion generation performance when user interaction data is sparse, semantic enhancement and regularization are performed on sparse user latent vectors before the latent space diffusion process is executed.

[0039] Optionally, adaptive augmentation weights are first calculated based on the sparsity of user interactions. When users are sparse, similar users are obtained based on rating similarity to supplement latent vector information. The sparsity is calculated for each user:

[0040]

[0041] in This indicates the number of a user's historical interactions. That is, the maximum number of interactions. A larger value indicates sparser user interactions, and then sparse weights are generated. ,in For the Sigmoid function, and For trainable parameters, such that The latent vector enhancement formula is as follows:

[0042]

[0043] in For a set of users similar to the user, For latent vectors of similar users, The weights are based on the cosine similarity of the ratings. To prevent the enhanced latent vectors from deviating too far from the original distribution, a KL regularization term is added.

[0044]

[0045] in For Kullback-Leibler divergence, This represents the regularization loss of the similar user latent vector augmentation module, used to constrain the distribution of the augmented latent variables. Distribution of original latent variables Consistency.

[0046] Optionally, after obtaining the enhanced user latent vector representation, parameter transformation is performed during the forward process of the diffusion step. ,in, and represent the mean and variance of the Gaussian distribution, respectively. Controlling every step The added Gaussian noise scale. Using reparameterization techniques and the additivity of two independent Gaussian noises, it is possible to directly obtain the scale from... get The details are as follows:

[0047]

[0048] in This indicates the proportion of signal retained at each diffusion step. This represents the cumulative signal hold rate from step 1 to step t, and then it can be... Reparameterization, where To regulate The added noise is 1- A linear noise scheme was designed, namely,

[0049] ,

[0050] hyperparameters Controlling the noise scale, two hyperparameters This indicates the upper and lower limits for adding noise.

[0051] S4: Guide the reverse process of diffusion by extracting the user's most recent n interaction history records, and use the cross-attention conditional diffusion model to learn how to gradually recover the original low-dimensional representation from the data contaminated by Gaussian noise;

[0052] Optionally, in order to make full use of the user's historical interaction signals during the training and reverse generation of the diffusion model and enhance the personalization and controllability of the denoising process, the reverse process is based on the Transformer's historical interaction sequence encoder and cross-attention conditional module.

[0053] Optionally, in conventional diffusion recommendation models, directly denoising latent vectors lacks guidance from historical interactions, which can easily lead to a lack of consistency in the generated item representations with user features. To address this issue, the user's historical interaction sequence is first analyzed. Perform embedding mapping to obtain the corresponding item embedding sequence. Then, a multi-layer Transformer encoder is used to process the sequence. Encode, that is Interactive sequence encoding As a guiding condition, the noise reduction process should be guided in the following manner:

[0054]

[0055] Among them, time step embedding It is obtained by mapping sinusoidal time-position encoding with a multilayer perceptron, and a cross-attention conditional diffusion module is designed based on this. , The output is taken as the mean of a Gaussian distribution, representing the model's performance at time step embedding t. The prediction is used to iteratively learn the denoising process from the latent space. This is done to restore the user representation. The cross-attention conditional diffusion module includes a feedforward sub-layer and a layer normalization structure, and residual connections are used in the denoising network to ensure training stability. The specific implementation is as follows:

[0056]

[0057]

[0058]

[0059]

[0060]

[0061] in (query) is determined by the current noisy latent variable. It is obtained by linearly combining it with the time step embedding t and mapping it to represent the semantic information that the model wants to query at the current time step; (key) is encoded by the user's historical interactions. It is obtained through linear mapping, representing the key semantics of each position in the historical behavior sequence, and matched with q for attention; (value) by The mapped information represents the content information that can be aggregated into the denoising path, and is used to construct the attention-weighted output. , , Learnable parameter matrix. For attention weights, The denoising mean function is used; the MLP is a feedforward sublayer that explicitly injects the user's historical sequence encoding into the denoising path. This step achieves deep integration of time-step conditionalization and user history conditionalization through a cross-attention mechanism, enabling the diffusion model to dynamically consider user preference information when generating latent representations.

[0062] S5: Reconstructed from the reverse-guided denoising process Categorized by item type Then each Input to by In the parameterized decoder, through predict .

[0063] S6: Model training and inference, constructing the model's total loss and generating Top-K recommendations.

[0064] Optionally, in order to learn the parameters of the cross-attention network Conditional diffusion models aim to maximize the generation in the latent space. ELBO (Expected Log-Likelihood):

[0065]

[0066] The reconstruction term is represented in The negative reconstruction error reflects whether the conditional diffusion model can accurately reconstruct the original latent variables at the end of denoising. The prior matching term is a constant without trainable parameters, so it can be ignored during optimization. The denoising matching term is forced by KL divergence. Approximately tractable distributions In this way, Optimized to iterate from recover .

[0067] The estimation of the denoised matching term, according to the Bayesian rule, It can be rewritten in the following form:

[0068]

[0069]

[0070] in It is the denoised mean of the posterior distribution, representing the value of a given noisy latent variable. Original latent variables Given the diffusion time step embedding t, the true posterior distribution of the forward diffusion process Mean estimate; This is the denoised variance of the posterior distribution.

[0071] Alternatively, to maintain training stability and simplify computation, the following can be ignored: Learning in China Instead, set it directly to After that, the noise reduction matching term It can be calculated as follows:

[0072]

[0073] Therefore near Decompose in this way ,get:

[0074]

[0075] in Based on Historical Interaction and time Predicted It is instantiated by a parameterized multilayer perceptron, thus yielding the diffusion loss. ,

[0076]

[0077] Optionally, the variational autoencoder is trained according to the MultiVAE method, by optimizing the... and A set of variational autoencoders yields:

[0078]

[0079] in Used for control The strength of the regularization, followed by the diffusion loss. The total loss is obtained by combining the variational autoencoder loss and the sparse user latent vector augmentation loss. hyperparameters , Ensure that the magnitude of each loss is the same.

[0080] Specifically, the model training phase process is as follows:

[0081] First, enter the interaction history of all users on the item set. and random initialization parameters , , ;

[0082] S1. Sample a batch of user interaction history ;

[0083] S2. Use K-means clustering to partition the item set. kind;

[0084] S3. For all Follow the procedure below;

[0085] S3.1 will Division and via encoder Rear series Obtain the latent vector ;

[0086] S3.2 Semantic enhancement and regularization of latent vectors to improve sparse user representations. ;

[0087] S3.3 Truncation Recent Interactions get ;

[0088] S3.4 Sampling or , Reparameterization computation ;

[0089] S3.5 Calculate the total loss Gradient descent optimization , , ;

[0090] S4. Repeat steps S3.1-S3.5 as above until convergence;

[0091] Finally, output the optimized parameters. , , .

[0092] In the reasoning stage, firstly... Divided into Then each Compressed into a deterministic variable Without considering variance, then Connect into Then, the diffusion process is performed by reconstructing... The input is fed into the decoder, and the obtained This is used for item ranking and generating the top K recommendations. Specifically, the inference phase process is as follows:

[0093] First, input the model parameters. , , and user u's interaction history ;

[0094] S5. Use K-means clustering to divide the item set into... kind;

[0095] S6. Will Division and via encoder Rear series Obtain the latent vector ;

[0096] S7. Perform semantic enhancement and regularization on latent vectors to improve sparse user representations. ;

[0097] S8. Given , calculate At the same time, ;

[0098] S9. Iterative calculation of traversal time t = ;

[0099] S10. Final latent vector Classification and decoded Reconstruction ;

[0100] The final output generates Top-K recommended items based on the ranking of the reconstructed items.

[0101] Based on the recommendation generation method based on latent space semantic enhancement and conditional diffusion modeling provided in this embodiment, a specific simulation embodiment is provided, wherein the method of the present invention is referred to as LSE-CDM.

[0102] Optionally, the original data used in this embodiment are input data from three widely used real-world scenarios: Amazon-book comes from the Amazon review dataset, which covers rich user interactions with a large number of books. Movilens-1M is a popular movie rating benchmark dataset. Yelp is a representative commercial dataset containing user reviews of different restaurants. This invention is compared with competing methods, including generative recommendation methods and non-generative recommendation methods, in which: (1) MF is one of the most representative collaborative filtering methods based on matrix factorization; (2) LightGCN learns user and item representations through linear domain aggregation on graph convolutional networks; (3) CDAE trains an autoencoder to recover original user interactions from randomly corrupted interactions; (4) MultiDAE uses dropout to corrupt interactions and recovers them through AEs with multinomial likelihood; (5) MultiVAE utilizes variational autoencoders to simulate the interaction generation process, where the posterior is approximated by the encoder; (6) DODIGEM is a generative model that uses a diffusion process, which uses multiple AEs to simulate inverse generation, but only uses the first AE for interaction prediction. (7) DiffRec infers the user's interaction probability by denoising, directly adding noise to the user's interaction history and iteratively recovering the user's original interaction through a parameterized neural network. (8) DreamRec reshapes sequence recommendation into a learning generation task through a guided diffusion model. In this embodiment, recall (Recall@K) and normalized depreciation cumulative efficiency (NDCG@K) are used as evaluation metrics for the service recommendation model. Recall@K represents the proportion of the number of actual relevant services to the total number of all relevant services in the first K recommendation results. In the experiment, it is set to K={10,20}. NDCG can reflect the ranking quality of the recommendation list.

[0103] Table 1 Comparison of experimental results

[0104]

[0105] The final experimental results are shown in Table 1. In most test cases across the three datasets, the LSE-CDM method of this invention performed best, demonstrating the effectiveness of the proposed model. On the Recall@20 and NDCG@20 metrics, the model showed significant advantages over traditional recommendation models. Furthermore, on different metrics and datasets, the best performance was achieved by multiple different models besides the proposed method, while the performance of the proposed method remained relatively stable. This comparison clearly demonstrates the effectiveness of the proposed method.

[0106] In one exemplary embodiment, a recommendation generation system based on latent spatial semantic enhancement and conditional diffusion modeling is provided, comprising:

[0107] The collection module is used to collect multi-source interaction data of users in the service system, sort it by timestamp and then filter it to construct the interaction vector between users and items and obtain the user interaction sequence.

[0108] The encoder module is used to obtain the embedded representation of each item from the item set using the LightGCN model, classify the items by K-means clustering, and compress the interaction vector between the user and the item into a low-dimensional vector using variational autoencoder.

[0109] The forward process module is used to generate the latent vector representation of the user by connecting the low-dimensional vectors, perform semantic enhancement and regularization on the sparse user latent vectors, and perturb its historical interaction information in the forward process of the conditional diffusion model on the user latent vectors in the low-dimensional latent space.

[0110] The reverse process module is used to encode the user interaction sequence through the Transformer historical interaction sequence encoder, and then dynamically guide the conditional diffusion model to reverse denoising through the cross-attention mechanism, so as to gradually generate the original latent vector representation from the data contaminated by Gaussian noise.

[0111] The decoder module is used to divide the generated original latent vector representations into item categories, and predict the interaction vectors between users and items through the decoder of the corresponding category;

[0112] The training and inference module is used to train and infer the conditional diffusion model, construct its total loss, and generate Top-K recommended items based on the item ranking during the inference phase.

[0113] In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above-described method for recommender generation based on latent spatial semantic enhancement and conditional diffusion modeling.

[0114] In one exemplary embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned recommendation generation method based on latent spatial semantic enhancement and conditional diffusion modeling.

[0115] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described recommendation generation method based on latent spatial semantic enhancement and conditional diffusion modeling.

[0116] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A recommendation generation method based on latent spatial semantic enhancement and conditional diffusion modeling, characterized in that, Includes the following steps: Collect multi-source interaction data of users in the service system, sort and filter it by timestamp to construct the interaction vector between users and items and obtain the user interaction sequence; The LightGCN model is used to obtain the embedded representation of each item from the item set. The items are classified by K-means clustering. At the same time, variational autoencoder is used to compress the interaction vector between the user and the item into a low-dimensional vector. By connecting the low-dimensional vectors, the latent vector representation of the user is generated. The sparse user latent vectors are semantically enhanced and regularized. The historical interaction information of the user latent vectors is perturbed during the forward process of the conditional diffusion model in the low-dimensional latent space. The user interaction sequence is encoded by the Transformer historical interaction sequence encoder, and the conditional diffusion model is dynamically guided by the cross-attention mechanism to denoise in reverse, so that the data contaminated by Gaussian noise is gradually generated into the original latent vector representation. The generated original latent vector representations are divided by item category, and the interaction vectors between users and items are predicted by the decoder of the corresponding category; The conditional diffusion model is trained and inferred to construct its total loss, and Top-K recommended items are generated based on the item ranking during the inference stage.

2. The method according to claim 1, characterized in that, The user interaction sequence This represents the user's most recent n interactions, arranged chronologically. After encoding, these interactions are used to guide the reverse process of the conditional diffusion model to capture the dynamic changes in the user's short-term interests and immediate preferences.

3. The method according to claim 1 or 2, characterized in that, The K-means clustering can identify latent categories in interactive data, enabling the conditional diffusion model to better capture category information during propagation in the latent space after clustering. This ensures that the generation process retains the category structure and user-personalized information, while supporting parallel computation of training and generation in the low-dimensional latent space.

4. The method according to claim 1, characterized in that, When performing the semantic enhancement and regularization on the sparse user latent vector representation during the forward pass of the conditional diffusion model in the low-dimensional latent space, the following steps are performed: First, adaptive augmentation weights are calculated based on the sparsity of user interactions. If interactions are sparse, latent vector information is supplemented using similar users. Secondly, KL regularization constraints are added to ensure that the distribution of the enhanced latent variables is consistent with the original latent variable distribution; Finally, parameters are adjusted during the forward diffusion process, and the noise scale is controlled by reparameterization techniques and Gaussian noise superposition.

5. The method according to claim 1 or 4, characterized in that, When denoising the latent vector representation using the Transformer history interaction sequence encoder and the cross-attention mechanism, the following steps are performed: First, embedding and mapping are performed on the user's historical interaction sequence to obtain the item embedding sequence; Secondly, the item embedding sequence is encoded using a Transformer encoder, and the interactive sequence encoding is used as a condition to guide the denoising process. Finally, by using a cross-attention mechanism, the time step embedding and interactive sequence encoding are applied to the attention layer of the denoising network to achieve conditional denoising.

6. The method according to claim 5, characterized in that, The conditional diffusion module of the cross-attention mechanism includes a feedforward sub-layer and a layer normalization structure, and residual connections are used in the denoising network to improve training stability.

7. A recommendation generation system based on latent spatial semantic enhancement and conditional diffusion modeling, characterized in that, include: The collection module is used to collect multi-source interaction data of users in the service system, sort it by timestamp and then filter it to construct the interaction vector between users and items and obtain the user interaction sequence. The encoder module is used to obtain the embedded representation of each item from the item set using the LightGCN model, classify the items by K-means clustering, and compress the interaction vector between the user and the item into a low-dimensional vector using variational autoencoder. The forward process module is used to generate the latent vector representation of the user by connecting the low-dimensional vectors, perform semantic enhancement and regularization on the sparse user latent vectors, and perturb its historical interaction information in the forward process of the conditional diffusion model on the user latent vectors in the low-dimensional latent space. The reverse process module is used to encode the user interaction sequence through the Transformer historical interaction sequence encoder, and then dynamically guide the conditional diffusion model to reverse denoising through the cross-attention mechanism, so as to gradually generate the original latent vector representation from the data contaminated by Gaussian noise. The decoder module is used to divide the generated original latent vector representations into item categories, and predict the interaction vectors between users and items through the decoder of the corresponding category; The training and inference module is used to train and infer the conditional diffusion model, construct its total loss, and generate Top-K recommended items based on the item ranking during the inference phase.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images