A social hint cold start recommendation method based on an SPGCLRec framework

Abstract

The application discloses a social prompt cold start recommendation method based on an SPGCLRec framework, and comprises the following steps: step one, constructing an interaction graph and pre-training a recommendation model; step two, optimizing a social prompt based on a diffusion model; step three, enhancing a cold start user representation; and step four, generating a cold start recommendation and jointly optimizing a model. The learnable social soft prompt designed by the application embeds social information into a representation space of a pre-training model, and can realize efficient parameter adaptation without fine-tuning core weights of the model, thereby making up for the deficiency of the prior art which only relies on a single auxiliary clue and realizing efficient parameter adaptation of the pre-training model. The diffusion model is used to filter noise social relationships in a latent vector space, thereby improving user representation accuracy. By introducing a contrast learning mechanism and introducing Gaussian noise in a latent space, the semantic consistency of the optimized social prompt and the item representation is strengthened, and the robustness of the user representation is improved.

Description

Technical Field

[0001] This invention relates to the field of chip testing technology, specifically to a social prompt cold start recommendation method based on the SPGCLRec framework. Background Technology

[0002] Cold start recommendation is a long-standing core technical bottleneck in practical recommendation systems, especially on platforms with rapidly growing users. New users lack historical interaction data such as clicks, purchases, and ratings after registration, making it difficult to build accurate user profiles. While pre-trained recommendation models trained on massive amounts of active user data encode general user-item interaction patterns, they suffer from semantic misalignment between cold start user representation and model representation space.

[0003] Existing solutions mainly fall into four categories: data augmentation, graph-based social recommendation, contrastive learning, and generative models. Each has the following drawbacks: 1. Data augmentation methods supplement user data with auxiliary information, but lack compatibility with the pre-trained model space and cannot filter noise; 2. Graph-based social recommendation utilizes social homogeneity to enhance user representations, but it only optimizes on the original social graph and does not filter noise that conflicts with pre-trained representations; 3. While contrastive learning can enhance representation robustness, most methods do not integrate social information and cannot align with the pre-trained model; 4. In generative models, GANs suffer from training instability and pattern collapse, while diffusion models only operate on graph or sequence spaces and do not incorporate cue learning to adapt to the pre-trained model. Furthermore, existing cue learning methods for adapting to pre-trained models rely solely on single auxiliary cues such as age and gender, failing to utilize social preference homogeneity and lacking effective social noise filtering mechanisms. Consequently, existing technologies cannot simultaneously address the three core issues of semantic misalignment, social noise filtering, and pre-trained model adaptation, making it difficult to effectively improve the accuracy and stability of cold-start recommendations. Summary of the Invention

[0004] The purpose of this invention is to provide a social prompt cold start recommendation method based on the SPGCLRec framework to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a social cue cold start recommendation method based on the SPGCLRec framework, comprising the following steps: Step 1, constructing an interaction graph and pre-training a recommendation model; Step 2, optimizing social cueing based on a diffusion model; Step 3, enhancing cold start user representation; Step 4, jointly optimizing cold start recommendation generation and the model.

[0006] In step one above, a user-item collaboration graph and a user-user social graph are constructed. Based on the historical interaction data of active users, the recommendation model is pre-trained to obtain the pre-trained user representation, item representation, and trainable parameters of the model.

[0007] In step two above, initial user social vectors are extracted from the user-user social graph constructed in step one. Through the forward noise addition process and the reverse noise removal process of the diffusion model, noisy social relationships are filtered in the latent vector space, and the denoised social vectors are reconstructed. The denoised social vectors are then converted into optimized social prompts that are adapted to the pre-trained recommendation model.

[0008] In step three above, the optimized social cues obtained in step two are superimposed on the original representation of the cold-start user to obtain a socially cued enhanced user representation. By contrastive learning, the semantic consistency between the socially cued enhanced user representation and the pre-trained item representation is strengthened to obtain a refined final cold-start user representation.

[0009] In step four above, based on the final cold-start user representation and pre-trained item representation obtained in step three, user-item recommendation scores are calculated, a joint loss function is constructed, model parameters are iteratively optimized, and Top-K recommendation results for cold-start users are generated by sorting them in descending order of recommendation scores.

[0010] In step one, the pre-trained recommendation model uses LightGCN, and BPR Loss is used to optimize the model parameters; the iterative message passing process of LightGCN is implemented through the following formula:

[0011]

[0012] in, For the adjacency matrix of the user-item collaboration graph, Adjacency matrix The corresponding degree diagonal matrix, The graph convolution kernel weight matrix, For the first Embedding matrix of subgraph convolution iteration For second normal form embedding of normalized functions, The number of convolutional layers in the graph. The total embedding matrix is ​​obtained after multiple rounds of iteration and summation. After splitting, the pre-trained user representation and item representation are obtained.

[0013] In step two, the initial user social vector is obtained by encoding the user-user social graph using a graph convolutional encoder. Let the initial time of the diffusion process be... The social vector is ,in This is the initial user social vector.

[0014] In step two, the forward noise addition process of the diffusion model is as follows: from to Gaussian noise is injected into the social vector sequentially. The total number of diffusion steps, The social vector at time step is calculated using the following formula:

[0015]

[0016] in, It follows a Gaussian distribution. for Noise intensity at any given time It is the identity matrix;

[0017] The reverse denoising process of the diffusion model is as follows: from to The conditional probability is estimated using a neural network consisting of two fully connected layers to recover valid social signals from noisy social vectors. The conditional probability formula is:

[0018]

[0019] in, The mean of the Gaussian distribution predicted by the neural network. Let be the standard deviation of a Gaussian distribution, and let be the input to the neural network. Momentary Social Vectors and time step vector The formula for mean prediction is: , This involves two consecutive fully connected layer operations; followed by reverse denoising iterations to... The reconstructed denoised social vectors are obtained. .

[0020] In step two, the conversion formula for optimizing social prompts is:

[0021]

[0022] in, To optimize social notifications, Let be the transformation function from social vectors to social cues. The vector is the denoised social vector; the optimization loss of the diffusion model is the reconstruction loss, and the formula is:

[0023]

[0024] in, For distribution-based Expectations It is a norm 2. This is the initial user social vector.

[0025] In step three, the formula for calculating the user representation enhanced by social cues is as follows:

[0026]

[0027] in, Enhanced user representation for social cues The original characteristics of a cold-start user. To optimize social notifications.

[0028] In step three, the contrastive learning process includes: injecting Gaussian noise into the socially cued enhanced user representation and the pre-trained item representation respectively to obtain the perturbed user representation and the perturbed item representation; using InfoNCE loss as the optimization objective to enhance the semantic consistency between the socially cued enhanced user representation and its perturbed version.

[0029] The sampling formula for the Gaussian noise is:

[0030]

[0031] in, For noise vectors, It follows a Gaussian distribution. for 3D identity matrix As a representation dimension;

[0032] The formulas for calculating the perturbed user representation and item representation are as follows:

[0033]

[0034] in, This represents the user profile after the disturbance. To represent the object after the disturbance. , Let be the Gaussian noise vectors for the user and the item, respectively. Enhanced user representation for social cues This is for the representation of pre-trained items.

[0035] The formula for calculating the InfoNCE loss is as follows:

[0036]

[0037] in, For the cold start user set, This is the cosine similarity calculation function. The temperature coefficient is used for comparative learning.

[0038] In step four, the formula for calculating the user-item recommendation score is:

[0039]

[0040] in, For recommended score, This is the final representation of a cold-start user. For pre-trained item representation, The transpose of the object's representation;

[0041] The formula for calculating the joint loss function is:

[0042]

[0043] in, It is the Sigmoid activation function. These are positive sample items that users have interacted with during the cold start. For cold start, negative sample items that users have not interacted with. Optimize the weighting coefficients of tasks for social cues. To compare the weight coefficients of the learning tasks, Optimize losses for social prompts. To compare learning loss.

[0044] Compared with existing technologies, the beneficial effects of this invention are as follows: The learnable social soft cue designed in this invention embeds social information into the representation space of the pre-trained model, achieving efficient parameter adaptation without fine-tuning the core weights of the model. This overcomes the shortcomings of existing technologies that rely solely on a single auxiliary cue, thus achieving efficient parameter adaptation of the pre-trained model. Furthermore, by filtering noisy social relationships in the latent vector space through a diffusion model, the accuracy of user representation is improved. Finally, by introducing a contrastive learning mechanism and introducing Gaussian noise into the latent space, the semantic consistency between the optimized social cue and the item representation is strengthened, enhancing the robustness of the user representation. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method of the present invention;

[0046] Figure 2 This is a schematic diagram illustrating the principle of enhancing cold-start user representation based on social cues.

[0047] Figure 3 This is a schematic diagram of the SPGCLRec architecture;

[0048] Figure 4 The results of the parameter sensitivity analysis for the social prompt optimization module are shown in the figure; (a) Ciao-HR; (b) Ciao-NDCG; (c) Douban-HR; (d) Douban-NDCG; (e) Yelp-HR; (f) Yelp-NDCG;

[0049] Figure 5The following are the results of the parameter sensitivity analysis for the comparative learning module: (a) Ciao-HR; (b) Ciao-NDCG; (c) Douban-HR; (d) Douban-NDCG; (e) Yelp-HR; (f) Yelp-NDCG;

[0050] Figure 6 This is a visualization analysis diagram of the user-item representation T-SNE based on the present invention. Detailed Implementation

[0051] 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.

[0052] Please see the appendix Figure 1 -Appendix Figure 6 The present invention provides an embodiment of a social cue cold start recommendation method based on the SPGCLRec framework, comprising the following steps: Step 1, constructing an interaction graph and pre-training a recommendation model; Step 2, optimizing social cuees based on a diffusion model; Step 3, enhancing cold start user representation; Step 4, jointly optimizing cold start recommendation generation and the model.

[0053] In step one above, a user-item collaboration graph and a user-user social graph are constructed. The recommendation model is pre-trained based on historical interaction data of active users, yielding pre-trained user representations, item representations, and trainable model parameters. The pre-trained recommendation model uses LightGCN, and BPR Loss is used to optimize the model parameters. The iterative message passing process of LightGCN is implemented using the following formula:

[0054]

[0055] in, For the adjacency matrix of the user-item collaboration graph, Adjacency matrix The corresponding degree diagonal matrix, The graph convolution kernel weight matrix, For the first Embedding matrix of subgraph convolution iteration For second normal form embedding of normalized functions, The number of convolutional layers in the graph. The total embedding matrix after multiple rounds of iteration and summation is split to obtain pre-trained user representations and item representations;

[0056] In step two above, initial user social vectors are extracted from the user-user social graph constructed in step one. Through the forward denoising and backward denoising processes of the diffusion model, noisy social relationships are filtered in the latent vector space, reconstructing denoised social vectors. These denoised social vectors are then converted into optimized social cues adapted to the pre-trained recommendation model. The initial user social vectors are obtained by encoding the user-user social graph using a graph convolutional encoder. Let the initial time of the diffusion process be... The social vector is ,in The initial user social vector; the forward noise addition process of the diffusion model is as follows: from to Gaussian noise is injected into the social vector sequentially. The total number of diffusion steps, The social vector at time step is calculated using the following formula:

[0057]

[0058] in, It follows a Gaussian distribution. for Noise intensity at any given time It is the identity matrix;

[0059] The reverse denoising process of the diffusion model is as follows: from to The conditional probability is estimated using a neural network consisting of two fully connected layers to recover valid social signals from noisy social vectors. The conditional probability formula is:

[0060]

[0061] in, The mean of the Gaussian distribution predicted by the neural network. Let be the standard deviation of a Gaussian distribution, and let be the input to the neural network. Momentary Social Vectors and time step vector The formula for mean prediction is: , This involves two consecutive fully connected layer operations; followed by reverse denoising iterations to... The reconstructed denoised social vectors are obtained. The optimized conversion formula for social prompts is:

[0062]

[0063] in, To optimize social notifications, Let be the transformation function from social vectors to social cues. The vector is the denoised social vector; the optimization loss of the diffusion model is the reconstruction loss, and the formula is:

[0064]

[0065] in, For distribution-based Expectations It is a norm 2. This serves as the initial user social vector;

[0066] In step three above, the optimized social cues obtained in step two are superimposed on the original representation of the cold-start user to obtain a socially cued enhanced user representation. Through contrastive learning, the semantic consistency between the socially cued enhanced user representation and the pre-trained item representation is strengthened, resulting in a refined final cold-start user representation. The calculation formula for the socially cued enhanced user representation is as follows:

[0067]

[0068] in, Enhanced user representation for social cues The original characteristics of a cold-start user. To optimize social notifications.

[0069] In step three, the contrastive learning process includes: injecting Gaussian noise into the socially-enhanced user representation and the pre-trained item representation, respectively, to obtain perturbed user representations and perturbed item representations; using InfoNCE loss as the optimization objective, the semantic consistency between the socially-enhanced user representation and its perturbed version is strengthened; the sampling formula for Gaussian noise is:

[0070]

[0071] in, For noise vectors, It follows a Gaussian distribution. for 3D identity matrix As a representation dimension;

[0072] The formulas for calculating the perturbed user representation and item representation are as follows:

[0073]

[0074] in, This represents the user profile after the disturbance. To represent the object after the disturbance. , Let be the Gaussian noise vectors for the user and the item, respectively. Enhanced user representation for social cues For pre-trained item representation; the formula for calculating InfoNCE loss is:

[0075]

[0076] in, For the cold start user set, This is the cosine similarity calculation function. For comparative learning of temperature coefficients;

[0077] In step four above, based on the final cold-start user representation and pre-trained item representation obtained in step three, user-item recommendation scores are calculated, a joint loss function is constructed, model parameters are iteratively optimized, and Top-K recommendation results for cold-start users are generated by sorting the recommendation scores in descending order. The formula for calculating the user-item recommendation score is as follows:

[0078]

[0079] in, For recommended score, This is the final representation of a cold-start user. For pre-trained item representation, The transpose of the object's representation;

[0080] The formula for calculating the joint loss function is:

[0081]

[0082] in, It is the Sigmoid activation function. These are positive sample items that users have interacted with during the cold start. For cold start, negative sample items that users have not interacted with. Optimize the weighting coefficients of tasks for social cues. To compare the weight coefficients of the learning tasks, Optimize losses for social prompts. To compare learning loss.

[0083] Experimental Example 1:

[0084] To verify the effectiveness of this invention, the following experiments were conducted: Four mainstream cold-start recommendation methods were selected as comparative examples, including traditional cold-start methods: MF, DropoutNet, Heater; GNN-based cold-start methods: LightGCN, XSimGCL, FS-GNN; socially-aware cold-start method: MHCN; and diffusion-based recommendation method: SGSR. Experiments were conducted using three real-world public datasets. The statistical information of these datasets is shown in Table 1. Ciao is a consumer review platform dataset, containing user item ratings and explicit trust relationships between users; Doubt... `an` is a dataset from a movie social platform, containing user movie ratings and user social following relationships; `Yelp` is a dataset from a merchant review platform, containing user merchant ratings and user friend relationships. In the experiment, users with fewer than 10 historical interactions were defined as cold-start users, and the rest were active users. Active user data was used for pre-training, and cold-start users were divided into validation and test sets in a 1:2 ratio. The pre-trained model used LightGCN, with user and item representation dimensions uniformly set to 64 dimensions, and initialized using Xavier. In the diffusion model, the optimizer was Adam, and the learning rate was set to 0.001. The total number of diffusion steps was set to 20. In the comparative learning module, the optimizer was Adam, the learning rate was set to 0.0001, the L2 regularization coefficient was set to 0.0001, and the temperature coefficient was set to 0.2. The evaluation metrics used were the hit rate HR@k and the normalized cumulative gain NDCG@k, which are common in recommendation systems, with k set to 10 and 20, respectively. The experimental results are shown in Tables 2 and 3. The core metrics of this method (SPGCLRec) are better than all the baselines on the three datasets. The specific improvement is as follows: Ciao dataset: HR@10 reaches 0.045, which is 4.6% better than the second-best model. The NDCG@10 score reached 0.052, a 4.0% improvement over the second-best model; on the Douban dataset, HR@10 reached 0.064, an 8.4% improvement over the second-best model; NDCG@10 reached 0.099, an 8.7% improvement over the second-best model; on the Yelp dataset, HR@10 reached 0.040, a 5.2% improvement over the second-best model; NDCG@10 reached 0.051, a 6.2% improvement over the second-best model. This method achieves a significant improvement of 4.0%-8.7% over the current best method for HR@10 and NDCG@10 metrics for cold-start users, validating the effectiveness of the proposed method.

[0085] Experimental Example 2:

[0086] To verify the role of each core module of SPGCLRec, ablation experiments were conducted. Variant models were constructed by removing the corresponding modules. The test results under the condition of k=10 are shown in Table 3. The experimental results show that: after removing the social cue optimization module, the model performance dropped significantly, verifying that this module is the core of improving the cold start recommendation effect; after removing the contrastive learning module, the performance dropped slightly, verifying that this module enhances semantic alignment and representation robustness; the two modules work together to achieve the optimal cold start recommendation effect.

[0087] Experimental Example 3:

[0088] To verify the representation optimization and semantic alignment effects of this invention, the T-SNE algorithm was used to perform dimensionality reduction and visualization analysis on the user and item representations in the Yelp dataset processed by the method of this invention. The results are as follows: Figure 6 As shown, the cold-start user representation optimized by this method can be effectively clustered around active users and user-preferred items. User-preferred items are concentrated in the user cluster center, while irrelevant items are distributed around the cluster periphery. This intuitively verifies that this method can effectively filter social noise, enrich the cold-start user preference representation, and achieve accurate semantic alignment between user representation and item representation.

[0089] Table 1 Dataset

[0090] Table 2. Recommended performance comparison for cold starts (Part 1)

[0091] Table 3. Recommended performance comparison for cold starts (Part 2)

[0092] Table 4 Ablation Experiment Results

[0093] Based on the above, the advantages of this invention are as follows: When used, by integrating social cues, diffusion model denoising, and contrastive learning techniques, it fundamentally solves the problems of semantic misalignment between social information and pre-trained models, and distortion of preference signals by noisy social relationships in cold-start recommendation. This achieves accurate and efficient adaptation of the pre-trained recommendation model to cold-start users. The invention designs learnable social soft cues, embedding social information into the representation space of the pre-trained model. This allows for efficient parameter adaptation without fine-tuning the model's core weights, effectively compensating for the shortcomings of existing cue learning that relies solely on a single auxiliary cue. It fully utilizes the homogeneity of social preferences to enrich the cold-start user profile, expanding the ability of the pre-trained recommendation model to handle cold-start tasks. Optimizing social cues through diffusion models in the latent vector space rather than the original social graph accurately filters out weak associations and false or noisy social relationships incompatible with the pre-trained model, retaining effective social signals and significantly improving the accuracy of cold-start user representation. Introducing a contrastive learning mechanism and Gaussian noise into the latent space strengthens the semantic consistency between the optimized social cues and item representations, improving the robustness of user representations and the accuracy of preference matching between cold-start users and related items.

[0094] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A cold-start recommendation method based on the SPGCLRec framework using social prompts, comprising the following steps: Step 1: Construct an interaction graph and pre-train a recommendation model; Step 2: Optimize social prompts based on a diffusion model; Step 3: Enhance cold-start user representation; Step 4: Jointly optimize cold-start recommendation generation and the model; Its key features are: In step one above, a user-item collaboration graph and a user-user social graph are constructed. Based on the historical interaction data of active users, the recommendation model is pre-trained to obtain the pre-trained user representation, item representation, and trainable parameters of the model. In step two above, initial user social vectors are extracted from the user-user social graph constructed in step one. Through the forward noise addition process and the reverse noise removal process of the diffusion model, noisy social relationships are filtered in the latent vector space, and the denoised social vectors are reconstructed. The denoised social vectors are then converted into optimized social prompts that are adapted to the pre-trained recommendation model. In step three above, the optimized social cues obtained in step two are superimposed on the original representation of the cold-start user to obtain a socially cued enhanced user representation. By contrastive learning, the semantic consistency between the socially cued enhanced user representation and the pre-trained item representation is strengthened to obtain a refined final cold-start user representation. In step four above, based on the final cold-start user representation and pre-trained item representation obtained in step three, user-item recommendation scores are calculated, a joint loss function is constructed, model parameters are iteratively optimized, and Top-K recommendation results for cold-start users are generated by sorting them in descending order of recommendation scores.

2. The social cue cold start recommendation method based on the SPGCLRec framework according to claim 1, characterized in that: In step one, the pre-trained recommendation model uses LightGCN, and BPR Loss is used to optimize the model parameters; the iterative message passing process of LightGCN is implemented through the following formula: in, For the adjacency matrix of the user-item collaboration graph, Adjacency matrix The corresponding degree diagonal matrix, The graph convolution kernel weight matrix, For the first Embedding matrix of subgraph convolution iteration For second normal form embedding of normalized functions, The number of convolutional layers in the graph. The total embedding matrix is ​​obtained after multiple rounds of iteration and summation. After splitting, the pre-trained user representation and item representation are obtained.

3. The social cue cold start recommendation method based on the SPGCLRec framework according to claim 1, characterized in that: In step two, the initial user social vector is obtained by encoding the user-user social graph using a graph convolutional encoder. Let the initial time of the diffusion process be... The social vector is ,in This is the initial user social vector.

4. The social cue cold start recommendation method based on the SPGCLRec framework according to claim 1, characterized in that: In step two, the forward noise addition process of the diffusion model is as follows: from to Gaussian noise is injected into the social vector sequentially. The total number of diffusion steps, The social vector at time step is calculated using the following formula: in, It follows a Gaussian distribution. for Noise intensity at any given time It is the identity matrix; The reverse denoising process of the diffusion model is as follows: from to The conditional probability is estimated using a neural network consisting of two fully connected layers to recover valid social signals from noisy social vectors. The conditional probability formula is: in, The mean of the Gaussian distribution predicted by the neural network. Let be the standard deviation of a Gaussian distribution, and let be the input to the neural network. Momentary Social Vectors and time step vector The formula for mean prediction is: , This involves two consecutive fully connected layer operations; followed by reverse denoising iterations to... The reconstructed denoised social vectors are obtained. .

5. The social cue cold start recommendation method based on the SPGCLRec framework according to claim 1, characterized in that: In step two, the conversion formula for optimizing social prompts is: in, To optimize social notifications, Let be the transformation function from social vectors to social cues. The vector is the denoised social vector; the optimization loss of the diffusion model is the reconstruction loss, and the formula is: in, For distribution-based Expectations It is a 2-norm. This is the initial user social vector.

6. The social cue cold start recommendation method based on the SPGCLRec framework according to claim 1, characterized in that: In step three, the formula for calculating the user representation enhanced by social cues is as follows: in, Enhanced user representation for social cues The original characteristics of a cold-start user. To optimize social notifications.

7. The social cue cold start recommendation method based on the SPGCLRec framework according to claim 1, characterized in that: In step three, the contrastive learning process includes: injecting Gaussian noise into the socially cued enhanced user representation and the pre-trained item representation respectively to obtain the perturbed user representation and the perturbed item representation; using InfoNCE loss as the optimization objective to enhance the semantic consistency between the socially cued enhanced user representation and its perturbed version.

8. The social cue cold start recommendation method based on the SPGCLRec framework according to claim 7, characterized in that: The sampling formula for the Gaussian noise is: in, For noise vectors, It follows a Gaussian distribution. for 3D identity matrix As a representation dimension; The formulas for calculating the perturbed user representation and item representation are as follows: in, This represents the user profile after the disturbance. To represent the object after the disturbance. , Let be the Gaussian noise vectors for the user and the item, respectively. Enhanced user representation for social cues This is for the representation of pre-trained items.

9. A social cue cold start recommendation method based on the SPGCLRec framework according to claim 7, characterized in that: The formula for calculating the InfoNCE loss is as follows: in, For the cold start user set, This is the cosine similarity calculation function. The temperature coefficient is used for comparative learning.

10. A social cue cold start recommendation method based on the SPGCLRec framework according to claim 1, characterized in that: In step four, the formula for calculating the user-item recommendation score is: in, For recommended score, This is the final representation of a user during a cold start. For pre-trained item representation, The transpose of the object's representation; The formula for calculating the joint loss function is: in, It is the Sigmoid activation function. These are positive sample items that users have interacted with during the cold start. For cold start, negative sample items that users have not interacted with. Optimize the weighting coefficients of tasks for social cues. To compare the weight coefficients of the learning tasks, Optimize losses for social prompts To compare learning loss.

Images