User interest prediction method, device and system and storage medium

Abstract

The invention discloses a user interest prediction method, device and system, and a storage medium. The method comprises the following steps: S1, acquiring historical interaction data of a user; s2, according to historical interaction data of the user, mapping the representation in the Euclidean space to different geometric spaces; s3, performing diffusion depolarization on the object embedding in the different spaces to obtain depolarized user representations of the different spaces; and S4, mapping the user representations in different spaces back to the Euclidean space for reconstruction and matching to obtain the actual interests of the users. By adopting the technical scheme provided by the invention, the interest and preference of the user can be predicted more accurately.

Description

Technical Field

[0001] The present invention belongs to the technical field of information recommendation, and in particular relates to a method and device, a system, and a storage medium for predicting user interests. Background Art

[0002] With the advancement of Internet technology, users have access to a large amount of information, but this also brings new challenges in how to efficiently filter out content that users are interested in. Traditional recommendation systems usually rely on collaborative filtering technology in Euclidean space. Although this technology is effective in some aspects, it fails to fully consider the complexity and multidimensionality of the interaction between users and items. It tends to simplify this interaction into a linear relationship, while in reality, users' interests are diverse and complex, and their interactions with items occur in a more complex multidimensional space. The geometric structure of this space is much more complex than traditional models can capture, resulting in traditional methods not being effective in dealing with highly nonlinear interactions between users and items. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a user interest prediction method and device, system, and storage medium, which go beyond the limitations of the traditional Euclidean space model and can more accurately predict the user's interests and preferences; can better understand and simulate the user's complex selection behavior in the ocean of information, and provide more personalized and accurate recommendations.

[0004] To achieve the above object, the present invention adopts the following technical solution:

[0005] A user interest prediction method, comprising:

[0006] Step S1, obtaining the user's historical interaction data;

[0007] Step S2: Mapping the representation in the Euclidean space to different geometric spaces according to the user's historical interaction data;

[0008] Step S3, performing diffusion debiasing on the embeddings of items in different spaces to obtain user representations after debiasing in different spaces;

[0009] Step S4: Map the user representations in different spaces back to the Euclidean space for reconstruction and matching to obtain the user's actual interests.

[0010] Preferably, the geometric space adopts hyperbolic space, Euclidean space and spherical space to understand the intention behind the user behavior.

[0011] Preferably, in step S3, the embeddings of items in different spaces are subjected to denoising and homogenization, and then iterative denoising is performed to obtain user representations after debiasing in different spaces.

[0012] Preferably, in step S4, reconstruction and matching are performed in a Euclidean space where similarity is better calculated, and a top-20 recommendation method is used to recommend items ranked in top 20 similarity to the user.

[0013] The present invention also provides a user interest prediction device, comprising:

[0014] The acquisition module is used to obtain the user's historical interaction data;

[0015] A first processing module is used to map the representation in the Euclidean space to different geometric spaces according to the historical interaction data of the user;

[0016] The second processing module is used to diffuse and debias the embeddings of items in different spaces to obtain user representations after debiasing in different spaces;

[0017] The third processing module is used to map the user representations in different spaces back to the Euclidean space for reconstruction and matching to obtain the user's actual interests.

[0018] Preferably, the geometric space adopts hyperbolic space, Euclidean space and spherical space to understand the intention behind the user behavior.

[0019] The present invention also provides a user interest prediction system, characterized in that it includes: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a user interest prediction method when executed by the processor.

[0020] The present invention also provides a storage medium, on which a computer program is stored, and the computer program executes the user interest prediction method when running.

[0021] The present invention improves the accuracy of personalized recommendations of the recommendation system through user interest decoupling, diffusion debiasing and user interest reconstruction. Geometric decoupling is used to break through the traditional Euclidean space limitations and can adapt to and express various geometric structures of data. The forward denoising and reverse denoising strategies of diffusion debiasing are used to process the long-tail deviation, remove the deviation and effectively optimize the user representation. In summary, the present invention effectively improves the effectiveness of the recommendation system in complex data scenarios and enhances the flexibility and accuracy of the system through geometric decoupling and debiasing processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0023] Figure 1 A flowchart of a method for predicting user interests according to an embodiment of the present invention;

[0024] Figure 2 A schematic diagram of a method for predicting user interests according to an embodiment of the present invention;

[0025] Figure 3 Schematic diagram for decoupling user interests;

[0026] Figure 4 Schematic diagram for diffusion;

[0027] Figure 5 Reconstruct and recommend schematic diagrams for interest;

[0028] Figure 6 Schematic diagram of the long tail experiment. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] Embodiment 1:

[0032] like Figure 1 As shown, an embodiment of the present invention provides a method for predicting user interests, including:

[0033] Step S1, obtaining the user's historical interaction data;

[0034] Step S2: Mapping the representation in the Euclidean space to different geometric spaces through user interest decoupling based on the user's historical interaction data;

[0035] Step S3, performing diffusion debiasing on the embeddings of items in different spaces to obtain user representations after debiasing in different spaces;

[0036] Step S4: Map the user representations in different spaces back to the Euclidean space for reconstruction and matching to obtain the user's actual interests.

[0037] As an implementation method of an embodiment of the present invention, in step S2, the interaction between users and items can be a unidirectional tree structure or a bidirectional ring structure. In order to more accurately calculate the similarity and recommendation relationship between them, a geometric decoupling method is used. For the tree structure of fast-growing nodes, a hyperbolic space is used; for the closed ring structure, a spherical space is used. Different manifold spaces are defined, and specific mapping methods (such as e-exponential mapping and logarithmic mapping) are used to process these spaces. The concept of parallel transfer is also introduced to connect the vectors of the manifold space and the tangent space. By applying these concepts to the embedding of users and items, calculations can be performed on different manifold spaces, thereby achieving more accurate geometric space decoupling.

[0038] like Figure 3 As shown in Figure 2, in user interest decoupling, three manifold spaces are defined, H represents hyperbolic space, E represents Euclidean space, and S represents spherical space. At the same time, prototypes corresponding to the three spaces are defined, where prototypes can be regarded as representative examples of concepts, which capture the key features of the concepts.

[0039] The manifold space is defined as follows:

[0040]

[0041] Where <·,·> represents the curvature-aware scalar product, k is the curvature magnitude, and when k>0 When k<0

[0042] Then the item vectors in the original space are mapped to the three geometric spaces respectively, the method of mapping from Euclidean space to manifold space is simplified to an exponential mapping formula, and the reverse mapping from manifold space back to Euclidean space is simplified to a logarithmic mapping formula.

[0043] The manifold space mapping is as follows:

[0044]

[0045] in, is the tangent vector norm, and the curvature-aware trigonometric functions are as follows:

[0046]

[0047] After mapping the item vector to the manifold space, the similarity between the item and the prototype is calculated in different spaces and grouped. In the Euclidean space, this formula is used to calculate the length between two points as the similarity, while in the non-Euclidean space, the geodesic distance calculated by this formula on different manifolds is used as the standard for measuring similarity. Finally, the probability of different items being assigned to different geometric spaces is obtained, thereby achieving the purpose of decoupling user interests.

[0048] The European space allocation of items is as follows:

[0049] ||log x (y)||=||yx||=d E (x,y)

[0050] Among them, ||log x (y)|| is the Euclidean distance between the object and the prototype.

[0051] The non-Euclidean space allocation of items is as follows:

[0052]

[0053] As an implementation method of an embodiment of the present invention, in step S3, diffusion debiasing uses the diffusion mechanism in Dreamrec, the purpose of which is to convert the user embedded data into an unbiased Gaussian distribution through the process of adding noise and denoising, thereby improving the accuracy of the recommendation system and the quality of personalized services. It can effectively remove the bias in the embedded data, generate more representative and reliable user vectors, and thus improve the performance of the recommendation algorithm.

[0054] like Figure 4 As shown in the figure, in the conditional diffusion model, three inputs are received: input vector x, guide vector h, and step vector t. The input vector x is calculated through a graph convolutional network, and the item embeddings in different spaces are convolved to obtain an aggregated vector. This vector then undergoes a noise addition and denoising process in the diffusion model. The guide vector h is randomly initialized to ensure that it follows a normal distribution, which helps the diffusion model learn and generalize user preferences. The embedding vector t represents the number of layers in the debiasing process and is set by randomly selecting values ​​within a specific range. These three inputs together provide information for the recommendation task. By optimizing these inputs, it is possible to better meet user personalization needs and improve the performance of the recommendation system.

[0055] In diffusion debiasing, there are two parts, namely forward denoising and reverse denoising. First, the embedding of items grouped in each geometric space is aggregated through graph convolution to obtain the input vector. The guide vector is a vector randomly initialized under normal distribution. First, the forward denoising process is performed. The formula in the figure below is iterative denoising. In order to simplify the operation, the denoising intensity is obtained by the following equation: The input vector can be added to the homogenized vector in one step. The second is the reverse denoising process, where the homogenized vector is input into the denoising part, and the iterative denoising process is performed, and finally the debiased user embedding x' is output. In order to prevent the loss of some user interest information during the iteration process, a residual connection is added to the output part, and the input vector is connected to the output part to retain the user interest information.

[0056] User interest code:

[0057]

[0058] in, is an unnormalized matrix; for The degree matrix of W (l) represents the weight matrix of the lth layer; H (l) represents the node feature vector of the lth layer, H (0) is a randomly initialized feature matrix; finally, after multi-layer graph convolution operations, a node representation that aggregates multi-order neighborhood information is obtained.

[0059] The forward noise is:

[0060] Among them, β t Indicates the noise intensity at each step in the iteration process, I is the unit matrix, x t With x t-1 They are the output x vector after asynchronous noise addition, that is, the embedding of the input vector x after noise addition with different intensities,

[0061] The reverse denoising is:

[0062] in, and are the mean and variance of the Gaussian distribution fitted by the MLP.

[0063] The output is: x output =x′+x

[0064] Among them, x′ is the user embedding generated by iterative debiasing. In order to make the user embedding representation of the recommendation task better, the input vector x is added to the embedding to obtain the final representation x for the recommendation task. output .

[0065] As an implementation of the embodiment of the present invention, in step S4, Figure 5 As shown in the figure, after iterative denoising, interest reconstruction and recommendation optimization are performed. Because the generated user embedding is in different geometric spaces, and the item embedding to be matched with it is in Euclidean space, for simplicity, the user representation in different geometric spaces is reversely mapped back to Euclidean space, and reconstruction and matching are performed in Euclidean space where similarity is better calculated. The method used to calculate the similarity is to perform the inner product of the user vector mapped back to Euclidean space and the item vector, and use the top-20 recommendation method to recommend the top 20 items with similarity ranking to the user.

[0066] Matches:

[0067] in, For user characterization, h i Representation of items.

[0068] In terms of optimization and recommendation, cross entropy loss is used for optimization, and the loss function is:

[0069]

[0070] Among them, p i is the probability of user-item interaction, t i is the input user-item interaction matrix, which reduces the gap between the predicted user-item interaction and the input interaction, thereby achieving a more accurate prediction.

[0071] The present invention is based on geometric diffusion debiasing decoupling, and analyzes the historical interaction data of users to more accurately understand their interests and preferences. First, geometric decoupling is used to map the representation in Euclidean space to different geometric spaces, and hyperbolic space, Euclidean space and spherical space are used to understand the intention behind user behavior and reveal the diversity and level of user interests. Secondly, the embedding of items in different spaces is aggregated to obtain the embedding input into the generative model, and a diffusion debiasing operation is performed on it. The embedding is first denoised and homogenized, and then iterative denoising is performed to finally obtain the user representation after debiasing in different spaces, which can clean up the noise in the user behavior data and help discover the potential connection between user behaviors to more accurately capture the real interests of users. Finally, in the process of user interest representation reconstruction and matching recommendation, the user representation in different spaces is mapped back to the Euclidean space for reconstruction and matching operations, so that it is closer to the actual interests of users, thereby improving the personalization level and accuracy of the recommendation system. This method can provide users with more accurate and high-quality information services, meet their personalized needs, and improve user experience and satisfaction.

[0072] The embodiment of the present invention selects two representative public data sets for experiments, and verifies the effectiveness of the method of the embodiment of the present invention through comparative experiments, ablation experiments and long-tail experiments.

[0073] In order to verify the effectiveness of the method of the embodiment of the present invention, two public datasets, Last-fm and CiteULike-a, were used for experiments. The interaction records and density characteristics of these datasets are different, which helps to comprehensively evaluate the performance of the model of the embodiment of the present invention in different scenarios. Last-fm is a dataset widely used in the study of music recommendation systems. It contains 1,884 users and 17,632 items, and comes from the music sharing community Last.fm. The dataset records information such as users' listening behavior, preference tags, and social relationships, and provides valuable resources for studying user interest patterns and music recommendation algorithms. CiteULike-a is a dataset commonly used in the study of academic recommendation systems. It contains 3,277 users and 16,807 items, and comes from the academic document sharing platform Citeulikea. The dataset records users' reading, collection, and commenting on academic documents, and provides important data support for studying academic interest patterns and recommendation algorithms. The method of the embodiment of the present invention divides the two datasets into training and test sets in a ratio of 8:2, and adopts a positive and negative example ratio of 1:1 to ensure that the model can fully consider samples of different categories in both training and testing, as shown in Table 1

[0074] Table 1

[0075] Last-fm CiteULike-a Number of users 1,884 3,277 Item quantity 17,632 16,807 Number of interactions 92,826 178,062 density 0.279％ 0.293％

[0076] In the experiment, the embedding vectors of users and items are randomly initialized to 100 dimensions. The Adam optimizer is used for training, with a batch size of 512, a learning rate of 0.001, a diffusion step size of 200, and only the first 20 recommended items are considered. When evaluating, two common indicators, Recall@K and NDCG@K, are used to measure the proportion of items of interest to users in the recommendation list and the ranking quality of the list, respectively. All models and comparison methods are implemented in Python 3.9 and PyTorch1.12.0.

[0077] First, a comparative experiment is performed, in which GDID-4 is a method initialized with uniform distribution, GDID-1 is a method that uses the vector obtained by aggregating the original space embedding as a guide vector, and GDID-2 is a method that uses the embedding of each space item as a guide vector. Based on the results of the comprehensive comparative experiment, all indicators of the embodiment of the present invention on the two data sets have reached the highest level, which proves the effectiveness of the model. In the comparative experiment, the four methods of GDID-1, GDID-2, GDID-3, and GDID-4 are compared, and it is found that the four methods have certain differences in performance, indicating that different guide vectors will have corresponding effects on the recommendation performance, and when the guide vector is normally distributed, the model can better capture user preferences. It also shows that iterative denoising with a more balanced guidance method can generate better user representations, as shown in Table 2.

[0078] Table 2

[0079]

[0080] Secondly, in the ablation experiment, different geometric spaces were removed for evaluation, and the experiment was divided into a single space, a dual-space joint, and a triple-space joint group. The experimental results show that compared with other groups, the recommendation performance of the method of the embodiment of the present invention is improved on both data sets. The reason is that different geometric spaces can only solve the data structure corresponding to them. Therefore, when a certain space is removed, the model will inappropriately process some data structures, resulting in a decrease in performance, which makes it impossible to better capture the user intention under the mixed geometric structure. At the same time, it also shows that the reasonable fusion of different spaces can more comprehensively capture the relationship between user items and thus improve the overall performance, as shown in Table 3.

[0081] Table 3

[0082]

[0083]

[0084] Finally, in the long-tail experiment, the performance of the model proposed in the present invention was calculated on groups of items with different numbers of interactions, and the performance of the method of the embodiment of the present invention in effectively removing long-tail deviations was explored. In the experiment, 5 groups were set up, and the number of interactions of items in the groups gradually decreased from left to right. From the experimental results, it can be seen that with the change of groups, the performance of the method of the embodiment of the present invention has advantages. In groups with a higher number of interactions, the performance of the method of the embodiment of the present invention is comparable to or higher than that of other methods, but in groups with a lower number of interactions, the performance of the method of the embodiment of the present invention is better than that of other methods. This shows that the method of the embodiment of the present invention has significant advantages in effectively removing long-tail deviations, such as Figure 6 shown.

[0085] In summary, the method of the embodiment of the present invention evaluates the model performance through comparative experiments, ablation experiments and long-tail experiments. The comparative experiments show that different guide vectors have little effect on the model performance, indicating that the method of the embodiment of the present invention is flexible and robust in the selection of guide vectors. The ablation experiment shows that different geometric spaces have a significant impact on the model performance, especially high-dimensional space can better capture the complex relationship between users and items. The long-tail experiment verifies the effectiveness of the model in reducing long-tail deviations, and shows that in groups of items with different numbers of interactions, the method of the embodiment of the present invention shows significant advantages.

[0086] Embodiment 2:

[0087] The embodiment of the present invention further provides a user interest prediction device, comprising:

[0088] The acquisition module is used to obtain the user's historical interaction data;

[0089] A first processing module is used to map the representation in the Euclidean space to different geometric spaces according to the historical interaction data of the user;

[0090] The second processing module is used to diffuse and debias the embeddings of items in different spaces to obtain user representations after debiasing in different spaces;

[0091] The third processing module is used to map the user representations in different spaces back to the Euclidean space for reconstruction and matching to obtain the user's actual interests.

[0092] As an implementation method of the embodiment of the present invention, the geometric space adopts hyperbolic space, Euclidean space and spherical space to understand the intention behind the user behavior.

[0093] Embodiment 3:

[0094] An embodiment of the present invention further provides a user interest prediction system, characterized in that it includes: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a user interest prediction method when executed by the processor.

[0095] Embodiment 4:

[0096] An embodiment of the present invention further provides a storage medium, on which a computer program is stored, and the computer program executes the user interest prediction method when running.

[0097] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for predicting user interests, characterized in that: include: Step S1, obtaining the user's historical interaction data; Step S2: Mapping the representation in the Euclidean space to different geometric spaces according to the user's historical interaction data; Step S3, performing diffusion debiasing on the object embeddings in different spaces to obtain user representations after debiasing in different spaces; Step S4: Map the user representations in different spaces back to the Euclidean space for reconstruction and matching to obtain the user's actual interests.

2. The user interest prediction method according to claim 1, characterized in that: Geometric space uses hyperbolic space, Euclidean space, and spherical space to understand the intentions behind user behaviors.

3. The user interest prediction method according to claim 2, characterized in that: Step S3: Noise-homogenize the embeddings of items in different spaces, and then perform iterative denoising to obtain user representations after debiasing in different spaces.

4. The user interest prediction method according to claim 3, characterized in that: In step S4, reconstruction and matching are performed in a Euclidean space where similarity is easier to calculate, and a top-20 recommendation method is used to recommend the top 20 items in similarity ranking to the user.

5. A user interest prediction device, characterized in that: include: The acquisition module is used to obtain the user's historical interaction data; A first processing module is used to map the representation in the Euclidean space to different geometric spaces according to the historical interaction data of the user; The second processing module is used to diffuse and debias the embeddings of items in different spaces to obtain user representations after debiasing in different spaces; The third processing module is used to map the user representations in different spaces back to the Euclidean space for reconstruction and matching to obtain the user's actual interests.

6. The user interest prediction device according to claim 5, characterized in that: Geometric space uses hyperbolic space, Euclidean space, and spherical space to understand the intentions behind user behaviors.

7. A user interest prediction system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the user interest prediction method according to any one of claims 1 to 4 is executed.

8. A storage medium, characterized in that: The storage medium stores a computer program, which executes the user interest prediction method according to any one of claims 1 to 4 when running.

Images