Object recommendation method, device and equipment based on artificial intelligence, computer readable storage medium and computer program product

Abstract

The invention provides an object recommendation method, device and equipment based on artificial intelligence, a computer readable storage medium and a computer program product. The method comprises the steps of determining user scoring features and object scoring features based on first scoring data of a user and a to-be-recommended object; generating user embedding features and object embedding features based on the user attribute data and the object attribute data; performing multi-head attention mechanism feature cross processing on the user embedded features and the object embedded features to obtain multi-head attention features corresponding to the user embedded features; performing fusion processing on the multi-head attention features, the user score features and the object score features to obtain object recommendation features and user recommendation features; and mapping the object recommendation feature and the user recommendation feature into second score data of the user and the to-be-recommended object, and executing a recommendation operation of the user and the to-be-recommended object based on the second score data. Through the application, the accuracy of predicting the score of the to-be-recommended object by the user and recommending the object can be improved.

Description

Technical Field

[0001] The present application relates to artificial intelligence technology, and in particular to an object recommendation method, device, equipment, computer-readable storage medium and computer program product based on artificial intelligence. Background Art

[0002] Recommendation algorithms are an important research direction in the field of artificial intelligence. Their purpose is to mine potential patterns and preferences from a large amount of interaction data between users and projects, so as to provide personalized recommendations for content, products or services that users may be interested in. This technology is of great significance in improving user experience, enhancing user stickiness, optimizing resource allocation, and promoting the effective dissemination of information. Through recommendation algorithms, platforms can better understand user needs and provide more accurate and considerate services.

[0003] In related technologies, traditional recommendation algorithms mainly rely on historical interaction information between users and items. However, in large cloud storage websites, this can easily lead to a sparse rating matrix. Especially when new users use it for the first time, the recommendation effect will be significantly reduced due to the lack of historical data. In addition, traditional recommendation algorithms simply splice the features of user rating data and item text data, which will lead to the loss of important information and cause the problem of information cocoon. Summary of the invention

[0004] In order to solve the technical problems existing in the related technologies, the embodiments of the present application provide an object recommendation method, device, equipment, computer-readable storage medium and computer program product based on artificial intelligence.

[0005] To achieve the above purpose, the technical solution of the embodiment of the present application is implemented as follows:

[0006] The present application embodiment provides an object recommendation method based on artificial intelligence, the method comprising:

[0007] Determining user rating features and object rating features based on first rating data between multiple users and multiple objects to be recommended;

[0008] Based on the user attribute data of the multiple users and the object attribute data of the multiple objects to be recommended, generating multiple user embedding features and multiple object embedding features;

[0009] Performing feature cross processing based on a multi-head attention mechanism on the multiple user embedding features and the multiple object embedding features to obtain a multi-head attention feature corresponding to each of the user embedding features;

[0010] Fusing the plurality of the multi-head attention features, the user rating features, and the object rating features to obtain an object recommendation feature and a user recommendation feature;

[0011] The object recommendation features and the user recommendation features are mapped into second rating data between the multiple users and the multiple objects to be recommended, and a recommendation operation between the multiple users and the multiple objects to be recommended is performed based on the second rating data.

[0012] The embodiment of the present application provides an object recommendation device based on artificial intelligence, comprising:

[0013] A rating feature determination module, used to determine user rating features and object rating features based on first rating data between multiple users and multiple objects to be recommended;

[0014] An embedding feature determination module, configured to generate a plurality of user embedding features and a plurality of object embedding features based on the user attribute data of the plurality of users and the object attribute data of the plurality of objects to be recommended;

[0015] A multi-head attention feature determination module, used to perform feature cross processing based on a multi-head attention mechanism on the multiple user embedding features and the multiple object embedding features to obtain a multi-head attention feature corresponding to each of the user embedding features;

[0016] A recommendation feature determination module, used to fuse the multiple multi-head attention features, the user rating features and the object rating features to obtain an object recommendation feature and a user recommendation feature;

[0017] A recommendation module is used to map the object recommendation features and the user recommendation features into second rating data between the multiple users and the multiple objects to be recommended, and perform a recommendation operation between the multiple users and the multiple objects to be recommended based on the second rating data.

[0018] An embodiment of the present application provides an electronic device, including:

[0019] A memory for storing computer executable instructions;

[0020] The processor is used to implement the artificial intelligence-based object recommendation method provided in the embodiment of the present application when executing the computer executable instructions stored in the memory.

[0021] An embodiment of the present application also provides a computer-readable storage medium having computer-executable instructions stored thereon, wherein the computer-executable instructions, when executed by a processor, implement the artificial intelligence-based object recommendation method provided in the embodiment of the present application.

[0022] An embodiment of the present application also provides a computer program product, including computer executable instructions, characterized in that when the computer executable instructions are executed by a processor, the object recommendation method based on artificial intelligence provided by the embodiment of the present application is implemented.

[0023] The embodiments of the present application provide an artificial intelligence-based object recommendation method, apparatus, device, computer-readable storage medium, and computer program product. By acquiring user rating features, object rating features, user embedding features, and object embedding features, a multi-head attention mechanism is introduced to perform feature cross-processing, and object recommendation features and user recommendation features are obtained. This can assign higher weights to key features of users and objects, and use a multi-head attention mechanism to cross user and object features in multiple feature subspaces to obtain richer feature combination information, more accurately predict user rating data for recommended objects, and thereby improve the pertinence and accuracy of object recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1-Figure 7 A flowchart of an artificial intelligence-based object recommendation method according to an embodiment of the present application;

[0025] Figure 8 A system block diagram of an object recommendation method based on artificial intelligence according to an embodiment of the present application;

[0026] Fig. 9 A diagram showing the predicted ratings of cloud storage items by users in an embodiment of the present application;

[0027] Fig.10 This is a schematic diagram of the composition structure of an object recommendation device based on artificial intelligence according to an embodiment of the present application;

[0028] Fig.11 A schematic diagram of the hardware structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] The present application is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0031] The rapid development of information technology has led to an explosive growth in the amount of information on the Internet. How to quickly and high-quality find the information you need in the vast ocean of data is a huge challenge. Therefore, recommendation algorithms used to solve the problem of information overload have emerged. Traditional recommendation algorithms are divided into content-based recommendation algorithms and collaborative filtering recommendation algorithms. Content-based recommendation algorithms implicitly represent user interest characteristics through user historical interaction records. Since historical interaction records often reflect the user's preferences at historical moments, it is difficult for this algorithm to mine new user interests. Collaborative filtering algorithms are mainly divided into user-based collaborative filtering algorithms and project-based collaborative filtering methods. User-based collaborative filtering algorithms are mainly expressed as "word of mouth", that is, similar people have the same interests and hobbies, and recommend projects that are of interest to similar people and have not interacted with the target user to the target user. Project-based collaborative filtering is based on the historical interaction behavior between users to obtain the similarity between different projects, so as to recommend projects similar to the project of interest to the target user. This algorithm has strong interpretability and is widely used.

[0032] Since its development, cloud computing has been receiving extensive attention. Cloud computing can provide users with dynamic, scalable and virtualized network resources. Cloud storage, which provides basic support for cloud computing services, is the key to cloud computing to achieve high performance, high security, high availability and low cost. Cloud storage uses large-scale distributed storage technology to store a large amount of user data in the cloud, allowing users to access and share data anytime and anywhere, providing data support for cloud computing. Cloud storage stores user data on different servers through multiple copies, achieving load balancing, reducing fault redundancy, improving storage efficiency and achieving high reliability. Users do not need to purchase expensive hardware facilities, and can experience high-performance storage devices at a low cost. The common cloud hard disks used to store user structured data are snapshot and backup modes. The backup mode can save data copies of user data at any time by comparing snapshot data at different times. The cloud hard disk backup uploads user data to the bucket of object storage, and realizes asynchronous disaster recovery by separating backup data from original disk data, further ensuring the security and reliability of data. With the development of cloud storage, various products have emerged in an endless stream, which can be roughly divided into object storage that can store unstructured data, cloud hard disks that store sensitive data and have high read and write speeds, and file storage suitable for shared access. There are many types of cloud hard drives alone, including high-performance cloud hard drives, ultra-fast cloud hard drives, performance-optimized cloud hard drives, and capacity-based cloud hard drives. Different types of cloud hard drives have different bandwidths and latencies, and different application scenarios. If the user requires a lower latency and a large amount of data transmitted per second, a high-performance cloud hard drive can be recommended to the user; if the user has a low budget, a performance-optimized cloud hard drive can be provided to the user on time. In order to recommend the most suitable cloud storage product to users and to increase the stickiness of cloud computing platform users, an embodiment of the present application provides a multi-feature fusion cloud storage recommendation method based on a multi-head self-attention mechanism.

[0033] In many large cloud computing websites, there are not only a large amount of rating data that can express the direct emotions of users, but also a large amount of attribute data. User attribute data mainly includes user ID, user age, user occupation, location, etc. Cloud storage attribute data mainly includes important indicators for measuring storage device performance, such as iops, latency, price, reliability, and scenarios. Compared with explicit rating data, the attribute data of users and cloud storage can deeply mine the attribute characteristics of users and cloud storage products, and have richer semantic features. In the open source software project developer recommendation method based on the secondary attention mechanism, the secondary attention mechanism is used to model the relationship between project team members and project text description information, and finally recommend suitable developers for open source projects. The recommendation system model algorithm based on convolutional neural networks in deep learning uses two convolutional neural networks with the same structure to extract user comment text information and item comment text information respectively, and couples the user comment text information and item comment text information in the last layer to predict the user's score for the item. This method alleviates the sparse problem of the rating matrix in the context of large models.

[0034] However, there are certain problems in the related technologies. First, the traditional recommendation algorithm only uses the historical interaction information between users and items. In large-scale cloud storage websites, it is easy to produce a sparse rating matrix problem. At the same time, when a new user registers a cloud computing website and the website lists a new cloud storage product, since the user's historical interaction information with cloud storage has not yet been generated at this time, a cold start problem will occur. The recommendation algorithm has no historical interaction data to refer to, and the recommendation effect drops rapidly. Secondly, the recommendation algorithm that integrates rating data and text data does not explore the deep connection between user rating data and item text data. It simply splices the features of two different granularities without deep modeling of the two types of information data, which will lead to the loss of important information and the problem of information cocoon. In addition, more and more recommendation algorithms use methods that integrate multiple auxiliary information to improve the accuracy of recommendations. However, the traditional recommendation algorithm often uses convolutional neural networks to extract feature vectors for attribute data, ignoring the key attributes and non-key attributes contained in the attribute data. For example, the user's student attribute often has more important semantic information than the user's gender attribute. Giving the same weight to all attribute information is likely to weaken the key attribute features, especially the key attributes of cloud storage greatly distinguish the application scenarios of different cloud storage products.

[0035] Based on this, an embodiment of the present application proposes a recommendation method. In various embodiments of the present application, by obtaining user rating features, object rating features, user embedding features, and object embedding features, a multi-head attention mechanism is introduced for feature cross-processing, and object recommendation features and user recommendation features are obtained. This can give higher weights to key features of users and objects, and use a multi-head attention mechanism to cross user and object features in multiple feature subspaces to obtain richer feature combination information, more accurately predict users' rating data for recommended objects, and thereby improve the pertinence and accuracy of object recommendations.

[0036] The present application embodiment provides an object recommendation method based on artificial intelligence. Figure 1 , Figure 1 is a flowchart of an object recommendation method based on artificial intelligence provided by an embodiment of the present application, which will be combined with Figure 1 Steps 101 to 106 are shown for explanation.

[0037] In step 101, based on first rating data between a plurality of users and a plurality of objects to be recommended, user rating features and object rating features are determined.

[0038] In some embodiments, step 101 can be implemented by the following technical solutions: obtaining a rating matrix corresponding to the first rating data; performing matrix decomposition processing based on a latent semantic mechanism on the rating matrix to obtain the user rating features and the object rating features.

[0039] As an example, the latent semantic mechanism assumes that there are some latent features or factors between users and objects, which can explain why users are interested in certain objects. Based on these latent features, the latent semantic model can make personalized recommendations for users. In the latent semantic mechanism, two matrices are usually constructed: user rating features and object rating features. The user rating features represent the user's interest in each latent feature, and the object rating features represent the weight of the recommended object on each latent feature. The user rating features and object rating features obtained by matrix decomposition processing based on the latent semantic mechanism can approximately represent the user's preference for the recommended object and explore the potential and difficult to directly observe associations between users and recommended objects.

[0040] As an example, the user rating feature Output generated by matrix decomposition based on latent semantic mechanism u_Rating And object scoring features Output i_Rating , the score matrix of the initial input data first score data is decomposed into two latent vector matrices to represent, namely R m×n =UI T , the rating matrix R m×nRepresents the ratings of m users on n recommended objects. This rating can be represented by the product of the user rating feature U and the object rating feature I. Each row of the user rating feature U represents a user, using the k-dimensional vector Output u_Rating_i For the specific expression, please refer to the following formula (1).

[0041]

[0042] Each row in the object score feature I represents an object to be recommended, using the k-dimensional vector Output i_Rating_i For the specific expression, please refer to the following formula (2).

[0043]

[0044] In some embodiments, see Figure 2 , Figure 1 Before step 101 shown in the figure, the following steps 107 to 109 may also be performed to obtain the first scoring data. Figure 2 Provide explanation.

[0045] In step 107, for each of the objects to be recommended, the weight of the object to be recommended in each object category is obtained, and for each of the users, the preference of the user for each of the object categories is obtained.

[0046] As an example, it is necessary to classify the recommended objects into categories and obtain the weight of the recommended objects in each category, that is, the importance of the recommended objects in the corresponding category. After classifying the recommended objects into categories, the user's preference for each category of the recommended objects is obtained.

[0047] As an example, the recommended object i is in the object category f1, f2, ..., f k The weights in are q i1 ,q i2 ,……,q ik ; User u has an object category f1,f2,...,f k The preference levels are p u1 ,p u2 ,……,p uk .

[0048] In step 108, for any object to be recommended and any user, based on the weight of the object to be recommended in each object category, the preference of the user for each object category is weighted and summed to obtain the first predicted score of the user for the object to be recommended. The specific calculation formula for the first predicted score data of any user u for any object to be recommended i is shown in the following formula (3).

[0049]

[0050] in represents the first prediction score data, p uf represents the preference of user u for latent factor, i.e., object category f, q if represents the weight of the recommended object i in the latent factor f, and l represents the number of latent factors, that is, the categories into which the recommended object is divided. The smaller l is, the fewer categories the recommended object can be divided into, and the higher the generalization of the model; the larger l is, the more categories the recommended object can be divided into, the more complex the model is, and the lower the generalization of the model is.

[0051] As an example, user u2 treats the recommended object categories f1,f2,...,f k The preference levels are p 21 ,p 22 ,……,p 2k , and the recommended object i1 is in the object category f1,f2,…,f k The weights in are q 11 ,q 12 ,……,q 1k , multiplying the preference of user u2 and the weight of the recommended object i1 in all object categories can obtain the user's first predicted score data for the recommended object: By analogy, we can further obtain the first predicted rating data of user u2 for all recommended objects. Different categories of recommended objects in the latent semantic algorithm are different latent factors. The first predicted rating data of user u2 for the recommended object i1 is:

[0052] In step 109, the first predicted scores of the multiple users for the multiple objects to be recommended are combined into first score data between the multiple users and the multiple objects to be recommended.

[0053] Through step 107 to step 109, first rating data between multiple users and multiple recommended objects are obtained, which can intuitively show the explicit data of the user's satisfaction with the recommended objects. The higher the rating, the higher the user's satisfaction with the object.

[0054] Continue to see Figure 1 In step 102, based on the user attribute data of the multiple users and the object attribute data of the multiple objects to be recommended, a plurality of user embedding features and a plurality of object embedding features are generated.

[0055] In some embodiments, see Figure 3 , Figure 1 Step 102 shown in FIG. 102 can be implemented by steps 1021 to 1022 as follows. Figure 3 Provide explanation.

[0056] In step 1021, at least one user attribute code of each user is obtained based on the user attribute data of each user. At least one object attribute code of each object to be recommended is obtained based on the object attribute data of each object to be recommended.

[0057] As an example, the user attribute data and the object attribute data are encoded, which may be encoded as one-hot encoding or multi-hot encoding, etc., which is not limited here. By encoding, the user attribute data and the object attribute data are converted into numerical data that can be processed by the machine learning model.

[0058] In step 1022, embedding mapping processing based on a word dictionary is performed on each of the user attribute codes and each of the object attribute codes to obtain a user embedding feature corresponding to each of the user attribute codes and an object embedding feature corresponding to each of the object attribute codes.

[0059] As an example, it is necessary to map the user attribute encoding and the object attribute encoding into the geometric space, and then better calculate the word vectors of different feature spaces, which can provide better input features for the processing of the multi-head self-attention mechanism. The word vector representation technology GloVe for natural language processing and the pre-trained language representation model BERT are widely used. In the embodiment of the present application, the embedding mapping process based on the word dictionary uses the GloVe word embedding model to embed the word vector from the high-dimensional space to the low-dimensional space for display. The calculation formula for obtaining the embedded feature based on the embedding mapping process of the word dictionary is shown in the following formula (4).

[0060] d i =W i e i (4)

[0061] Where i represents the i-th user attribute data or object attribute data, W i is the word dictionary matrix, e i Represents the feature vector of user attribute data or object attribute data after one-hot encoding or multi-hot encoding, that is, user attribute encoding or object attribute encoding, di Embed features for users or embed features for objects.

[0062] Through step 1021 to step 1022, the user attribute data and the object attribute data can be mapped into the geometric space, and then better calculated with the word vectors of different feature spaces, so as to better provide input features for the multi-head self-attention mechanism processing. The vectorized representation of attribute data can be realized, overcoming the dimensionality curse, while retaining the potential semantic features and correlation between the attribute data.

[0063] In some embodiments, Figure 1 Before the shown step 102, the following technical scheme can also be executed to implement the recommendation operation of adapted cloud storage products for users: obtain multiple cloud storage products as the multiple objects to be recommended; obtain at least one of the following data of the cloud storage products as the object attribute data: latency data, price data, stability evaluation data, application scenario data, unit operation number; the recommendation operation between the multiple users and the multiple objects to be recommended based on the second rating data includes: performing the following processing for each of the users: based on the second rating data corresponding to each of the cloud storage products by the user, obtain the cloud storage products adapted to the user, and perform the recommendation operation of the adapted cloud storage products for the user.

[0064] As an example, cloud storage products can be used as recommended objects, attribute data of cloud storage products can be obtained, second rating data of users on different cloud storage products can be predicted, and cloud storage products suitable for users can be recommended based on the second rating data. This can further improve the accuracy of recommended products and make the recommended new products more in line with user needs.

[0065] Continue to see Figure 1 In step 103, feature cross-processing based on a multi-head attention mechanism is performed on the multiple user embedding features and the multiple object embedding features to obtain a multi-head attention feature corresponding to each of the user embedding features.

[0066] As an example, the multi-head self-attention mechanism improves the interpretability of deep learning. The multi-head self-attention mechanism can give higher weights to word vector combinations that are more helpful for decision-making from multiple angles and in all directions during the feature combination process, and finally generate text vectors that can better reflect user attribute data and object attribute data, that is, attention-enhanced features.

[0067] In some embodiments, see Figure 4 , Figure 1 Step 103 shown in FIG. 103 can be implemented by steps 1031 to 1033 as follows. Figure 4It should be noted that step 1031 to step 1033 are performed for each user embedding feature.

[0068] In step 1031, the importance of combining the user embedding feature with each of the object embedding features is determined.

[0069] In some embodiments, step 1031 can be implemented by the following technical solutions: determining the attention weight between the user embedded feature and each of the object embedded features; summing the attention weights corresponding to multiple object embedded features to obtain a first summation result; based on the attention weight between the user embedded feature and each of the object embedded features and the first summation result, determining the importance of combining the user embedded feature with each of the object embedded features.

[0070] As an example, the importance of combining the user embedding feature with the object embedding feature can be obtained by the following formula (5).

[0071]

[0072] where d r represents the user embedding feature of the rth user attribute data, d S represents the object embedding feature of the Sth object attribute data, R is the total number of object attribute data; h represents the hth attention space; Indicates the attention correlation coefficient, i.e., the degree of importance; (h) Represents the method of calculating attention weights; Ψ (h) (d r ,d s ) represents the r ,d S The new matrix obtained by projecting the original vector space into multiple attention spaces. exp(Ψ (h) (d r ,d s )) is the attention weight, This is the first summation result.

[0073] In step 1032, a linear mapping process is performed in each attention space based on the importance levels corresponding to the plurality of object embedding features and the user embedding features to obtain a feature combination result corresponding to each of the attention spaces.

[0074] In some embodiments, see Figure 5 , Figure 4 Step 1032 shown in FIG. 1032 can be implemented by steps 1032A to 1032C as follows. Figure 5It should be noted that steps 1032A to 1032C are performed in each attention space.

[0075] In step 1032A, the user embedding feature is mapped based on the parameters corresponding to the attention space to obtain a first mapping result.

[0076] As an example, the calculation method for mapping the user embedding features based on the parameters corresponding to the attention space is shown in the following formula (6).

[0077]

[0078] in, is the first mapping result, V (h) is the parameter corresponding to the attention space, d r Represents the user embedding features of the r-th user attribute data.

[0079] In step 1032B, the importance levels corresponding to the plurality of object embedding features are summed to obtain a second summation result.

[0080] As an example, the specific calculation formula for summing the importance levels corresponding to the multiple object embedding features is shown in the following formula (7).

[0081]

[0082] in, represents the attention correlation coefficient, that is, the importance obtained by formula (5), R is the total number of attribute data of the object to be recommended, and the second summation result is a obtained by formula (7).

[0083] In step 1032C, based on the first mapping result and the second summation result, a feature combination result corresponding to the attention space is determined.

[0084] As an example, it is necessary to linearly combine the first mapping result and the second summation result in the attention space h to obtain a feature combination result of the attention space. For the specific calculation formula, see the following formula (8).

[0085]

[0086] in, represents the feature combination result of linearly combining the r-th user attribute data in the attention space h, a is the second summation result obtained by formula (7), is the first mapping result obtained by formula (6).

[0087] Continue to see Figure 4In step 1033, the feature combination results of multiple attention spaces are spliced ​​to obtain a multi-head attention feature corresponding to the user embedded feature.

[0088] As an example, we need to embed the user of the rth user attribute data into feature d r In all attention spaces, That is, the feature combination results of multiple attention spaces are combined to obtain d′ r That is, the multi-head attention features corresponding to the user embedding features. The specific combination method is shown in the following formula (9).

[0089]

[0090] Among them, R is the total number of attention spaces.

[0091] Through step 1031 to step 1033, a multi-head self-attention mechanism is introduced to improve the interpretability of deep learning. The multi-head self-attention mechanism can give higher weights to word vector combinations that are more helpful for decision-making from multiple angles and in all directions during the feature combination process.

[0092] Continue to see Figure 1 In step 104, the plurality of multi-head attention features, the user rating features and the object rating features are fused to obtain object recommendation features and user recommendation features.

[0093] In some embodiments, see Figure 6 , Figure 1 Step 104 shown can be implemented by the following steps 1041 to 1042, as shown below. Figure 6 Provide explanation.

[0094] In step 1041, feature enhancement processing is performed on the multi-head attention feature corresponding to each of the user embedded features to obtain an attention enhancement feature.

[0095] As an example, the multi-head attention feature of the user embedded feature is enhanced using the convolutional layer of the convolutional neural network. Assume that the number of neurons is m and neuron j uses a convolution kernel with a word window size of t. Apply formula (8) to obtain d′ r After convolution kernel K j The expression after the effect of is shown in the following formula (10).

[0096] K j =f(M′ u *d′ r +b j ) (10)

[0097] Where M′ u Represents the parameters of the convolution kernel, d′ r It is obtained from formula (8), that is, the multi-head attention feature corresponding to each user embedding feature, b j represents the bias; f represents the activation function, and the rectified linear unit is used for nonlinear mapping operation. The specific calculation method is shown in the following formula (11).

[0098] f(x)=max{0,x} (11)

[0099] As an example, the pooling layer can reduce the dimension of the information extracted by the convolution layer, reduce the amount of data, and save limited computing resources. Here, the maximum pooling is used to compress all attention-enhanced features to a fixed length n. c , only extract the main features to reduce overfitting. Through the maximum pooling layer, the convolution layer outputs K j will become a fixed vector O j , for specific calculation method, please refer to the following formula (12).

[0100] O j =max{K1,K2,…,K n-t+1} (12)

[0101] As an example, s convolution kernels are used to perform feature enhancement processing on the multi-head attention features of the user embedded features. The output vector of the s convolution kernels is expressed as O = (O1, O2, ... O s ), which is the attention-enhancing feature.

[0102] In step 1042, based on the attention enhancement feature, the user rating feature, and the object rating feature, an object recommendation feature and a user recommendation feature are determined.

[0103] In some embodiments, see Figure 7 , Figure 6 Step 1042 shown in FIG. 1044 can be implemented by steps 1042A to 1042C as follows. Figure 7 Provide explanation.

[0104] In step 1042A, the object scoring feature and the attention enhancement feature are added to obtain a first object feature, and the user scoring feature and the attention enhancement feature are added to obtain a first user feature.

[0105] As an example, the most direct and effective fusion method of the scoring feature and the attention enhancement feature is to directly add the corresponding dimensions of the two vectors. In this way, the low-order features are directly constructed. The calculation method of the first user feature U_low is shown in the following formula (13).

[0106] Ulow =Output u_Rating ⊙O (13)

[0107] Among them, ⊙ represents the corresponding dimension addition operation, Output u_Rating The user rating feature is obtained by formula (1), O is the output vector representation of s convolution kernels, that is, the attention enhancement feature. Similarly, the first object feature I_low can be obtained.

[0108] In step 1042B, the object scoring feature and the attention enhancement feature are deeply fused to obtain a second object feature, and the user scoring feature and the attention enhancement feature are deeply fused to obtain a second user feature.

[0109] As an example, the second user feature and the second object feature are extracted through a deep neural network. The deep neural network is composed of a fully connected multi-layer neural network and has a strong modeling ability. It can be used to extract high-order features. For the specific calculation process, see the following formulas (14) to (18).

[0110]

[0111] in, is the vector output by the i-th layer, b i is the bias of the i-th layer, w i is the weight parameter of the i-th layer, σ i is the activation function of the i-th layer, Output u_Rating It is obtained from formula (1), that is, the user rating feature, O is the attention enhancement feature, and U high is the second user feature. Similarly, the second object feature can be obtained as I high .

[0112] In step 1042C, the first user feature and the second user feature are fused into the user recommendation feature, and the first object feature and the second object feature are fused into the object recommendation feature.

[0113] As an example, the final user recommendation feature can be obtained by concatenating the first user feature and the second user feature. For the specific calculation formula, see the following formula (19).

[0114]

[0115] Among them U feature Recommend features for users, U low is the first user feature, U high is the second user feature, and the object recommendation feature I can be obtained in the same way. feature .

[0116] Through step 1041 to step 1042, data enhancement is performed through the convolution layer and the pooling layer to obtain the first user feature, the second user feature, the first object feature and the second object feature, and they are fused to obtain the user recommendation feature and the object recommendation feature, so as to obtain more effective recommendation information and further improve the accuracy of the rating prediction.

[0117] Continue to see Figure 1 In step 105, the object recommendation feature and the user recommendation feature are mapped into second rating data between the multiple users and the multiple objects to be recommended.

[0118] As an example, the user recommendation feature U feature and object recommendation features I feature Directly concatenate and send to the fully connected layer, so that the user recommendation features and object recommendation features are linearly combined in a manner with different weights, and finally regress the second rating data of the user for the recommended object. The specific calculation method is shown in the following formula (20).

[0119]

[0120] in represents the predicted second score data, W is the weight, and multiply(a,b) represents an operation that can multiply the corresponding elements of two vectors of equal dimensions. When W is 1, formula (20) is the standard latent factor model. U feature Recommend features for users, I feature Recommend features for objects.

[0121] In step 106, a recommendation operation is performed between the multiple users and the multiple objects to be recommended based on the second rating data.

[0122] An artificial intelligence-based object recommendation method provided in an embodiment of the present application is implemented by an artificial intelligence-based object recommendation model. In the training process of the artificial intelligence-based object recommendation model, user rating features and object rating features are determined based on the first rating data between multiple users and multiple objects to be recommended; multiple user embedding features and multiple object embedding features are generated based on the user attribute data of the multiple users and the object attribute data of the multiple objects to be recommended; the multiple user embedding features and the multiple object embedding features are subjected to feature cross-processing based on a multi-head attention mechanism to obtain multi-head attention features corresponding to each of the user embedding features; the multiple multi-head attention features, the user rating features and the object rating features are fused to obtain object recommendation features and user recommendation features; the object recommendation features and the user recommendation features are mapped to the second rating data between the multiple users and the multiple objects to be recommended. The loss function is designed by the difference between the second rating data and the real rating data, specifically referring to the following formula (21).

[0123]

[0124] Among them, T is the set of all samples, r ui Represents the user’s true rating of the recommended object. Scoring data for the first prediction.

[0125] In order to minimize the loss function, the embodiment of the present application uses the deep learning parameter optimization algorithm Adam optimizer for optimization. Adam can automatically adjust the learning rate according to the gradient change during the training phase and converge quickly. The training process is to minimize the loss function L, and the learning rate corresponding to the minimized loss function L is synchronously updated to the object recommendation model based on artificial intelligence, completing the training process of the object recommendation model based on artificial intelligence.

[0126] The embodiments of the present application provide an artificial intelligence-based object recommendation method, apparatus, device, computer-readable storage medium, and computer program product. By acquiring user rating features, object rating features, user embedding features, and object embedding features, a multi-head attention mechanism is introduced to perform feature cross-processing, and object recommendation features and user recommendation features are obtained. This can assign higher weights to key features of users and objects, and use a multi-head attention mechanism to cross user and object features in multiple feature subspaces to obtain richer feature combination information, more accurately predict user rating data for recommended objects, and thereby improve the pertinence and accuracy of object recommendations.

[0127] The present application is described below in conjunction with application examples.

[0128] The rapid development of information technology has led to an explosive growth in the amount of information on the Internet. How to quickly and high-quality find the information you need in the vast ocean of data is a huge challenge. Therefore, recommendation algorithms used to solve the problem of information overload have emerged. Traditional recommendation algorithms are divided into content-based recommendation algorithms and collaborative filtering recommendation algorithms. Content-based recommendation algorithms implicitly represent user interest characteristics through user historical interaction records. Since historical interaction records often reflect the user's preferences at historical moments, it is difficult for this algorithm to mine new user interests. Collaborative filtering algorithms are mainly divided into user-based collaborative filtering algorithms and project-based collaborative filtering methods. User-based collaborative filtering algorithms are mainly expressed as "word of mouth", that is, similar people have the same interests and hobbies, and recommend projects that are of interest to similar people and have not interacted with the target user to the target user. Project-based collaborative filtering is based on the historical interaction behavior between users to obtain the similarity between different projects, so as to recommend projects similar to the project of interest to the target user. This algorithm has strong interpretability and is widely used.

[0129] Since its development, cloud computing has been receiving extensive attention. Cloud computing can provide users with dynamic, scalable and virtualized network resources. Cloud storage, which provides basic support for cloud computing services, is the key to cloud computing to achieve high performance, high security, high availability and low cost. Cloud storage uses large-scale distributed storage technology to store a large amount of user data in the cloud, allowing users to access and share data anytime and anywhere, providing data support for cloud computing. Cloud storage stores user data on different servers through multiple copies, achieving load balancing, reducing fault redundancy, improving storage efficiency and achieving high reliability. Users do not need to purchase expensive hardware facilities, and can experience high-performance storage devices at a low cost. The common cloud hard disks used to store user structured data are snapshot and backup modes. The backup mode can save data copies of user data at any time by comparing snapshot data at different times. The cloud hard disk backup uploads user data to the bucket of object storage, and realizes asynchronous disaster recovery by separating backup data from original disk data, further ensuring the security and reliability of data. With the development of cloud storage, various products have emerged in an endless stream, which can be roughly divided into object storage that can store unstructured data, cloud hard disks that store sensitive data and have high read and write speeds, and file storage suitable for shared access. There are many types of cloud hard drives alone, including high-performance cloud hard drives, ultra-fast cloud hard drives, performance-optimized cloud hard drives, and capacity-based cloud hard drives. Different types of cloud hard drives have different bandwidths and latencies, and different application scenarios. If the user requires a lower latency and a large amount of data transmitted per second, a high-performance cloud hard drive can be recommended to the user; if the user has a low budget, a performance-optimized cloud hard drive can be provided to the user on time. In order to recommend the most suitable cloud storage product to users and to increase the stickiness of cloud computing platform users, an embodiment of the present application provides a multi-feature fusion cloud storage recommendation method based on a multi-head self-attention mechanism.

[0130] In many large cloud computing websites, there are not only a large amount of rating data that can express the direct emotions of users, but also a large amount of attribute data. User attribute data mainly includes user ID, user age, user occupation, location, etc. Cloud storage attribute data mainly includes important indicators for measuring storage device performance, such as iops, latency, price, reliability, and scenarios. Compared with explicit rating data, the attribute data of users and cloud storage can deeply mine the attribute characteristics of users and cloud storage products, and have richer semantic features. In the open source software project developer recommendation method based on the secondary attention mechanism, the secondary attention mechanism is used to model the relationship between project team members and project text description information, and finally recommend suitable developers for open source projects. The DeepCoNN algorithm, a recommendation system model based on convolutional neural networks in deep learning, uses two convolutional neural networks with the same structure to extract user comment text information and item comment text information respectively, and couples the user comment text information and item comment text information in the last layer to predict the user's score for the item. This method alleviates the sparse problem of the rating matrix in the context of large models.

[0131] However, there are certain problems in the related technologies. First, the traditional recommendation algorithm only uses the historical interaction information between users and items. In large-scale cloud storage websites, it is easy to produce a sparse rating matrix problem. At the same time, when a new user registers a cloud computing website and the website lists a new cloud storage product, since the user's historical interaction information with cloud storage has not yet been generated at this time, a cold start problem will occur. The recommendation algorithm has no historical interaction data to refer to, and the recommendation effect drops rapidly. Secondly, the recommendation algorithm that integrates rating data and text data does not explore the deep connection between user rating data and item text data. It simply splices the features of two different granularities without deep modeling of the two types of information data, which will lead to the loss of important information and the problem of information cocoon. In addition, more and more recommendation algorithms use methods that integrate multiple auxiliary information to improve the accuracy of recommendations. However, the traditional recommendation algorithm often uses convolutional neural networks to extract feature vectors for attribute data, ignoring the key attributes and non-key attributes contained in the attribute data. For example, the user's student attribute often has more important semantic information than the user's gender attribute. Giving the same weight to all attribute information is likely to weaken the key attribute features, especially the key attributes of cloud storage greatly distinguish the application scenarios of different cloud storage products.

[0132] The multi-feature fusion cloud storage recommendation algorithm based on the multi-head self-attention mechanism proposed in the embodiment of the present application introduces user attribute and cloud storage attribute information on the basis of alleviating the sparsity of rating data. When processing user and cloud storage attribute features, a multi-head self-attention mechanism is introduced to give higher weights to key attribute features and cross-reference user and cloud storage product attribute features in multiple feature subspaces to obtain richer feature combination information. In the feature fusion module, the embodiment of the present application fully combines the advantages of the scalability of matrix decomposition and the advantages of deep learning in extracting deep-level features to obtain low-order features and high-order features, and better connect different features. Ultimately, the purpose of deeply mining user preferences and cloud storage product attributes is achieved. See Figure 8 The embodiment of the present application is mainly divided into: a rating data extraction module 201, an attribute feature extraction module 202, and a fusion and prediction module 203.

[0133] 1. Rating data extraction module: Rating data is a kind of explicit data that can intuitively show the user's satisfaction with the cloud storage product. The higher the score, the higher the user's satisfaction with the cloud storage product. The embodiment of the present application uses a simple and high-performance latent semantic algorithm to preliminarily process the user-cloud storage rating matrix data. The specific implementation is as follows: first, all cloud storage products are classified according to a certain standard, and then the preference of any user for cloud storage products in different categories and the weight of the cloud storage product in the current category are calculated. The larger the weight obtained, the higher the matching degree between the cloud storage product and the current classified category.

[0134] See also Fig. 9 ,user u2 has cloud storage categories f1,f2,...,f k The degree of preference is p 21 ,p 22 ,……,p 2k , and cloud storage product i1 is in cloud storage category f1,f2,…,f k The weights in are q 11 ,q 12 ,……,q 1k , multiplying the preference of user u2 and the weight of cloud storage product i1 in all cloud storage product categories can get the user's predicted score for the cloud storage product as By analogy, we can further obtain the predicted scores of user u2 for all cloud storage products. Different categories of cloud storage products in the latent semantic algorithm are different latent factors. The predicted scores of user u2 for cloud storage products are:

[0135] Finally, the specific calculation formula for obtaining the scores of all users u for any cloud storage product i can be found in the following formula (22).

[0136]

[0137] in represents the predicted score, p uf represents the preference of user u for latent factor f, q if represents the weight of cloud storage product i in latent factor f, and l represents the number of latent factors, that is, the categories into which cloud storage is divided. The smaller l is, the fewer categories cloud storage can be divided into, and the higher the generalization of the model; the larger l is, the more categories cloud storage can be divided into, the more complex the model is, and the lower the generalization of the model is.

[0138] The fusion and prediction module of the embodiment of the present application is responsible for the final prediction score, so the user score latent vector Output generated by matrix decomposition is simply used in the score data extraction module. u_Rating And cloud storage product rating latent vector Output i_Rating , the initial input data user-cloud storage product rating matrix is ​​decomposed into two latent vector matrices to represent, namely R m×n =UI T , the rating matrix R m×n Represents the ratings of m users on n cloud storage products. This rating can be represented by the product of the user latent vector matrix U and the cloud storage product latent vector matrix I. Each row of the matrix U represents a user, and the k-dimensional vector Output u_Rating_i For the specific expression, please refer to the following formula (23).

[0139]

[0140] Each row in the matrix I represents a cloud storage product, using the k-dimensional vector Output i_Rating_i For the specific expression, please refer to the following formula (24).

[0141]

[0142] 2. Attribute feature extraction module: In large cloud computing websites, user rating data for cloud storage products is relatively sparse and highly subjective. If only a single rating data is used, a more accurate recommendation effect cannot be obtained. Therefore, the embodiment of the present application combines highly explanatory user attributes and attribute feature data of cloud storage products to alleviate the sparsity of the rating matrix to improve the accuracy of cloud storage product recommendations. The embodiment of the present application uses an embedding layer to compress the high-dimensional and sparse user attribute information feature vector and cloud storage product attribute information feature vector into a low-dimensional dense vector. At the same time, a multi-head self-attention mechanism is introduced in the feature crossover process to assign different weights to different features, so that the model can better simulate the preference of users with different attribute characteristics for cloud storage products.

[0143] (1) Embedding layer

[0144] The purpose of the word embedding layer is to map human language into geometric space. In the embodiment of the present application, the purpose of the word embedding layer is to map user attributes such as gender, address, age, etc. into geometric space, and then better calculate the word vectors of different feature spaces, which can provide better input features for the processing of the multi-head self-attention mechanism in the task of the entire model training. In the field of word embedding layer, the two pre-trained models of GloVe, a word vector representation technology for natural language processing, and BERT, a pre-trained language representation model, are the most widely used. In the embodiment of the present application, the word embedding layer of the model uses the GloVe word embedding model to embed word vectors from high-dimensional space to low-dimensional space for display. Words with similar semantics are close in distance in the vector space, and operations such as addition, subtraction, multiplication and division can be performed to achieve vectorized representation of text, overcome the dimensionality curse, and retain the potential semantic features and correlations between words.

[0145] First, encode the user attributes and cloud storage product attributes into one-hot encoding or multi-hot encoding. Assume According to the word dictionary W i The function that maps words to J-dimensional vectors, K i is the dimension number of the i-th feature of the user or cloud storage product. The calculation formula of the embedding vector is shown in the following formula (25).

[0146] d i =W i e i (25)

[0147] Among them, i represents the i-th attribute feature, W i is the dictionary matrix, e i Represents the feature vector of the attribute feature after one-hot encoding, d i is the corresponding embedding vector obtained through the dictionary matrix.

[0148] (2) Multi-head self-attention mechanism

[0149] In the process of combining user and cloud storage product attribute features, the embodiment of the present application introduces a multi-head self-attention mechanism that can distinguish between noise information and valuable information, thereby improving the interpretability of deep learning. The multi-head self-attention mechanism can give higher weights to word vector combinations that are more helpful for decision-making from multiple angles and in all directions during the feature combination process, and ultimately generate text vectors that can better reflect user attribute information and cloud storage product attribute features. Taking the rth user feature and the sth cloud storage product feature as an example, the feature crossover process is as follows:

[0150] First, the importance of the combination of the above two features under the attention head h is obtained. The specific calculation formula is shown in the following formula (26).

[0151]

[0152] Where r and s are used to represent the rth user and sth cloud storage product features respectively, R is the total number of cloud storage product attribute features; h represents the hth attention head; represents the attention correlation coefficient; Ψ (h) Indicates the method of calculating attention weight; d r ,d s denote the embedding vector of the rth feature of the user and the embedding vector of the sth feature of the cloud storage product respectively; Ψ (h) (d r ,d s ) represents the r ,d s The new matrix obtained by projecting the original vector space into multiple attention head spaces.

[0153] Secondly, the rth feature is linearly combined in the attention head space h. The specific calculation formula is shown in the following formula (27).

[0154]

[0155] in, represents the result of linearly combining the rth feature in the attention head space h. represents the attention correlation coefficient, V (h) represents the linear combination parameter, d r represents the embedding vector of the rth feature of the user, s represents the sth feature of the cloud storage product, and R is the total number of cloud storage product attribute features.

[0156] Next, d r The vector obtained in all attention head spaces Combine to get d′ r , for specific combination methods, please refer to the following formula (28).

[0157]

[0158] Among them, d′ r is the concatenation operation, and R is the total number of attention heads. r for The concatenated result vector, d′ r It will be used as the input of the convolutional neural network.

[0159] (3) Convolutional layer

[0160] The convolution layer is the core layer of the convolutional neural network. This layer extracts feature information through convolution kernels. Different filters can extract multiple features. After the convolution operation of the convolution layer, the original input features will be enhanced. The embodiment of the present application uses the convolution layer to enhance the key features of users and cloud storage products. Assume that the number of neurons is m, and neuron j uses a convolution kernel with a word window size of t. The word vector matrix d′ obtained by formula (28) r After convolution kernel K j The expression after the effect of is shown in the following formula (29).

[0161] K j =f(M′ u *d′ r +b j ) (29)

[0162] Where M′ u Represents the parameters of the convolution kernel, d′ r From formula (28), b j represents the bias, f represents the activation function, and the rectified linear unit is used for nonlinear mapping operation. The specific calculation method is shown in the following formula (30).

[0163] f(x)=max{0,x} (30)

[0164] (4) Pooling layer

[0165] The pooling layer can reduce the dimension of the information extracted by the convolution layer, reduce the amount of data, and save limited computing resources. In the embodiment of the present application, the maximum pooling is used to compress all the text sets of a single user to a fixed length n. c , which is convenient for subsequent text processing and only extracts the main features, reducing overfitting. Through the maximum pooling layer, the convolution layer outputs K j will become a fixed vector O j , for specific calculation method, please refer to the following formula (31).

[0166] O j=max{K1,K2,…,K n-t+1} (31)

[0167] Use s convolution kernels to extract different text features respectively, then the output vector of s convolution kernels is expressed as O = (O1, O2, ... O s ).

[0168] 3. Fusion and prediction module

[0169] (1) Feature Fusion

[0170] The user and cloud storage product rating data and user and cloud storage attribute feature data used in the present embodiment are non-homologous data and exist in different feature spaces. They cannot be simply connected end to end to form a vector. u_i_text 、Output u_Rating 、Output i_Rating Perform low-order feature combination and high-order feature extraction. In order to facilitate the interaction and fusion of feature vectors from different sources, the four feature vectors are all k-dimensional vectors. u_i_text O=(O1,O2,…O s ).

[0171] Low-order features: Relevant research shows that the most direct and effective way to fuse text vectors and rating latent vectors is to directly add the corresponding dimensions of the two vectors. This method is used to directly construct low-order features. The calculation method of the user's low-order feature U_low is shown in the following formula (32).

[0172] U low =Output u_Rating ⊙Output u_i_text (32)

[0173] Among them, ⊙ represents the addition operation of the corresponding dimension. Similarly, the low-order feature I_low of the stored product can be obtained. Output u_Rating It can be obtained by formula (23).

[0174] High-order features: The embodiment of the present application extracts high-order features of users and stored products through a deep neural network. The deep neural network is composed of a fully connected multi-layer neural network, has a strong modeling ability, and can be used to extract nonlinear high-order features. The user vector and the user and stored product attribute feature vectors are used as inputs and sent to the deep neural network respectively to learn high-order features. The specific calculation process is shown in the following formulas (33) to (37).

[0175]

[0176] in, is the vector output by the i-th layer, b i is the bias of the i-th layer, w i is the weight parameter of the i-th layer, σ i is the activation function of the i-th layer. Similarly, the high-order features of the item can be obtained as I high .

[0177] The final user depth feature can be obtained by concatenating the first user feature and the second user feature. For the specific calculation formula, please refer to the following formula (38).

[0178]

[0179] Among them U feature is the final user deep feature. Similarly, we can get the cloud storage product deep feature I feature .

[0180] (2) Rating prediction

[0181] In order to improve the recommendation accuracy, the embodiment of the present application obtains the final user depth feature vector U in the rating prediction stage. feature and cloud storage product deep feature vector I feature Directly concatenate and send to the fully connected layer, so that the user's deep features and the cloud storage product's deep features are linearly combined with different weights, and finally regress the user's rating of the item. The specific calculation method is shown in the following formula (39).

[0182]

[0183] in represents the final prediction score of the model, W is the weight, and multiply(a,b) represents an operation that can multiply the corresponding elements of two vectors of equal dimensions. When W is 1, formula (39) is a standard latent factor model.

[0184] Since the recommendation method proposed in the embodiment of the present application is ultimately intended to obtain the user's predicted score for the storage product, and since it is essentially a regression problem, the design of the objective function is specifically referred to the following formula (40).

[0185]

[0186] Among them, T is the set of all samples, r ui Represents users' real ratings of cloud storage products.

[0187] In order to minimize the loss function, the embodiment of the present application uses the deep learning parameter optimization algorithm Adam optimizer for optimization. Adam can automatically adjust the learning rate according to the gradient change during the training phase and converge quickly. The training process is to minimize the loss function L.

[0188] The overall algorithm is as follows:

[0189] (1) Obtain the user-cloud storage product rating matrix from the original data based on the user ID and cloud storage product ID, and use the latent factor model to perform matrix decomposition to obtain the user rating feature vector and the cloud storage product rating feature vector;

[0190] (2) Preprocess the user attribute information and cloud storage attribute information, obtain the word vector matrix through the embedding layer, and then use multi-head attention to cross user features and cloud storage product features to give different attention levels to different user and cloud storage product features. Finally, send them to the convolutional neural network to obtain the user and cloud storage attribute text feature vectors;

[0191] (3) The user and cloud storage product feature vectors obtained in (1) and (2) are processed by simple concatenation and deep neural network to obtain low-order features and high-order features respectively, and then the low-order features and high-order features are concatenated to obtain the final user deep features and cloud storage product deep features;

[0192] (4) Concatenate the user deep features and cloud storage product deep features in (3) to obtain the final feature vector, which is then sent to the fully connected layer to obtain the predicted score.

[0193] Compared with the methods of related technologies, this application has the following beneficial effects:

[0194] The embodiments of the present application can fuse multiple data, namely rating data and attribute data, to better improve the accuracy of recommendations. In the attribute feature extraction module, a multi-head self-attention mechanism is used to give different attention to different feature combinations, focusing on more valuable semantic features in the attribute text, so as to more accurately obtain user interest preferences and cloud storage attribute features. In the fusion and rating prediction layer, low-order features and high-order features are established for data in two different feature spaces, and finally the high-order features and low-order features are deeply fused to obtain more effective attribute text and rating feature information, further improving the accuracy of rating prediction.

[0195] The advantage of the embodiment of the present application is that it can integrate various auxiliary information, so as to better extract user features and cloud storage information feature information, and improve the pertinence and accuracy of recommendations. When extracting user and cloud storage product attribute features, a multi-head self-attention mechanism is introduced to give different attention to different feature combinations, realize the effective distinction between noise data and valuable data, and deepen the weight feature information of key feature texts. It is possible to extract low-order features and high-order features of two types of non-homologous features located in different feature spaces, and further optimize feature extraction.

[0196] Recommendation algorithms have important application prospects in the field of artificial intelligence. The embodiment of this application applies recommendation algorithms to cloud storage products, helping cloud computing platform users to more conveniently and quickly select cloud storage products that they are interested in and have high cost performance among a large number of cloud storage products. This can improve user stickiness and revenue for cloud computing platform service providers, promote the development of the cloud computing service market, and has important commercial value.

[0197] In order to implement the object recommendation method based on artificial intelligence in the embodiment of the present application, the embodiment of the present application also provides an object recommendation device based on artificial intelligence. Fig.10 This is a schematic diagram of the composition structure of an object recommendation device based on artificial intelligence according to an embodiment of the present application. The object recommendation device based on artificial intelligence includes:

[0198] The scoring feature determination module 301 is used to determine the user scoring feature and the object scoring feature based on the first scoring data between the multiple users and the multiple objects to be recommended;

[0199] An embedding feature determination module 302, configured to generate a plurality of user embedding features and a plurality of object embedding features based on the user attribute data of the plurality of users and the object attribute data of the plurality of objects to be recommended;

[0200] A multi-head attention feature determination module 303 is used to perform feature cross processing based on a multi-head attention mechanism on the multiple user embedded features and the multiple object embedded features to obtain a multi-head attention feature corresponding to each of the user embedded features;

[0201] A recommendation feature determination module 304 is used to fuse the multiple multi-head attention features, the user rating features and the object rating features to obtain an object recommendation feature and a user recommendation feature;

[0202] The recommendation module 305 is used to map the object recommendation features and the user recommendation features into second rating data between the multiple users and the multiple objects to be recommended, and perform recommendation operations between the multiple users and the multiple objects to be recommended based on the second rating data.

[0203] In some embodiments, the artificial intelligence-based object recommendation device also includes a first scoring data acquisition module, which is used to obtain the weight of each object to be recommended in each object category, and to obtain the user's preference for each object category for each user; for any object to be recommended and any user, based on the weight of any object to be recommended in each object category, a weighted summation process is performed on the preference of any user for each object category to obtain a first predicted score of any user for any object to be recommended; and the first predicted scores of multiple users for multiple objects to be recommended are combined into the first scoring data between the multiple users and the multiple objects to be recommended.

[0204] In some embodiments, the rating feature determination module in the artificial intelligence-based object recommendation device is also used to obtain a rating matrix corresponding to the first rating data; perform matrix decomposition processing on the rating matrix based on a latent semantic mechanism to obtain the user rating features and the object rating features.

[0205] In some embodiments, the embedded feature determination module in the artificial intelligence-based object recommendation device is also used to obtain at least one user attribute code for each of the users based on the user attribute data of each of the users, and to obtain at least one object attribute code for each of the objects to be recommended based on the object attribute data of each of the objects to be recommended; and to perform embedding mapping processing based on a word dictionary on each of the user attribute codes and each of the object attribute codes to obtain user embedded features corresponding to each of the user attribute codes and object embedded features corresponding to each of the object attribute codes.

[0206] In some embodiments, the multi-head attention feature determination module in the artificial intelligence-based object recommendation device is also used to perform the following processing for each of the user embedded features: determine the importance of combining the user embedded feature with each of the object embedded features; perform linear mapping processing in each attention space based on the importance corresponding to a plurality of the object embedded features and the user embedded features, to obtain a feature combination result corresponding to each of the attention spaces; and splice the feature combination results of a plurality of the attention spaces to obtain a multi-head attention feature corresponding to the user embedded feature.

[0207] In some embodiments, the multi-head attention feature determination module in the artificial intelligence-based object recommendation device is also used to determine the attention weight between the user embedded feature and each of the object embedded features; sum the attention weights corresponding to the multiple object embedded features to obtain a first summation result; based on the attention weight between the user embedded feature and each of the object embedded features and the first summation result, determine the importance of combining the user embedded feature with each of the object embedded features.

[0208] In some embodiments, the multi-head attention feature determination module in the artificial intelligence-based object recommendation device is also used to perform the following processing for each of the attention spaces: mapping the user embedded features based on the parameters corresponding to the attention space to obtain a first mapping result; summing the importance levels corresponding to multiple user embedded features to obtain a second summation result; and determining the feature combination result corresponding to the attention space based on the first mapping result and the second summation result.

[0209] In some embodiments, the recommendation feature determination module in the artificial intelligence-based object recommendation device is also used to perform feature enhancement processing on the multi-head attention features corresponding to each of the user embedded features to obtain attention enhancement features; and determine the object recommendation features and the user recommendation features based on the attention enhancement features, the user rating features and the object rating features.

[0210] In some embodiments, the recommendation feature determination module in the artificial intelligence-based object recommendation device is also used to add the object scoring feature and the attention enhancement feature to obtain a first object feature, and add the user scoring feature and the attention enhancement feature to obtain a first user feature; perform deep fusion processing on the object scoring feature and the attention enhancement feature to obtain a second object feature, and perform deep fusion processing on the user scoring feature and the attention enhancement feature to obtain a second user feature; fuse the first user feature and the second user feature into the user recommendation feature, and fuse the first object feature and the second object feature into the object recommendation feature.

[0211] In some embodiments, the artificial intelligence-based object recommendation device also includes a cloud storage product recommendation module, which is used to obtain multiple cloud storage products as the multiple objects to be recommended; obtain at least one of the following data of the cloud storage products as the object attribute data: delay data, price data, stability evaluation data, application scenario data, unit operation number; the recommendation operation between the multiple users and the multiple objects to be recommended based on the second rating data includes: performing the following processing for each of the users: based on the second rating data corresponding to each of the cloud storage products by the user, obtaining a cloud storage product adapted to the user, and performing a recommendation operation on the adapted cloud storage product for the user.

[0212] In actual application, the scoring feature determination module 301, the embedding feature determination module 302, the multi-head attention feature determination module 303 and the recommendation feature determination module 304 can be implemented by a processor in an object recommendation device based on artificial intelligence; the recommendation module 305 can be implemented by a communication interface in the object recommendation device based on artificial intelligence.

[0213] It should be noted that: the artificial intelligence-based object recommendation device provided in the above embodiment only uses the division of the above-mentioned program modules as an example when performing object recommendation. In actual applications, the above-mentioned processing can be assigned to different program modules as needed, that is, the internal structure of the device can be divided into different program modules to complete all or part of the processing described above.

[0214] Based on the hardware implementation of the above program modules, and in order to implement the object recommendation method based on artificial intelligence in the embodiment of the present application, the embodiment of the present application also provides an electronic device, Fig.11 The hardware structure diagram of the electronic device of the embodiment of the present application is as follows. The electronic device 400 includes:

[0215] Communication interface 401, capable of exchanging information with other electronic devices;

[0216] The processor 402 is connected to the communication interface 401 to implement information interaction with other electronic devices, and is used to execute the above-mentioned artificial intelligence-based object recommendation method when running a computer program, and the computer program is stored in the memory 403.

[0217] Specifically, the processor 402 is used to determine user rating features and object rating features based on first rating data between multiple users and multiple objects to be recommended; generate multiple user embedding features and multiple object embedding features based on user attribute data of the multiple users and object attribute data of the multiple objects to be recommended; perform feature cross-processing based on a multi-head attention mechanism on the multiple user embedding features and the multiple object embedding features to obtain multi-head attention features corresponding to each of the user embedding features; and perform fusion processing on the multiple multi-head attention features, the user rating features and the object rating features to obtain object recommendation features and user recommendation features.

[0218] The communication interface 401 is used to map the object recommendation features and the user recommendation features into second rating data between the multiple users and the multiple objects to be recommended, and perform recommendation operations between the multiple users and the multiple objects to be recommended based on the second rating data.

[0219] It should be noted that the specific processing process of the communication interface 401 and the processor 402 can be understood by referring to the above-mentioned object recommendation method based on artificial intelligence.

[0220] Of course, in actual application, the various components in the electronic device 400 are coupled together through the bus system 404. It is understandable that the bus system 404 is used to realize the connection and communication between these components. In addition to the data bus, the bus system 404 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, Fig.11 Various buses are labeled as bus system 404 .

[0221] The memory 403 in the embodiment of the present application is used to store various types of data to support the operation of the electronic device 400. Examples of such data include: any computer program used to operate on the electronic device 400.

[0222] The object recommendation method based on artificial intelligence provided in the above embodiment of the present application can be applied to the processor 402, or implemented by the processor 402. The processor 402 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above-mentioned object recommendation method based on artificial intelligence can be completed by the hardware integrated logic circuit or software instructions in the processor 402. The above-mentioned processor 402 may be a general-purpose processor, a digital signal processor (DSP, Digital Signal Processor), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The processor 402 can implement or execute the object recommendation method, steps and logic block diagram based on artificial intelligence disclosed in the embodiment of the present application. A general-purpose processor may be a microprocessor or any conventional processor, etc. In combination with the steps of the object recommendation method based on artificial intelligence disclosed in the embodiment of the present application, it can be directly embodied as a hardware decoding processor to execute, or it can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium, which is located in the memory 403, and the processor 402 reads the information in the memory 403, and completes the steps of the above-mentioned object recommendation method based on artificial intelligence in combination with its hardware.

[0223] In an exemplary embodiment, the electronic device 400 can be implemented by one or more application specific integrated circuits (ASIC), DSP, programmable logic device (PLD), complex programmable logic device (CPLD), field programmable gate array (FPGA), general processor, controller, microcontroller (MCU), microprocessor, or other electronic components to execute the aforementioned artificial intelligence-based object recommendation method.

[0224] In an exemplary embodiment, the present application also provides a storage medium, namely a computer storage medium, specifically a computer-readable storage medium, for example, a memory 403 storing a computer program, and the computer program can be executed by a processor 402 in an electronic device 400 to complete the steps of the object recommendation method based on artificial intelligence described in the embodiment of the present application. The computer-readable storage medium can be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, FlashMemory, magnetic surface storage, optical disk, or CD-ROM.

[0225] It should be noted that: "first", "second", "third", etc. are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0226] In addition, the technical solutions described in the embodiments of the present application can be combined arbitrarily without conflict.

[0227] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An object recommendation method based on artificial intelligence, characterized in that: The method comprises: Determining user rating features and object rating features based on first rating data between multiple users and multiple objects to be recommended; Based on the user attribute data of the multiple users and the object attribute data of the multiple objects to be recommended, generating multiple user embedding features and multiple object embedding features; Performing feature cross processing based on a multi-head attention mechanism on the multiple user embedding features and the multiple object embedding features to obtain a multi-head attention feature corresponding to each of the user embedding features; Fusing the plurality of the multi-head attention features, the user rating features, and the object rating features to obtain an object recommendation feature and a user recommendation feature; The object recommendation features and the user recommendation features are mapped into second rating data between the multiple users and the multiple objects to be recommended, and a recommendation operation between the multiple users and the multiple objects to be recommended is performed based on the second rating data.

2. The method according to claim 1, characterized in that Before determining the user scoring features and the object scoring features based on the first scoring data between the plurality of users and the plurality of objects to be recommended, the method further includes: For each of the objects to be recommended, the weight of the object to be recommended in each object category is obtained, and for each of the users, the preference of the user for each of the object categories is obtained; For any object to be recommended and any user, based on the weight of the object to be recommended in each object category, a weighted summation process is performed on the preference of the user for each object category to obtain a first predicted score of the user for the object to be recommended; The first predicted scores of the multiple users for the multiple objects to be recommended are combined into first score data between the multiple users and the multiple objects to be recommended.

3. The method according to claim 1, characterized in that The determining of user rating features and object rating features based on first rating data between multiple users and multiple objects to be recommended includes: Obtaining a scoring matrix corresponding to the first scoring data; The rating matrix is ​​subjected to matrix decomposition processing based on a latent semantic mechanism to obtain the user rating feature and the object rating feature.

4. The method according to claim 1, characterized in that: The generating a plurality of user embedding features and a plurality of object embedding features based on the user attribute data of the plurality of users and the object attribute data of the plurality of objects to be recommended includes: Based on the user attribute data of each of the users, obtaining at least one user attribute code of each of the users, and based on the object attribute data of each of the objects to be recommended, obtaining at least one object attribute code of each of the objects to be recommended; Each of the user attribute codes and each of the object attribute codes is respectively subjected to embedding mapping processing based on a word dictionary to obtain a user embedding feature corresponding to each of the user attribute codes and an object embedding feature corresponding to each of the object attribute codes.

5. The method according to claim 1, characterized in that Performing feature cross processing based on a multi-head attention mechanism on the multiple user embedding features and the multiple object embedding features to obtain a multi-head attention feature corresponding to each of the user embedding features, including: The following processing is performed for each of the user embedded features: determining the importance of combining the user embedding feature with each of the object embedding features; Based on the importance levels respectively corresponding to the plurality of object embedding features and the user embedding features, a linear mapping process is performed in each attention space to obtain a feature combination result corresponding to each of the attention spaces; The feature combination results of multiple attention spaces are concatenated to obtain a multi-head attention feature corresponding to the user embedding feature.

6. The method according to claim 5, characterized in that The determining the importance of combining the user embedding feature with each of the object embedding features comprises: Determining an attention weight between the user embedding feature and each of the object embedding features; Summing the attention weights corresponding to the plurality of object embedding features to obtain a first summation result; Based on the attention weight between the user embedding feature and each of the object embedding features and the first summation result, the importance of combining the user embedding feature with each of the object embedding features is determined.

7. The method according to claim 5, characterized in that The linear mapping process in each attention space is performed based on the importance levels respectively corresponding to the plurality of object embedding features and the user embedding features to obtain a feature combination result corresponding to each of the attention spaces, including: The following processing is performed for each of the attention spaces: Mapping the user embedding feature based on the parameters corresponding to the attention space to obtain a first mapping result; Summing the importance levels respectively corresponding to the plurality of object embedding features to obtain a second summation result; Based on the first mapping result and the second summation result, a feature combination result corresponding to the attention space is determined.

8. The method according to claim 1, characterized in that The fusing of the plurality of multi-head attention features, the user rating features and the object rating features to obtain object recommendation features and user recommendation features includes: Performing feature enhancement processing on the multi-head attention feature corresponding to each of the user embedded features to obtain an attention enhancement feature; Based on the attention enhancement feature, the user rating feature and the object rating feature, an object recommendation feature and a user recommendation feature are determined.

9. The method according to claim 8, characterized in that The determining of the object recommendation feature and the user recommendation feature based on the attention enhancement feature, the user rating feature and the object rating feature includes: Adding the object scoring feature and the attention enhancement feature to obtain a first object feature, and adding the user scoring feature and the attention enhancement feature to obtain a first user feature; Performing deep fusion processing on the object scoring feature and the attention enhancement feature to obtain a second object feature, and performing deep fusion processing on the user scoring feature and the attention enhancement feature to obtain a second user feature; The first user feature and the second user feature are fused into the user recommendation feature, and the first object feature and the second object feature are fused into the object recommendation feature.

10. The method according to claim 1, characterized in that Before generating a plurality of user embedding features and a plurality of object embedding features based on the user attribute data of the plurality of users and the object attribute data of the plurality of objects to be recommended, the method includes: Acquire multiple cloud storage products as the multiple objects to be recommended; Acquire at least one of the following data of the cloud storage product as the object attribute data: latency data, price data, stability evaluation data, application scenario data, and unit operation times; The performing the recommendation operation between the plurality of users and the plurality of objects to be recommended based on the second rating data includes: The following processing is performed for each of the users: Based on the second rating data corresponding to each of the cloud storage products by the user, a cloud storage product adapted to the user is acquired, and a recommendation operation of the adapted cloud storage product is performed for the user.

11. An object recommendation device based on artificial intelligence, characterized in that: The device comprises: A rating feature determination module, used to determine user rating features and object rating features based on first rating data between multiple users and multiple objects to be recommended; an embedding feature determination module, configured to generate a plurality of user embedding features and a plurality of object embedding features based on the user attribute data of the plurality of users and the object attribute data of the plurality of objects to be recommended; A multi-head attention feature determination module, used to perform feature cross processing based on a multi-head attention mechanism on the multiple user embedding features and the multiple object embedding features to obtain a multi-head attention feature corresponding to each of the user embedding features; A recommendation feature determination module, used for fusing the plurality of the multi-head attention features, the user rating features and the object rating features to obtain an object recommendation feature and a user recommendation feature; A recommendation module is used to map the object recommendation features and the user recommendation features into second rating data between the multiple users and the multiple objects to be recommended, and perform a recommendation operation between the multiple users and the multiple objects to be recommended based on the second rating data.

12. An electronic device, characterized in that: include: A memory for storing computer executable instructions; A processor, configured to implement the method according to any one of claims 1 to 10 when executing computer executable instructions stored in the memory.

13. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: When the computer executable instructions are executed by a processor, the method according to any one of claims 1 to 10 is implemented.

14. A computer program product comprising computer executable instructions, characterized in that: When the computer executable instructions are executed by a processor, the method according to any one of claims 1 to 10 is implemented.

Images