Method and system for content recommendation through transfer learning based on source domain features

Abstract

The present invention provides a transfer learning video recommendation method based on source domain features, the method comprising roughly dividing source domain variables into blocks, performing PCA algorithm calculation on the formed blocks, performing spectral clustering using nodes, features and adjacency matrices, performing Bayesian fusion on the formed multiple statistics to form the final source domain features, and using the source domain features as the main trust parameters in transfer learning to transfer the existing model to the model required for the target.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method and system for content recommendation based on transfer learning of source domain features. Background Art

[0002] In the contemporary society, information is king. Faced with a huge amount of video programs, users are often at a loss. How to effectively help users search and filter out useful information from a huge amount of videos has become an urgent problem to be solved. Personalized recommendation systems have emerged as a result. Commonly used personalized recommendation methods are mainly divided into two types:

[0003] 1. Collaborative filtering recommendation method

[0004] The main idea of ​​collaborative filtering is to collect relevant information about users, analyze their interests and preferences through this information, and then find other users with similar interests and hobbies based on different methods. The collaborative filtering model is very simple and can be easily applied in practical environments. Its biggest advantage is that it can basically recommend all items that can be recommended, regardless of whether the items or projects can be recognized by the machine, such as music, movies, pictures, etc.

[0005] 2. Content-based recommendation method

[0006] The content-based recommendation system mainly extracts the content features of items, then learns and builds the user's interests based on the attribute features of items previously purchased by the user, and finally recommends items that highly match the user's interests to the user. Content-based recommendation technology also calculates the similarity of items, but compared to collaborative filtering, it determines the similarity by analyzing the attribute content of items.

[0007] However, these methods require a large amount of historical data to build models, making it difficult to meet the needs of building models without data.

[0008] Therefore, a method is needed to improve the deficiencies in the prior art. Summary of the invention

[0009] This Summary is provided to introduce some concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.

[0010] In order to solve the problem of lack of available models in the current personalized video recommendation, the present invention provides a transfer learning video recommendation method based on source domain features. Compared with the traditional transfer learning method, the features of the source domain are introduced as the main reference model of transfer learning, while the influence of the target domain is reduced, thereby solving the negative transfer effect caused by transfer learning.

[0011] Specifically, in one embodiment of the present invention, a method for content recommendation through transfer learning based on source domain features is provided, the method comprising:

[0012] Preprocess the original model;

[0013] Dividing the source domain variables of the original model into blocks to form variable blocks;

[0014] For each variable block, find the related variable blocks to determine the adjacency matrix;

[0015] Perform principal component analysis (PCA) on each variable block to extract the first variable features;

[0016] Using the nodes, features and the adjacency matrix of the original model to perform spectral clustering on each variable block to perform a second variable feature extraction;

[0017] The multiple variable features extracted twice are fused in a Bayesian manner to form source domain features;

[0018] Performing transfer learning on the original model based on the source domain features to obtain a target model; and

[0019] The target model is used to perform content recommendation.

[0020] In one embodiment of the present invention, preprocessing the original model includes preprocessing redundant structures of the original model to form a model that conforms to the target data.

[0021] In this embodiment of the present invention, preprocessing the original model further includes deleting and cleaning useless and obvious noise data of the original model.

[0022] In one embodiment of the present invention, dividing the source domain variables of the original model into blocks includes finding related variables for each individual variable to form variable blocks.

[0023] In one embodiment of the present invention, performing PCA processing on each variable block includes executing PCA algorithm calculation on the formed variable block, wherein the PCA algorithm calculation is used to divide the variable block by correlation and perform PCA dimensionality reduction processing on the divided variable block to eliminate variable blocks with weaker correlation and retain the main variable blocks.

[0024] In one embodiment of the present invention, the second variable feature extraction is performed on the clustered variable blocks.

[0025] In one embodiment of the present invention, the multiple variable features extracted twice are fused as Bayesian statistics.

[0026] In another embodiment of the present invention, a content recommendation system is provided, the system comprising:

[0027] An original model preprocessing module, which is configured to preprocess the original model;

[0028] Principal component analysis PCA feature extraction module, which is configured to:

[0029] Dividing the source domain variables of the original model into blocks to form variable blocks;

[0030] Finding related blocks of variables for each block of variables to determine an adjacency matrix; and

[0031] Perform principal component analysis (PCA) on each variable block to extract the first variable features;

[0032] A spectral clustering feature extraction module, configured to perform spectral clustering on each variable block by using the nodes, features and the adjacency matrix of the original model to perform a second variable feature extraction;

[0033] A Bayesian fusion module, configured to perform Bayesian fusion on the multiple variable features extracted twice to form source domain features;

[0034] A transfer learning module, configured to perform transfer learning on the original model based on the source domain features to obtain a target model; and

[0035] The content recommendation module is configured to use the target model to perform content recommendation.

[0036] In one embodiment of the present invention, the original model preprocessing module is further configured to preprocess redundant structures of the original model to form a model that conforms to the target data.

[0037] In this embodiment of the present invention, the original model preprocessing module is further configured to delete and clean up useless and obvious noise data of the original model.

[0038] In one embodiment of the present invention, the PCA feature extraction module is further configured to find related variables for each individual variable to form a variable block.

[0039] In one embodiment of the present invention, the PCA feature extraction module is further configured to perform PCA algorithm calculation on the formed variable blocks, and the PCA algorithm calculation is used to divide the variable blocks by correlation and perform PCA dimensionality reduction processing on the divided variable blocks to eliminate variable blocks with weaker correlation and retain the main variable blocks.

[0040] In one embodiment of the present invention, the spectral clustering feature extraction module is further configured to perform a second variable feature extraction after performing spectral clustering on each variable block.

[0041] In one embodiment of the present invention, the Bayesian fusion module is further configured to perform Bayesian fusion on the multiple statistics extracted twice to form source domain features.

[0042] In yet another embodiment of the present invention, a system for content recommendation by transfer learning based on source domain features is provided, the system comprising:

[0043] A device for preprocessing the original model;

[0044] A device for dividing the source domain variables of the original model into blocks to form variable blocks;

[0045] means for finding related variable blocks for each variable block to determine an adjacency matrix;

[0046] A device for performing principal component analysis (PCA) processing on each variable block to perform a first variable feature extraction;

[0047] A device for performing spectral clustering on each variable block by using the nodes, features and the adjacency matrix of the original model to perform a second variable feature extraction;

[0048] A device for performing Bayesian fusion on the multiple variable features extracted twice to form source domain features;

[0049] A device for performing transfer learning on the original model based on the source domain features to obtain a target model; and

[0050] A device for performing content recommendation using the target model.

[0051] In one embodiment of the present invention, the means for preprocessing the original model includes means for preprocessing redundant structures of the original model to form a model that conforms to the target data.

[0052] In this embodiment of the present invention, the means for preprocessing the original model further comprises means for deleting and cleaning useless and obvious noise data of the original model.

[0053] In one embodiment of the present invention, the means for dividing the source domain variables of the original model into blocks includes means for finding related variables for each individual variable to form variable blocks.

[0054] In one embodiment of the present invention, the device for performing PCA processing on each variable block includes a device for performing PCA algorithm calculation on the formed variable block, wherein the PCA algorithm calculation is used to divide the variable blocks according to correlation and perform PCA dimensionality reduction processing on the divided variable blocks to eliminate variable blocks with weaker correlation and retain the main variable blocks.

[0055] In yet another embodiment of the present invention, a computer readable medium storing computer executable instructions is provided, the instructions comprising:

[0056] Instructions for preprocessing the original model;

[0057] Instructions for dividing source domain variables of the original model into blocks to form variable blocks;

[0058] Instructions for finding related blocks of variables for each block of variables to determine the adjacency matrix;

[0059] Instructions for performing principal component analysis (PCA) on each variable block to extract first variable features;

[0060] An instruction for performing spectral clustering on each variable block by using the nodes, features and the adjacency matrix of the original model to perform a second variable feature extraction;

[0061] Instructions for performing Bayesian fusion on the multiple variable features extracted twice to form source domain features;

[0062] Instructions for performing transfer learning on the original model based on the source domain features to obtain a target model; and

[0063] Instructions for using the target model to perform content recommendation.

[0064] In one embodiment of the present invention, the instruction for preprocessing the original model includes an instruction for preprocessing redundant structures of the original model to form a model that conforms to the target data.

[0065] In this embodiment of the present invention, the instructions for preprocessing the original model further include instructions for deleting and cleaning useless and obvious noise data of the original model.

[0066] In one embodiment of the present invention, the instructions for dividing the source domain variables of the original model into blocks include instructions for finding related variables for each individual variable to form variable blocks.

[0067] In one embodiment of the present invention, the instructions for performing PCA processing on each variable block include instructions for executing PCA algorithm calculation on the formed variable blocks, wherein the PCA algorithm calculation is used to divide the variable blocks by correlation and perform PCA dimensionality reduction processing on the divided variable blocks to eliminate variable blocks with weaker correlation and retain the main variable blocks.

[0068] After studying the description of the specific exemplary embodiments of the present invention below in conjunction with the accompanying drawings, other aspects, features and embodiments of the present invention will be apparent to those of ordinary skill in the art. Although features of the present invention may be discussed below with respect to certain embodiments and accompanying drawings, all embodiments of the present invention may include one or more of the advantageous features discussed herein. In other words, although one or more embodiments may be discussed to have certain advantageous features, one or more of such features may also be used according to the various embodiments of the present invention discussed herein. In a similar manner, although exemplary embodiments may be discussed below as device, system or method embodiments, it should be appreciated that such exemplary embodiments may be implemented in various devices, systems, and methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to understand in detail the manner in which the above-stated features of the present disclosure are used, a more particular description of the content briefly summarized above may be made with reference to various aspects, some of which are illustrated in the accompanying drawings. It should be noted, however, that the accompanying drawings illustrate only certain typical aspects of the present disclosure and are not to be considered limiting of its scope, as the description may admit to other equally effective aspects.

[0070] Figure 1 is a block diagram of a content recommendation system according to an embodiment of the present disclosure.

[0071] Figure 2 A block diagram of components for secondary feature extraction according to an embodiment of the present disclosure is shown.

[0072] Figure 3 A flowchart of a method for content recommendation through transfer learning based on source domain features according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0073] The various embodiments will be described in more detail below with reference to the accompanying drawings that form part of the present invention and illustrate various specific exemplary embodiments. However, the various embodiments can be implemented in many different forms and should not be interpreted as limiting the various embodiments set forth herein; on the contrary, these embodiments are provided to make the present disclosure thorough and complete, and to fully convey the scope of these embodiments to those of ordinary skill in the art. The various embodiments may be implemented in accordance with methods, systems, or devices. Therefore, these embodiments may be implemented in hardware, in full software, or in a combination of software and hardware. Therefore, the following specific embodiments are not limiting.

[0074] The steps in each flow chart may be performed by hardware (e.g., processor, engine, memory, circuit), software (e.g., operating system, application, driver, machine / processor executable instructions), or a combination thereof. As will be appreciated by those skilled in the art, the methods involved in each embodiment may include more or fewer steps than shown.

[0075] Figure 1 is a block diagram of a content recommendation system 100 according to an embodiment of the present disclosure.

[0076] like Figure 1 As shown, the content recommendation system 100 includes an original model preprocessing module 102, a PCA feature extraction module 104, a spectral clustering feature extraction module 106, a Bayesian fusion module 108, a transfer learning module 110, and a content recommendation module 112. These modules will be described in detail below.

[0077] The original model preprocessing module 102 is configured to preprocess the original model to preliminarily form a model that conforms to the target data. Specifically, the original model preprocessing module is further configured to preprocess the redundant structure of the original model, and the preprocessing includes deleting and cleaning useless and obvious noise data of the original model; combining data from multiple data sources for unified storage to establish a data warehouse; and structuring and normalizing the sorted data.

[0078] The PCA feature extraction module 104 is configured to search for related variables for a single variable in the source domain, divide the variable into blocks, and then perform PCA processing on each block to complete the first variable feature extraction. The details of the PCA feature extraction module 104 will be combined with Figure 2 Further description.

[0079] The spectral clustering feature extraction module 106 is configured to use nodes, features, and adjacency matrices to perform secondary variable feature extraction on variables using a spectral clustering method. The details of the spectral clustering feature extraction module 106 will be described in conjunction with Figure 2 Further description.

[0080] The combination of the functions of the PCA feature extraction module 104 and the spectral clustering feature extraction module 106 realizes secondary extraction of variable features, thereby being able to better characterize source domain features.

[0081] The Bayesian fusion module 108 is configured to perform Bayesian fusion on the multiple variable features extracted twice, so as to use the formed source domain features in transfer learning.

[0082] The transfer learning module 110 is configured to perform transfer learning on the original model to obtain the target model. Specifically, the transfer learning module 110 is configured to perform transfer learning on the original model based on the source domain features from the Bayesian fusion module 108 to obtain the target model.

[0083] The content recommendation module 112 is configured to use the learned model to perform the content recommendation task, that is, to use the obtained target model to perform content recommendation.

[0084] Figure 2 A block diagram of components for secondary feature extraction according to an embodiment of the present disclosure is shown.

[0085] like Figure 2 As shown, the PCA feature extraction module 104 includes a source domain variable partitioning component 202, an adjacency matrix determination component 204, and a PCA processing component 206. The spectral clustering feature extraction module 106 includes a spectral clustering component 208 and a variable feature extraction component 210. These components will be described in more detail below.

[0086] The source domain variable block component 202 in the PCA feature extraction module 104 blocks the source domain variables of the original model to form variable blocks. Specifically, the source domain variable block component 202 forms variable blocks by finding related variables for each individual variable to complete variable blocking.

[0087] Next, the adjacency matrix determination component 204 finds related variable blocks for each variable block formed to determine an adjacency matrix.

[0088] Then, the PCA processing component 206 performs PCA processing on each variable block by executing PCA algorithm calculation on the formed variable block to complete the first variable feature extraction. The PCA algorithm calculation is used to divide the variable block by correlation and perform PCA dimensionality reduction processing on the divided variable block to eliminate the variable block with weak correlation and retain the main variable block.

[0089] Subsequently, the spectral clustering component 208 in the spectral clustering feature extraction module 106 performs spectral clustering on each variable block using the nodes and features of the original model and the adjacency matrix determined by the adjacency matrix determination component 204. As those skilled in the art will appreciate, any suitable spectral clustering method known in the art may be used to perform spectral clustering on each variable block, and the present invention is not limited to any specific spectral clustering method.

[0090] Finally, the variable feature extraction component 210 performs a second variable feature extraction operation on the clustered variable blocks.

[0091] Figure 3 A flowchart of a method 300 for content recommendation through transfer learning based on source domain features according to an embodiment of the present disclosure is shown.

[0092] As shown, method 300 begins at step 302. At 302, the original model is preprocessed. In one embodiment of the present invention, preprocessing the original model includes preprocessing redundant structures of the original model to form a model that conforms to the target data. In this embodiment of the present invention, preprocessing the original model further includes deleting and cleaning useless and obvious noise data of the original model.

[0093] Next, the method 300 proceeds to step 304. At step 304, the source domain variables of the original model are divided into blocks to form variable blocks. In one embodiment of the present invention, dividing the source domain variables of the original model into blocks includes finding related variables for each individual variable to form variable blocks.

[0094] Then, the method 300 continues to step 306. At 306, for each variable block, related variable blocks are found to determine an adjacency matrix. As can be appreciated by those skilled in the art, any suitable adjacency matrix determination method in the art can be used to determine the adjacency matrix, and the present invention is not limited to any specific method.

[0095] The method 300 proceeds to step 308. At step 308, PCA processing is performed on each variable block to perform a first variable feature extraction. In one embodiment of the present invention, performing PCA processing on each variable block includes executing PCA algorithm calculation on the formed variable block.

[0096] The method 300 then proceeds to step 310, where each variable block is spectrally clustered using the nodes, features and adjacency matrix of the original model to perform a second variable feature extraction. In one embodiment of the present invention, the second variable feature extraction is performed on the clustered variable blocks.

[0097] The method 300 then proceeds to step 312, where the multiple variable features extracted twice are subjected to Bayesian fusion to form source domain features. In one embodiment of the present invention, the multiple variable features extracted twice are fused as Bayesian statistics.

[0098] The method 300 then proceeds to step 314 , where transfer learning is performed on the original model based on the source domain features to obtain a target model.

[0099] Finally, the method 300 ends at step 316 , where the obtained target model is used to perform content recommendation.

[0100] In summary, the present invention provides a content recommendation method based on transfer learning of source domain features, and adds a source domain feature transfer learning module to the content recommendation system. In the absence of a data model scenario, transfer learning can be performed on an existing model, and the target model can be quickly obtained and applied to the content recommendation function. In addition, the negative transfer phenomenon that may be caused by transfer learning is optimized through the method of source domain features, and the original model is placed at a higher level of trust, thereby achieving a better content recommendation effect. Moreover, the feature extraction of the source domain adopts two feature extraction methods of block PCA extraction and spectral clustering, and Bayesian fusion is used for multiple variable features to better characterize the source domain features.

[0101] Embodiments of the present invention have been described above with reference to block diagrams and / or the operating instructions of methods, systems, and computer program products according to embodiments of the present invention. Each function / action noted in the frame may occur in an order different from that shown in any flow chart. For example, depending on the function / action involved, two frames shown in succession may actually be performed substantially simultaneously, or these frames may sometimes be performed in an opposite order.

[0102] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for content recommendation by transfer learning based on source domain features, the method comprising: Preprocess the original model; Dividing the source domain variables of the original model into blocks to form variable blocks; For each variable block, find the related variable blocks to determine the adjacency matrix; Perform principal component analysis (PCA) on each variable block to extract the first variable features; Using the nodes, features and the adjacency matrix of the original model to perform spectral clustering on each variable block to perform a second variable feature extraction; The multiple variable features extracted twice are fused in a Bayesian manner to form source domain features; Performing transfer learning on the original model based on the source domain features to obtain a target model; as well as Using the target model to perform content recommendation; The step of dividing the source domain variables of the original model into blocks includes finding related variables for each individual variable to form variable blocks.

2. The method of claim 1, wherein preprocessing the original model comprises preprocessing redundant structures of the original model to form a model that conforms to the target data.

3. The method of claim 2, wherein preprocessing the original model further comprises deleting and cleaning useless and obvious noise data of the original model.

4. The method as claimed in claim 1, wherein performing PCA processing on each variable block includes executing PCA algorithm calculation on the formed variable block, wherein the PCA algorithm calculation is used to divide the variable blocks according to correlation and perform PCA dimensionality reduction processing on the divided variable blocks to eliminate variable blocks with weaker correlation and retain the main variable blocks.

5. The method as claimed in claim 1, wherein the second variable feature extraction is performed on the clustered variable blocks.

6. The method according to claim 1, wherein the multiple variable features extracted twice are fused as Bayesian statistics.

7. A content recommendation system, the system comprising: An original model preprocessing module, configured to preprocess the original model; Principal component analysis PCA feature extraction module, which is configured to: Dividing the source domain variables of the original model into blocks to form variable blocks; For each variable block, find the related variable blocks to determine the adjacency matrix; as well as Perform principal component analysis (PCA) on each variable block to extract the first variable features; A spectral clustering feature extraction module, configured to perform spectral clustering on each variable block by using the nodes, features and the adjacency matrix of the original model to perform a second variable feature extraction; A Bayesian fusion module, configured to perform Bayesian fusion on the multiple variable features extracted twice to form source domain features; A transfer learning module, configured to perform transfer learning on the original model based on the source domain features to obtain a target model; as well as A content recommendation module, configured to perform content recommendation using the target model; The step of dividing the source domain variables of the original model into blocks includes finding related variables for each individual variable to form variable blocks.

8. A computer system comprising means for performing the method as claimed in any one of claims 1 to 6.

9. A computer-readable storage medium having instructions which, when executed, cause a machine to perform the method of any one of claims 1-6.

Images