An information processing method and apparatus for recommendation systems

Abstract

This application discloses an information processing method and apparatus for a recommender system. The method first obtains the recommendation feature information of each product in the recommended product list, such as user ID, product ID, and product brand. Then, based on the recommendation feature information and a product recommendation rating model, a predicted rating value for each product is obtained. The product recommendation rating model extracts features not only in the channel domain but also in the spatial domain, and employs multiple convolutional kernels when extracting features in the channel domain. This allows for more comprehensive and accurate feature extraction, highlighting important features and weakening unimportant features, thereby obtaining more accurate product recommendation ratings. This, in turn, arranges recommended products in an order that better aligns with user preferences and expectations, helping users find the products they truly want more quickly.

Description

Technical Field

[0001] This application relates to computer information processing technology, and more particularly to an information processing method and apparatus for use in recommendation systems. Background Technology

[0002] As e-commerce continues to expand, the number and variety of goods sold online are increasing daily. Users often have to browse a large number of products to find the ones they really want to buy, which undoubtedly wastes a lot of time and energy and leads to a poor online shopping experience.

[0003] In this context, recommendation systems emerged. A recommendation system is a system that recommends information and products that a user might be interested in based on their interests and purchasing behavior.

[0004] However, since there are still many products to recommend, the order in which the recommended products are sorted to better reflect the user's actual needs will directly determine the user experience.

[0005] Currently, most recommendation systems use the following methods for feature extraction: 1) Feature extraction using a compressed extraction network (SENet), but this approach only considers channel domain features and not spatial domain features, resulting in low accuracy; 2) Feature extraction using a convolutional block attention module (CBAM). Although CBAM considers both channel and spatial domains, the channel domain uses fixed-function convolutional kernels, limiting the features that can be extracted and failing to fully and comprehensively represent three-dimensional semantic features. Therefore, its accuracy is also low.

[0006] Therefore, improving the accuracy of feature extraction is a technical problem that recommendation systems urgently need to solve in order to make the product ranking in the recommendation system more in line with the user's real needs. Summary of the Invention

[0007] The inventors of this application have creatively provided an information processing method and apparatus for use in recommendation systems.

[0008] According to a first aspect of the embodiments of this application, an information processing method for a recommendation system is provided. The method includes: obtaining recommendation feature information of each product in a recommended product list; obtaining a predicted rating value for each product based on the recommendation feature information and a product recommendation rating model, wherein the feature extraction performed by the product recommendation rating model on the recommendation feature information includes feature extraction in the channel domain and feature extraction in the spatial domain, wherein the feature extraction in the channel domain includes feature extraction based on multi-kernel convolution; and sorting the recommended product list according to the predicted rating value of each product.

[0009] According to one embodiment of this application, obtaining a predicted rating value for each product based on recommendation feature information and a product recommendation rating model includes: performing compressed sparse coding transformation on the recommendation feature information to obtain compressed feature coding information; performing feature extraction on the compressed feature coding information and mapping the compressed feature coding information into a first embedding vector, wherein the feature extraction includes feature extraction in the channel domain and feature extraction in the spatial domain, wherein the feature extraction in the channel domain includes feature extraction based on multi-kernel convolution calculation; and calculating a predicted rating value for each product based on the first embedding vector.

[0010] According to one embodiment of this application, feature extraction is performed on compressed feature-encoded information, and the compressed feature-encoded information is mapped into a first embedding vector. This includes: performing a first feature extraction on the feature-encoded information to map the feature-encoded information into a second embedding vector; performing a second feature extraction on all or part of the vectors in the second embedding vector to obtain a third embedding vector, wherein the second feature extraction includes feature extraction in the channel domain and feature extraction in the spatial domain, wherein the feature extraction in the channel domain includes feature extraction based on multi-kernel convolution calculation; and merging the second embedding vector and the third embedding vector to obtain a first embedding vector.

[0011] According to one embodiment of this application, before merging the second embedding vector and the third embedding vector to obtain the first embedding vector, the method further includes: performing feature enhancement processing on the second embedding vector and / or the third embedding vector.

[0012] According to one embodiment of this application, feature enhancement processing includes performing an inner product and a Hadamazé operation with the feature weight value vector.

[0013] According to one embodiment of this application, feature extraction includes feature extraction based on the attention weights of each feature, wherein the attention weights of each feature are obtained through self-learning of a product recommendation rating model.

[0014] According to one embodiment of this application, calculating the predicted rating value for each product based on a first embedding vector includes: using a fully connected layer and a classifier layer to perform neural network calculations to obtain the predicted rating value for each product based on the first embedding vector.

[0015] According to one embodiment of this application, calculating the predicted rating value for each product based on a first embedding vector includes: calculating the predicted rating value for each product using a neural network based on a multi-layer neural network and a classifier layer, based on the first embedding vector.

[0016] According to one embodiment of this application, after calculating the predicted rating value of each product based on the first embedding vector, the method further includes: calculating the value of the loss function based on a preset loss function, the expected rating value of each product, and the predicted rating value; and optimizing the parameters based on the value of the loss function and the backpropagation mechanism.

[0017] According to a second aspect of the embodiments of this application, an information processing apparatus for a recommendation system is provided. The apparatus includes: a recommendation feature information acquisition module, used to acquire recommendation feature information of each product in a recommended product list; a product recommendation scoring module, used to obtain a predicted score value for each product based on the recommendation feature information and a product recommendation scoring model, wherein the feature extraction performed by the product recommendation scoring model on the recommendation feature information includes feature extraction in the channel domain and feature extraction in the spatial domain, wherein the feature extraction in the channel domain includes feature extraction based on multi-kernel convolution calculation; and a product list sorting module, used to sort the recommended product list according to the predicted score value of each product.

[0018] This application provides an information processing method and apparatus for a recommendation system. The method first obtains recommendation feature information for each product in the recommended product list, such as user ID, product ID, and product brand. Then, based on the recommendation feature information and a product recommendation rating model, a predicted rating value for each product is obtained. The product recommendation rating model extracts features not only in the channel domain but also in the spatial domain, and employs multiple convolutional kernels when extracting features in the channel domain.

[0019] Using multiple convolutional kernels for feature extraction in the channel domain, compared to using fixed convolutional kernels, can yield features of multiple granularities from coarse to fine, thus making the extracted features more comprehensive and accurate. In terms of feature extraction in the spatial domain, features within a certain range can be further extracted and enhanced, thereby strengthening important features and weakening unimportant features.

[0020] Accordingly, using this product recommendation rating model can yield more accurate product recommendation ratings, thereby arranging recommended products in an order that better aligns with user preferences and expectations, helping users find the products they truly want more quickly.

[0021] It should be understood that the implementation of this application does not need to achieve all the beneficial effects described above. Rather, a specific technical solution can achieve a specific technical effect, and other embodiments of this application can also achieve beneficial effects not mentioned above. Attached Figure Description

[0022] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, in which:

[0023] In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts.

[0024] Figure 1 This is a schematic diagram illustrating the implementation flow of an embodiment of the information processing method applied to a recommendation system according to this application;

[0025] Figure 2 This is a schematic diagram of KCBAM used in another embodiment of the information processing method applied to a recommendation system in this application;

[0026] Figure 3 This is a schematic diagram of a product recommendation scoring model used in another embodiment of the information processing method applied to a recommendation system in this application;

[0027] Figure 4 This is a schematic diagram of the composition structure of an embodiment of the information processing device applied to a recommendation system according to this application. Detailed Implementation

[0028] To make the objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0030] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0031] Figure 1 The implementation flow of an embodiment of the information processing method of this application applied to a recommendation system is shown. (Reference) Figure 1 This embodiment provides an information processing method for a recommendation system. The method includes: operation 110, obtaining recommendation feature information of each product in the recommended product list; operation 120, obtaining a predicted rating value for each product based on the recommendation feature information and a product recommendation rating model, wherein the feature extraction performed by the product recommendation rating model on the recommendation feature information includes feature extraction in the channel domain and feature extraction in the spatial domain, wherein the feature extraction in the channel domain includes feature extraction based on multi-kernel convolution calculation; and operation 130, sorting the recommended product list according to the predicted rating value of each product.

[0032] In operation 110, the recommended product list is generated through user queries or based on user historical behavior. The information processing method applied in this application is mainly used to reorder the products in the obtained recommended product list, so that the products that truly meet the user's needs and preferences are placed first, thereby reducing the time users spend browsing products and finding products that meet their expectations more quickly.

[0033] Recommendation features typically refer to the characteristics that determine whether a product is selected for a recommendation list. These mainly include user features (e.g., age, gender, and preferences), interaction features (e.g., keywords entered in the query, user click history over a certain time period, and user visit duration), and product features (e.g., brand, color, size, performance, and price). User and interaction features are used to create user profiles and understand user preferences and needs. The product recommendation list is precisely the list of products that may meet specific criteria, retrieved from thousands of products based on user profiles, preferences, and needs.

[0034] Recommendation feature information can be obtained from user profiles, introduction information, etc.; while interaction feature information can be obtained by acquiring user actions and network requests and generating cookies, etc.; product feature information can be obtained from product attributes in the product database, etc.

[0035] The aforementioned features are not only the determining factors for a product to be included in the recommended product list, but also an important factor to consider in determining its ranking.

[0036] In Operation 120, the product recommendation rating model is used to determine the degree to which a product meets the user's expectations based on the recommendation feature information. In other words, it is the probability that the user will click on the product. The products are then arranged in descending order of their degree of compliance (click probability), so that products that better meet the user's expectations are placed first, thus helping the user find the product they really want more quickly.

[0037] Therefore, the more accurate the prediction results of the product recommendation rating model, the higher the products that users really want will be ranked, the less time users spend, and the better the user experience.

[0038] For the accuracy of product recommendation rating model predictions, feature extraction of recommendation information is particularly important. Since feature extraction is the data foundation for model computation and the basis for further adjustments and optimizations, the more comprehensive and representative the extracted features, the more accurate the product recommendation rating model predictions will be.

[0039] The product recommendation rating model used in this application includes feature extraction in the spatial domain and feature extraction in the spatial domain when extracting features from the recommendation feature information.

[0040] Feature extraction in the channel domain mainly refers to performing corresponding spatial transformations on spatial domain information to extract the most critical information. For example, in image recognition, this involves identifying the regions of interest in an image and extracting the important information from those regions through transformations such as rotation, scaling, and denoising. Similarly, in semantic recognition, this involves finding the key words or phrases in a sentence and extracting the most important semantic information from the sentence through necessary transformations and denoising.

[0041] Features in the channel domain represent the components of recommendation feature information on different convolutional kernels. Feature extraction in the channel domain is equivalent to decomposing recommendation feature information into components in multiple channels. The contribution of each channel to the key information will vary. Therefore, by adding a weight to each component in each channel to represent the relevance of that channel to the key information, the weight of each channel can be learned and the weight can be used to highlight the components that are more important to the key information, thereby strengthening important features and weakening unimportant features.

[0042] In existing technologies, when performing feature extraction in the channel domain, a fixed single convolution kernel is usually used, such as 3*3 or 5*5, and the extraction is completed in one pass.

[0043] For images, whether important information in the picture can be accurately extracted mainly depends on the size of the convolutional kernel. Therefore, once a relatively appropriate convolutional kernel size is determined according to the pixel size, clarity or blur degree of the picture, extracting once in the channel domain can meet the requirements of most image processing.

[0044] However, for a product recommendation system that mainly makes predictions through semantics, using a convolutional kernel of a fixed size is very likely to fail to extract all the semantics of a sentence.

[0045] For semantic feature extraction, the process of convolving semantic information with a convolutional kernel of a certain size is similar to the process of extracting words from a sentence with a fixed length. For example, for the sentence "Chanel-style dress", if words are extracted with a one-character length, the words "small", "fragrant", "wind", "style", "skirt", "dress" will be extracted; if words are extracted with a two-character length, the words "small fragrant", "fragrant wind", "wind wind", "style", "style skirt", "skirt dress" will be extracted; if words are extracted with a three-character length, the words "Chanel style", "fragrant wind wind", "wind style", "style skirt", "style dress" will be extracted.

[0046] Thus, no matter which single character length is used to extract words, some important information may be missed. Correspondingly, in the feature extraction process of semantic analysis, if only a convolutional kernel with a fixed step size is used, it is very difficult to capture all important information with just one feature extraction.

[0047] For this reason, the embodiments of this application creatively use multiple convolutional kernels for convolution calculation. In this way, more channels with different granularities can be obtained without missing important information, and thus the accuracy of the entire product recommendation scoring model is higher.

[0048] After that, through operation 140, the recommended product list can be sorted according to the scoring prediction value of each product and returned to the user. In this way, the recommended products can be arranged in an order that better meets the user's preferences and expectations, enabling the user to find the products they really want faster.

[0049] It should be noted that Figure 1 The shown embodiment is only the most basic basic embodiment of the information processing method of this application applied to the recommendation system, and implementers can further refine and expand it based on this.

[0050] According to one embodiment of this application, obtaining a predicted rating value for each product based on recommendation feature information and a product recommendation rating model includes: performing compressed sparse coding transformation on the recommendation feature information to obtain compressed feature coding information; performing feature extraction on the compressed feature coding information and mapping the compressed feature coding information into a first embedding vector, wherein the feature extraction includes feature extraction in the channel domain and feature extraction in the spatial domain, wherein the feature extraction in the channel domain includes feature extraction based on multi-kernel convolution calculation; and calculating a predicted rating value for each product based on the first embedding vector.

[0051] Typically, machine learning models require numerical data because only numerical data can be used for computation. Therefore, various specific feature values ​​need to be encoded accordingly, which is also a quantization process.

[0052] For product recommendation systems, the total number of product items is very large, while a single user typically interacts with only a very small subset of all products. Therefore, if simple encoding (e.g., One-Hot encoding or Label Encoding) is used to represent users (rows) and behaviors (columns) when encoding recommendation features, a large sparse feature matrix consisting of many zero values ​​will be formed. This large sparse feature matrix undoubtedly consumes more memory and makes feature extraction extremely complex, requiring enormous computational power.

[0053] Therefore, large sparse feature matrices can be further compressed, for example, by compressing sparse rows (CSR), thereby saving a significant amount of memory and simplifying the complexity of feature extraction.

[0054] In this embodiment, the compressed sparse coding conversion refers to a conversion in which a sparse matrix is ​​obtained by encoding the recommended feature information, and then the sparse matrix is ​​subjected to sparse compression processing.

[0055] Compressed sparse coding transformation can be performed in two steps: first, the recommendation feature information is encoded, and then the resulting sparse feature matrix is ​​further sparsely encoded; or it can be performed in a single step, using a function or model to achieve the end-to-end transformation process.

[0056] According to one embodiment of this application, feature extraction is performed on compressed feature-encoded information, and the compressed feature-encoded information is mapped into a first embedding vector. This includes: performing a first feature extraction on the feature-encoded information to map the feature-encoded information into a second embedding vector; performing a second feature extraction on all or part of the vectors in the second embedding vector to obtain a third embedding vector, wherein the second feature extraction includes feature extraction in the channel domain and feature extraction in the spatial domain, wherein the feature extraction in the channel domain includes feature extraction based on multi-kernel convolution calculation; and merging the second embedding vector and the third embedding vector to obtain a first embedding vector.

[0057] Feature extraction based on multi-kernel convolution is often more suitable for categorical features, such as user ID, user gender, product ID, and product model. However, it is less effective for other continuous features or features that are difficult to classify, such as visit duration and product price.

[0058] At this point, performing feature extraction in both the channel domain and the spatial domain, and using a feature extraction method based on multi-kernel convolution for feature extraction in the channel domain, may not be able to completely replace the original feature extraction method, or the replacement cost may be too high (the high cost of converting continuous features into categorical features). In such cases, it can be used as a supplement to the original feature extraction method, that is, the original feature extraction method is used for the first feature extraction; then, the feature extraction method that performs both the channel domain and the spatial domain, and uses a feature extraction method based on multi-kernel convolution for feature extraction in the channel domain, is used as the second feature extraction.

[0059] In this way, the advantages of the first and second feature extractions can be fully combined to obtain more complete and comprehensive features.

[0060] According to one embodiment of this application, before merging the second embedding vector and the third embedding vector to obtain the first embedding vector, the method further includes: performing feature enhancement processing on the second embedding vector and / or the third embedding vector.

[0061] When describing or representing a specific problem, some features are more important than others. After completing the initial feature extraction, it may be that as many features as possible have been obtained to prevent information from being missed. However, it is also possible that too many features have been extracted, causing important features to be overwhelmed or unable to stand out. This makes it difficult to capture important features in subsequent processing, blurring the image or problem and thus greatly increasing the difficulty of information recognition.

[0062] In this embodiment, feature enhancement processing aims to highlight important features and weaken unimportant features in order to obtain the truly needed important information or typical features, and quickly identify or classify information.

[0063] Specifically, this can be achieved by assigning values ​​to features based on their importance or by removing unimportant features.

[0064] According to one embodiment of this application, feature enhancement processing includes performing an inner product and a Hadamazé operation with the feature weight value vector.

[0065] By assigning corresponding weight values ​​to each feature based on its importance to specific information, and using these weight values ​​and the feature's position in the vector, a feature weight vector can be constructed. Performing an inner product or Hadamardian operation using this vector and the feature vector can highlight important features and weaken less important features, thereby enhancing the features.

[0066] According to one embodiment of this application, feature extraction includes feature extraction based on the attention weights of each feature, wherein the attention weights of each feature are obtained through self-learning of a product recommendation rating model.

[0067] In this embodiment, an attention mechanism is introduced during feature extraction, and the weight of each feature is dynamically learned through the self-learning process of the product recommendation rating model. Compared with manually set feature weights, this weight value can be more accurate and refined after training and verification with a large amount of data.

[0068] According to one embodiment of this application, calculating the predicted rating value for each product based on a first embedding vector includes: using a fully connected layer and a classifier layer to perform neural network calculations to obtain the predicted rating value for each product based on the first embedding vector.

[0069] If the product recommendation rating model uses a shallow neural network, after obtaining the first embedding vector, a fully connected layer and a classifier layer (e.g., softmax) can be used to classify and inductively obtain the predicted rating value for each product.

[0070] According to one embodiment of this application, calculating the predicted rating value for each product based on a first embedding vector includes: calculating the predicted rating value for each product using a neural network based on a multi-layer neural network and a classifier layer, based on the first embedding vector.

[0071] If the product recommendation rating model uses a deep neural network, after obtaining the first embedding vector, the predicted rating value for each product can be obtained by neural network calculation based on a multi-layer neural network (such as DeepFM, Deep Cross Network, etc.) and a classifier layer (such as softmax).

[0072] According to one embodiment of this application, after calculating the predicted rating value of each product based on the first embedding vector, the method further includes: calculating the value of the loss function based on a preset loss function, the expected rating value of each product, and the predicted rating value; and optimizing the parameters based on the value of the loss function and the backpropagation mechanism.

[0073] In this way, the product recommendation and rating model can converge more quickly to obtain a model with the expected accuracy and be put into practical applications more quickly.

[0074] The above embodiments illustrate how to Figure 1 The above examples illustrate how to further refine and expand upon the basic embodiments shown. Implementers can also combine various implementation methods in the above embodiments to form new embodiments according to specific implementation conditions and needs, so as to achieve a more ideal implementation effect.

[0075] The following will combine Figures 2 to 3 This application describes a specific implementation of an information processing method for a recommendation system according to another embodiment of this application.

[0076] like Figure 2 As shown, in this embodiment of the application, the product recommendation rating model is based on a K-kernel convolutional block attention module (KCBAM) 20 to achieve feature extraction in the channel domain and feature extraction in the spatial domain. The KCBAM includes a feature input 201, a channel domain K-kernel attention module 202, a spatial domain attention module 203, and a feature output 204.

[0077] Unlike traditional CBAM, KCBAM uses K convolutional kernels of different lengths (e.g., 3*1, 5*1, 7*1, etc.) to perform convolutional calculations on the 3D recommendation data to extract features in the channel domain.

[0078] Furthermore, since traditional CBAM four-dimensional data is mainly in the form of images, it targets thought data. However, this application is applied to a recommendation system, where the received data does not have length and width and is mainly three-dimensional data. Therefore, all conv2d operations are converted to conv1d (i.e., all convolutions are converted to one-dimensional operations).

[0079] Figure 3 This embodiment of the present application is shown based on Figure 2 The product recommendation and rating model constructed using KCBAM is shown below. Figure 3 As shown, the product recommendation rating model includes: input layer 301, embedding layer 302, bilinear cross layer 303, combination layer 304, and output layer 305. Among them,

[0080] Input layer 301 receives sparse features extracted from the recommendation feature information: sparse feature 1, sparse feature 2 and sparse feature 3, and compresses the sparse features to obtain the compressed feature code.

[0081] The embedding layer 302 includes the original embedding module 3021 and the KCBAM embedding module 3022. The compressed feature code is input into the original embedding module 3021 for a first feature extraction and mapping to a second embedding vector. Then, the second embedding vector is input into the KCAM module, and after passing through the KCBM embedding module 3022, a second feature extraction and mapping is performed to obtain a third embedding vector.

[0082] The bilinear cross layer 303 uses the second or third embedding vector and the feature weight value vector to perform Hadamard operation to enhance the features of the second or third embedding vector. The feature weight value is an importance weight learned by using the compressed statistics of the feature group based on the attention mechanism.

[0083] Combination layer 304 concatenates the second and third embedding vectors to obtain the result vector, which is the first embedding vector.

[0084] Output layer 305: The first embedding vector is processed by a multi-layer neural network (Layer1, Layer2, ..., LayerN) and a softmax layer to obtain the predicted rating value for each product.

[0085] In addition, the product recommendation rating model calculates the loss function value by comparing the predicted results with the actual values, and then updates the gradient through backpropagation until it converges.

[0086] This embodiment of the application is based on the recommended feature information, such as... Figure 3 The product recommendation rating model shown obtains a predicted rating value for each product; then, the predicted rating values ​​of each product are used to sort the list of recommended products, so that products that meet the user's expectations are placed at the top, thereby shortening the time required for the user to find the product they really want and improving the user experience.

[0087] The inventors of this application conducted numerous comparative experiments using the publicly available dataset criteo, and the results are shown in Table 1:

[0088] Model Loss AUC DeepFM 0.4513 79.96 FiBiNet 0.4490 80.19 CBAM 0.4481 80.54 KCBAM 0.4438 80.94

[0089] Table 1

[0090] The parameters used are: embedding_dim = 4, Adam lr = 1e-4, batch_size = 1024, drop_out = 0.5, and dnn_hidden = (128, 128).

[0091] It should be noted that, Figures 2 to 3 The embodiments shown are merely illustrative examples of the information processing method applied to the recommendation system in this application, and are not intended to limit the implementation methods or application scenarios of the embodiments in this application. Implementers may adopt any applicable implementation methods and apply them to any applicable application scenario according to specific implementation needs and conditions.

[0092] Furthermore, embodiments of this application also provide an information processing apparatus applied to a recommendation system, such as... Figure 4 As shown, the device 40 includes: a recommendation feature information acquisition module 401, used to acquire recommendation feature information of each product in the recommended product list; a product recommendation scoring module 402, used to obtain a predicted score value for each product based on the recommendation feature information and the product recommendation scoring model, wherein the feature extraction performed by the product recommendation scoring model on the recommendation feature information includes feature extraction in the channel domain and feature extraction in the spatial domain, wherein the feature extraction in the channel domain includes feature extraction based on multi-kernel convolution calculation; and a product list sorting module 403, used to sort the recommended product list according to the predicted score value of each product.

[0093] According to one embodiment of this application, the product recommendation and rating module 402 includes: a feature encoding submodule, used to perform compressed sparse encoding conversion on the recommendation feature information to obtain compressed feature encoding information; a feature extraction submodule, used to extract features from the compressed feature encoding information and map the compressed feature encoding information into a first embedding vector, wherein the feature extraction includes feature extraction in the channel domain and feature extraction in the spatial domain, wherein the feature extraction in the channel domain includes feature extraction based on multi-kernel convolution calculation; and a rating prediction submodule, used to calculate the rating prediction value for each product based on the first embedding vector.

[0094] According to an embodiment of this application, the feature extraction submodule includes: a first feature extraction unit, used to perform a first feature extraction on the feature encoding information and map the feature encoding information into a second embedding vector; a second feature extraction unit, used to perform a second feature extraction on all or part of the vectors in the second embedding vector to obtain a third embedding vector, the second feature extraction including feature extraction in the channel domain and feature extraction in the spatial domain, wherein the feature extraction in the channel domain includes feature extraction based on multi-kernel convolution calculation; and a feature combination unit, used to merge the second embedding vector and the third embedding vector to obtain a first embedding vector.

[0095] According to one embodiment of this application, the feature extraction submodule further includes a feature enhancement unit for performing feature enhancement processing on the second embedding vector and / or the third embedding vector.

[0096] According to one embodiment of this application, the feature enhancement unit is specifically used to perform an inner product and a hadamardi operation with the feature weight value vector.

[0097] According to one embodiment of this application, the feature extraction submodule is specifically used for feature extraction based on the attention weights of each feature, wherein the attention weights of each feature are obtained through self-learning of the product recommendation rating model.

[0098] According to one embodiment of this application, the rating prediction submodule is specifically used to calculate the rating prediction value for each product using a neural network with a fully connected layer and a classifier layer based on a first embedding vector.

[0099] According to one embodiment of this application, the rating prediction submodule is specifically used to calculate the rating prediction value for each product using a neural network based on a multi-layer neural network and a classifier layer, based on a first embedding vector.

[0100] According to one embodiment of this application, the product recommendation and rating module 402 further includes: a loss function value calculation submodule, used to calculate the value of the loss function based on a preset loss function, the expected rating value of each product, and the predicted rating value; and a parameter optimization submodule, used to optimize parameters based on the value of the loss function and the backpropagation mechanism.

[0101] According to a third aspect of the present application, a computer storage medium is provided, the storage medium including a set of computer-executable instructions, which, when executed, are used to perform any of the above-described information processing methods applied to a recommendation system.

[0102] It should be noted that the above descriptions of the information processing apparatus embodiments applied to recommendation systems and the computer storage medium embodiments are similar to the descriptions of the foregoing method embodiments, and have similar beneficial effects, therefore they will not be repeated. For technical details not disclosed in the descriptions of the information processing apparatus embodiments applied to recommendation systems and the computer storage medium embodiments of this application, please refer to the descriptions of the foregoing method embodiments of this application for understanding; to save space, they will not be repeated here.

[0103] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0104] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another device, or some features may be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0105] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0106] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0107] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage media, read-only memory (ROM), magnetic disks, or optical disks.

[0108] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage media, ROM, magnetic disks, or optical disks.

[0109] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1.An information processing method applied to a recommendation system, the method comprising: obtaining recommendation feature information of each item in a recommended item list; performing compressed sparse coding conversion on the recommendation feature information to obtain compressed feature coding information; performing first feature extraction on the feature coding information to map the feature coding information into a second embedding vector; performing second feature extraction on all or part of the second embedding vectors to obtain a third embedding vector, wherein the second feature extraction comprises feature extraction in a channel domain and feature extraction in a spatial domain, and the feature extraction in the channel domain comprises feature extraction based on multi-convolution kernel convolution calculation; merging the second embedding vector and the third embedding vector to obtain a first embedding vector; calculating a rating prediction value of each item according to the first embedding vector; and sorting the recommended item list according to the rating prediction value of each item. 2.The method of claim 1, before the merging of the second embedding vector and the third embedding vector to obtain a first embedding vector, the method further comprises: performing feature enhancement processing on the second embedding vector and / or the third embedding vector. 3.The method of claim 2, wherein the feature enhancement processing comprises: performing inner product and Hadamard product operations with a feature weight value vector. 4.The method of any one of claims 1 to 3, wherein the feature extraction comprises feature extraction based on attention weights of respective features, wherein the attention weights of the respective features are obtained through self-learning of an item recommendation rating model. 5.The method of claim 1, wherein the calculating of the rating prediction value of each item according to the first embedding vector comprises: using a fully connected layer and a classifier layer to perform neural network calculation to obtain the rating prediction value of each item according to the first embedding vector. 6.The method of claim 1, wherein the calculating of the rating prediction value of each item according to the first embedding vector comprises: using a multi-layer neural network and a classifier layer to perform neural network calculation to obtain the rating prediction value of each item according to the first embedding vector. 7.The method of claim 1, after the calculating of the rating prediction value of each item according to the first embedding vector, the method further comprises: calculating a value of a preset loss function according to the rating prediction value and an expected rating value of each item; performing parameter optimization according to the value of the loss function and a back propagation mechanism. 8.An information processing device applied to a recommendation system, the device comprising: a recommendation feature information obtaining module configured to obtain recommendation feature information of each item in a recommended item list; an item recommendation rating module configured to obtain a rating prediction value of each item according to the recommendation feature information and an item recommendation rating model, wherein feature extraction performed by the item recommendation rating model on the recommendation feature information comprises feature extraction in a channel domain and feature extraction in a spatial domain, and the feature extraction in the channel domain comprises feature extraction based on multi-convolution kernel convolution calculation. The commodity list sorting module is configured to sort the recommended commodity list according to the score prediction value of each commodity. The commodity recommendation scoring module comprises: a feature encoding submodule configured to perform compressed sparse coding conversion on the recommended feature information to obtain compressed feature encoding information; a feature extraction submodule configured to perform feature extraction on the compressed feature encoding information, and map the compressed feature encoding information into a first embedding vector, wherein the feature extraction comprises feature extraction in a channel domain and feature extraction in a spatial domain, and the feature extraction in the channel domain comprises feature extraction based on multi-convolution kernel convolution calculation; and a score prediction value submodule configured to calculate a score prediction value of each commodity according to the first embedding vector. The feature extraction submodule comprises: a first feature extraction unit configured to perform first feature extraction on the feature encoding information, and map the feature encoding information into a second embedding vector; and a second feature extraction unit configured to perform second feature extraction on all or part of the vectors in the second embedding vector to obtain a third embedding vector, wherein the second feature extraction comprises feature extraction in a channel domain and feature extraction in a spatial domain, and the feature extraction in the channel domain comprises feature extraction based on multi-convolution kernel convolution calculation; and a feature combination unit configured to combine the second embedding vector and the third embedding vector to obtain the first embedding vector. The feature extraction submodule comprises: a first feature extraction unit configured to perform first feature extraction on the feature encoding information, and map the feature encoding information into a second embedding vector; and a second feature extraction unit configured to perform second feature extraction on all or part of the vectors in the second embedding vector to obtain a third embedding vector, wherein the second feature extraction comprises feature extraction in a channel domain and feature extraction in a spatial domain, and the feature extraction in the channel domain comprises feature extraction based on multi-convolution kernel convolution calculation; and a feature combination unit configured to combine the second embedding vector and the third embedding vector to obtain the first embedding vector.

Images