Method and apparatus, device, medium, and product for personalized recommendation of commodities

Abstract

This application relates to a method and its device, equipment, medium, and product for personalized recommendation of goods in the field of e-commerce information technology. The method includes: obtaining the personalized information of a user, which includes personal characteristic information, commodity preference characteristic information, and spatio-temporal characteristic information; respectively determining, using different neural network models, a first popularity characteristic corresponding to the personalized characteristic information, a second popularity characteristic corresponding to the commodity characteristic information of candidate goods, and a third popularity characteristic corresponding to the comprehensive characteristic information formed by the personalized information and the commodity characteristic information; calculating and obtaining an effective popularity based on the first and second popularity characteristics to intervene in the third popularity characteristic; and determining whether to push the candidate goods according to the effective popularity. This application can better restore the popularity of goods in the commodity database, match more interesting goods that better meet the personal preferences of users, enable more potential long-tail goods to have the opportunity to be exposed, and reduce the Matthew effect in the process of goods recommendation.

Description

Technical Field

[0001] This application relates to the field of e-commerce information technology, and in particular, to a method for personalized product recommendation, as well as a corresponding device, computer device, computer-readable storage medium, and computer program product. Background Art

[0002] Product ranking algorithms have proven their value in various fields, and users increasingly rely on systems for product recommendations. At the same time, there are a series of biases in the system itself, such as exposure bias, selection bias, and popularity bias.

[0003] Regarding popularity bias, its definition is that popular items become more popular. It is caused by the current training paradigm and is also the root cause of the "information cocoon" for users. The existence of popularity bias will reduce the degree of personalization, reduce the fairness of recommendations, and exacerbate the Matthew effect. There are three categories of existing methods for eliminating popularity bias:

[0004] Transfer learning: Using the idea of domain adaptation, transferring the knowledge learned from popular items to long-tail products, alleviating the problem of uneven sample distribution, such as the ESAM algorithm;

[0005] Adversarial learning: Adversarial learning between the generator G and the discriminator D. D learns the implicit management of popular products and long-tail products, while G captures more long-tail products related to the user's history, so as to recommend long-tail products;

[0006] Introducing unbiased data: The KDD algorithm, which uses a multi-hop random walk algorithm to explore unbiased data to increase the product candidate set, and at the same time introduces a popularity penalty in the edge construction process to eliminate bias.

[0007] However, the above methods only focus on how to adjust the weights of long-tail products, lack consideration of how popularity affects each specific interaction, and lack a systematic understanding of the popularity bias mechanism. Therefore, their actual performance is limited in practice. Therefore, there is still room for exploration and improvement in related technologies on how to eliminate the information cocoon in product recommendations and give more long-tail products the opportunity to be discovered. Summary of the Invention

[0008] The purpose of this application is to solve at least one of the above problems, and to provide a method for personalized product recommendation, as well as a corresponding device, computer device, computer-readable storage medium, and computer program product.

[0009] To achieve the various purposes of this application, the following technical solutions are adopted:

[0010] On the one hand, to achieve one of the purposes of this application, a method for personalized product recommendation is provided, including the following steps:

[0011] Obtain the personalized information of the user, where the personalized information includes the personal characteristic information of the user, the commodity preference characteristic information in the historical behavior data, and the spatio-temporal characteristic information describing the access environment of the user;

[0012] Use different neural network models pre-trained to the convergence state to respectively determine the popularity, including the first popularity feature corresponding to the personal characteristic information, the second popularity feature corresponding to the commodity characteristic information of the candidate commodity, and the third popularity feature corresponding to the comprehensive characteristic information composed of the personalized information and the commodity characteristic information of the candidate commodity;

[0013] Intervene in the calculation of the third popularity feature according to the first popularity feature and the second popularity feature to obtain the effective popularity of the candidate commodity;

[0014] Determine whether to push the candidate commodity to the user according to the effective popularity.

[0015] In some deepened embodiments, obtaining the personalized information of the user includes the following steps:

[0016] Obtain the personal information of the user, and extract the corresponding personal characteristic information therefrom, where the personal characteristic information includes any of the following: user ID, gender, age;

[0017] Obtain the historical behavior data of the user, extract the label statistical information of the commodities historically visited by the user therefrom, and construct it into the commodity preference characteristic information of the user;

[0018] Extract the time information and device information corresponding to the historical access behavior of the user from the historical behavior data, and construct it into the spatio-temporal characteristic information of the user.

[0019] In some extended embodiments, before the step of using different neural network models pre-trained to the convergence state to respectively determine the popularity, the following steps are included: Execute a training task to perform synchronous training on the respective neural network models, and share the same supervision label in each training.

[0020] In some deepened embodiments, executing a training task to perform synchronous training on the respective neural network models includes the following steps:

[0021] Call a single training sample from a preset data set, where each training sample includes the personalized information of a user and the commodity characteristic information of a candidate commodity in the commodity database, and the personalized information includes the personal characteristic information of the user, the commodity preference characteristic information in the historical behavior data, and the spatio-temporal characteristic information describing the access environment of the user;

[0022] Input the vectorized representation of the personal feature information in the training sample into the first neural network model for processing to obtain the first popularity feature;

[0023] Input the vectorized representation of the product feature information of the candidate product in the training sample into the second neural network model for processing to obtain the second popularity feature;

[0024] Input the personalized information in the training sample and the vectorized representation of the product feature information into the third neural network model for processing to obtain the third popularity feature;

[0025] After using a fusion layer to fuse the first popularity feature, the second popularity feature, and the third popularity feature to obtain the overall feature, calculate the overall popularity;

[0026] Use the supervision label corresponding to the training sample to jointly supervise the overall popularity, so as to realize the gradient update of each neural network model until each neural network model is trained to a convergence state. When the convergence state is not reached, continue to call the next training sample in the dataset to perform iterative training on each neural network model.

[0027] In a specific partial embodiment, using the supervision label corresponding to the training sample to jointly supervise the overall popularity to realize the gradient update of each neural network model includes the following steps:

[0028] According to the supervision label corresponding to the training sample, calculate the model loss values corresponding to the first popularity feature, the second popularity feature, and the overall feature respectively based on each neural network model;

[0029] Determine the model loss value corresponding to it according to the overall popularity obtained by classifying and mapping the overall feature;

[0030] Calculate whether the model loss value corresponding to the overall popularity reaches a preset threshold. When it reaches the preset threshold, it is determined that each neural network model reaches the convergence state and the training task is terminated. Otherwise, it is determined that each neural network model does not reach the convergence state, and gradient update and iterative training are respectively implemented on each neural network model according to the respective model loss values.

[0031] In a further partial embodiment, calculating the effective popularity of the candidate product by intervening the third popularity feature according to the first popularity feature and the second popularity feature includes the following steps:

[0032] Activate, fuse, and match weights for the first popularity feature and the second popularity feature respectively to obtain a deviation feature;

[0033] Subtract the deviation feature from the third popularity feature to obtain the overall feature;

[0034] Perform classification mapping according to the overall characteristics to obtain the effective popularity.

[0035] In a further partial embodiment, determining whether to push the candidate product to the user according to the effective popularity includes the following steps:

[0036] Obtain the effective popularity corresponding to each candidate product in the preset product database for the user, and construct the candidate product and its effective popularity into list data;

[0037] Screen each candidate product in the list data according to the effective popularity, and determine that some candidate products whose effective popularity meets the preset conditions are constructed into a product recommendation list;

[0038] Push the product recommendation list to the user's terminal device.

[0039] On the other hand, a product personalized recommendation device provided to meet one of the purposes of the present application includes an information acquisition module, a popularity acquisition module, a popularity optimization module, and a product optimization module, wherein: the information acquisition module is used to acquire the personalized information of the user, and the personalized information includes the personal characteristic information of the user, the product preference characteristic information in the historical behavior data, and the spatio-temporal characteristic information describing the user's access environment; the popularity acquisition module is used to respectively determine the popularity by using different neural network models pre-trained to the convergence state, including the first popularity feature corresponding to the personal characteristic information, the second popularity feature corresponding to the product characteristic information of the candidate product, and the third popularity feature corresponding to the comprehensive characteristic information composed of the personalized information and the product characteristic information of the candidate product; the popularity optimization module is used to intervene in the calculation of the third popularity feature according to the first popularity feature and the second popularity feature to obtain the effective popularity of the candidate product; the product optimization module is used to determine whether to push the candidate product to the user according to the effective popularity.

[0040] In a further partial embodiment, the information acquisition module includes: a personal information acquisition unit for acquiring the personal information of the user and extracting the corresponding personal characteristic information therefrom, and the personal characteristic information includes any of the following: user ID, gender, age; a preference information acquisition unit for acquiring the historical behavior data of the user and extracting the label statistical information of the products historically visited by the user to construct the product preference characteristic information of the user; a spatio-temporal information acquisition unit for extracting the time information and device information corresponding to the user's historical access behavior from the historical behavior data to construct the spatio-temporal characteristic information of the user.

[0041] In an extended partial embodiment, prior to the heat acquisition module, there is a model training module for performing a training task to synchronously train the respective neural network models and sharing the same supervision label in each training.

[0042] In a deepened partial embodiment, the model training module includes: a sample calling unit for calling a single training sample from a preset dataset, where each training sample includes the personalized information of a user and the product feature information of a candidate product in the product database, and the personalized information includes the personal feature information of the user, the product preference feature information in the historical behavior data, and the spatio-temporal feature information describing the access environment of the user; a first processing unit for inputting the vectorized representation of the personal feature information in the training sample into a first neural network model for processing to obtain a first heat feature; a second processing unit for inputting the vectorized representation of the product feature information of the candidate product in the training sample into a second neural network model for processing to obtain a second heat feature; a third processing unit for inputting the vectorized representation of the personalized information and the product feature information in the training sample into a third neural network model for processing to obtain a third heat feature; a fusion processing unit for fusing the first heat feature, the second heat feature, and the third heat feature by using a fusion layer to obtain an overall feature and then calculating the overall popularity heat; a joint supervision unit for jointly supervising the overall popularity heat by using the supervision label corresponding to the training sample to realize the gradient update of each neural network model until each neural network model is trained to a convergence state, and when the convergence state is not reached, continuing to call the next training sample in the dataset to perform iterative training on each neural network model.

[0043] In a specific partial embodiment, the joint supervision unit includes: a single loss calculation sub-unit for calculating the model loss values corresponding to the first heat feature, the second heat feature, and the overall feature respectively based on each neural network model according to the supervision label corresponding to the training sample; a summary loss calculation module for determining the model loss value corresponding to it according to the overall popularity heat obtained by classifying and mapping the overall feature; an iterative decision sub-unit for calculating whether the model loss value corresponding to the overall popularity heat reaches a preset threshold. When the preset threshold is reached, it is determined that each neural network model reaches the convergence state and the training task is terminated. Otherwise, it is determined that each neural network model does not reach the convergence state, and gradient update and iterative training are respectively implemented on each neural network model according to the respective model loss values.

[0044] In some of the further embodiments, the heat optimization module includes: a deviation calculation unit configured to respectively activate, fuse, and match weights for the first heat feature and the second heat feature to obtain a deviation feature; a feature fusion unit configured to subtract the deviation feature from the third heat feature to obtain an overall feature; and a heat determination unit configured to perform a classification mapping based on the overall feature to obtain an effective popular heat.

[0045] In some of the further embodiments, the product preference module includes: a heat annotation unit configured to obtain the effective popular heat corresponding to each candidate product in a preset product database for the user, and construct a list data with the candidate product and its effective popular heat; a product screening unit configured to screen each candidate product in the list data according to the effective popular heat, and determine a part of the candidate products whose effective popular heat meets a preset condition to construct a product recommendation list; and a product recommendation unit configured to push the product recommendation list to the user's terminal device.

[0046] In another aspect, a computer device provided to meet one of the purposes of the present application includes a central processing unit and a memory, and the central processing unit is configured to call and run a computer program stored in the memory to execute the steps of the product personalized recommendation method described in the present application.

[0047] In another aspect, a computer-readable storage medium provided to meet another purpose of the present application stores a computer program implemented according to the product personalized recommendation method in the form of computer-readable instructions, and when the computer program is called and run by a computer, it executes the steps included in the method.

[0048] In another aspect, a computer program product provided to meet another purpose of the present application includes a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the method described in any one of the embodiments of the present application are implemented.

[0049] Compared with the prior art, the present application has multiple advantages, at least including the following aspects:

[0050] First, the present application respectively uses different neural network models to respectively determine the first, second, and third popularity features according to the user's personal feature information, the product feature information of the candidate product, and the comprehensive feature information that combines the user's personalized information and the product feature information of the candidate product. The personalized information includes the personal feature information related to the user, the product preference feature information, and the spatio-temporal feature information. It can be seen from this that the first popularity feature is independent data predicted based on the user-side feature information, the second popularity feature is independent data predicted based on the feature information of the candidate product itself, and the third popularity feature is independent data predicted by integrating richer feature information on the user side and the candidate product side. There is a causal relationship between the third popularity feature and the first and second popularity features. The first and second popularity features are predictions of the popular popularity without considering other information, while the third popularity feature fully considers more comprehensive information to represent the popular popularity. Therefore, the first and second popularity features are used to filter the third popularity feature to achieve intervention and intervention in the third popularity feature, which is equivalent to correcting the hypothesis conditions of the third popularity feature to obtain the effective popular popularity, constructing a deviation correction mechanism for the popular popularity, so that the effective popular popularity can provide more representative prediction data for the ranking of candidate products, which is more accurate.

[0051] Secondly, the present application, through a systematic understanding of the popularity deviation mechanism, rather than blindly increasing the weight of long-tail products, quantifies the direct impact brought by the popularity deviation through the causal relationship between information. On the basis of the main model for calculating the third popularity feature, a user model for calculating the first popularity feature and a product model for calculating the second popularity feature are added, and are pre-trained to the convergence state through multi-task learning. Finally, when it takes effect online, through the correction mechanism, the deviation is subtracted from the popular popularity obtained by the main model, which well restores the actual interests of the user and can match more interesting products that conform to the user's personal preferences for the user.

[0052] In addition, under the action of the popularity deviation mechanism of the present application, more long-tail products have the opportunity to be exposed, which can not only achieve cold start, but also enable more products with sales potential to obtain recommendation opportunities, greatly reducing the Matthew effect of the product recommendation mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above and / or additional aspects and advantages of the present application will become obvious and easy to understand from the following description of the embodiments in conjunction with the drawings, in which:

[0054] Figure 1 is a schematic flowchart of a typical embodiment of the product personalized recommendation method of the present application;

[0055] Figure 2It is a principle block diagram of a network architecture for determining the effective popularity of the present application;

[0056] Figure 3 It is a flowchart showing the multi-task synchronous training process of the network architecture of the present application;

[0057] Figure 4 In the embodiment of the present application, it is a flowchart showing the process of calculating the loss values of each model to determine the iterative training of the decision network model;

[0058] Figure 5 In the embodiment of the present application, it is a flowchart showing the process of determining the effective popularity;

[0059] Figure 6 In the embodiment of the present application, it is a flowchart showing the process of exemplarily pushing a product recommendation list to a user;

[0060] Figure 7 It is a principle block diagram of the product personalized recommendation device of the present application;

[0061] Figure 8 It is a structural diagram of a computer device adopted by the present application. Detailed implementation manners

[0062] The embodiments of the present application are described in detail below. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals indicate the same or similar elements or elements with the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary and are only used to explain the present application, and cannot be construed as a limitation to the present application.

[0063] Those skilled in the art of the present technology can understand that unless specifically stated otherwise, the singular forms "a", "an", "the" and "said" used herein may also include the plural forms. It should be further understood that the term "including" used in the specification of the present application means the presence of the described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups. It should be understood that when we say that an element is "connected" or "coupled" to another element, it can be directly connected or coupled to other elements, or there may also be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The phrase "and / or" used herein includes all or any unit and all combinations of one or more related listed items.

[0064] Those skilled in the art can understand that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as the general understanding of those of ordinary skill in the art to which this application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0065] Those skilled in the art can understand that the "client", "terminal", and "terminal device" used herein include both devices with a wireless signal receiver that only has the ability to receive and no ability to transmit, and devices with receiving and transmitting hardware that can perform two-way communication on a two-way communication link. Such devices can include: cellular or other communication devices such as personal computers, tablets, etc., which have a single-line display or a multi-line display or a cellular or other communication device without a multi-line display; PCS (Personal Communications Service), which can combine voice, data processing, fax, and / or data communication capabilities; PDA (Personal Digital Assistant), which can include a radio frequency receiver, pager, Internet / intranet access, web browser, notepad, calendar, and / or GPS (Global Positioning System) receiver; conventional laptop and / or palm-top computers or other devices, which are conventional laptop and / or palm-top computers or other devices with and / or including a radio frequency receiver. The "client", "terminal", and "terminal device" used herein can be portable, transportable, installed in a vehicle (air, sea, and / or land), or suitable for and / or configured to operate locally and / or in a distributed manner at any other location on the earth and / or in space. The "client", "terminal", and "terminal device" used herein can also be a communication terminal, an Internet access terminal, a music / video playback terminal, for example, can be a PDA, MID (Mobile Internet Device), and / or a mobile phone with music / video playback function, or can also be devices such as smart TVs, set-top boxes, etc.

[0066] The hardware referred to by names such as "server", "client", and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer. It is a hardware device with the necessary components revealed by the von Neumann principle, including a central processing unit (including an arithmetic unit and a controller), a memory, an input device, and an output device. The computer program is stored in its memory, and the central processing unit loads the program stored in the external memory into the internal memory for execution, executes the instructions in the program, and interacts with the input / output devices to complete specific functions.

[0067] It should be noted that the concept of "server" in this application can similarly be extended to apply to the case of a server cluster. According to the network deployment principles understood by those skilled in the art, the various servers should be logically divided. Physically, these servers can either be independent of each other but can be invoked through an interface, or integrated into a single physical computer or a set of computer clusters. Those skilled in the art should understand this flexibility and should not use it to restrict the implementation manner of the network deployment method of this application.

[0068] One or several technical features of this application, unless expressly specified, can either be deployed on the server and accessed by the client remotely invoking the online service interface provided by the server, or directly deployed and run on the client for access.

[0069] The neural network models cited or possibly cited in this application, unless expressly specified, can either be deployed on a remote server and remotely invoked on the client, or deployed on a client with sufficient device capabilities for direct invocation. In some embodiments, when it runs on the client, its corresponding intelligence can be obtained through transfer learning to reduce the requirements for the client's hardware operating resources and avoid excessive occupation of the client's hardware operating resources.

[0070] All kinds of data involved in this application, unless expressly specified, can either be remotely stored on the server or stored on the local terminal device, as long as it is suitable for being invoked by the technical solution of this application.

[0071] Those skilled in the art should be aware that although the various methods of this application are described based on the same concept and thus show commonality with each other, unless otherwise specified, these methods can all be executed independently. Similarly, for the various embodiments disclosed in this application, they are all proposed based on the same inventive concept. Therefore, for concepts with the same expression, as well as concepts that are only appropriately transformed for convenience although the concept expressions are different, they should be equivalently understood.

[0072] For each embodiment to be disclosed in this application, unless explicitly stated to be mutually exclusive, the relevant technical features involved in each embodiment can be cross - combined to flexibly construct new embodiments, as long as such combination does not deviate from the inventive spirit of this application and can meet the requirements in the prior art or solve certain deficiencies in the prior art. Those skilled in the art should be aware of this flexibility.

[0073] A method for personalized recommendation of goods in this application can be programmed as a computer program product and deployed to run on a client or a server. For example, in the exemplary application scenario of this application, it can be deployed and implemented in the server of an e - commerce platform. Thus, by accessing the interface opened after the computer program product runs, human - computer interaction can be performed with the process of the computer program product through a graphical user interface to execute this method.

[0074] Please refer to Figure 1 , in the typical embodiment of the method for personalized recommendation of goods in this application, the following steps are included:

[0075] Step S1100: Obtain the personalized information of the user. The personalized information includes the personal characteristic information of the user, the preference characteristic information of goods in the historical behavior data, and the spatio - temporal characteristic information describing the access environment of the user.

[0076] In an exemplary application scenario, a consumer user submits a goods recommendation request to the server corresponding to an online store in an e - commerce platform through their terminal device. After receiving the request, the server uses the corresponding personalized information of the user to match the goods suitable for recommending to the user.

[0077] The personalized information is mainly the information with characteristic functions related to the user himself, including the personal information of the user, the preference information reflected by historical access behaviors, the spatio - temporal information reflected by historical access behaviors, etc. An exemplary process for collecting the personalized information may include the following steps:

[0078] Step S1110: Obtain the personal information of the user and extract the corresponding personal characteristic information from it. The personal characteristic information includes any of the following: user ID, gender, age.

[0079] To obtain the personal information of the user, it can be collected from their personal account registration information. For example, collect any one or more of their user ID, gender, age, points, identity level, etc. that play a role in representing personal characteristics. It is not difficult to understand that information such as the user's gender, age, and identity level may all imply information about the user's shopping ability. Therefore, these information can be flexibly selected and organized according to certain preset rules to form the corresponding personal characteristic information.

[0080] Step S1120: Obtain the historical behavior data of the user, extract the label statistical information of the products historically visited by the user therefrom, and construct it into the product preference feature information of this user:

[0081] In order to obtain the preference information of the user, the historical behavior data generated by the behavior events of the products visited by the user on the e-commerce platform historically can be statistically analyzed to obtain the label statistical information of the labels corresponding to the products visited by this user, so as to present the preference structure of this user, and based on this, the product preference feature information of this user is constructed.

[0082] Since the user's access behavior may act on each business link of the shopping business process of the e-commerce platform, therefore, the historical behavior data of different business links can be comprehensively used to obtain the label statistical information. For example, for different business links such as the business link where the user clicks to view the product details, the business link where the user clicks to purchase and create the bill corresponding to the product, the business link where the user adds the product to the shopping cart, and the business link where the user settles the bill and completes the payment for a certain product, etc., the labels corresponding to the products visited by this user can be respectively statistically analyzed under the corresponding business link. Thus, for each business link, multiple labels with a relatively high access ratio are collected accordingly to form the feature data corresponding to this business link. By extension, the label statistical information corresponding to this user in all business links is obtained. The so-called label generally refers to the attribute label obtained after data profiling of the product and is pre-labeled in the corresponding product information, so it can be directly statistically analyzed.

[0083] After obtaining the label statistical information of the user, organizing it according to the preset rules can obtain the corresponding product preference feature information.

[0084] Step S1130: Extract the time information and device information corresponding to the user's historical access behavior from the historical behavior data, and construct it into the spatio-temporal feature information of this user:

[0085] In the historical behavior data generated by the user adapting to each of its access events, usually, the type of the terminal device, device UID, etc. of the device relied on by the user when triggering the access event, the time information of the user triggering the access, the spatial geographical location information, etc. will also be recorded. These information can reflect the shopping habits of the user under the conditions of the device at a specific time and place. Accordingly, such information can be appropriately selected and organized according to the preset rules to obtain the corresponding spatio-temporal feature information.

[0086] The above exemplarily reveals the organization process of the personalized information of the user in this application. It is not difficult to understand that those skilled in the art can, according to the principles revealed here, flexibly select various basic data constituting the personalized information to construct the corresponding personalized information.

[0087] To meet the requirements of this application for working based on a neural network model, at the data level, the personal feature information, product preference feature information, and spatio-temporal feature information can be encoded into a vectorized representation form.

[0088] Step S1200: Use different neural network models pre-trained to a converged state to respectively determine the popularity, including a first popularity feature corresponding to the personal feature information, a second popularity feature corresponding to the product feature information of the candidate product, and a third popularity feature corresponding to the comprehensive feature information composed of the personalized information and the product feature information of the candidate product.

[0089] To recommend products in an online store to the user, the product feature information of the candidate products can be obtained one by one from the product database of the online store, and based on this product feature information and the user's personalized information, it is comprehensively determined whether the candidate product is suitable for recommendation to the user. The product feature information refers to the feature information with the ability to semantically describe the product, and can be obtained by previously using a mature feature extraction model to extract the deep semantic information from the text information such as the product title and product details of the corresponding candidate product, and / or from the image information such as the default product image and detail images. The product feature information extracted by the feature extraction model can be stored in the product database and directly called here.

[0090] On this basis, please refer to Figure 2 , this embodiment prepares three neural network models pre-trained to a converged state, which are used to determine the popularity features corresponding to the relative popularity represented under the corresponding feature information depending on different aspects of the feature information, specifically including:

[0091] The first neural network model is trained to be suitable for predicting a first popularity feature according to the personal feature information in the personalized information. For ease of understanding, the first neural network model can be regarded as a user model.

[0092] The second neural network model is trained to be suitable for predicting a second popularity feature according to the product feature information of the candidate product. For ease of understanding, the first neural network model can be regarded as a product model.

[0093] The third neural network model is trained to be suitable for predicting a third popularity feature according to the comprehensive feature information composed of the personalized information and the product feature information of the candidate product. For ease of understanding, the first neural network model can be regarded as a main model.

[0094] Each of the neural network models can be flexibly selected and implemented by those skilled in the art. For example, for the first and second neural network models, basic network models such as TextCNN and Bert can be selected. For the third neural network model, a model such as the DIN model can be used. Since the DIN model applies the multi-head attention mechanism, it can determine the weights of the user's commodity preference feature information based on the commodity feature information of the candidate commodity, and can extract key feature information more effectively to accurately calculate the third heat feature mentioned above.

[0095] As for the training process of each neural network model, it will be further disclosed in the subsequent embodiments of this application and will not be elaborated here for the time being.

[0096] Step S1300: Intervene in the calculation of the third heat feature based on the first heat feature and the second heat feature to obtain the effective popularity of the candidate commodity.

[0097] It can be understood that the third heat feature represents the popularity trend obtained when the candidate commodity is closely related to the user. The first heat feature represents the popularity trend obtained when the user is independent. The second heat feature represents the popularity trend when the candidate commodity is independent. Since the popularity trend feature data obtained by the user and the candidate commodity under their respective independent conditions are themselves influencing factors of the feature data obtained under the combined condition of the two, based on this causal relationship and applying counterfactual reasoning, the interference factors of the first heat feature and the second heat feature are filtered out from the third heat feature to realize the intervention in the third heat feature, and it is expected to obtain the effective popularity of the candidate commodity for the user.

[0098] Specifically, corresponding to Figure 2 As shown, the outputs of the first and second neural network models are connected to the fusion layer of the third neural network model for fusion. The first and second heat features can be activated respectively in advance, and each data therein is mapped to the same confidence interval. After fusion, the final overall feature is obtained, and then the third neural network model performs classification mapping to obtain the overall popularity. In the process of fusing the first heat feature and the second heat feature input into the fusion layer with the third heat feature, it is configured to subtract the deviation feature determined according to the first heat feature and the second heat feature from the third heat feature, so that the overall popularity becomes the effective popularity.

[0099] Step S1400: Determine whether to push the candidate commodity to the user according to the effective popularity.

[0100] After determining the effective popularity, the representativeness of the recommended value of the candidate product is realized. Thus, according to a preset threshold, it can be judged whether the effective popularity is greater than the preset threshold. When it is greater than the preset threshold, the candidate product is recommended to the user; otherwise, the candidate product is not pushed to the user.

[0101] Subsequent embodiments of this application will further disclose a more rigorous product recommendation process, which will not be elaborated here for the time being.

[0102] According to the embodiments disclosed herein, it can be seen that compared with the prior art, this application has multiple advantages, at least including the following aspects:

[0103] First, this application respectively uses different neural network models to respectively determine the first, second, and third heat characteristics according to the user's personal characteristic information, the product characteristic information of the candidate product, and the comprehensive characteristic information that combines the user's personalized information and the product characteristic information of the candidate product. The personalized information includes the personal characteristic information related to the user, the product preference characteristic information, and the spatio-temporal characteristic information. Accordingly, it can be known that the first heat characteristic is independent data predicted based on the user-side characteristic information, the second heat characteristic is independent data predicted based on the characteristic information of the candidate product itself, and the third heat characteristic is independent data predicted by integrating richer characteristic information on the user side and the candidate product side. There is a causal relationship between the third heat characteristic and the first and second heat characteristics. The first and second heat characteristics are predictions of the popularity without considering other information, while the third heat characteristic fully considers more comprehensive information to represent the popularity. Therefore, the first and second heat characteristics are used to filter the third heat characteristic to realize the intervention and intervention in the third heat characteristic, which is equivalent to correcting the hypothesis conditions of the third heat characteristic to obtain the effective popularity, constructing a deviation correction mechanism for the popularity, so that the effective popularity can provide more representative prediction data for the ranking of the candidate product and is more accurate.

[0104] Secondly, this application, through a systematic understanding of the popularity deviation mechanism, rather than blindly increasing the weight of long-tail products, quantifies the direct impact brought by the popularity deviation through the causal relationship between information. On the basis of the main model for calculating the third heat characteristic, a user model for calculating the first heat characteristic and a product model for calculating the second heat characteristic are added, and they are pre-trained to the convergence state through multi-task learning. Finally, when it takes effect online, through the correction mechanism, the deviation is subtracted from the popularity obtained by the main model, which well restores the actual interests of the user and can match more interesting products that conform to the user's personal preferences for the user.

[0105] In addition, under the action of the popularity deviation mechanism of the present application, more long-tail products have the opportunity to be exposed, which can not only achieve cold start, but also enable more products with sales potential to obtain recommendation opportunities, greatly reducing the Matthew effect of the product recommendation mechanism.

[0106] In some extended embodiments, before the steps S1200 of respectively determining the popularity by using different neural network models pre-trained to a convergent state, the following steps are included:

[0107] Step S2000: Execute a training task to perform synchronous training on each of the neural network models, and share the same supervision label in each training:

[0108] As Figure 2 shown in the network architecture, each of the neural network models therein can be applied to multi-task synchronous training, thereby improving the training efficiency. In the multi-task training process, the same set of training samples is used. After obtaining the input parameters required by each neural network model from it and calculating their respective heat characteristics, the model loss values corresponding to each heat characteristic are respectively calculated according to the same supervision label corresponding to this set of training samples. Taking the linear sum result of each model loss value reaching a preset threshold as the goal, the entire network architecture is iteratively trained, so as to realize the training of the entire network architecture with the shared same supervision label.

[0109] According to the embodiments herein, it can be known that when each neural network model of the present application is subjected to multi-task synchronous training, the heat characteristics obtained by each model are all subject to the same supervision label, and thus are correlated. Finally, there will also be a causal relationship between the prediction capabilities of each neural network model. Therefore, it provides a factual basis for the application of counterfactual reasoning, and can ensure the determination of effective popularity according to the mutual constraints between each heat characteristic.

[0110] Please refer to Figure 3 , in some further embodiments, step S2000 of executing a training task to perform synchronous training on each of the neural network models includes the following steps:

[0111] Step S2100: Call a single training sample from a preset dataset. Each training sample includes the personalized information of a user and the product feature information of a candidate product in the product database. The personalized information includes the personal feature information of the user, the product preference feature information in the historical behavior data, and the spatio-temporal feature information describing the access environment of the user:

[0112] For the purpose of executing Figure 2Synchronous training of the network architecture shown. Prepare a data set, which contains a sufficient number of training samples required to train the network architecture to a convergent state. The training samples are collected in advance and manually labeled, for example, labeled with two labels: popular or unpopular, thus constituting positive samples and negative samples.

[0113] For the training samples, data is collected to adapt to the input parameters required by each neural network model in the network architecture disclosed in this application. Therefore, each training sample includes the personalized information of a user and the product feature information of a candidate product in the product database of the online store. Similarly, the personalized information includes the personal feature information of the user, the product preference feature information in the historical behavior data, and the spatio-temporal feature information describing the user's access environment.

[0114] Those skilled in the art should be aware that there is a correspondence between the input parameters used in the training stage and the inference stage of the neural network model. Therefore, the encoding rules applied to the input parameters in the training stage also apply to the inference stage, and vice versa. Accordingly, when those skilled in the art implement training, they only need to collect the information of the training samples by applying the corresponding rules to adapt to the rule requirements implemented in the model inference stage.

[0115] Step S2200: Input the vectorized representation of the personal feature information in the training sample into the first neural network model for processing to obtain the first heat feature.

[0116] According to the embodiments of the application of the model inference stage described above in this application, for the first neural network model of the network architecture, the personal feature information in the personalized information of a single training sample called for training can be input into the first neural network model in the form of a vectorized representation for feature extraction and then fully connected to obtain its corresponding first heat feature.

[0117] Step S2300: Input the vectorized representation of the product feature information of the candidate product in the training sample into the second neural network model for processing to obtain the second heat feature.

[0118] According to the embodiments of the application of the model inference stage described above in this application, for the second neural network model of the network architecture, the product feature information of the candidate product in a single training sample called for training can be input into the second neural network model in the form of a vectorized representation for feature extraction and then fully connected to obtain its corresponding second heat feature.

[0119] Step S2400: Input the vectorized representation of the personalized information and the product feature information in the training sample into the third neural network model for processing to obtain the third heat feature.

[0120] Similarly, according to the embodiments of the application regarding the application in the model inference stage described above, for the third neural network model of the network architecture, the personalized information and the product feature information of the candidate product in a single training sample called for training can be integrated to obtain integrated feature information represented in a vector form, and the model can perform a fully connected operation based on the integrated feature information to obtain its third popularity feature.

[0121] Among them, the integrated feature information of the application can be generated in the following manner: The product preference feature information of the user and the product feature information of the candidate product are input into the preset multi-head attention layer of the third neural network model for in-depth feature interaction, so as to highlight the internal features in the product preference feature information that have a high correlation with the product feature information of the candidate product. Then, these internal features are concatenated with the personal feature information, spatio-temporal feature information in the training sample, and the product feature information of the candidate product in multiple channels and expanded into high-dimensional vectors, thereby obtaining the integrated feature information. Subsequently, the third neural network model can perform a fully connected operation based on this integrated feature information to obtain the corresponding third popularity feature.

[0122] Step S2500: After using a fusion layer to fuse the first popularity feature, the second popularity feature, and the third popularity feature to obtain an overall feature, calculate the overall popularity:

[0123] To facilitate the joint supervision of the three neural network models, activation layers can be respectively used to activate and output the first popularity feature and the second popularity feature, and then the two are concatenated into a deviation feature through a linear layer associated with learnable weights. Then, through a linear layer, calculate the overall feature after subtracting the deviation feature from the third popularity feature, and finally perform a classification mapping based on the overall feature to obtain the overall popularity corresponding to the classification mapped to represent popularity.

[0124] Step S2600: Use the supervision label corresponding to the training sample to perform joint supervision on the overall popularity, so as to realize the gradient update of each neural network model until each neural network model is trained to a convergence state. When the convergence state is not reached, continue to call the next training sample in the dataset to perform iterative training on each neural network model.

[0125] It is not difficult to understand from Step S2200 to Step S2500 that, essentially, each neural network model can obtain the corresponding popularity through its corresponding popularity feature, and by means of the popularity features of the three neural network models, the overall feature can be comprehensively determined and the overall popularity can be obtained based on the overall feature. Therefore, through the supervised training of the overall popularity, the supervised training of the three neural network models can be realized.

[0126] Based on this idea, the model loss values of each neural network model can be calculated according to the supervision labels specified by the training samples. Among them, the model loss value of the third neural network model can be obtained by calculating its corresponding model loss value based on the overall feature. Furthermore, by linearly aggregating the model loss value of the first neural network model, the model loss value of the second neural network model, and the model loss value corresponding to the overall feature, the model loss value corresponding to the entire network architecture can be obtained, which is the corresponding model loss value obtained by classifying according to the overall feature. This model loss value can be used to determine the convergence state of the three neural network models and can be used for backpropagation to the third neural network model, while the first neural network model and the second neural network model can also perform backpropagation according to their own model loss values. That is, when the model loss value corresponding to the overall popularity obtained after the classification of the third neural network model reaches the preset threshold for determining the convergence of the entire network architecture, it can be regarded that each neural network has been trained to the convergence state. Otherwise, continue to train each neural network model and perform gradient update on the three neural network models during each training process.

[0127] It is not difficult to understand that in this case, when the model loss value corresponding to the overall popularity does not reach the preset threshold, it means that the entire network architecture has not reached the convergence state. Therefore, the model loss value can be used for backpropagation to each neural network model to update the weight parameters of each link by gradient, and then continue to call the next training sample from the dataset to continue the iterative training of the entire network architecture, and so on, until the model loss value corresponding to the overall popularity reaches the preset threshold, at which point it is regarded that each neural network model in the entire network architecture has been trained to the convergence state, and thus the multi-task synchronous training can be terminated.

[0128] According to the embodiments disclosed herein, it can be understood that by performing multi-task synchronous training on the entire network architecture including three neural network models, it not only ensures the association of the three neural network models in maintaining representation learning and reasoning capabilities, but also provides a reliable data basis for applying counterfactual reasoning to determine the effective popularity according to the causal relationship between data. Thus, a practical and effective model for product recommendation is provided.

[0129] Please refer to Figure 4 , in specific partial embodiments, the step S2600 of using the supervision label corresponding to the training sample to jointly supervise the overall popularity to achieve gradient update of each neural network model includes the following steps:

[0130] Step S2610: According to the supervision label corresponding to the training sample, calculate the model loss values corresponding to the first popularity feature, the second popularity feature, and the overall feature respectively based on each neural network model:

[0131] Please refer to Figure 2 After a training sample is trained, the corresponding first heat feature, second heat feature, and third heat feature can be obtained from the classification of the representation popularity of each neural network model. Accordingly, the model loss value of each neural network model can be calculated based on its corresponding supervision label.

[0132] Since the third neural network model determines its overall popularity by classifying and mapping the overall feature obtained by interfering with its third heat feature with the deviation feature composed of the first heat feature and the second heat feature, the model loss value corresponding to the overall feature can be determined as the model loss value corresponding to the third neural network model itself.

[0133] Based on this principle, the following formula can be applied to implement the fusion layer. After activating the first heat feature of the first neural network model and the second heat feature of the second neural network model and fusing them with the third heat feature of the third neural network model the overall feature is obtained:

[0134]

[0135] Thus, the outputs of the three models are unified through the overall feature Therefore, using the supervision label corresponding to the training sample, unified gradient update control of the three neural network models can be achieved.

[0136] To calculate the model loss value corresponding to each neural network model, the following formula can be applied respectively for each neural network model:

[0137]

[0138]

[0139]

[0140] where represents the model loss value of the overall feature obtained by the third neural network model, represents the model loss value corresponding to the second heat feature of the second neural network model, represents the model loss value corresponding to the first heat feature of the first neural network model; represents user-side feature information, represents candidate product-side feature information, represents the application of the activation function, Represents a data set, represents a supervision label, represents the actual output value of the model.

[0141] Step S2620: Determine the corresponding model loss value according to the overall popularity obtained by the overall feature classification mapping:

[0142] After obtaining each model loss value, the model loss value of the entire network architecture can be calculated, that is, the model loss value corresponding to the classification result finally obtained by the classification mapping of the third neural network model , and the exemplary formula is as follows:

[0143]

[0144] Wherein, , are preset weights, which can be flexibly configured by those skilled in the art or obtained through training and learning.

[0145] Step S2630: Calculate whether the model loss value corresponding to the overall popularity reaches a preset threshold. When the preset threshold is reached, it is determined that each neural network model reaches a convergence state and the training task is terminated. Otherwise, it is determined that each neural network model does not reach a convergence state, and gradient update and iterative training are performed on each neural network model according to the respective model loss values:

[0146] So far, the model loss value corresponding to the output after the overall feature is classified and mapped is determined. Just by comparing this model loss value with the preset threshold for judging whether the model converges, it can be determined whether the entire network architecture reaches a convergence state. When the entire network architecture has not converged, continue to call the next training sample to perform iterative training on the network architecture, and at the same time perform backpropagation on the entire network architecture according to the model loss value to achieve gradient update. Otherwise, when the entire network architecture has converged, the training can be terminated.

[0147] According to the embodiments herein, since each neural network model of the entire network architecture calculates the model loss value under the same supervision label using the same set of training samples during the training process to realize the gradient update of the model, the prediction capabilities learned by each neural network model are correlated and can subsequently be used for the calculation of effective popularity to obtain more accurate prediction data.

[0148] Please refer to Figure 5 , in some deepened embodiments, the step S1300: Intervene the third heat feature according to the first heat feature and the second heat feature to calculate the effective popularity of the candidate commodity, includes the following steps:

[0149] Step S1310: After activating the first heat feature and the second heat feature respectively, fuse them and match weights to obtain a deviation feature:

[0150] According to the counterfactual reasoning logic, it is not difficult to understand that the first heat feature obtained by the first neural network model and the second heat feature obtained by the second neural network model are actually a deviation data for the third heat feature of the third neural network model. Therefore, in the model inference stage, the corresponding deviation popularity of the first heat feature and the second heat feature can be determined by matching preset weights respectively.

[0151] Step S1320: Subtract the deviation feature from the third heat feature to obtain an overall feature:

[0152] In order to correct the third heat feature and obtain an overall feature that takes into account the causal relationship and performs counterfactual reasoning, directly subtract the deviation popularity from the third heat feature, that is, obtain the overall feature. An exemplary formula is as follows:

[0153]

[0154] According to the latter part of this formula, the outputs of the first neural network model and the second neural network model are respectively passed through the activation layer and fused, and then configured with preset weights. These weights can be flexibly set by those skilled in the art according to actual measurements or experience, and then fused with the overall feature originally obtained by the third neural network model, that is, the latter subtracts the former, so as to obtain a new overall feature for implementing classification mapping.

[0155] According to this formula, different from the entire network architecture in the training stage, in the inference stage here, the outputs of the first neural network model and the second neural network model are used to intervene in the output of the third neural network model to obtain an overall feature. Thus, the overall effect obtained by the third neural network model, minus the total direct effect of the first neural network model and the second neural network model, the final effect obtained will be the actual effect after eliminating the popularity deviation, and it is expected to obtain a more effective popularity.

[0156] Step S1330: Perform classification mapping according to the overall feature to obtain an effective popularity:

[0157] Finally, perform classification mapping according to the overall feature, and the confidence level mapped to the classification representing popularity can be obtained, which can be used as the effective popularity corresponding to the candidate product.

[0158] As can be seen from the embodiments herein, the effective popularity is obtained after correcting the third popularity feature using the first popularity feature and the second popularity feature, filtering out some interfering factors, which can better represent the popularity trend of candidate products relative to a specific user. Therefore, more long-tail products also have the opportunity to be recommended and exposed.

[0159] Please refer to Figure 6 , in the in-depth partial embodiments, the step S1400 of determining whether to push the candidate product to the user according to the effective popularity includes the following steps:

[0160] Step S1410: Obtain the effective popularity corresponding to each candidate product in the preset product database for the user, and construct the candidate products and their effective popularity into list data:

[0161] In this embodiment, when multiple products need to be recommended to a user, for each candidate product in the product database, according to the process from step S1100 to step S1300, the effective popularity corresponding to each candidate product can be obtained one by one. On this basis, the mapping relationship data between each candidate product in the product database and its effective popularity can be obtained, and it can be constructed into list data.

[0162] Step S1420: Screen each candidate product in the list data according to the effective popularity, and determine some candidate products whose effective popularity meets the preset conditions to construct a product recommendation list:

[0163] In order to select some candidate products that better meet the requirements, the candidate products in the list data can be sorted in reverse according to the effective popularity, and then the first N in the reverse order can be used as the preset condition, where N can be flexibly preset, and the first several candidate products can be obtained to construct a product recommendation list.

[0164] Step S1430: Push the product recommendation list to the user's terminal device:

[0165] Finally, push the product recommendation list to the user's terminal device for display to complete the response to the user's product recommendation request.

[0166] As can be known from the embodiments herein, each embodiment of the present application can serve the product recommendation service, match products with a popularity trend for users, enable long-tail products with potential sales value to have the opportunity to be exposed, and make the recommended products better match the personalized needs of users.

[0167] Please refer to Figure 7, a personalized product recommendation device provided to meet one of the purposes of this application, is a functional embodiment of the personalized product recommendation method of this application. The device includes an information acquisition module 1100, a popularity acquisition module 1200, a popularity optimization module 1300, and a product selection module 1400, where: The information acquisition module 1100 is used to acquire the personalized information of the user, and the personalized information includes the personal characteristic information of the user, the product preference characteristic information in the historical behavior data, and the spatio-temporal characteristic information describing the access environment of the user; The popularity acquisition module 1200 is used to respectively determine the popularity by using different neural network models pre-trained to a converged state, including a first popularity characteristic corresponding to the personal characteristic information, a second popularity characteristic corresponding to the product characteristic information of the candidate product, and a third popularity characteristic corresponding to the comprehensive characteristic information composed of the personalized information and the product characteristic information of the candidate product; The popularity optimization module 1300 is used to intervene in the calculation of the third popularity characteristic according to the first popularity characteristic and the second popularity characteristic to obtain the effective popularity of the candidate product; The product selection module 1400 is used to determine whether to push the candidate product to the user according to the effective popularity.

[0168] In some deepened embodiments, the information acquisition module 1100 includes: a personal information acquisition unit, which is used to acquire the personal information of the user and extract the corresponding personal characteristic information therefrom. The personal characteristic information includes any of the following: user ID, gender, age; a preference information acquisition unit, which is used to acquire the historical behavior data of the user and extract the label statistical information of the products historically visited by the user, and construct it as the product preference characteristic information of the user; a spatio-temporal information acquisition unit, which is used to extract the time information and device information corresponding to the historical access behavior of the user from the historical behavior data, and construct it as the spatio-temporal characteristic information of the user.

[0169] In some extended embodiments, prior to the popularity acquisition module 1200, there is a model training module, which is used to execute training tasks to implement the synchronous training of the respective neural network models and share the same supervision label in each training.

[0170] In some embodiments of further refinement, the model training module includes: a sample calling unit configured to call a single training sample from a preset dataset, where each training sample includes personalized information of a user and product feature information of a candidate product in a product database. The personalized information includes personal feature information of the user, product preference feature information in historical behavior data, and spatio-temporal feature information describing the user's access environment; a first processing unit configured to input a vectorized representation of the personal feature information in the training sample into a first neural network model for processing to obtain a first popularity feature; a second processing unit configured to input a vectorized representation of the product feature information of the candidate product in the training sample into a second neural network model for processing to obtain a second popularity feature; a third processing unit configured to input a vectorized representation of the personalized information and the product feature information in the training sample into a third neural network model for processing to obtain a third popularity feature; a fusion processing unit configured to use a fusion layer to fuse the first popularity feature, the second popularity feature, and the third popularity feature to obtain an overall feature, and then calculate an overall popularity; a joint supervision unit configured to perform joint supervision on the overall popularity using a supervision label corresponding to the training sample to update the gradients of each neural network model until each neural network model is trained to a convergence state. When the convergence state is not reached, the next training sample in the dataset is continuously called to perform iterative training on each neural network model.

[0171] In some embodiments of further specification, the joint supervision unit includes: a single loss calculation sub-unit configured to calculate model loss values corresponding to the first popularity feature, the second popularity feature, and the overall feature respectively based on each neural network model according to the supervision label corresponding to the training sample; a summary loss calculation module configured to determine a model loss value corresponding to it according to the overall popularity obtained by classifying and mapping the overall feature; an iterative decision sub-unit configured to calculate whether the model loss value corresponding to the overall popularity reaches a preset threshold. When the preset threshold is reached, it is determined that each neural network model reaches a convergence state and the training task is terminated. Otherwise, it is determined that each neural network model does not reach a convergence state, and gradient update and iterative training are respectively performed on each neural network model according to the respective model loss values.

[0172] In some embodiments of further refinement, the popularity optimization module 1300 includes: a deviation calculation unit configured to activate, fuse, and match weights of the first popularity feature and the second popularity feature respectively to obtain a deviation feature; a feature fusion unit configured to subtract the deviation feature from the third popularity feature to obtain an overall feature; a popularity determination unit configured to perform classification mapping according to the overall feature to obtain an effective popularity.

[0173] In a deepened partial embodiment, the commodity preference module 1400 preferably includes: a popularity marking unit configured to obtain the effective popularity of each candidate commodity in a preset commodity database corresponding to the user, and construct a list data with the candidate commodities and their effective popularity; a commodity screening unit configured to screen each candidate commodity in the list data according to the effective popularity, and determine some candidate commodities whose effective popularity meets a preset condition to construct a commodity recommendation list; and a commodity recommendation unit configured to push the commodity recommendation list to the terminal device of the user.

[0174] To solve the above technical problems, an embodiment of the present application further provides a computer device. As Figure 8 shown, it is a schematic internal structure diagram of the computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected through a system bus. Among them, the computer-readable storage medium of the computer device stores an operating system, a database, and computer-readable instructions. Control information sequences can be stored in the database. When the computer-readable instructions are executed by the processor, the processor can implement a commodity personalized recommendation method. The processor of the computer device is used to provide computing and control capabilities to support the operation of the entire computer device. Computer-readable instructions can be stored in the memory of the computer device. When the computer-readable instructions are executed by the processor, the processor can execute the commodity personalized recommendation method of the present application. The network interface of the computer device is used to communicate with the terminal. Those skilled in the art can understand that Figure 8 the structure shown is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0175] In this embodiment, the processor is used to execute Figure 7 the specific functions of each module and its sub-modules. The memory stores the program codes and various types of data required to execute the above modules or sub-modules. The network interface is used for data transmission between the user terminal and the server. The memory in this embodiment stores the program codes and data required to execute all modules / sub-modules in the commodity personalized recommendation device of the present application. The server can call the program codes and data of the server to execute the functions of all sub-modules.

[0176] The present application further provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, one or more processors are caused to execute the steps of the commodity personalized recommendation method according to any embodiment of the present application.

[0177] The present application also provides a computer program product, including computer programs / instructions which, when executed by one or more processors, implement the steps of the method according to any embodiment of the present application.

[0178] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments of the present application can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the aforementioned storage medium can be a computer-readable storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0179] In summary, the present application can better restore the popularity of products in the product database, match more interesting products that better meet the personal preferences of users, enable more potential long-tail products to have the opportunity to be exposed, and reduce the Matthew effect in the process of product recommendation.

[0180] Those skilled in the art of the present technology can understand that the steps, measures, and solutions in the various operations, methods, and processes discussed in the present application can be alternated, changed, combined, or deleted. Further, the other steps, measures, and solutions in the various operations, methods, and processes discussed in the present application can also be alternated, changed, rearranged, decomposed, combined, or deleted. Further, the steps, measures, and solutions in the prior art that are the same as those disclosed in the various operations, methods, and processes in the present application can also be alternated, changed, rearranged, decomposed, combined, or deleted.

[0181] The above are only some embodiments of the present application. It should be noted that for those of ordinary skill in the art of the present technology, without departing from the principle of the present application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. A method for personalized recommendation of commodities, characterized in that, It includes the following steps: Obtain the personalized information of the user, where the personalized information includes the personal characteristic information of the user, the commodity preference characteristic information in the historical behavior data, and the spatio-temporal characteristic information describing the access environment of the user; Use different neural network models pre-trained to a converged state to respectively determine the popularity, including the first popularity feature corresponding to the personal characteristic information, the second popularity feature corresponding to the commodity characteristic information of the candidate commodity, and the third popularity feature corresponding to the comprehensive characteristic information composed of the personalized information and the commodity characteristic information of the candidate commodity; Intervene in the calculation of the third popularity feature according to the first popularity feature and the second popularity feature to obtain the effective popularity of the candidate commodity, including: respectively activating and fusing the first popularity feature and the second popularity feature and matching weights to obtain a deviation feature; subtracting the deviation feature from the third popularity feature to obtain an overall feature; performing classification mapping according to the overall feature to obtain the effective popularity; Determine whether to push the candidate commodity to the user according to the effective popularity; Before the step of using different neural network models pre-trained to a converged state to respectively determine the popularity, perform a training task to implement synchronous training of the respective neural network models, including the following steps: Call a single training sample from a preset dataset, where each training sample includes the personalized information of a user and the commodity characteristic information of a candidate commodity in the commodity database, and the personalized information includes the personal characteristic information of the user, the commodity preference characteristic information in the historical behavior data, and the spatio-temporal characteristic information describing the access environment of the user; Input the vectorized representation of the personal characteristic information in the training sample into the first neural network model for processing to obtain the first popularity feature; Input the vectorized representation of the commodity characteristic information of the candidate commodity in the training sample into the second neural network model for processing to obtain the second popularity feature; Input the vectorized representation of the personalized information and the commodity characteristic information in the training sample into the third neural network model for processing to obtain the third popularity feature; After using a fusion layer to fuse the first popularity feature, the second popularity feature, and the third popularity feature to obtain an overall feature, calculate the overall popularity; Use the supervision label corresponding to the training sample to perform joint supervision on the overall popularity to realize the gradient update of each neural network model until each neural network model is trained to a converged state, and when the converged state is not reached, continue to call the next training sample in the dataset to perform iterative training on each neural network model.

2. The personalized recommendation method for commodities according to claim 1, wherein Obtain the personalized information of the user, including the following steps: Obtain the personal information of the user and extract the corresponding personal characteristic information therefrom, where the personal characteristic information includes any of the following: user ID, gender, age; Obtain the historical behavior data of the user and extract the label statistical information of the commodities historically visited by the user, and construct it into the commodity preference characteristic information of the user; Extract the time information and device information corresponding to the user's historical access behavior from the historical behavior data, and construct them into the spatio-temporal feature information of the user.

3. The personalized recommendation method for commodities according to claim 1, wherein Use the supervision label corresponding to the training sample to jointly supervise the overall popularity, so as to realize the gradient update of each neural network model, including the following steps: According to the supervision label corresponding to the training sample, calculate the model loss values corresponding to the first heat feature, the second heat feature, and the overall feature based on each neural network model respectively; Determine the corresponding model loss value according to the overall popularity obtained by the classification mapping of the overall feature; Calculate whether the model loss value corresponding to the overall popularity reaches a preset threshold. When it reaches the preset threshold, it is determined that each neural network model reaches the convergence state and the training task is terminated. Otherwise, it is determined that each neural network model does not reach the convergence state, and gradient update and iterative training are performed on each neural network model according to the respective model loss values.

4. The personalized recommendation method for goods according to any one of claims 1 to 3, characterized in that Determine whether to push the candidate product to the user according to the effective popularity, including the following steps: Obtain the effective popularity corresponding to each candidate product in the preset product database and the user, and construct the candidate product and its effective popularity into list data; Screen each candidate product in the list data according to the effective popularity, and determine some candidate products whose effective popularity meets the preset conditions to construct a product recommendation list; Push the product recommendation list to the user's terminal device.

5. A commodity personalized recommendation device, characterized in that, Including: An information acquisition module, configured to acquire the personalized information of the user, where the personalized information includes the personal feature information of the user, the product preference feature information in the historical behavior data, and the spatio-temporal feature information describing the user's access environment; A heat acquisition module, configured to use different neural network models pre-trained to the convergence state to respectively determine the popularity, including the first heat feature corresponding to the personal feature information, the second heat feature corresponding to the product feature information of the candidate product, and the third heat feature corresponding to the comprehensive feature information composed of the personalized information and the product feature information of the candidate product; A heat optimization module, configured to intervene in the third heat feature according to the first heat feature and the second heat feature to calculate the effective popularity of the candidate product, including: respectively activating and fusing the first heat feature and the second heat feature and matching weights to obtain a deviation feature; subtracting the deviation feature from the third heat feature to obtain an overall feature; performing classification mapping according to the overall feature to obtain the effective popularity; A product optimization module, configured to determine whether to push the candidate product to the user according to the effective popularity; Prior to the heat acquisition module, there is a model training module, configured to execute a training task to perform synchronous training on each neural network model, and share the same supervision label in each training; the model training module includes: A sample calling unit, configured to call a single training sample from a preset dataset. Each training sample includes the personalized information of a user and the product feature information of a candidate product in a product database. The personalized information includes the personal feature information of the user, the product preference feature information in historical behavior data, and the spatio-temporal feature information describing the access environment of the user. A first processing unit, configured to input the vectorized representation of the personal feature information in the training sample into a first neural network model for processing to obtain a first popularity feature. A second processing unit, configured to input the vectorized representation of the product feature information of the candidate product in the training sample into a second neural network model for processing to obtain a second popularity feature. A third processing unit, configured to input the personalized information in the training sample and the vectorized representation of the product feature information into a third neural network model for processing to obtain a third popularity feature. A fusion processing unit, configured to use a fusion layer to fuse the first popularity feature, the second popularity feature, and the third popularity feature to obtain an overall feature, and then calculate an overall popularity. A joint supervision unit, configured to perform joint supervision on the overall popularity using the supervision label corresponding to the training sample to implement gradient update of each neural network model until each neural network model is trained to a convergence state. When the convergence state is not reached, the next training sample in the dataset is continuously called to perform iterative training on each neural network model.

6. The personalized recommendation device for goods according to claim 5, characterized in that, The information acquisition module includes: A personal information acquisition unit, configured to acquire the personal information of the user and extract the corresponding personal feature information therefrom. The personal feature information includes any of the following: user ID, gender, age. A preference information acquisition unit, configured to acquire the historical behavior data of the user and extract the label statistical information of the products historically accessed by the user therefrom, and construct the product preference feature information of the user. A spatio-temporal information acquisition unit, configured to extract the time information and device information corresponding to the historical access behavior of the user from the historical behavior data and construct the spatio-temporal feature information of the user.

7. The personalized recommendation device for goods according to claim 5, wherein The joint supervision unit includes: A single loss calculation sub-unit, configured to calculate the model loss values corresponding to the first popularity feature, the second popularity feature, and the overall feature respectively based on each neural network model according to the supervision label corresponding to the training sample. An aggregated loss calculation module, configured to determine the model loss value corresponding thereto according to the overall popularity obtained by classifying and mapping the overall feature. An iterative decision sub-unit, configured to calculate whether the model loss value corresponding to the overall popularity reaches a preset threshold. When the preset threshold is reached, it is determined that each neural network model reaches a convergence state and the training task is terminated. Otherwise, it is determined that each neural network model does not reach a convergence state, and gradient update and iterative training are respectively performed on each neural network model according to the respective model loss values.

8. The personalized recommendation device for goods according to any one of claims 5 to 7, characterized in that The product optimization module includes: A heat annotation unit, configured to obtain the effective popularity of each candidate product in a preset product database corresponding to the user, and construct a list data with the candidate products and their effective popularity; A product screening unit, configured to screen each candidate product in the list data according to the effective popularity, and determine a part of the candidate products whose effective popularity meets the preset conditions to construct a product recommendation list; A product recommendation unit, configured to push the product recommendation list to the user's terminal device.

9. A computer device, comprising a central processing unit and a memory, characterized in that, The central processing unit is configured to call and run a computer program stored in the memory to execute the steps of the method according to any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, It stores a computer program implemented according to the method according to any one of claims 1 to 4 in the form of computer-readable instructions. When the computer program is called and run by a computer, it executes the steps included in the corresponding method.

Images