Recommended model training method, recommended method, device and storage medium

Abstract

The application provides a recommendation model training method, a recommendation method, equipment and a storage medium. The method comprises the following steps: obtaining a first recommendation model and a deep cross-scene feature network unit; constructing a second recommendation model based on the first recommendation model and the deep cross-scene feature network unit, wherein the deep cross-scene feature network unit is configured to cross and combine non-scene features and scene features; collecting a first sample data set, wherein the data of each sample in the first sample data set comprises scene features, non-scene features and corresponding recommendation result labels; and training the second recommendation model based on the first sample data set. The application develops a deep network unit of deep cross-scene features, which can be flexibly embedded into a recall and / or sorting deep model of a recommendation system, so as to achieve the purpose of sharing multiple scene data and better matching the distribution of scene data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a recommendation model training method, a recommendation method, a device and a storage medium. BACKGROUND

[0002] The recommendation system is widely used in various APPs to enhance the personalized experience of users and increase the revenue of suppliers. The recall and ranking algorithms, as two important modules of the recommendation system, directly affect the final recommendation result and revenue.

[0003] The algorithm of the recommendation system completely depends on user behavior data, and the data is the upper limit of the effect, and the model only approximates the upper limit as much as possible. However, the recommendation system has always had the problem of data sparsity. For the same recommendation candidate set, the data distribution of each scene is not completely consistent, resulting in that the data used for the model of each scene is self-sufficient, and the data is more sparse. For example, there are multiple scene-related data of scenic spots in an online travel agency APP, such as hotel post-scenic spot recommendation, list page scenic spot ranking, etc. The characteristic is to use the same scenic spot candidate set, but the user behavior data distribution of each scene is completely different. The user in the hotel post-scenic spot recommendation scene may focus on the information related to the hotel, and may be a hotel for some scenic spots, and at the same time, because it is a hotel post-scene, the data will be relatively sparse. The user in the list page scenic spot ranking scene is more likely to focus on the scenic spots in the current city, or just want to find some scenic spots for daily friends gathering, such as KTV, table game bar, etc., and at the same time, the data in this scene is relatively more. Generally, there are three kinds of conventional solutions: 1) model according to the data of each scene to obtain multiple models respectively used for online reasoning. However, this method has obvious shortcomings, one is that multiple models waste resources and are not easy to maintain, and the other is that the data of each scene is sparse, resulting in model underfitting and poor effect; 2) in order to solve the shortcomings of solution 1, the data of multiple scenes is mixed together for modeling, and finally one model is used for online reasoning for the same candidate set. The disadvantage of this solution is that merging the training data of multiple scenes completely destroys the data distribution of each scene, and the model tends to learn the data distribution of the scene with more data, resulting in inconsistent effects of each scene; 3) the purpose of this solution is to improve the shortcomings of solution 2, and simply add scene features during modeling to enable the model to learn to distinguish the data distribution of each scene. However, the deep model is not omnipotent, and it is difficult to find a balance point between data sharing and self-adapting to the data distribution of each scene by simply giving scene features, and the actual landing effect is not obvious.

[0004] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] In view of the problems in the prior art, the present application aims to provide a recommendation model training method, a recommendation method, a device and a storage medium, aiming at the same recommendation candidate set in different scenarios having respective user behavior distributions, and the problem of online effect difference of each scenario caused by merging multi-scenario data modeling, a deep network unit of deep cross-scenario features is developed, which can be flexibly embedded into a recall and / or sorting deep model of a recommendation system, so as to achieve the purposes of sharing multi-scenario data and better matching the data distribution of each scenario.

[0006] The embodiment of the present application provides a recommendation model training method, comprising the following steps:

[0007] obtaining a first recommendation model and a deep cross-scenario feature network unit;

[0008] constructing a second recommendation model based on the first recommendation model and the deep cross-scenario feature network unit, wherein the deep cross-scenario feature network unit is configured to cross-combine non-scenario features and scenario features;

[0009] collecting a first sample data set, wherein the data of each sample in the first sample data set comprises scenario features, non-scenario features and corresponding recommendation result labels;

[0010] training the second recommendation model based on the first sample data set.

[0011] In some embodiments, the second recommendation model is configured to splice the scenario features and the non-scenario features to obtain first features, input the scenario features and the non-scenario features into the deep cross-scenario feature network unit to obtain second features, and input the first features and the second features into the first recommendation model, wherein the output result of the first recommendation model is used as a predicted recommendation result.

[0012] In some embodiments, the first recommendation model is a first recall model, and the second recommendation model is a second recall model; or,

[0013] the first recommendation model is a first sorting model, and the second recommendation model is a second sorting model.

[0014] In some embodiments, the deep cross-scenario feature network unit comprises a scenario feature side and a non-scenario feature side, and the weight matrix and the bias term between each processing layer in the non-scenario feature side are input by the scenario feature side.

[0015] In some embodiments, the scene feature side of the deep cross scene feature network unit comprises a first processing layer, an input scene feature is processed by Embedding and full connection to obtain a first vector, and the first processing layer splits the first vector to obtain a first weight matrix, a first bias term, a second weight matrix and a second bias term;

[0016] The non-scene feature side comprises a second processing layer, a third processing layer and a fourth processing layer, the first weight matrix and the first bias term are input between the second processing layer and the third processing layer, and the second weight matrix and the second bias term are input between the third processing layer and the fourth processing layer.

[0017] In some embodiments, the first recommendation model is a first recall model, and the second recommendation model is a second recall model.

[0018] After training the second recommendation model based on the sample data set, the method further comprises the following steps:

[0019] A third recommendation model is obtained, and the third recommendation model is a third sorting model.

[0020] A fourth recommendation model is constructed based on the third recommendation model and the deep cross scene feature network unit, and the fourth recommendation model is a fourth sorting model.

[0021] A second sample data set is collected, and the data of each sample in the second sample data set comprises scene features, non-scene features and corresponding recommendation result labels.

[0022] The fourth recommendation model is trained based on the second sample data set.

[0023] In some embodiments, the non-scene features comprise at least one of user features, item features and context features, the predicted recommendation result of the second recommendation model comprises a probability of whether to recall, and the predicted recommendation result of the fourth recommendation model comprises a click probability of a user.

[0024] In some embodiments, the fourth recommendation model is configured to splice scene features and non-scene features to obtain a third feature, input the scene features and non-scene features into the deep cross scene feature network unit to obtain a fourth feature, and input the third feature and the fourth feature into the third recommendation model, and the output result of the third recommendation model is taken as a predicted recommendation result.

[0025] The embodiment of the application further provides a recommendation method, which adopts the second recommendation model and the fourth recommendation model, and the recommendation method comprises the following steps:

[0026] A user's request to view an item is received.

[0027] obtain a scene feature, a user feature and a context feature, combine the user feature, the context feature and a candidate item feature to obtain a first non-scene feature;

[0028] input the scene feature and the first non-scene feature into the second recommendation model;

[0029] determine a recall item list according to an output of the second recommendation model;

[0030] combine the user feature, the context feature and a recall item feature to obtain a second non-scene feature;

[0031] input the scene feature and the second non-scene feature into the fourth recommendation model;

[0032] determine a recall item ranking according to an output of the fourth recommendation model.

[0033] Embodiments of the present application also provide a recommendation system for implementing the recommendation model training method, and the system comprises:

[0034] a request receiving module configured to receive a user's viewing item request;

[0035] an item recall module configured to obtain a scene feature, a user feature and a context feature, combine the user feature, the context feature and a candidate item feature to obtain a first non-scene feature, input the scene feature and the first non-scene feature into the second recommendation model, and determine a recall item list according to an output of the second recommendation model;

[0036] an item ranking module configured to combine the user feature, the context feature and a recall item feature to obtain a second non-scene feature, input the scene feature and the second non-scene feature into the fourth recommendation model, and determine a recall item ranking according to an output of the fourth recommendation model.

[0037] Embodiments of the present application also provide a recommendation device, which comprises:

[0038] a processor;

[0039] a memory having executable instructions of the processor stored therein;

[0040] The processor is configured to execute the steps of the recommendation method by executing the executable instructions.

[0041] Embodiments of the present application also provide a computer readable storage medium for storing a program, and the program is executed by a processor to implement the steps of the recommendation method.

[0042] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory and are not restrictive of the disclosure.

[0043] The recommendation model training method, the recommendation method, the device and the storage medium of the present application have the following beneficial effects:

[0044] The present application develops a deep network unit of cross-scene features, which can be flexibly embedded into a first recommendation model of a recommendation system to obtain a second recommendation model that better integrates scene features. The first recommendation model can be a traditional recall and / or ranking model, thereby achieving the purpose of sharing multi-scene data and better matching the distribution of each scene data. BRIEF DESCRIPTION OF DRAWINGS

[0045] Other features, objects, and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0046] Figure 1 is a flowchart of a recommendation model training method according to an embodiment of the present application;

[0047] Figure 2 is a structural schematic diagram of a deep cross-scene feature network unit according to an embodiment of the present application;

[0048] Figure 3 is a structural schematic diagram of a second recommendation model according to an embodiment of the present application;

[0049] Figure 4 is a flowchart of a recommendation method according to an embodiment of the present application;

[0050] Figure 5 is a structural schematic diagram of a recommendation system according to an embodiment of the present application;

[0051] Figure 6 is a structural schematic diagram of a recommendation device according to an embodiment of the present application;

[0052] Figure 7 is a structural schematic diagram of a computer-readable storage medium according to an embodiment of the present application. DETAILED DESCRIPTION

[0053] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these implementations are provided so that this disclosure will be thorough and complete, and will fully convey the scope of example implementations to those skilled in the art. Features described in the description, structures, or characteristics may be combined in any suitable manner in one or more implementations.

[0054] Furthermore, the accompanying drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification. The drawings are not necessarily to scale, the dimensions of certain features being purposely exaggerated in order to illustrate aspects not readily apparent from the other features. In the drawings, the same reference numbers signify the same or similar parts. For the purposes of the description hereinafter, the terms "upper", "lower", "right", "left", "vertical", "horizontal", "top", "bottom", "lateral", "length", "width", "thickness", "end", "edge", "surface", "side", "front", "rear", "back", "forward", "rearward", and "port" refer to the apparatus as oriented in the drawing figures. However, it is to be understood that unless otherwise specified, the apparatus can assume any orientation and the terms are not to be construed as limiting.

[0055] The flowcharts shown in the drawings are merely illustrative and do not necessarily include all steps. For example, some steps can be further divided, and some steps can be combined or partially combined, so the actual execution order can be changed according to the actual situation.

[0056] As Figure 1 shown, the embodiment of the present application provides a recommendation model training method, comprising the following steps:

[0057] S110: obtaining a first recommendation model and a deep cross-scene feature network unit;

[0058] S120: constructing a second recommendation model based on the first recommendation model and the deep cross-scene feature network unit, the deep cross-scene feature network unit being configured to cross-combine non-scene features and scene features;

[0059] S130: collecting a first sample data set, the data of each sample in the first sample data set including scene features, non-scene features and corresponding recommendation result labels;

[0060] S140: training the second recommendation model based on the first sample data set.

[0061] The present application develops a deep network unit of deep cross-scene features by using the recommendation model training method, which can be flexibly embedded into the first recommendation model of the recommendation system to obtain a second recommendation model that better fuses scene features. The first recommendation model can be in a traditional recall and / or sorting model, thereby achieving the purpose of sharing multi-scene data and better matching the distribution of each scene data, and solving the problem of data sparsity that has always existed in the recommendation system.

[0062] In this embodiment, the first recommendation model is a first recall model, and the second recommendation model is a second recall model, that is, a recall model that better matches the scene data is obtained through the recommendation model training method. The recall model is used to recall recommended items according to user-related non-scene features and scene features. The first recall model is a traditional recall model in the prior art, such as a recall model constructed based on a deep learning network. For example, a plurality of sample data are collected in advance, and recall labels are marked for each sample data. If the recall is marked as 1, and if the recall is not marked as 0, the sample data is used as the recall label. Then the first recall model is pre-trained using the sample data, that is, the sample data is input into the first recall model to obtain the recall probability predicted by the first recall model. A loss function is constructed according to the recall label, and the first recall model is iteratively trained based on the loss function value.

[0063] Alternatively, the first recommendation model is a first ranking model, and the second recommendation model is a second ranking model, that is, a ranking model that better matches the scene data is obtained through the recommendation model training method. The ranking model is used to sort the recalled recommended items according to user-related non-scene features and scene features. The first ranking model is a traditional ranking model in the prior art, such as a ranking model constructed based on a deep learning network. For example, a plurality of sample data are collected in advance, and click probability labels are marked for each sample data. If the click is marked as 1, and if the click is not marked as 0, the sample data is used as the click label. Then the first ranking model is pre-trained using the sample data, that is, the sample data is input into the first ranking model to obtain the click probability predicted by the first ranking model. A loss function is constructed according to the click probability label, and the first ranking model is iteratively trained based on the loss function value.

[0064] The first recommendation model is a first recall model, and the second recommendation model is a second recall model, which is used as an example for description, but the present application is not limited thereto.

[0065] In this embodiment, the second recommendation model is configured to splice the scene features and the non-scene features to obtain first features, input the scene features and the non-scene features into the deep cross-scene feature network unit to obtain second features, and input the first features and the second features into the first recommendation model. The output result of the first recommendation model is used as a predicted recommendation result.

[0066] For example, Figure 2As shown, it is a structural schematic diagram of the deep cross scene feature network unit of the embodiment. In the embodiment, the deep cross scene feature network unit (DCSU) includes a scene feature side and a non-scene feature side, and the weight matrix and bias term between each processing layer in the non-scene feature side are input by the scene feature side. The non-scene feature is normally input for several layers of full connection, but the weight matrix and bias term (dimension) of the hidden layer x2, x3 and x4 in the middle are directly generated by the scene feature side, so as to achieve the interaction purpose of the non-scene feature and the scene feature.

[0067] In the embodiment, the scene feature side of the deep cross scene feature network unit includes a first processing layer, the input scene feature is processed by Embedding and full connection to obtain a first vector, and the first processing layer splits the first vector to obtain a first weight matrix, a first bias term, a second weight matrix and a second bias term. The non-scene feature side includes a second processing layer, a third processing layer and a fourth processing layer, the first weight matrix and the first bias term are input between the second processing layer and the third processing layer, and the second weight matrix and the second bias term are input between the third processing layer and the fourth processing layer. In the embodiment, the first processing layer, the second processing layer, the third processing layer and the fourth processing layer are full connection layers respectively.

[0068] Figure 3 It is a structural schematic diagram of the second recommendation model of an embodiment of the application, which includes a first recommendation model and the deep cross scene feature network unit. As shown in Figure 2 and Figure 3 As shown, specifically, on the scene feature side, the scene feature is first processed by Embedding and full connection into an x1 layer A-dimensional vector, the x1 layer A-dimensional vector is split into a b*c-dimensional matrix, a 1*c-dimensional bias term and a c*d-dimensional matrix, a 1*d-dimensional bias term, and the two weight matrices and bias terms are respectively used as the weight matrix and bias term between the hidden layers x2, x3 and x4. Finally, the interaction result of the scene feature and the non-scene feature is output through a layer of full connection. The deep cross scene feature network unit is ingenious in that the scene feature is used as the weight matrix of the non-scene feature to adaptively interact, the model is iteratively self-learned, and the balance point of the scene feature and the non-scene feature is found. In Figure 3 In the embodiment, the scene feature and the non-scene feature on the left side are output as a first feature after passing through a concat layer, the deep cross scene feature network unit on the right side outputs a second feature, and the first feature and the second feature are input into the first recommendation model after passing through a concat layer, and the first recommendation model is a conventional recall model network structure.

[0069] The step S140 of training the second recommendation model based on the sample data set further comprises the following steps:

[0070] A third recommendation model is obtained, and the third recommendation model is a third ranking model.

[0071] A fourth recommendation model is constructed based on the third recommendation model and the deep cross-scene feature network unit, and the fourth recommendation model is a fourth ranking model.

[0072] A second sample data set is collected, and the data of each sample in the second sample data set comprises scene features, non-scene features and corresponding recommendation result labels.

[0073] The fourth recommendation model is trained based on the second sample data set.

[0074] The third ranking model is a conventional ranking model in the prior art, such as a ranking model constructed based on a deep learning network. For example, a plurality of sample data are collected in advance, and a click probability label is marked for each sample data. If clicked, it is marked as 1, and if not clicked, it is marked as 0, as a click label of the sample data. Then the third ranking model is pre-trained using the sample data, that is, the sample data is input into the third ranking model to obtain the click probability predicted by the third ranking model, a loss function is constructed according to the click probability label, and the third ranking model is iteratively trained based on the loss function value. The architecture of the deep cross-scene feature network unit can adopt the architecture shown in Figure 2 The architecture of the fourth recommendation model can adopt the architecture shown in Figure 3 In this embodiment, the fourth recommendation model is configured to splice the scene features and the non-scene features to obtain third features, input the scene features and the non-scene features into the deep cross-scene feature network unit to obtain fourth features, and input the third features and the fourth features into the third recommendation model, and the output result of the third recommendation model is used as a predicted recommendation result. That is, when the architecture of Figure 3 The scene features and the non-scene features on the left side are output as third features after passing through the concat layer, the deep cross-scene feature network unit on the right side outputs fourth features, and the third features and the fourth features are input into the fourth recommendation model after passing through the concat layer, and the fourth recommendation model is a conventional ranking model network structure.

[0075] Therefore, by adopting the method of this embodiment, a second recall model and a fourth ranking model can be trained. In this embodiment, the non-scene features include at least one of user features, item features and context features. The predicted recommendation result of the second recall model includes a probability of whether to recall, and the predicted recommendation result of the fourth ranking model includes a click probability of the user. The user features may, for example, include user age, registration duration, gender, residence, etc. The item features may, for example, include name, category, price, characteristics, etc. of the item, and the item can be a commodity and / or service for sale / rental. The scene features can include a current scene category. For example, the scene can be divided into multiple categories, the first category is a hotel booking scene, the second category is a list page scene, and so on.

[0076] As shown in Figure 4 The embodiment of the present application also provides a recommendation method, which adopts the second recommendation model and the fourth recommendation model, and includes the following steps:

[0077] S210: receiving a user's request for viewing an item, in the present application, the item can be a commodity, a service, etc. searched by the user, and the user can send the request for viewing the item by inputting the name of the item to be searched in a search box, or the user can be recommended an item when opening a recommendation page;

[0078] S220: obtaining scene features, user features and context features, and combining the user features, the context features and candidate item features to obtain first non-scene features;

[0079] S230: inputting the scene features and the first non-scene features into the second recommendation model, the second recommendation model being trained by the corresponding steps of the above recommendation model training method, and the second recommendation model being a second recall model;

[0080] S240: determining a recall item list according to the output of the second recommendation model;

[0081] S250: combining the user features, the context features and the recall item features to obtain second non-scene features;

[0082] S260: inputting the scene features and the second non-scene features into the fourth recommendation model, the fourth recommendation model being trained by the corresponding steps of the above recommendation model training method, and the fourth recommendation model being a fourth ranking model;

[0083] S270: determining a ranking of the recall item according to the output of the fourth recommendation model.

[0084] The application can push the sorted recall item information in the form of a list to the user through the second recommendation model to obtain a recall item list and the fourth recommendation model to obtain a recall item ranking, so that the user can conveniently select and view.

[0085] As shown in Figure 5 The application also provides a recommendation system for implementing the recommendation model training method, and the system comprises:

[0086] The request receiving module M100 is configured to receive a view item request of the user.

[0087] The item recall module M200 is configured to obtain a scene feature, a user feature and a context feature, combine the user feature, the context feature and a candidate item feature to obtain a first non-scene feature, input the scene feature and the first non-scene feature into the second recommendation model, and determine a recall item list according to the output of the second recommendation model.

[0088] The item ranking module M300 is configured to combine the user feature, the context feature and a recall item feature to obtain a second non-scene feature, input the scene feature and the second non-scene feature into the fourth recommendation model, and determine a recall item ranking according to the output of the fourth recommendation model.

[0089] The application can push the sorted recall item information in the form of a list to the user through the second recommendation model to obtain a recall item list and the fourth recommendation model to obtain a recall item ranking, so that the user can conveniently select and view. The second recommendation model and the fourth recommendation model both combine the scene feature and the non-scene feature, so that the final recall result and the ranking result are more in line with the user demand and the specific scene demand, and the user experience of the recommendation system is improved.

[0090] In the recommendation system, the functions of the modules can be implemented by the specific embodiments of the recommendation method, which will not be described here.

[0091] The application also provides a recommendation device comprising a processor, a memory having executable instructions of the processor stored therein, wherein the processor is configured to execute the steps of the recommendation method by executing the executable instructions.

[0092] Those skilled in the art can understand that each aspect of the present application can be implemented as a system, a method or a program product. Therefore, each aspect of the present application can be embodied in a form of entirely hardware, entirely software (including firmware, microcode, etc.), or a combination of hardware and software, which can be generically referred to as "circuitry", "module" or "platform".

[0093] The electronic device 600 according to this embodiment of the present application will be described below with reference to Figure 6 Figure 6 The electronic device 600 shown is merely an example and should not limit the function and scope of use of the embodiments of the present application.

[0094] As shown in Figure 6 , the electronic device 600 is in the form of a general computing device. The components of the electronic device 600 can include, but are not limited to, at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different system components (including the storage unit 620 and the processing unit 610), a display unit 640, etc.

[0095] The storage unit stores program codes which can be executed by the processing unit 610, so that the processing unit 610 performs the steps according to various exemplary embodiments of the present application described in the recommended method part of the present specification. For example, the processing unit 610 can perform the steps as shown in Figure 4

[0096] The storage unit 620 can include a readable medium in the form of a volatile storage unit, such as a random access memory (RAM) 6201 and / or a cache memory 6202, and can further include a read-only memory (ROM) 6203.

[0097] The storage unit 620 can further include program / utilities 6204 having a set of (at least one) program modules 6205, such as an operating system, one or more application programs, other program modules, and program data, each of which or some combination of which can include implementation of a network environment.

[0098] The bus 630 can represent one or more of several types of bus structures, including a storage unit bus or storage unit controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of a variety of bus structures.

[0099] ​​The electronic device 600 can also communicate with one or more external devices 700 such as a keyboard or a pointing device, a Bluetooth device, etc.; other devices that enable a user to interact with the electronic device 600; and / or any devices (e.g., a router, a modem, a switch, etc.) that enable the electronic device 600 to communicate with one or more other computing devices. Such communication can occur via an input / output (I / O) interface 650. Still yet, the electronic device 600 can communicate with one or more networks, such as one or more local area networks (LANs), wide area networks (WANs), and / or the Internet, through a network adapter 660. The network adapter 660 can communicate with the other components of the electronic device 600 via the bus 630. It should be appreciated that the electronic device 600 has many of the same components as the computing device 100. The same components of the electronic device 600 will have the same reference numbers as those of the computing device 100 but with a following digit that is different. As described, those components of the electronic device 600 that are the same as those of the computing device 100, including the bus 630, the processor 602, the memory 604, the storage component 606, the input component 612, the output component 614, the display 616, the communication interface 620, and the network adapter 660, can perform the functions of the components of the computing device 100 as described above and as known to those of skill in the art.

[0100] The program in the memory, when executed by the processor, implements the steps of the recommendation method, and thus the device can also obtain the technical effects of the recommendation method.

[0101] The embodiment of the present application also provides a computer readable storage medium for storing a program, which, when executed by a processor, implements the steps of the recommendation method. In some possible implementation manners, various aspects of the present application can also be implemented in the form of a program product, which includes program codes for causing a terminal device to perform the steps described in the above method part of the specification according to various exemplary embodiments of the present application when the program product is executed on the terminal device.

[0102] Reference Figure 7 As shown, a program product 800 for implementing the above method according to the embodiment of the present application is described, which can adopt a portable compact disc read-only memory (CD-ROM) and include program codes, and can be executed on a terminal device, such as a personal computer. However, the program product of the present application is not limited to this, and in this document, the readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, device or apparatus.

[0103] The program product can employ any combination of one or more computer-readable media. The computer-readable media can be a computer-readable storage medium or a computer-readable signal medium. The computer-readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include the following: an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0104] The computer-readable storage medium can include data signals on a carrier wave, propagated over a propagation medium, in which the computer-readable program code is embodied. Such propagated signals can take a wide variety of forms, including but not limited to, electro-magnetic signals, optical signals, or any suitable combination thereof. The computer-readable storage medium can also be any computer-readable medium other than a computer-readable storage medium that can be a source of computer-readable program code, which can be read by a computer. The computer-readable program code embodied on the computer-readable storage medium can be transmitted using any appropriate medium, including but not limited to wireless, wired, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0105] The program code for carrying out operations of the present application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++, etc., and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's device and partly on a remote computing device or entirely on the remote computing device or server. In the latter scenario, the remote computing device can be connected to the user's device through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computing device, such as through the Internet using an Internet Service Provider (ISP).

[0106] The program in the computer storage medium is executed by the processor to implement the steps of the recommendation method, and therefore the computer storage medium can also obtain the technical effects of the recommendation method.

[0107] The above is further detailed description of the present application in combination with specific preferred embodiments, and cannot be deemed as limitation of the specific implementation of the present application to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, and all should be deemed as falling within the protection scope of the present application.

Claims

1. A method for training a recommendation model, characterized in that, The method comprises the following steps: obtain a first recommendation model and a deep cross-scene feature network unit, the deep cross-scene feature network unit comprising a scene feature side and a non-scene feature side, the weight matrix and bias term between each processing layer in the non-scene feature side being input by the scene feature side; construct a second recommendation model based on the first recommendation model and the deep cross-scene feature network unit, the deep cross-scene feature network unit being configured to cross-combine non-scene features and scene features; wherein the first recommendation model is a first recall model, and the second recommendation model is a second recall model; collect a first sample data set, the data of each sample in the first sample data set comprising scene features, non-scene features, and corresponding recommendation result labels, the scene features comprising a current scene category, the scene category being a hotel post-booking scenic spot recommendation scene or a list page scenic spot sorting scene, the non-scene features comprising at least one of user features, item features, and context features; train the second recommendation model based on the first sample data set, the predicted recommendation result of the second recommendation model comprising a probability of whether to recall; obtain a third recommendation model, the third recommendation model being a third sorting model; construct a fourth recommendation model based on the third recommendation model and the deep cross-scene feature network unit, the fourth recommendation model being a fourth sorting model; collect a second sample data set, the data of each sample in the second sample data set comprising scene features, non-scene features, and corresponding recommendation result labels; train the fourth recommendation model based on the second sample data set, the predicted recommendation result of the fourth recommendation model comprising a user's click probability; wherein the second recommendation model is configured to concatenate scene features and non-scene features to obtain first features, input the scene features and non-scene features into the scene feature side and the non-scene feature side of the deep cross-scene feature network unit respectively to obtain second features, and input the first features and the second features into the first recommendation model, the output result of the first recommendation model being used as a predicted recommendation result; wherein the scene feature side of the deep cross-scene feature network unit comprises a first processing layer, the input scene features being processed by Embedding and full connection to obtain a first vector, the first processing layer splitting the first vector to obtain a first weight matrix, a first bias term, a second weight matrix, and a second bias term; the non-scene feature side comprising a second processing layer, a third processing layer, and a fourth processing layer, the first weight matrix and the first bias term being input between the second processing layer and the third processing layer, and the second weight matrix and the second bias term being input between the third processing layer and the fourth processing layer.

2. The recommendation model training method of claim 1, wherein, The fourth recommendation model is configured to concatenate scene features and non-scene features to obtain a third feature, input the scene features and non-scene features into the deep cross-scene feature network unit to obtain a fourth feature, input the third feature and the fourth feature into the third recommendation model, and use the output of the third recommendation model as the prediction recommendation result.

3. A recommendation method characterized by, Using the second and fourth recommendation models as described in claim 2, the recommendation method includes the following steps: Received a user's request to view items; Obtain scene features, user features, and context features, and combine the user features, context features, and candidate item features to obtain a first non-scene feature; The scene features and the first non-scene features are input into the second recommendation model; The list of items to be recalled is determined based on the output of the second recommendation model; The user characteristics, the context characteristics, and the recalled item characteristics are combined to obtain the second non-scenario characteristic; The scene features and the second non-scene features are input into the fourth recommendation model; The order of recalled items is determined based on the output of the fourth recommendation model.

4. A recommendation system characterized in that, The system for implementing the recommended method of claim 3 includes: The request receiving module is used to receive user requests to view items. The item recall module is used to acquire scene features, user features and context features, combine the user features, the context features and candidate item features to obtain a first non-scene feature, input the scene features and the first non-scene feature into the second recommendation model, and determine the recall item list based on the output of the second recommendation model. The item sorting module is used to combine the user features, the context features, and the recalled item features to obtain a second non-scenario feature, input the scenario features and the second non-scenario feature into the fourth recommendation model, and determine the order of recalled items based on the output of the fourth recommendation model.

5. A recommendation device, characterized by include: processor; A memory in which executable instructions of the processor are stored; The processor is configured to perform the steps of the recommended method of claim 3 by executing the executable instructions.

6. A computer readable storage medium for storing a program, characterized in that, When the program is executed by the processor, it implements the steps of the recommended method as described in claim 3.

Images