A game recommendation method and system based on hybrid model

Abstract

The embodiment of the present application provides a game recommendation method and system based on a hybrid model, the method comprising: inputting the feature vectors of the candidate game's historical click records, historical search records, and historical login activation records into a deep recall model, and obtaining the initial screening results of the game through the deep recall model; obtaining the ranking results of each game in the initial screening results through a deep sorting model; and recommending games based on the ranking results of the game. When the embodiment of the present application uses the deep recall model for recall, in addition to using the candidate game's historical click records and historical search records, it also uses the candidate game's historical login activation records, thereby improving the recall recommendation accuracy, which can further enhance the recommendation effect.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computer technology, and in particular to a game recommendation method, system, computer device, and computer-readable storage medium based on a hybrid model. Background Art

[0002] With the development of the Internet, various games have been continuously proposed. As there are more and more games, it is difficult for users to quickly find the games they are interested in. In order to facilitate users to obtain the games they are interested in, many game platforms will actively recommend some games to users:

[0003] When a user scrolls down and views games on the game center homepage provided by the game platform, the background of the game platform returns the game recommendation results to the user after a series of calculations and presents them on the game center homepage.

[0004] At present, the architecture diagram of the intelligent game recommendation system used by the game platform is as follows: Figure 1 As shown, it usually includes: data collection layer, data storage layer, data calculation layer, feature layer, recall layer, filtering layer, sorting layer, and application service layer from bottom to top.

[0005] The key links are the recall layer and the ranking layer; the recall layer is responsible for the initial screening of candidate games, selecting hundreds of games from the pool of tens of thousands of games for downstream use; the ranking layer is responsible for sorting the hundreds of recalled games. In the current intelligent game recommendation system, the recall layer and the ranking layer use the YouTubeDNN (YouTube Deep Neural Network) deep recall model and the YouTubeDNN (YouTube Deep Neural Network) deep ranking model respectively.

[0006] However, the existing intelligent game recommendation system has the following limitations in the game center recommendation scenario: although games with user historical viewing and historical search are recommended on the game center homepage, games with a large amount of historical play but no user behavior of viewing and searching are not recommended.

[0007] Therefore, it is necessary to provide a game recommendation system that can further improve the recommendation effect. Summary of the invention

[0008] The purpose of the embodiments of the present application is to provide a game recommendation method, system, computer device and computer-readable storage medium based on a hybrid model. When using the deep recall model for recall, in addition to using the historical click records and historical search records of the candidate games, the historical login activation records of the candidate games are also used, thereby improving the accuracy of the recall recommendation, and the recommendation effect can be further improved.

[0009] One aspect of an embodiment of the present application provides a game recommendation method based on a hybrid model, comprising:

[0010] Inputting feature vectors of the candidate game's historical click records, historical search records, and historical login activation records into a deep recall model, and obtaining a preliminary screening result of the game through the deep recall model;

[0011] The preliminary screening results of the games are subjected to a deep sorting model to obtain a sorting result of each game in the preliminary screening results;

[0012] The game is recommended according to the ranking results of the games.

[0013] Optionally, the deep recall model is specifically an improved YouTube deep neural network deep recall model, and its model structure is based on the YouTube deep neural network deep recall model and further expands the following neurons:

[0014] The input layer neurons are expanded from two groups to three groups;

[0015] The first hidden layer neurons are expanded from two to three, connecting three groups of input layer neurons respectively;

[0016] The neurons in the second hidden layer are expanded from two to three, which are connected to the three neurons in the first hidden layer respectively.

[0017] Optionally, in the improved YouTube deep neural network deep recall model,

[0018] The first, second and third groups of neurons in the input layer are used to input feature vectors of the historical click records, historical search records and historical login activation records of the candidate games respectively.

[0019] Optionally, the neurons in the first hidden layer are used to perform weighted average calculation on the feature vectors input by a group of input layer neurons connected thereto according to the feature values ​​of the timeliness of the user behavior in the feature vectors.

[0020] Optionally, the neurons in the second hidden layer are used to form a vector with weighted average calculation results of the outputs of the neurons in the first hidden layer connected thereto.

[0021] Optionally, the deep sorting model is specifically an improved deep factorization machine DeepFM model; and

[0022] The improved DeepFM model obtains the ranking results of the games according to the interest level of each game in the initial screening results.

[0023] Optionally, the step of applying the preliminary screening results of the games to a deep sorting model to obtain a sorting result of each game in the preliminary screening results specifically includes:

[0024] Determining the interest level of each game in the preliminary screening results;

[0025] Sort the games by their interest level.

[0026] Optionally, determining the interest level of each game in the preliminary screening result specifically includes:

[0027] For each game in the preliminary screening result, when the game is judged to be a new game, the interest level of the game is determined according to the release time of the game;

[0028] If the game is not a new game and is a low-exposure game, determining the interest level of the game based on the historical click-through rate and exposure of the game;

[0029] If the game is not a new game and is not a low-exposure game, the interest level of the game is determined based on the difference between the current time and the user behavior time of the game.

[0030] One aspect of an embodiment of the present application provides a game recommendation system based on a hybrid model, including:

[0031] The recall layer is used to input the feature vectors of the candidate game's historical click records, historical search records, and historical login activation records into the deep recall model, and obtain the initial screening results of the game through the deep recall model;

[0032] A sorting layer, used to obtain the sorting results of each game in the preliminary screening results by applying the preliminary screening results of the games to a deep sorting model;

[0033] The application service layer is used to recommend games based on the ranking results of the games.

[0034] One aspect of an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned hybrid model-based game recommendation method when executing the computer program.

[0035] One aspect of an embodiment of the present application provides a computer-readable storage medium, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the game recommendation method based on the hybrid model are implemented.

[0036] The hybrid model-based game recommendation method, system, device, and computer-readable storage medium provided in the embodiments of the present application input the historical click records, historical search records, and historical login activation records of candidate games into the deep recall model, and obtain the initial screening results of the games through the deep recall model; obtain the ranking results of each game in the initial screening results through the deep sorting model; and recommend games based on the ranking results of the games. Because when the deep recall model is used for recall, in addition to the historical click records and historical search records of the candidate games, the historical login activation records of the candidate games are also used, thereby improving the recall recommendation accuracy, and the recommendation effect can be further improved.

[0037] Furthermore, the improved YouTube DNN deep recall model has undergone certain structural modifications based on the existing YouTube DNN deep recall model. On the one hand, the improved YouTubeDNN deep recall model can process not only the historical click records and historical search records of the candidate games, but also the historical login activation records of the candidate games, and the output recalled games can include games with historical login activation records. On the other hand, the modification plan can be implemented at a lower cost.

[0038] Furthermore, the deep sorting model adopts an improved DeepFM model. Compared with the existing DeepFM model, when determining the interest of a game, it not only takes into account the decay of user interest over time, but also takes into account new games and high-quality games with low exposure and high click-through rates, which further improves the recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic diagram of the architecture of an intelligent game recommendation system in the prior art;

[0040] Figure 2 A diagram schematically shows an application environment of a game recommendation method based on a hybrid model according to an embodiment of the present application;

[0041] Figure 3a , 3b The following schematically shows an architecture diagram of a game recommendation system based on a hybrid model according to the first embodiment of the present application;

[0042] Figure 4 The flowchart of the game recommendation method based on the hybrid model according to the second embodiment of the present application is schematically shown;

[0043] Figure 5a A schematic diagram of the internal structure of the existing YoutubeDNN deep recall model;

[0044] Figure 5bThe schematic diagram of the internal structure of the improved YoutubeDNN deep recall model according to the second embodiment of the present application is shown;

[0045] Figure 6 The flowchart of the method for determining the interest level of a game according to the improved DeepFM model of the second embodiment of the present application is schematically shown;

[0046] Figure 7 The hardware architecture diagram of a computer device suitable for implementing a game recommendation method based on a hybrid model according to the third embodiment of the present application is schematically shown. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0048] It should be noted that the descriptions involving "first", "second", etc. in the embodiments of the present application are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0049] In the description of the present application, it should be understood that the numerical labels before the steps do not indicate the order in which the steps are executed, but are only used to facilitate the description of the present application and to distinguish each step, and therefore should not be understood as a limitation on the present application.

[0050] Figure 2 The following schematic diagram shows an environmental application diagram according to an embodiment of the present application. Figure 2 As shown:

[0051] The server 2 provided by the game service provider can be connected to multiple terminal devices 6 via the network 4.

[0052] The server 2 can provide game services for the terminal device 6. The game center homepage provided by the server 2 can recommend games that the user may be interested in. For example, the server 2 can display the recommended results of the game on its game center homepage, and the recommended results can be obtained by the game recommendation method based on the hybrid model in the embodiment of the present application.

[0053] The server 2 may be the following devices, such as: a rack server, a blade server, a tower server or a cabinet server (including an independent server, or a server cluster composed of multiple servers), etc.

[0054] Server 2 may be located in a data center such as a single location, or distributed in different geographic locations (e.g., in multiple locations). Server 2 may provide services via one or more networks 4. Network 4 includes various network devices, such as routers, switches, multiplexers, hubs, modems, bridges, repeaters, firewalls, proxy devices, and / or the like. Network 4 may include physical links, such as coaxial cable links, twisted pair cable links, fiber optic links, combinations thereof, and the like. Network 4 may include wireless links, such as cellular links, satellite links, Wi-Fi links, and the like.

[0055] The terminal device 6 may be configured to access the server 2. The terminal device 6 may include any type of terminal device 6, such as: a mobile device, a tablet device, a laptop computer, a smart device (e.g., a smart watch, smart glasses), a virtual reality headset, a gaming device, a set-top box, a digital streaming device, a robot, a vehicle-mounted terminal, a smart TV, an e-book reader, an MP4 (Moving Picture Experts Group Audio Layer IV) player, etc.

[0056] Terminal device 6, can run Windows system, Android TM ) system and other operating systems. Users can also install various application installation packages as needed to obtain applications that implement various specific functions.

[0057] This application provides a new game recommendation solution based on a hybrid model for the following situations:

[0058] The applicant analyzed the data of the game center and found that half of the users in the game center have played games but have no viewing or search records. For these users, if only the existing intelligent game recommendation system is used, the values ​​of the browse and search towers of the games played by these users are all 0, which means that the YoutubeDNN deep recall model in the existing intelligent game recommendation system will not recall games that are played by users but have no browsing and search records. As a result, the model degenerates into games that are statistically related based on demographic characteristics, and loses personalized recall and recommendation effects. In addition, some users have different degrees of preference for games with historical clicks or searches (such as download preference is greater than clicks), and the existing YoutubeDNN deep recall model does not take these preferences into account.

[0059] The new hybrid model-based game recommendation solution provided in the present application, when using the deep recall model for recall, in addition to using the historical click records and historical search records of the candidate games, also uses the historical login activation records of the candidate games, thereby improving the accuracy of the recall recommendation and further improving the recommendation effect.

[0060] The following will introduce a game recommendation scheme based on a hybrid model through multiple embodiments. The scheme can be implemented by a computer device 1300, which can be a server 2 or a computing node thereof.

[0061] Embodiment 1

[0062] Figure 3a The schematic diagram shows an architecture diagram of a game recommendation system based on a hybrid model according to the first embodiment of the present application.

[0063] like Figure 3a As shown, the game recommendation system based on the hybrid model includes: a recall layer 305, a ranking layer 307 and an application service layer 308; wherein:

[0064] The recall layer 305 is used to input the feature vectors of the candidate game's historical click records, historical search records, and historical login activation records into the deep recall model, and obtain the initial screening results of the game through the deep recall model; as a more optimal implementation method, the deep recall model can adopt the improved YoutubeDNN (YouTube Deep Neural Network) deep recall model provided by the present invention, and the specific structure and function of the model will be introduced in detail later.

[0065] The sorting layer 307 is used to obtain the sorting results of each game in the preliminary screening results through a deep sorting model. As a better implementation, the deep sorting model can adopt the improved DeepFM model provided by the present invention. The specific structure and function of the model will be introduced in detail later.

[0066] The application service layer 308 is used to recommend games according to the ranking results of the games.

[0067] Furthermore, the game recommendation system based on the hybrid model may also include: a data collection layer 301, a data storage layer 302, a data calculation layer 303, a feature layer 304, and a filtering layer 306;

[0068] In the game recommendation system based on the hybrid model of the present application, from bottom to top, they are: data collection layer 301, data storage layer 302, data calculation layer 303, feature layer 304, recall layer 305, filtering layer 306, sorting layer 307 and application service layer 308.

[0069] The data collection layer 301 may include: a client tracking unit, a server tracking unit, a game SDK (standard development kit) tracking unit, and a recommendation API log unit, so as to collect and obtain relevant information such as historical click records, historical search records, historical login activation records, etc. of candidate games;

[0070] Compared with the data collection layer in the existing intelligent game recommendation system, the data collection layer 301 adds a game SDK (standard development kit) tracking unit, so that compared with the data collection layer in the existing intelligent game recommendation system, it can obtain the historical login activation record information of the candidate game, that is, the historical play information of the game.

[0071] In the feature layer 304 , the historical click records, historical search records, and historical login activation records of the candidate games can be feature extracted and mapped to the encoding space to obtain the feature vectors of the historical click records, historical search records, and historical login activation records of the candidate games and input them into the recall layer 305 .

[0072] Compared with the feature layer in the existing intelligent game recommendation system, the feature layer 304 adds the operation of extracting features from the historical login activation records of the candidate games and mapping them to the encoding space, and outputs a feature vector with the historical login activation records of the candidate games.

[0073] The functions of the data storage layer 302, the data calculation layer 303, and the filtering layer 306 are respectively the same as those of the data storage layer, the data calculation layer, and the filtering layer in the existing intelligent game recommendation system, and are not described in detail here.

[0074] Furthermore, the game recommendation system based on the hybrid model may also include: an effect evaluation layer 309.

[0075] The effect evaluation layer 309 may perform effect evaluation on the game recommendation results output by the application service layer 308 .

[0076] Figure 3b Another way of expressing the structure of the game recommendation system based on the hybrid model is illustrated.

[0077] The game recommendation system based on the hybrid model in the first embodiment of the present application, when using the deep recall model for recall, in addition to using the historical click records and historical search records of the candidate games, also uses the historical login activation records of the candidate games, thereby improving the accuracy of the recall recommendation and further improving the recommendation effect.

[0078] Embodiment 2

[0079] Figure 4 The flowchart of the game recommendation method based on the hybrid model according to the second embodiment of the present application is schematically shown.

[0080] like Figure 4 As shown, the game recommendation method based on the hybrid model may include the following steps S401 to S405, wherein:

[0081] Step S401: Obtain the historical click records, historical search records, and historical login activation records of the candidate games.

[0082] Specifically, when a user browses the game center homepage provided by the game platform in server 2 through terminal device 6, and scrolls down to view games; the hybrid model-based game recommendation system in server 2 can obtain information on historical click records, historical search records, and historical login activation records of candidate games.

[0083] In an exemplary embodiment, the data collection layer 301 of the hybrid model-based game recommendation system in server 2 can obtain the user's games with historical click records, historical search records, or historical login activation records as candidate games; in fact, games with historical login activation records indicate that the game has been played by the user.

[0084] In another exemplary embodiment, the data collection layer 301 of the hybrid model-based game recommendation system in the server 2 can obtain all users' games with historical click records, historical search records, or historical login activation records as candidate games.

[0085] Step S402: Generate feature vectors of the candidate game's historical click records, historical search records, and historical login activation records.

[0086] In an exemplary embodiment, the feature layer 304 of the hybrid model-based game recommendation system in the server 2 can perform feature extraction and coding mapping on the historical click records of the candidate games, the historical search records of the candidate games, and the historical login activation records of the candidate games, and obtain feature vectors of the historical click records of the candidate games, feature vectors of the historical search records of the candidate games, and feature vectors of the historical login activation records of the candidate games, respectively. How to perform feature extraction and coding mapping is a technology well known to those skilled in the art and will not be described in detail here.

[0087] Step S403: inputting the feature vectors of the historical click records, historical search records, and historical login activation records of the candidate games into the deep recall model, and obtaining the initial screening results of the games through the deep recall model.

[0088] In an exemplary embodiment, the recall layer 305 of the hybrid model-based game recommendation system in server 2 uses an improved YoutubeDNN deep recall model for game recall.

[0089] Figure 5a This is a schematic diagram of the internal structure of the existing YoutubeDNN deep recall model; Figure 5b Schematic diagram of the internal structure of the improved YoutubeDNN deep recall model of this application; Figure 5a and Figure 5b It can be seen from the comparison that the improved YoutubeDNN deep recall model of the present application has a model structure that is based on the YouTube deep neural network deep recall model and further expands the following neurons:

[0090] The input layer neurons are expanded from two groups to three groups;

[0091] The first hidden layer neurons are expanded from two to three, connecting three groups of input layer neurons respectively;

[0092] The neurons in the second hidden layer are expanded from two to three, which are connected to the three neurons in the first hidden layer respectively.

[0093] Specifically, the structure of the improved YoutubeDNN deep recall model of the present application includes the following layers from bottom to top: input layer 501, first hidden layer 502, second hidden layer 503, at least one activation (ReLU) layer 504, and normalization (softmax) layer 505;

[0094] Among them, there can be three groups of neurons in the input layer 501. The first, second and third groups of neurons in the input layer 501 are used to input the feature vectors of the historical click records of the candidate games, the feature vectors of the historical search records of the candidate games, and the feature vectors of the historical login activation records of the candidate games, respectively.

[0095] As a more optimal implementation, the feature vectors of the candidate game's historical click records, the feature vectors of the candidate game's historical search records, and the feature vectors of the candidate game's historical login activation records can also be weighted in the input layer 501 according to the following parameters: the interest of the candidate game, and the timeliness of the video material of the candidate game. In this way, when the improved YoutubeDNN deep recall model of the present application determines the recalled game based on these feature vectors, it can also refer to the interest of the candidate game and the timeliness of the video material of the candidate game, thereby further improving the effect of game recommendation.

[0096] The interest of the candidate game includes clicks, downloads, comments, and forwarding of the candidate game, and the weight coefficient of the interest of the candidate game can be automatically learned by the model;

[0097] The timeliness coefficient of the video material of the candidate game is calculated based on the inverse of the difference between the current time and the update time of the video material of the candidate game; that is:

[0098] The timeliness coefficient of the video material of the candidate game = (1 / current time - update time of the video material of the candidate game);

[0099] When weighting the feature vector of the historical click / historical search / historical login activation record of a candidate game, the feature vector of the historical click / historical search / historical login activation record of the candidate game can be multiplied by the weight coefficient of the interest of the candidate game and the coefficient of the timeliness of the video material of the candidate game: the feature vector of the historical click / historical search / historical login activation record of the candidate game × the weight coefficient of the interest of the candidate game × the coefficient of the timeliness of the video material of the candidate game.

[0100] Furthermore, the input layer 501 inputs the weighted feature vectors of the candidate game's historical click records, the feature vectors of the candidate game's historical search records, and the feature vectors of the candidate game's historical login activation records into the first hidden layer 502 .

[0101] Accordingly, the first hidden layer 502 may include three neurons, which are respectively connected to the three groups of neurons in the input layer 501 .

[0102] The neurons in the first hidden layer 502 are used to perform weighted averaging of the feature vectors input by a group of input layer neurons connected to it based on the feature values ​​of the timeliness of user behavior in the feature vectors. Compared with the neurons in the first hidden layer of the existing YoutubeDNN deep recall model that are used to average feature vectors, the neurons in the first hidden layer of the improved YoutubeDNN deep recall model of the present application perform weighted averaging of feature vectors based on the timeliness of user behavior, which can make the recalled games more timely, thereby further improving the recommendation effect.

[0103] Specifically, the first neuron of the first hidden layer 502 is used to perform weighted average calculation on the feature vector of the historical click record of the candidate game based on the feature value of the timeliness of the user's click behavior in the feature vector;

[0104] The second neuron of the first hidden layer 502 is used to perform weighted average calculation on the feature vector of the historical search record of the candidate game based on the feature value of the timeliness of the user's search behavior in the feature vector;

[0105] The third neuron of the first hidden layer 502 is used to perform weighted average calculation on the feature vector of the historical login activation record of the candidate game based on the feature value of the timeliness of the user's login activation behavior in the feature vector.

[0106] Accordingly, the second hidden layer 503 may include three neurons, which are respectively connected to the three neurons of the first hidden layer 502 .

[0107] The neurons of the second hidden layer 503 are used to form a vector with the weighted average calculation results output by the neurons of the first hidden layer 502 connected thereto.

[0108] Specifically, the first neuron of the second hidden layer 503 is used to output the weighted average calculation result output by the first neuron of the first hidden layer 502 into the clicked game vector to the activation (ReLU) layer 504;

[0109] The second neuron of the second hidden layer 503 is used to output the searched game vector composed of the weighted average calculation result output by the second neuron of the first hidden layer 502 to the activation (ReLU) layer 504;

[0110] The third neuron of the second hidden layer 503 is used to output the played game vector composed of the weighted average calculation result output by the third neuron of the first hidden layer 502 to the activation (ReLU) layer 504 .

[0111] In addition, the second hidden layer 503 may also include other neurons for outputting relevant information of the user, such as the user's age, gender, etc., to the activation (ReLU) layer 504.

[0112] The activation (ReLU) layer 504 may be multiple layers, for example, three layers; the activation (ReLU) layer 504 is used to perform an activation (ReLU) function operation on the input data and then output it.

[0113] The normalization (softmax) layer 505 is used to normalize the input data and then output it.

[0114] Among them, the activation (ReLU) layer 504 and the normalization (softmax) layer 505 have the same functions as the activation (ReLU) layer and the normalization (softmax) layer in the existing YoutubeDNN deep recall model, respectively.

[0115] This application uses an improved YoutubeDNN deep recall model as the deep recall model of the recall layer 305. Compared with the existing YoutubeDNN deep recall model, it introduces historical play information (i.e., game login activation information) and inputs it into the model as a separate set of feature vectors. In addition, considering that the user’s interest in games that have interacted in the past is different, the average of historical interactive games is changed to a weighted average based on interest and timeliness characteristics. Thus, the improved YoutubeDNN deep recall model solves the problem of users who have played games but have no viewing and search records losing personalization under the original model. At the same time, the introduction of timeliness and interest has greatly improved the recommendation effect.

[0116] Step S404: The preliminary screening results of the games are subjected to a deep sorting model to obtain a sorting result of each game in the preliminary screening results.

[0117] Specifically, the initial screening results of the games output by the recall layer 305 pass through the filtering layer 306 and then reach the sorting layer 307 , and are sorted by the deep sorting model of the sorting layer 307 to obtain the sorting results of the games.

[0118] As a more optimal implementation, a DeepFM (Deep Factor Machine) model may be used as a deep ranking model in the ranking layer 307 to obtain the ranking results of the games according to the interest level of each game in the initial screening results.

[0119] Compared with the YoutubeDNN (YouTube Deep Neural Network) deep sorting model used in the sorting layer of existing intelligent game recommendation systems, the DeepFM (Deep Factorization Machine) model can process high-dimensional features that often appear in recommendation problems in different regions to reduce the complexity of the algorithm. A deep neural network is added to combine and analyze the processing results of each region of high-order feature combinations, and FM functions are added to analyze low-order feature combinations and then merge the outputs.

[0120] Although the DeepFM recommendation algorithm of the DeepFM model has the above advantages, and does not require pre-training FM to obtain latent vectors, does not require artificial feature engineering, and can simultaneously learn low-order and high-order combined features; however, its main shortcomings are reflected in two aspects: First, this method is not friendly to low-exposure games, making the overall recommendations narrower and narrower, which is not conducive to the maintenance of the recommendation ecosystem; second, it is weak in dealing with the cold start problem of new games.

[0121] In view of the above-mentioned deficiencies of the existing DeepFM model, in the technical solution of the present application, an improved DeepFM (Deep Factorization Machine) model is used as a deep sorting model in the sorting layer 307 to solve the above-mentioned deficiencies.

[0122] In the process of sorting games, the improved DeepFM (deep factorization machine) model first determines the interest of each game in the initial screening results, and then sorts the games according to the interest of each game. Compared with the DeepFM recommendation algorithm of the existing DeepFM model, the improvement is mainly made in the method of determining the interest of the game; the specific method flow of the improved DeepFM (deep factorization machine) model to determine the interest of the game is as follows: Figure 6 As shown, it includes the following sub-steps S601 to S605:

[0123] Sub-step S601: for each game in the preliminary screening result, determine whether the game is a new game; if so, execute the following sub-step S602; otherwise, execute the following sub-step S603;

[0124] In this sub-step, for each game in the initial screening results, determine whether the difference between the listing time of the game and the current time exceeds the set value; if so, determine that the game is not a new game, and execute the following sub-step S603; otherwise, determine that the game is a new game, and execute the following sub-step S602.

[0125] For example, a game that has been launched within 30 days may be determined to be a new game.

[0126] Sub-step S602: Determine the interest level of the game according to the release time of the game.

[0127] In an exemplary embodiment, the interest level Y of a new game may be determined according to the following formula 1:

[0128] Y=1-t0 / a (Formula 1)

[0129] In formula 1, t0 represents the time when the game is released, which can be expressed in days; a represents the set time value, which can be set to 30 days, for example.

[0130] Sub-step S603: further determine whether the game is a low-exposure game; if so, execute the following sub-step S604; otherwise, execute the following sub-step S605.

[0131] In this sub-step, it is determined whether the game is a low-exposure game based on the exposure rate of the game; if so, the following sub-step S604 is executed; otherwise, the following sub-step S605 is executed.

[0132] Sub-step S604: Determine the interest level of the game based on the historical click-through rate and exposure of the game.

[0133] In an exemplary embodiment, the interest level Y of the low-exposure game may be determined according to the following formula 2:

[0134] Y=s×(1 / w) (Formula 2)

[0135] In Formula 2, s represents the normalized historical click rate of the game, and w represents the normalized exposure of the game.

[0136] Sub-step S605: Determine the interest level of the game according to the difference between the current time and the user behavior time of the game.

[0137] In an exemplary embodiment, the interest level Y of the low-exposure game may be determined according to the following formula 3:

[0138] Y=0.5×(1 / (t-t1)) (Formula 3)

[0139] In Formula 3, t represents the current time, and t1 represents the user behavior time of the game.

[0140] When building the improved YouTubeDNN deep recall model, if a player interacts with the game, the player's interest in the game will be set to 1 (i.e., a positive sample). This method has three shortcomings: first, this method does not take into account that the user's interest in the game changes over time; second, this method is not friendly to those high-quality games with low exposure and high clicks; finally, considering the business, new games should get more exposure, but this method does not take this into consideration.

[0141] It can be seen from the above algorithm for determining the interest of a game that, compared to the DeepFM recommendation algorithm of the existing DeepFM model, the DeepFM model of the present application not only takes into account the decay of user interest over time when determining the interest of a game, but also takes into account new games and high-quality games with low exposure and high click-through rates, thereby further improving the recommendation effect.

[0142] Step S405: recommending games according to the ranking results of the games.

[0143] In an exemplary embodiment, the application service layer 308 of the hybrid model-based game recommendation system in the server 2 generates game recommendation results based on the game ranking results output by the ranking layer 307, and sends them to the homepage of the game center opened on the terminal device 6 for presentation to the user.

[0144] If in step S401, a game with a historical click record, a historical search record, or a historical login activation record of the user is obtained as a candidate game, then the game recommended in this step is obtained based on the behavioral characteristics of the user, including the historical click records, the historical search records, or the historical login activation behavioral characteristics, that is, the game recommendation result reflects the games that the user may be interested in.

[0145] If in step S401, games with historical click records, historical search records, or historical login activation records of all users are obtained as candidate games, then the games recommended in this step are recommendation results that can reflect the games that all users are interested in.

[0146] The game recommendation method based on a hybrid model in the second embodiment of the present application, when using the deep recall model for recall, in addition to using the historical click records and historical search records of the candidate games, also uses the historical login activation records of the candidate games, thereby improving the accuracy of the recall recommendation and further improving the recommendation effect.

[0147] Preferably, in the game recommendation method based on the hybrid model of the second embodiment of the present application, the deep recall model used is an improved YoutubeDNN deep recall model, which introduces historical play information (i.e., game login activation information) compared to the existing YoutubeDNN deep recall model, and inputs it into the model as a separate tower. In addition, considering that the user's interest in games that have interacted in the past is different, the average of historical interactive games is changed to a weighted average based on interest and timeliness characteristics. Thus, the improved YoutubeDNN deep recall model solves the problem of users who have played games but have no viewing and search records losing personalization under the original model. At the same time, the introduction of timeliness and interest has greatly improved the recommendation effect. Experiments have shown that compared with the existing YoutubeDNN deep recall model, the improved YoutubeDNN deep recall model used in the technical solution of this application can improve the recall recommendation accuracy by 21%.

[0148] More preferably, in the game recommendation method based on the hybrid model of Example 2 of the present application, the sorting model adopted is the DeepFM model. Compared with the YoutubeDNN (YouTube Deep Neural Network) deep sorting model adopted in the sorting layer of the existing intelligent game recommendation system, the DeepFM (Deep Factorization Machine) model can process the high-dimensional features that often appear in the recommendation problems in different regions to reduce the complexity of the algorithm, add a deep neural network to combine and analyze the processing results of each region of the high-order feature combination, and add the FM function to analyze the low-order feature combination and then merge the output.

[0149] Preferably, in the game recommendation method based on a hybrid model of the second embodiment of the present application, the sorting model used is an improved DeepFM model. Compared with the existing DeepFM model, when determining the interest of the game, the automatic score enhancement of the new game in the time attention scenario is used to improve the moderate exposure tilt of DeepFM to new games, and the exposure bias adaptive tilt method is used for low-exposure games to improve the long-tail recommendation ability of the DeepFM model. In other words, the improved DeepFM model adopted in the technical solution of the present application not only takes into account the decay of user interest over time, but also takes into account new games and high-quality games with low exposure and high click-through rates, so that the recommendation effect is further improved. Experiments have shown that the improved DeepFM model adopted in the technical solution of the present application can improve the sorting accuracy by 33% compared with the existing YoutubeDNN (YouTube Deep Neural Network) deep sorting model, and can improve the exposure rate by 36% compared with the existing DeepFM model.

[0150] Embodiment 3

[0151] Figure 7 The schematic diagram of the hardware architecture of a computer device 1300 suitable for implementing a game recommendation method based on a hybrid model according to the third embodiment of the present application is shown schematically. In this embodiment, the computer device 1300 is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions. For example, it can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a rack server, a blade server, a tower server or a cabinet server (including an independent server, or a server cluster composed of multiple servers), etc. Figure 7 As shown, the computer device 1300 includes at least but is not limited to: a memory 1310, a processor 1320, and a network interface 1330 which can communicate with each other via a system bus. Among them:

[0152] The memory 1310 includes at least one type of computer-readable storage medium, and the readable storage medium includes a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory 1310 may be an internal storage module of the computer device 1300, such as a hard disk or a memory of the computer device 1300. In other embodiments, the memory 1310 may also be an external storage device of the computer device 1300, such as a plug-in hard disk equipped on the computer device 1300, a smart memory card (Smart Media Card, referred to as SMC), a secure digital (Secure Digital, referred to as SD) card, a flash card, etc. Of course, the memory 1310 may also include both the internal storage module of the computer device 1300 and its external storage device. In this embodiment, the memory 1310 is generally used to store the operating system and various application software installed on the computer device 1300, such as the program code of the game recommendation method based on the hybrid model, etc. In addition, the memory 1310 can also be used to temporarily store various data that have been output or will be output.

[0153] In some embodiments, the processor 1320 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 1320 is generally used to control the overall operation of the computer device 1300, such as performing control and processing related to data interaction or communication with the computer device 1300. In this embodiment, the processor 1320 is used to run the program code stored in the memory 1310 or process data.

[0154] The network interface 1330 may include a wireless network interface or a wired network interface, and the network interface 1330 is generally used to establish a communication link between the computer device 1300 and other computer devices. For example, the network interface 1330 is used to connect the computer device 1300 to an external terminal through a network, and to establish a data transmission channel and a communication link between the computer device 1300 and the external terminal. The network may be a wireless or wired network such as an intranet, the Internet, the Global System of Mobile communication (GSM), Wideband Code Division Multiple Access (WCDMA), 4G network, 5G network, Bluetooth, Wi-Fi, etc.

[0155] It should be pointed out that Figure 7 Only a computer device having components 1310 - 1330 is shown, but it should be understood that implementing all of the components shown is not a requirement, and more or fewer components may alternatively be implemented.

[0156] In this embodiment, the over-drawing test method based on color feature recognition stored in the memory 1310 can also be divided into one or more program modules and executed by one or more processors (processor 1320 in this embodiment) to complete the embodiment of the present application.

[0157] Embodiment 4

[0158] An embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the game recommendation method based on the hybrid model in the embodiment are implemented.

[0159] In this embodiment, the computer-readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the computer-readable storage medium can be an internal storage unit of a computer device, such as a hard disk or memory of the computer device. In other embodiments, the computer-readable storage medium can also be an external storage device of a computer device, such as a plug-in hard disk equipped on the computer device, a smart memory card (Smart Media Card, referred to as SMC), a secure digital (Secure Digital, referred to as SD) card, a flash card, etc. Of course, the computer-readable storage medium can also include both the internal storage unit of the computer device and its external storage device. In this embodiment, the computer-readable storage medium is generally used to store an operating system and various application software installed on the computer device, such as the program code of the game recommendation method based on the hybrid model in the embodiment. In addition, the computer-readable storage medium can also be used to temporarily store various types of data that have been output or are to be output.

[0160] Obviously, those skilled in the art should understand that the modules or steps of the above-mentioned embodiments of the present application can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order from that herein, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present application are not limited to any specific combination of hardware and software.

[0161] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A game recommendation method based on a hybrid model, characterized in that: include: Inputting feature vectors of the candidate game's historical click records, historical search records, and historical login activation records into a deep recall model, and obtaining a preliminary screening result of the game through the deep recall model; The preliminary screening results of the games are subjected to a deep sorting model to obtain a sorting result of each game in the preliminary screening results; Recommending games according to the ranking results of the games; The deep sorting model is specifically an improved deep factorization machine DeepFM model; and The improved DeepFM model obtains the ranking results of the games according to the interest level of each game in the initial screening results; The process of obtaining the ranking results of each game in the preliminary screening results by using a deep ranking model includes: Determining the interest level of each game in the preliminary screening results; Sort the games by their interest level; Determining the interest level of each game in the preliminary screening results specifically includes: For each game in the preliminary screening result, when the game is judged to be a new game, the interest level of the game is determined according to the release time of the game; If the game is not a new game and is a low-exposure game, determining the interest level of the game based on the historical click-through rate and exposure of the game; If the game is not a new game and is not a low-exposure game, the interest level of the game is determined based on the difference between the current time and the user behavior time of the game.

2. The method according to claim 1, characterized in that The deep recall model is specifically an improved YouTube deep neural network deep recall model, and its model structure is based on the YouTube deep neural network deep recall model, and further expands the following neurons: The input layer neurons are expanded from two groups to three groups; The first hidden layer neurons are expanded from two to three, connecting three groups of input layer neurons respectively; The neurons in the second hidden layer are expanded from two to three, which are connected to the three neurons in the first hidden layer respectively.

3. The method according to claim 2, characterized in that In the improved YouTube deep neural network deep recall model, The first, second and third groups of neurons in the input layer are used to input feature vectors of the historical click records, historical search records and historical login activation records of the candidate games respectively.

4. The method according to claim 3, characterized in that The neurons of the first hidden layer are used to perform weighted average calculation on the feature vectors input by a group of input layer neurons connected thereto according to the feature values ​​of the timeliness of the user behavior in the feature vectors.

5. The method according to claim 4, characterized in that The neurons in the second hidden layer are used to form a vector of the weighted average calculation results of the outputs of the neurons in the first hidden layer connected thereto.

6. The method according to claim 1, characterized in that The feature vectors of the candidate game's historical click / historical search / historical login activation records are also weighted according to the following parameters: The interest level of the candidate games, and the timeliness of the video materials of the candidate games.

7. A game recommendation system based on a hybrid model, characterized in that: include: The recall layer is used to input the feature vectors of the candidate game's historical click records, historical search records, and historical login activation records into the deep recall model, and obtain the initial screening results of the game through the deep recall model; A sorting layer, used to obtain the sorting results of each game in the preliminary screening results by applying the preliminary screening results of the games to a deep sorting model; An application service layer, used for recommending games according to the ranking results of the games; Wherein, the sorting layer is also used for: The deep sorting model is specifically an improved deep factorization machine DeepFM model; and The improved DeepFM model obtains the ranking results of the games according to the interest level of each game in the initial screening results; The process of obtaining the ranking results of each game in the preliminary screening results by using a deep ranking model includes: Determining the interest level of each game in the preliminary screening results; Sort the games by their interest level; Determining the interest level of each game in the preliminary screening results specifically includes: For each game in the preliminary screening result, when the game is judged to be a new game, the interest level of the game is determined according to the release time of the game; If the game is not a new game and is a low-exposure game, determining the interest level of the game based on the historical click-through rate and exposure of the game; If the game is not a new game and is not a low-exposure game, the interest level of the game is determined based on the difference between the current time and the user behavior time of the game.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it is used to implement the steps of the game recommendation method based on the hybrid model as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which can be executed by at least one processor to enable the at least one processor to perform the steps of the game recommendation method based on the hybrid model as described in any one of claims 1 to 6.

Images