A game recommendation method based on big data

Abstract

The application relates to the technical field of network games, in particular to a game recommendation method based on big data, which comprises the following steps: constructing a game recommendation heterogeneous graph based on player and game history data, wherein the nodes in the graph are players and games; calculating the correlation distance between the nodes; inputting the player and game history data and the correlation distance between the nodes into a GP neural network, outputting the feature vectors of the player and game history data through an embedding layer; obtaining the update weight when each feature vector is convolved, outputting the behavior vectors of the player and game history data through a GP neural network aggregation layer according to the update weight and each feature vector; and outputting the recommendation result of the game through each behavior vector through a GP neural network output layer. The application considers the inherent attribute connection and social attribute connection between the game and the player, so that the game recommendation result is more in line with the demand of the player.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network games, and particularly relates to a game recommendation method based on big data. BACKGROUND

[0002] With the development of computer hardware and software and the improvement of people's living standards, electronic games have become the entertainment way of most people and have brought a lot of economic value and cultural value to the society. The game industry is booming, there are millions of games on the market, and the number of games is increasing by 1-3 million per year, and a large amount of big data is generated on various platforms. It is a tedious thing for people to find their favorite games. The recommendation system is an automated pre-selection technology to solve information overload. The current mainstream method is to develop a game popular list for players to choose their favorite games.

[0003] In the prior art, the game recommendation of major game platforms and application stores is recommended by big data, such as the popularity, attention and click volume of games, but these games may not meet the entertainment needs of all people, so the game itself needs to be concerned, and the needs of the players themselves also need to be concerned. Therefore, these games recommended by big data screening do not achieve good electronic game recommendation effect, causing waste of recommended resources and poor user experience for users. SUMMARY

[0004] In order to solve the above technical problems, the purpose of the present application is to provide a game recommendation method based on big data, and the technical solution adopted is as follows:

[0005] A game recommendation heterogeneous graph is constructed based on player and game history data, and the nodes in the graph are players and games; the correlation distance between the nodes is calculated, specifically including: the correlation distance between two player nodes is calculated based on the duration of common games between the two players; the correlation distance between the player node and the game node is calculated based on the duration of the player playing the game and the consumption amount in the game; the correlation distance between two game nodes is calculated based on the number of times of being listed on the list together;

[0006] The player and game history data and the correlation distance between the nodes are input into the GP neural network, and the feature vectors of the player and game history data are output through the embedding layer; the feature vectors are player attribute vectors, game label vectors, player preference vectors and game audience vectors;

[0007] The update weight of the GP neural network aggregation layer when performing convolution on each feature vector is obtained by using the player attribute vector, the game label vector, the player preference vector and the game audience vector, and the correlation distance between nodes; the behavior vector of the player and the game history data is output by the GP neural network aggregation layer according to the update weight and each feature vector; and the recommendation result of the game is output by the GP neural network output layer through each behavior vector.

[0008] Preferably, the game recommendation heterogeneous graph constructed based on the player and game history data is specifically:

[0009] The history data related to the player and the game is obtained, each player and each game is taken as a node respectively, and a graph structure is constructed; wherein, according to the history data, two game nodes on the same list are connected, two player nodes with common game records are connected, and the player node and the game node with game duration are connected, to form a game recommendation heterogeneous graph.

[0010] Preferably, the feature vector of the player and the game history data output by the embedding layer is specifically:

[0011] The game label vector is obtained according to the attribute label and the basic attribute of the game in the game history data, and the player attribute vector is obtained according to the basic information of the player in the player history data; the player preference vector is obtained according to the mean value of the label vectors of the top 5 games in terms of duration in all games played by the player in the player history data; and the game audience vector is obtained according to the mean value of the player attribute vectors of the top 30% of players in terms of game duration in all players of any one game in the game history data.

[0012] Preferably, the specific method for obtaining the update weight of the GP neural network aggregation layer when performing convolution on each feature vector is:

[0013] The update weight corresponding to the player attribute vector is obtained according to the similarity of the player attribute vectors of two players and the correlation distance between the two player nodes; the update weight corresponding to the game label vector is obtained according to the similarity of the game label vectors of two games and the correlation distance between the two game nodes; the update weight corresponding to the player preference vector is obtained according to the similarity of the player preference vector corresponding to the player and the game label vector corresponding to the game and the correlation distance between the player node and the game node; and the update weight corresponding to the game audience vector is obtained according to the similarity of the player attribute vector corresponding to the player and the game audience vector corresponding to the game and the correlation distance between the player node and the game node.

[0014] Preferably, the method for obtaining the update weight corresponding to the player attribute vector is specifically:

[0015]

[0016] wherein W ij denotes the update weight of the player attribute vector corresponding to the player node i, Pro i denotes the player attribute vector corresponding to the player node i, Pro j denotes the player attribute vector corresponding to the player node j, D ij denotes the correlation distance between the players corresponding to the player node i and the player node j, U denotes a set of all player nodes connected to the player node i on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

[0017] Preferably, the method for obtaining the update weight corresponding to the game label vector is specifically:

[0018]

[0019] wherein W ab denotes the update weight of the game label vector corresponding to the game node a, Tag a denotes the game label vector corresponding to the game node a, Tag b denotes the game label vector corresponding to the game node a, Y ab denotes the correlation distance between the games corresponding to the game node a and the game node b, V denotes a set of all game nodes connected to the game node a on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

[0020] Preferably, the method for obtaining the update weight corresponding to the player preference vector is specifically:

[0021]

[0022] wherein W ib denotes the update weight of the player preference vector corresponding to the player node i, Fav i denotes the player preference vector corresponding to the player node i, Tag b denotes the game label vector corresponding to the game node a, L ib denotes the correlation distance between the player node i and the game node b, R denotes a set of all game nodes connected to the player node i on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

[0023] Preferably, the method for obtaining the update weight corresponding to the game audience vector is specifically:

[0024]

[0025] wherein W aj denotes the update weight of the game audience vector corresponding to the game node a, Aim aPro represents a game audience vector corresponding to the game node a j L represents a player attribute vector corresponding to the player node j aj G represents a correlation distance between the game node a and the player node j, G represents a set of all player nodes connected to the game node a on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

[0026] Preferably, the GP neural network comprises an embedding layer, an aggregation layer and an output layer; wherein the embedding layer is used to output feature vectors of the player and the game history data; the aggregation layer is used to update each feature vector to output behavior vectors of the player and the game history data; and the output layer is used to process each behavior vector to output a game recommendation result.

[0027] The embodiments of the present application have at least the following beneficial effects:

[0028] The present application constructs a game recommendation heterogeneous graph containing two kinds of nodes and three kinds of relationships, and obtains the correlation distance between the nodes on the graph, takes the history data of the player and the game as input, uses the GP neural network embedding layer to output the feature vectors of the player and the game history data; the aggregation layer updates each feature vector to output the behavior vectors of the player and the game history data; and the output layer processes each behavior vector to output a game recommendation result. The present application fully considers the inherent attribute relationship between the game and the player and the social attribute relationship, so that the game recommendation result is more in line with the needs of the player, and the player user has a better user experience. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0030] Figure 1 is a method flowchart of a game recommendation method based on big data. DETAILED DESCRIPTION

[0031] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined invention purpose, the following combines the drawings and the preferred embodiments to specifically describe the specific implementation, structure, features and effects of the game recommendation method based on big data according to the present application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0033] The following will specifically describe a specific scheme of a game recommendation method based on big data provided by the application in combination with the drawings.

[0034] Embodiment:

[0035] Please refer to Figure 1 , which shows a method flowchart of a game recommendation method based on big data provided by an embodiment of the application, and the method comprises the following steps:

[0036] Step one, constructing a game recommendation heterogeneous graph based on player and game history data, and the nodes in the graph are players and games; calculating the correlation distance between nodes, specifically including: calculating the correlation distance between two player nodes based on the duration of common games between the two players; calculating the correlation distance between the player node and the game node based on the duration of the player playing the game and the consumption amount in the game; calculating the correlation distance between two game nodes based on the number of times the two games are listed together.

[0037] Firstly, it is determined that the customized game recommendation to all players on a certain game platform is the scene, wherein the game platform can be Steam, WeGame, Epic, etc., and the implementer can set it according to the actual situation. Then, the historical data of players and games on the game platform is obtained. At the same time, the games involved in this embodiment are all network electronic games.

[0038] Among them, the historical data of players and games includes: the total duration of all players playing all games; the duration of any two players playing games together; the duration of all players playing each game respectively; obtaining the interaction data in the game, and obtaining the intimacy attribute between two players according to the interaction data; the total consumption amount of all players in all games; the consumption amount of all players in each game respectively; the number of all game lists; the number of times any two games are listed together.

[0039] It should be noted that, due to the increasing social attribute of games, many game companies will count a game statistical data for players every certain period of time, so in this embodiment, the historical data of players and games in a set time period needs to be obtained, and the value of the set time period can be set to 1 year, and the implementer can also set it according to the actual situation. At the same time, the interaction data in the game can specifically represent the number of times of teaming up together in the game played by two players together, or the number of monsters defeated together, or the number of game achievements obtained together, etc., and the implementer can obtain the specific interaction data according to the actual situation.

[0040] Meanwhile, the historical data of the players and the games also includes attribute tags of all the games and basic attributes, where the basic attributes are, for example, a publisher of the game, a price of the game, and a hot degree of the game, and the attribute tags are, for example, role-playing game (RPG), action game, multiplayer, and the like. In the embodiment, the attribute tags of the games on the selected game platform are set to 500 kinds, and the kinds of the attribute tags of the games can be set according to actual conditions, or can be obtained according to the selected game platform. For example, the kinds of the attribute tags of the games on the game platform Steam in 2022 are not more than 700 kinds. The historical data of the players and the games also includes basic information of all the players, and in the embodiment, the basic information includes gender, age, region, device kind, registration time length, and game praise rate of the players, and the basic information can be obtained according to actual conditions.

[0041] It should be noted that there are various game recommendation lists on the game platform, such as monthly best game list, new game list this month, and annual selected action game list, and the like. Various recommendation lists list many games for players to choose. The games on the same list are more likely to be seen at the same time, and the list is selected by the game publisher or the game platform, and has a certain reference value. In addition, on some specific game platforms, the players need to purchase the game to play the game, and the purchase can be free, which is determined according to the specific selected game platform.

[0042] Then, according to the historical data related to the players and the games, each player and each game is taken as a node to construct a graph structure. According to the historical data, two game nodes on the same list are connected, two player nodes with a record of playing the game together are connected, and a player node with a game time length and a game node are connected to form a game recommendation heterogeneous graph. Wherein, two game nodes on the same list at least once are connected by a straight line, two player nodes with a record of playing the game together are connected by a straight line, and a game node and a player node corresponding to a purchase record or a game time length are connected by a straight line. The obtained heterogeneous graph contains two kinds of nodes and three kinds of relationships.

[0043] Finally, the correlation distance between the nodes is calculated. Specifically, according to the time length of two players playing the game together and the intimacy attribute between the two players, the correlation distance between the two player nodes is calculated, which is expressed by the formula as follows:

[0044]

[0045] Wherein, D ijrepresents the correlation distance between player node i and player node j, represents the time length of the game played by the players corresponding to player node i and player node j, represents the total time length of all games played by the player corresponding to player node i, represents the total time length of all games played by the player corresponding to player node j, represents the affinity attribute between the players corresponding to player node i and player node j, k1 and k2 represent weight coefficients, and in the embodiment, the values of the weight coefficients are k1=0.6 and k2=0.4, which can be set according to actual conditions by the implementer.

[0046] MAX(·) represents taking the maximum value, so as to avoid the situation that the time length of all games played by the player is short, and the time length of all games played by the player is simultaneously the time length of the game played with another player, which lacks reference value. The longer the time length of all games played by the two players, and the greater the affinity attribute between the two players, that is, the more intimate the two players are, the greater the correlation distance between the two player nodes.

[0047] According to the time length of any game played by the player and the consumption amount of the player in the game, the correlation distance between the player node and the game node is calculated, which is represented by the formula:

[0048]

[0049] wherein, L ia represents the correlation distance between player node i and game node a, represents the time length of the game corresponding to game node a played by the player corresponding to player node i, represents the total time length of all games played by the player corresponding to player node i, represents the consumption amount of the player corresponding to player node i in the game corresponding to game node a, represents the total consumption amount of the player corresponding to player node i in all games. k3 and k4 are weight coefficients, and in the embodiment, the values of the weight coefficients are k3=0.8 and k4=0.2.

[0050] The longer the time length of the game played by the player, and the greater the consumption amount in the game, the greater the correlation distance between the player and the game, which indicates that the player has a greater degree of love for the game. At the same time, the consumption amount in the game is considered in view of the fact that the game company is inclined to recommend games to players who are willing to pay.

[0051] According to the number of times that two games are simultaneously on the list, the correlation distance between the two game nodes is calculated, which is represented by the formula:

[0052]

[0053] wherein Y ab represents the correlation distance between game node a and game node b, V ab represents the number of times that game node a and game node b correspond to the same game on the same game list, V z represents the number of all game lists. The greater the correlation distance between two game nodes, the more likely that the games corresponding to the two game nodes appear on the same list, and the more likely that the two games are simultaneously seen by players. Moreover, the game list in the game platform is analyzed by a human being, and thus has a certain reference value.

[0054] In step two, the player and game history data and the correlation distance between nodes are input into the GP neural network, and the feature vectors of the player and game history data are output through an embedding layer. The feature vectors are a player attribute vector, a game label vector, a player preference vector, and a game audience vector.

[0055] The GP neural network is constructed, and in this embodiment, the GP neural network includes an embedding layer, an aggregation layer, and an output layer. The player and game history data and the correlation distance between nodes obtained in step one are input into the GP neural network, and first, the feature vectors of the player and game history data are output through the embedding layer. The feature vectors are a player attribute vector, a game label vector, a player preference vector, and a game audience vector.

[0056] Specifically, the game label vector is obtained according to the attribute labels and basic attributes of the games in the game history data. In this embodiment, the game platform contains 500 attribute labels of games, and both the attribute labels and the basic attributes of the games are used as the labels of the games to form the game label vector. For example, the game label vector of game a can be represented as Tag a = [Publisher, Prc, Hot, RPG, Action, Multiplayer, ……], wherein the elements represent the game publisher Publisher, the game price Prc, the game popularity Hot, the role-playing game RPG, the action game Action, and the multiplayer game Multiplayer. The labels of the games are used as the elements to form the game label vector. The implementer can obtain the labels of the games according to the actual situation. Different labels of the games result in different game label vectors. Meanwhile, the implementer can set the arrangement order of the elements, for example, arranging the first letters of the elements in the order of the English alphabet. The value of each element is set to 1 or 2. The value of 1 of an element indicates that the game does not belong to the label corresponding to the element, and the value of 2 of the element indicates that the game belongs to the label corresponding to the element.

[0057] According to the basic information of the players in the player historical data, a player attribute vector is obtained. In this embodiment, the basic information of the players includes the gender, age, region, device type, registration duration and game evaluation rate of the players. The player attribute vector of the player i is denoted as Pro i =[gender is male, gender is female, age, region, device type, registration duration, game evaluation rate]. Different gender players prefer different types of games, and different ages may also prefer different types of games. The gender is divided into two elements to prevent the loss of the proportion of men and women in social relations.

[0058] According to the mean of the game tag vectors of the top 5 games in terms of duration in all games played by the players in the player historical data, a player preference vector is obtained. Specifically, for a player, the game tag vectors corresponding to the top 5 games in terms of game duration in all games played by the player are obtained, and the mean of each element in the game tag vectors corresponding to the 5 games is obtained to obtain the player preference vector corresponding to the player, which roughly reflects the general situation of the player's game preference. The number of games with top duration can be set by the implementer according to actual conditions.

[0059] According to the mean of the player attribute vectors of the top 30% players in terms of game duration in any game in the game historical data, a game audience vector is obtained. Specifically, for a game, the player attribute vectors corresponding to the top 30% players in terms of game duration in the game are obtained, and the mean of each element in the player attribute vectors corresponding to the top 30% players is obtained to obtain the game audience vector corresponding to the game, which roughly reflects the general situation of the game audience group. The number of players with top duration can be set by the implementer according to actual conditions.

[0060] Finally, the feature vectors output by the embedding layer have a unified format, which is convenient for subsequent calculation, and similar vectors will be as close as possible in the parameter space.

[0061] Step three, using the player attribute vector, game tag vector, player preference vector and game audience vector and the correlation distance between nodes, obtaining the update weight of the GP neural network aggregation layer when convolving each feature vector; according to the update weight and each feature vector, outputting the behavior vector of the player and game historical data through the GP neural network aggregation layer; outputting the recommendation result of the game through the GP neural network output layer.

[0062] First, using the correlation distance between each feature vector and the node, the update weight of the GP neural network aggregation layer when convolving each feature vector is obtained, and the update weight is used to update each feature vector.

[0063] Specifically, the friend features of the players are aggregated to update the player attribute vectors of the players, and the friend circle neighborhood of each player is constructed according to the game recommendation heterogeneous graph, the friend circle neighborhood including all player nodes connected to the player on the game recommendation heterogeneous graph. The update weight corresponding to the player attribute vector is obtained according to the similarity of the player attribute vectors corresponding to two players and the correlation distance between the two player nodes, and is expressed by a formula as follows:

[0064]

[0065] wherein, W ij represents the update weight of the player attribute vector corresponding to the player node i, Pro i represents the player attribute vector corresponding to the player node i, Pro j represents the player attribute vector corresponding to the player node j, D ij represents the correlation distance between the player corresponding to the player node i and the player corresponding to the player node j, U represents a set composed of all player nodes connected to the player node i on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

[0066] The update weight of all player attribute vectors corresponding to the player node i is obtained, wherein each player node in the set U composed of all player nodes connected to the player node i on the game recommendation heterogeneous graph corresponds to an update weight, and then convolution operation is performed to update the player attribute vector of the player node, which is expressed by a formula as follows:

[0067] Pro′ i = Pro i + Σ j∈U W ij Pro j

[0068] wherein, Pro′ i represents the updated player attribute vector corresponding to the player node i, Pro i represents the player attribute vector corresponding to the player node i before updating, W ij represents the update weight of the player attribute vector corresponding to the player node i, Pro j represents the player attribute vector corresponding to the player node j, and U represents a set composed of all player nodes connected to the player node i on the game recommendation heterogeneous graph.

[0069] The interaction attributes of the games are aggregated to update the game label vectors corresponding to the games, and the relevant game neighborhood of each game is constructed according to the game recommendation heterogeneous graph, the relevant game neighborhood including all game nodes directly connected to the game on the game recommendation heterogeneous graph. The update weight corresponding to the game label vector is obtained according to the similarity of the game label vectors corresponding to two games and the correlation distance between the two game nodes, and is expressed by a formula as follows:

[0070]

[0071] wherein W ab represents the update weight of the game tag vector corresponding to game node a, Tag a represents the game tag vector corresponding to game node a, Tag b represents the game tag vector corresponding to game node a, Y ab represents the correlation distance between the game corresponding to game node a and game node b, V represents the set of all game nodes connected to game node a on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

[0072] An update weight of all game tag vectors corresponding to game node a is obtained, wherein each game node in the set V of all game nodes connected to game node a on the game recommendation heterogeneous graph corresponds to an update weight, and then convolution operation is performed to update the game tag vector of the game node, which is expressed by the formula as follows:

[0073] Tag′ a = Tag a +∑ b∈V W ab Tag b

[0074] wherein Tag′ a represents the updated game tag vector corresponding to game node a, Tag a represents the game tag vector corresponding to game node a before updating, W ab represents the update weight of the game tag vector corresponding to game node a, Tag b represents the game tag vector corresponding to game node a, and V represents the set of all game nodes connected to game node a on the game recommendation heterogeneous graph.

[0075] The preference features of the players are aggregated to update the player preference vector corresponding to the player node, the game library neighborhood of the player is constructed according to the game recommendation heterogeneous graph, and the game library neighborhood contains all game nodes directly connected to the player. The update weight of the player preference vector is obtained according to the similarity between the player preference vector corresponding to the player and the game tag vector corresponding to the game and the correlation distance between the player node and the game node, which is expressed by the formula as follows:

[0076]

[0077] wherein W ib represents the update weight of the player preference vector corresponding to player node i, Fav i represents the player preference vector corresponding to player node i, Tagb denotes the game label vector corresponding to game node a, L ib denotes the correlation distance between player node i and game node b, R denotes the set of all game nodes connected to player node i on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

[0078] An update weight of all player preference vectors corresponding to player node i is obtained, wherein each game node in the set R of all game nodes connected to player node i on the game recommendation heterogeneous graph corresponds to an update weight, and then a convolution operation is performed to update the player preference vector of the player node, which is expressed by a formula as follows:

[0079] Fav′ i = Fav i +Σ b∈R W ib Tag b

[0080] wherein Fav′ i denotes the updated player preference vector corresponding to player node i, Fav i denotes the player preference vector corresponding to player node i before updating, Tag b denotes the game label vector corresponding to game node b, W ib denotes the update weight of the player preference vector corresponding to player node i, and R denotes the set of all game nodes connected to player node i on the game recommendation heterogeneous graph.

[0081] The audience characteristics of the game are aggregated to update the game audience vector of the game node, a target audience neighborhood of the game is constructed according to the game recommendation heterogeneous graph, and the target audience neighborhood contains all player nodes directly connected to the game. An update weight corresponding to the game audience vector is obtained according to the similarity between the player attribute vector corresponding to the player and the game audience vector corresponding to the game and the correlation distance between the player node and the game node, which is expressed by a formula as follows:

[0082]

[0083] wherein W aj denotes the update weight of the game audience vector corresponding to game node a, Aim a denotes the game audience vector corresponding to game node a, Pro i denotes the player attribute vector corresponding to player node j, L aj denotes the correlation distance between game node a and player node j, G denotes the set of all player nodes connected to game node a on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

[0084] An update weight of all game audience vectors corresponding to the game node a is acquired, wherein each player node in a set G of all player nodes connected to the game node a on the game recommendation heterogeneous graph corresponds to an update weight, and then a convolution operation is performed to update the game audience vector of the game node, which is expressed by a formula as follows:

[0085] Aim' a = Aim a + Π j∈G W aj Pro j

[0086] wherein Aim' a represents the game audience vector corresponding to the updated game node a, Aim a represents the game audience vector corresponding to the game node a before being updated, Pro j represents the player attribute vector corresponding to the player node j, W aj represents the update weight of the game audience vector corresponding to the game node a, and G represents a set of all player nodes connected to the game node a on the game recommendation heterogeneous graph.

[0087] It should be noted that in the embodiment, the similarity between vectors is calculated according to the cosine similarity between vectors, and the adjustment coefficient k is set to adjust the weight of the cosine similarity and the correlation distance. If the value of the correlation distance is larger, the cosine similarity is given a larger weight. The implementer can set the value of the adjustment coefficient k according to the actual situation, and can also select other methods to calculate the similarity between vectors.

[0088] Then, the updated player attribute vector and the player preference vector corresponding to the player node i and the game label vector and the game audience vector corresponding to the game node a are spliced together to obtain the behavior vector corresponding to the player node i and the game node a output by the GP neural network aggregation layer, denoted as X ia = [Pro' i , Tag' a , Fav' i , Aim' a ], and further to obtain the behavior vector corresponding to all player nodes and game nodes, and the recommendation result of the game is output by the GP neural network output layer through each behavior vector.

[0089] Specifically, the success rate of the recommendation behavior of recommending the game a to the player i is output by the GP neural network output layer according to the behavior vector corresponding to the player node i and the game node a, and the success rate of the recommendation behavior of recommending all games to the player i is obtained. The success rate of the recommendation behavior of recommending the game to the player is obtained by the Softmax classifier according to the behavior vector corresponding to the player node and the game node, that is, a player can be regarded as a category, and the probability of each game belonging to the category is obtained by the classifier, and the probability is the success rate of the recommendation behavior.

[0090] The success rates of all games are arranged in descending order to obtain the game recommendation list corresponding to the player i, and then the game recommendation list corresponding to all players can be obtained. The game recommendation list is the recommendation result of the game. The obtained game recommendation list fully considers the inherent attribute relationship and social attribute relationship between the game and the player, so that the game recommendation result is more in line with the needs of the player.

[0091] It should be noted that the GP neural network includes an embedding layer, an aggregation layer and an output layer; wherein the embedding layer is used to output the feature vector of the player and the game historical data; the aggregation layer is used to update each feature vector to output the behavior vector of the player and the game historical data; and the output layer is used to process each behavior vector to output the game recommendation result.

[0092] Finally, the training process of the GP neural network is described. Specifically, the Steam Video Games data set provided by the machine learning data platform Kaggle is used in the experiment, that is, the Steam video game data set. The data collects the historical data of players and games on the game platform Steam. The data set contains a total of 200000 record information, 12393 player users and 5155 games, and then a game recommendation heterogeneous graph is constructed as a verification game recommendation heterogeneous graph. Then the connection between the recommended successful game node and the player node is deleted, and they are labeled as the recommendation success behavior as the positive sample label, and the player and the game without the recommendation behavior are used as the negative sample label to obtain the test game recommendation heterogeneous graph. The game recommendation is performed on the test game recommendation heterogeneous graph, and the verification game recommendation heterogeneous graph is used for verification. The loss function of the network is the cross-entropy loss function, and the optimizer uses Adam. In this way, the GP neural network is preliminarily trained.

[0093] Further, the data of all users on the game platform actually used is obtained, and the above data set generation method is used for training. 80% of the data set is randomly selected as the training set, and 20% is selected as the test set, so as to fine-tune the neural network. In the formal application process, the heterogeneous graph is constructed in real time, the success rate of the recommendation behavior is obtained in the above manner, the priority ranking result is generated, and the recommendation behaviors with the same success rate can be recommended at the same time.

[0094] The above-described embodiments are only used to illustrate the technical solutions of the present application, but not limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A big data-based game recommendation method, characterized by, The method comprises the following steps: Based on the game and game history data, a game recommendation heterogeneous graph is constructed, and the nodes in the graph are players and games; the correlation distance between the nodes is calculated, specifically including: the correlation distance between two player nodes is calculated based on the length of time of common games between the two players; the correlation distance between a player node and a game node is calculated based on the length of time of the player playing the game and the consumption amount in the game; the correlation distance between two game nodes is calculated based on the number of times the two games have been on the list together; The player and game history data and the correlation distance between the nodes are input into the GP neural network, and the feature vectors of the player and game history data are output through the embedding layer; the feature vectors are player attribute vectors, game label vectors, player preference vectors and game audience vectors; The player attribute vectors, game label vectors, player preference vectors and game audience vectors, and the correlation distance between the nodes are used to obtain the update weight of the GP neural network aggregation layer when each feature vector is convolved; the behavior vectors of the player and game history data are output through the GP neural network aggregation layer according to the update weight and each feature vector; the recommendation result of the game is output through the GP neural network output layer through each behavior vector; The specific method for obtaining the update weight of the GP neural network aggregation layer when each feature vector is convolved is: The update weight corresponding to the player attribute vector is obtained according to the similarity of the player attribute vectors corresponding to the two players and the correlation distance between the two player nodes; The update weight corresponding to the game label vector is obtained according to the similarity of the game label vectors corresponding to the two games and the correlation distance between the two game nodes; The update weight corresponding to the player preference vector is obtained according to the similarity of the player preference vector corresponding to the player and the game label vector corresponding to the game and the correlation distance between the player node and the game node; The update weight corresponding to the game audience vector is obtained according to the similarity of the player attribute vector corresponding to the player and the game audience vector corresponding to the game and the correlation distance between the player node and the game node; The specific method for obtaining the update weight corresponding to the player attribute vector is: wherein, represents the update weight of the player attribute vector corresponding to the player node i, represents the player attribute vector corresponding to the player node i, represents the player attribute vector corresponding to the player node j, represents the correlation distance between the players corresponding to the player node i and the player node j, U represents a set of all player nodes connected to the player node i on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

2. The game recommendation method based on big data according to claim 1, characterized in that, The specific method for constructing the game recommendation heterogeneous graph based on the player and game history data is: The history data related to the player and the game is obtained, each player and each game is taken as a node, and a graph structure is constructed; wherein, according to the history data, two game nodes on the same list are connected, two player nodes with common game records are connected, and player nodes and game nodes with game length are connected, to form a game recommendation heterogeneous graph.

3. The game recommendation method based on big data according to claim 1, characterized in that, The specific method for outputting the feature vectors of the player and game history data through the embedding layer is: According to the attribute label and the basic attribute of the game in the game history data, a game label vector is obtained, and according to the basic information of the player in the player history data, a player attribute vector is obtained; according to the mean of the game label vectors of the top 5 games in terms of time length in all the games played by the player in the player history data, a player preference vector is obtained; and according to the mean of the player attribute vectors of the top 30% of players in terms of game time length in any one game in the game history data, a game audience vector is obtained.

4. The game recommendation method based on big data according to claim 1, characterized in that, The method for obtaining the update weight corresponding to the game label vector is specifically as follows: wherein, represents the update weight of the game label vector corresponding to the game node a, represents the game label vector corresponding to the game node a, represents the game label vector corresponding to the game node a, represents the correlation distance between the game corresponding to the game node a and the game corresponding to the game node b, V represents a set of all game nodes connected to the game node a on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

5. The game recommendation method based on big data according to claim 1, characterized in that, The method for obtaining the update weight corresponding to the player preference vector is specifically as follows: wherein, represents the update weight of the player preference vector corresponding to the player node i, represents the player preference vector corresponding to the player node i, represents the game label vector corresponding to the game node a, represents the correlation distance between the player node i and the game node b, R represents a set of all game nodes connected to the player node i on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

6. The game recommendation method based on big data according to claim 1, characterized in that, The method for obtaining the update weight corresponding to the game audience vector is specifically as follows: wherein, represents the update weight of the game audience vector corresponding to the game node a, represents the game audience vector corresponding to the game node a, represents the player attribute vector corresponding to the player node j, represents the correlation distance between the game node a and the player node j, G represents a set of all player nodes connected to the game node a on the game recommendation heterogeneous graph, and k is an adjustment coefficient.

7. The game recommendation method based on big data according to claim 1, characterized in that, The GP neural network comprises an embedding layer, an aggregation layer and an output layer; wherein the embedding layer is used to output the feature vectors of the player and the game history data; the aggregation layer is used to update each feature vector to output the behavior vectors of the player and the game history data; and the output layer is used to process each behavior vector to output a game recommendation result.

Images