Method for extracting user interests in field of game personalized recommendation

Abstract

The invention provides a method for extracting user interests in the field of game personalized recommendation, which comprises the following steps of: obtaining a user historical game behavior sequence, and generating a historical game vector list; determining a to-be-predicted game vector; carrying out vector splicing processing, obtaining an importance degree weight representing the association degree of the historical game vector to the game vector to be predicted through processing of a multi-layer perceptron, and carrying out normalization on the importance degree weight; and performing weighted summation on the historical game vector list and the normalized importance weight list to obtain a user interest vector. When the user interest vector is obtained, the relevance between the user historical game behavior and the to-be-estimated game is considered, and the user interest change can be dynamically captured. According to the method, the difference of the importance of each behavior in the historical game behaviors of the user is considered, and the importance weight of each behavior is calculated, so that each historical game behavior accurately plays a role in the interest degree of the user, the accuracy of the interest vector of the user is improved, and an accurate user game recommendation list is favorably formed.

Description

Technical Field

[0001] The present invention belongs to the field of information processing technology, and in particular relates to a method for extracting user interests in the field of personalized game recommendations. Background Art

[0002] In the field of personalized game recommendations, deep neural network models are widely used to analyze users' historical behavior data to determine their gaming preferences; based on their gaming preferences, a list of recommended games for them is then determined.

[0003] The traditional method mainly includes: 1) mapping each game in the user's historical behavior sequence into a fixed-length game vector to obtain a game vector list. 2) calculating the mean vector of all game vectors in the game vector list, which is the user's interest vector; and then determining the user's game recommendation list based on the user's interest vector.

[0004] The traditional method has the following shortcomings: since the user interest vector finally obtained is the mean vector of all game vectors corresponding to the historical behavior sequence, each historical behavior plays the same role in calculating the user's game preference, and the different roles of each behavior in the historical behavior sequence are not considered, which reduces the accuracy of the user interest vector and makes it difficult for the user interest vector to accurately reflect the user's game interest, thereby reducing the accuracy of the game recommendation results. Summary of the invention

[0005] In view of the defects of the prior art, the present invention provides a method for extracting user interests in the field of personalized game recommendation, which can effectively solve the above problems.

[0006] The technical solution adopted by the present invention is as follows:

[0007] The present invention provides a method for extracting user interests in the field of personalized game recommendation, comprising the following steps:

[0008] Step S1, collecting the game IDs of all games in the game recommendation platform; for each game ID, generating a unique corresponding game vector of uniform length, thereby obtaining a game vector set formed by the game vectors of all games;

[0009] Step S2, for the current user to be analyzed, obtain a user historical game behavior sequence; the user historical game behavior sequence is n user historical game behaviors arranged in chronological order, and each user historical game behavior is the user's access or download behavior of a game with a certain game ID; therefore, extract the n user historical game behaviors to obtain n historical game IDs arranged in chronological order;

[0010] According to the game vector set of step S1, each historical game ID is mapped to a corresponding historical game vector, thereby generating a historical game vector list, which is expressed as: in, Respectively represent the first historical game vector, the second historical game vector, …, the nth historical game vector;

[0011] Step S3, for the current game to be predicted, obtain the game ID of the game to be predicted; according to the game vector set of step S1, map the game ID of the game to be predicted to the corresponding game vector to obtain the game vector to be predicted

[0012] Step S4, for the historical game vector list Each historical game vector in i=1,2,...,n, all perform vector splicing processing to obtain the game splicing vector The method is:

[0013] Comprehensive historical game vector and the game vector to be predicted Get the game splicing vector Among them, the concat function represents the vector concatenation operation;

[0014] Step S5, for each game splicing vector After processing by a multi-layer perceptron, we get a vector representing the historical game. Treat the prediction game vector The importance weight of the association degree a i ;

[0015] Step S6, therefore, for the historical game vector list Each historical game vector in The corresponding importance weight a is obtained i , thus we get the importance weight list [a1,a2,…,a n ];

[0016] Step S7: For the importance weight list [a1, a2, ..., a n ] are normalized to obtain a normalized importance weight list, expressed as [a1', a'2, ..., a' n ]; where a1'+a'2+…+a' n =1;

[0017] Step S8: Use the following formula to list the historical game vectors: and the normalized importance weight list [a1',a'2,…,a' n ] to perform weighted summation:

[0018]

[0019] Thus, the user's interest vector for the current game to be predicted is obtained.

[0020] Preferably, in step S1, each game ID generates a corresponding game vector through a random generation method.

[0021] Preferably, step S5 is specifically:

[0022] Step S5.1, the multilayer perceptron has K layers of perceptrons, and each layer of perceptrons has a weight matrix and a bias vector; therefore, the weight matrices of the K layers of perceptrons are W1, W2, …, W K ; The bias vectors of the K-layer perceptron are

[0023] in:

[0024] For the weight matrix W j , j∈[1,K], whose height and width are m respectively j and n j , and the following conditions are met: m1=4D, n K =1,n j =m j+1 ; Where D is the length of each game vector in step S1;

[0025] For the bias vector j∈[1,K], whose length is n j ;

[0026] Step S5.2, the multilayer perceptron uses the activation function σ; the game concatenation vector obtained in step S4 is It is expressed as:

[0027] Then for the j-th layer perceptron, j∈[1,K], the input game splicing vector Using the formula: Output game stitching vector

[0028] Step S5.3, therefore, after being processed by the K-layer perceptron, the game splicing vector is output Game mosaic vector Is a vector of length 1, the value contained in the vector is the historical game vector The importance weight a i .

[0029] Preferably, it also includes:

[0030] Step S9, establish another multi-layer perceptron, called user interest score prediction multi-layer perceptron; transform the user interest vector and the game vector to be predicted To splice, the method is: Get the concatenation vector Concatenate vectors The length of is 2D;

[0031] Step S10, concatenate the vector Input to the user interest score prediction multilayer perceptron, the user interest score prediction multilayer perceptron output represents the user interest vector Treat the prediction game vector The higher the score, the more interested the current user is in the predicted game vector The higher the level of interest.

[0032] The method for extracting user interests in the field of personalized game recommendation provided by the present invention has the following advantages:

[0033] 1) The present invention vectorizes the user's historical game behavior and the game to be predicted into numerical form, thereby obtaining a game ID in vector form, which can then be used as an input for models such as a multi-layer perceptron;

[0034] 2) When acquiring the user interest vector, the present invention considers the correlation between the user's historical game behavior and the game to be estimated, which is conducive to dynamically capturing the changes in user interests.

[0035] 3) The present invention takes into account the differences in the importance of each behavior in the user's historical game behavior, and can automatically calculate the importance weight of each behavior, so that each historical game behavior accurately plays a role in the user's interest level, improves the accuracy of the user's interest vector, and is conducive to forming an accurate user game recommendation list. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A flow chart of a method for extracting user interests in the field of personalized game recommendation provided by the present invention;

[0037] Figure 2 This is a flow chart of weight value calculation provided by the present invention. DETAILED DESCRIPTION

[0038] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention more clearly understood, the present invention 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 invention and are not used to limit the present invention.

[0039] The present invention provides a method for extracting user interests in the field of personalized game recommendation, which can more accurately predict the user's preference for different games to be predicted, thereby forming an accurate user game recommendation list. Figure 1-2 , including the following steps:

[0040] Step S1, game vector initialization:

[0041] Collect the game IDs of all games in the game recommendation platform; for each game ID, generate a unique corresponding game vector of uniform length, thereby obtaining a game vector set formed by the game vectors of all games;

[0042] In this step, each game ID generates a corresponding game vector through random generation, and the length of the game vector is D.

[0043] Step S2, historical game ID vectorization:

[0044] For the current user to be analyzed, a user historical game behavior sequence is obtained; the user historical game behavior sequence is n user historical game behaviors arranged in chronological order, and each user historical game behavior is the user's access or download behavior of a game with a certain game ID; therefore, the n user historical game behaviors are extracted to obtain n historical game IDs arranged in chronological order;

[0045] According to the game vector set of step S1, each historical game ID is mapped to a corresponding historical game vector, thereby generating a historical game vector list, which is expressed as: in, Respectively represent the first historical game vector, the second historical game vector, …, the nth historical game vector;

[0046] Step S3, for the current game to be predicted, obtain the game ID of the game to be predicted; according to the game vector set of step S1, map the game ID of the game to be predicted to the corresponding game vector to obtain the game vector to be predicted

[0047] Step S4, for the historical game vector list Each historical game vector in i=1,2,...,n, all perform vector splicing processing to obtain the game splicing vector The method is:

[0048] Comprehensive historical game vector and the game vector to be predicted Get the game splicing vector Among them, the concat function represents the vector concatenation operation; the length of the concatenated vector is 4D.

[0049] Step S5, for each game splicing vector After processing by a multi-layer perceptron, we get a vector representing the historical game. Treat the prediction game vector The importance weight of the association degree a i ;

[0050] Step S5 is as follows: use a multilayer perceptron (MLP) to map each game splicing vector Processing is performed to obtain the weight value corresponding to each historical game vector in the user's historical game behavior sequence. The specific method is as follows:

[0051] Step S5.1, the multilayer perceptron (MLP) has K layers of perceptrons, and each layer of perceptrons has a weight matrix and a bias vector; therefore, the weight matrices of the K layers of perceptrons are W1, W2, …, W K ; The bias vectors of the K-layer perceptron are

[0052] in:

[0053] For the weight matrix W j , j∈[1,K], whose height and width are m respectively j and n j , and the following conditions are met: m1=4D, n K =1,n j =m j+1 ; Where D is the length of each game vector in step S1;

[0054] For the bias vector j∈[1,K], whose length is n j ;

[0055] During the sample training process, the weight matrix W j and the bias vector Random initialization is used for assignment, and gradient descent is used for updating during the training process until a trained multi-layer perceptron is obtained. In this step, the trained multi-layer perceptron is used.

[0056] Step S5.2, the multilayer perceptron uses an activation function σ; the purpose of using the activation function is to introduce nonlinear factors into the multilayer perceptron so that it can complete more complex tasks; the game splicing vector obtained in step S4 is It is expressed as:

[0057] Then for the j-th layer perceptron, j∈[1,K], the input game splicing vector Using the formula: Output game stitching vector therefore, and They are the output results of the perceptrons of the previous layer and the current layer, respectively, with lengths of n j-1 and n j . It is the input of the first layer of perceptron, with a length of 4D.

[0058] Step S5.3, therefore, after being processed by the K-layer perceptron, the game splicing vector is output Game mosaic vector Is a vector of length 1, the value contained in the vector is the historical game vector The importance weight a i .

[0059] Step S6, therefore, for the historical game vector list Each historical game vector in The corresponding importance weight a is obtained i , thus we get the importance weight list [a1,a2,…,a n ];

[0060] Step S7, the importance weight list [a1, a2, ..., a n ] are normalized to obtain a normalized importance weight list, expressed as [a1', a'2, ..., a' n ]; where a1'+a'2+…+a' n =1;

[0061] Step S8: Use the following formula to list the historical game vectors: and the normalized importance weight list [a1',a'2,…,a' n ] to perform weighted summation:

[0062]

[0063] Thus, the user's interest vector for the current game to be predicted is obtained.

[0064] Also includes:

[0065] Step S9, establish another multi-layer perceptron, called user interest score prediction multi-layer perceptron; transform the user interest vector and the game vector to be predicted To splice, the method is: Get the concatenation vector Concatenate vectors The length of is 2D;

[0066] Step S10, concatenate the vector Input to the user interest score prediction multilayer perceptron, the user interest score prediction multilayer perceptron output represents the user interest vector Treat the prediction game vector The higher the score, the more interested the current user is in the predicted game vector The higher the level of interest.

[0067] The method for extracting user interests in the field of personalized game recommendation provided by the present invention has the following advantages:

[0068] 1) The present invention vectorizes the user's historical game behavior and the game to be predicted into numerical form, thereby obtaining a game ID in vector form, which can then be used as an input for models such as a multi-layer perceptron;

[0069] 2) When acquiring the user interest vector, the present invention considers the correlation between the user's historical game behavior and the game to be estimated, which is conducive to dynamically capturing the changes in user interests.

[0070] 3) The present invention takes into account the differences in the importance of each behavior in the user's historical game behavior, and can automatically calculate the importance weight of each behavior, so that each historical game behavior accurately plays a role in the user's interest level, improves the accuracy of the user's interest vector, and is conducive to forming an accurate user game recommendation list.

[0071] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.

Claims

1. A method for extracting user interests in the field of personalized game recommendations, characterized in that: The following steps are involved: Step S1, collecting the game IDs of all games in the game recommendation platform; for each game ID, generating a unique corresponding game vector of uniform length, thereby obtaining a game vector set formed by the game vectors of all games; Step S2, for the current user to be analyzed, obtain a user historical game behavior sequence; the user historical game behavior sequence is n user historical game behaviors arranged in chronological order, and each user historical game behavior is the user's access or download behavior of a game with a certain game ID; therefore, extract the n user historical game behaviors to obtain n historical game IDs arranged in chronological order; According to the game vector set of step S1, each historical game ID is mapped to a corresponding historical game vector, thereby generating a historical game vector list, which is expressed as: in, Respectively represent the first historical game vector, the second historical game vector, …, the nth historical game vector; Step S3, for the current game to be predicted, obtain the game ID of the game to be predicted; according to the game vector set of step S1, map the game ID of the game to be predicted to the corresponding game vector to obtain the game vector to be predicted Step S4, for the historical game vector list Each historical game vector in i=1,2,...,n, all perform vector splicing processing to obtain the game splicing vector The method is: Comprehensive historical game vector and the game vector to be predicted Get the game splicing vector Among them, the concat function represents the vector concatenation operation; Step S5, for each game splicing vector After processing by a multi-layer perceptron, we get a vector representing the historical game. Treat the prediction game vector The importance weight of the association degree a i ; Step S6, therefore, for the historical game vector list Each historical game vector in The corresponding importance weight a is obtained i , thus we get the importance weight list [a1,a2,…,a n ]; Step S7: For the importance weight list [a1, a2, ..., a n ] are normalized to obtain a normalized importance weight list, expressed as [a'1,a'2,…,a' n ]; among them, a'1+a'2+…+a' n =1; Step S8: Use the following formula to list the historical game vectors: and the normalized importance weight list [a'1,a'2,…,a' n ] to perform weighted summation: Thus, the user's interest vector for the current game to be predicted is obtained.

2. A method for extracting user interests in the field of personalized game recommendation according to claim 1, characterized in that: In step S1, each game ID generates a corresponding game vector through random generation.

3. A method for extracting user interests in the field of personalized game recommendation according to claim 1, characterized in that: Step S5 is specifically as follows: Step S5.1, the multilayer perceptron has K layers of perceptrons, and each layer of perceptrons has a weight matrix and a bias vector; therefore, the weight matrices of the K layers of perceptrons are W1, W2, …, W K ; The bias vectors of the K-layer perceptron are in: For the weight matrix W j , j∈[1,K], whose height and width are m respectively j and n j , and the following conditions are met: m1=4D, n K =1,n j =m j+1 ; Where D is the length of each game vector in step S1; For the bias vector j∈[1,K], whose length is n j ; Step S5.2, the multilayer perceptron uses the activation function σ; the game concatenation vector obtained in step S4 is It is expressed as: Then for the j-th layer perceptron, j∈[1,K], the input game splicing vector Using the formula: Output game stitching vector Step S5.3, therefore, after being processed by the K-layer perceptron, the game splicing vector is output Game mosaic vector Is a vector of length 1, the value contained in the vector is the historical game vector The importance weight a i .

4. A method for extracting user interests in the field of personalized game recommendation according to claim 1, characterized in that: Also includes: Step S9, establish another multi-layer perceptron, called user interest score prediction multi-layer perceptron; transform the user interest vector and the game vector to be predicted To splice, the method is: Get the concatenation vector Concatenate vectors The length of is 2D; Step S10, concatenate the vector Input to the user interest score prediction multilayer perceptron, the user interest score prediction multilayer perceptron output represents the user interest vector Treat the prediction game vector The higher the score, the more interested the current user is in the predicted game vector. The higher the level of interest.

Images