CN112569608B - Table game hybrid recommendation method based on multi-source heterogeneous data - Google Patents

Table game hybrid recommendation method based on multi-source heterogeneous data Download PDF

Info

Publication number
CN112569608B
CN112569608B CN202011531148.XA CN202011531148A CN112569608B CN 112569608 B CN112569608 B CN 112569608B CN 202011531148 A CN202011531148 A CN 202011531148A CN 112569608 B CN112569608 B CN 112569608B
Authority
CN
China
Prior art keywords
game
user
matrix
variance
cnn
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202011531148.XA
Other languages
Chinese (zh)
Other versions
CN112569608A (en
Inventor
李绍利
杨传颖
石宝
雷小涵
李亚龙
王成龙
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Inner Mongolia University of Technology
Original Assignee
Inner Mongolia University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Inner Mongolia University of Technology filed Critical Inner Mongolia University of Technology
Priority to CN202011531148.XA priority Critical patent/CN112569608B/en
Publication of CN112569608A publication Critical patent/CN112569608A/en
Application granted granted Critical
Publication of CN112569608B publication Critical patent/CN112569608B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A63—SPORTS; GAMES; AMUSEMENTS
    • A63F—CARD, BOARD, OR ROULETTE GAMES; INDOOR GAMES USING SMALL MOVING PLAYING BODIES; VIDEO GAMES; GAMES NOT OTHERWISE PROVIDED FOR
    • A63F13/00—Video games, i.e. games using an electronically generated display having two or more dimensions
    • A63F13/70—Game security or game management aspects
    • A63F13/79—Game security or game management aspects involving player-related data, e.g. identities, accounts, preferences or play histories
    • A63F13/798—Game security or game management aspects involving player-related data, e.g. identities, accounts, preferences or play histories for assessing skills or for ranking players, e.g. for generating a hall of fame
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/958—Organisation or management of web site content, e.g. publishing, maintaining pages or automatic linking
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • A—HUMAN NECESSITIES
    • A63—SPORTS; GAMES; AMUSEMENTS
    • A63F—CARD, BOARD, OR ROULETTE GAMES; INDOOR GAMES USING SMALL MOVING PLAYING BODIES; VIDEO GAMES; GAMES NOT OTHERWISE PROVIDED FOR
    • A63F2300/00—Features of games using an electronically generated display having two or more dimensions, e.g. on a television screen, showing representations related to the game
    • A63F2300/60—Methods for processing data by generating or executing the game program
    • A63F2300/61—Score computation

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • General Business, Economics & Management (AREA)
  • Computer Security & Cryptography (AREA)
  • Business, Economics & Management (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a table game hybrid recommendation algorithm based on multi-source heterogeneous data, aiming at solving the problem that the game characteristics and the user characteristics are not expressed sufficiently due to the sparse scoring matrix. The invention provides improvement on the basis of the traditional probability matrix decomposition algorithm (PMF), and the improvement points comprise: 1. the representation of the game feature vector is enhanced using game description text and game attribute information, 2. the representation of the user features is enhanced using the features of the user's scored game in combination with an attention mechanism. Experiments show that compared with a baseline model, the model provided by the invention has smaller score prediction error.

Description

Table game hybrid recommendation method based on multi-source heterogeneous data
Technical Field
The invention belongs to the technical field of big data and artificial intelligence, and particularly relates to a table game hybrid recommendation method based on multi-source heterogeneous data.
Background
The interest in board games has gradually begun to revive in the beginning of the 21 st century, the share of board games in the entertainment market has increased day by day, and this trend is global. The interest in table games is newer than watching movies, and there are currently fewer students researching the application recommended by table games. With the increasing number of table game players and table game players, the generated user interaction information and table game description information are more and more, and a good table game recommendation algorithm is needed to better meet the needs of people.
At present, various recommended applications are unclear, but the recommendation research of the board game is rare. Amazon and Barnes & Noble provide table game recommendation methods, but their recommendation methods are based on user purchase patterns and collaborative filtering from merchandise to merchandise, and do not involve game features to tailor the user's preferences, and do not achieve personalized recommendations. Yiu-Kai Ng et al propose the PeGRec model, which takes into account the characteristics of the game when recommending a new game to a user. The features comprise types, complexity, game time and the like, are integrated, and finally determine a recommendation list, so that the personalized recommendation of the user is realized, the cold start problem is solved to a certain extent, and the user satisfaction is improved. But requires the user to create a user profile and provide a favorite designated game by himself and does not achieve a fully intelligent recommendation. Jan Zalewski et al cluster the board games, calculate the user's likeability for each category, and recommend games to the user using a collaborative filtering idea based on the user. The method is suitable for occasions with few users, but in practical scenes, the users are many, and the cost of calculating the similarity matrix of the users is high. Michaelon et al propose a board game recommendation system that functions as a user entering a set of games they like and the system makes recommendations. The article provides a collaborative filtering-based and content-based hybrid recommendation method, and recommendation accuracy, diversity and novelty are improved. However, this approach requires the user to enter a set of games they like, on the one hand, for novice users who do not know what games they like, and on the other hand, does not utilize machine learning methods to analyze user preferences from user scored game assistance information.
Along with table game and the number of table game users is more and more, the data sparsity problem is more and more serious, traditional matrix decomposition algorithm only uses score data to recommend, when the score matrix is sparse, the performance of the score matrix is limited, and the traditional matrix decomposition algorithm does not consider: 1. enhancement of game feature expression by game multi-source heterogeneous auxiliary information. 2. User scored game features enhance user feature expression.
Disclosure of Invention
In order to overcome the defects of the prior art and further improve the effect of table game recommendation, the invention aims to provide a table game hybrid recommendation method based on multi-source heterogeneous data, which combines multi-source heterogeneous auxiliary information to solve the problem of inaccurate game characteristic and user characteristic representation caused by sparse scoring data, so that games which are more satisfactory can be recommended for users.
In order to achieve the purpose, the invention adopts the technical scheme that:
1. setting the conditional distribution p of the scoring matrix R as:
Figure BDA0002852150840000021
wherein R represents a scoring matrix, and R ═ Rij}，RijScoring a game j for a user i, i being a user number, i being 1,2,3.. L, L being a total number of users, j being a game number, j being 1,2,3.. M, M being a total number of games, U being a user feature matrix, U being { U }i}，UiA feature vector representing a user i, V is a game feature matrix, and V is { V ═ V }j}，VjFeature vector, alpha, representing game j2Is RijThe variance of (a) is determined,
Figure BDA0002852150840000022
represents RijObey mean value of Ui TVjVariance is α2With T as a transposed symbol, IijIs a matrix, if user I scores game j, then IijIs 1, otherwise is 0;
2. using convolutional neural networks to describe text from a board gameIn the method, the characteristic vector of the table game is extracted, and the formal expression is as follows: CNN (X)j,Wcnn) Wherein X isjIs a description text of game j, WcnnIs a weight parameter inside the convolutional neural network;
3. extracting the table game characteristic vector from the attribute information of the table game by using a multilayer perceptron, and formally expressing the table game characteristic vector as MLP (Y)j,Wmlp) Wherein Y isjIs attribute information of game j, WmlpRepresenting weight parameters inside the multi-layer perceptron network;
4. mixing CNN (X)j,Wcnn) And MLP (Y)j,Wmlp) Fusion was performed, denoted by CM as follows: CM (X)j,Yj,Wcnn,Wmlp,Wcm)＝relu(Wcm*concatenate(CNN(Xj,Wcnn),MLP(Yj,Wmlp)))
Wherein concatenate is a splicing function, relu is an activation function, WcmIs a fusion layer weight parameter;
5. feature vector V of game jjIs defined as: vj＝CM(Xj,Yj,WV)+ξj，WV＝{Wcnn,Wmlp,Wcm}, setting WVThe condition distribution of
Figure BDA0002852150840000031
I.e. for WVEach parameter w inkMean of 0 and variance of
Figure BDA0002852150840000032
k is WVNumber of medium-weight parameter, k ═ 1,2,3. | wk|，|wkL is WVNumber of middle weight parameters, ξjXi characteristics learned from the scoring matrixjObedience mean 0, variance
Figure BDA00028521508400000316
Of a Gaussian spherical distribution, i.e.
Figure BDA0002852150840000033
I is an identity matrix;
6. the condition distribution of the game feature matrix is as follows:
Figure BDA0002852150840000034
wherein X is a game description text set, Y is a game attribute information set,
Figure BDA0002852150840000035
is a VjThe variance of each of the components in the signal,
Figure BDA0002852150840000036
represents VjObey mean value CM (X)j,Yj,WV) Variance is
Figure BDA0002852150840000037
(ii) a gaussian spherical distribution;
7. features of a user's scored game, in combination with an attention mechanism, are used to enhance the user's feature representation, denoted UA,
Figure BDA0002852150840000038
wherein
Figure BDA0002852150840000039
a represents the use of a self-attention mechanism
Figure BDA00028521508400000310
Information is summarized to obtain an attention vector,
Figure BDA00028521508400000311
a matrix of feature vectors, W, for user i's scored gamesa1And Wa2For the weight parameter, softmax and tanh are activation functions;
8. the feature vector of user i is defined as:
Figure BDA00028521508400000312
wherein, WUWeight parameter, W, for the previous step in extracting user featuresU＝{Wa1,Wa2}, setting WUThe condition distribution of
Figure BDA00028521508400000313
I.e. for WUEach parameter w intMean of 0 and variance of
Figure BDA00028521508400000314
t is WUSequence number of medium weight parameter, t ═ 1,2,3. | wt|，|wtL is WUNumber of middle weight parameters, δiRepresenting features of the user learned from the scoring matrix, δiObedience mean 0 and variance
Figure BDA00028521508400000315
Of a Gaussian spherical distribution, i.e.
Figure BDA0002852150840000041
9. The condition distribution of the user characteristic matrix is as follows:
Figure BDA0002852150840000042
wherein VUA matrix set formed by the feature vectors of the scored games of each user;
Figure BDA0002852150840000043
is UiThe variance of each of the components in (a),
Figure BDA0002852150840000044
represents UiObey mean value of
Figure BDA0002852150840000045
Variance of
Figure BDA0002852150840000046
(ii) a gaussian spherical distribution;
10. and (3) estimating variables and parameters in the optimization model by using the maximum posterior, wherein the formula is as follows:
Figure BDA0002852150840000047
11. by taking the negative logarithm, the above equation is expressed as:
Figure BDA0002852150840000048
wherein the content of the first and second substances,
Figure BDA0002852150840000049
12. updating V by coordinate descent methodj:
Vj←(UIjUT+λUI)-1(URj+λVCM(Xj,Yj,WV))
Wherein, IjIs a diagonal matrix whose diagonal elements are IijMiddle j column element, RjIs a vector of value RijThe value of the element in the j-th column;
13. updating U by coordinate descent methodi
Figure BDA00028521508400000410
Wherein, IiIs a diagonal matrix whose diagonal elements are IijElement of row i of (1), RiIs a vector of value RijThe ith row element value;
14. according to epsilon (W)V) Updating W using a back propagation algorithmV：
Figure BDA00028521508400000411
15. According to epsilon (W)U) Updating W using a back propagation algorithmU：
Figure BDA0002852150840000051
16. Circularly executing the variables and parameters in the optimization model in the steps of 12-15 until convergence;
17. using optimized U, V, WU、WVPredicting the unknown score of the user on the game:
Figure BDA0002852150840000052
18. to evaluate the performance of the models herein, Root Mean Square Error (RMSE) was used as an evaluation criterion, with lower RMSE values representing better results. RMSE is defined as:
Figure BDA0002852150840000053
s is the score in the test set.
The board game attribute information includes: a minimum number of players, a maximum number of players, a mean time of play, a category of play, and a game play scheme. Dividing the table game attribute information into category information and numerical information, wherein the category information comprises: game category and game mechanism, the numerical information is: the minimum number of people in the game, the maximum number of people in the game and the average time of the game. And for category information, dimension reduction is carried out on the category information through embedding, Min-Max standardization is carried out on numerical value information, and the processed attribute information is spliced and then a multi-layer perceptron is used for extracting the table game feature vector from the attribute information.
Compared with the prior art, the invention has the beneficial effects that:
1. the problem of inaccurate game feature expression due to sparse scoring matrix is solved.
2. The problem of inaccurate user feature expression due to sparse scoring matrix is solved.
Drawings
FIG. 1 is a schematic diagram of a model of the present invention.
Detailed Description
The embodiments of the present invention will be described in detail below with reference to the drawings and examples.
The invention integrates the multi-source heterogeneous auxiliary information into the traditional PMF algorithm. And integrating the feature vectors extracted from the table game description text and the table game attribute information into the feature representation of the game end. Because the scoring matrix is sparse, the expression of the user characteristics is also inaccurate, and the invention enhances the expression of the user characteristics by combining the characteristics of the scored game of the user with the attention mechanism. The overall frame is shown in fig. 1.
The specific steps of one embodiment of the invention are as follows:
1. the data set adopted by the embodiment is from https:// boardgamegeek.com website which is the largest table game website in the world, and scores of 9021 games of 142308 users are crawled by using a crawler technology, and are shown in table 1. The game auxiliary information comprises game description text and game attribute information, wherein the attribute information comprises: the minimum number of people in the game, the maximum number of people in the game, the average time of the game, the game category and the game mechanism.
TABLE 1
Number of users Number of table games Score of Degree of sparseness
142308 9021 854848 99.93％
2. The text data is preprocessed as follows:
1) setting the maximum length of the game description text to be 200;
2) deleting the stop word;
3) selecting words with frequency more than 5 to construct a vocabulary list;
4) the game description is represented as a word index vector.
3. Setting the conditional distribution p of the scoring matrix R as:
Figure BDA0002852150840000061
wherein R represents a scoring matrix, and R ═ Rij}，RijScoring a game j for a user i, i being a user number, i being 1,2,3.. L, L being a total number of users, j being a game number, j being 1,2,3.. M, M being a total number of games, U being a user feature matrix, U being { U }i}，UiA feature vector representing a user i, V is a game feature matrix, and V is { V ═ V }j}，VjFeature vector, alpha, representing game j2Is RijThe variance of (a) is determined,
Figure BDA0002852150840000062
represents RijObey mean value of Ui TVjVariance is α2With T as a transposed symbol, IijIs a matrix, if user I scores game j, then IijIs 1, otherwise is 0;
4. extracting a table game feature vector from the table game description text by using a Convolutional Neural Network (CNN), and formally expressing the table game feature vector as follows: CNN (X)j,Wcnn) Wherein X isjIs a description text of game j, WcnnIs a weight parameter inside the CNN network. The CNN network consists of an embedded layer, a convolutional layer, a pooling layer and a full-connection layer 4, and the experimental settings are as follows:
1) the dimension of the word vector is 300, and the word vector is trained through an optimization process;
2) in the convolutional layer, 64 filters were used, each filter having a window size set to 3, 4, 5;
3) the output dimension is set to 200.
5. The game attribute information can be divided into two types, namely category information and numerical information, wherein the category information comprises: game category and game mechanism, the numerical information is: the minimum number of people in the game, the maximum number of people in the game and the average time of the game. For the category information, dimension reduction is performed through embedding, because high-dimensional and sparse vectors influence the performance of the model. Min-Max standardization is carried out on numerical value information, the processed attribute information is spliced, a multi-layer perceptron (MLP) is used for extracting a table game feature vector from the table game attribute information, and the table game feature vector is expressed as MLP (Y) in a formalized modej,Wmlp) Wherein Y isjIs attribute information of game j, WmlpRepresenting the weight parameters inside the MLP network. In this embodiment, the MLP network includes two hidden layers, and the output dimension is set to 50.
6. Fusing the feature vectors extracted from the game description text and the attribute information, and expressing the feature vectors by using CM: CM (X)j,Yj,Wcnn,Wmlp,Wcm)＝relu(Wcm*concatenate(CNN(Xj,Wcnn),MLP(Yj,Wmlp) Conate is the splicing function and relu is the activation function. WcmIs a fusion layer weight parameter. The output dimension is set to 50.
7. Feature vector V of game jjIs defined as: vj＝CM(Xj,Yj,WV)+ξjFor convenience of illustration, this embodiment uses WVA weight parameter indicating the above-mentioned extraction of the game feature vector from the plurality of auxiliary data,
WV＝{Wcnn,Wmlp,Wcm}, setting WVThe condition distribution of
Figure BDA0002852150840000071
I.e. for WVEach parameter w inkMean of 0 and variance of
Figure BDA0002852150840000072
k is WVNumber of medium-weight parameter, k ═ 1,2,3. | wk|，|wkL is WVNumber of middle weight parameters, ξjXi characteristics learned from the scoring matrixjObedience mean 0, variance
Figure BDA0002852150840000073
Of a Gaussian spherical distribution, i.e.
Figure BDA0002852150840000074
I is an identity matrix, ξjHas a dimension of 50.
8. Thus, the condition distribution of the game feature matrix is:
Figure BDA0002852150840000075
wherein X is a game description text set, Y is a game attribute information set,
Figure BDA0002852150840000081
is a VjThe variance of each of the components in the signal,
Figure BDA0002852150840000082
represents VjObey mean value CM (X)j,Yj,WV) Variance is
Figure BDA0002852150840000083
Gaussian spherical distribution.
9. Since the scoring data is sparse, the expression of the user features is inaccurate, and the game played by the user reflects the interest of the user, however, the contribution of different games to the user features is different, so the embodiment uses the features of the scored game of the user in combination with the attention mechanism to enhance the user features, which are denoted as UA,
Figure BDA0002852150840000084
wherein
Figure BDA0002852150840000085
a represents the use of a self-attention mechanism
Figure BDA0002852150840000086
Information is summarized to obtain an attention vector,
Figure BDA0002852150840000087
a matrix of feature vectors, W, for user i's scored gamesa1And Wa2For the weight parameters, softmax and tanh are activation functions, and the output dimension of UA is set to 50.
10. The feature vector of user i is defined as:
Figure BDA0002852150840000088
wherein, WUWeight parameter, W, for the previous step in extracting user featuresU＝{Wa1,Wa2}, setting WUThe condition distribution of
Figure BDA0002852150840000089
I.e. for WUEach parameter w intMean of 0 and variance of
Figure BDA00028521508400000810
t is WUSequence number of medium weight parameter, t ═ 1,2,3. | wt|，|wtL is WUNumber of middle weight parameters, δiRepresenting features of the user learned from the scoring matrix, δiObedience mean 0 and variance
Figure BDA00028521508400000811
Of a Gaussian spherical distribution, i.e.
Figure BDA00028521508400000812
δiHas a dimension of 50.
11. The condition distribution of the user characteristic matrix is as follows:
Figure BDA00028521508400000813
wherein VUA matrix set formed by the feature vectors of the scored games of each user;
Figure BDA00028521508400000814
is UiThe variance of each of the components in (a),
Figure BDA00028521508400000815
represents UiObey mean value of
Figure BDA00028521508400000816
Variance of
Figure BDA00028521508400000817
(ii) a gaussian spherical distribution;
12. to optimize variables, parameters in the model, maximum a posteriori estimation (MAP) is used. The formula is as follows:
Figure BDA00028521508400000818
13. by taking the negative logarithm, the above equation is expressed as:
Figure BDA0002852150840000091
wherein the content of the first and second substances,
Figure BDA0002852150840000092
14 updating by coordinate descent
Figure BDA0002852150840000093
Wherein, IjIs a diagonal matrix whose diagonal elements are IijMiddle j column element, RjIs a vector of value RijThe value of the element in the j-th column.
15. Updating by coordinate descent
Figure BDA0002852150840000094
IiIs a diagonal matrix whose diagonal elements are IijElement of row i of (1), RiIs a vector of value RijRow i element value.
16. According to epsilon (W)V) Updating W using a back propagation algorithmV：
Figure BDA0002852150840000095
17. According to epsilon (W)U) Updating W using a back propagation algorithmU：
Figure BDA0002852150840000096
18. And circularly executing the variables and the parameters in the optimization model in the steps 14-17 until convergence.
19. Using optimized U, V, WU、WVPredicting the unknown score of the user on the game:
Figure BDA0002852150840000097
20. in order to measure the score prediction effect of the model of this embodiment, the embodiment uses Root Mean Square Error (RMSE) to verify the difference between the predicted score and the actual score obtained by the model of this embodiment and the baseline model. RMSE is defined as:
Figure BDA0002852150840000098
s is the score in the test set.
The results of the experiment are shown in table 2.
TABLE 2
Algorithm RMSE
PMF 1.617
CDL 1.605
ConvMF 1.590
Model of the invention 1.381
It can be seen from table 2 that the model of the present invention is superior to the PMF model, the CDL model and the Convmf model, the PMF model only uses the scoring matrix for scoring prediction, the effect is the worst, the CDL model uses the SDAE to process the text information and fuses in the PMF, the scoring prediction effect is superior to the PMF model, the Convmf model uses the CNN network to process the text information and fuses in the PMF, and the error of scoring prediction is further reduced. However, no auxiliary information is added to the user terminal in the above baseline method, so that the score prediction error is further reduced. In the method for referencing Convmf, CNN is used for processing text information, MLP network processing attribute information is added, the text information and the MLP network processing attribute information are integrated and then are fused into a PMF model to enhance game feature representation, and user feature representation is enhanced by combining user scored game features and an attention mechanism, so that the error of scoring prediction is further reduced.
21. To verify the parameter lambda in the model hereinUAnd λVInfluence on score prediction result by setting different lambdaUAnd λVThe results of the experiments are shown in Table 3
TABLE 3
λU λV RMSE
90 10 1.43616
50 50 1.40019
10 90 1.38192
10 100 1.38281
20 90 1.38094
Experiments have shown that the result of the score prediction is subject to λUAnd λVOf suitable lambdaUAnd λVThe error of the score prediction can be further reduced, and as can be seen from the above table, when lambda is usedUAnd λVThe score prediction error is minimal when the values of (a) and (b) are 20 and 90, respectively.

Claims (4)

1. A board game hybrid recommendation method based on multi-source heterogeneous data is characterized by comprising the following steps:
step 1, setting the conditional distribution p of a scoring matrix R as:
Figure FDA0003502243820000011
wherein R represents a scoring matrix, and R ═ Rij}，RijScoring a game j for a user i, i being a user number, i being 1,2,3.. L, L being a total number of users, j being a game number, j being 1,2,3.. M, M being a total number of games, U being a user feature matrix, U being { U }i}，UiA feature vector representing a user i, V is a game feature matrix, and V is { V ═ V }j}，VjFeature vector, alpha, representing game j2Is RijThe variance of (a) is determined,
Figure FDA0003502243820000012
represents RijObey mean value of Ui TVjVariance is α2With T as a transposed symbol, IijIs a matrix, if user I scores game j, then IijIs 1, otherwise is 0;
step 2, extracting the table game feature vector from the table game description text by using a convolutional neural network, wherein the table game feature vector is represented in a formalized way as follows: CNN (X)j,Wcnn) Wherein X isjIs a description text of game j, WcnnIs a weight parameter inside the convolutional neural network;
and 3, extracting the table game characteristic vector from the table game attribute information by using a multilayer perceptron, and formally expressing the table game characteristic vector as MLP (Y)j,Wmlp) Wherein Y isjIs attribute information of game j, WmlpRepresenting weight parameters inside the multi-layer perceptron network;
step 4, CNN (X)j,Wcnn) And MLP (Y)j,Wmlp) Fusion was performed, denoted by CM as follows:
CM(Xj,Yj,Wcnn,Wmlp,Wcm)＝relu(Wcm*concatenate(CNN(Xj,Wcnn),MLP(Yj,Wmlp)))
wherein concatenate is a splicing function, relu is an activation function, WcmIs a fusion layer weight parameter;
step 5, the feature vector V of the game jjIs defined as: vj＝CM(Xj,Yj,WV)+ξj，WV＝{Wcnn,Wmlp,Wcm}, setting WVThe condition distribution of
Figure FDA0003502243820000013
I.e. for WVEach parameter w inkMean of 0 and variance of
Figure FDA0003502243820000014
k is WVNumber of medium-weight parameter, k ═ 1,2,3. | wk|，|wkL is WVNumber of middle weight parameters, ξjXi characteristics learned from the scoring matrixjObedience mean 0, variance
Figure FDA00035022438200000221
Of a Gaussian spherical distribution, i.e.
Figure FDA0003502243820000021
I is an identity matrix;
step 6, the condition distribution of the game feature matrix is as follows:
Figure FDA0003502243820000022
wherein X is a game description text set, Y is a game attribute information set,
Figure FDA0003502243820000023
is a VjThe variance of each of the components in the signal,
Figure FDA0003502243820000024
represents VjObey mean value CM (X)j,Yj,WV) Variance is
Figure FDA0003502243820000025
(ii) a gaussian spherical distribution;
step 7, using the features of the user's scored game in combination with an attention mechanism to enhance the user feature representation, denoted UA,
Figure FDA0003502243820000026
wherein
Figure FDA0003502243820000027
a represents the use of a self-attention mechanism
Figure FDA0003502243820000028
Information is summarized to obtain an attention vector,
Figure FDA0003502243820000029
a matrix of feature vectors, W, for user i's scored gamesa1And Wa2For the weight parameter, softmax and tanh are activation functions;
step 8, defining the feature vector of the user i as:
Figure FDA00035022438200000210
wherein, WUWeight parameter, W, for the previous step in extracting user featuresU＝{Wa1,Wa2}, setting WUThe condition distribution of
Figure FDA00035022438200000211
I.e. for WUEach parameter w intMean of 0 and variance of
Figure FDA00035022438200000212
t is WUSequence number of medium weight parameter, t ═ 1,2,3. | wt|，|wtL is WUNumber of middle weight parameters, δiRepresenting features of the user learned from the scoring matrix, δiObedience mean 0 and variance
Figure FDA00035022438200000213
Of a Gaussian spherical distribution, i.e.
Figure FDA00035022438200000214
Step 9, the condition distribution of the user feature matrix is as follows:
Figure FDA00035022438200000215
wherein VUA matrix set formed by the feature vectors of the scored games of each user;
Figure FDA00035022438200000216
is UiThe variance of each of the components in (a),
Figure FDA00035022438200000217
represents UiObey mean value of
Figure FDA00035022438200000218
Variance of
Figure FDA00035022438200000219
(ii) a gaussian spherical distribution;
step 10, using maximum posteriori to estimate variables and parameters in the optimization model, wherein the formula is as follows:
Figure FDA00035022438200000220
step 11, by taking the negative logarithm, the above formula is expressed as:
Figure FDA0003502243820000031
wherein the content of the first and second substances,
Figure FDA0003502243820000032
step 12, updating V by adopting a coordinate descent methodj:
Vj←(UIjUT+λUI)-1(URj+λVCM(Xj,Yj,WV))
Wherein, IjIs a diagonal matrix whose diagonal elements are IijMiddle j column element, RjIs a vector of value RijThe value of the element in the j-th column;
step 13, updating U by adopting a coordinate descent methodi
Figure FDA0003502243820000033
Wherein, IiIs a diagonal matrix whose diagonal elements are IijElement of row i of (1), RiIs a vector of value RijThe ith row element value;
step 14, according to epsilon (W)V) Updating W using a back propagation algorithmV：
Figure FDA0003502243820000034
Step 15, according to epsilon (W)U) Updating W using a back propagation algorithmU：
Figure FDA0003502243820000035
Step 16, circularly executing the variables and parameters in the optimization model in the steps 12-15 until convergence;
step 17, using the optimized U, V, WU、WVPredicting the unknown score of the user on the game:
Figure FDA0003502243820000036
and step 18, in order to evaluate the performance of the model, the Root Mean Square Error (RMSE) is used as an evaluation standard, and the lower the root mean square error value is, the better the representation effect is.
2. The multi-source heterogeneous data-based board game hybrid recommendation method according to claim 1, wherein the board game attribute information comprises: a minimum number of players, a maximum number of players, a mean time of play, a category of play, and a game play scheme.
3. The multi-source heterogeneous data-based board game hybrid recommendation method according to claim 1, wherein the board game attribute information is divided into category information and numerical information, and the category information includes: game category and game mechanism, the numerical information is: the minimum number of people playing, the maximum number of people playing and the average time of playing; and for category information, dimension reduction is carried out on the category information through embedding, Min-Max standardization is carried out on numerical value information, and the processed attribute information is spliced and then a multi-layer perceptron is used for extracting the table game feature vector from the attribute information.
4. The multi-source heterogeneous data-based board game hybrid recommendation method according to claim 1, wherein the root mean square error is defined as:
Figure FDA0003502243820000041
s is the score in the test set.
CN202011531148.XA 2020-12-22 2020-12-22 Table game hybrid recommendation method based on multi-source heterogeneous data Active CN112569608B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011531148.XA CN112569608B (en) 2020-12-22 2020-12-22 Table game hybrid recommendation method based on multi-source heterogeneous data

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011531148.XA CN112569608B (en) 2020-12-22 2020-12-22 Table game hybrid recommendation method based on multi-source heterogeneous data

Publications (2)

Publication Number Publication Date
CN112569608A CN112569608A (en) 2021-03-30
CN112569608B true CN112569608B (en) 2022-03-25

Family

ID=75138898

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011531148.XA Active CN112569608B (en) 2020-12-22 2020-12-22 Table game hybrid recommendation method based on multi-source heterogeneous data

Country Status (1)

Country Link
CN (1) CN112569608B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106909540A (en) * 2015-12-23 2017-06-30 神州数码信息系统有限公司 A kind of smart city citizen's preference discovery technique based on Cooperative Study
CN108959603A (en) * 2018-07-13 2018-12-07 北京印刷学院 Personalized recommendation system and method based on deep neural network
CN109255049A (en) * 2018-09-05 2019-01-22 广州小楠科技有限公司 A kind of wisdom music recommender system

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109635989B (en) * 2018-08-30 2022-03-29 电子科技大学 Social network link prediction method based on multi-source heterogeneous data fusion

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106909540A (en) * 2015-12-23 2017-06-30 神州数码信息系统有限公司 A kind of smart city citizen's preference discovery technique based on Cooperative Study
CN108959603A (en) * 2018-07-13 2018-12-07 北京印刷学院 Personalized recommendation system and method based on deep neural network
CN109255049A (en) * 2018-09-05 2019-01-22 广州小楠科技有限公司 A kind of wisdom music recommender system

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Wide and Deep Model of Multi-Source Information-Aware Recommender System;WEIHUA YUAN 等;《IEEE Access》;20180928;第6卷;第49385-49398页 *
基于深度学习的融合多源异构数据的推荐模型;冀振燕 等;《北京邮电大学学报》;20191217;第42卷(第6期);第35-42页 *
融合多源异构数据的推荐模型与系统;邱丰羽;《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》;20190715(第07期);第I138-1508页 *

Also Published As

Publication number Publication date
CN112569608A (en) 2021-03-30

Similar Documents

Publication Publication Date Title
CN107330049B (en) News popularity estimation method and system
CN107992531A (en) News personalization intelligent recommendation method and system based on deep learning
CN110362738B (en) Deep learning-based individual recommendation method combining trust and influence
CN111127142B (en) Article recommendation method based on generalized nerve attention
CN112417306B (en) Method for optimizing performance of recommendation algorithm based on knowledge graph
CN107608956A (en) A kind of reader's mood forecast of distribution algorithm based on CNN GRNN
CN110837578B (en) Video clip recommendation method based on graph convolution network
CN109271629B (en) Method for generating text abstract based on reinforcement learning
CN109978660A (en) A kind of recommender system off-line training method based on intensified learning frame
CN112925977A (en) Recommendation method based on self-supervision graph representation learning
CN111241394A (en) Data processing method and device, computer readable storage medium and electronic equipment
CN110457711B (en) Subject word-based social media event subject identification method
CN114519145A (en) Sequence recommendation method for mining long-term and short-term interests of users based on graph neural network
CN110930219A (en) Personalized merchant recommendation method based on multi-feature fusion
CN114332519A (en) Image description generation method based on external triple and abstract relation
CN112569608B (en) Table game hybrid recommendation method based on multi-source heterogeneous data
CN111897999B (en) Deep learning model construction method for video recommendation and based on LDA
CN113449200B (en) Article recommendation method and device and computer storage medium
CN115809374B (en) Method, system, device and storage medium for correcting mainstream deviation of recommendation system
CN116385077A (en) Multi-behavior recommendation system based on behavior perception fusion graph convolution network
CN108182288A (en) A kind of recommendation method based on artificial immunity
CN114461928A (en) Session recommendation method based on graph neural network
WO2021189583A1 (en) Restricted boltzmann machine-driven interactive personalized search method
CN114611013A (en) Collaborative filtering recommendation algorithm based on adaptive combination of user interest and scoring preference difference
CN113362034A (en) Position recommendation method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant