CN114580529A - System for selecting and training football players - Google Patents

Abstract

The invention discloses a system for selecting and training football players, and belongs to the field of data processing. The invention collects the characteristics of the players, and classifies and evaluates the various aspects of the abilities of the players according to the characteristic data to obtain the states of the players and find the aspects of the players needing to be improved. In calculating the item that the athlete needs to improve most, not only is the performance of the athlete on the item taken into account, but the relative importance of the item is additionally taken into account. It is of great practical significance to find the most needed items for each player, so that a set of personalized training scheme can be made for each player to improve the overall level of the team. The invention can find out the athletes with special characteristics from the athletes with poor labels, and can improve the core competitiveness of the team by carrying out strengthening training or weak item supplement; on the other hand, the full-functional players are few and few, and the finding of the potential special-color players through potential mining has great practical significance.

Description

System for selecting and training football players

Technical Field

The invention belongs to the technical field of machine learning, and relates to a system for improving the overall level by analyzing the body of an athlete and selecting excellent athletes through training data and proposing a targeted training suggestion for each individual based on a random forest classification model.

Background

In real life, selection of excellent football players is a process of comprehensively considering various index factors of players. Whether the football player can become an excellent football player depends on the kicking skills of the player, such as ball carrying, shooting, passing, visual field, team cooperation ability and the like, and is also influenced by physiological factors such as height, weight, endurance, age and the like. With the advancement of Chinese football toward the occupational, the traditional multi-match manner is adopted by the green training clubs or football schools to screen potential outstanding players. The world football strong nations pay high attention to the early selection of talent athletes, and the Germany, Spain, Italy and other football strong nations have a set of comprehensive material selection systems which are suitable for the characteristics of the athletes in the country and comprise physical quality, football technology, tactical awareness, psychological quality and team spirit, and are implemented by professional staff. Firstly, the ball probe widely collects information, regularly goes to schools, football clubs and the like to watch children and teenagers to play balls, train and match, and finds seedlings with talents; then tests such as physical quality, football technology, tactical awareness, psychological ability and the like are carried out, and long-term tracking and investigation are carried out, and the selection and investigation of some athletes can be as long as months, even years.

The prior art has the following disadvantages:

the traditional football player selection method not only depends on the expert experience of selection personnel, has larger influence by subjective factors, but also is time-consuming and labor-consuming and is difficult to implement in relatively undeliverable areas. With the rise of data mining technology, many learners begin to adopt methods such as machine learning and neural network based on classification to perform personnel screening, but the methods can not effectively solve the problem of football player selection. Specifically, the following points are included: firstly, for training data only containing labels with discreteness such as excellence, good and passing or passing, etc., the traditional machine learning classifier or the current neural network classifier can only obtain discreteness results, and the real level of the movement staff cannot be represented by a continuous value; secondly, for the players at the middle and low levels, it is not enough to just classify them, and we would rather want to understand where each of their individuals is different from the players at the high level, i.e. explicitly showing the advantages and disadvantages of each player in football, to be able to provide referable suggestions for the subsequent training; finally, both traditional player selection methods and classification-based machine learning methods lack the analysis and mining of the potential of soccer players.

Disclosure of Invention

Aiming at the defects of the existing player selection method, the invention provides a system for selecting excellent players and providing adaptive training suggestions for each individual based on the continuous discrete indexes of a random forest classification model. In the aspect of feature screening, the importance of the features is sequenced by adopting a method of fusing multiple feature extractions, so that the contingency and deviation caused by a single method are reduced; from the aspect of training data, the invention adopts five-fold cross validation, balances the possibility of training and testing each sample, and optimizes the composition of training samples; from the output result, the invention provides a continuous evaluation result for each athlete, not only can more finely classify the athlete group, but also can evaluate the potential of all the athletes, analyze the advantages and disadvantages of the athletes of medium level and provide training suggestions for the athletes of medium level, and excavate the potential of the athletes of low level.

The technical scheme of the invention is a system for selecting and training football players, which comprises: the system comprises a data input module, a preprocessing module, a prediction module, a potential evaluation module, a major-minor situation analysis module, a potential mining module and an output module; the selecting and drawing system can be divided into two stages according to whether the model is trained or not: a training phase and a prediction phase; the system is in a training stage initially, and input data are required to be height, weight, strength, pass and shooting data of a quantified athlete and a final expert evaluation result, wherein all the input data except the expert evaluation result are continuous fraction values or discrete values, and the evaluation result is required to be a discrete value; after the training is finished, the system can enter a prediction stage, and input data at the moment are all other items except the evaluation result; the data input module receives the input and transmits the data to the preprocessing module;

the preprocessing module is used for carrying out normalization processing on data except the label by adopting a 0-1 normalization method; the preprocessing module transmits the preprocessed data to the prediction module;

the prediction module is a random forest classification model; in the system training stage, a random forest classification model receives labeled data output by a preprocessing module, random seeds are arranged in the system, and multi-classification training is carried out on the model by adopting a training form of five-fold cross validation matched with grid search parameters; the random forest classification model in the prediction module has different structural parameters corresponding to different inputs, the model can automatically store the parameter combination with the highest accuracy in the training stage, and the module does not output in the training stage; in the prediction phase, the prediction module receives the unlabeled data output from the preprocessing module and the user-input threshold K₁,K₂The method comprises the steps of calculating the excellent probability P of each sport and classifying athletes into four classes according to a threshold value, wherein if P is larger than or equal to K₂Is sufficiently excellent that P is 0.5. ltoreq. P < K₂Is excellent but to be improved if K₁P is less than or equal to P and less than 0.5, the ratio is not excellent but can be improved, and P is less than K₁Classification as not excellent; the prediction module directly inputs the data of the athletes classified as being excellent enough to the output module in the prediction stage, inputs the data corresponding to the athletes classified as being excellent but to be improved to the advantage and disadvantage analysis module, inputs the data corresponding to the athletes classified as not being excellent but to be improved to the potential evaluation module, and inputs the data corresponding to the athletes classified as not being excellent to the potential mining module;

the potential evaluation module sets a parameter n after acquiring the data output by the prediction module₁，0≤n₁< N, wherein N₁Representing the maximum tolerable number of items of poor performance of the athlete by the user, N representing the total number of test items of the athlete; by calculating the top n of the athlete most in need of improvement₁An item, then the athlete is on n₁Replacing the true value on each project with the average level value of an excellent athlete, predicting the athlete by adopting the prediction module again, if the athlete is predicted to be excellent enough, changing the athlete into the excellent type but to be improved, and transmitting the result to the superiority and inferiority analysis module, otherwise, classifying the athlete into the non-excellent type, and transmitting the athlete data to the potential mining module;

after the superiority and inferiority analysis module obtains data, a parameter n is set₂，0≤n₂N or less, wherein N₂The potential evaluation module is used for calculating the top n which needs to be improved most for the athlete by the same method of representing the relatively important item number which is wanted by the user and is not good for the athlete₂The project transmits the result to an output module, so that subsequent targeted training is facilitated;

after the potential mining module receives the data, the potential mining module calculates and analyzes the special project of the athlete according to the data, so that the court position distribution of the athlete is facilitated, if no special project exists, the special project is directly eliminated, the result is transmitted to the output module, and the special project is that the level of the project of the athlete is higher than the average level of excellent athletes.

Further, the specific process of the normalization process in the preprocessing module is as follows:

in order to perform statistics and analysis on various indexes, data needs to be preprocessed into a format which can be received by a prediction module. Firstly, converting Chinese discrete values such as excellent, good and passing characters into Arabic numerals such as 1, 2, 3 and 4, and requiring the correlation between the size of the numerals and the actual value. In order to make features of different dimensions in the same order of magnitude and reduce the effect of variance on the features, feature scaling is performed using 0-1 normalization.

Further, the weight coefficient of each feature is calculated in the preprocessing module, and the calculation method is as follows:

firstly, selecting features by using a plurality of methods respectively, carrying out independence test on the features and the labels by using chi-square test, and sequencing according to the independence from small to large; calculating the information gain of each characteristic by using an information gain method, and sequencing the information gains from large to small; the Lasso method regards the process as multivariate linear regression, each feature has a corresponding weight coefficient coef, the positive and negative values of the weight coefficient coef represent whether the feature is positively correlated or negatively correlated with a target value, and the features are sorted from large to small according to the absolute value of coef; calculating a correlation coefficient of each feature and a target value by a pearson correlation coefficient method, and sorting according to the absolute value of the correlation coefficients from large to small; and finally, endowing each feature with a weight value from 1 to n according to the sorting result, adding the weight values of the four sorting results, and taking the reciprocal of the sum to obtain a weight coefficient of each feature, wherein the greater the weight coefficient, the more important the weight coefficient is.

Further, the specific process of calculating the athlete excellence probability in the prediction module is as follows:

the prediction result of the adopted decision tree is excellent or not excellent, a plurality of decision trees are adopted for prediction, and the value of p is the ratio of the number of excellent decision trees to the total number of all decision trees.

Further, the specific process of calculating the top n items of the athlete most needing improvement is calculated in the potential evaluation and superiority and inferiority analysis module:

firstly, reading an average level vector of excellent athletes obtained by calculation of a preprocessing module in a system training stage, then calculating a difference value between the excellent level vector and a sample feature vector, and performing dot product on a feature weight vector consisting of the difference vector and a feature weight value to obtain a result vector; the item with larger value in the result vector represents that the athlete has poor performance on the item and the item has large influence on the final judgment result, and the item with the largest value in the top n is taken as the top n items which need to be improved most.

Further, the specific process of calculating the characteristics of the athlete in the potential mining module is as follows:

firstly, reading the average level vector of the excellent athlete calculated by the preprocessing module in the system training stage, then calculating the difference vector of the excellent level vector and the feature vector corresponding to the sample, and taking all items greater than 0 in the difference vector as the characteristic of the sample, namely, the characteristic represents that the overall level of the athlete does not reach the standard, but the excellent level is reached on the characteristic items.

Compared with the prior art, the invention has the beneficial effects that:

1. the invention adopts a mode of fusing a plurality of feature selection methods to calculate the weight coefficient of each feature, thereby avoiding the deviation and the contingency brought by a single feature selection method. It was verified that the results sorted out separately using different feature selection methods are significantly different. Compared with the prior art, the characteristic sequencing result of the invention has more universality and persuasion.

2. The present invention uses data with only discrete tags, but can predict a continuous type of tag value for each sample. The ratio of the number of decision trees classified as 1 to the total number of decision trees represents the excellent probability of the sample in step 3, and assuming that the used random forest classification model has 100 decision trees, the number of decision trees classified as 1 is a continuous value between 0 and 100. A continuous label value can further refine and classify the sample category, and can solve the problem that similar excellent athletes can be better at all. Different from the traditional machine learning or deep learning classification method, the invention can provide discrete classification results and also can obtain continuous results, and the output form is more effective and diversified.

3. When the project which needs to be improved by the athlete is calculated, the performance of the athlete on the project is considered, the relative importance of the project, namely the influence degree on the final overall evaluation, is also considered, and the traditional experience-based athlete selecting method only can focus on one project and lacks a scientific and specific evaluation method. It is of great practical significance to find the most needed items for each player, so that a set of personalized training scheme can be made for each player to improve the overall level of the team.

4. Compared with the traditional selection method, the invention can dig out the players with the special length. In the traditional selecting process, one athlete is judged to be poor and usually represents that the whole game is rejected, but some athletes possibly have special talents in aspects of certain projects such as shooting, passing and the like, the invention can find out the athlete with special length from the athletes marked as the poor, and the core competitiveness of the team can be improved by carrying out strengthening training or weak item supplement; on the other hand, the full-functional players are few and few, and the finding of the potential special-color players through potential mining has great practical significance.

Drawings

Fig. 1 is a block diagram of a football player selection system according to the present invention.

FIG. 2 is a flow chart of a method for performing feature sorting by combining multiple methods.

FIG. 3 is a flow chart of a method for refining re-classification when predicting results.

FIG. 4 is a flow chart of a method for calculating the most desirable items for each athlete to improve.

Detailed description of the preferred embodiments

In order to make the selecting process and the innovation point of the present invention clearer, the present invention is further described in detail below with reference to the accompanying drawings.

Fig. 2 visually shows the process of performing feature sorting and calculating to obtain feature weights by fusing multiple feature selection methods, and the specific algorithm flow is as follows:

firstly, respectively using chi-square test, pearson correlation coefficient method, Lasso and information gain method to make feature selection, sorting according to feature importance from large to small to obtain table T₁，T₂，T₃，T₄. Wherein T is_i＝{f₁，f₂，f₃…f_n}；

② setting characteristic f_iThe rank values in the four tables are r₁，r₂，r₃，r₄Then the composite rank of the ith feature is S_i＝∑r_iSorting according to the comprehensive ranking from small to large to obtain a comprehensive ranking list of characteristicsT₀＝{f₁，f₂，f₃…f_nIs corresponding to a composite score of S₀＝{s₁，s₂，s₃…s_n}；

③ the weight calculation mode of the characteristic i is w_i＝1/s_iAnd according to the characteristic weight table W corresponding to the characteristic comprehensive ranking table₀＝{w₁，w₂，w₃…w_n}

FIG. 3 shows a process of training a random forest classification model with discrete label data, performing continuous label prediction, and then performing refinement and classification.

FIG. 4 shows in detail the process of calculating the most desirable items for each athlete, with the specific algorithm flow as follows:

firstly, the original test training data of the athlete i is set as V_i＝{f₁，f₂，f₃…f_nAnd f, taking all the players with the labels as excellent players, and calculating the average level of the excellent players on each item to obtain V₀＝{f_1’，f_2’，f_3’…f_n’}

② calculating a difference vector V_g＝V₀-V_iI.e. V_g＝{f₁-f_1’，f₂-f_2’，f₃-F_3’f_n…F_n’}；

Thirdly, calculating the result vector V_r＝V_g·W₀；

And fourthly, sorting the elements of the result vector from large to small to obtain a final result, wherein the feature corresponding items with larger element values need to be improved.

The above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should be considered as the protection scope of the present invention.

Claims (6)

1. A system for selecting and training a soccer player, the system comprising: the system comprises a data input module, a preprocessing module, a prediction module, a potential evaluation module, a major-minor situation analysis module, a potential mining module and an output module; the selecting and drawing system can be divided into two stages according to whether the model is trained or not: a training phase and a prediction phase; the system is in a training stage initially, and input data are required to be height, weight, strength, pass and shooting data of a quantified athlete and a final expert evaluation result, wherein all the input data except the expert evaluation result are continuous fraction values or discrete values, and the evaluation result is required to be a discrete value; after the training is finished, the system can enter a prediction stage, and input data at the moment are all other items except the evaluation result; the data input module receives the input and transmits the data to the preprocessing module;

the preprocessing module is used for carrying out normalization processing on data except the label by adopting a 0-1 normalization method; the preprocessing module transmits the preprocessed data to the prediction module;

the prediction module is a random forest classification model; in the system training stage, a random forest classification model receives labeled data output by a preprocessing module, random seeds are arranged in the system, and multi-classification training is carried out on the model by adopting a training form of five-fold cross validation matched with grid search parameters; the random forest classification model in the prediction module has different structural parameters corresponding to different inputs, the model can automatically store the parameter combination with the highest accuracy in the training stage, and the module does not output in the training stage; in the prediction phase, the prediction module receives the unlabeled data output from the preprocessing module and the user-input threshold K₁,K₂The method comprises the steps of calculating the excellent probability P of each sport and classifying athletes into four classes according to a threshold value, wherein if P is larger than or equal to K₂Is sufficiently excellent that P is 0.5. ltoreq. P < K₂Is excellent but to be improved if K₁P is less than or equal to P and less than 0.5, the ratio is not excellent but can be improved, and P is less than K₁Classification as not excellent; the prediction module directly inputs the athlete data classified as excellent enough to the output module in the prediction stage, and the athlete classified as excellent but to be improvedCorresponding data is input to a superiority and inferiority analysis module, data corresponding to athletes classified as not superior but capable of being improved is input to a potential evaluation module, and data corresponding to athletes classified as not superior is input to a potential mining module;

the potential evaluation module sets a parameter n after acquiring the data output by the prediction module₁，0≤n₁< N, wherein N₁Representing the maximum tolerable number of items for the athlete to underperform by the user, N representing the total number of test items for the athlete; by calculating the top n of the athlete most in need of improvement₁An item, then the athlete is on n₁Replacing the true value on each project with the average level value of an excellent athlete, predicting the athlete by adopting the prediction module again, if the athlete is predicted to be excellent enough, changing the athlete into the excellent type but to be improved, and transmitting the result to the superiority and inferiority analysis module, otherwise, classifying the athlete into the non-excellent type, and transmitting the athlete data to the potential mining module;

after the superiority and inferiority analysis module obtains data, a parameter n is set₂，0≤n₂N or less, wherein N₂The potential evaluation module is used for calculating the top n which needs to be improved most for the athlete by the same method of representing the relatively important item number which is wanted by the user and is not good for the athlete₂The project transmits the result to an output module, so that subsequent targeted training is facilitated;

after the potential mining module receives the data, the potential mining module calculates and analyzes the special project of the athlete according to the data, so that the court position distribution of the athlete is facilitated, if no special project exists, the special project is directly eliminated, the result is transmitted to the output module, and the special project is that the level of the project of the athlete is higher than the average level of excellent athletes.

2. A system for selecting and training an athlete in soccer according to claim 1, wherein said normalization process in said preprocessing module is performed by:

in order to perform statistics and analysis on various indexes, data needs to be preprocessed into a format which can be received by a prediction module. Firstly, converting excellent, good, and qualified Chinese discrete values into 1, 2, 3, 4, and other Arabic numerals, and requiring the correlation between the size of the numerals and the actual values. In order to make the features of different dimensions in the same order of magnitude, and reduce the effect of variance on the features, feature scaling is performed using 0-1 normalization.

3. A system for selecting and training an athlete in soccer according to claim 1, wherein the weight coefficients for each feature are recalculated in the preprocessing module by:

firstly, selecting features by using a plurality of methods, carrying out independence test on the features and the labels by using chi-square test, and sequencing according to the independence from small to large; calculating the information gain of each characteristic by using an information gain method, and sequencing the information gains from large to small; the Lasso method regards the process as multivariate linear regression, each feature has a corresponding weight coefficient coef, the positive and negative values of the weight coefficient coef represent whether the feature is positively correlated or negatively correlated with a target value, and the features are sorted from large to small according to the absolute value of coef; calculating a correlation coefficient of each feature and a target value by a pearson correlation coefficient method, and sorting according to the absolute value of the correlation coefficients from large to small; and finally, endowing each feature with a weight value from 1 to n according to the sorting result, adding the weight values of the four sorting results, and taking the reciprocal of the sum to obtain a weight coefficient of each feature, wherein the greater the weight coefficient, the more important the weight coefficient is.

4. A system for selecting and training an athlete in soccer according to claim 1, wherein said specific process of calculating the athlete's excellence probability in the prediction module is:

the prediction result of the adopted decision tree is excellent or not excellent, a plurality of decision trees are adopted for prediction, and the value of p is the ratio of the number of excellent decision trees to the total number of all decision trees.

5. A system for culling and training a soccer player according to claim 1, wherein the specific process of calculating the player's traits in the potential mining module is:

firstly, reading the average level vector of the excellent athlete calculated by the preprocessing module in the system training stage, then calculating the difference vector of the excellent level vector and the feature vector corresponding to the sample, and taking all items greater than 0 in the difference vector as the characteristic of the sample, namely, the characteristic represents that the overall level of the athlete does not reach the standard, but the excellent level is reached on the characteristic items.

6. A system for selecting and training a soccer player according to claim 3, wherein the specific process of calculating the top n items that the player most needs to improve is performed in the potential assessment and merit analysis module:

firstly, reading an average level vector of excellent athletes obtained by calculation of a preprocessing module in a system training stage, then calculating a difference value between the excellent level vector and a sample feature vector, and performing dot product on a feature weight vector consisting of the difference vector and a feature weight value to obtain a result vector; the item with larger value in the result vector represents that the athlete has poor performance on the item and the item has large influence on the final judgment result, and the item with the largest value in the top n is taken as the top n items which need to be improved most.

Images