Competition prediction and fitness behavior based on GA-SVM algorithm and PCA model

Abstract

Competitive sports require athletes to operate in real time, and there are many uncertainties. At present, there are few applications of artificial intelligence in the prediction of competitive sports, and the relevant literature about fitness motivation is rare. Based on this, this study is based on the machine learning algorithm and uses the support vector machine to build the competitive sports model and fitness motivation evaluation. At the same time, this study combines the actual situation to construct a corresponding factor analysis model for racing sports, and this factor analysis is a combination of data mining and machine learning. Only by adopting appropriate measures can students’ motivation of physical fitness be effectively fostered and stimulated, their active participation in physical exercise and lifelong fitness habits be fostered. On the basis of traditional SVM method, PCA-SVM model is constructed to further improve the prediction accuracy and validity of fitness motivation. In this paper, the principal components of eight kinds of operation behavior are extracted; fitness motivation is not only the direct reason for college students to participate in fitness exercise, but also the motive force of fitness behavior. Grid Search algorithm is selected to optimize the parameters of SVM. The recognition rate of Grid Search-SVM is 94.79%, and satisfactory results are obtained.

Keywords

Support vector machine racing sports regression model GA-SVM algorithm

1 Introduction

From the current literature on competitive sports, it is found that most literature are qualitative descriptions, which are mostly commentary and speculation. Even with some quantitative literature, these documents use a single method to analyze technical and tactical characteristics. In order to expand the diversity of research and improve the research level of technology and tactics [1], this paper describes this purpose. This paper describes in detail the methods of establishing various model methods, analyzes the connotations of various models, and integrates the characteristics of various models. Moreover, this quantitative diagnostic study has positive implications for tennis skills and tactics.

The establishment of the model not only provides convenience for researchers, but also has an extraordinary effect on the training and competition of athletes. What the athletes want most before the game is to restrain the opponent’s technology, and then strengthen it to grasp the initiative on the field and finally win. At the same time, the establishment of the model and the comparison of the two information, can quickly get the strengths and weaknesses of the players, and thus achieve a good state of knowing oneself and knowing each other, adding a competitive sports system to the competition is a typical nonlinear, time-varying, uncertain system. In addition, it is difficult to describe it accurately by traditional motion mechanics analysis [2]. Different from the traditional thinking about racing sports research, the research object of this paper is the collected data about the racing sports and the actual training environment, rather than the dynamic analysis system.

Motivation is the psychological or internal motivation to promote a person’s activities, which can stimulate and maintain human activities, and make the activities towards a certain goal to meet individual needs, wishes or ideals. Fitness motivation is the internal reason or motive force to promote and maintain college students’ participation in fitness. It plays a role of orientation, initiation, regulation, strengthening and maintenance of College Students’ fitness behavior, and has an important impact on the effect of fitness activities.

The formation of College Students’ fitness motivation is influenced by the internal needs and external conditions of College students. It is based on the needs of College Students’ physical activities and is influenced by external stimulation or inducement. Physical fitness, improving skills and self-expression may be the need for students to participate in fitness activities. Excellent venues, equipment, equipment, beautiful physical exercise environment, the positive influence of families, schools, teachers, classmates, and rich and colorful physical exercise activities can all become an attractive environmental factor for students to participate in fitness sports. When students have the need to engage in sports activities, they will have the idea of meeting the needs of sports activities. When this requirement is adapted to the motivation of sports environment and conditions, it becomes a driving force to promote students to participate in sports. College students’ internal fitness motivation is difficult to observe from the outside, but when its intensity reaches a certain level, it will stimulate them to participate in fitness exercises [3].

2 Related work

Through the correlation analysis and cluster analysis,Wen Fang et al., “Establishment and Application of Special Performance Prediction Models for Aerobics Athletes” and Wang Ni’s “Establishment and Application of Special Performance Prediction Models Based on Neural Network Aerobics Athletes”, selected a number of major body shapes and quality indicators most closely related to the special action technique from the 32 body shape and quality indicators of the primary selection, and established a multiple regression equation for major body shape, quality and athletic performance. The variance test shows that the evaluation model of body shape and quality of bodybuilders established by multiple stepwise regression can accurately predict the results of aerobics. At the same time, the case analysis shows that the evaluation model can effectively predict the special sports performance of aerobics athletes, assess the development level of athletes’ physical quality, and test the training effect of athletes [4].

In view of the problems existing in the traditional multi-regression analysis model for the Olympic Games performance prediction based on the principle of econometrics, Wang Guofan et al. ’s “Analysis of Olympic Performance Prediction Based on GA and Regression Analysis” proposes a prediction model that combines GA, competitive sports strength assessment and regression analysis. The model uses GA to conduct dynamic and supervised optimization evaluation of competitive sports strength. At the same time, based on this, combined with the regression analysis and the global optimization ability of GA, the complex nonlinear relationship between the achievements and impact factors of the participating countries is established. Through GA, the effective supervision and calculation of the competitive sports strength level of participating countries (regions) can be realized, and the optimal competitive sports strength evaluation can be dynamically explored to optimize the predictive model based on competitive sports strength. At the same time, it improves the objectivity of the prediction model, and has high precision and good stability in the medal (gold medal) number prediction [5].

Song Ailing and others’ “High-level women’s 100m hurdle special performance prediction model based on BP neural network” proposes a clear function mapping relationship between athletes’ special performance and physical quality training level. By establishing a theoretical model of the relationship between athletes’ special grades and physical fitness, accurate predictions of athletes’ special grades can be achieved, and a scientific theoretical basis for developing physical fitness preparation training can be provided. At the same time, according to the special quality value of China’s 100 m hurdle athlete Liu Jing from 1995 to 1998, they used Matlab simulation analysis platform and artificial neural network function mapping method to construct and select the appropriate BP neural network to establish a special prediction model for women’s 100m hurdle athletes. The model has high prediction accuracy, and its extended application can provide important theoretical guidance for the scientific training of women’s 100 m hurdle athletes [6].

Fu Yan and Li Feng’s “Application of ARMA Model in China’s Sports Stock Price Forecast” took the 1510 daily closing price index of China Sports Industry Co., Ltd. as a research sample. Moreover, they conducted time series analysis and established an autoregressive moving average model of the stocks of the Chinese industry (ie ARMA model). The fluctuations in the stock price of the Chinese industry are affected by many uncertain factors, and these or some of these intricate correlations exist. Therefore, it is very difficult to theoretically predict and analyze the stock price trend of the physical industry by relying on traditional linear regression analysis. The time series model in the field of econometrics is to examine the changing trends of variables in a period of time. Moreover, this study is based on the stochastic change trend of the daily stock market price index of the Chinese sports industry in a certain period of time, and through the description and analysis of the changes to achieve the purpose of prediction. Applying the time series model to the forecast and analysis of the stock market daily closing price index for the Chinese industry is still the first time in the field of sports science research, and the ideal results have been obtained through research, and the difference between the obtained model prediction values and the actual observation values is very small [6].

In 1968, Buchan and Flewitt used nonlinear theory to analyze the steering system characteristics of the kayaking itself. It mainly studies the mechanism of the automatic rudder, and the dynamic model at this time is very simplified [7].

The earliest analysis of the performance of racing sports is the Cuniss. He discussed wind pressure angle and yaw in detail in his published academic papers and proposed a model that is considered a linear system [8].

TitloW first proposed six degrees of freedom analysis for maneuvering kayaking. This analysis is based on a linearized equation of motion and explores a combination of velocity, sail force coefficient, rudder characteristics, displacement, eigenvalue stability, and model shape, but only considers downwind navigation and fixed control [9].

The first study of the effects of balancing and maneuvering performance on hull attachments and devices was done in 1974 by Letcherp. He has done a lot of experiments on the movement of rigs, rudders and keels [10].

Dougal Harris studied the performance of racing sports in the wind. At the same time, he compiled a speed prediction program to predict the average speed of racing sports in the downwind. The software is mainly used to study the design parameters of hulls, appendages, sails and rigging that affect the racing sports in the waves [11].

In 2004, the French K.Roncin and J.M.Kobus analyzed the simulation of two racing sports in the windward segment of the competition in “Dynamic Simulation of Two Racing Sports Competitions”. The results of several simulation experiments show that the simulation results are very compatible with the speed limit diagram of this type of racing sports [12].

Li Jing et al. [13] used a gradient histogram (HOG)-based support vector machine (SVM) algorithm and a Haar-based Adaboost cascade classifier to identify the racing process. Due to the limited number of samples, the HOG-SVM detection algorithm has a high recognition rate in this experiment, but its recognition time is more than four times that of Adaboost. Therefore, choosing the Adaboost algorithm with a sufficient number of samples will be more practical than the SVM algorithm.

Lu Shengfang [14] proposed a CLBP eigen operator under the Sobell gradient and added this feature as a training set to the SVM for training. Li Juan [15] proposed a motion detection algorithm based on neural network, which realizes the optimization of neural network threshold by parallel genetic algorithm. Wu Yingyong [16] proposed a SVM motion recognition algorithm based on SIFT features.

Through the collection and analysis of data, it can be seen that many experts and scholars have already thoroughly studied the mathematical modeling of racing sports, but there are fewer articles applied to natural sciences, and the articles applied to forecasting research are rare.

Traditional statistical modeling simply compares existing data between players and compares them between data. Moreover, it is unable to formulate corresponding tactical strategies according to the situation of the opponent, and targeted training to strengthen the level of individual sports skills.

A questionnaire survey was conducted among 2250 subjects. Indicators of each respondent included: source, gender, height, weight, staying up late, regularity of three meals, nutrition, snacks, smoking, active health preservation, health awareness, self-evaluation compared with the previous year’s health status, physical self-awareness, self-concern about their health status. Situation, physical health self-concern, whether to understand diet and nutrition, regular physical examination every year, self-evaluation of sports awareness-physical exercise can enhance physical fitness and prevent diseases, physical health will produce a sense of urgency for exercise, whether to understand exercise-related knowledge and skills, whether to understand how to rationally arrange exercise plans, exercise scientifically, master more than two kinds of luck Technical points of sports, know how to prevent injuries in sports, exercise internal driving force, feel the excitement and excitement brought by sports, want to maintain and improve sports skills, want to learn new sports skills, want to control weight for better looking, want to exercise muscles for better looking, want to enhance physical and mental health, whether or not To improve cardiovascular function and health, friends want me to participate in fitness, make new friends through exercise, self-efficacy - even in the rain, snow, haze and other weather, I will still adhere to exercise a total of 86 indicators.

3 Theoretical analysis

3.1 Support vector regression machine

Before discussing the support vector regression machine, the regression problem is first looked at: A training data set T ={ (x₁, y₁) , ⋯ , (x_i, y_i) , ⋯ , (x_k, y_k) } is given. Among them, x_i ∈ Rⁿ, y ∈ R. Accordingly, a real-valuedfunction z (x) is looked up on Rⁿ such that y = z (x) can be used to infer the output Y [17] corresponding to any input X.

As can be seen from the above, the regression problem is the same as the structure of the classification problem. The difference is that the output of the two is different, the output of the regression problem can be any real number, and the classification problem only allows binary output (for multi-class partitioning problems, finite-value output is allowed).

The goal of the regression problem is to infer the output of the new given input data from the knowledge learned from the training data set. That is, the curve that approximates each input point is found (Fig. 1). As can be seen from Fig. 2, the training data set for the linear regression machine is T ={ (x₁, y₁) , ⋯ , (x_i, y_i) , ⋯ , (x_k, y_k) } x_i ∈ Rⁿ. The goal is to find the fitting function y = w • x + b and at the same time minimize ɛ. By adding or subtracting each item of the training set, the regression machine problem can be transformed into a classification problem, and then the convex programming can be used to solve such problems. Therefore, the original problem is equivalently transformed into the following question [18]:

Fig.1

Schematic diagram of the geometric meaning of the regression machine.

Fig.2

Schematic diagram of linear regression machine.

$min_{w_{\cdot} b} \frac{1}{2} {∥ w ∥}^{2}$ (1) $| w • x_{i} + b - y_{i} | ⩽ ɛ i = 1, 2, \dots, k$

By introducing the slack variable ξ and the penalty factor C, the linear regression support vector machine can be transformed into a convex quadratic programming problem [19]: $\begin{matrix} min_{w_{\cdot} b_{\cdot} ξ} \frac{1}{2} {∥ w ∥}^{2} + c \sum_{i}^{k} (ξ_{i} + ξ_{i}^{'}) \\ s . t . - (ɛ + ξ_{i}^{'}) ⩽ w • x_{i} + b - y_{i} ⩽ ɛ + ξ_{i} \\ ξ ⩾ 0, ξ^{T} = (ξ_{i}, ξ_{1}^{'}, \dots, ξ_{K}, ξ_{K}^{'}), i = 1, 2, \dots, k \end{matrix}$ (2)

For the line ɛ support vector regression machine, the following algorithm can be established ^[20]:

Algorithm 1 (Linear ɛ - Support vector regression machine)

Step 1: Training sample data is given $T = {(x_{1}, y_{1}), \dots, (x_{i}, y_{i}), \dots, (x_{k}, y_{k})}$ (3)

Among them, x_i ∈ Rⁿ, y ∈ R. Moreover, the penalty factor C and the loss parameter ɛ are selected.

Step 2: The convex function plan is constructed: $\begin{matrix} min_{λ \in R^{2 K}} \frac{1}{2} \sum_{i, j}^{k} (λ_{i}^{'} - λ_{i}) (λ_{j}^{'} - λ_{j}) (x_{i} \cdot x_{j}) \\ + ɛ \sum_{i = 1}^{k} (λ_{i}^{'} + λ_{i}) - \sum_{i = 1}^{k} y_{i} (λ_{i}^{'} - λ_{i}) \\ s . t . \sum_{i = 1}^{k} (λ_{i} - λ_{i}^{'}) = 0, 0 ⩽ λ_{i} ⩽ c \\ i = 1, 2, \dots, k \end{matrix}$ (4)

Get $\bar{λ} = {(\bar{λ_{1}}, {\bar{λ_{1}}}^{'}, \dots, \bar{λ_{K}}, {\bar{λ_{K}}}^{'})}^{T}$

Step 3: Calculate: $\bar{b}$ :

The component $\bar{λ_{j}}$ or $\bar{λ_{k}}$ of $\bar{λ}$ located in (0, c) is selected.

If $\bar{λ_{j}}$ is selected, then $\bar{b} = y_{j} - \sum_{i = 1}^{k} (\bar{λ_{i}} - {\bar{λ_{i}}}^{'}) (x_{i} \cdot x_{j}) + ɛ$ (5)

If $\bar{λ_{k}}$ is selected, then $\bar{b} = y_{k} - \sum_{i = 1}^{k} (\bar{λ_{i}} - {\bar{λ_{i}}}^{'}) (x_{i} \cdot x_{k}) + ɛ$ (6)

Step 4: The decision function is obtained: $y = h (x) = \sum_{i = 1}^{k} ({\bar{λ_{i}}}^{'} - \bar{λ_{i}}) (x_{i} \cdot x) + \bar{b}$ (7)

3.2 Nonlinear support vector regression machine

For linear indivisible problems, we can map the input space to the high-dimensional feature space (Hilbert): Φ : x = Φ (x). In this way, the algorithm returns to the linear support vector machine method described above. After the transformation, the problem of the original nonlinear support vector regression machine is transformed into the convex quadratic programming problem of Hilbert space, which can be described as [21]:

$\begin{matrix} min_{w . b . ξ} \frac{1}{2} {∥ w ∥}^{2} + c \sum_{i}^{k} (ξ_{i}^{'} + ξ_{i}) \\ s . t . - (ɛ + ξ_{i}^{'}) ⩽ w \cdot Φ (x_{i}) + b - y_{i} ⩽ ɛ + ξ_{i} \end{matrix}$ (8)

Among them, $ξ ⩾ 0, ξ^{T} = (ξ_{1}, ξ_{1}^{'}, \dots, ξ_{k}, ξ_{k}^{'}), i = 1, 2, \dots, k$

Similarly, the Lagrange function can be introduced to transform it into a dual problem.

Assuming that $λ = {(λ_{1}, λ_{1}^{'}, \dots, λ_{K}, λ_{K}^{'})}^{T}$

$η = (η_{1}, η_{1}^{'}, \dots, η_{K}, η_{K}^{'})$ represents the Lagrange multiplier vector. Then, the dual problem of the original nonlinear support vector machine regression problem can be derived as follows [22]:

$\begin{matrix} min_{λ, η \in R^{2 K}} \frac{1}{2} \sum_{i, j}^{k} (λ_{i}^{'} - λ_{i}) (λ_{j}^{'} - λ_{j}) (Φ (x_{i}) \cdot Φ (x_{j})) \\ + ɛ \sum_{j = 1}^{k} (λ_{i}^{'} + λ) - \sum_{i = 1}^{k} y_{i} (λ_{i}^{'} - λ_{i}) \end{matrix}$ (9)

$\begin{matrix} s . t . \sum_{i = 1}^{k} (λ_{i} - λ_{i}^{'}) = 0 \\ c - λ_{i} - η_{i} = 0, i = 1, 2, \dots, K \\ λ_{i} ⩾ 0, η_{i} ⩾ 0, i = 1, 2, \dots, K \end{matrix}$ (10)

For the nonlinear support vector regression machine, the following algorithm can be established:

Algorithm 2 (Nonlinear Support Vector Regression)

Step 1: Training sample data is given

$T = {(x_{1}, y_{1}), \dots, (x_{i}, y_{i}), \dots, (x_{k}, y_{k})}$ (11)

Among them, x_i ∈ Rⁿ, y ∈ R. A suitable transform Φ : x = Φ (x) is chosen such that the input space can be mapped to the Hilbert space and the penalty factor C and the loss parameter ɛ are selected.

Step 2: The convex function plan is constructed [23]: $\begin{matrix} min_{λ, η \in R^{2 K}} \frac{1}{2} \sum_{i, j}^{k} (λ_{i}^{'} - λ_{i}) (λ_{j}^{'} - λ_{j}) (Φ (x_{i}) \cdot Φ (x_{j})) \\ + ɛ \sum_{j = 1}^{k} (λ_{i}^{'} + λ) - \sum_{i = 1}^{k} y_{i} (λ_{i}^{'} - λ_{i}) \end{matrix}$ (12) $\begin{matrix} s . t . \sum_{i = 1}^{k} (λ_{i} - λ_{i}^{'}) = 0 \\ 0 ⩽ λ_{i} ⩽ c, i = 1, 2, \dots, K \end{matrix}$ (13)

Step 3: Calculate $\bar{b}$ :

The component $\bar{λ_{j}}$ or $\bar{λ_{k}}$ of $\bar{λ}$ located in (0, c) is selected.

If $\bar{λ_{j}}$ is selected, then

$\bar{b} = y_{j} - \sum_{i = 1}^{k} (\bar{λ_{i}} - {\bar{λ_{i}}}^{'}) (Φ (x_{i}) \cdot Φ (x_{j})) + ɛ$ (14)

If $\bar{λ_{k}}$ is selected, then

$\bar{b} = y_{k} - \sum_{i = 1}^{k} (\bar{λ_{i}} - {\bar{λ_{i}}}^{'}) (Φ (x_{i}) \cdot Φ (x_{j})) + ɛ$ (15)

Step 4: The decision function is obtained:

$y = h (x) = \sum_{i = 1}^{k} ({\bar{λ_{i}}}^{'} - \bar{λ_{i}}) (Φ (x_{i}) \cdot Φ (x_{j})) + \bar{b}$ (16)

3.3 Kernel function

It can be seen from the calculation process of the previous nonlinear algorithm that the transformation effect of Φ : x = Φ (x) always appears in the form of inner product. If Φ (x_i) · Φ (x) does not appear separately, we can define the inner product Φ (•) • Φ (•) in Rⁿ × Rⁿ space as the kernel function. It is specifically written as [24]: $K (x, x) = Φ (x) \cdot Φ (x)$ (17)

Among them, (•) is the inner product of the Hilbert space. The regression problem has the same structure as the classification problem, except that the output has a different range of values. The classification problem is generally a binary problem (or finite value), and the output of the regression machine can be any real number.

According to the pattern recognition theory, it is known that although the linear separability can be achieved by transforming the low-dimensional nonlinear function into the high-dimensional feature linear space, when the dimension of the high dimension is very large, it will lead to “dimensionality disaster” and make the whole learning process quite complicated. Moreover, the use of the kernel function can subtly achieve a high-dimensional effect by calculating the inner product in a low dimension. It does not even need to know the form and parameters of Φ : x = Φ (x) specifically.

Next, we will introduce several commonly used kernel functions [25]:

Subpolynomial kernel function: $K (x, x^{'}) = {(x \cdot x)}^{D}$ (18) In the above formula, D is a positive integer.

Nonhomogeneous polynomial kernel function: $K (x, x^{'}) = {(x \cdot x + 1)}^{D}$ (19) In the above formula, D is a positive integer.

Radial basis kernel function: $K (x, x^{'}) = exp (- \frac{{∥ x^{'} - x ∥}^{2}}{σ^{2}}), σ > 0$ (20)

Neural network kernel function: $K (x, x^{'}) = tanh (ρ 〈 x, x^{'} 〉 + c)$ (21)ρ and C are certain constants;

Fourier sequence kernel function: $K (x, x^{'}) = \frac{sin (N + \frac{1}{2}) (x - x^{'})}{sin (\frac{1}{2} (x - x^{'}))}$ (22)

B-spline kernel function: $K (x, x^{'}) = B_{2 N + 1} (x - x^{'})$ (23)

The kernel function has a large impact on the results of regression prediction. Therefore, it is necessary to select a kernel function that can make prediction performance better before making predictions. In this paper, linear kernel function, polynomial kernel function, RBF kernel function and sigmoid kernel functions are selected. The experimental results are as follows:It can be seen from the experimental results that the RBF kernel function is the choice of the optimal kernel function. When other parameters are fixed, it performs best in terms of correlation coefficient and mean square error. Therefore, the RBF kernel function is selected in the next simulation experiment.

3.4 Principal component analysis

The idea of Principal Component Analysis (PCA) is to reduce dimensionality, reduce data volume and remove overlapping information. In this chapter, PCA is used to reduce the dimension and eliminate the correlation, so as to minimize the interference between the data.

$\begin{matrix} X & = & (x_{ij})_{n \times p} = [\begin{matrix} x_{11} & x_{12} & \dots & x_{1 p} \\ x_{21} & x_{22} & \dots & x_{2 p} \\ \dots & \dots & \dots \\ x_{n 1} & x_{n 2} & \dots & x_{np} \end{matrix}] \\ = & (\begin{matrix} e_{1}^{T} \\ e_{1}^{T} \\ ⋮ \\ e_{n}^{T} \end{matrix}) = (x_{1}, x_{2}, \dots, x_{p}) \end{matrix}$ (24)

It can be abbreviated as:

$\begin{matrix} F_{j} & = & \sum_{k = 1}^{p} a_{jk} x_{k} = a_{j 1} x_{1} + a_{j 2} x_{2} + \dots \\ + a_{jp} x_{p} (j = 1, 2, \dots, p) \end{matrix}$ (25)

The objective of principal component analysis is to determine the coefficient vectors so as to maximize the variance of Fj. Var (Fj) denotes the degree of dispersion of the transformed values of all samples relative to the mean. The calculation is as follows: $Var (F_{j}) = \frac{1}{n} \sum_{i = 1}^{n} {(F_{ij} - \bar{F_{j}})}^{2}$ (26)

4 VMG prediction model

Based on the collected sports data information, starting from the history, law and current situation of racing sports data, this paper uses SVM support vector regression machine to achieve the prediction of the target value we are interested in, which is of great significance to the scientific training of athletes.

In view of the variety of data and the large amount of noise involved in the data, the measured data needs to be preprocessed. We divide the measured data into two parts: training data and data to be predicted. The learning machine model is trained and completed through the training data, and then the test results are used to verify the prediction results. The forecasting process of the entire racing sports data mainly includes the following key points.

Data preprocessing

In order to improve the accuracy of data prediction, it is necessary to denoise and normalize the data. The new training data obtained is $T_{n} ew = {(x_{1}^{'}, y_{1}^{'}), \dots,$ $(x_{i}^{'}, y_{i}^{'}), \dots, (x_{k}^{'}, y_{k}^{'})}$ . Among them, $x_{i}^{'} \in R^{n}, y^{'} \in R$

Training and prediction of SVM prediction model

The training points of the training data T_new after the processing, such as (x_i, y_i), are respectively moved up and down by a suitable distance $\bar{ɛ}$ to the new training point $(x_{i}^{+}, y_{i}^{+})$ , $(x_{i}^{-}, y_{i}^{-})$ . This turns the original regression problem into a classification problem. Then the original problem can be transformed into a convex programming problem, and the Lagrange method is used to solve the decision function.

Algorithm 3 (SVM prediction algorithm)

Input: Training sample data and data to be predicted for the time series data of the racing sports.

Output: The predicted result of the time data of the sports class.

Step 1: The racing sports data is collected, and the data is pre-processed and normalized;

Step 2: Appropriate training samples are selected, and the error parameters MSE (Mean Squared Error, MSE), penalty factor c and other related parameters are set;

Step 3: Select the kernel function. This paper chooses RBF (radial basis kernel function): $K (x, x') = \exp (- \frac{γ {‖ x' - x ‖}^{2}}{σ^{2}})$ . According to the training sample, the prediction function is obtained, and the model training is completed;

Step 4: Predictive data is substituted to complete the prediction of the time series;

Step 5: The mean squared error MSE and the Square Correlation Coefficient (SCC) are used to evaluate the prediction results. If the result does not meet expectations, the program jumps to Step.3 and retrains the model.

Step 6: The algorithm ends.

$\begin{matrix} MSE = \frac{1}{K} \sum_{K = 1}^{K} {(f (x_{k}) - y_{k})}^{2}; \\ γ^{2} = \\ \frac{{(K \sum_{K = 1}^{K} f (x_{k}) y_{k} - \sum_{K = 1}^{K} f (x_{k}) \sum_{K = 1}^{K} y_{k})}^{2}}{(K \sum_{K = 1}^{K} f {(x_{k})}^{2} - \sum_{K = 1}^{K} f {(x_{k})}^{2}) (K \sum_{K = 1}^{K} y_{i}^{2} (\sum_{K = 1}^{K} y_{i}))} \end{matrix}$ (27)

In the process of SVM model training, K-fold cross validation is used. K-fold cross-validation is used to validate and train each sample data in K data sets. The basic principle of the grid search method is to divide the parameters C and G into grids within a certain range of values and traverse all points in the grid. Finally, the group C and G with the highest accuracy of the training set are selected as the best parameters.

RBF is used as the kernel function of grid search support vector machine. The range of C is set to [2–5 , 25], the range of G is set to [2–10 , 25], and the step distance is set to 0.1. The K-fold cross-validation method was used to test the training set, where K = 10, and the local optimal parameters C = 4.69842 and g = 0.7325 were obtained. Fig. 3 is a three-dimensional graph of optimization parameters for grid search.

Fig.3

Grid-search optimization parameter selection results.

5 Model verification

Firstly, the traditional support vector machine (SVM) method is used to solve practical problems. Because of the high correlation between the characteristic indexes of training samples, the principal component analysis (PCA) method is adopted to reduce the dimensionality and eliminate the interference caused by data overlap, thus solving the traditional support vector machine (SVM) method to explain the redundancy problem.

According to the scale of influencing factors of adolescent sport adherence behavior, each psychological type has more effective incentive methods to promote adolescent sport adherence formation. The adolescent sport behavior is divided into eight psychological types:

Conformity Psychology (131): Participation in sports has a strong follow-up. Motivation methods: group sports, sports circle of friends,etc.

Admiration Psychology (110): Affected by sports stars and other admirers, they will participate in the corresponding sports. Motivation methods: admirers’movement information push, inspirational story push, etc.

Achievement Psychology (197): Get a certain sense of achievement through the sports meeting. Motivation methods: sports ranking, sports sharing rewards, etc.

Social Psychology (139): Promote family relationships through sports or maintaining friendships. Motivation: Develop and organize family and collective sports programs.

Showing off Psychology (207): Through the completion of the movement to meet their own desire for honor. Motivation methods: Sports Got Talent List, Sports Rank.

Practical Psychology (1139): Through sports to meet their own goals. Motivation method: Make personalized sports program, reasonable arrangement of sports time and space.

Enjoyment Psychology (87): Relax your body and mind by exercising, and reduce your stress. Motivation method: Make personalized sports program, reasonable arrangement of sports time and space.

Knowledge-seeking Psychology (240): Through sports learning related sports knowledge, master sports skills. Incentive methods: scientific fitness knowledge push, sports skills mastery test, etc.

Combining with the characteristics of actual data, a method of combining PCA with optimized SVM prediction model is constructed, which improves the accuracy and validity of model classification. 60% of the samples randomly selected from each class are used as training sets. The number of training samples was 118, 52, 66, 83, 78, 683, 124 and 144 respectively. The remaining samples were 79, 35, 44, 56, 53, 456, 83 and 96 as test sets.

Table 1
Test results of nuclear function selection

Kernel function type r² (%) MSE Convergence time (s) Other parameters (c,g)

linear kernel function 90.65 0.005 0.84 (2,0.02)

polynomial kernel function 41.40 0.085 0.85 (2,0.02)

RBF kernel function 93.77 0.005 0.84 (2,0.02)

sigmoid kernel functions 88.44 0.006 0.85 (2,0.02)

Kernel function type	r² (%)	MSE	Convergence time (s)	Other parameters (c,g)
linear kernel function	90.65	0.005	0.84	(2,0.02)
polynomial kernel function	41.40	0.085	0.85	(2,0.02)
RBF kernel function	93.77	0.005	0.84	(2,0.02)
sigmoid kernel functions	88.44	0.006	0.85	(2,0.02)

Table 2 is the classification result of traditional SVM model, in which the accuracy of training model is 100%.

Table 2

Classification results of traditional SVM models

type	Predictive group relationship
	A	B	C	D	E	F	G	H
Training set	Correct recognition rate (%)
A	78								100%
B		66
C			118
D				83
E					124
F						683
G							52
H								144
Test set	Correct recognition rate (%)
A	48	2		2	1				87.92%
B		39	2				1	2
C	4	2	66	2	2	3
D		2		48		3		3
E		3	2		62		2	4
F	8	4	11		7	410	16
G				1			33	1
H	1	3		3		2		87

Table 3

Classification results of Grid Search-SVM model

type	Predictive group relationship
	A	B	C	D	E	F	G	H
Training set	Correct recognition rate (%)
A	78								100%
B		66
C			118
D				83
E					124
F						683
G							52
H								144
Test set	Correct recognition rate (%)
A	49	2		1	1				94.79%
B	1	40	1					2
C		2	74	1	2
D				53		2	1
E		2	1	1	77		2
F	5		4	7		436	4
G							35
H			3		2			91

The experimental results show that GA-SVM is superior to Grid Search-SVM in recognition rate of 94.79%. It can be seen that GA-SVM undoubtedly does better in generalization ability. In the process of computer operation, it is found that the traditional SVM model may miss the optimal solution in the search process. Grid Search-SVM is better than the traditional SVM in global variable search. Therefore, Grid Search-SVM is an efficient, parallel and comprehensive classification model. Its automatic control search process determines the optimal solution by self-adapting, and has good convergence.

This study will use SVM to establish a regression prediction model to regression fit the speed and direction of the race-like movement. Before doing regression prediction, we assume that the value of each second in speed and direction is related to multiple factors in the previous second. That is, factors such as the speed and direction of the previous second are treated as independent variables.

The simulation data of the following test is based on the training data of athletes on July 18, 2013, with a total of 891 articles (The recording frequency of the data is 1 / sec, and the data contained in one data has the moving speed, the moving direction, the wind speed, and the wind direction). In this paper, the first 889 data are used for the training of the speed SVM regression prediction model. Figs. 4 and 5 depict the original training data (speed and direction).

Fig.4

Speed initial data.

Fig.5

Direction initial data.

According to the above steps, the data is first normalized, the kernel function is selected, and then the optimal parameters C and g are selected according to the cross-validation algorithm (CV) and the grid search algorithm (Gs). Then, these parameters are used to predict speed and direction. It should be pointed out that in the simulation, the rough range of C and g is set to [-8, 8] and [-8, 8]. Then, according to the rough selection result graph, the precise selection is performed. At this time, the search step is 0.5, and the test set is divided into three parts. On this basis, the comparison between the original velocity obtained and the regression prediction results is shown in Fig. 6.

Fig.6

Comparison of original speed and regression prediction results.

Fig.7

Comparison of original direction and regression prediction results.

It can be seen from the above simulation result diagram that, by using the grid search method and cross-validation, the optimal parameters c, g for predictive performance are C = 1, g = 0.5. At this time, the mean square error and correlation coefficient are: MSE = 0.00397899, R = 95.2402%. In addition, we used the trained model to predict the speed of the next second. The actual value of the next second speed is 5.6 m/s, and the result predicted by SVM regression is 5.35 m/s, and the prediction error is 4.5%. In addition to the speed prediction, we used the same sample data to predict the direction of the next second. The simulation results are as follows: the mean square error MSE = 0.00872579 and the correlation coefficient R = 92.3323%.

6 Analysis of results

This study first established a data warehouse for the collected data on racing sports and made relevant explanations and pre-processing. Then, this paper uses the LIBSVM tool to make regression predictions on the speed and direction of the sports. After that, the kernel function of SVM is selected in this paper, and the optimal penalty factor C and kernel function parameter g are selected by cross-validation and grid search algorithm to optimize the performance of the prediction regression machine. The experimental simulation results show that when the kernel function is selected as RBF and the cross-validation method and the grid search algorithm are used to search for the appropriate C, g, the performance of the regression predictor can be achieved, which can be seen by the error graph between the actual data and the predicted data, thereby obtaining the corresponding experimental research results. At the same time, based on the above research, the corresponding system structure can be constructed.

First, we can build a parameter database related to the race class movement. In the next simulation experiment, the factors (wind speed, wind direction, etc.) that have a large influence on the objective function (speed, direction) are first selected from the database, and then these parameters are used as inputs to predict the output of the objective function. It should be pointed out that, unless otherwise specified, the wind speed and wind direction mentioned in this paper refer to the wind speed of the relative athletes and the wind direction of the relative athletes.

There are many toolboxes for SVM, such as Weka, SVMlit, Matlab’s own SVM toolbox, and so on. The functions of Matlab that implement SVM only support classification problems. However, LIBSVM can not only solve the classification problem, but also support the regression prediction problem. Therefore, this article uses LIBSVM, and the implementation of SVM in this paper is done through the LmSVM toolbox.

The LIBSVM toolbox is a software package invented by Lin Zhiren et al. to solve classification problems and regression problems. This package has the advantage of requiring less adjustment of the SVM parameters required to be adjusted (many default parameters are set). Moreover, the package contains many language versions. In addition, there is no uniform standard in the international selection of SVM parameters and kernel functions. In other words, the choice of parameters and kernel functions that can make the SVM learner optimal is only based on experience, experimental comparison, or relying on the interactive check function of the software package.

The performance of the SVM regression machine is mainly affected by the error penalty parameter c and the choice of the kernel function and its parameters (such as g). (1) The penalty parameter c is essentially a compromise between the complexity of the algorithm and the accuracy of prediction of the prediction result. In other words, adjusting the size of C is a ratio that changes the empirical risk of the SVM learner and the predicted confidence range. When the value of C is relatively large, it means that the penalty for empirical error is relatively small. At this time, the complexity of the learning machine is low but the empirical risk value is relatively large. However, when the value of c exceeds the fixed value, the complexity of the learning machine becomes very large and the empirical risk value and promotion ability hardly change. (2) Different kernel functions have different effects on regression performance. The experiments in this paper will select the appropriate c and g through cross-validation and grid search to optimize the performance of the SVM.

The selection of the kernel function of the support vector machine, the penalty factor, and the kernel function parameters have a significant impact on the performance of the SVM. Therefore, the relevant experimental analysis is done in this paper. By comparing the effects of several kernel functions on the SVM modeling side, it is found that the RBF kernel function can make the SVM’s prediction performance the best when dealing with the speed prediction problem. In addition, regarding the selection of parameters, there is no unified conclusion in the world at present, and the commonly used ones are the trial and error method and the cross-validation method. Moreover, this paper chooses cross-validation and grid search method for optimization. The experimental results show that this parameter optimization method can significantly improve the performance of the SVM regression machine. The prediction model based on support vector machine can effectively improve the prediction direction, and the process can be expressed as the form shown in Fig. 8.

Fig.8

data processing process of support vector machine.

With the increasing popularity of racing sports, research work related to sports will also attract more and more research enthusiasts. At the same time, sports centers in various countries will also make scientific plans for athlete management and training. Before making these plans, they must be based on forecasts. Therefore, the prediction has a greater role in promoting the research of racing sports. This paper has made good research results in the use of SVM to predict the speed of racing sports. If we can further improve and advance on the basis of existing work, it will have a significant driving impact on the research of the entire competitive sports.

7 Conclusion

With the increasing popularity of racing sports, research work related to sports will also attract more and more research enthusiasts of researchers. This study will use SVM to establish a regression prediction model to regression fit the speed and direction of the race-like motion. Before doing regression prediction, this paper assumes that the value of each second in speed and direction is related to multiple factors in the previous second. That is, this article considers factors such as speed and direction of the previous second as independent variables.

Fitness motivation is the key factor to encourage college students to actively participate in sports learning and physical exercise. The cultivation and stimulation of motivation are multi-factor and multi-level. They influence each other and run through various activities. In the process of cultivating and stimulating college students’ fitness motivation, teachers are the inducers and trainers of motivation, students are the main and core of motivation. Only by fully respecting students’ physical and mental development characteristics and internal needs and adopting appropriate measures, can teachers effectively cultivate and stimulate students’ fitness motivation, promote their active participation in physical exercise and develop lifelong fitness habits. In this paper, adolescents’ operational behavior is classified into eight psychological types, and good results are obtained. Principal Component Analysis (PCA) is used to extract the spectral information characteristics of the original sample, which is combined with Support Vector Machine (SVM) for pattern recognition and classification.

Then, the paper predicts the speed and direction, and the relevant parameters can be selected from the established data warehouse, and further, the relevant data can be analyzed and processed. After that, this paper builds a parameter database related to the racing movement. In the following simulation experiments, this paper first selects the factors (wind speed, wind direction, etc.) that have a large influence on the objective function (speed, direction) from the database, and then uses these parameters as input to predict the output of the objective function. The experimental results show that this parameter optimization method can significantly improve the performance of SVM regression machine, and the prediction model based on support vector machine can effectively improve the prediction direction.

Footnotes

Acknowledgments

This research is supported by Young Scholars Program of Shandong University (2018WLJH16); and Fundamental Research Funds of Shandong University (2016GN002), Xianliang Zhang and Lei Wang are co- corresponding authors.

References

Usharani

, Rao

T.K.R.K.

, Reddy

R.K.K.

, An Efficient Machine Learning Regression Model for Rainfall Prediction International Journal of Computer Applications, 115(23) (2015), 24–30.

Allyn Jér?me , Nicolas

, Pascal

, et al., A Comparison of a Machine Learning Model with EuroSCORE II in Predicting Mortality after Elective Cardiac Surgery: A Decision Curve Analysis, PLOS ONE 12(1) (2017), e0169772.

Ghali

, Sébastien

Ouellet

and Frasson

, LewiSpace: An Exploratory Study with a Machine Learning Model in an Educational Game, Journal of Education & Training Studies 4(1), 2015.

Lee

, Kim

S.G.

, Park

H.W.

, et al., Pre-launch new product demand forecasting using the Bass model: A statistical and machine learning-based approach, Technological Forecasting and Social Change 86(86) (2014), 49–64.

Najeebullah , Zameer

, Khan

, et al., Machine Learning based short term wind power prediction using a hybrid learning model, Computers & Electrical Engineering (2014), S0045790614001876.

Jiang

, Hamer

, Wang

, et al., SecureLR: Secure Logistic Regression Model via a Hybrid Cryptographic Protocol IEEE/ACM Transactions on Computational Biology and Bioinformatics, (2018), 1.

Hristopulos

D.T.

, Stochastic Local Interaction (SLI) Model: Interfacing Machine Learning and Geostatistics Computers & Geosciences, 85(PB) (2015), 26–37.

Taniguchi

, Sato

, Shirakawa

, A machine learning model with human cognitive biases capable of learning from small and biased datasets, Scientific Reports 8(1) (2018), 7397.

Cornejo-Bueno

, Casanova-Mateo

, Sanz-Justo

, et al., Efficient Prediction of Low-Visibility Events at Airports Using Machine-Learning Regression Boundary-Layer Meteorology, 2017.

10.

Das

S.P.

, Padhy

, A new hybrid parametric and machine learning model with homogeneity hint for European-style index option pricing, Neural Computing and Applications 28(12) (2016), 1–17.

11.

Shortridge

J.E.

, Guikema

S.D.

, Zaitchik

B.F.

, Machine learning methods for empirical streamflow simulation: A comparison of model accuracy, interpretability, and uncertainty in seasonal watersheds Hydrology and Earth System Sciences, 20,7 (2016-07-04), 20(7) (2016), 2611–2628.

12.

Yamada

, Sugiyama

, Sese

, Least-Squares Independence Regression for Non-Linear Causal Inference under Non-Gaussian Noise Machine Learning, 96(3) (2014), 249–267.

13.

Stylianou

, Akbarov

, Kontopantelis

, et al., Mortality risk prediction in burn injury: Comparison of logistic regression with machine learning approaches Burns:, Journal of the International Society for Burn Injuries 39(5), 2015.

14.

Jin

, Wu

, Vidyanti

, et al., Predicting Depression among Patients with Diabetes Using Longitudinal Data, A Multilevel Regression Model Methods of Information in Medicine 54(06) (2015), 553–559.

15.

Povak

N.A.

, Hessburg

P.F.

, Mcdonnell

T.C.

, et al., Machine learning and linear regression models to predict catchment-level base cation weathering rates across the southern Appalachian Mountain region, USA Water Resources Research 50(4) (2014), 2798–2814.

16.

Ulfenborg

, Klingalevan

, Björn.

, Genome-wide discovery of miRNAs using ensembles of machine learning algorithms and logistic regression, International Journal of Data Mining & Bioinformatics 13(4) (2015), 338.

17.

Barzegar

, Moghaddam

A.A.

, Deo

, et al., Mapping groundwater contamination risk of multiple aquifers using multi-model ensemble of machine learning algorithms, Science of The Total Environment 621 (2018), 697–712.

18.

Yip

C.F.

, Ma

A.J.

, Wong

W.S.

, et al., Laboratory parameter-based machine learning model for excluding non-alcoholic fatty liver disease (NAFLD) in the general population Alimentary Pharmacology & Therapeutics, 2017.

19.

Bai

Y.Q.

, Shen

K.J.

, Alternating Direction Method of Multipliers for ell₁ - ell₂ -Regularized Logistic Regression Model Journal of the Operations Research Society of China, 4(2) (2016), 243–253.

20.

Feng

, Huang

, Shi

, et al., Learning with the maximum correntropy criterion induced losses for regression, Journal of Machine Learning Research 16(1) (2015), 993–1034.

21.

Chen

W.C.

, Maitra

, Model-based clustering of regression time series data via APECM— an AECM algorithm sung to an even faster beat, Statistical Analysis and Data Mining 4(6) (2011), 567–578.

22.

Kitsikoudis

, Sidiropoulos

, Hrissanthou

, Machine Learning Utilization for Bed Load Transport in Gravel-Bed Rivers Water Resources Management, 28(11) (2014), 3727–3743.

23.

Eric

Bastos Görgens

, Montaghi

, Rodriguez

LCE

. A performance comparison of machine learning methods to estimate the fast-growing forest plantation yield based on laser scanning metrics Computers & Electronics in Agriculture, 116(C) (2015), 221–227.

24.

Miller

A.A.

, Bloom

J.S.

, Richards

J.W.

, et al., A Machine-learning method to infer fundamental stellar parameters from photometric light curves, The Astrophysical Journal 798(2) (2015), 122.

25.

Taylor

R.A.

, Pare

J.R.

, Venkatesh

A.K.

, et al., Prediction of In-hospital Mortality in Emergency Department Patients With Sepsis: A Local Big Data– Driven, Machine Learning Approach Academic Emergency Medicine 23(3) (2016), 10.