Research on the optimization system of athlete selection and training effect based on big data

K-means clustering algorithm is an iterative relocation algorithm, the algorithm is generally divided into two steps: the first step is iterative, the distance between each sample point is calculated, and then the corresponding sample point is divided into the closest cluster to complete the initial clustering. The second step is relocation. Recalculate the clustering center of each cluster, and divide the closest sample points into their corresponding clusters. Repeat this operation until the clustering center does not change. The algorithm is to obtain the clustering results by means of multiple iterations. The basic idea is: first of all, randomly select K initial clustering center in the given data set. Then calculate the distance of all sample points, according to the principle of nearest distance, the sample points are divided into the corresponding clusters. The clustering center of the cluster is recalculated during each iteration. When the clustering center no longer changes or reaches the specified conditions, no more clustering is carried out, to which the clustering is ended and the clustering results are output [19].

Denote the average sample distance of dataset D as: 1MeanDist(D)=1Cn2∑i=1n∑j=1id(xi,xj) where n is the number of sample points in data set D, Cn2 is the number of all optional combinations of two randomly selected sample points out of n, and d(x_i, x_j) denotes the Euclidean distance from sample point x_i to sample point x_j.

The sum of error squares for data set D is denoted as: 2SSE=∑i=1k∑x∈cid(x,ci)2 where K is the number of clusters, c_i is the clustering center of the i rd cluster C_i, and d(x, c_i) refers to the dissimilarity between data points x and c_i. Different computations of dissimilarity often lead to different clustering results, and here the Euclidean distance metric is usually used.The size of SSE can indicate how dense the sample points are.

The Euclidean distance is a distance named after the ancient Greek mathematician Euclid. Therefore, the Euclidean distance, also known as the Euclidean distance, is a common distance metric that measures the true distance between two points in n-dimensional space.

The Euclidean distance in 2D and 3D space is the distance between two points, and the 2D Euclidean distance formula is: 3d=(x1−x2)2+(y1−y2)2

The 3D Euclidean distance formula is: 4d=(x1−x2)2+(y1−y2)2+(z1−z2)2

Generalizing to n-dimensional space, the Euclidean distance is given by: 5d=∑i=1n(xi−yj)2

Euclidean distance is a measure of the distance between two sample points. The closer the distance between two sample points, the more similar these two sample points are. Conversely, the less similar these two sample points are [20].

In this section, a K-means clustering algorithm based on optimizing the initial clustering centers and profile coefficients is designed. For a better description of the algorithm, it is assumed that the dataset to be clustered is D = {x_i | x_i ∈ R_p, i = 1,2,…n} and each sample element is denoted as x_i = (x_i1,x_i2,…,x_ip),1 ≤ i ≤ n, where p is the dimension of the sample element.

The inter-cluster similarity is defined as: 6CluSim=min(d(ci,cj))

Where, i = 1,2,…,k–1, j = i+1,…,k, c_i are the cluster centers of cluster C_i, c_j is the cluster center of cluster C_j, and d(c_i, c_j) denotes the Euclidean distance between cluster center c_i and cluster center c_j. The smaller the CluSim value is, the closer the distance between two clusters is, and the more similar the clusters are to each other. Conversely, the further the distance between clusters the less similarity between clusters.

In order to better show the characteristics of the sample points, this section introduces the median for the calculation of the profile coefficient. Assuming that sample point x_i is clustered into cluster C, the median of the distance between sample points is defined as: 7ai=median(d(xi,xj)) 8bi=minp,p≠c[median(d(xi,xj))] where a_i denotes the median distance between sample point x_i and all other sample points belonging to the same cluster C, and b_i denotes the median distance between sample point x_i and all sample points in the nearest cluster that do not belong to cluster C.

Assuming that sample point x_i is clustered into cluster C, the median-based profile coefficient is defined as: 9si=bi−aimax(ai,bi)

Assuming that sample point x_i is clustered into cluster C, the median-based mean profile coefficient is defined as: 10S=1n∑i=1nsi

The median-based average profile coefficient combines intraclass and interclass distances to evaluate the reasonableness of the overall effect of clustering, and takes a value between -1 and 1. If the value is close to 1, it means that the intraclass distance of the sample is much smaller than the minimum interclass distance, which indicates that the clustering result for this data set is optimal.

The five categories of male athletes for each physical quality are summarized in Table 2. The first category had the largest number of male athletes with 401. This category had moderate test results in total score, body mass index, lung capacity, long jump, and forward body flexion endurance run, but had poorer results in sprinting and higher results in pull-ups.

The second group of male athletes had low total scores and poor results in lung capacity, body mass index, standing long jump, and forward body bends. But performed better in sprint and endurance running events. This category had the least number of male athletes with only 114.

The third category of male athletes had the lowest body mass index, higher performance in sprints, and best performance in pull-ups. There were 295 male athletes in this category.

The fourth category of male athletes had poor performance in endurance running and sprinting, better performance in long jump, forward body flexion, lunges, poor sprinting quality, and poor endurance quality, and there were 320 male athletes in this category.

The fifth category of male athletes had the highest total score, highest long jump, forward body flexion, highest lung capacity and body mass index, but lower scores in pull-ups and endurance running. Although it is the category with the highest total score, it is not well-rounded and lacks more endurance and upper body strength.

Table 3 shows the five categories of female athletes for each physical fitness. According to Table 3, it can be seen the results of female athletes' physical fitness test clustered 5 categories. The first category of female athletes had the highest total score of 91, and this category of female athletes had the best results in long jump, forward bend, endurance run and one-minute sit-up in the selection test, but this category of female athletes had the worst results in the sprint test. The best results in sprinting were obtained by female athletes in categories III and V. The lowest total scores were obtained by female athletes in category III. Category III female athletes had the lowest total score and were the worst in BMI, lung capacity, long jump, forward bending, endurance run, and one minute sit-up, but this category of female athletes had the best results in sprinting, and this category had the lowest number of female athletes, which was only 65. Category V female athletes excelled in sprinting and endurance running but were average in lunges, forward bending, long jump, and one-minute sit-ups. There were 222 female athletes in this category. Category 2 female athletes performed average in every event tested, with balanced scores in each event, no outstanding scores and no overly poor scores. The fourth category of female athletes had higher total scores, better scores in every test, and more balanced scores in every test event.

Overall, the female athletes clustered results in more complex categorical variables, mainly sprint scores, which are more different from endurance running scores. Female athletes with higher overall scores had lower sprint scores, while female athletes with lower overall scores had higher sprint scores.

The degree of construction of the athlete model directly affects the recommendation effect of the whole recommendation system, the construction of the sports model not only considers the basic information filled in by the athlete when he/she first registers in the system, and constructs the model for the athlete using the basic information, but also takes into account the change of the athlete's interest in using the system in the short term or the long term, which has to go to the updating of the athlete model, so that the athlete can achieve the purpose of accurate recommendation.

The athlete model is made up of a model of the athlete's basic physical traits and a model of their rating matrix for the program. The basic athlete profile model includes basic attribute information and interests of the athlete. Here the athlete keywords are extracted, each keyword is a basic information of the athlete, the basic information of the athlete is age, gender, BMI, sports category of interest, etc., the athlete model is represented as shown in equation (11): 11Uu→={(ku1,wu1),(ku2,wu2),(ku3,wu3),⋯,(kun,wun)}

Where, Uu→ denotes the feature vector of athlete u, kun denotes the n th keyword of athlete u, and wun denotes the weight of the n th keyword of athlete u.

The athlete-to-program rating matrix model represents the athlete's liking of the sport he or she is interested in, i.e., the rating scale. The athlete's rating level of the sport contains access, like, share, and favorite. The rating representation is implicitly obtained and expresses the athlete's preference for the recommended sport through the athlete's interface actions.

Because the application domains of recommender systems are different, there is no standard guideline to establish a unified modeling standard for each application. This shows that recommender system modeling has a great impact on recommender systems. In this paper, from the perspective of sports, according to the role of sports can be in the human body can be divided into the upper limbs, trunk and lower limbs. According to sports, the qualities of sports can be divided into strength, speed, flexibility, endurance, and sensitivity: five major sports qualities. If the sports have both upper limb and lower limb sports, the keyword of the object is set to 1, then the trunk sports is set to 0. If the sports can have the three major sports qualities of strength, speed, flexibility, their weights are set to 1, then the two major sports qualities of endurance and agility are set to 0. Eventually, the model of the sports is as shown in equation (12): 12Iu→={(pu1,wu1),(pu2,wu2),(pu3,wu4),...,(pu7,wu7),(pu8,wu8)}

Where pun represents the n nd keyword, wun represents the weight of the n th keyword, and the weight of the keyword is either 0 or 1, that is, whether the sport has this keyword or not.

The UB-CF algorithm measures and scores the degree of liking for these items based on the historical behavior of the athletes, and then calculates the relationship between the athletes' attitudes and preferences for the same item.

In this paper, Pearson correlation coefficient is used to calculate the correlation between two athletes, and the range of the calculated correlation result should be in [-1, +1], -1 means that there is an inverse influence between the two, +1 means that there is a positive influence, and 0 means that there is no correlation, and the formula is as follows: 13sim(u,v)=∑p∈Pu,v(ru,p−r¯u)(rv,p−r¯v)∑p∈Pu,v(ru,p−r¯u)2∑p∈Pu,v(rv,p−r¯v)2 where P_u,_v denotes the joint rating of athlete u and athlete v on the item, r¯u and r¯v denote the average ratings of athlete u and athlete v, p is the rating term of the item, and r_u,p denotes the rating of athlete u on item p.

By using the above formula, the set of similar athletes or the nearest neighbor set of athletes of the target athlete to be recommended can be calculated, and by denoting the set by N, it can be predicted that the athlete u has a rating of pred(u, p) for item p on the set of N, and then the formula for pred(u, p) is as follows: 14pred(u,p)=r¯u+∑v∈Nsim(u,v)*(rv,p−r¯v)∑v∈Nsim(u,v)

Where, sim(u, v) denotes the similarity of athlete u and athlete v, r¯u and r¯v denote the average ratings of athlete u and athlete v, and r_v,_p denotes the rating of athlete v for program p.

Finally, based on the predicted values of the athletes for the items, Top-N programs can be selected for the target users based on the descending list.

The CB recommendation algorithm requires only one important feature, i.e., labels, which is needed to decompose the sports into a series of features that are sufficient to indicate the sport and to give a relationship between the sport and the user based on the user's behavior in the system (checking out, sharing, rating, and favoriting) [21].

The cosine similarity formula for CB recommendation algorithm is as follows: 15sim(A,B)=∑i=1nAi*Bi∑i=1n(Ai)2*∑i=1n(Bi)2

Where A_i indicates the degree of preference of the user for the type of sport, and B_i indicates the type of sport to which each sport belongs, and this belonging relationship is non-zero or one.

Finally, according to the similarity of the sports in descending order, select the Top-N to the user can be.

Training intensity is one of the main factors in determining training tolerance, and it usually refers to the degree of exhaustion and strain on the body during training. Training intensity is usually measured by physiological indicators such as enzyme shedding, heart rate, and maximum oxygen consumption. It has been found that training intensity has an impact on the raw materials used by the body and the adaptations carried out by the body. Reasonable training intensity not only improves the heart and respiratory capacity, but also enhances muscle strength, bone density, the body's responsiveness to the outside world, and reduces one's own sense of depression, thus improving the human body's physical fitness.

The exercise heart rate equation is as follows: 16THR=(HRmax−HRrest)*EIdesired

Where THR stands for Exercise Heart Rate, HR_max stands for Maximum Exercise Heart Rate, which can be obtained from the absolute value of the difference between 220 and age, HR_rest stands for Resting Heart Rate, and EI_desired stands for Desired Training Intensity.

The resting heart rate is obtained by measuring the pulse rate for one minute in the morning while awake and before getting out of bed, or it can be measured on all three mornings and the average of the three measurements is obtained. The intensity of training for cravings is now more commonly categorized into three levels, minimum, optimal and maximum intensity, which are calculated as in the formula below: 17THR=(HRmax−HRrest)*α+HRrest

Eq. α When is 0.9, the THR obtained is the highest training intensity. When it is 0.8, the THR obtained is the best training intensity: when it is 0.7, the THR obtained is the lowest training intensity. When the training intensities were calculated, the overall trend of the statistical results was decreasing from year to year according to the unit of years, since the age was increasing from year to year. The three calculated intensity results are displayed in a visual way, as well as compared with the athlete's training heart rate each time, it can be intuitively prompted to the athlete to adjust their training intensity in a timely manner.

Eight preparatory athletes were selected as experimental subjects and a questionnaire was designed to collect data on the effect of training recommendations. The questionnaire was in the form of a Likert scale with options categorized as 1, 2, 3, 4, 5, and 6, where 6 represents the highest level of satisfaction with the training recommendation and 1 represents the lowest level of satisfaction with the training recommendation.

The athletes' data were automatically divided into training set (70%) and test set (30%), the training set was used to predict the rating matrix of the test set, and then the root mean square error (RMSE) between the true and predicted values was calculated for evaluating the prediction quality of the algorithm, and the smaller the RMSE, the better the model. Since the training and test sets are randomly divided, we use the mean value of 200 experiments to represent the prediction results, which is more convincing. Figure 7 displays the experimental results and algorithm comparison.

Figure 7.

Comparison of algorithms

The personalized training recommendation algorithm proposed in this paper can achieve the lowest RMSE value in the four methods with the best results. Under the same similarity calculation model, lower RMSE values can be obtained by using the new recommendation method, which proves the rationality of the personalized training recommendation method proposed in this study.

Table 4 shows the real recommendation results of the test set, while Table 5 shows the recommendations results of the proposed algorithm in this paper. The actual top-2 recommended sports of the 5th athlete are badminton and swimming, while the algorithm recommended badminton and basketball. Other than that, the actual top2 recommendations and predicted top2 recommendations of all other athletes are completely consistent, which also shows the superiority of this algorithm.

It can be seen that this paper's algorithm recommendation results and the actual results are very close to the 8 preparatory athletes' sports top2 recommendation, for which 7 athletes' algorithm recommendation results coincide with their actual situation, and 1 athlete's algorithm recommendation results are different from the actual situation with one sport.

The traditional athlete collaborative filtering recommendation algorithm only considers whether the athlete likes it or not, which often does not match the actual training scenarios. Experimental results show that the algorithm in this paper has a very superior recommendation effect.