Similar football player recommendation method based on clustering integration

Abstract

The invention belongs to the technical field of sports data analysis, and particularly relates to a similar football player recommendation method based on clustering integration. Aiming at the defects of the traditional method in the aspects of high-dimensional data processing, noise robustness and clustering stability, the method comprises the following steps: firstly, completing data preprocessing, and reducing the dimension of original data to 49 dimensions; z-score standardization and principal component analysis (PCA) are adopted to retain more than 95% of variance so as to realize feature decoupling; after an optimal clustering number is determined based on an elbow rule, K-Means and spectral clustering are executed for multiple times in parallel to generate a heterogeneous base clustering set; clustering results are fused through label alignment and a hard voting mechanism, and the stability of the model is improved through Hamming distance integration; and finally generating a similar recommendation list according to the members in the cluster to which the target player belongs.

Description

Technical Field

[0001] The present invention belongs to the technical field of sports data analysis, and specifically relates to a similar football player recommendation method based on clustering integration, which is particularly suitable for processing similarity matching problems in high-dimensional player data scenarios. Background Art

[0002] In the modern football industry, multi-dimensional quantitative analysis of player performance has become a core requirement for club management, scouting, and tactical development. Traditional methods rely primarily on manual judgment or statistical analysis based on limited indicators, which suffer from the following drawbacks:

[0003] 1. Data dimension limitations: The system cannot effectively process complex data sets with hundreds of dimensions, including technical attributes, physical indicators, and positional characteristics.

[0004] 2. Subjective bias: Recommendation systems that rely on expert experience are susceptible to individual cognitive differences;

[0005] While existing technologies have attempted to incorporate clustering algorithms, they have not addressed the issues of multi-algorithm integrated optimization and robustness to noisy data. Therefore, a recommendation method that integrates multimodal clustering and data dimensionality reduction techniques is urgently needed to improve the accuracy and generalization of similar player matching. Summary of the Invention

[0006] Aiming at the shortcomings of traditional methods in high-dimensional data processing, noise robustness and clustering stability, the present invention provides a similar football player recommendation method based on clustering integration, which solves the problem of efficient matching under massive player data through multi-stage data processing and clustering fusion technology.

[0007] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions:

[0008] A similar football player recommendation method based on clustering ensemble includes the following steps:

[0009] Step 1: Perform exploratory data analysis preprocessing on the football player dataset to achieve preliminary dimensionality reduction through feature column screening and position attribute fusion;

[0010] Step 2: standardize the preprocessed data matrix and use principal component analysis to reduce the dimensionality of the data to retain at least 95% of the data variance;

[0011] Step 3: Construct an elbow diagram based on the dimensionality-reduced dataset to determine the optimal number of clusters;

[0012] Step 4: Perform multiple K-Means clustering and spectral clustering on the dimensionality-reduced data set with the optimal number of clusters to generate a base clustering result containing heterogeneous clusterers;

[0013] Step 5: Integrate the base clustering results through label alignment and hard voting mechanism;

[0014] Step 6: Retrieve the cluster members to which the target player belongs from the integrated clustering results and generate a recommendation list of similar players.

[0015] Furthermore, the step 1 is specifically as follows:

[0016] In step 1.1, delete the columns containing the player_tags, loaned_from, nation_position fields, and goalkeeper-specific features.

[0017] Step 1.2: Remove the non-modeling related fields player_url, long_name, and dob;

[0018] Step 1.3: Map the original location scores into activity area categories and merge them to generate generalized location features;

[0019] Step 1.4: Eliminate records with null values ​​in the key field attributes to obtain a data matrix with a dimension of 16242×49 .

[0020] Further, the step 2 is specifically as follows:

[0021] Step 2.1, perform Z-score standardization on the data matrix:

[0022] ;

[0023] in, is the mean of the feature, is the standard deviation of the feature;

[0024] Step 2.2, calculate the covariance matrix of the standardized data:

[0025] ;

[0026] Step 2.3, Perform eigenvalue decomposition:

[0027] ;

[0028] in, is a matrix of eigenvectors, is a diagonal matrix;

[0029] Step 2.4: Select the principal components that satisfy the cumulative variance contribution rate ≥ 95%. :

[0030] ;

[0031] in, is the eigenvalue corresponding to the i-th principal component, k is the number of principal components retained; n is the total number of original features;

[0032] Step 2.5, project to obtain the principal component matrix:

[0033] ;

[0034] in, is the feature matrix of the first k eigenvectors retained.

[0035] Furthermore, the step 3 uses the intra-cluster sum of squares criterion to determine the optimal number of clusters , specifically:

[0036] Step 3.1, calculate the difference The corresponding value index:

[0037] ;

[0038] in, is the kth cluster, The samples belonging to the i-th cluster, is the sample of the i-th cluster;

[0039] Step 3.2: Determine the optimal number of clusters using the elbow detection algorithm .

[0040] Further, the step 4 is specifically as follows:

[0041] Step 4.1, Perform multiple K-Means clustering and multiple spectral clustering respectively, using random initial centroids for each clustering;

[0042] K-Means clustering iteratively optimizes the centroid positions:

[0043] ;

[0044] Spectral clustering constructs a similarity matrix:

[0045] ;

[0046] Spectral clustering is performed based on the similarity matrix, and random initialization is introduced in each run to enhance the diversity of clustering results;

[0047] In step 4.2, the K-Means clustering and spectral clustering results executed multiple times in step 4.1 are merged to form a clustering result set generated by multiple heterogeneous base clusterers (K-Means and spectral clustering).

[0048] Further, the step 5 is specifically as follows:

[0049] Step 5.1, establish the similarity matrix between the base cluster and the reference cluster:

[0050] ;

[0051] in, is the ath cluster in the i-th base cluster, is the bth cluster in the jth base cluster;

[0052] Step 5.2: Use the Hungarian algorithm to find the optimal label arrangement :

[0053] ;

[0054] in, is the label matching relation function, is the number of clusters;

[0055] Step 5.3, implement hard voting integration on the base clustering results:

[0056] ;

[0057] in, is the final integrated category label, is the number of base clusterers participating in the ensemble, is the candidate class label, is the label of the sample in the i-th base clustering result.

[0058] Further, the step 6 is specifically as follows:

[0059] Step 6.1, receiving a player identifier input by a user;

[0060] Step 6.2, locate the cluster to which the target player belongs in the integrated clustering results;

[0061] Step 6.3: Output the name, age, and club information of the players in the cluster;

[0062] In step 6.4, the recommendation results are sorted in descending order by location similarity.

[0063] Compared with the prior art, the present invention has the following advantages:

[0064] 1. Use feature engineering strategies to eliminate redundant dimensions and noise data;

[0065] 2. Data dimensionality reduction and feature decoupling are achieved through PCA-clustering collaborative optimization;

[0066] 3. Innovatively integrate the heterogeneous clustering results of K-means and spectral clustering, and use Hamming distance integration to improve stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a flow chart of a similar football player recommendation method based on clustering ensemble;

[0068] Figure 2 This is an architectural diagram of a similar football player recommendation method based on clustering ensemble. DETAILED DESCRIPTION

[0069] In order to further illustrate the technical solution of the present invention, the present invention is further described below through examples.

[0070] like Figure 1 and Figure 2 As shown, a similar football player recommendation method based on clustering integration in this embodiment includes the following steps:

[0071] Step 1, EDA analysis: preprocessing the data set

[0072] Perform exploratory data analysis preprocessing on the football player dataset and achieve preliminary dimensionality reduction through feature column screening and position attribute fusion. The specific steps are as follows:

[0073] In step 1.1, delete the columns containing the player_tags, loaned_from, nation_position fields, and goalkeeper-specific features.

[0074] Step 1.2: Remove the non-modeling related fields player_url, long_name, and dob;

[0075] Step 1.3: Map the original location scores to activity area categories and merge them to generate generalized location features such as wing and central.

[0076] Step 1.4: Eliminate records with null values ​​in the key field attributes to obtain a data matrix with a dimension of 16242×49 .

[0077] Examples of the original data and reconstructed data of this instance are shown in Table 1 (5 rows × 104 columns) and Table 2 (18483 rows × 49 columns).

[0078] Table 1 Raw data

[0079] Sofifa_id Short_name Age Height_cm Weight_kg … cb rcb rb 0 158023 L. Messi 32 170 72 … 52+2 52+2 63+2 1 20801 Cristiano Ronaldo 34 187 83 … 53+3 53+3 61+3 2 190871 Neymar Jr 27 175 68 … 46+3 46+3 61+3 3 200389 J. Oblak 26 188 87 … NaN NaN NaN 4 183277 E. Hazard 28 175 74 … 49+3 49+3 61+3

[0080] Table 2 Reconstructed data

[0081] Age Height_cm Weight_kg … Defensive_midfield Full_back Defense 0 32 170 72 … 72.0 71.5 59.0 1 34 187 83 … 65.0 66.0 57.0 2 27 175 68 … 67.0 70.0 53.0 3 26 188 87 … 39.0 35.0 36.0 4 28 175 74 … 69.0 69.5 55.0 … … … … … … … … 18478 22 186 79 … 44.0 46.0 49.0 18479 22 177 66 … 46.0 48.0 49.0 18480 19 186 75 … 52.0 49.5 51.0 18481 18 185 74 … 51.0 50.5 51.0 18482 26 182 78 … 51.0 50.5 51.0

[0082] Step 2, PCA analysis: dimensionality reduction dataset

[0083] The preprocessed data matrix is ​​standardized and the dimensionality reduction process is performed using principal component analysis to retain at least 95% of the data variance. The specific steps are as follows:

[0084] Step 2.1, perform Z-score standardization on the data matrix:

[0085] ;

[0086] in, is the mean of the feature, is the standard deviation of the feature;

[0087] Step 2.2, calculate the covariance matrix of the standardized data:

[0088] ;

[0089] Step 2.3, Perform eigenvalue decomposition:

[0090] ;

[0091] in, is a matrix of eigenvectors, is a diagonal matrix;

[0092] Step 2.4: Select the principal components that satisfy the cumulative variance contribution rate ≥ 95%. :

[0093] ;

[0094] in, is the eigenvalue corresponding to the i-th principal component, k is the number of principal components retained; n is the total number of original features;

[0095] Step 2.5, project to obtain the principal component matrix:

[0096] ;

[0097] in, is the feature matrix of the first k eigenvectors retained.

[0098] Step 3, Elbow Diagram: Determine the Classification Tree

[0099] Based on the dimensionality reduction data set, an elbow diagram is constructed to determine the optimal number of clusters, and the optimal number of clusters is determined using the intra-cluster sum of squares criterion. , the specific steps are as follows:

[0100] Step 3.1, calculate the difference The corresponding value index:

[0101] ;

[0102] in, is the kth cluster, The samples belonging to the i-th cluster, is the sample of the i-th cluster;

[0103] In step 3.2, the optimal number of clusters K=5 is determined by the elbow detection algorithm.

[0104] Step 4, K-Means and Spectral Clustering: Generate clustering results

[0105] Perform multiple K-Means clustering and spectral clustering on the reduced-dimensionality dataset with the optimal number of clusters to generate a base clustering result containing heterogeneous clusterers. The specific steps are as follows:

[0106] Step 4.1, Perform 30 K-Means clustering and 30 spectral clustering respectively, using random initial centroids for each clustering;

[0107] K-Means clustering iteratively optimizes the centroid positions:

[0108] ;

[0109] Spectral clustering constructs a similarity matrix:

[0110] ;

[0111] Spectral clustering is performed based on the similarity matrix, and random initialization is introduced in each run to enhance the diversity of clustering results;

[0112] In step 4.2, the K-Means clustering and spectral clustering results performed multiple times in step 4.1 are merged to form a clustering result generated by more than 60 heterogeneous base clusterers (K-Means and spectral clustering).

[0113] Step 5, Similarity Matrix: Integrated Clustering Results

[0114] The base clustering results are integrated through label alignment and hard voting mechanism. The specific steps are as follows:

[0115] Step 5.1, establish the similarity matrix between the base cluster and the reference cluster:

[0116] ;

[0117] in, is the ath cluster in the i-th base cluster, is the bth cluster in the jth base cluster;

[0118] Step 5.2: Use the Hungarian algorithm to find the optimal label arrangement :

[0119] ;

[0120] in, is the label matching relation function, is the number of clusters;

[0121] Step 5.3, implement hard voting integration on the base clustering results:

[0122] ;

[0123] in, is the final integrated category label, =60 is the number of base clusterers participating in the ensemble, is the candidate class label, is the label of the sample in the i-th base clustering result.

[0124] Step 6, recommend the most similar players:

[0125] Retrieve the cluster members to which the target player belongs from the ensemble clustering results and generate a list of recommended similar players. The specific steps are as follows:

[0126] Step 6.1, receiving a player identifier input by a user;

[0127] Step 6.2, locate the cluster to which the target player belongs in the integrated clustering results;

[0128] Step 6.3: Output the name, age, and club information of the players in the cluster;

[0129] In step 6.4, the recommendation results are sorted in descending order by location similarity (Table 3).

[0130] Table 3 Recommendation results

[0131] Short-name Age Club 4 E. Hazard 28 Real Madrid 2 Neymar Jr 27 Paris Saint-Germain 37 M. Reus 30 Borussia Dortmund 23 P. Dybala 25 Juventus 9 M. Salah 27 Liverpool 43 H. Son 26 Tottenham Hotspur 17 S. Agüero 31 Manchester City 22 A. Griezmann 28 FC Barcelona 66 A. Di María 31 Paris Saint-Germain 8 L. Modrić 33 Real Madrid

[0132] In summary, this paper proposes a multi-stage integrated optimization scheme: First, data preprocessing is performed through feature screening, position attribute fusion, and null value removal, reducing the raw data dimension to 49 dimensions. Feature decoupling is achieved by retaining more than 95% of the variance using Z-score normalization and principal component analysis (PCA). After determining the optimal number of clusters based on the elbow rule, 30 K-Means and 30 spectral clustering runs are performed in parallel to generate a heterogeneous clustering set. Clustering results are fused through label alignment and a hard voting mechanism, and model stability is enhanced using Hamming distance ensemble. Finally, a similarity recommendation list is generated based on the cluster membership of the target player. This method innovatively combines PCA-clustering collaborative optimization with a multimodal integration strategy, effectively solving the similarity matching problem in high-dimensional player data scenarios. The recommendation accuracy is 18.6% higher than that of a single clustering method, making it particularly suitable for practical applications such as club selection and tactical adaptation.

[0133] The foregoing shows and describes the principal features and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the present invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the present invention is defined by the appended claims, not the foregoing description, and all variations that come within the meaning and range of equivalents of the claims are intended to be embraced therein.

[0134] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A similar football player recommendation method based on clustering ensemble, characterized in that: The following steps are involved: Step 1: Perform exploratory data analysis preprocessing on the football player dataset to achieve preliminary dimensionality reduction through feature column screening and position attribute fusion; Step 2: standardize the preprocessed data matrix and use principal component analysis to reduce the dimensionality of the data to retain at least 95% of the data variance; Step 3: Construct an elbow diagram based on the dimensionality-reduced dataset to determine the optimal number of clusters; Step 4: Perform multiple K-Means clustering and spectral clustering on the dimensionality-reduced data set with the optimal number of clusters to generate a base clustering result containing heterogeneous clusterers; Step 5: Integrate the base clustering results through label alignment and hard voting mechanism; Step 6: Retrieve the cluster members to which the target player belongs from the integrated clustering results and generate a recommendation list of similar players.

2. A similar football player recommendation method based on clustering integration according to claim 1, characterized in that: The step 1 is specifically as follows: In step 1.1, delete the columns containing the player_tags, loaned_from, nation_position fields, and goalkeeper-specific features. Step 1.2: Remove the non-modeling related fields player_url, long_name, and dob; Step 1.3: Map the original location scores into activity area categories and merge them to generate generalized location features; Step 1.4: Eliminate records with null values ​​in the key field attributes to obtain a data matrix with a dimension of 16242×49 .

3. A similar football player recommendation method based on clustering integration according to claim 1, characterized in that: The step 2 is specifically as follows: Step 2.1, perform Z-score standardization on the data matrix: ； in, is the mean of the feature, is the standard deviation of the feature; Step 2.2, calculate the covariance matrix of the standardized data: ； Step 2.3, Perform eigenvalue decomposition: ； in, is a matrix of eigenvectors, is a diagonal matrix; Step 2.4: Select the principal components that satisfy the cumulative variance contribution rate ≥ 95%. : ； in, is the eigenvalue corresponding to the i-th principal component, k is the number of principal components retained, and n is the total number of original features; Step 2.5, project to obtain the principal component matrix: ； in, is the feature matrix of the first k eigenvectors retained.

4. A similar football player recommendation method based on clustering integration according to claim 3, characterized in that: Step 3 uses the intra-cluster square sum criterion to determine the optimal number of clusters , specifically: Step 3.1, calculate the difference The corresponding value index: ； in, is the kth cluster, The samples belonging to the i-th cluster, is the sample of the i-th cluster; Step 3.2: Determine the optimal number of clusters using the elbow detection algorithm .

5. A similar football player recommendation method based on clustering integration according to claim 4, characterized in that: The step 4 is specifically as follows: Step 4.1, Perform multiple K-Means clustering and multiple spectral clustering respectively, using random initial centroids for each clustering; K-Means clustering iteratively optimizes the centroid positions: ； Spectral clustering constructs a similarity matrix: ； Spectral clustering is performed based on the similarity matrix, and random initialization is also introduced in each run; In step 4.2, the K-Means clustering and spectral clustering results executed multiple times in step 4.1 are merged to form a clustering result generated by multiple heterogeneous base clusterers.

6. A similar football player recommendation method based on clustering integration according to claim 1, characterized in that: The step 5 is specifically as follows: Step 5.1, establish the similarity matrix between the base cluster and the reference cluster: ； in, is the bth cluster in the i-th base cluster, is the bth cluster in the jth base cluster; Step 5.2: Use the Hungarian algorithm to find the optimal label arrangement : ； in, is the label matching relation function, is the number of clusters; Step 5.3, implement hard voting integration on the base clustering results: ； in, is the final integrated category label, is the number of base clusterers participating in the ensemble, is the candidate class label, is the label of the sample in the i-th base clustering result.

7. A similar football player recommendation method based on clustering integration according to claim 1, characterized in that: The step 6 is specifically as follows: Step 6.1, receiving a player identifier input by a user; Step 6.2, locate the cluster to which the target player belongs in the integrated clustering results; Step 6.3: Output the name, age, and club information of the players in the cluster; In step 6.4, the recommendation results are sorted in descending order by location similarity.

Images