Behavior feature extraction method of college students’ social network in sports field based on clustering algorithm

2.1 Social network information dissemination mechanism in college students’ sports field

The spread range of social information in the field of college students’ sports was affected by many factors such as the scale of the social network of information publishers, the number of strong contact chains, the frequency of contact, the communication power of contact objects and so on. In the social network of college students’ sports field, a message sent by the initial node V 0 and forwarded by its adjacent node V 1 generates a directed chain from V 0 to V 1 and V 1 represents the first level dissemination node of this message. When the adjacent node V 2 V 1 forwarded the message V 1 again, it generated a directed chain from V 1 to V 2 ; V 2 representing the secondary dissemination node in information dissemination, and so on. Thus, the information from the initial node forms a directed propagation path V 0 → V 1 → V 2 → … . The depth of information transmission of college students’ sports social networks was also the longest forwarding path. The distance between the initial node and the final forwarding node was called the length of the propagation chain. Because the forwarding from nonfans was almost zero in the actual social network platform in the field of college students’ sports, this part can be ignored. It is considered that information was forwarded layer by layer by fans from the publisher. The forwarding tree as shown in the figure can be used to describe the process of information dissemination, and the propagation depth refers to the progression of the forwarding tree. The information forwarding tree of social networks in college students’ sports field is shown in Figure 1.

Figure 1

The information forwarding tree of social networks in college students’ sports field.

In addition, the range of nodes affected by the information dissemination in the social network of college students’ sports field refers to the collection of all nodes forwarding the information from the initial node V 0 . If the set of paths formed by the dissemination is represented by E, the range of information dissemination can be represented by W ( V , E ) the way of graph theory. When a single message was forwarded, the size of the forwarding group represented the breadth of information dissemination. From the forwarding tree in Figure 1, the forwarding scale was the number of all intermediate nodes except the root node and all non-forwarding leaf nodes.

2.2 Getting social network behavior data

Under the social network of college students’ sports field, all users’ real-time data information was obtained and stored. Before obtaining the user information on Microblog, it was necessary to determine a Microblog user as the starting point of the crawling data, by entering the user name or nickname, and searching among the crawled users. If the user name existed, it was the starting point to start crawling data; if the user name did not exist, it was necessary to change the starting user name [6]. Taking this point as the starting point of data acquisition, the web crawler was used to achieve real-time data acquisition. The specific data acquisition process is shown in Figure 2.

Figure 2

Real-time data acquisition flow chart of the social network.

Most popular social networks in the field of college students’ Sports contain videos and images. For Microblogs containing images, you can find “feed”_ list_ media_. Then we can use the regular expression to match the ID of the image and get the real address of the image http://ww1.sinaima.cn/bmiddle/pic_id [7]. At the same time, the crawler system can resolve the video link address accordingly. After obtaining the page address of the image or video, we can download its web page, and use regular expressions to obtain detailed information about the image or video, such as the title and upload time.

2.3 Social network behavior data processing

The obtained social network behavior data were processed from two aspects of data cleaning and de-duplication. Data cleaning mainly completes data format standardization, abnormal data removal, error correction, repeated data removal, and so on. The specific aspect was to remove the noise and irrelevant data in the original data set, fill in the missing data and fields, and identify and delete outliers. In data filtering, data can be cleaned and corrected according to some rules [8]. For example, if the user account field is missing, the record will lose the user identity information; if the URL field is empty, the record will lose the social network site information; if the number of timestamp fields was incomplete, the exception will appear after format conversion. This kind of problem can be solved by appropriate abandonment and repair.

In social network behavior data de-reprocessing, bloom filter, and 2 m bit vector were used to express the data set. The first m bits corresponded to k mutually independent hash functions h k , whose value range is [0, m − 1] and is an integer, and the last m bits corresponded to another k mutually independent hash functions h 2 k , whose value range is [m, 2m − 1] and is an integer. When new data r i were inserted, it was calculated with the function h _i to get the hash value h i ( r i ) , and the corresponding bit M in the bit vector m was set to 1. When querying whether a data formation s i existed in the set, it was calculated with the hash function h _i to get the h i ( s i ) , and then it was checked that the corresponding bit m [ h i ( s i ) ] in the bit vector group M was all 1 and whether s _i belonged to the data set or not [9]. The optimal values of k and m can be obtained from the number of data set elements n and the maximum allowable error rate p, the calculation method is as follows:

When the parameter k reached the optimal value, the social network behavior data was de-reprocessed.

2.4 Using a clustering algorithm to determine the type of social network user behavior

According to the processing results of the social network user behavior data in the field of college students’ sports, the network users, the records in the log data that were triggered by the same internet user were identified and were connected to the access records of the same user to get the user’s browsing sequence during this period [10]. In the weblog files stored by the operators, due to the summary of different users’ online records, there was no identification of independent users, so the first task after data cleaning was to identify independent users. On this basis, a clustering algorithm was used to classify user behavior. First, the cluster center was selected randomly and then the cluster center was transferred to a higher density region iteratively until convergence. The cluster center and node distribution of cluster selection are shown in Figure 3.

Figure 3

Distribution map of the cluster center and other cluster nodes.

By superimposing the offset vector on the current cluster center coordinate vector, the cluster center was updated, as follows:

shift ( t ) represents the offset vector, which represents the average distance from all samples in the current cluster to the current cluster center, the basic form is as follows:

The formula S ( t ) represents the set of sample points in the current cluster, and any x i satisfies the following conditions:

That is, the distance from all sample points to the current cluster center was less than the key parameter, and the parameter bandwidth represents the bandwidth [11]. On this basis, Mean-shift aimed to make the offset vector less than a stop threshold:

In the formula, stopthresh was the set stop threshold to replace the above process until all sample points find the most suitable cluster center. At the same time, the clusters with close cluster centers were merged to form the final clustering results.

2.5 Set social network behavior characteristics

This study observed user characteristic attributes from the aspects of user access commodity category attributes, access frequency, website stay time, user status and location, and historical access commodity category attributes [12]. The characteristic quantity of network communication behavior selection is shown in Table 1.

When users browse different pages, each category will visit multiple or multiple times to the same page. This paper studies the user propensity of College Students under different categories, and the ratio of the number of users visiting such pages to the number of users visiting various pages [13]. The access frequency characteristic parameters can be expressed as:

u is all categories, i is one of them, ∑ i ( PageView ) and ∑ u ( PageView ) is the total number of visits of category i and the number of visits of the user in all categories. Users stayed longer on the pages they were interested in, but shorter on the pages they were not interested in. The longer the users stayed, the higher was the tendency of browsing such pages. Because of the browsing tendency of users in each page category, it was necessary to summarize and calculate the average residence time of users in the same page category [14]. Due to the randomness of the order of users browsing the pages, each category of pages will be interleaved, while the operator’s user online records were stored in the order of Timestamp, so it needed to be calculated separately. The residence time of a user on a social network page can be expressed as:

In the formula, ΔTimeStamp is the time difference between two adjacent records in the cleaned data, ∑ i PageView is the number of class i browsing times of the user and t i is the average residence time of the user on class i goods.

2.6 Extraction of social network behavior characteristics of college students in the sports field

Based on the setting results of social network behavior characteristics, the quantitative extraction of social network behavior was realized. First, the weight of network behavior was calculated. Taking the weight of a user’s entry and single sentence as an example, the formula of word weight is as follows:

In the above formula, W i is the weight of the entry, TF _i is the number of times the entry appears in the current social process, n i is the number of documents containing the entry in the dataset, and N is the total number of all documents in the dataset [15]. In the social process of college students in the field of sports, if an entry appeared more frequently in social behavior and less frequently in other documents, it was more likely to express the core meaning of this study [16]. In Chinese language habits, only some words in a sentence have practical meaning, while others are only used to support the composition of the sentence, which has no contribution to the substance of the sentence. In the same way, we can get the formula of sentence weight:

In the formula, W ( i , j ) is the weight of the entry in a single sentence, and the weight of a single sentence is the accumulation of the weight of the notional words in the sentence. Since long entries often contain more words with real meaning, in order to prevent the accumulation of entry weights, the weights of long sentences are naturally selected as feature quantities, it was necessary to calculate the average value of the weight of single sentences, where n represents the number of meaningful entries in a single sentence. In addition, the parameter λ total represents the score of each weight factor of a single sentence. Thus, the social network behavior of college students in the field of sports can be expressed by documents and their weights, and the quantification of behavior characteristics can be realized [17]. In theory, in order to avoid the differences caused by different servers and time, the data set was the best for the user records of the same server in a small time window. However, this would lead to the sparsity of the user behavior data set and the poor performance of the intelligent detection algorithm [18]. Therefore, it was necessary to process the input behavior feature vector, which can be processed by normalization and interval alignment. Therefore, it was necessary to process the input behavior feature vector, which can be processed by normalization and interval alignment. The normalization selection min-max normalization was a linear transformation of the original data to map the value to the [0,1] interval. The conversion function is as follows:

In the formula, max is the maximum value of feature sample data, min is the minimum value of feature sample data. The horizontal detection was to compare the user’s behavior with other users’ behavior, while the vertical detection was to compare the user’s current behavior with historical behavior.

3.1 Experimental environment and parameter setting

There were four machines in the whole operation platform; each machine was equipped with a dual-core Intel (R) core (TM) i5-2400 CPU, the main frequency was 3.1 GHz, the network bandwidth was 100 Mbps, 4 G memory, and 500 GB hard disk [19]. The operating system of each machine was Ubuntu 12.04, the Hadoop version was 1.0.2, and the JDK version was java-1.6.0–495. The IP configuration and node configuration of the four machines are shown in Table 2.

The configuration process is as follows: Hadoop 1.0.2 tar.gz store in-home/Hadoop and unzip the file. The command was “$tar zvxfhadoop-1.0.2 tar.gz”. Set the environment variable; Hadoop in/home/Hadoop/Hadoop conf directory- env.sh Set the environment variables needed by the Hadoop platform running in Java_ Home was a necessary variable, and Hadoop was not the home variable is optional [20,21]. To configure the conf/Masters and slaves files, you need to configure them on existing machines. Change the configuration file, such as the core site, configure the IP and port number of the machine cluster where the namenode is located – site, and configure the IP and port number of the cluster computer where the job tracker is located.

3.2 The experimental data set

The data sets prepared in the experiment were all from the social network of college students in the field of sports, in which the user’s common friend relationship and user tag information was used to describe the degree of similarity between users. The significance was to reduce the number of users in the network and improve the calculation efficiency of the model. There were a large number of value corpse powder users in the social network of college students’ sports field, ignoring the one-way attention connection can minimize the impact of advertising users and corpse powder users on the accuracy of the model prediction. In social networks, a user tag is an important user attribute information. A user is allowed to enter an open user attribute tag. These user tags are used to express users’ interests and the topics they are concerned about. Through the social network user tag query, the input user can return all the tag information of the user and the corresponding weight of each tag. The specific settings of the user’s original data set in the experiment are shown in Table 3.

3.3 Set test index

In this study, the performance of feature extraction was analyzed from the aspects of feature extraction efficiency and integrity. Generally speaking, the result of feature extraction should not be less than 10% of the social network behavior data. By calculating the ratio of the amount of feature extraction result data to the amount of social network behavior data, it can be concluded that the integrity of feature extraction can be reflected. The closer the ratio was to 10%, the more complete was the feature extraction. The efficiency of feature extraction was mainly the running time cost of the social network behavior feature extraction program, which can be directly obtained by retrieving the background data of the running program.

3.4 The process of the experiment

Since the clustering algorithm was applied in the designed behavior feature extraction method, the algorithm needed to transform into program code and imported into the experimental environment. The goal of K-means was to divide the data points into k clusters, s to find the center of each cluster and minimize the function:

Among them, β i is the center of the i cluster, the above formula is that each data point is as close as the center of the cluster they belong to.

The calculated matrix D made its diagonal the sum of the value of the column (or row) corresponding to the A matrix and the rest of 0.

1. Makes the B = D − A;

2. Seeks the first k eigenvalues and eigenvectors of the B matrix to project the data points into a k-dimensional space. The j value of the i eigenvector represents the projection of the j data point in the i dimension in the k dimension space. That is, if the k feature vectors are combined into a matrix of N*k, then each row represents the coordinates of a data point in the k-dimensional space.

3. Clusters the data in the k dimension space of each data point, based on the k dimension space coordinates, using K-means or other clustering algorithms.

After the operation of the designed feature extraction method, the results of behavior feature extraction of college students’ social networks in the sports field were obtained, and the quantitative extraction results of the number of social topics are shown in Figure 4.

Figure 4

Quantitative extraction results of social network topic number behavior characteristics.

In Figure 4, CDF is a quantitative indicator of the number of social network topics.

In addition, in order to reflect the advantages of the design method, the traditional feature extraction method and the feature extraction method based on worm detection were set as the two contrast methods of the experiment. The contrast method was imported into the experimental environment in the same way, and the final feature extraction result was obtained by calling the prepared data set.

3.5 Comparative analysis of experimental results

After the operation of the three methods, the quantitative comparison results are obtained, as shown in Table 4.

The amount of social network behavior data of college students in the field of sports set in the experiment was 2,000 MB, then through the calculation of the data in Table 4, it can be concluded that the integrity of feature extraction obtained by the three methods was 9.08, 9.37 and 9.93% respectively. In terms of timeliness, the average time cost of the three methods was 0.868, 0.616, and 0.344 s respectively. It was clear that the Social network behavior feature extraction method designed here had better completeness and the feature extraction time was the shortest. This is because this study collected social network behavior data according to the social network information transmission mechanism to ensure the integrity and correlation of data collection, then processes the raw data with Cologne filtering to ensure more clear and complete data. Using a clustering algorithm to extract the social network behavior characteristics of college students in the sports industry, normalized and transformed the extracted behavioral feature vectors, which improved the accuracy of traditional feature extraction.