CN106295702B - A social platform user classification method based on individual emotional behavior analysis - Google Patents

Abstract

The invention discloses a social platform user classification method based on emotional behavior analysis, comprising the following steps: S1, constructing a forwarding tree; S2, constructing user history records, extracting nodes with the same user ID to construct the user's individual forwarding history records; S3. Build user emotional behavior description features; S4. Use the features in S3 to build a user role classification model given a decision tree, and complete social platform user classification based on emotional behavior analysis. The invention constructs a relatively comprehensive user emotional behavior description model, which can more comprehensively consider the user's personal history information; the method makes full use of the user's personal information, dissemination structure information, emotional information and dynamic time domain information in microblogs. Due to the adoption of the above measures, the present invention can obtain better classification accuracy.

Description

A social platform user classification method based on individual emotional behavior analysis

technical field

The invention belongs to the technical field of social networks, in particular to a method for classifying social platform users based on individual emotional behavior analysis.

Background technique

With the development of Internet technology, online social networks represented by Weibo have been used on a large scale. Users can publish information on it themselves, and can also interact with other information by forwarding, commenting, liking, etc. Like real social networks, user behavior on online social networks conveys not only literal information, but also The user's emotional attitude is different due to the user's personal background and habits, and runs through all the user's interactive behaviors. We call this emotional characteristic of the user the user's emotional role.

The current research on online social network users mainly includes the following aspects, 1. Mining of user influence. This type of research focuses on the analysis of users' personal attributes and information dissemination characteristics to establish a model or algorithm to describe users' social influence. , realize the calculation of user influence, and discover social leaders; 2. Prediction of user online behavior. This type of research models users by considering factors such as user history, context, and social relations, and realizes user-specific behaviors or preferences. predictions, such as whether to participate in forwarding, whether to be interested, etc. 3. User sentiment analysis. This kind of research takes what kind of emotion the user will have as a starting point at a certain moment, and realizes the user's sentiment through a variety of data sources (including text, pictures, videos, music, etc.), online and offline integration, and social relations. Sentiment Analysis and Prediction. The above studies have revealed the rules of users' online behavior and the internal operation rules of social networks to a certain extent, but lack of comprehensive consideration of users' emotions.

Contents of the invention

In view of the above problems, the present invention provides a social platform user classification method based on individual emotional behavior analysis by analyzing from the perspective of user's personal emotion, and the specific technical solution is:

A social platform user classification method based on individual emotional behavior analysis, comprising the following steps:

S1. Construct forwarding tree: extract forwarding information of social platform users, and establish social platform forwarding tree based on tree topology;

S2. Constructing user history records: perform sentiment calculation on forwarding information of nodes in the forwarding tree, and classify the results into positive, negative, and neutral according to sentiment; extract nodes with the same user ID to construct the individual forwarding history records of this user;

S3. Construct user emotional behavior description features: including user tendency description features: individual and group emotional relationship ER _u , user's personal historical emotional preference HP _u ; user emotional impact description feature EI _u ;

S4. Use the features in S3 to build a user role classification model given to the decision tree. First, construct the input vector U _u =<ER _u ,HP _u ,EI _u >, and then calculate the information entropy for each feature U ^j is the jth feature, select the feature with the largest information gain to construct the current decision node, and recurse the remaining features layer by layer to obtain the final decision tree model, and then complete the user classification based on emotion.

Further, forwarding information in a social platform user classification method S1 based on individual emotional behavior analysis includes original text information, forwarded text information, and individual information of participating users.

Further, a social platform user classification method S1 based on the analysis of individual emotional behaviors analyzes text sentiment from bottom to top according to the level, adds forwarding nodes layer by layer, and constructs a forwarding tree.

Further, a social platform user classification method S2 based on individual emotional behavior analysis adopts a multi-rule set model for emotional calculation, and establishes emotional dictionaries and grammatical rules from the bottom up through the mutual information of text points. The bottom-up refers to Analyze in order from words, phrases, short sentences, and whole sentences.

Further, the individual and group emotional relationship described in S3, a social platform user classification method based on individual emotional behavior analysis, is based on individual emotional selection and group emotional distribution, and is described as an emotional relationship factor between an individual and a current piece of text information ER _u (w), its value ranges from -1 to 1, the larger the value, the more positive the current relationship is, and the smaller the value, the more negative the current relationship is, as follows:

Among them, N(w), P(w), and O(w) respectively represent the negative sentiment distribution, positive sentiment distribution, and neutral sentiment distribution in the current forwarding tree, and S(w) represents the scale of the forwarding tree.

Further, a social platform user classification method based on the analysis of individual emotional behavior S3, the individual historical emotional preference HP _u (e) is based on the emotional distribution in the user's historical records and the user comment participation C _u (w) in the historical forwarding , expressed by the following formula:

Among them, exp{-θ ₁ (t ₀ -t _w )} is the time decay to control user preference, and log(C _u (w)+2) is the description of user participation through comment length.

Further, the emotional impact EI _u described in a social platform user classification method based on individual emotional behavior analysis S3 is based on the structural characteristics of the forwarding tree SF _u (w), the time domain impact of the forwarding tree TF _u (w), The user's emotional change EI _u (w) is expressed as follows:

HR _u represents the number of internal nodes forwarded by users, and NR _u represents the number of leaf nodes forwarded by users.

Further, the structural characteristics of the forwarding tree SF _u (w) in a social platform user classification method based on individual emotional behavior analysis is based on the absolute scale S(w) of the forwarding tree, the relative scale S _u (w) and the subtree depth DP _u (w), as follows:

Further, in a social platform user classification method based on individual emotional behavior analysis, the time-domain influence TF _u (w) of the forwarding tree is the contribution of the forwarding tree to information dissemination in the perspective of time, and the contribution is reflected in the subtree relative to the entire The survival time of the forwarding tree and the time delay of the subtree relative to the original text;

Among them, LP _u (w) is the subtree life cycle, LP (w) is the forwarding tree life cycle, is the survival time of the subtree relative to the entire forwarding tree, and exp{-ε(t _u -t _w )} is the time domain delay of the appearance of the subtree;

Furthermore, in a social platform user classification method based on individual emotional behavior analysis, the user's emotional change EF _u (w) takes the current user's forwarding behavior as the time boundary point, calculates the difference in emotional distribution before and after the user's forwarding, and uses the index The function normalizes the parameters, represented by the following formula:

Among them, B _u (w, e) and A _u (w, e) are the emotion distribution of the user before and after retweeting respectively.

The present invention has the following beneficial effects:

In order to systematically describe the online emotional behavior of users, the present invention defines six types of microblog user emotional roles, namely active leader, active follower, passive leader, passive follower, neutral leader, and neutral follower, and proposes A social platform user classification method based on individual emotional behavior analysis, which establishes a user emotional behavior description model from two dimensions (emotional tendency and emotional impact).

Due to the adoption of the user's emotional tendency features and user influence features in the technical solution, a relatively comprehensive user emotional behavior description model can be constructed, which can more comprehensively consider the user's personal historical information; this method makes full use of the user's personal information in Weibo , to propagate structural information, emotional information, and dynamic temporal information. Due to the adoption of the above measures, the present invention can obtain better classification accuracy.

Description of drawings

Fig. 1 is a flow chart of a method for classifying social platform users based on individual emotional behavior analysis of the present invention;

Fig. 2 a kind of social platform user classification method user historical record instance based on individual emotional behavior analysis of the present invention;

Fig. 3 is a kind of social platform user classification method structural characteristic distribution based on individual emotional behavior analysis of the present invention;

Fig. 4 is a kind of social platform user classification method time-domain characteristic distribution based on individual emotional behavior analysis of the present invention;

Fig. 5 is a parameter learning result of a social platform user classification method based on individual emotional behavior analysis of the present invention;

Fig. 6 presents a kind of social platform user classification method based on individual emotional behavior analysis emotion change characteristic distribution;

Fig. 7 is a social platform user classification method based on individual emotional behavior analysis of the present invention and the distribution of personal and macro-emotional relationships;

Fig. 8 is a social platform user classification method based on individual emotional behavior analysis of the present invention, and the historical emotional preference result distribution;

Fig. 9 is the emotional impact result of a social platform user classification method based on individual emotional behavior analysis of the present invention.

Detailed ways

In order to make the objects and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the examples. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

Example

S1. Build forwarding tree: extract forwarding information of social platform users, and build social platform forwarding tree based on tree topology.

Taking Weibo as an example, grab the forwarded data on Weibo, keep the user information in the data, the forwarded information and the original Weibo information, according to the identifier "//@" forwarded by Weibo and the nickname of the upper-level user, according to the level by Bottom-up text analysis, adding reposting nodes layer by layer, and constructing Weibo reposting tree. A total of 19,389 user information was collected, and 7,096 forwarding trees were constructed.

S2. Construct user history records: perform sentiment calculation on the forwarding information of nodes in the forwarding tree, and classify the results into positive, negative, and neutral according to sentiment; extract nodes with the same user ID to construct the user's individual forwarding history.

Using the multi-rule set model, the emotional calculation is carried out on the text information contained in each node in the forwarding tree, and three results are obtained, which are positive, negative and neutral. Afterwards, using the user information contained in each Weibo forwarding node, the nodes with the same user ID are extracted to construct the user's personal historical forwarding records and stored in the form of XML files. Figure 2 is an example of a user's history record, <uid_1796678344> represents a user, <retweet> is a retweet of the current user, <org_id>, <org_text>, <org_time>, <org_emotion>, <p_name>, <p_id> , <w_id>, <w_test>, <w_time>, <w_emotion> indicate the relevant attributes of the corresponding forwarding.

S3. Construct user emotional behavior description features: including user tendency description features: individual and group emotional relationship ER _u , user's personal historical emotional preference HP _u ; user emotional impact description feature EI _u .

The user's emotional tendency is constructed from two perspectives, the personal and macro emotional relationship and the user's personal historical emotional preference. For the former, ER _u (w) represents the emotional relationship factor between the user and the current Weibo. The value range is between -1 and 1. , the larger the value, the more positive the current relationship is, and vice versa. In order to make the neutral emotion near 0, the origins of positive and negative are set to 0.5 and -0.5 respectively.

N(w), P(w), and O(w) respectively represent the three types of emotion distributions (negative, positive, neutral) in the current forwarding tree, and S(w) represents the scale of the forwarding tree.

The user's personal historical emotional preference HP _u (e) is based on the emotional distribution in the user's historical records and the user comment participation C _u (w) in the historical retweeting. The index part is used to control the time decay of the user's preference, published in the most recent Weibo Time t ₀ is used as a reference point, and the logarithmic part describes the user's participation degree through the comment length:

Describe the user's emotional impact from the perspective of forwarding structural characteristics, time-domain characteristics, and emotional changes. The structural characteristics of Weibo forwarding SF _u (w) weigh the absolute size S(w) of the forwarding tree, the relative size S _u (w) and the subtree Depth DP _u (w):

Figure 3 describes the distribution of calculation results of SF _u (w). We believe that with the same forwarding scale, the deeper the subtree means the sparser the subtree, and vice versa, the denser the subtree, and the denser subtree often has wider impact.

Different from the structural characteristics, the time-domain influence TF _u (w) is used to describe the contribution of the forwarding tree to information dissemination in the perspective of time. This contribution is mainly reflected in two aspects. First, the survival time of the subtree relative to the entire forwarding; Second, the time delay of the subtree relative to the original microblog. TF _u (w) comprehensively considers the subtree life cycle LP _u (w), the forwarding tree life cycle LP(w) and the time domain delay exp{-ε(t _u -t _w )} of the subtree. The ε map is used to control the rate of decay:

In this method, the test accuracy is set to 0.2, and Fig. 4 depicts the distribution of calculation results of TFu(w). α ₁ and β ₁ are the learning parameters. Through the separate classification and verification of the features, the value with the highest accuracy is selected as the actual value of the parameter with a step size of 0.1. The test results of the classification method using the decision tree are shown in Figure 5. , so the parameter values are set to 0.6 and 0.7, respectively.

Emotional change EF _u (w) takes the current user’s retweeting behavior as the time cut-off point, and the emotional distribution before and after the user’s retweeting is represented by B _u (w,e) and A _u (w,e) respectively, through |B _u (w, e)-Au( _w ,e)| Computes the sentiment distribution difference and normalizes the parameters by an exponential function:

Figure 6 depicts the distribution of calculated results of EF _u (w).

S4. Use the features in S3 to build a user role classification model given to the decision tree. First, construct the input vector U _u =<ER _u ,HP _u ,EI _u >, and then calculate the information entropy for each feature U ^j is the jth feature, select the feature with the largest information gain to construct the current decision node, and recurse the remaining features layer by layer to obtain the final decision tree model, and then complete the user classification based on emotion.

Perform feature fusion according to the results obtained in S3, and obtain the features that comprehensively describe the user's emotional tendency ER _u , HP _u and emotional impact EI _u as model input:

Among them, EI _u fuses the three types of influence features, and considers that the leaf nodes do not have any influence, introduces As a denoising factor, HR _u represents the number of internal nodes forwarded by users, and NR _u represents the number of leaf nodes forwarded by users. Figure 7 shows the distribution of calculation results of the current data set ER _u , and Figure 8 shows the calculation of HP _u Result distribution, Fig. 9 shows the distribution of calculation results of EI _u . Finally, through the classification method based on the decision tree, six emotional role classifications are obtained, and the classification results are shown in Table 1.

Table 1 embodiment emotion role classification result

emotional role Accuracy Positive Leader (PL) 0.87 Positive Follower (PF) 0.90 Neutral Leader (OL) 0.83 Neutral Follower (OF) 0.86 Negative Leader (NL) 0.91 Negative Follower (NF) 0.92

Claims (7)

1. A social platform user classification method based on individual emotional behavior analysis, comprising the following steps: S1. Build a forwarding tree: extract the forwarding information of social platform users, and establish a social platform forwarding tree based on tree topology: S2. Construct user history records: perform sentiment calculation on forwarding information of nodes in the forwarding tree, and classify the results into positive, negative, and neutral according to sentiment: extract nodes with the same user ID to construct the individual forwarding history records of this user: S3. Construct user emotional behavior description features: including user tendency description features: individual and group emotional relationship ER _u , user personal historical emotional preference HP _u : user emotional impact description feature EI _u ; The emotional relationship between the individual and the group is based on the emotional choice of the individual and the distribution of group emotion, and is described as the emotional relationship factor ER _u (w) between the individual and the current piece of text information, and its value ranges from -1 to 1. The larger the value, the more positive the current relationship is, and the smaller the value, the more negative the current relationship is, as follows: Among them, N(w), P(w), and O(w) respectively represent the negative sentiment distribution, positive sentiment distribution, and neutral sentiment distribution in the current forwarding tree, and S(w) represents the scale of the forwarding tree; The individual historical emotional preference HP _u (e) is based on the emotional distribution in user history records and the user comment participation degree C _u (w) in historical forwarding, expressed by the following formula: Among them, exp[-θ ₁ (t ₀ -t _w )} is to control the time decay of user preference, and log(c _u (w)+2) is to describe the user's participation degree through the comment length; The emotional impact EIu is based on the structural characteristics SF _u (w) of the forwarding tree, the time domain impact TF _u (w) of the forwarding tree, and the user's emotional change EF _u (w), as follows: Among them, HR _u represents the number of internal N nodes forwarded by users, and NR _u represents the number of leaf nodes forwarded by users; S4. Use the features in S3 to build a user role classification model given to the decision tree. First, construct the input vector U _u =< ER _u , HP _u , EI _u >, and then calculate the information entropy for each feature U ^j is the jth feature, select the feature with the largest information gain to construct the current decision node, and recurse the remaining features layer by layer to obtain the final decision tree model, and then complete the user classification based on emotion. 2. A social platform user classification method based on individual emotional behavior analysis according to claim 1, characterized in that: said forwarding information in S1 includes original text information, forwarding text information, and individual information of participating users. 3. a kind of social platform user classification method based on individual emotional behavior analysis according to claim 1, it is characterized in that: described S1 carries out text emotion analysis according to hierarchy from bottom to top, adds forwarding node layer by layer, builds forwarding tree . 4. a kind of social platform user classification method based on individual emotion behavior analysis according to claim 1, it is characterized in that: the emotion computing among the described S ₂ adopts multi-rule set model, bottom-up by text point mutual information Establishing emotional dictionaries and grammar rules, the bottom-up refers to analyzing words, phrases, short sentences, and whole sentences in sequence. 5. a kind of social platform user classification method based on individual emotional behavior analysis according to claim 1, is characterized in that: the structural characteristic SF _u (w) of described forwarding tree is based on the absolute scale S(w) of forwarding tree , relative scale S _u (w) and subtree depth DP _u (w), expressed as follows: 6. a kind of social platform user classification method based on individual emotion behavior analysis according to claim 1, it is characterized in that: the time domain influence TF _u (w) of described forwarding tree is forwarding tree to information propagating in angle of time Contribution, which is reflected in two aspects: the survival time of the subtree relative to the entire forwarding tree, and the time delay of the subtree relative to the original text; Among them, LP _u (w) is the subtree life cycle, LP (w) is the forwarding tree life cycle, is the survival time of the subtree relative to the entire forwarding tree, and exp{-e(t _u -t _w )} is the time-domain delay of the appearance of the subtree. 7. a kind of social platform user classification method based on individual emotional behavior analysis according to claim 1, is characterized in that: described user's emotional change EF _u (w) takes current user's forwarding behavior as time demarcation point, By calculating the emotional distribution difference before and after the user reposts, and standardizing the parameters through the exponential function, it is expressed by the following formula: Among them, B _u (w, e) and A _u (w, e) are the emotion distribution before and after the user forwards, respectively.