Multi-model integrated network hotspot topic propagation mode classification method and system

Abstract

The application discloses a network hotspot topic propagation mode classification method based on comprehensive multi-models, which comprises the following steps: step one, collecting post data of a target hotspot topic in a platform t time period; step two, calculating post basic dimension information based on the post data; step three, drawing a topic heat change curve about time-post quantity; step four, extracting competitive topics of the target hotspot topic and calculating the number of the competitive topics; step five, quantifying the information of a propagation role of the target hotspot topic based on the post data; and step six, performing propagation mode identification and classification of the target hotspot topic through a funnel model and a network model. The application constructs a universal network hotspot topic propagation mode identification and classification method, which can effectively determine the propagation type of a topic and still has a good classification result even in the case that part of dimension data of the topic is missing.

Description

Technical Field

[0001] This invention relates to the field of data mining technology. More specifically, this invention relates to a method and system for classifying the spread patterns of trending topics on the internet using a comprehensive multi-model approach. Background Technology

[0002] With the continuous development of social networks, people are becoming increasingly closer, and topics are spreading faster and faster. Especially now that the internet has become a hub for topic exchange, people can express their thoughts, forward topics, or comment on events anytime, anywhere. However, due to differences in user preferences, topic characteristics, and other factors, different topics circulate at different speeds, have different participants, and exhibit varying development trends online. This presents many users with the challenge of understanding the popularity and dissemination patterns of a topic to determine its relevance and the level of public interest. Therefore, finding a scientific, reasonable, and adaptable method to classify and assess the dissemination patterns of various topics on social media is of significant practical importance.

[0003] In recent years, the ways in which online topics are disseminated have undergone tremendous changes. Compared with the past, they are characterized by rapid response, fast dissemination speed, wide dissemination channels, and diversified platforms. Currently, domestic microblogs and foreign platforms are important platforms for topic dissemination. The hot searches and trending topics on these platforms are topics of concern to most people. Therefore, it is important to determine whether a topic's dissemination model is effective on only one platform and whether it can be extended to multiple data sources in order to meet more complex needs. Summary of the Invention

[0004] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0005] Another objective of this invention is to provide a comprehensive multi-model classification method for the propagation patterns of trending online topics. This method innovatively classifies the propagation patterns of topics into normal propagation, diffusion propagation, and suppression propagation, and constructs a universal method for identifying and classifying the propagation patterns of trending online topics. This method can effectively determine the propagation type of a topic and still achieve good classification results even when some dimensions of the topic data are missing.

[0006] To achieve these objectives and other advantages according to the present invention, a comprehensive multi-model classification method for the propagation patterns of trending online topics is provided, comprising the following steps:

[0007] Step 1: Collect post data on the target trending topic within the platform's time period t;

[0008] Step 2: Calculate basic dimension information of posts based on post data, including total number of posts, number of forwarded posts, number of original posts, number of hyped posts, percentage of forwarded posts, percentage of original posts, and percentage of hyped posts;

[0009] Step 3: Based on the basic dimension information of the posts, count the number of posts within a unit time period t1, and plot the topic popularity change curve with respect to time and number of posts;

[0010] Step 4: Extract competing topics for the target trending topic and calculate the number of competing topics;

[0011] Step 5: Based on post data, quantify the information of the dissemination roles of the target hot topic. Specifically, calculate the number of people corresponding to each type of dissemination role of the target hot topic. The dissemination roles include topic initiators, main disseminators, opinion leaders, and dissemination followers.

[0012] Step Six: Based on the basic dimensional information of the posts, the topic popularity curve, the number of competing topics, and the number of people corresponding to various dissemination roles, the dissemination pattern of the target hot topic is identified and classified using funnel and network models.

[0013] Preferably, the network hot topic propagation pattern classification method of the integrated multi-model is obtained by fusing four basic models: RandomForestClassifier, LightGBM, XGBoost, and LogisticRegression.

[0014] Preferably, in the comprehensive multi-model classification method for the dissemination patterns of trending online topics, step six specifically includes:

[0015] S61. Based on the basic dimension information of the post, the topic popularity change curve, the number of competitive topics, and the number of people corresponding to various dissemination roles, the dissemination pattern of the target hot topic is identified and classified through the funnel model layer by layer. If the funnel model identification and classification obtains the first result of the dissemination pattern of the target hot topic, then proceed to S62; otherwise, if there is no result, proceed to S63.

[0016] S62. The first result is output as the identification and classification result;

[0017] Alternatively, the basic dimensions of the post, the number of competing topics, and the number of people corresponding to each type of dissemination role can be used as inputs into the network model to obtain four secondary results of the dissemination patterns of the target hot topics corresponding to the four basic models. The primary results and the four secondary results are then classified according to various dissemination patterns, and the number of each type of dissemination pattern is counted. The dissemination pattern with the most numbers is selected as the final identification and classification result for output.

[0018] S63. Using the basic dimension information of posts, the number of competitive topics, and the number of people corresponding to various dissemination roles as inputs into the network model, four third results of the dissemination patterns of the target hot topics corresponding to the four basic models are obtained; the four third results are classified according to various dissemination patterns and the number of each dissemination pattern is counted, and the dissemination pattern with the most numbers is selected as the final identification and classification result for output.

[0019] Preferably, the integrated multi-model classification method for the spread patterns of trending online topics further includes:

[0020] Step 7: Describe the characteristics of the propagation pattern of the target trending topic using a funnel model, specifically including:

[0021] Step 71: Obtain the target hot topic popularity feature description through the funnel model: If the total number of posts for the target hot topic b < 0.1a, the target hot topic popularity feature description is extremely low; if 0.1a ≤ b < a, the target hot topic popularity feature description is relatively low; if a ≤ b < 4a, the target hot topic popularity feature description is relatively high; if 4a ≤ b, the target hot topic popularity feature description is extremely high; a is a preset post count threshold.

[0022] Step 72: Obtain the location feature description of the target hot topic posts through the funnel model: Classify the posting regions of the target hot topic posts and count the number of posts corresponding to each posting region. If the sum of the number of posts corresponding to three posting regions is greater than the total number of posts of the target hot topic, the location feature description of the target hot topic posts is that the post location distribution is uneven; otherwise, it is described as that the post location distribution is balanced.

[0023] Step 73: Obtain the feature description of the role in the spread of the target hot topic through the funnel model: The number of people corresponding to each type of role in the spread of the target hot topic is described as a feature;

[0024] Step 74: Output the characteristic description of the dissemination pattern of the target hot topic along with the identification and classification results.

[0025] Preferably, in the comprehensive multi-model network hot topic propagation pattern classification method, step S61, based on the basic dimension information of posts, the topic popularity change curve, the number of competitive topics, and the number of people corresponding to various propagation roles, uses a funnel model to filter layer by layer to identify and classify the propagation pattern of the target hot topic, specifically as follows:

[0026] S610. Based on the basic dimension information of posts, calculate the average number of posts d. Based on the topic popularity change curve, obtain the minimum slope of the curve and the number of posts e and f corresponding to the next point next to it. If (fe) > 3d, and the number of posts of the next n consecutive points after the minimum slope of the curve is not greater than f, then the propagation mode of the target hot topic is identified as suppressed propagation and is taken as the first result; otherwise, there is no result. Where, d = N / (t / t1), and N is the total number of posts.

[0027] S611. Calculate the average number of competitive topics h, and calculate the number of competitive topics i within the preset time threshold. If i > 3h, the propagation mode of the target hot topic is identified as suppressed propagation and is taken as the first result; otherwise, there is no result.

[0028] S612. If either S610 or S611 yields the first result, then the filtering ends; otherwise, proceed to S613.

[0029] S613. If the proportion of forwarded posts is greater than the first proportion threshold and / or if the proportion of hype posts is greater than the second proportion threshold, then the propagation mode of the target hot topic is identified as diffusion and propagation, and this is taken as the first result. If the proportion of original posts is greater than the third proportion threshold, then the propagation mode of the target hot topic is identified as normal propagation and this is taken as the first result. Otherwise, there is no result.

[0030] Preferably, the comprehensive multi-model network hot topic propagation pattern classification method extracts competitive topics of the target hot topic, specifically as follows:

[0031] Collect post data from all other topics within the platform's time period t. Based on tags and a filter term list, filter out candidate competitive topics from the post data of all other topics. Sort the candidate competitive topics in descending order of the number of posts and extract the top M% of candidate competitive topics as candidate competitive topics.

[0032] Based on the post attributes of the topic, obtain the set of post IDs of the target hot topic and the set of post IDs of the candidate competing topics, and calculate the post time series of the target hot topic and the post time series of the candidate competing topics; based on the user attributes of the topic, obtain the set of user IDs of the target hot topic and the set of user IDs of the candidate competing topics, and calculate the user time series of the target hot topic and the user time series of the candidate competing topics.

[0033] Using the dynamic time warping algorithm, the post time series similarity x1 and user time series similarity x2 between the target hot topic and the candidate competing topics are calculated based on the post time series of the target hot topic and the post time series of the candidate competing topics, respectively.

[0034] Calculate the number of intersections between the set of post IDs of the target trending topic and the set of post IDs of the candidate competing topics, and obtain the post ID similarity between the target trending topic and the candidate competing topics x3;

[0035] Calculate the number of intersections between the user ID set of the target hot topic and the user ID set of the candidate competing topics, and obtain the user ID similarity between the target hot topic and the candidate competing topics x4;

[0036] Based on Formula 1, the post time series similarity x1, user time series similarity x2, post ID similarity x3, and user ID similarity x4 are standardized respectively;

[0037]

[0038] Based on Formula 2, calculate the overall similarity St between the target hot topic and the candidate competing topics;

[0039] Formula 2: St = r1*x1 + r2*x2 + r3*x3 + r4*x4;

[0040] In Formula 2, r1, r2, r3, and r4 are weighting coefficients, and r1 + r2 + r3 + r4 = 1;

[0041] Candidate competing topics are sorted from high to low according to their overall similarity St, and the top M% of candidate competing topics are extracted as competing topics for the target hot topic.

[0042] This invention also provides a comprehensive multi-model classification system for the propagation patterns of trending online topics, which is used to implement the above-mentioned classification method. The classification system includes:

[0043] The data collection module is used to collect post data on target trending topics within a time period t on the platform.

[0044] The first calculation module is used to calculate basic dimension information of posts based on post data, including the total number of posts, the number of forwarded posts, the number of original posts, the number of hyped posts, the percentage of forwarded posts, the percentage of original posts, and the percentage of hyped posts.

[0045] The second calculation module is used to count the number of posts within a unit time period t1 based on the basic dimension information of the posts, and to draw a topic popularity change curve with respect to time and number of posts.

[0046] The third calculation module is used to extract competing topics of the target hot topic and calculate the number of competing topics.

[0047] The fourth calculation module is used to quantify the information of the dissemination roles of the target hot topic based on post data. Specifically, it calculates the number of people corresponding to various dissemination roles of the target hot topic. The dissemination roles include topic initiators, main disseminators, opinion leaders, and dissemination followers.

[0048] The identification and classification module is used to identify and classify the propagation patterns of target hot topics based on the basic dimensional information of posts, the topic popularity change curve, the number of competitive topics, and the number of people corresponding to various propagation roles, using funnel and network models.

[0049] The present invention has at least the following beneficial effects:

[0050] 1. Considering that data from different platforms have different attributes, and using both funnel and network models, this application classifies topic propagation patterns based on the basic dimension information of posts, topic popularity change curves, the number of competitive topics, and the attribute information of the number of people corresponding to various propagation roles. Different models and specific attributes are selected according to the attributes of the data, which enables the invention to achieve better classification results when data attributes are sufficient, and to accurately classify even when data attributes are limited.

[0051] 2. This invention uses funnel model, RandomForestClassifier, LightGBM, XGBoost, and LogisticRegression model, which have fast training speed and can quickly determine the topic propagation pattern, resulting in high time efficiency;

[0052] 3. This invention can not only accurately identify and classify the dissemination patterns of trending topics, but also effectively describe the various dimensions of the topic's attributes in the funnel model, helping users understand why the topic belongs to a certain dissemination pattern.

[0053] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the integrated multi-model classification method for the propagation patterns of trending topics on the internet as described in this invention.

[0055] Figure 2 This is a trend curve showing the change in popularity of the target trending topic in an embodiment of the present invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments, so that those skilled in the art can implement it based on the description.

[0057] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0058] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0059] In the description of this invention, the terms "lateral", "longitudinal", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0060] like Figure 1 As shown, this invention provides a method for classifying the spread patterns of trending topics on the internet using a comprehensive multi-model approach, which includes the following steps:

[0061] Step 100: Collect post data on the target trending topic within the platform's time period t;

[0062] Before starting the task, you need to provide the topic name of the target trending topic, the time span of its spread (i.e., the start and end times), and the time period t. This application can be applied to the analysis of topics across multiple channels (one channel corresponds to one task for classifying the spread patterns of a target trending topic), such as Weibo, Facebook, or Twitter. All post data under the target topic will be extracted from the database based on the topic name and time span. To ensure rapid task deployment, the topic data can be stored in the database before starting the task, allowing for faster data retrieval. Elasticsearch distributed storage database can be selected. Each post data (basic attributes) includes the post's repost / comment information, originality attribute, post text content, post publisher, and post publication time. For reposted or commented posts, the original poster's username should also be provided.

[0063] Step 200: Calculate the basic dimension information of posts based on the post data. The basic dimension information of posts includes the total number of posts, the number of forwarded posts, the number of original posts, the number of hyped posts, the percentage of forwarded posts, the percentage of original posts, and the percentage of hyped posts.

[0064] The post data is quantified. Based on the data of each post (basic attributes), the number and proportion of posts of each type of topic within a given time range are obtained (the proportion of each type of post is equal to the ratio of the number of posts of that type to the total number of posts), and indicators such as geographical location and changes in popularity are obtained.

[0065] Hype posts are defined as posts with high repetition rates. The identification process involves calculating the unique set of all text and then counting the number of times each text appears in all posts. Text with a high repetition rate is identified as hype content. The detailed identification steps are as follows: First, the post data undergoes preprocessing (based on a filtering vocabulary to remove invalid information), removing useless text such as web links, emoticons, and stop words. Then, the unique set of all text is calculated, and the repetition rate of each text is counted. Text with a repetition rate exceeding 100 is identified as hype post.

[0066] Step 300: Based on the basic dimension information of the posts, count the number of posts within a unit time period t1, and plot the topic popularity change curve with respect to time and number of posts;

[0067] The time span is subdivided into smaller time windows, i.e., unit time periods t1 (for example, time t is from 9:00 AM to 9:00 PM, and the unit time period t1 is one hour, including 12 time windows: 9-10:00, 10-11:00, 11-12:00, 12-13:00, 13-14:00, 14-15:00, 15-16:00, 16-17:00, 17-18:00, 18-19:00, 19-20:00, and 20-21:00). The number of posts for the topic in each smaller time window is counted to obtain the number of posts for the topic within the time period. The topic popularity curve is plotted with the time period as the X-axis and the number of posts as the Y-axis. The popularity of the continuous smaller time windows reflects the change in the topic's popularity.

[0068] Step 400: Extract competing topics for the target trending topic and calculate the number of competing topics; specifically, extracting competing topics for the target trending topic involves:

[0069] S41. Collect post data from all other topics within the platform's time period t. Based on tags and a filter term list, filter out candidate competitive topics from the post data of all other topics. Sort the candidate competitive topics in descending order of the number of posts, and extract the top M% of candidate competitive topics as candidate competitive topics. The extraction of candidate competitive topics is specifically as follows: extract tags based on the target hot topic, select the top 50% of tags as keywords to extract optional competitive topics from all other topics, and filter out candidate competitive topics from the optional competitive topics based on the filter term list.

[0070] S42. Based on the post attributes (post ID, post count) of the topic, obtain the set of post IDs for the target hot topic and the set of post IDs for the candidate competing topics, and calculate the post time series for the target hot topic and the post time series for the candidate competing topics; based on the user attributes (user ID, user count) of the topic, obtain the set of user IDs for the target hot topic and the set of user IDs for the candidate competing topics, and calculate the user time series for the target hot topic and the user time series for the candidate competing topics.

[0071] S43. Using the dynamic time warping algorithm, based on the post time series of the target hot topic and the post time series of the candidate competing topics, respectively, and using Formula 3, calculate the post time series similarity x1 and user time series similarity x2 between the target hot topic and the candidate competing topics.

[0072]

[0073] Where W = (w1, w2, ..., w k () represents a path from the beginning to the end of the sequence coordinates; w k This represents a sequence Q and a sequence C stored at a certain grid point along the path, where two points (q) are represented. i c j The square of the Euclidean distance w between ) k =(q i -c j ) 2 Thus, the post time series similarity x1 (post-sequence-num) and the user time series similarity x2 (user-sequence-num) are obtained.

[0074] S44. Calculate the intersection of the post ID set of the target trending topic and the post ID set of the candidate competing topics, and obtain the post ID similarity between the target trending topic and the candidate competing topics multiplied by 3.

[0075] (post-intersections-num), as shown in Formula 4;

[0076] post-intersections-num

[0077] =len(set(raw-post-id)∩set(candidate-post-id)) Formula 4

[0078] S45. Calculate the number of intersections between the user ID set of the target hot topic and the user ID set of the candidate competing topics to obtain the user ID similarity between the target hot topic and the candidate competing topics x4(user-intersections-num), as shown in Formula 5;

[0079] user-intersections-num

[0080] =len(set(raw-user-id)∩set(candidate-user-id)) Formula 5

[0081] In Formulas 4 and 5, set(raw-post-id) and set(raw-user-id) are the set of post IDs and user IDs for the target topic, respectively, while set(candidate-post-id) and set(candidate-user-id) are the set of post IDs and user IDs for the candidate competing topics, respectively.

[0082] S46. Based on Formula 1, standardize the post time series similarity x1, user time series similarity x2, post ID similarity x3, and user ID similarity x4 respectively;

[0083]

[0084] X in the formula min X max The minimum and maximum similarity values ​​among multiple candidate competing topics are respectively identified.

[0085] S47. Based on Formula 2, calculate the total similarity St between the target hot topic and the candidate competing topics;

[0086] St = r1*x1 + r2*x2 + r3*x3 + r4*x4 (Formula 2)

[0087] In Formula 2, r1, r2, r3, and r4 are weighting coefficients, and r1 + r2 + r3 + r4 = 1;

[0088] For example, the weighted sum of the standardized temporal similarity and set similarity is used to obtain the overall similarity, as shown in Formula 6:

[0089] total-similarity

[0090] =0.3*post-sequence-similarityy+0.3*user-sequence-similarity+0.2*post-set-similarity+0.2*userset-similarity Formula 6

[0091] In Formula 6, post-sequence-similarity and user-equence-similarity are the standardized post time series similarity and user time series similarity, respectively; post-set-similarity and user-set-similarity are the standardized post set similarity and user set similarity, respectively; here, r1, r2, r3, and r4 are set to 0.3, 0.3, 0.2, and 0.2, respectively.

[0092] S48. Sort the candidate competitive topics in descending order of overall similarity St, and extract the top M% of candidate competitive topics as competitive topics for the target hot topic; the preferred value of M% is 10%.

[0093] Step 500: Based on post data, quantify the information of the dissemination roles of the target hot topic. Specifically, calculate the number of people corresponding to each type of dissemination role of the target hot topic. The dissemination roles include topic initiators, main disseminators, opinion leaders, and dissemination followers.

[0094] The identification of communication roles is mainly divided into four categories: initiator, main communicator, opinion leader, and communicator follower. Their identification mainly depends on the set rules.

[0095] The identification process involves constructing a data table based on task data, containing columns such as all user posts, the earliest posting time of the user's original posts, dissemination influence, degree of opinion matching, number of posts, average post length, and whether there is original content creation. Then, each user is identified according to the following rules:

[0096] Originator: Has an original post, and the earliest posting time of the original post is in the first 1% of the selected time period;

[0097] Key disseminators: The top K% of users are selected based on the number of reposts and comments.

[0098] Opinion Leaders: First, calculate the degree of matching with the opinion's keywords; then, sort by the earliest publication date and select the top K% of users as leaders.

[0099] Followers: Users are sorted by the number of posts [ascending order], average post length [ascending order], and whether they have created original content [ascending order], and the top K% of users are selected as followers; K% can be selected from 1% to 10% based on actual analysis;

[0100] Step 600: Based on the basic dimensional information of the posts, the topic popularity curve, the number of competing topics, and the number of people corresponding to various dissemination roles, the dissemination pattern of the target hot topic is identified and classified using funnel and network models, specifically including:

[0101] S61. Based on the basic dimensional information of posts, the topic popularity curve, the number of competing topics, and the number of people corresponding to various dissemination roles, the dissemination pattern of the target hot topic is identified and classified through a funnel model for layer-by-layer filtering:

[0102] S610. Based on the basic dimension information of posts, calculate the average number of posts d. Based on the topic popularity change curve, obtain the point with the minimum slope of the curve and the number of posts e and f corresponding to the next adjacent point. If

[0103] If (fe)>3d, and the number of posts in the next n consecutive points after the point with the minimum slope of the curve is not greater than f, then the propagation mode of the target hot topic is identified as suppressed propagation and is taken as the first result; otherwise, there is no result; where d=N / (t / t1), and N is the total number of posts;

[0104] S611. Calculate the average number of competitive topics h, and calculate the number of competitive topics i within the preset time threshold (t). If i > 3h, the propagation mode of the target hot topic is identified as suppressed propagation and is taken as the first result; otherwise, there is no result. h is the average number of competitive topics corresponding to multiple topic data of existing propagation modes in the database.

[0105] S612. If either S610 or S611 yields the first result, then the filtering ends; otherwise, proceed to S613.

[0106] S613. If the proportion of forwarded posts is greater than the first proportion threshold and / or if the proportion of hype posts is greater than the second proportion threshold, the propagation mode of the target hot topic is identified as diffusion and propagation and is taken as the first result. If the proportion of original posts is greater than the third proportion threshold, the propagation mode of the target hot topic is identified as normal propagation and is taken as the first result. Otherwise, there is no result.

[0107] If the funnel model identification and classification yields the first result of the propagation pattern of the target hot topic, proceed to S62; otherwise, if there is no result, proceed to S63.

[0108] S62. The first result is output as the identification and classification result;

[0109] Alternatively, the basic dimensions of the post, the number of competing topics, and the number of people corresponding to each type of dissemination role can be used as inputs into the network model to obtain four secondary results of the dissemination patterns of the target hot topics corresponding to the four basic models. The primary results and the four secondary results are then classified according to various dissemination patterns, and the number of each type of dissemination pattern is counted. The dissemination pattern with the most numbers is selected as the final identification and classification result for output.

[0110] S63. Input the basic dimension information of the post, the number of competitive topics, and the number of people corresponding to each type of dissemination role into the network model to obtain four third results of the dissemination mode of the target hot topic corresponding to the four basic models; classify the four third results according to each type of dissemination mode and count the number of each type of dissemination mode; select the dissemination mode with the most numbers as the final identification and classification result for output.

[0111] Step 700: Describe the characteristics of the propagation pattern of the target hot topic using a funnel model, specifically including:

[0112] Step 71: Obtain the target hot topic popularity feature description through the funnel model: If the total number of posts for the target hot topic b < 0.1a, the target hot topic popularity feature description is extremely low; if 0.1a ≤ b < a, the target hot topic popularity feature description is relatively low; if a ≤ b < 4a, the target hot topic popularity feature description is relatively high; if 4a ≤ b, the target hot topic popularity feature description is extremely high; a is a preset post count threshold.

[0113] Step 72: Obtain the location feature description of the target hot topic posts through the funnel model: Classify the posting regions of the target hot topic posts and count the number of posts corresponding to each posting region. If the sum of the number of posts corresponding to three posting regions is greater than the total number of posts of the target hot topic, the location feature description of the target hot topic posts is that the post location distribution is uneven; otherwise, it is described as that the post location distribution is balanced.

[0114] Step 73: Obtain the feature description of the role in the spread of the target hot topic through the funnel model: The number of people corresponding to each type of role in the spread of the target hot topic is described as a feature;

[0115] Step 74: Output the characteristic description of the dissemination pattern of the target hot topic along with the identification and classification results.

[0116] After the four models in the network model are trained and tested, the results of the four models are voting, and the category with the most occurrences is selected as the final topic propagation category. Generally speaking, if the funnel model can correctly determine the topic propagation type, the funnel model's judgment result is used by default; otherwise, the network model's propagation pattern classification is used. The propagation pattern classification results of the target hot topic are stored in the database for easy display to users and for further analysis and use in the future.

[0117] Funnel models and network models are highly dependent on training data, meaning that different training sets will lead to different propagation patterns identified by the method. Since propagation patterns are inherently subjective, if the selected training set has generally high topic popularity and forwarding rates, the method will adaptively raise the benchmark for judging propagation; conversely, if the artificially added training set includes topics with relatively low popularity as propagation targets, the method will correspondingly lower the benchmark for judging propagation.

[0118] On the other hand, the use of network models and funnel models can be adjusted arbitrarily. You can choose to use only the funnel model or only the network model, or you can set weights for the results of the two models and choose the weighted value of the two results (the one with the most propagation types is selected as the result) as the final propagation mode classification result.

[0119] This invention also provides a comprehensive multi-model classification system for the propagation patterns of trending online topics, which is used to implement the above-mentioned classification method. The classification system includes:

[0120] The data collection module is used to collect post data on target trending topics within a time period t on the platform.

[0121] The first calculation module is used to calculate basic dimension information of posts based on post data, including the total number of posts, the number of forwarded posts, the number of original posts, the number of hyped posts, the percentage of forwarded posts, the percentage of original posts, and the percentage of hyped posts.

[0122] The second calculation module is used to count the number of posts within a unit time period t1 based on the basic dimension information of the posts, and to draw a topic popularity change curve with respect to time and number of posts.

[0123] The third calculation module is used to extract competing topics of the target hot topic and calculate the number of competing topics.

[0124] The fourth calculation module is used to quantify the information of the dissemination roles of the target hot topic based on post data. Specifically, it calculates the number of people corresponding to each type of dissemination role of the target hot topic. The dissemination roles include topic initiators, main disseminators, opinion leaders, and dissemination followers.

[0125] The identification and classification module is used to identify and classify the propagation patterns of target hot topics based on the basic dimensional information of posts, the topic popularity change curve, the number of competitive topics, and the number of people corresponding to various propagation roles, using funnel and network models.

[0126] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0127] Example

[0128] The trending topics related to the "Qiao Renliang incident" on Weibo from 00:00 on November 18, 2016 to 00:00 on November 19, 2016.

[0129] Step 1: First, extract all post data of the target trending topic from Weibo within the time period t (November 18, 2016, 00:00 to November 19, 2016, 00:00) using "Qiao Renliang" as the keyword;

[0130] Step 2: Based on the post data, calculate the total number of posts, the number of forwarded posts, the number of original posts, and the number of hype posts for the target hot topic. Further calculate the percentage of original posts (29.23%) (the third threshold is 37.1%), the percentage of forwarded posts (47.09%) (the first threshold is 53.37%), and the percentage of hype posts (14.15%) (the second threshold is 34.82%).

[0131] Step 3: Based on the basic dimension information of the posts, count the number of posts within a unit time period t1 (20 minutes), and plot the topic popularity curve as a function of time and post count, such as... Figure 2 As shown;

[0132] Step 4: According to the method provided by the present invention, competitive topics of the target hot topic are extracted from Weibo, and the number of competitive topics is calculated; 13 dissimilar competitive topics are obtained.

[0133] Step 5: Based on post data, quantify the information of the dissemination roles of the target hot topic. Specifically, calculate the number of people corresponding to each type of dissemination role of the target hot topic. The dissemination roles include topic initiators, main disseminators, opinion leaders, and dissemination followers. The calculation shows that there are 100 topic initiators, 4 main disseminators, 61 opinion leaders, and 500 dissemination followers.

[0134] Step Six: Based on the funnel model, the topic's popularity is relatively low (total number of target trending topics: 3844, preset post threshold: 11661; conforms to the formula 0.1a≤b<a, indicating low popularity), and the geographical distribution is relatively balanced (the sum of posts from the three regions with the most users does not exceed 50% of the total number of target posts), therefore it does not constitute suppression of propagation (analysis). Figure 2 It can be seen that the topic did not remain at a low level after the popularity dropped, so it is judged that the topic is spreading normally (this can be analyzed based on the data in step two); at the same time, the fusion model calculation shows that the topic is spreading normally, so the topic's spread pattern category is normal spread.

[0135] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A multi-model classification method for the dissemination patterns of trending online topics, characterized in that: Includes the following steps: Step 1: Collect post data on the target trending topic within the platform's time period t; Step 2: Calculate basic dimension information of posts based on post data, including total number of posts, number of forwarded posts, number of original posts, number of hyped posts, percentage of forwarded posts, percentage of original posts, and percentage of hyped posts; Step 3: Based on the basic dimension information of the posts, count the number of posts within a unit time period t1, and plot the topic popularity change curve with respect to time and number of posts; Step 4: Extract competing topics for the target trending topic and calculate the number of competing topics; Step 5: Based on post data, quantify the information of the dissemination roles of the target hot topic. Specifically, calculate the number of people corresponding to each type of dissemination role of the target hot topic. The dissemination roles include topic initiators, main disseminators, opinion leaders, and dissemination followers. Step Six: Based on the basic dimensional information of the posts, the topic popularity curve, the number of competitive topics, and the number of people corresponding to various dissemination roles, the dissemination pattern of the target hot topic is identified and classified through funnel model and network model; the network model is obtained by fusing four basic models: RandomForestClassifier, LightGBM, XGBoost, and LogisticRegression.

2. The method for classifying the spread patterns of trending topics on the internet using a comprehensive multi-model approach as described in claim 1, characterized in that, Step Six, specifically includes: S61. Based on the basic dimension information of the post, the topic popularity change curve, the number of competitive topics, and the number of people corresponding to various dissemination roles, the dissemination pattern of the target hot topic is identified and classified through the funnel model layer by layer. If the funnel model identification and classification obtains the first result of the dissemination pattern of the target hot topic, then proceed to S62; otherwise, if there is no result, proceed to S63. S62. The first result is output as the identification and classification result; Alternatively, the basic dimensions of the post, the number of competing topics, and the number of people corresponding to each type of dissemination role can be used as inputs into the network model to obtain four secondary results of the dissemination patterns of the target hot topics corresponding to the four basic models. The primary results and the four secondary results are then classified according to various dissemination patterns, and the number of each type of dissemination pattern is counted. The dissemination pattern with the most numbers is selected as the final identification and classification result for output. S63. Using the basic dimension information of posts, the number of competitive topics, and the number of people corresponding to various dissemination roles as inputs into the network model, four third results of the dissemination patterns of the target hot topics corresponding to the four basic models are obtained; the four third results are classified according to various dissemination patterns and the number of each dissemination pattern is counted, and the dissemination pattern with the most numbers is selected as the final identification and classification result for output.

3. The method for classifying the spread patterns of trending topics on the internet using a comprehensive multi-model approach as described in claim 2, characterized in that... Also includes: Step 7: Describe the characteristics of the propagation pattern of the target trending topic using a funnel model, specifically including: Step 71: Obtain the target hot topic popularity feature description through the funnel model: If the total number of posts for the target hot topic b < 0.1a, the target hot topic popularity feature description is extremely low; if 0.1a ≤ b < a, the target hot topic popularity feature description is relatively low; if a ≤ b < 4a, the target hot topic popularity feature description is relatively high; if 4a ≤ b, the target hot topic popularity feature description is extremely high; a is a preset post count threshold. Step 72: Obtain the location feature description of the target hot topic posts through the funnel model: Classify the posting regions of the target hot topic posts and count the number of posts corresponding to each posting region. If the sum of the number of posts corresponding to three posting regions is greater than the total number of posts of the target hot topic, the location feature description of the target hot topic posts is that the post location distribution is uneven; otherwise, it is described as that the post location distribution is balanced. Step 73: Obtain the feature description of the role in the spread of the target hot topic through the funnel model: The number of people corresponding to each type of role in the spread of the target hot topic is described as a feature; Step 74: Output the characteristic description of the dissemination pattern of the target hot topic along with the identification and classification results.

4. The method for classifying the spread patterns of trending topics on the internet using a comprehensive multi-model approach as described in claim 3, characterized in that... In step S61, based on the basic dimensional information of the posts, the topic popularity curve, the number of competing topics, and the number of people corresponding to various dissemination roles, the dissemination pattern of the target hot topic is identified and classified through a funnel model of layer-by-layer filtering. Specifically: S610. Based on the basic dimension information of posts, calculate the average number of posts d. Based on the topic popularity change curve, obtain the minimum slope of the curve and the number of posts e and f corresponding to the next point next to it. If (fe) > 3d, and the number of posts of the next n consecutive points after the minimum slope of the curve is not greater than f, then the propagation mode of the target hot topic is identified as suppressed propagation and is taken as the first result; otherwise, there is no result. Where, d = N / (t / t1), and N is the total number of posts. S611. Calculate the average number of competitive topics h, and calculate the number of competitive topics i within the preset time threshold. If i > 3h, the propagation mode of the target hot topic is identified as suppressed propagation and is taken as the first result; otherwise, there is no result. S612. If either S610 or S611 yields the first result, then the filtering ends; otherwise, proceed to S613. S613. If the proportion of forwarded posts is greater than the first proportion threshold and / or if the proportion of hype posts is greater than the second proportion threshold, then the propagation mode of the target hot topic is identified as diffusion and propagation, and this is taken as the first result. If the proportion of original posts is greater than the third proportion threshold, then the propagation mode of the target hot topic is identified as normal propagation and this is taken as the first result. Otherwise, there is no result.

5. The method for classifying the spread patterns of trending topics on the internet using a comprehensive multi-model approach as described in claim 4, characterized in that... Extract competitive topics related to the target trending topic, specifically: Collect post data from all other topics within the platform's time period t. Based on tags and a filter term list, filter out candidate competitive topics from the post data of all other topics. Sort the candidate competitive topics in descending order of the number of posts and extract the top M% of candidate competitive topics as candidate competitive topics. Based on the post attributes of the topic, obtain the set of post IDs of the target hot topic and the set of post IDs of the candidate competing topics, and calculate the post time series of the target hot topic and the post time series of the candidate competing topics; Based on the user attributes of the topics, obtain the user ID set of the target hot topic and the user ID set of the candidate competing topics, and calculate the user time series of the target hot topic and the user time series of the candidate competing topics; Using the dynamic time warping algorithm, the post time series similarity x1 and user time series similarity x2 between the target hot topic and the candidate competing topics are calculated based on the post time series of the target hot topic and the post time series of the candidate competing topics, respectively. Calculate the number of intersections between the set of post IDs of the target trending topic and the set of post IDs of the candidate competing topics, and obtain the post ID similarity between the target trending topic and the candidate competing topics x3; Calculate the number of intersections between the user ID set of the target hot topic and the user ID set of the candidate competing topics, and obtain the user ID similarity between the target hot topic and the candidate competing topics x4; Based on Formula 1, the post time series similarity x1, user time series similarity x2, post ID similarity x3, and user ID similarity x4 are standardized respectively; Formula 1 Based on Formula 2, calculate the overall similarity St between the target hot topic and the candidate competing topics; St = r1*x1 + r2*x2 + r3*x3 + r4*x4 (Formula 2) In Formula 2, r1, r2, r3, and r4 are weighting coefficients, and r1 + r2 + r3 + r4 = 1; Candidate competing topics are sorted from high to low according to their overall similarity St, and the top M% of candidate competing topics are extracted as competing topics for the target hot topic.

6. A comprehensive multi-model classification system for the propagation patterns of trending online topics, used to implement the classification method described in any one of claims 1 to 5, characterized in that, Classification systems include: The data collection module is used to collect post data on target trending topics within a time period t on the platform. The first calculation module is used to calculate basic dimension information of posts based on post data, including the total number of posts, the number of forwarded posts, the number of original posts, the number of hyped posts, the percentage of forwarded posts, the percentage of original posts, and the percentage of hyped posts. The second calculation module is used to count the number of posts within a unit time period t1 based on the basic dimension information of the posts, and to draw a topic popularity change curve with respect to time and number of posts. The third calculation module is used to extract competing topics of the target hot topic and calculate the number of competing topics. The fourth calculation module is used to quantify the information of the dissemination roles of the target hot topic based on post data. Specifically, it calculates the number of people corresponding to various dissemination roles of the target hot topic. The dissemination roles include topic initiators, main disseminators, opinion leaders, and dissemination followers. The identification and classification module is used to identify and classify the propagation patterns of target hot topics based on the basic dimensional information of posts, the topic popularity change curve, the number of competitive topics, and the number of people corresponding to various propagation roles, through funnel models and network models. The network model is obtained by fusing four basic models: RandomForestClassifier, LightGBM, XGBoost, and LogisticRegression.

Images