CN-122020314-A - Online community creator feedback prediction method and system based on dynamic reputation graph and text analysis

CN122020314ACN 122020314 ACN122020314 ACN 122020314ACN-122020314-A

Abstract

The invention relates to an online community creator feedback prediction method and system based on dynamic reputation graph and text analysis. The method comprises the steps of S1 obtaining and preprocessing historical interaction data containing comment texts, user IDs, time stamps and praise point stepping records, S2 constructing a praise and praise double-view interaction map, S3 dividing the data according to days, calculating daily positive and negative reputation values through a PageRank algorithm, S4 introducing a time decay function, weighting and summing the daily granularity reputation values in adjacent time windows to obtain dynamic reputation, S5 extracting text semantic features through RoBERTa, splicing and fusing the text semantic features with the dynamic reputation features, and inputting the praise or praise probability output by a fully-connected neural network. The method has the advantages of taking the positive and negative reputation and time dynamics into consideration, making up the defect of pure text prediction, remarkably improving the feedback prediction precision, especially optimizing the prediction effect of the point stepping behavior, and providing support for community atmosphere guidance and network violence early warning.

Inventors

DENG YUANYUAN
GAO XU

Assignees

赞噢(苏州)科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260212

Claims (9)

1. An online community creator feedback prediction method based on dynamic reputation graph and text analysis is characterized by comprising the following steps: s1, acquiring and preprocessing data, namely acquiring historical interaction data of a target online community, wherein the historical interaction data comprises comment texts, context contents, author IDs, reviewer IDs, time stamps and historical praise and click records; S2, constructing a multi-view interaction map, namely respectively constructing an interaction map formed by a praise interaction map and a point stepping interaction map according to praise and point stepping records in the historical interaction data, wherein nodes of the praise interaction map and the point stepping interaction map are users, and directed edges represent interaction directions among the users; s3, calculating a daily granularity PageRank reputation value, namely cutting the historical interaction data according to the day, and respectively running a PageRank algorithm aiming at a praise interaction diagram and a point stepping interaction diagram of each day to obtain a positive reputation value and a negative reputation value of each user on the same day; S4, calculating time attenuation dynamic reputation, namely selecting a day granularity positive reputation value and a day granularity negative reputation value in a time window close to a target date to be predicted, introducing a time attenuation function, and carrying out weighted summation according to a rule that the weight is larger when the date is close to the target date to obtain the dynamic positive reputation and the dynamic negative reputation of a user; S5, multi-mode feature fusion and prediction, namely extracting text semantic features of comment texts and contexts thereof through a RoBERTa text encoder, splicing and fusing text vectors formed by the text semantic features with reputation vectors of reputation features constructed by dynamic positive reputation and dynamic negative reputation, inputting the reputation vectors into a fully-connected neural network, and outputting a probability prediction value of clicking or praying the comments by an author.
2. An online community creator feedback prediction system based on dynamic reputation graph and text analysis, for executing the prediction method of claim 1, wherein the system comprises a hardware carrier and a software functional module on the hardware carrier, the hardware carrier comprises a processor, a communication interface, a memory and a bus, the processor, the communication interface and the memory are connected with each other through the bus, the memory stores a computer program, the processor executes the computer program to drive the software functional module to work, and the software functional module comprises: The data acquisition and preprocessing module is used for acquiring historical interaction data of the target online community and completing preprocessing, wherein the historical interaction data comprises comment texts, context contents, author IDs, reviewer IDs, time stamps, historical praise and click records, and outputting preprocessed standardized interaction data; The multi-view interaction map construction module is used for receiving the standardized interaction data output by the data acquisition and preprocessing module, respectively constructing a praise interaction map and a pedal interaction map according to praise records and pedal records in the standardized interaction data, wherein nodes of the praise interaction map and the pedal interaction map are users, directed edges represent interaction directions among the users, and outputting a double-view interaction map; the daily granularity PageRank reputation calculation module is used for receiving the double-view interaction pattern output by the multi-view interaction pattern construction module, dividing standardized interaction data according to the day, respectively running PageRank algorithm aiming at the praise interaction pattern and the point stepping interaction pattern of each day, calculating to obtain the positive reputation value and the negative reputation value of each user on the same day, and outputting a daily granularity reputation data set; the time attenuation dynamic reputation calculation module is used for receiving the daily granularity reputation data set output by the daily granularity PageRank reputation calculation module, selecting a daily granularity positive reputation value and a daily granularity negative reputation value in a time window close to a target date before the date aiming at the target date to be predicted, introducing a time attenuation function, weighting and summing according to a rule that the weight is larger when the date is close to the target date, obtaining the dynamic positive reputation and the dynamic negative reputation of a user, and outputting dynamic reputation characteristics; And the multi-mode feature fusion and prediction module is used for receiving comment texts and contexts output by the data acquisition and preprocessing module and the dynamic reputation features output by the time attenuation dynamic reputation calculation module, extracting text semantic features through a RoBERTa text encoder, splicing and fusing the text semantic features and the dynamic reputation features, inputting the text semantic features and the dynamic reputation features into a fully-connected neural network, and outputting a probability prediction value of an author praise or trample the comment.
3. The online community creator feedback prediction method based on dynamic reputation graph and text analysis according to claim 1, wherein the praise interaction graph constructed in S2 Interaction map for pedal In the case of the user To the user Praise and praise, then In which there is a directed edge If the user To the user Point stepping In which there is a directed edge Where U is the user set.
4. The online community creator feedback prediction method based on dynamic reputation graph and text analysis according to claim 1, wherein the damping coefficient d of the PageRank algorithm in S3 takes a value of 0.85, and the computation formula of the PageRank algorithm is as follows: where M (u) is the set of users pointing to user u, L (v) is the number of edges that user v links out, and N is the total number of users.
5. The online community creator feedback prediction method based on dynamic reputation graph and text analysis according to claim 1, wherein the time decay function w (i) in S4 is a linear decay function or an exponential decay function, and the linear decay function expression is The exponential decay function expression is Where i is the difference in days from the target date, e is a natural constant (about 2.71828), a, b, c are positive coefficients, and w (i) >0 is satisfied.
6. The online community creator feedback prediction method based on dynamic reputation graph and text analysis of claim 5, wherein the dynamic forward reputation in S4 The calculation formula is as follows: similarly available dynamic negative reputation 。
7. The online community creator feedback prediction method based on dynamic reputation graph and text analysis of claim 1, wherein the preprocessing operation in S1 comprises data deduplication, missing value filling, timestamp format normalization, and user ID unicoding.
8. The online community creator feedback prediction method based on dynamic reputation graph and text analysis according to claim 1 is characterized in that the fully-connected neural network in S5 comprises 1-3 layers of hidden layers, a ReLU function is adopted by a hidden layer activation function, a probability prediction value of a 0-1 interval is obtained by mapping an output layer through a Sigmoid function, the probability prediction value output by the Sigmoid function is set to be 0.5, when the prediction value is more than or equal to 0.5, the feedback behavior of an author on the comment is judged to be praise, and when the prediction value is less than 0.5, the feedback behavior of the author on the comment is judged to be trample.
9. The online community creator feedback prediction system based on dynamic reputation graph and text analysis according to claim 2, wherein the communication interface supports multiple data transmission protocols, including HTTP, webSocket and FTP protocols, and can adapt to the data source interface types of different online communities, and meanwhile, the communication interface has a data encryption transmission function and performs encryption processing on transmitted historical interaction data and prediction result information.

Description

Online community creator feedback prediction method and system based on dynamic reputation graph and text analysis Technical Field The invention relates to the technical field of feedback prediction, in particular to a feedback prediction method and a feedback prediction system for an online community creator based on dynamic reputation graph and text analysis. Background With the explosive growth of online social communities, user-generated content grows exponentially. In community interaction, feedback of comments received by content creators not only reflects personal preferences of the creators, but also is an important mechanism for community atmosphere guidance and content screening. In the prior art, predictions for user feedback or emotional tendency rely primarily on textual content analysis. Conventional practice utilizes natural language processing techniques, such as models of TF-IDF, LSTM or BERT, to extract semantic features of comment text to predict the response of the author. However, existing text-only approaches have the significant disadvantage that the user's feedback behavior depends not only on what the comment was written, but also on the author's own "network status" or "reputation". The interaction pattern of the high reputation user with the low reputation user may be quite different. The existing reputation or influence models often ignore hostile or repulsive information contained in 'click-through'. The reputation of the user is not constant. Recent liveness and community status changes of users cannot be captured using static global impact indicators. Disclosure of Invention The invention aims to provide an online community creator feedback prediction method and system based on dynamic reputation graph and text analysis, so as to solve the problems in the background technology. In order to achieve the purpose, the invention provides the technical scheme that the feedback prediction method for the online community creator based on dynamic reputation graph and text analysis comprises the following steps: s1, acquiring and preprocessing data, namely acquiring historical interaction data of a target online community, wherein the historical interaction data comprises comment texts, context contents, author IDs, reviewer IDs, time stamps and historical praise and click records; S2, constructing a multi-view interaction map, namely respectively constructing an interaction map formed by a praise interaction map and a point stepping interaction map according to praise and point stepping records in the historical interaction data, wherein nodes of the praise interaction map and the point stepping interaction map are users, and directed edges represent interaction directions among the users; s3, calculating a daily granularity PageRank reputation value, namely cutting the historical interaction data according to the day, and respectively running a PageRank algorithm aiming at a praise interaction diagram and a point stepping interaction diagram of each day to obtain a positive reputation value and a negative reputation value of each user on the same day; S4, calculating time attenuation dynamic reputation, namely selecting a day granularity positive reputation value and a day granularity negative reputation value in a time window close to a target date to be predicted, introducing a time attenuation function, and carrying out weighted summation according to a rule that the weight is larger when the date is close to the target date to obtain the dynamic positive reputation and the dynamic negative reputation of a user; S5, multi-mode feature fusion and prediction, namely extracting text semantic features of comment texts and contexts thereof through a RoBERTa text encoder, splicing and fusing text vectors formed by the text semantic features with reputation vectors of reputation features constructed by dynamic positive reputation and dynamic negative reputation, inputting the reputation vectors into a fully-connected neural network, and outputting a probability prediction value of clicking or praying the comments by an author. Further, the praise interaction map constructed in the S2Interaction map for pedalIn the case of the userTo the userPraise and praise, thenIn which there is a directed edgeIf the userTo the userPoint steppingIn which there is a directed edgeWhere U is the user set. Further, the damping coefficient d of the PageRank algorithm in the S3 is 0.85, and the calculation formula of the PageRank algorithm is as follows: where M (u) is the set of users pointing to user u, L (v) is the number of edges that user v links out, and N is the total number of users. Further, the time attenuation function w (i) in S4 is a linear attenuation function or an exponential attenuation function, and the expression of the linear attenuation function is thatThe exponential decay function expression isWhere i is the difference in days from the target date, e is a natural constant (about 2.71828