A Live Streaming Content Recognition System and Method Based on Virtual Localization and Multimodal Reasoning

Abstract

This invention discloses a live streaming content recognition system and method based on virtual positioning and multimodal reasoning, relating to the field of internet content analysis technology. By analyzing the difference between virtual and actual positioning, and combining it with bullet screen and tipping behavior patterns, this invention forms a cross-modal association evidence chain. It comprehensively calculates a live streaming anomaly index by considering the proportion of fraudulent activity, the proportion of abnormal bullet screens, and the proportion of tipping spam. Each indicator corresponds to an independent weight and allowable threshold, achieving a quantitative classification of the degree of anomaly. This more accurately reflects the risk level of the live streaming room and solves the problem in existing technologies that rely solely on single-modal data for analysis, making it impossible to comprehensively and accurately identify abnormal behavior during the live streaming process and prone to misjudgment.

Description

Technical Field

[0001] This invention relates to the field of internet content analysis technology, and in particular to a live streaming content recognition system and method based on virtual positioning and multimodal reasoning. Background Technology

[0002] With the booming development of internet technology, the live streaming industry has experienced explosive growth and has become a key platform for content dissemination, commercial marketing, and social interaction. However, with the prosperity of the live streaming industry, various problems have also emerged, seriously affecting the healthy development of the industry. In order to pursue traffic and commercial interests, some streamers and merchants use virtual location technology to fake the geographical distribution of viewers. For example, they may virtually locate viewers who are actually concentrated in the same data center IP segment as being from different cities, thereby creating the illusion of popularity to attract target customers.

[0003] However, existing live streaming content recognition systems still have the following shortcomings when used to analyze fake data in live streaming rooms:

[0004] On the one hand, existing technologies that rely solely on single-modal data for analysis, such as IP location or bullet screen text, cannot comprehensively and accurately identify abnormal behavior during live streaming, and are prone to misjudgment or omission.

[0005] On the other hand, risk assessments using existing technologies are often based on data from a single live stream, lacking comprehensive analysis of historical data and trend tracking, resulting in a rather one-sided risk assessment of live streams.

[0006] To address this, a live content recognition system and method based on virtual positioning and multimodal reasoning have been developed. Summary of the Invention

[0007] In view of this, the present invention provides a live content recognition system and method based on virtual positioning and multimodal reasoning to solve the problems mentioned in the background art.

[0008] The objective of this invention can be achieved through the following technical solution: a live content recognition system based on virtual positioning and multimodal reasoning, comprising:

[0009] Multimodal data transmission module: collects location data and user behavior data during the current live broadcast; the location data includes the virtual location set by the user and the actual location of the device; the user behavior data includes the bullet screen text in the live broadcast room, the sending timestamp, and the reward record; the reward record includes the type and quantity of gifts;

[0010] Multimodal reasoning module: After extracting location data and user behavior data and processing them comprehensively, the module determines the proportion of hot topic fraud, abnormal bullet comments, and tipping spam. It then performs cross-modal correlation analysis on the proportion of hot topic fraud, abnormal bullet comments, and tipping spam to determine the live broadcast abnormality index (PHE) during the current live broadcast.

[0011] Live Streaming Risk Module: Receives the Live Streaming Anomaly Index (PHE) during the current live stream and selectively triggers risk warning signals. After triggering the risk warning signal, it combines the historical data stored in the database to perform corresponding steps to determine the current live stream's trend anomaly level. The trend anomaly level includes a general trend level and a risk trend level.

[0012] In some embodiments, the specific steps for determining the proportion of trending fraud are as follows:

[0013] The system identifies the number of online users at each point in time during the current live stream. Based on the location data during the live stream, it extracts the virtual location set by each online user at each point in time and the actual device location. If the virtual location and the actual device location of an online user are different, the user is marked as an abnormal user and a triplet is generated. The triplet is represented as (virtual location, actual device location, abnormal user). For each triplet generated at each point in time, the actual device location is extracted and matched. The number of triplets that successfully match the actual device location is counted and recorded as the number of fake users.

[0014] Obtain the account registration information of each fake user, extract the account registration time from the account registration information, and calculate the time difference between the time difference and the current live broadcast time to obtain the account registration duration of each fake user;

[0015] The number of fake users whose account registration time is less than the set registration time is counted and recorded as the number of fake verifications; the number of fake verifications marked at each time point is extracted, multiplied by the preset number correction coefficient, and then rounded to obtain the corrected number of fake verifications at each time point in the current live broadcast room;

[0016] The ratio of the number of online users and the number of fake corrections at each time point during the current live stream is calculated. That is, the number of fake corrections is used as the numerator and the number of online users is used as the denominator to obtain the proportion of fake users at each time point during the current live stream. The average of the proportion of fake users at each time point is taken as the proportion of popularity fraud during the current live stream.

[0017] In some embodiments, the specific steps for determining the proportion of abnormal bullet comments are as follows:

[0018] Extract the text and timestamp of the bullet comments from the user behavior data during the current live broadcast; count the total number of bullet comments during the current live broadcast; count the number of bullet comments with the same content from each bullet comment; and mark the bullet comment with the highest number of bullet comments as an abnormal bullet comment password.

[0019] For abnormal bullet screen passwords, identify users who have sent abnormal bullet screen passwords in the live broadcast room and record them as spam users. Extract the number of times each spam user sent abnormal bullet screen passwords and the timestamp of each abnormal bullet screen password sent.

[0020] If a user who spams the screen sends more abnormal barrage commands than the set threshold number of times within a set time window, they are identified as a user with abnormal frequency. The actual device location of each user with abnormal frequency is extracted. For users with abnormal frequency who belong to the same location, the total number of abnormal barrage commands sent is counted, and the percentage of the total number of barrages is calculated and recorded as the percentage of abnormal barrages.

[0021] In some embodiments, the specific steps for determining the percentage of screen-scrolling rewards are as follows:

[0022] Extract tipping records from user behavior data during the current live stream; count the total number and types of gifts during the current live stream; count the number of gifts belonging to the same gift type to obtain the number of gifts received by each gift type in the current live stream; mark the gift type with the highest number of gifts as the gift pattern type; and mark the number of gift pattern types as the gift pattern quantity.

[0023] Identify users who have given gifts in the live stream. Users whose gifts belong to a regular gift type are recorded as reward pattern users. Extract the number of gifts given by each reward pattern user. Extract the actual device location of each reward pattern user. For reward pattern users belonging to the same location, calculate the total number of gifts given and the percentage of the total number of gifts, which is recorded as the reward flooding percentage.

[0024] In some embodiments, determining the Public Anomaly Index (PHE) during the current live stream process specifically involves:

[0025] The percentages of fraudulent popularity, abnormal bullet comments, and excessive tipping observed during the current live stream are respectively tagged as follows: ;

[0026] Using formula The Live Streaming Abnormality Index (PHE) is calculated by comprehensively analyzing the percentage of fraudulent activity, abnormal comments, and excessive tipping during the current live stream. These are the preset percentages for allowed popularity fraud, allowed percentages for allowed abnormal bullet comments, and allowed percentages for allowed tipping spam. The weighting coefficients for the percentage of trending fraud, the percentage of abnormal bullet comments, and the percentage of tipping spam.

[0027] In some embodiments, receiving the Live Abnormality Index (PHE) during the current live broadcast and selectively triggering a risk warning signal specifically involves:

[0028] The live broadcast anomaly index (PHE) during the current live broadcast is compared with the set live broadcast anomaly threshold index. If the live broadcast anomaly index (PHE) during the current live broadcast is higher than the live broadcast anomaly threshold index, a risk warning signal is triggered.

[0029] In some embodiments, the step of performing corresponding steps after triggering the risk warning signaling, combined with historical data stored in the database, specifically includes:

[0030] After triggering the risk warning signal, extract the live streaming data from the database that is c times closer to the current live streaming room. The live streaming data includes the live streaming time and the live streaming anomaly index (PHE); where c > 3.

[0031] The number of times the current live stream room triggered risk warning signals during the last c live streams is recorded as the number of risk events. The percentage of the number of risk events in c is calculated as the current live stream room's continuous risk ratio.

[0032] For each group of live broadcast anomaly indexes in the current live broadcast room's last c live broadcasts, the average value is taken and then the ratio is calculated with the set live broadcast anomaly threshold index. That is, the average value is used as the numerator and the set live broadcast anomaly threshold index is used as the denominator to obtain the risk quantification ratio of the current live broadcast room.

[0033] The historical anomaly index (PLS) of the current live stream is obtained by comprehensively calculating the continuous risk ratio and risk quantification ratio of the current live stream.

[0034] In some embodiments, the specific process of obtaining the historical dissent index (PLS) of the current live stream is as follows:

[0035] Using formula The historical anomaly index (PLS) of the current live stream is obtained by comprehensively calculating the continuous risk ratio and the risk quantification ratio of the current live stream; among which... These represent the current live stream's ongoing risk ratio and risk quantification ratio, respectively. These are the weighting coefficients set for the continuous risk ratio and the risk quantification ratio, respectively.

[0036] In some embodiments, determining the current trend anomaly level of the live stream specifically involves:

[0037] The historical anomaly index PLS of the current live stream is compared with the set threshold index range. If the historical anomaly index PLS of the current live stream is within the threshold index range, the trend anomaly level of the current live stream is determined to be the general trend level, and a warning signal is sent to the current live stream.

[0038] If the historical anomaly index (PLS) of the current live stream is higher than the threshold index range, the current live stream is determined to be at the trend risk level, the live stream permission of the current live stream is banned, and the ban is set for an extended period.

[0039] A live streaming content recognition method based on virtual localization and multimodal reasoning includes:

[0040] Data collection: Collect location data and user behavior data during the current live broadcast. Location data includes the virtual location set by the user and the actual location of the device. User behavior data includes the text of the chat in the live broadcast room, the timestamp of the message, and the reward record. The reward record includes the type and quantity of gifts.

[0041] Data inference: Extract location data and user behavior data during the current live broadcast process and perform cross-modal correlation analysis to determine the live broadcast anomaly index (PHE) during the current live broadcast process;

[0042] Risk Determination: Receive the PHE (Problem Index) of the live broadcast during the current live broadcast process, and selectively trigger risk warning signals. After triggering the risk warning signal, perform corresponding steps in combination with the historical data stored in the database to determine the trend anomaly level of the current live broadcast.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] This invention analyzes the difference between virtual and actual positioning, and combines it with bullet screen and reward behavior patterns to form a cross-modal association evidence chain. It comprehensively calculates the live stream anomaly index by combining the proportion of hot topic fraud, the proportion of abnormal bullet screens, and the proportion of reward spam. Each indicator corresponds to an independent weight and allowable threshold, realizing the quantitative classification of the degree of anomaly. It can more accurately reflect the risk level of the live stream and solve the problem that the existing technology relies on only a single modality of data for analysis, which cannot comprehensively and accurately identify abnormal behavior in the live stream process and is prone to misjudgment.

[0045] This invention calculates a historical anomaly index by comprehensively calculating the continuous risk ratio and the risk quantification ratio. The historical anomaly index is then compared with a set threshold index range to determine the level of trend abnormality. If it is within the threshold index range, it is judged as a general trend level and a warning signal is sent. If it is higher than the threshold index range, it is judged as a trend risk level, and the live streaming privilege is banned and rectification is required. This achieves differentiated handling, urging rectification for minor violations and deterring serious violations.

[0046] This invention improves the risk assessment capability of live streaming rooms by continuously accumulating historical risk data, establishing a dynamically updated database, and analyzing abnormal behavior trends in live streaming rooms over a long period of time. Attached Figure Description

[0047] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which:

[0048] Figure 1 This is a schematic diagram of the principle of the present invention;

[0049] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0050] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0051] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0052] Example 1

[0053] Please see Figure 1 As shown, the live content recognition system based on virtual positioning and multimodal reasoning includes a multimodal data transmission module, a multimodal reasoning module, and a live risk module.

[0054] The multimodal data transmission module is used to collect location data and user behavior data during the current live broadcast. The location data includes the virtual location set by the user and the actual location of the device. The user behavior data includes the bullet screen text, sending timestamps, and reward records in the live broadcast room. The reward records include the type and quantity of gifts.

[0055] Additional information: The multimodal data acquisition process is as follows:

[0056] The client SDK obtains the virtual location information actively set by the user (such as "select city" in the APP), along with the timestamp.

[0057] The device includes mobile and PC terminals. Mobile terminals obtain physical coordinates through GPS, base station positioning, and Wi-Fi fingerprinting (scanning the MAC address and signal strength of surrounding APs).

[0058] On the PC side: the location is inferred by resolving the IP address (calling geographic libraries such as MaxMind) and network operator information (such as broadband location), and the network interface information obtained by WebRTC is used to assist in verification;

[0059] Bullet screen text: Real-time capture of the live stream bullet screen API stream, along with the sending timestamp, account ID, and device ID;

[0060] Tipping records: Monitor payment system callbacks and collect tipping amounts, item types (such as "rockets" and "flowers"), tipping times, and tipper account information;

[0061] Account registration information: Synchronize account registration time, historical activity (such as number of broadcasts, interaction frequency), real-name authentication status, etc. from the user center database;

[0062] At the same time, user behavior data is preprocessed by default, including but not limited to denoising bullet comments: filtering emojis and special characters, and extracting the main text.

[0063] Standardize tipping behavior: Convert tipping amounts and items into uniform numerical characteristics (such as "1 yuan = 1 unit value").

[0064] The multimodal reasoning module is used to extract location data and user behavior data during the current live broadcast process and perform cross-modal correlation analysis to determine the live broadcast anomaly index (PHE) during the current live broadcast process.

[0065] Specifically:

[0066] Identify the number of online users at each time point during the current live broadcast. Based on the location data during the current live broadcast, extract the virtual location set by each online user at each time point and the actual device location. If the virtual location and the actual device location of an online user are different, mark it as an abnormal user and generate a triplet. The triplet is represented as (virtual location, actual device location, abnormal user).

[0067] For each triple generated at each time point, the actual location of the device is extracted and matched. The number of triples that are successfully matched by the actual location of the device is counted and recorded as the number of fake users.

[0068] For example, if 80% of the viewers in a live stream are actually located in Hangzhou, but their virtual locations are displayed as "Shanghai", "Shenzhen", or "Chengdu", this can be considered an anomaly in location clustering.

[0069] Obtain the account registration information of each fake user, extract the account registration time from the account registration information, and calculate the time difference between the time difference and the current live broadcast time to obtain the account registration duration of each fake user;

[0070] The number of fake users whose account registration duration is less than the set registration duration is counted as the number of fake verifications.

[0071] Extract the number of false verifications marked at each time point, multiply it by the preset quantity correction coefficient, and then round it to get the number of false verifications corrected at each time point in the current live broadcast room; the quantity correction coefficient is set to 0.8, and can be dynamically adjusted according to the actual situation in the future.

[0072] Additional explanation: By using a preset quantity correction coefficient, the number of real users in the number of fake verifications can be corrected. Some users may have just created accounts and started using the service. This helps to prevent real users from being mistakenly identified as fake users because they share the same positioning as most fake users.

[0073] The ratio of the number of online users and the number of fake corrections at each time point during the current live broadcast is calculated. That is, the number of fake corrections is used as the numerator and the number of online users is used as the denominator to obtain the proportion of fake users at each time point during the current live broadcast.

[0074] The average percentage of fake users at each time point is taken as the percentage of fraudulent activity in the current live stream.

[0075] Extract bullet screen text and sending timestamps from user behavior data during the current live stream;

[0076] Additional notes: The bullet screen text is preprocessed, including removing bullet screen text caused by the lucky bag password set by the streamer.

[0077] The system counts the total number of bullet comments during the current live stream, counts the number of bullet comments with the same content from each bullet comment, and marks bullet comments with a high number of bullet comments as abnormal bullet comment codes.

[0078] For abnormal bullet screen passwords, identify users who have sent abnormal bullet screen passwords in the live broadcast room and record them as spam users. Extract the number of times each spam user sent abnormal bullet screen passwords and the timestamp of each abnormal bullet screen password sent.

[0079] If a user who spams the screen sends more abnormal barrage commands than the set threshold number of times within a set time window, they are identified as a user with abnormal frequency.

[0080] Additional explanation: For example, if an account sends ≥8 abnormal bullet screen codes within one minute;

[0081] Extract the actual device location of each frequency-abnormal user. For frequency-abnormal users belonging to the same location, count the total number of abnormal barrage passwords sent, and calculate the percentage of the total number of barrages, which is recorded as the abnormal barrage percentage.

[0082] Extract tipping records from user behavior data during the current live stream;

[0083] The total number and types of gifts received during the current live stream are counted. The number of gifts belonging to the same gift type is counted to obtain the number of gifts received by each gift type in the current live stream. Gift types with a high number of gifts are marked as gift pattern types, and the number of gift pattern types is marked as gift pattern numbers.

[0084] Identify users who have given gifts in the live stream, and record users whose gifts belong to a gift pattern as reward pattern users. Extract the number of gifts given by each reward pattern user.

[0085] Extract the actual device location of each user with a tipping pattern. For users with the same tipping pattern, count the total number of gifts given and calculate the percentage of the total number of gifts given. This percentage is recorded as the tipping screen refresh rate.

[0086] The percentages of fraudulent popularity, abnormal bullet comments, and excessive tipping observed during the current live stream are respectively tagged as follows: ;

[0087] Using formula The Live Streaming Abnormality Index (PHE) is calculated by comprehensively analyzing the percentage of fraudulent activity, abnormal comments, and excessive tipping during the current live stream. These are the preset percentages for allowed popularity fraud, allowed percentages for allowed abnormal bullet comments, and allowed percentages for allowed tipping spam. The weighting coefficients for the percentage of trending fraud, the percentage of abnormal bullet comments, and the percentage of tipping spam, respectively;

[0088] Further explanation: By analyzing the difference between virtual and actual locations (triple tagging) and combining it with bullet screen / tipping behavior patterns, a cross-modal association evidence chain is formed;

[0089] The live streaming risk module is used to receive the live streaming anomaly index (PHE) during the current live streaming process and selectively trigger risk warning signals. After triggering the risk warning signal, it combines the historical data stored in the database to perform corresponding steps to determine the trend anomaly level of the current live streaming room. The trend anomaly level includes the general trend level and the trend risk level.

[0090] Specifically:

[0091] The live broadcast anomaly index (PHE) during the current live broadcast is compared with the set live broadcast anomaly threshold index. If the live broadcast anomaly index (PHE) during the current live broadcast is higher than the live broadcast anomaly threshold index, a risk warning signal is triggered.

[0092] After triggering the risk warning signal, the system extracts the live streaming data from the database that is c times closer to the current live streaming room. The live streaming data includes the live streaming time and the live streaming anomaly index (PHE); where c > 3, and the specific settings are determined by technical personnel.

[0093] The number of times the current live stream room triggered risk warning signals during the last c live streams is recorded as the number of risk events. The percentage of the number of risk events in c is calculated as the current live stream room's continuous risk ratio.

[0094] To add further explanation, the higher the current continuous risk ratio of a live stream, the higher the level of false data in the live stream in the last c live streams, which is likely to mislead the consumption tendencies of normal users and result in a higher level of risk for the live stream.

[0095] For each group of live broadcast anomaly indexes in the current live broadcast room's last c live broadcasts, the average value is taken and then the ratio is calculated with the set live broadcast anomaly threshold index. That is, the average value is used as the numerator and the set live broadcast anomaly threshold index is used as the denominator to obtain the risk quantification ratio of the current live broadcast room.

[0096] To clarify, if the risk quantification ratio of the current live stream is greater than 1 and relatively high, it indicates that the amount of false data in the last c live streams is consistently high, resulting in a high level of risk for the live stream.

[0097] Using formula The historical anomaly index (PLS) of the current live stream is obtained by comprehensively calculating the continuous risk ratio and the risk quantification ratio of the current live stream; among which... These represent the current live stream's ongoing risk ratio and risk quantification ratio, respectively. These are the weighting coefficients set for the continuous risk ratio and the risk quantification ratio, respectively.

[0098] The historical anomaly index PLS of the current live stream is compared with the set threshold index range. If the historical anomaly index PLS of the current live stream is within the threshold index range, the trend anomaly level of the current live stream is determined to be the general trend level, and a warning signal is sent to the current live stream to remind it to stop falsifying live stream data.

[0099] If the historical anomaly index (PLS) of the current live stream is higher than the threshold index range, the current live stream is determined to be at the trend risk level, the live stream permission of the current live stream is banned for a set duration, and a rectification signal is sent to remind the live stream to rectify and learn.

[0100] Additional explanation: By comprehensively calculating the historical anomaly index PLS, the risk level is divided into "moderate trend" and "risk trend" to achieve differentiated handling: Moderate trend level: a warning signal is sent to urge rectification and avoid minor violations from escalating;

[0101] Trend Risk Level: Directly banning live streaming privileges and requiring rectification, serving as a deterrent to serious violations;

[0102] Automatically triggering corresponding measures based on data indicators reduces the cost of manual intervention while ensuring consistent handling standards and avoiding subjective bias. For example, automatic blocking is performed when PLS exceeds the threshold range, thus improving platform management efficiency.

[0103] Example 2

[0104] Please see Figure 2 As shown, based on the live content recognition system based on virtual positioning and multimodal reasoning provided in Embodiment 1 of this application, Embodiment 2 of this application proposes a live content recognition method based on virtual positioning and multimodal reasoning. Embodiment 2 is merely a preferred embodiment of Embodiment 1, and the implementation of Embodiment 2 will not affect the individual implementation of Embodiment 1.

[0105] Specifically, the live content recognition method based on virtual positioning and multimodal reasoning provided in Embodiment 2 of this application differs in that it includes:

[0106] Data collection: Collect location data and user behavior data during the current live broadcast. Location data includes the virtual location set by the user and the actual location of the device. User behavior data includes the text of the chat in the live broadcast room, the timestamp of the message, and the reward record. The reward record includes the type and quantity of gifts.

[0107] Data inference: Extract location data and user behavior data during the current live broadcast process and perform cross-modal correlation analysis to determine the live broadcast anomaly index (PHE) during the current live broadcast process;

[0108] Risk Determination: Receive the PHE (Problem Index) of the live broadcast during the current live broadcast process, and selectively trigger risk warning signals. After triggering the risk warning signal, perform corresponding steps in combination with the historical data stored in the database to determine the trend anomaly level of the current live broadcast.

[0109] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A live streaming content recognition system based on virtual positioning and multimodal reasoning, characterized in that, include: Multimodal data transmission module: collects location data and user behavior data during the current live broadcast process; The location data includes the virtual location set by the user and the actual location of the device; the user behavior data includes the bullet screen text in the live broadcast room, the sending timestamp, and the reward record; the reward record includes the type and quantity of gifts; Multimodal reasoning module: After extracting location data and user behavior data and processing them comprehensively, the module determines the proportion of hot topic fraud, abnormal bullet comments, and tipping spam. It then performs cross-modal correlation analysis on the proportion of hot topic fraud, abnormal bullet comments, and tipping spam to determine the live broadcast abnormality index (PHE) during the current live broadcast. The specific steps to determine the percentage of abnormal bullet comments are as follows: Extract bullet screen text and sending timestamps from user behavior data during the current live stream; The number of bullet comments with the same content is counted, and the bullet comment with the highest number of bullet comments is marked as an abnormal bullet comment password; Users who send abnormal bullet screen codes in the live broadcast room are identified as spammers. The number of times each spammer sends abnormal bullet screen codes and the sending timestamp are extracted. If a spammer sends abnormal bullet screen codes more than a set threshold number of times within a set time window, it is determined to be a user with abnormal frequency. Extract the actual device location of each frequency-abnormal user. For frequency-abnormal users belonging to the same location, count the total number of abnormal barrage passwords sent, and calculate the percentage of the total number of barrages, which is recorded as the abnormal barrage percentage. Live Streaming Risk Module: Receives the Live Streaming Anomaly Index (PHE) and selectively triggers risk warning signals. After triggering the risk warning signal, it combines the historical data stored in the database to perform corresponding steps to determine the current trend anomaly level of the live streaming room.

2. The live content recognition system based on virtual positioning and multimodal reasoning according to claim 1, characterized in that, The specific steps for determining the proportion of fraudulent activity related to trending topics are as follows: Identify the number of online users at each time point during the current live broadcast. Extract the virtual location set by each online user at each time point and the actual device location. If they are different, mark them as abnormal users and generate a triplet representing the virtual location, actual device location, and abnormal user. For each triplet generated at each time point, extract the actual device location and match it. Count the number of triplets that match the actual device location successfully and record it as the number of fake users. Obtain the account registration information of each fake user, extract the account registration time from the account registration information, and calculate the time difference between the time difference and the current live broadcast time to obtain the account registration duration of each fake user; The number of fake users whose account registration time is less than the set registration time is the number of fake verifications; the number of fake verifications marked at each time point is extracted, multiplied by the preset number correction coefficient, and then rounded to obtain the corrected number of fake verifications at each time point in the current live broadcast room; Calculate the ratio of the number of online users to the number of fake corrections at each time point during the current live broadcast to obtain the proportion of fake users at each time point during the current live broadcast. The average value is taken as the percentage of fraudulent activity in the current live stream.

3. The live content recognition system based on virtual positioning and multimodal reasoning according to claim 1, characterized in that, The specific steps for determining the percentage of screen-scrolling rewards are as follows: Extract tipping records from user behavior data during the current live stream; count the total number and types of gifts during the current live stream, count the number of gifts for each gift type, mark the gift type with the highest number of gifts as the gift pattern type, and mark the number of gift pattern types as the gift pattern quantity; Identify users who have given gifts in the live stream. Users whose gifts belong to a gift pattern are recorded as reward pattern users. Extract the number of gifts given by each reward pattern user. Extract the actual device location of each reward pattern user. Calculate the total number of gifts given by reward pattern users belonging to the same location. Calculate their percentage of the total number of gifts and record it as the reward spam percentage.

4. The live content recognition system based on virtual positioning and multimodal reasoning according to claim 1, characterized in that, The determination of the Public Anomaly Index (PHE) during the current live stream process is specifically as follows: The percentages of fraudulent popularity, abnormal comments, and excessive tipping observed during the current live stream are respectively tagged as follows: ; ； in These are the preset percentages for allowed popularity fraud, allowed percentages for allowed abnormal bullet comments, and allowed percentages for allowed tipping spam. The weighting coefficients for the percentage of trending fraud, the percentage of abnormal bullet comments, and the percentage of tipping spam.

5. The live content recognition system based on virtual positioning and multimodal reasoning according to claim 1, characterized in that, Receive the Live Streaming Anomaly Index (PHE) and selectively trigger risk warning signaling, specifically as follows: The live broadcast anomaly index (PHE) during the current live broadcast is compared with the set live broadcast anomaly threshold index. If the live broadcast anomaly index (PHE) during the current live broadcast is higher than the live broadcast anomaly threshold index, a risk warning signal is triggered.

6. The live content recognition system based on virtual positioning and multimodal reasoning according to claim 1, characterized in that, The steps to be performed after triggering the risk warning signaling, combined with the historical data stored in the database, are as follows: After triggering the risk warning signal, extract the live streaming data from the database that is c times closer to the current live streaming room. The live streaming data includes the live streaming time and the live streaming anomaly index (PHE). Where c > 3; The number of times the current live stream room triggered risk warning signals during the last c live streams is recorded as the number of risk events. The percentage of the number of risk events in c is calculated as the current live stream room's continuous risk ratio. The risk quantification ratio of the current live stream is obtained by taking the average of the abnormal indexes of each group of live streams in the most recent c live streams and the set abnormal comment threshold index. The historical anomaly index (PLS) of the current live stream is obtained by comprehensively calculating the continuous risk ratio and risk quantification ratio of the current live stream.

7. The live content recognition system based on virtual positioning and multimodal reasoning according to claim 6, characterized in that, The specific process for obtaining the historical dissent index (PLS) of the current live stream is as follows: ;in These represent the current live stream's ongoing risk ratio and risk quantification ratio, respectively. These are the weighting coefficients set for the continuous risk ratio and the risk quantification ratio, respectively.

8. The live content recognition system based on virtual positioning and multimodal reasoning according to claim 7, characterized in that, The determination of the current trend anomaly level in the live stream is specifically as follows: The historical anomaly index PLS of the current live stream is compared with the set threshold index range. If the historical anomaly index PLS of the current live stream is within the threshold index range, the trend anomaly level of the current live stream is determined to be the general trend level, and a warning signal is sent to the current live stream. If the historical anomaly index (PLS) of the current live stream is higher than the threshold index range, the current live stream is determined to be at the trend risk level, and the live stream permission of the current live stream is banned for a set duration.

9. A live content recognition method based on virtual localization and multimodal reasoning, applied to the live content recognition system based on virtual localization and multimodal reasoning as described in any one of claims 1-8, characterized in that, include: Data collection: Collect location data and user behavior data during the current live broadcast. The location data includes the virtual location set by the user and the actual location of the device; the user behavior data includes the bullet screen text in the live broadcast room, the sending timestamp, and the reward record; the reward record includes the type and quantity of gifts; Data inference: Extract location data and user behavior data during the current live broadcast process and perform cross-modal correlation analysis to determine the live broadcast anomaly index (PHE) during the current live broadcast process; Risk Determination: Receive the PHE (Problem Index) of the live broadcast during the current live broadcast process, and selectively trigger risk warning signals. After triggering the risk warning signal, perform corresponding steps in combination with the historical data stored in the database to determine the trend anomaly level of the current live broadcast.

Images