Intelligent live broadcast recommendation method and system based on big data

Abstract

The invention relates to the technical field of live broadcast recommendation, in particular to an intelligent live broadcast recommendation method and system based on big data, and the method comprises the following steps: obtaining user social interaction data, recording text comments, picture sharing and like behaviors, extracting keywords and mapping the keywords to interest classifications, combining the like frequency and interest labels to update dynamic interest records, and recommending the like. And matching the time sequence interest tag set to generate a cross-platform interest map, counting distribution of watching contents of social circle members, adding recommended contents, and generating a live broadcast recommendation list. According to the invention, the social interaction data of the user is acquired and the specific behavior is recorded, the creation of the dynamic interest label record can more accurately reflect the current interest and behavior mode of the user, and the watching type and duration of the live broadcast platform and the keyword in the comment are combined to ensure that the recommended live broadcast content is highly matched with the actual interest of the user, so that the user experience is improved. The social correlation of recommendation is enhanced, and the accuracy of live broadcast recommendation and the interaction rate of users are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of live broadcast recommendation, and particularly to an intelligent live broadcast recommendation method and system based on big data. Background Art

[0002] The technical field of live broadcast recommendation includes the analysis of user viewing habits and preferences as well as the personalized push of real-time content. The purpose of such technologies is to enhance the user viewing experience and increase user interaction. This field relies on big data analysis, user behavior research, and real-time data processing technologies. By analyzing user viewing data, interaction records, and other relevant information, the system can predict and recommend live broadcast content that users are interested in. The core content of the technical field includes data collection, user profile establishment, the design of recommendation algorithms, and the implementation of real-time recommendation systems. The components jointly support the comprehensive development of the live broadcast recommendation system.

[0003] Among them, the intelligent live broadcast recommendation method refers to using machine learning algorithms to analyze and process a large amount of user data in order to automatically recommend the most suitable live broadcast content for each user. The technical matters involved in this patent theme include the collection and analysis of user data, the application of recommendation algorithms, and the optimization of content matching strategies. Specifically, the intelligent live broadcast recommendation method identifies the user's interest points by analyzing the user's viewing records, search habits, and interaction behaviors, and uses this information to push the most relevant live broadcast content. The process mainly relies on structured data processing and precise algorithm operations.

[0004] The existing technologies rely on big data analysis and user behavior research. Although they can process and analyze a large amount of user data, they ignore the rapid changes in user interests and social influence factors. This results in the recommendation system being unable to adjust and reflect the user's latest interest points in a timely manner, causing a deviation between the recommended content and the user's actual needs, and reducing the user's viewing experience and interactivity. Traditional recommendation systems mostly focus on the analysis of user behavior data and rarely consider real-time interaction data and changes in the social environment. This leads to untimely or low-correlation recommendations in the rapidly changing live broadcast scenario, affecting user stickiness and the activity of the platform. Through analysis, it can be seen that although the existing technologies build a basic recommendation framework, there is still room for improvement in the flexibility and adaptability of the processing logic. Summary of the Invention

[0005] The purpose of the present invention is to solve the disadvantages existing in the prior art, and to propose an intelligent live broadcast recommendation method and system based on big data.

[0006] In order to achieve the above purpose, the present invention adopts the following technical solutions. The intelligent live broadcast recommendation method based on big data includes the following steps:

[0007] S1: Obtain the user's social interaction data, record the user's text comments, picture sharing, and like behaviors, extract keywords from the text content and map them to the preset live broadcast interest classification, assign initial weights according to the like behavior frequency, and combine with the interest tags of the user ID to obtain the dynamic interest tag record;

[0008] S2: Based on the timestamp sequence associated with the user ID in the dynamic interest tag record, calculate the sliding weighted value of the tag interaction times within the recent period, sort and adjust the tag weights, and generate the time-series adjusted interest tag set;

[0009] S3: Call the data of the watched content type, single viewing duration, and comment keywords in the user behavior of the live broadcast platform, extract the standard interest classification corresponding to the content type, match the similarity between the standard classification and the time-series adjusted interest tag set, and assign live broadcast behavior weights to the successfully matched tags to generate a cross-platform interest map;

[0010] S4: Use the tag list corresponding to the user ID in the cross-platform interest map to calculate the similarity score of the live broadcast content tags, correct the score in combination with the live broadcast real-time interaction data and sort it to generate the user real-time recommended content queue.

[0011] As a further solution of the present invention, the dynamic interest tag record includes keyword weights, social interaction types, and interest tag mapping relationships. The time-series adjusted interest tag set includes tag sliding weighted values, interest tag time decay rates, and tag weight sorting. The cross-platform interest map includes content type similarity, tag association degree, and live broadcast behavior weights. The user real-time recommended content queue includes tag real-time similarity, content real-time interaction scores, and recommendation sorting levels.

[0012] As a further solution of the present invention, the acquisition steps of the dynamic interest tag record are specifically as follows:

[0013] S111: Obtain the user's social interaction data, record the user's text comment content, picture sharing annotation words, and like behaviors, perform word segmentation processing on the comment text, extract keywords and perform category mapping with the set live broadcast interest classification table, and identify the term combinations that have a matching relationship with the interest tag library to obtain the user interest association category set;

[0014] S112: According to the user interest association category set, call the user like behavior record, count the frequency of user likes in each type of interest-associated content, screen the interest categories whose frequency exceeds the like behavior interest bias threshold, and combine the number of user comment words and like frequencies in multiple categories to obtain the live broadcast content preference ranking sequence;

[0015] S113: Based on the live content preference ranking sequence, call the live room tag information, screen the live room content that matches the previous item of the ranking sequence, perform a combined score on the successfully matched live rooms according to the activity ratio and the popularity of the live room, and adjust the priority of the recommendation results according to the scores to obtain a dynamic interest tag record.

[0016] As a further solution of the present invention, the acquisition steps of the time-sequence adjusted interest tag set are specifically as follows:

[0017] S211: Based on the timestamp sequence associated with the user ID in the dynamic interest tag record, extract the set of interaction time points of the tags, calculate the time difference between the time points and the moments, and generate a set of time decay factors;

[0018] S212: Call the set of time decay factors, combine with the number of interactions of the tags within the period, and use the formula:

[0019]

[0020] Calculate the sliding weighted value of the tag, and generate a sliding weighted value sequence in descending order;

[0021] Among them, W c represents the sliding weighted value of tag c, λ t represents the time decay factor of time unit t, I t represents the number of interactions of the tag in time unit t, σ c represents the dispersion of the number of interactions of tag c, μ c represents the average value of the number of interactions of tag c, ν t represents the interaction frequency offset of time unit t, and T represents the set of all time units;

[0022] S213: Call the sliding weighted value sequence, perform interpolation fusion on the real-time period tag weights and the weight sequence, screen the tags whose weights exceed the dynamic threshold according to the interpolation results, and generate a time-sequence adjusted interest tag set.

[0023] As a further solution of the present invention, the acquisition steps of the cross-platform interest map are specifically as follows:

[0024] S311: According to the time-sequence adjusted interest tag set, call the viewing content type data of the live broadcast platform, match the keyword mapping and semantic information, extract the standard interest classification corresponding to each viewing record, and generate a classification mapping set;

[0025] S312: Based on the classification mapping set, call the vectorized expression of the time-sequence adjusted interest tag set, calculate the similarity between the vector of each standard interest classification and the tag vector in the tag set, and use the formula:

[0026]

[0027] Calculate the label matching degree;

[0028] Among them, S ij represents the matching degree between the i-th content and the j-th label, a i represents the single viewing duration of the i-th content, k j represents the number of comment keyword matches of the j-th label, c ib represents the occurrence frequency of the i-th content type in the b-th period, v j and v b represent the vector norms of label j and label b respectively, and n is the number of periods;

[0029] S313: According to the label matching degree, superimpose the duration weight and comment keyword weight of the corresponding viewing content type for each matching label, calculate the total value, and screen the labels whose total weight exceeds the benchmark of the live broadcast platform to generate a cross-platform interest map.

[0030] As a further solution of the present invention, the step of obtaining the user real-time recommended content queue is specifically as follows:

[0031] S411: Call the label list corresponding to the user ID in the cross-platform interest map, calculate the cosine similarity between each pair of live content labels, classify and store the calculation results according to the user ID, and generate an initial label similarity value;

[0032] S412: Based on the initial label similarity value, obtain the like rate, comment rate and viewing duration ratio in the live real-time interaction data, and use the formula:

[0033]

[0034] Calculate the label corrected similarity, sum the multiple label correction values of each live content and then take the average value to generate an interaction corrected similarity;

[0035] Among them, M op is the label corrected similarity, AB op represents the initial label similarity value between user o and live content p, Q1 represents the like rate of live content p, Q2 represents the comment rate of live content p, and T represents the time difference decay factor;

[0036] S413: Call the interaction corrected similarity, sort the values, and combine the reverse weighted value of the user click behavior in the real-time interaction data to perform a secondary adjustment on the sorting result to generate a user real-time recommended content queue.

[0037] As a further solution of the present invention, the method further includes step S5:

[0038] S5: Extract the cross - platform interest graph tags corresponding to the user's social circle member list, count the distribution of the content types watched by the members, add points to the content scores of the matching types in the user's real - time recommended content queue, and generate a live broadcast recommendation list;

[0039] The live broadcast recommendation list includes the social circle tag distribution weight, the content type matching score, and the social interest addition coefficient.

[0040] As a further solution of the present invention, the steps for obtaining the live broadcast recommendation list are specifically as follows:

[0041] S511: Use the user's real - time recommended content queue to extract the data of the user's social circle members, identify the user's interest preferences, analyze the content and interest manifestations of the user's interaction with the social circle members, and generate cross - platform interest tags for the user's social circle;

[0042] S512: Based on the cross - platform interest tags of the user's social circle, count the content types watched by each member on multiple live broadcast platforms, classify and summarize the viewing frequencies according to the time period, and combine the frequencies of the viewing types to obtain the distribution of the content types watched by the members;

[0043] S513: According to the distribution of the content types watched by the members and the interest tags, combined with the user's real - time recommended content queue, add points to the matching content types, identify the recommendation priority of each content, and update the content recommendation list according to the score addition to generate a live broadcast recommendation list.

[0044] The intelligent live broadcast recommendation system based on big data is used to execute the above - mentioned intelligent live broadcast recommendation method based on big data. The system includes:

[0045] The social interest recognition module calls the social platform API interface to obtain the user's text comments, picture sharing, and like behavior data, extracts the keywords in the text comments and maps them to the preset live broadcast interest classification, and assigns an initial weight according to the like behavior frequency to obtain a dynamic interest tag record;

[0046] The weight dynamic update module performs a sliding window weighted sum on the interaction times of the same interest tag based on the time - stamp sequence of the user ID in the dynamic interest tag record, calculates the tag activity index, sorts and adjusts after superimposing the initial weight to generate a time - series adjusted interest tag set;

[0047] The cross - platform fusion module calls the viewing content type, single - viewing duration, and comment keyword data corresponding to the user ID in the live broadcast platform behavior log, extracts the standard interest tags corresponding to the content type, calculates the cosine similarity between the standard tags and the tags in the time - series adjusted interest tag set, and generates a cross - platform interest graph;

[0048] The content matching module traverses the live broadcast titles and classification tags based on the cross-platform interest graph, calculates the similarity scores between the live broadcast content tags and the user interest tags, and corrects the scores by combining the like growth rate in the real-time interaction data to generate a user real-time recommended content queue;

[0049] The social circle correction module uses the user real-time recommended content queue, calls the user social circle member list, extracts the weighted interest tags of the members in the cross-platform interest coefficient library, statistically analyzes the distribution ratio of the content types watched by the members recently, and performs a fixed ratio addition to the content matching scores of the same type in the content matching priority queue to generate a live broadcast recommendation list.

[0050] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0051] In the present invention, by obtaining the social interaction data of users and recording specific behaviors such as text comments and likes, the creation of dynamic interest tag records can more accurately reflect the current interests and behavior patterns of users. Using the data to calculate the sliding weighted values of the tags and perform time series adjustment makes the interest tag set closer to the real-time interest changes of users, improves the relevance of content recommendation and user satisfaction. By combining the viewing type and duration of the live broadcast platform and the keywords in the comments, the content recommendation is further refined to ensure that the recommended live broadcast content highly matches the actual interests of users. By analyzing the viewing preferences of the members in the user's social circle and using the information to strengthen personalized recommendation, the recommendation is not only limited to the user's behavior but also incorporates social influence factors, enhances the social relevance of the recommendation, improves the accuracy of the recommendation and the interaction rate of users, because it can better capture and reflect the interests of users changing over time. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a schematic diagram of the working process of the present invention;

[0053] Figure 2 is a flowchart for obtaining dynamic interest tag records in the present invention;

[0054] Figure 3 is a flowchart for obtaining a time series adjusted interest tag set in the present invention;

[0055] Figure 4 is a flowchart for obtaining a cross-platform interest graph in the present invention;

[0056] Figure 5 is a flowchart for obtaining a user real-time recommended content queue in the present invention;

[0057] Figure 6 is a flowchart for obtaining a live broadcast recommendation list in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0058] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0059] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present invention. In addition, in the description of the present invention, the meaning of "a plurality of" is two or more, unless otherwise specifically defined.

[0060] Please refer to Figure 1 , the present invention provides a technical solution, an intelligent live broadcast recommendation method based on big data, including the following steps:

[0061] S1: Obtain the social interaction data of the user, record the user's text comments, picture sharing and like behaviors, extract keywords from the text content and map them to the preset live broadcast interest classification, assign an initial weight according to the like behavior frequency, and combine the interest tags of the user ID to obtain a dynamic interest tag record;

[0062] S2: Based on the time stamp sequence associated with the user ID in the dynamic interest tag record, calculate the sliding weighted value of the tag interaction times in the recent period, sort and adjust the tag weights, and generate a time series adjusted interest tag set;

[0063] S3: Call the data of the viewing content type, single viewing duration and comment keywords in the user behavior of the live broadcast platform, extract the standard interest classification corresponding to the content type, match the similarity between the standard classification and the time series adjusted interest tag set, and assign a live broadcast behavior weight to the successfully matched tags to generate a cross-platform interest map;

[0064] S4: Use the tag list corresponding to the user ID in the cross-platform interest map to calculate the similarity score of the live broadcast content tags, correct the score in combination with the live broadcast real-time interaction data and sort, and generate a user real-time recommendation content queue;

[0065] S5: Extract the cross-platform interest map tags corresponding to the user social circle member list, count the distribution of the member viewing content types, add points to the content scores of the matching types in the user real-time recommendation content queue, and generate a live broadcast recommendation list;

[0066] The dynamic interest tag record includes keyword weights, social interaction types, and interest tag mapping relationships. The time-series adjusted interest tag set includes tag sliding weighted values, interest tag time decay rates, and tag weight sorting. The cross-platform interest graph includes content type similarity, tag association degree, and live broadcast behavior weights. The user real-time recommended content queue includes tag real-time similarity, content real-time interaction scores, and recommendation ranking levels. The live broadcast recommendation list includes social circle tag distribution weights, content type matching scores, and social interest bonus coefficients.

[0067] Please refer to Figure 2 , and the specific steps for obtaining the dynamic interest tag record are as follows:

[0068] S111: Obtain the user's social interaction data, record the user's text comment content, picture sharing annotation words, and like behavior, perform word segmentation on the comment text, extract keywords, and perform category mapping with the set live broadcast interest classification table to identify the term combinations that have a matching relationship with the interest tag thesaurus, and obtain the user interest association category set;

[0069] Obtain the social interaction data of the user. The operations of recording the user's text comment content, picture sharing annotation words, and like behavior are to capture all the social interaction records of user ID A1009 in the past 7 days through the platform server. The statistical results show that the user has submitted a total of 321 interaction records, including 187 text comments, 58 picture shares, and 76 like behaviors. Subsequently, semantic word segmentation processing is performed on each text comment content. Using the built-in Chinese word segmentation dictionary, all comments are split into 603 non-repeating word items, and their word frequency distributions are counted. The word frequency threshold is set to 3 times, and low-frequency word items that appear less than 3 times are removed. 143 word items are retained as the candidate word set for interest determination. Call the 16 types of live interest tags set in the live interest classification table, including tags such as food, games, travel, music, variety shows, and technology. According to the word meanings, one-to-one mapping and matching are performed between the candidate word items and the interest tags. Set the candidate word items "barbecue", "dessert", "takeout", and "recipe" to be mapped to the "food" category, "ranking", "clearing the level", "streak", and "bullet screen" to be matched to the "games" category, "route", "hotel", "homestay", and "scenery" to correspond to the "travel" category, and the rest such as "folk song" and "performance" belong to the "music" category, etc. After the matching is completed, count the number of hits by category. Among them, the matching keywords in the "food" category hit a total of 12 times, the "games" category 15 times, the "travel" category 9 times, the "music" category 3 times, and both the "variety show" category and the "technology" category are 0 times. Set the interest recognition category threshold to 5 times, exclude the categories with insufficient hit counts, and the categories entering the next stage of processing are "food", "games", and "travel". Identify the annotation information of the pictures uploaded by the user, map "eating live broadcast" to "food" and "scenery photo" to "travel", and according to the frequencies of 2 times and 1 time that appear in the picture sharing records respectively, add them to the corresponding categories of the matching interest tags as supplementary evidence. Retrieve the user's like behavior records, aggregate and count the live content liked by category. The content in the "food" category received 18 likes, the "games" category 21 likes, and the "travel" category 10 likes. Classify and organize the number of keyword hits, picture label times, and like frequencies that have appeared in the interest categories, and bind them under user ID A1009. The data in this set shows that the "food" category has 4 keyword hits, 2 picture annotations, and 18 likes, the "games" category has 4 keyword hits and 21 likes, the "travel" category has 4 keyword hits, 1 picture annotation, and 10 likes, and the remaining categories do not meet the word frequency and interaction quantity requirements and are not recorded. After the above matching and screening process, the user interest association category set is obtained.

[0070] S112: According to the user interest association category set, call the user's like behavior records, count the frequency of likes of the user in each type of interest-associated content, screen the interest categories whose frequencies exceed the like behavior interest bias threshold, and combine the number of comment words and like frequencies of the user in multiple categories to obtain the live content preference ranking sequence;

[0071] Call the user's like behavior record, archive the detailed like behavior of user ID A1009, and establish an index list of liked content according to the above three interest categories of "Food", "Games", and "Travel". Among them, the food category has 18 likes, involving a total of 7 live room numbers, the game category has 21 likes, involving 9 numbers, and the travel category has 10 likes, covering 4 numbers. Count the corresponding like frequencies under each interest category, and conduct a preliminary screening in combination with the like behavior interest bias threshold of 12 set by the platform. Both 18 times in the food category and 21 times in the game category exceed this threshold, and the travel category is temporarily excluded because it is lower than this threshold; further call the comment text processing record, and analyze that the total number of comment words of users under food category live broadcasts is 113, and that of the game category is 142. By performing normalized weighted calculation on the number of comment words and the number of likes, set the weight parameter of the like frequency weight to 0.6 and the weight of the number of comment words to 0.4, and use the following weighted formula to calculate the live broadcast interest activity ratio:

[0072]

[0073] Among them, R i is the activity ratio of interest category i, L i represents the number of likes in this category, max(L) is the maximum number of likes in all categories, C i is the total number of comment words in this category, max(C) is the maximum number of comment words in all categories. After substituting the data:

[0074] Food category:

[0075]

[0076] Game category:

[0077]

[0078] After sorting the above activity ratios, generate a preference intensity ranking. The sorting result is Games > Food, and mark the travel category as an interest record but do not enter the preference sequence. Arrange the user's interest categories in order according to the preference value, and thus obtain the live broadcast content preference sorting sequence. Among them, the game category content is the top recommended category, the food category is the second, and the travel category is used as a backup interest record but does not enter the recommended priority sorting.

[0079] S113: Based on the live broadcast content preference sorting sequence, call the live room tag information, screen the live room content that matches the previous item of the sorting sequence, combine and score the matching live rooms according to the activity ratio and the popularity of the live rooms, and adjust the priority of the recommendation results according to the scores to obtain the dynamic interest tag record;

[0080] After calling the currently available live room tag information, the platform classifies and organizes the live rooms labeled with "game" and "food" tags, screens out the live rooms whose tag fields contain the corresponding interest keywords. The preliminary matching results show that there are 112 game live rooms and 76 food live rooms. The system extracts the user behavior index data of these two types of live rooms within the past 48 hours, covering the average stay duration of viewers, the number of entries, the number of likes, and the number of bullet screens, etc. The platform extracts the index data fields item by item through the live content index table and generates a list of live room behavior feature vectors. Taking the game live room as an example, the average stay duration of the viewers of the live room numbered G301 within the statistical period is 43 minutes, the number of likes is 1,487 times, the number of bullet screens is 560, and the number of entries is 3,423 times. For the food live room numbered F207, the stay duration is 38 minutes, the number of likes is 1,222 times, the number of bullet screens is 411, and the number of entries is 2,780 times. The system standardizes the above behavior feature data so that it is comparable under the same numerical scale. After standardization, it calls the user preference ranking results to compare each live room according to the interest category, and preferentially selects the live content under the category with a higher preference ranking. For example, if "game" takes precedence over "food" in the current ranking, the system calculates the recommendation priority of the game live rooms first. Based on the distribution position of the audience behavior data on the standard scale, the live rooms higher than the 75th percentile are set as high-active candidates. Further, the live rooms with a suspension record within the past 48 hours or an interaction frequency lower than the platform's benchmark active value (such as a bullet screen density lower than 8 per minute) are excluded. The remaining candidate live rooms are marked with priorities according to the user stay trend and the like interaction frequency. The top ten live room contents with the highest scores under each category are loaded in sequence according to the user preference category and the recommendation output is generated. The recommendation order is combined with the original order in the preference ranking sequence and written into the user ID tag field to output the live content recommendation list corresponding to the user in the current time period, forming a dynamic interest tag record.

[0081] Please refer to Figure 3 , and the specific steps for obtaining the time series adjusted interest tag set are as follows:

[0082] S211: Based on the time stamp sequence associated with the user ID in the dynamic interest tag record, extract the set of interaction time points of the tags, calculate the time difference between the time points and the moments, and generate a set of time decay factors;

[0083] Extract the interaction time points of user U1 with tag L1 within the past three natural days to obtain the time stamp sequence

[0084] [[2025-04-09 14:30:00, 2025-04-10 09:15:00, 2025-04-11 16:45:00]], convert each time point to the time difference from the current moment (2025-04-11 17:00:00), calculate them as 2.10 days, 1.31 days, and 0.02 days respectively, take the reciprocal of each time difference to obtain the time decay factors [0.476, 0.763, 50.000], and generate a set of time decay factors. In the specific process, the time difference calculation is expressed in decimal days accurate to the hour. For example, 2 days and 2.5 hours is converted to 2.10 days, and the reciprocal value is retained to three significant figures;

[0085] Table 1 Calculation Table of Time Decay Factor for Label L1

[0086]

[0087]

[0088] As shown in Table 1, the time decay factor increases exponentially as the interaction time approaches the current moment, and the time decay factor for the most recent 15-minute interaction time reaches 50.000.

[0089] S212: Call the set of time decay factors, combine with the number of interactions within the period of the label, and use the formula:

[0090]

[0091] Calculate the sliding weighted value of the label, and generate a sliding weighted value sequence in descending order;

[0092] Among them, W c represents the sliding weighted value of label c, λ t represents the time decay factor of time unit t, I t represents the number of interactions of the label in time unit t, σ c represents the dispersion of the number of interactions of label c, μ c represents the mean value of the number of interactions of label c, ν t represents the interaction frequency offset of time unit t, and T represents the set of all time units;

[0093] Call the three periodic factors [0.476, 0.763, 50.000] of label L1 in the set of time decay factors, obtain the corresponding number of interactions in each period [3, 5, 8], and calculate the mean value of the number of interactions:

[0094] μ c =(3 + 5 + 8) / 3 = 5.333;

[0095] Standard deviation of the number of interactions:

[0096]

[0097] Interaction frequency offset ν t Take the absolute difference between the number of interactions in each period and the mean as [2.333, 0.333, 2.667], and substitute it into the formula:

[0098]

[0099] The first period:

[0100] The second period:

[0101] The third period:

[0102] Total sliding weighted value:

[0103] W c = 0.637 + 1.437 + 183.944 = 186.018;

[0104] Generate a sliding weighted value sequence.

[0105] S213: Call the sliding weighted value sequence, interpolate and fuse the real-time period label weights with the weight sequence, and screen the labels whose weights exceed the dynamic threshold according to the interpolation result to generate a time series adjusted interest label set;

[0106] Extract the sliding weighted value W of label L1 from all labels within the real-time period c = 186.018, obtain the weight distribution data of labels within the period (see Table 2 for example data), calculate the weight median M = 85.000, set the dynamic threshold to 1.5M = 127.500, and compare W c with the threshold. Since 186.018 > 127.500, it is determined that label L1 meets the screening conditions, add it to the candidate set, sort all labels in the candidate set in descending order of weights, and select the top 20% of the labels (if the candidate set contains 100 labels, take the top 20), and generate a time series adjusted interest label set;

[0107] Table 2 Example of label weight distribution representation

[0108] Label Name Sliding Weight Value L1 186.018 L2 92.450 L3 78.330 L4 145.670 L5 63.210

[0109] As shown in Table 2, the range of label weights in the current period is from 63.210 to 186.018. The median M = 85.000 is determined as the middle value (the 50th percentile) after sorting. The dynamic threshold 1.5M = 127.500 is used to filter low-active labels. During the weight calculation process, due to recent intensive interactions of label L1 (such as 8 interactions in the third period and a time decay factor of 50.000), its weighted value is significantly higher than the threshold;

[0110] In specific implementation, the calculation of the dynamic threshold uses box plot statistics: Calculate the 25th percentile Q1 = 72.500 and the 75th percentile Q3 = 120.800 of the weight distribution, and obtain the interquartile range IQR = Q3 - Q1 = 48.300. The dynamic threshold is set to

[0111] Q3 + 1.5×IQR = 120.800 + 72.450 = 193.250. If this method is adopted, the weight of label L1, 186.018, is still lower than the threshold 193.250. However, due to actual business requirements, the median multiple method is adjusted to improve the screening sensitivity. When generating the candidate set, the sorted label list is intercepted proportionally. For example, for

[0112] [186.018, 145.670, 92.450, 78.330, 63.210] after sorting, if the top 20% corresponds to 1 label (total 5), then only L1 is selected into the set.

[0113] Please refer to Figure 4 , and the steps for obtaining the cross-platform interest graph are specifically as follows:

[0114] S311: According to the time series adjustment of the interest label set, call the viewing content type data of the live broadcast platform, match the keyword mapping and semantic information, extract the standard interest classification corresponding to each viewing record, and generate a classification mapping set;

[0115] Call the viewing records of user A on the live broadcast platform, which include two types of content: "Digital product reviews" and "Fitness tutorials". The preset standard interest classification library includes categories such as "Technology" and "Health". For the "Digital product reviews" content, extract keywords "mobile phone", "parameters", "evaluation", and match them one by one with the keyword library of the "Technology" category in the standard classification library. The number of matching keywords is counted as 4, exceeding the preset matching threshold of 3, and map "Digital product reviews" to the "Technology" classification. For the "Fitness tutorials" content, extract keywords "fat burning", "actions", "diet", and the number of matches with the "Health" category keyword library is 5, exceeding the threshold of 3, and map it to the "Health" classification, generating a classification mapping set containing the mapping results of all viewing records of user A;

[0116] Table 3 Example table of content type mapping

[0117]

[0118]

[0119] As shown in Table 3, the number of keyword matches is achieved by counting the number of words that overlap with the standard classification library in the content description text. The preset matching threshold is set to 3 based on the average number of keywords in the classification library. When the number of matches exceeds the threshold, classification mapping is triggered to generate a classification mapping set.

[0120] S312: Based on the classification mapping set, call the vectorized expression of the time-series adjusted interest tag set, calculate the similarity between the vector of each standard interest classification and the tag vectors in the tag set, using the formula:

[0121]

[0122] Calculate the tag matching degree;

[0123] Among them, S ij represents the matching degree between the i-th piece of content and the j-th tag, a i represents the single-viewing duration of the i-th piece of content, k j represents the number of keyword matches of the j-th tag in the comments of the i-th piece of content, c ib represents the occurrence frequency of the i-th content type in the b-th period, v j and v b represent the vector norms of tag j and tag b respectively, and n is the number of periods;

[0124] Taking the "Technology" classification of user A as an example, its vector representation is [0.7, 0.3, 0.1]. The time-series adjusted interest tag set includes tags such as "Electronic Products" and "Software Development". Calculate the cosine similarity between "Technology" and "Electronic Products":

[0125] Parameter definition:

[0126] a i = 1200 seconds (the single-viewing duration of user A for technology-related content);

[0127] k j = 8 (the number of keyword matches of the "Electronic Products" tag in the comments of user A);

[0128] c ib = 0.6 (the frequency of user A watching technology-related content in the past 3 periods);

[0129] v j = 1.2 (the vector norm of the "Electronic Products" tag);

[0130] v b= 1.0 (average of the norms of other label vectors);

[0131] Calculate the numerator part:

[0132]

[0133] Denominator part calculation:

[0134]

[0135] Substitute into the formula for calculation:

[0136]

[0137] The dynamic benchmark is set to 6.0 according to the median of the label set similarity distribution. The matching degree of 8.85 is higher than the benchmark, and "electronic products" is added to the time series adjustment interest label set.

[0138] S313: According to the label matching degree, superimpose the duration weight and comment keyword weight of the corresponding viewing content type on each matching label, calculate the total value, screen the labels whose total weight exceeds the benchmark of the live broadcast platform, and generate a cross-platform interest map;

[0139] Superimpose the duration weight and comment keyword weight of the corresponding viewing content type on each matching label. Taking the "electronic products" label as an example, its duration weight is 1200 / 3600 = 0.33 (normalized viewing duration), and the comment keyword weight is 8 / 10 = 0.8 (keyword matching ratio). The total weight is 0.33 + 0.8 = 1.13. The platform benchmark is the median 0.95 of the total weights of all labels. Since 1.13 is higher than the benchmark, "electronic products" is added as a core node to the cross-platform interest map. The size of the nodes in the map is positively correlated with the total weight, and the edge weight is associated with the number of cross-platform behavior associations, generating a cross-platform interest map.

[0140] Please refer to Figure 5 , the steps for obtaining the user real-time recommended content queue are specifically as follows:

[0141] S411: Call the label list corresponding to the user ID in the cross-platform interest map, calculate the cosine similarity between each pair of live content labels, and store the calculation results classified by user ID to generate the initial label similarity value;

[0142] Extract the label set of user o as games, technology, programming, and the label set of live content p as technology, programming, education. Convert the label list into a numerical vector. The user label vector is 1, 1, 1, 0 (corresponding to games, technology, programming, education), and the live content label vector is 0, 1, 1, 1. Calculate the cosine similarity as the dot product of the vectors divided by the product of the vector norms, that is:

[0143]

[0144] Mark 0.666 as the initial label similarity value between user o and live broadcast p. Repeat this operation for all user and live content label pairs and store it as structured data.

[0145] Table 4 Example Table for Calculating the Similarity between Users and Live Broadcast Labels

[0146]

[0147] As shown in Table 4, the initial label similarity value between user o and live broadcast p is 0.666, and the similarity with live broadcast q is 0.408. The similarity quantification is completed through vectorization and normalization operations.

[0148] S412: Based on the initial label similarity value, obtain the like rate, comment rate, and viewing duration ratio in the real-time interaction data of the live broadcast, and use the formula:

[0149]

[0150] Calculate the label-corrected similarity. Sum the multiple label-corrected values of each live content and then take the average to generate the interaction-corrected similarity.

[0151] Among them, M op is the label-corrected similarity, AB op represents the initial label similarity value between user o and live content p, Q1 represents the like rate of live content p, Q2 represents the comment rate of live content p, and T represents the time-difference decay factor.

[0152] Meaning and calculation process of parameters:

[0153] AB op : The initial label similarity value between user o and live content p, obtained through cosine similarity calculation; Q1: The like rate of live content p, calculated by dividing the number of likes by the total number of viewers;

[0154] Q2: The comment rate of live content p, calculated by dividing the number of comments by the total number of viewers;

[0155] T: The time-difference decay factor, calculated as T = 1 / (1 + Δt / 24) based on Δt (the time difference between the start time of the live broadcast and the last interaction time of the user);

[0156] Based on the initial label similarity value AB between user o and live content p op= 0.666, collect the real-time interaction data of live content p: the number of likes is 1500 times (the total number of viewers is 10000), the number of comments is 800 (the total number of viewers is 10000), calculate the like rate Q1 = 1500 / 10000 = 0.15, the comment rate Q2 = 800 / 10000 = 0.08, set the start time of live content p as 2025-04-11 14:00, the last interaction time of user o is 2025-04-11 12:00, the time difference Δt = 2 hours, the time difference decay factor:

[0157] T = 1 / (1 + Δt / 24) = 1 / (1 + 2 / 24) = 0.923;

[0158] Substitute into the formula:

[0159]

[0160] Sum and average the multiple tag correction values of live p (for example, the technology, programming, and education tags correspond to 0.666, 0.500, and 0.333 respectively), to get:

[0161] M op = (0.666 + 0.500 + 0.333) / 3 = 0.500;

[0162] Generate the interaction correction similarity.

[0163] S413: Invoke the interaction correction similarity, sort the values, and combine the reverse weighted values of the user click behavior in the real-time interaction data to perform a secondary adjustment on the sorting result to generate a user real-time recommendation content queue;

[0164] Invoke the interaction correction similarity of live p as 0.500 and the correction value of live q as 0.300, sort them in descending order of value as p: 0.500, q: 0.300, combine the click behavior data of user o in the past 1 hour (the number of clicks on p is 2 times, the number of clicks on q is 1 time), and calculate the reverse weighted value W p = 1 / (1 + 2) = 0.333, W q

[0165] = 1 / (1 + 1) = 0.500, the adjusted score Score p = 0.500 × 0.333 = 0.166, Score q

[0166] = 0.300 × 0.500 = 0.150, re-sort as p: 0.166, q: 0.150, and generate the user real-time recommendation content queue as live p, live q;

[0167] Table 5 Example table for generating the real-time recommendation queue

[0168]

[0169] As shown in Table 5, live broadcast p is weighted by high interactive correction similarity and moderate click behavior, and is ranked preferentially in the recommendation queue.

[0170] Please refer to Figure 6 , and the steps for obtaining the live broadcast recommendation list are specifically as follows:

[0171] S511: Using the user's real-time recommended content queue, extract the data of the members of the user's social circle, identify the user's interest preferences, analyze the content and interest performance of the user's interaction with the members of the social circle, and generate cross-platform interest tags for the user's social circle;

[0172] By performing cluster analysis on the behavioral data of the members, identify the types of viewing content and interaction frequencies of each member on different platforms. For example, if the viewing frequency of a certain member on a certain platform mostly focuses on game content, it can be inferred that the member prefers game content. By analyzing the interaction records between the members of the social circle, analyze the interest similarity between the members. If a member often interacts with another member and watches similar content, the interest tags of the two are highly similar. By cross-analyzing the interest performance of all members, obtain a set of cross-platform interest tags. Each tag represents the user's preference behavior on multiple social platforms. If the user prefers to watch technology videos on platform A and anime content on platform B, then their cross-platform interest tags are "technology" and "anime", forming cross-platform interest tags for the user's social circle.

[0173] S512: Based on the cross-platform interest tags of the user's social circle, count the types of content viewed by each member on multiple live broadcast platforms, classify and summarize the viewing frequencies according to the time period, and combine the frequencies of the viewing types to obtain the distribution of the types of content viewed by the members;

[0174] According to the viewing records and interest tags, classify the content viewed by the members on different platforms into several categories, such as film and television, games, sports, etc. The viewing frequency of each category can be obtained by calculating the total viewing duration or the number of viewing times of the user for the content of this category within a certain time period. If a certain member watched game content for 5 hours on platform A and sports content for 3 hours on platform B in the past week, then the distribution of the types of content viewed by this member on each platform can be calculated. According to the viewing frequency data, classify and summarize the types of content viewed by each member in different time periods (such as daily, weekly, monthly), generate a viewing frequency distribution map for each category according to the time period, and calculate the viewing proportion of each content type. For example, if the game category accounts for 60% and the film and television category accounts for 30% of the content viewed by the member in a week, then the distribution of the types of content viewed by this member is 60% games and 30% film and television, obtaining the distribution of the types of content viewed by the members.

[0175] S513: Based on the distribution of the content types watched by members and interest tags, combined with the real-time recommended content queue for users, perform score addition to the matching content types, identify the recommended priority for each piece of content, and update the content recommendation list according to the score addition to generate a live broadcast recommendation list;

[0176] Match based on the distribution of the content types watched by members and the content tags in the real-time recommended content queue. If a member's interest tags are "games" and "anime", and the distribution of the content types they watched shows that they prefer game content more, then the game content in the real-time recommendation queue will receive a high addition. For each content type, assign a score addition according to its matching degree in the user's interest tags. For example, if the recommended game content highly matches the member's interest tags, the score addition for this content can be set to +1.5. If the recommended anime content has a lower matching degree with the tags, the addition is +0.8. Sort the recommended content according to the scores after addition to generate the recommended priority. For example, the content with a high score addition appears in the recommendation list first. Update the content recommendation list according to the score addition to generate a live broadcast recommendation list that meets the user's interests.

[0177] The intelligent live broadcast recommendation system based on big data is used to execute the above-mentioned intelligent live broadcast recommendation method based on big data. The system includes:

[0178] The social interest recognition module calls the social platform API interface to obtain user text comments, picture sharing, and like behavior data, extracts keywords in the text comments and maps them to the preset live broadcast interest classification, and assigns an initial weight according to the like behavior frequency to obtain a dynamic interest tag record;

[0179] The weight dynamic update module performs a sliding window weighted sum on the interaction times of the same interest tag based on the timestamp sequence of the user ID in the dynamic interest tag record, calculates the tag activity index, sorts and adjusts after superimposing the initial weight to generate a time-series adjusted interest tag set;

[0180] The cross-platform fusion module calls the data of the content types watched, the single viewing duration, and the comment keywords corresponding to the user ID in the live broadcast platform behavior log, extracts the standard interest tags corresponding to the content types, calculates the cosine similarity between the standard tags and the tags in the time-series adjusted interest tag set to generate a cross-platform interest map;

[0181] The content matching module traverses the live broadcast titles and classification tags based on the cross-platform interest map, calculates the similarity score between the live broadcast content tags and the user interest tags, and corrects the score in combination with the like growth rate in the real-time interaction data to generate a real-time recommended content queue for users;

[0182] The social circle correction module utilizes the user's real-time recommended content queue, calls the user's social circle member list, extracts the weighted interest tags of the members in the cross-platform interest coefficient library, statistically analyzes the distribution ratio of the types of content recently viewed by the members, and performs a fixed ratio addition to the content matching scores of the same type of content in the content matching priority queue to generate a live broadcast recommendation list.

[0183] The above are only the preferred embodiments of the present invention, and do not limit the present invention in other forms. Any person skilled in the art may use the technical content disclosed above to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as the technical content of the present invention is not departed from, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. An intelligent live broadcast recommendation method based on big data, characterized in that, It includes the following steps: S1: Obtain the user's social interaction data, record the user's text comments, picture sharing, and like behavior, extract keywords from the text content and map them to the preset live broadcast interest classification, assign initial weights according to the like behavior frequency, and combine with the interest tags of the user ID to obtain the dynamic interest tag record; S2: Based on the timestamp sequence associated with the user ID in the dynamic interest tag record, calculate the sliding weighted value of the tag interaction times within the recent period, sort and adjust the tag weights, and generate a time-series adjusted interest tag set; S3: Call the data of the watched content type, single-watch duration, and comment keywords in the live broadcast platform user behavior, extract the standard interest classification corresponding to the content type, match the similarity between the standard classification and the time-series adjusted interest tag set, and assign live broadcast behavior weights to the successfully matched tags to generate a cross-platform interest map; S4: Use the tag list corresponding to the user ID in the cross-platform interest map to calculate the similarity score of the live broadcast content tags, correct the score in combination with the live broadcast real-time interaction data and sort, and generate a user real-time recommended content queue.

2. The intelligent live broadcast recommendation method based on big data according to claim 1, wherein The dynamic interest tag record includes keyword weights, social interaction types, and interest tag mapping relationships. The time-series adjusted interest tag set includes tag sliding weighted values, interest tag time decay rates, and tag weight sorting. The cross-platform interest map includes content type similarity, tag association degree, and live broadcast behavior weights. The user real-time recommended content queue includes tag real-time similarity, content real-time interaction scores, and recommended sorting levels.

3. The intelligent live broadcast recommendation method based on big data according to claim 1, wherein The specific steps for obtaining the dynamic interest tag record are as follows: S111: Obtain the user's social interaction data, record the user's text comment content, picture sharing annotation words, and like behavior, perform word segmentation on the comment text, extract keywords and perform category mapping with the set live broadcast interest classification table, and identify the term combinations that have a matching relationship with the interest tag library to obtain the user interest association category set; S112: According to the user interest association category set, call the user like behavior record, count the frequency of user likes in each type of interest-associated content, screen the interest categories whose frequency exceeds the like behavior interest bias threshold, and combine the number of user comment words and like frequencies in multiple categories to obtain the live broadcast content preference ranking sequence; S113: Based on the live broadcast content preference ranking sequence, call the live broadcast room tag information, screen the live broadcast room content that matches the previous item in the ranking sequence, combine and score the successfully matched live broadcast rooms according to the activity ratio and live broadcast room popularity, and adjust the priority of the recommendation results according to the score to obtain the dynamic interest tag record.

4. The intelligent live broadcast recommendation method based on big data according to claim 3, wherein, The specific steps for obtaining the time-series adjusted interest tag set are as follows: S211: Based on the timestamp sequence associated with the user ID in the dynamic interest tag record, extract the set of interaction time points of the tags, calculate the time difference between the time points and the moments, and generate a set of time decay factors; S212: Call the set of time decay factors, combine with the number of interactions of the tags within the period, and use the formula: Calculate the sliding weighted value of the tags, and generate a sliding weighted value sequence in descending order; Among them, W c represents the sliding weighted value of label c, λ t represents the time decay factor of time unit t, I t represents the number of interactions of the label at time unit t, σ c represents the dispersion of the number of interactions of label c, μ c represents the mean value of the number of interactions of label c, ν t represents the interaction frequency offset of time unit t, and T represents the set of all time units; S213: Call the sliding weighted value sequence, interpolate and fuse the real-time periodic label weights with the weight sequence, and filter the labels whose weights exceed the dynamic threshold according to the interpolation result to generate a time-series adjusted interest label set.

5. The intelligent live broadcast recommendation method based on big data according to claim 4, characterized in that, The specific steps for obtaining the cross-platform interest graph are as follows: S311: Call the viewing content type data of the live platform according to the time-series adjusted interest label set, match the keyword mapping and semantic information, extract the standard interest classification corresponding to each viewing record, and generate a classification mapping set; S312: Based on the classification mapping set, call the vectorized expression of the time-series adjusted interest label set, calculate the similarity between the vector of each standard interest classification and the label vectors in the label set, using the formula: Calculate the label matching degree; Among them, S ij represents the matching degree between the i-th content and the j-th tag, a i represents the single-viewing duration of the i-th content, k j represents the number of comment keyword matches for the j-th tag, c ib represents the occurrence frequency of the i-th content type in the b-th period, v j and v b represent the vector norms of tag j and tag b respectively, and n is the number of periods; S313: According to the label matching degree, superimpose the duration weight and comment keyword weight of the corresponding viewing content type on each matching label, calculate the total value, and filter the labels whose total weight exceeds the benchmark of the live platform to generate a cross-platform interest graph.

6. The intelligent live broadcast recommendation method based on big data according to claim 5, characterized in that The specific steps for obtaining the user's real-time recommended content queue are as follows: S411: Call the label list corresponding to the user ID in the cross-platform interest graph, calculate the cosine similarity between each pair of live content labels, and store the calculation results classified by user ID to generate an initial label similarity value; S412: Based on the initial label similarity value, obtain the like rate, comment rate, and viewing duration ratio in the live real-time interaction data, using the formula: Calculate the label corrected similarity, sum the multiple label correction values of each live content and take the average to generate an interaction corrected similarity; Among them, M op is the label correction similarity, AB op represents the initial label similarity value between user o and live content p, Q1 represents the like rate of live content p, Q2 represents the comment rate of live content p, and T represents the time difference decay factor; S413: Call the interaction corrected similarity, sort the values, and combine the reverse weighted value of the user's click behavior in the real-time interaction data to perform a secondary adjustment on the sorting result to generate a user's real-time recommended content queue.

7. The intelligent live broadcast recommendation method based on big data according to claim 1, wherein The method further includes step S5: S5: Extract the cross-platform interest graph labels corresponding to the user's social circle member list, count the distribution of the viewing content types of the members, add points to the content scores of the matching types in the user's real-time recommended content queue, and generate a live broadcast recommendation list; The live broadcast recommendation list includes the social circle label distribution weight, the content type matching score, and the social interest addition coefficient.

8. The intelligent live broadcast recommendation method based on big data according to claim 7, wherein The specific steps for obtaining the live broadcast recommendation list are as follows: S511: Use the user's real-time recommended content queue to extract the data of the user's social circle members, identify the user's interest preferences, analyze the content and interest manifestations of the user's interaction with the social circle members, and generate cross-platform interest labels for the user's social circle; S512: Based on the cross-platform interest labels of the user's social circle, count the content types viewed by each member on multiple live platforms, classify and summarize the viewing frequencies according to the time period, and combine the frequencies of the viewing types to obtain the distribution of the viewing content types of the members; S513: According to the distribution of the viewing content types of the members and the interest labels, combine the user's real-time recommended content queue, add points to the matching content types, identify the recommended priority of each content, and update the content recommendation list according to the score addition to generate a live broadcast recommendation list.

9. An intelligent live broadcast recommendation system based on big data, characterized in that, The system is used to implement the big data-based intelligent live broadcast recommendation method described in any one of claims 1-8, and the system includes: The social interest recognition module calls the social platform API interface to obtain user text comments, picture sharing, and like behavior data, extracts keywords from the text comments and maps them to the preset live broadcast interest classification, assigns an initial weight according to the like behavior frequency, and obtains a dynamic interest tag record; The weight dynamic update module performs a sliding window weighted sum on the interaction times of the same interest tag based on the timestamp sequence of the user ID in the dynamic interest tag record, calculates the tag activity index, sorts and adjusts after superimposing the initial weight, and generates a time series adjusted interest tag set; The cross-platform fusion module calls the viewing content type, single viewing duration, and comment keyword data corresponding to the user ID in the live broadcast platform behavior log, extracts the standard interest tags corresponding to the content type, calculates the cosine similarity between the standard tags and the tags in the time series adjusted interest tag set, and generates a cross-platform interest map; The content matching module traverses the live broadcast title and classification tags based on the cross-platform interest map, calculates the similarity score between the live broadcast content tags and the user interest tags, and corrects the score in combination with the like growth rate in the real-time interaction data to generate a user real-time recommended content queue; The social circle correction module uses the user real-time recommended content queue, calls the user social circle member list, extracts the weighted interest tags of the members in the cross-platform interest coefficient library, counts the distribution ratio of the recent viewing content types of the members, and performs a fixed ratio addition on the content matching scores of the same type of content in the content matching priority queue to generate a live broadcast recommendation list.

Images