A cloud big data real-time streaming method and system based on intelligent analysis

Abstract

The application provides a cloud big data real-time streaming method and system based on intelligent analysis, comprising: collecting real-time interaction behaviors of users, including short drama viewing time and member page interaction records, and formatting to obtain standardized behavior data sets; identifying viewing time and picture quality preferences through the standardized behavior data sets, determining user traffic consumption types according to the viewing time and the picture quality preferences, and obtaining user package adaptation types; adjusting the display order of the traffic card package on the member page according to the optimized package combination scheme, setting the package scheme at the top according to the click frequency, and obtaining a personalized package streaming scheme; obtaining the user package click conversion rate from the personalized package streaming scheme, and identifying the user preference change trend to determine the recommendation weight configuration by continuously integrating the newly collected viewing time.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a cloud-based real-time big data delivery method and system based on intelligent analysis. Background Technology

[0002] The operational model combining short video content consumption with data services is becoming increasingly important in the digital marketing field, especially in scenarios that combine short drama membership promotion with data card sales. This model drives a dual improvement in user stickiness and commercial conversion by accurately meeting users' content viewing and network needs. However, existing methods in this field often face significant technical challenges in dealing with diverse user needs and real-time decision-making, particularly in how to effectively utilize user behavior data to achieve precise recommendations. Existing methods typically recommend fixed data card packages based on users' historical consumption records. However, this approach struggles to adapt to dynamic changes in users' short drama viewing habits, such as shifting from occasional to frequent viewing, or from standard definition to high definition. The limitation of this static recommendation is its inability to capture the real-time evolution of user behavior patterns, leading to a disconnect between recommended data card packages and users' actual needs, thus reducing purchase conversion rates. Furthermore, in the promotion of short drama memberships and the sale of data cards, user behavior data during short drama viewing, such as viewing time, number of episodes, image quality preferences, and daily active times, is highly heterogeneous and dynamic. For example, a user might initially choose a small-capacity monthly data plan, but as they watch more high-definition short dramas, their data needs could quickly shift to a larger-capacity quarterly or annual plan. This dynamic change in demand necessitates real-time optimization of data plan displays. Traditional recommendation systems, often based on fixed rules or historical data, cannot dynamically adjust the displayed data plan combinations based on real-time interactive behaviors such as user dwell time and click frequency when viewing plans on the membership page. For instance, a user might repeatedly view the details page of a large-capacity data plan on the membership page but ultimately not purchase it, possibly because the plan's presentation or bundled options failed to effectively stimulate their purchase intention. This combination of real-time interactive data analysis and plan optimization requires the system to have rapid response and dynamic adjustment capabilities, but current technology's shortcomings in this area lead to a reduced match between recommendation effectiveness and user needs. Therefore, how to dynamically adjust the display strategy and bundled options of data plans based on real-time analysis of user short drama viewing behavior and membership page interaction data has become a key issue in the scenario of combining short video content consumption with data services. Summary of the Invention

[0003] This invention provides a cloud-based real-time big data delivery method based on intelligent analysis, mainly including:

[0004] By collecting real-time user interaction behavior and formatting it, a standardized behavior dataset is obtained. This dataset includes short drama viewing time and member page interaction records. The standardized behavior dataset is used to identify user viewing time and image quality preferences, and based on these preferences, the user's data plan suitability type is determined. From the user's data plan suitability type, member page dwell time and data plan card click frequency are extracted to obtain the member page browsing path. This is then matched with the data plan preferences of similar user groups to generate a preliminary data plan recommendation list. Based on the member page browsing path, user page navigation behavior is analyzed to identify user behavior patterns, and real-time interaction is evaluated based on these patterns. The interaction is used to match the user's package type, sort the initial package recommendation list, and determine the optimized package combination scheme. Based on the optimized package combination scheme, the display order of data SIM card packages on the member page is adjusted, and packages are prioritized based on click frequency to generate a personalized package traffic distribution scheme. The user package click-through rate is extracted from the personalized package traffic distribution scheme, and combined with newly collected viewing time to identify user preference trends and determine the recommendation weight configuration. Based on the recommendation weight configuration, behavior pattern recognition is reprocessed, and based on the matching degree between dynamic demand changes and real-time interactive behavior, the final data SIM card package recommendation result is output, including package type, data capacity, and price tier.

[0005] Furthermore, the process of collecting real-time user interaction behavior and formatting it to obtain a standardized behavior dataset includes:

[0006] The system acquires real-time operation records of users on the short drama player and interaction records on the membership page, extracts short drama viewing duration, number of pauses, video quality switching records, package card click coordinates, and page dwell time, and aligns them according to timestamps to form an original behavior sequence; abnormal records are removed from the original behavior sequence, the data format is converted, and the numerical range is unified to the 0 to 1 interval through a normalization method to generate the standardized behavior dataset.

[0007] Furthermore, the process of identifying user viewing time and image quality preferences through a standardized behavioral dataset, and determining the user's package suitability type based on viewing time and image quality preferences, includes:

[0008] Viewing duration sequences and image quality selection records are extracted from the standardized behavior dataset. The ratio of the daily cumulative viewing duration to the preset duration is calculated as a viewing density index. The percentage of standard definition, high definition, and ultra-high definition selections is statistically analyzed to form an image quality preference distribution. These are combined to form a user viewing behavior feature vector. Based on the user viewing behavior feature vector, the unit time data consumption is calculated by combining the image quality selection frequency and bitrate standard. Based on the viewing density index and image quality preference distribution, the high data consumption type, high-speed data consumption type, or low data consumption type is determined, and the corresponding package capacity is matched to generate the user package adaptation type.

[0009] Furthermore, the step of extracting the dwell time on the member page and the click frequency of the package card from the user package adaptation type, obtaining the member page browsing path, matching the data card package preferences of similar user groups, and generating a preliminary package recommendation list includes:

[0010] Extract the dwell time on the member page and the click frequency of the package card from the user package adaptation type, record the page jump path, and generate a user page interaction behavior dataset; calculate the package attention value based on the user page interaction behavior dataset, combine the historical package purchase records of similar user groups to statistically calculate the group preference value, calculate the comprehensive recommendation score, and sort to generate the preliminary package recommendation list.

[0011] Furthermore, the step of analyzing user page navigation behavior based on member page browsing paths, identifying user behavior patterns, evaluating the matching degree between real-time interaction behavior and user package suitability based on user behavior patterns, sorting the initial package recommendation list, and determining the optimized package combination scheme includes:

[0012] Extract page jump sequences from the member page browsing path, calculate interaction frequency indicators and repeated access counts of package cards, and generate a user page interaction feature set; determine high-frequency interaction patterns based on the user page interaction feature set, and record package access sequences and dwell times; calculate the overlap ratio between the real-time access package type and the user package adaptation type as a consistency score, assign package priority values ​​based on the consistency score, sort the preliminary package recommendation list, and generate the optimized package combination scheme containing the main and alternative packages.

[0013] Furthermore, the process of adjusting the display order of data plan packages on the member page according to the optimized package combination scheme, prioritizing packages based on click frequency, and generating personalized data traffic allocation schemes includes:

[0014] The recommendation priority value is extracted from the optimized package combination scheme, and the comprehensive display weight is calculated by combining the user's historical click count to determine the package display order; the page display parameters are configured according to the comprehensive display weight, and the display area and visual effects of the top-weighted package, medium-weighted package and low-weighted package are set to generate the personalized package traffic delivery scheme that includes position sorting and style.

[0015] Furthermore, the process of adjusting the display order of data plan packages on the member page according to the optimized package combination scheme, prioritizing packages based on click frequency, and generating personalized data traffic allocation schemes includes:

[0016] The recommended weight value and user matching score are extracted from the optimized package combination scheme. The visual prominence coefficient is calculated by combining the user's historical click count and dwell time to generate a package visual configuration parameter set. The package display area, color saturation and border style are adjusted according to the package visual configuration parameter set, the number of page grids is allocated, dynamic effects are added, and a personalized page display layout including position, size and style is generated.

[0017] Furthermore, the step of extracting user package click-through rates from personalized traffic delivery plans, combining this with newly collected viewing time to identify user preference trends, and determining recommendation weight configuration includes:

[0018] Click counts and purchase counts are extracted from the personalized package traffic delivery plan, the click-through rate is calculated, and a time-series dataset is formed by combining the latest viewing time data; based on the time-series dataset, preference change points are detected, change trends are statistically analyzed, the recommendation weights of high-traffic packages or low-traffic packages are adjusted, and the recommendation weight configuration is generated.

[0019] Furthermore, the behavior pattern recognition is reprocessed according to the recommendation weight configuration, and based on the matching degree between dynamic demand changes and real-time interaction behavior, the final data plan recommendation result is output, including plan type, data capacity, and price tier, including:

[0020] Based on the recommended weight configuration, the behavior pattern recognition threshold is updated by combining user click and purchase behavior data, and a new behavior pattern identifier is generated; dynamic demand features are extracted based on the behavior pattern identifier, the traffic consumption change rate and the number of visits to the package details page are calculated, and the matching degree value is determined; based on the matching degree value, the package type, traffic capacity value and price tier are extracted to generate the structured final data card package recommendation result.

[0021] This invention provides a cloud-based real-time big data streaming system based on intelligent analysis, mainly comprising:

[0022] The behavior data acquisition module is used to collect users' real-time interactive behaviors, including short drama viewing time and member page interaction records, and process them into a standardized behavior dataset.

[0023] The user preference recognition module is used to identify viewing duration and image quality preferences through a standardized behavioral dataset, determine the user's data consumption type based on viewing duration and image quality preferences, and obtain the user's package matching type.

[0024] The initial recommendation generation module is used to extract statistics on the duration of the member page and the frequency of clicks on the package card from the user's package adaptation type. At the same time, it obtains the member page browsing path, matches the data card package preferences of similar user groups, and obtains an initial package recommendation list.

[0025] The behavior pattern analysis module is used to analyze user page jump behavior based on the member page browsing path, identify the current user behavior pattern, and if the user behavior pattern is a high-frequency interaction, evaluate the matching degree between the real-time interaction behavior and the user package adaptation type, prioritize the preliminary package recommendation list, and determine the optimized package combination scheme.

[0026] The personalized traffic delivery module is used to adjust the display order of data SIM card packages on the member page based on the optimized package combination scheme, and to place the package scheme at the top based on the click frequency to obtain the personalized package traffic delivery scheme; the weight configuration update module is used to obtain the user package click conversion rate from the personalized package traffic delivery scheme, and to identify the trend of user preference changes and determine the recommendation weight configuration by continuously incorporating newly collected viewing time.

[0027] The dynamic recommendation output module is used to reprocess behavioral pattern recognition based on the configuration of recommendation weights and a feedback mechanism. If the matching degree between dynamic demand changes and real-time interactive behavior improves, the final data plan recommendation output is obtained, including plan type, data capacity, price tier, and recommendation reason.

[0028] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0029] This invention discloses a cloud-based big data real-time streaming method and system based on intelligent analysis. Addressing the problem of accurately matching user data consumption needs in short drama viewing scenarios, it constructs a standardized behavioral dataset by collecting user short drama viewing time, image quality preferences, and member page interaction data. This dataset analyzes user data consumption types and forms a package-adaptive classification system. By cross-analyzing viewing time density and image quality preferences, this invention identifies user sensitivity and demand intensity regarding data consumption, categorizing them into high-data, lightweight, and high-speed consumption types. Combining member page browsing paths and interaction frequency, it dynamically adjusts package recommendation weights and display order, optimizing page layout to highlight highly compatible packages. By continuously incorporating user click-through rates and preference trends, this invention uses a feedback mechanism to dynamically update recommendation weights, ultimately outputting a personalized recommendation scheme that includes package type, data capacity, price, and recommendation reasons. This invention significantly improves the accuracy of data plan recommendations and user conversion rates, optimizing resource allocation efficiency in short drama scenarios. Attached Figure Description

[0030] Figure 1 This is a flowchart of a cloud-based big data real-time streaming method based on intelligent analysis, according to the present invention.

[0031] Figure 2 This is a schematic diagram of the structure of a cloud-based big data real-time streaming system based on intelligent analysis according to the present invention. Detailed Implementation

[0032] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0033] like Figures 1 to 2 This embodiment of a cloud-based big data real-time streaming method and system based on intelligent analysis may specifically include:

[0034] S101. By collecting real-time user interaction behavior, including short drama viewing time and member page interaction records, a standardized behavior dataset is obtained through formatted processing.

[0035] The system acquires real-time user activity logs on the short drama player, extracting viewing duration, pause counts, and video quality switching records from the player logs. Simultaneously, it collects click coordinates and page dwell time for package cards from the membership page. These data are then aligned temporally based on timestamps from each data source to obtain a raw behavior sequence containing playback actions and page interactions. This raw behavior sequence undergoes data cleaning, removing abnormal records with buffering times exceeding a preset threshold. Various behavior data types are transformed according to JSON format specifications, and the numerical range is standardized to the 0-1 interval using a maximum-minimum value normalization method, thus determining a standardized behavior dataset.

[0036] In one embodiment, when acquiring real-time user interaction behavior, the short drama player log records include a Unix timestamp for each playback operation, video ID, viewing duration in seconds, pause count, and image quality switching records. The image quality switching records specify the exact moment of switching from standard definition 360P to high definition 720P or ultra-high definition 1080P. Click coordinates on the membership page's package cards record the user's touch position on the screen, and page dwell time is counted from page load until the user navigates to or exits the page. Time alignment uses a unified time base, converting timestamps from different data sources to millisecond-level timestamps in the same time zone, arranging them chronologically to form a continuous sequence of behaviors.

[0037] For example, when a user switches to a membership page to browse a package after watching a short drama for 10 minutes, the original behavior sequence records the viewing behavior and page interaction behavior in chronological order. During data cleaning, a preset threshold for buffering time is set to 30 seconds; records exceeding this threshold are marked as invalid data due to network anomalies. JSON format conversion maps various types of behavioral data into key-value pair structures, with viewing time mapped to a duration field and click coordinates mapped to a position field. Normalization uses a maximum-minimum method, mapping viewing time from 0 to 3600 seconds to a 0-1 interval to ensure comparability of data with different dimensions.

[0038] S102. Identify viewing duration and image quality preferences through standardized behavioral datasets, determine user data consumption type based on viewing duration and image quality preferences, and obtain user package adaptation type.

[0039] Viewing duration and image quality selection records are extracted from a standardized behavioral dataset. The cumulative daily viewing duration is divided by 24 hours to obtain a viewing density index. The percentage of times users select standard definition (SD), high definition (HD), and ultra-high definition (UHD) is statistically analyzed to determine the image quality preference distribution. The viewing density index and image quality preference distribution are combined to form a user viewing behavior feature vector. Data consumption is evaluated based on this user viewing behavior feature vector. According to the bitrate standards of 500kbps for SD, 2Mbps for HD, and 4Mbps for UHD, the frequency of image quality selection is multiplied by the corresponding bitrate to obtain the data consumption per unit time. If the viewing density index exceeds 0.3 and the percentage of UHD selections exceeds 0.6, it is judged as a high-data consumption type. If viewing is concentrated between 8-11 PM, it is judged as a high-speed data consumption type; otherwise, it is judged as a low-data consumption type. Data plan capacity is matched according to the data consumption type: high-data consumption type corresponds to a monthly plan of 30GB or more; high-speed data consumption type corresponds to a 20GB plan including off-peak data; and low-data consumption type corresponds to a 10GB basic plan, thus determining the user's plan suitability type.

[0040] In one embodiment, when extracting user viewing behavior features from a standardized behavior dataset, the viewing density metric reflects the degree of user immersion in the short drama content.

[0041] Specifically, the cumulative viewing time of users over 24 hours is obtained, divided by 24 to obtain the hourly average, which serves as the base value for viewing density. Then, a weighted adjustment is made based on the continuity of viewing periods: periods of continuous viewing exceeding 30 minutes are weighted at 1.2, and periods of intermittent viewing are weighted at 0.8, resulting in the viewing density index. The image quality selection record includes the time point and selected image quality level for each user's active switching. The proportion of each image quality level selected relative to the total number of views is calculated to form an image quality preference distribution.

[0042] It should be noted that the user viewing behavior feature vector consists of two parts: a viewing density index and a picture quality preference distribution. The viewing density index is a single value, while the picture quality preference distribution is a three-dimensional vector, corresponding to the selection ratios of standard definition, high definition, and ultra-high definition, respectively.

[0043] Preferably, the data consumption assessment process determines the bitrate corresponding to different image qualities based on the video encoding standard. Standard definition (SD) content uses H.264 encoding with a bitrate of approximately 500kbps, high definition (HD) content has a bitrate increased to 2Mbps, and ultra-high definition (UHD) content reaches 4Mbps. The average data consumption rate is obtained by multiplying the user's viewing time at each image quality by the corresponding bitrate, summing the results, and then dividing by the total viewing time. When the viewing density index exceeds 0.3, it indicates that the user watches for more than 7 hours per day on average. Simultaneously, if the UHD selection percentage exceeds 0.6, it indicates that the user prefers high-quality content. This is considered a high-data consumption type, with an estimated monthly data demand exceeding 30GB.

[0044] For example, for users whose viewing time is concentrated between 8 pm and 11 pm, although the total viewing time may not be long, the bandwidth requirements are high during the peak network period. This is judged as a high-speed data consumption type, and a 20GB package that includes off-peak data is suitable for recommendation, with no speed limit on off-peak data at night.

[0045] In one possible implementation, the viewing behavior of users with light data consumption is fragmented, with each viewing session typically lasting less than 15 minutes. They often choose standard definition to save data, and their monthly data usage is less than 10GB, which can be met by matching a basic data plan.

[0046] Short drama viewing time data and image quality selection records are extracted from standardized behavioral datasets. The density of user viewing time and image quality preference are analyzed. The viewing time and image quality selection are cross-analyzed to identify the user's sensitivity to data consumption and the intensity of demand. Based on the frequency of viewing behavior and the level of image quality requirements, user data usage behavior is categorized into high data consumption type, low data consumption type, and high data consumption type. The correspondence between user data consumption behavior and package requirements is established to form a user package adaptation type classification system.

[0047] Short drama viewing duration sequences and image quality selection records are extracted from a standardized behavioral dataset. The 24-hour period is divided into 24 time slots, and the number of views and cumulative duration within each time slot are statistically analyzed. A duration density matrix is ​​constructed with time slots as rows and view counts and durations as columns. Simultaneously, the selection frequency ratio and switching sequence of standard definition (SD), high definition (HD), and ultra-high definition (UHD) image quality levels are statistically analyzed to obtain a basic feature set of user viewing behavior. Cross-feature extraction is performed on this basic feature set. Peak periods with viewing counts exceeding the average are identified from the duration density matrix. The ratio of high-quality image quality selection frequency to the total selection frequency within the peak period is calculated as traffic sensitivity. The proportion of total HD and UHD viewing time to total viewing time is defined as a demand intensity index, resulting in a user traffic consumption feature vector. Based on the user traffic consumption feature vector, consumption behavior is classified. If the demand intensity index exceeds 0.7 and the average daily viewing time exceeds 180 minutes, it is classified as a high-traffic consumption type. If the peak time is concentrated in the evening and the video quality switching frequency exceeds 3 times per hour, it is classified as a high-speed traffic consumption type. If the average daily viewing time is less than 60 minutes and the standard definition selection ratio exceeds 0.5, it is classified as a light-traffic consumption type. This determines the user traffic usage behavior category. Based on the user traffic usage behavior categories, a package demand mapping is established. The high-traffic consumption type corresponds to a monthly package of 30GB or more, the high-speed traffic consumption type corresponds to a 20GB peak speed-up package, and the light-traffic consumption type corresponds to a 10GB basic package. Based on the mapping relationship, a user package adaptation type classification system containing consumption type identifiers and package specifications is generated.

[0048] In one embodiment, when constructing the duration density matrix, a day of 24 hours is divided into 24 independent time periods, with each time period corresponding to a row of the matrix.

[0049] Specifically, the first column of the matrix records the number of times a user opened the short drama player during that time period, and the second column records the cumulative viewing time during that time period, in minutes. If a user opened the player 3 times between 2 AM and 3 AM, accumulating 45 minutes of viewing time, then the value in the third row of the matrix is ​​[3, 45]. By accumulating data over 7 consecutive days, the average value for each time period is calculated to form a stable duration density matrix. The picture quality selection record includes the timestamp of each user's active picture quality switch, the picture quality level before the switch, and the picture quality level after the switch. Statistics show that users switched from standard definition to high definition in 35% of the total number of switches, from high definition to ultra-high definition in 45% of the time, and the remaining 20% ​​were operations that reduced the picture quality. This switching sequence reflects the user's gradual demand for picture quality.

[0050] It should be noted that the user viewing behavior feature set is a multi-dimensional data structure comprising three components: a duration density matrix, a picture quality selection frequency distribution, and a picture quality switching sequence. These features characterize users' viewing habits from different perspectives. During the cross-feature extraction process, peak periods are identified based on statistical analysis of the number of views column in the duration density matrix. The mean and standard deviation of the number of views for 24 time periods are calculated; when the number of views for a given time period exceeds the mean plus one time deviation, it is marked as a peak period.

[0051] For example, the number of views between 8 PM and 10 PM were 8 and 7 respectively, while the daily average number of views was 3, with a standard deviation of 2. Both of these time periods exceeded the threshold of 5 views and were therefore identified as peak periods. Traffic sensitivity calculation focuses on image quality selection behavior during peak periods. If users select HD or Ultra HD 80% of the total selections during peak periods, the traffic sensitivity value is 0.8, indicating that users have high image quality requirements during high-frequency viewing periods.

[0052] In one possible implementation, the demand intensity index comprehensively considers users' overall preference for high-definition content. The base value of demand intensity is obtained by adding the total viewing time of high-definition (HD) content and ultra-high-definition (UHD) content, and then dividing by the total viewing time. If a user watches 100 minutes of standard-definition (SD) content, 300 minutes of HD content, and 200 minutes of UHD content within a week, the demand intensity index is 500 divided by 600, which is 0.83, indicating that the user has a strong demand for high-definition content.

[0053] For example, the user traffic consumption feature vector is a two-dimensional vector containing traffic sensitivity and demand intensity indicators. These two dimensions jointly determine the user's traffic consumption type. High-traffic-consumption users exhibit consistently high-intensity viewing behavior. A demand intensity index exceeding 0.7 means that over 70% of their viewing time is spent in HD or UHD quality, with an average daily viewing time exceeding 180 minutes, equivalent to watching more than 6 episodes of short dramas daily. These users typically watch at multiple times, not limited to a specific time, exhibiting a 24 / 7 traffic demand. High-speed traffic-consumption users are characterized by concentrated viewing behavior during peak network hours, typically between 7 PM and 11 PM. During this period, network congestion is severe, and users frequently switch between different quality levels to seek a smooth viewing experience. Switching more than 3 times per hour indicates that users are more sensitive to network speed than to quality. Light-traffic-consumption users exhibit relatively conservative viewing behavior, with an average daily viewing time of less than 60 minutes, typically watching only 2 to 3 episodes of short dramas. A standard definition selection rate exceeding 50% indicates that users prioritize saving data rather than pursuing a high-quality experience.

[0054] For example, a user with a demand intensity index of 0.75 and an average daily viewing time of 200 minutes is classified as a high-data-consumption user according to the classification rules. A monthly plan of 30GB or more is recommended for them, which includes enough high-speed data to meet their high-definition viewing needs all day long.

[0055] In one embodiment, the mapping relationship between package demand takes into account the actual data usage characteristics of users with different consumption types. High-data-consumption users, based on an average of 15 minutes per episode of a short drama with ultra-high definition and a bitrate of 4Mbps, consume approximately 450MB of data per episode. Watching 12 episodes daily would require 5.4GB of data, exceeding 160GB per month. However, considering WiFi usage scenarios, the actual mobile network data demand is approximately 30GB. High-speed data-consumption users, while having lower total data demand than high-data-consumption users, have higher requirements for network quality during specific times. The 20GB peak speed-up package provides a dedicated bandwidth channel during evening peak hours to ensure a smooth viewing experience.

[0056] Understandably, the user package matching type classification system achieves accurate product recommendations by combining consumption type identifiers and package specifications, thereby improving the efficiency of users in choosing suitable packages.

[0057] S103. Extract the membership page dwell time and package card click frequency statistics from the user package adaptation type, and at the same time obtain the membership page browsing path, match the data card package preferences of similar user groups, and obtain a preliminary package recommendation list.

[0058] The user page interaction behavior dataset is obtained by extracting the dwell time and click frequency of package cards from the user's package matching type. The cumulative dwell time and click count on each package details page are recorded. Simultaneously, the order in which users access package pages and the navigation paths are collected to obtain a user page interaction behavior dataset. Based on this dataset, the attention index for each package is calculated. The dwell time is multiplied by the click frequency to obtain the package attention value. By comparing the interaction behavior characteristics of user groups with the same package matching type, user sets with similar behaviors are identified, and their historical package purchase records are extracted. Based on these historical purchase records, the frequency of each package selection is used as a group preference value. The package attention value is multiplied by 0.6, and the group preference value is multiplied by 0.4 to calculate a comprehensive recommendation score. A preliminary package recommendation list is generated by sorting the comprehensive recommendation scores from high to low.

[0059] In one embodiment, when extracting page interaction data from the user's package compatibility type, the dwell time is recorded as the complete time from when the user enters the package details page to when they leave, accurate to the second. Package card click frequency is counted as the number of times the user clicks on each package card on the member page, including actions such as viewing details, expanding descriptions, and comparing functions. Access sequence records the time series of the user browsing different packages, and the navigation path tracks the user's complete journey from the package list page to the details page and then to the purchase page.

[0060] It should be noted that the package attention metric reflects the degree of user interest in a specific package.

[0061] Specifically, if a user spends 120 seconds on the details page of a 30GB monthly plan and clicks 3 times, the attention value of that plan is 360, while for a 10GB basic plan, spending 30 seconds and clicking once results in an attention value of only 30. The difference clearly reflects user preferences.

[0062] Preferably, the process of identifying users with similar behaviors is based on multi-dimensional feature comparison.

[0063] In one possible implementation, users with the same package suitability are extracted as a candidate set, and their page dwell time distribution, click frequency patterns, and access path similarity are compared. Users are considered to have similar behavior when the difference in dwell time between two users is within 20%, the correlation coefficient of their click frequency patterns exceeds 0.7, and their access paths contain more than 60% of the same nodes. The historical package purchase records of similar users identified in this way are highly valuable for reference, as similar browsing behavior often foreshadows similar purchasing decisions.

[0064] For example, the statistics of historical package purchase records cover transaction data within the past 30 days. If there are 100 users in a similar user set, and 60 of them purchased a 20GB peak speed-up package, 30 purchased a 30GB monthly package, and 10 purchased a 10GB basic package, then the group preference values ​​are 0.6, 0.3, and 0.1, respectively. Furthermore, the calculation of the comprehensive recommendation score adopts a weighted summation method. The package attention value weight of 0.6 reflects the importance of individual user behavior, while the group preference value weight of 0.4 reflects the reference value of collective wisdom.

[0065] For example, if a certain package has an attention value of 300 and a group preference value of 0.4, then the comprehensive recommendation score is 300×0.6+0.4×0.4×1000, where 1000 is the normalization coefficient.

[0066] Understandably, the initial package recommendation list is sorted in descending order of overall recommendation score, with packages with higher scores displayed first, thus improving the efficiency for users to find a suitable package.

[0067] S104. Analyze user page navigation behavior based on the member page browsing path, identify the current user behavior pattern, and if the user behavior pattern is a high-frequency interaction, evaluate the matching degree between the real-time interaction behavior and the user package adaptation type, prioritize the preliminary package recommendation list, and determine the optimized package combination scheme.

[0068] The user's navigation sequence between page nodes is extracted based on the browsing path on the member page. The number of page switches per unit time is divided by the access duration to obtain the interaction frequency index. The number of repeated accesses to package cards is counted to obtain the user page interaction feature set. Behavioral pattern recognition is performed on the user page interaction feature set. If the interaction frequency index exceeds 3 times per minute and the number of repeated accesses to a single package exceeds 5 times, it is determined to be a high-frequency interaction pattern. The package sequence and dwell time accessed by the user under the high-frequency interaction pattern are recorded to obtain the current behavioral pattern identifier and its associated data. The matching degree is evaluated based on the current behavioral pattern identifier and its associated data. The number of overlaps between the package type accessed by the user in real time and the package matching type is counted. The number of overlaps is divided by the total number of accessed packages to obtain a consistency score. The consistency score is multiplied by 100 and assigned as a priority value to the corresponding package in the preliminary package recommendation list to obtain the package priority sequence. The preliminary package recommendation list is sorted in descending order according to the package priority sequence. Packages with a priority value of more than 60 are selected as the main package, and packages with a priority value between 40 and 60 are selected as alternative packages. An optimized package combination scheme including the main and alternative packages is formed.

[0069] In one embodiment, the member page browsing path is extracted by recording the user's navigation behavior between different page nodes.

[0070] Specifically, when a user clicks from the package list page to enter the 20GB package details page, then returns to the list page and enters the 30GB package details page, this jump sequence is recorded as "List → 20GB Details → List → 30GB Details". The interaction frequency index is calculated based on the user's page switching behavior within a unit of time. If a user completes 15 page switches within 5 minutes, the interaction frequency index is 3 times per minute. The package card repeat visit count counts the user's repeated viewing behavior of the same package details page. When a user returns multiple times in a session to view the details, pricing information, and preferential terms of the same package, each visit is counted as a repeat visit. By integrating the jump sequence, interaction frequency index, and repeat visit count, a complete user page interaction feature set is formed. This feature set includes the temporal information, frequency characteristics, and focus of user browsing behavior.

[0071] It's important to note that the core of behavioral pattern recognition lies in determining whether a user is actively selecting a plan. When the interaction frequency exceeds 3 times per minute, it indicates that the user is quickly browsing and comparing different plans; when a single plan is accessed more than 5 times, it suggests that the user has a strong interest in that plan but may be hesitant in making a decision.

[0072] Preferably, the determination of high-frequency interaction modes adopts a dual-condition verification mechanism.

[0073] In one possible implementation, it is necessary not only to satisfy the numerical conditions of interaction frequency and repeated visits, but also to verify whether these behaviors occur within a continuous time window. If a user's high-frequency interaction behaviors are concentrated within 10 minutes, it is determined to be a valid high-frequency interaction pattern; if the behaviors are scattered over a longer period of time, it is not determined to be a high-frequency interaction even if the numerical conditions are met. The current behavior pattern identifier includes a pattern type marker and timestamp information, while the associated data records the complete sequence of packages accessed by the user under this pattern, including the access order, dwell time, and page depth of each package.

[0074] For example, the matching evaluation process achieves accurate judgment by comparing the user's real-time behavior with the preset data plan compatibility type. Assume the user's data plan compatibility type is high-data consumption, and the corresponding recommended data plans include a 30GB monthly plan, a 50GB quarterly plan, and an unlimited annual plan. In real-time interaction, the user sequentially accessed five data plans: a 30GB monthly plan, a 20GB monthly plan, a 50GB quarterly plan, a 30GB monthly plan, and a 10GB basic plan. The 30GB monthly plan and the 50GB quarterly plan fall within the high-data consumption recommendation range, with two overlapping plans. The consistency score is calculated as 2 divided by 5, which equals 0.4. Multiplying the consistency score by 100 yields a priority value of 40, which is assigned to the 30GB monthly plan and the 50GB quarterly plan. For the 30GB monthly plan that the user accesses repeatedly, its priority value is further increased by 5 times the number of repeated accesses. If this plan is accessed twice, the final priority value is 40 plus 10, which equals 50. Furthermore, the formation of the package priority sequence not only considers the priority value of a single package, but also introduces the correlation analysis between packages.

[0075] In one embodiment, when multiple plans have similar priority values, the comparison relationship between plans is determined by analyzing the user's access path. If a user views a 30GB monthly plan and then immediately views a 50GB quarterly plan, it indicates that the user is weighing capacity and duration. These two plans are marked as a comparison group and kept adjacent in the final sorting. The optimized plan combination scheme adopts a hierarchical display strategy. The featured plans are those with a priority value exceeding 60. These plans are highly matched to user needs and are prominently displayed as large cards on the member page, including a complete plan description and a one-click purchase button. Alternative plans have a priority value between 40 and 60, indicating a certain match but not a complete fit with user needs. They are displayed as small cards below the featured plans, providing basic information and an entry point to view details. Plans with a priority value below 40 are categorized as other optional plans and collapsed at the bottom of the page.

[0076] For example, after behavioral analysis, a user's priority is assigned as follows: 30GB monthly plan (75 priority), 20GB peak speed-up plan (65 priority), 50GB quarterly plan (45 priority), and 10GB basic plan (25 priority). The plan combination scheme displays the 30GB monthly plan and 20GB peak speed-up plan as the primary recommended plans at the top, the 50GB quarterly plan as a secondary option, and the 10GB basic plan collapsed among other options. This dynamic optimization mechanism adjusts the recommendation strategy based on real-time user behavior. When a user shows a clear plan preference, it quickly responds and optimizes the display order, reducing the user's selection cost and improving plan purchase conversion rates.

[0077] S105. Adjust the display order of data SIM card packages on the member page according to the optimized package combination scheme, and place the package scheme at the top according to the click frequency to obtain a personalized data SIM card allocation scheme.

[0078] Based on the optimized package combination scheme, the recommendation priority value of each package is extracted. Click counts for each package are statistically analyzed from user historical behavior data. The click counts are normalized and added to the recommendation priority value to obtain a comprehensive display weight. The package display order is determined according to this comprehensive display weight, from highest to lowest. Page display parameters are configured for this package display order. The top three packages with the highest comprehensive display weight are set to the top display, allocated large cards, and given prominent visual effects. Packages ranked fourth to eighth use medium-sized cards, and the remaining packages use a compact list format, resulting in a layered package display layout. A display scheme is generated based on this package display layout configuration. The display area of ​​the top package is set to 1.5 times that of a standard card and configured with dynamic effects. Medium-weight packages maintain the standard card size, and low-weight packages only display basic information, forming a personalized package delivery scheme that includes position sorting, size allocation, and visual intensity.

[0079] In one embodiment, the calculation of the overall display weights uses a weighted fusion method to achieve dynamic adjustment.

[0080] Specifically, the recommendation priority value extracted from the optimized package combination scheme ranges from 0 to 100, reflecting the degree of matching between the package and user needs. Click normalization maps historical click data to a range of 0 to 1, with the package most clicked having a normalized value of 1, and unclicked packages having a normalized value of 0. The overall display weight is calculated by adding the recommendation priority value to the normalized click count multiplied by 20, ensuring that the contribution of historical behavior to the final weight is controlled within a reasonable range. The package display order follows a weight-priority principle, with packages with higher overall display weights receiving better display positions, increasing the probability of capturing user attention. The tiered display layout configuration is dynamically adjusted based on the distribution characteristics of the overall display weight.

[0081] In one possible implementation, when the top three packages all have a weight value exceeding 80, they are displayed side-by-side at the top; if only one package has a weight value exceeding 80, it is displayed at the top alone, with the remaining packages arranged in descending order of weight. The top-ranked package features a large card containing the package name, data allowance, validity period, key benefits, and limited-time offers, occupying over 90% of the page width. Medium-sized cards display basic package information and prices, occupying 70% of the page width. A compact list format displays only the package name and price in a single line for quick browsing.

[0082] For example, the dynamic effects of the featured package include a gradient background, micro-animations, and a highlighted border. When a user scrolls the page, the featured package remains in a fixed position for 3 seconds before scrolling with the page, ensuring full exposure. Setting the display area to 1.5 times that of a standard card means a 50% increase in height while keeping the width unchanged, reserving space for additional promotional information and user reviews. Furthermore, the personalized package delivery scheme updates dynamically based on the user's real-time behavior.

[0083] For example, if a user lingers on a certain package deal card for more than 5 seconds, the display weight of that package deal will temporarily increase by 10 points, and it will be automatically adjusted to a more prominent position on the next page refresh.

[0084] Understandably, this tiered display mechanism guides user attention through visual hierarchy, increasing the conversion rate of highly relevant packages while retaining display opportunities for other packages, thus achieving a balance between recommendation efficiency and choice diversity.

[0085] The recommended weight and user matching degree of each package are obtained from the optimized package combination scheme. The packages are arranged in the display position on the member page according to the recommended weight and user matching degree. The package with the highest matching degree is given priority to be displayed in the focus area of ​​the user's vision. At the same time, the visual prominence and display area of ​​the package card are adjusted according to the user's historical click behavior. By dynamically adjusting the order and visual weight of package display, users are guided to focus on the data card plan that best meets their needs, forming a personalized page display layout.

[0086] From the optimized package combination scheme, the recommendation weight value and user matching score of each package are obtained. The packages are arranged in descending order of matching score. The number of clicks and page dwell time of each package in the user's historical click behavior are extracted to obtain the basic dataset of package display. Visual parameters are calculated for the basic dataset of package display. The vertical position of the package is determined according to the matching score. The package with the highest matching score is positioned in the top third of the page as the focus area. The visual prominence coefficient is calculated by multiplying the number of clicks by 0.7 and adding the dwell time in seconds by 0.3, resulting in the set of visual configuration parameters for the packages. The display attributes are adjusted based on the set of visual configuration parameters for the packages. If the visual prominence coefficient exceeds a preset threshold, the display area of ​​the package card is increased to 1.2 to 1.5 times the standard area, the color saturation is increased by 20%, and the border is widened by 2 pixels. The left and right order of the packages in the same row is adjusted according to the recommendation weight value to obtain the adjusted display attribute configuration. The page layout is constructed based on the adjusted display attribute configuration. The page is divided horizontally using a 12-column grid. The number of grids occupied by each package is allocated according to the display area. Gradient backgrounds and entrance animations are added to packages with high visual prominence, forming a personalized page display layout that includes position, size, and style.

[0087] In one embodiment, the construction of the underlying dataset for the package presentation involves the integration and processing of multi-source data.

[0088] Specifically, the recommendation weight value is derived from the output of the preceding recommendation algorithm, ranging from 0 to 100. Values ​​above 80 indicate strong recommendations, 60-80 indicate moderate recommendations, and below 60 indicate alternative recommendations. The user matching score is calculated by comparing user profiles with package feature vectors. A cosine similarity algorithm is used to match user traffic needs, price sensitivity, usage time distribution, and other multi-dimensional features with package attributes such as capacity, cost, and promotional periods. Historical click behavior data includes the number of times users clicked on each package in the past 30 days, the duration of page dwell after each click, whether they viewed the details page, and whether they added packages to the comparison list. This data undergoes time decay processing, with recent behaviors given higher weight, forming a comprehensive basic dataset for package display. The determination of the gaze focus area is based on the golden ratio principle from eye-tracking research. The top third of the page is the natural focal point of the user's gaze, with an information acquisition rate of over 75% in this area.

[0089] Preferably, the visual prominence coefficient is calculated using a weighted fusion method, which multiplies the number of clicks by a weight coefficient of 0.7 and the dwell time in seconds by a weight coefficient of 0.3, and then adds the two together to obtain the initial visual prominence coefficient.

[0090] In one possible implementation, if a user clicks on a package 10 times with an average dwell time of 60 seconds, the visual prominence coefficient is calculated as 10 × 0.7 + 60 × 0.3. To avoid the influence of extreme values, the calculation results are normalized, mapping the visual prominence coefficients of all packages to a standard range of 0 to 100. When the visual prominence coefficient exceeds 70, the package is considered a high-attention package; above 50, it is considered a medium-attention package; and below 50, it is considered a low-attention package. This tiered mechanism ensures that packages with different levels of attention receive differentiated visual presentations. High-attention packages attract user attention through enhanced visual effects, while low-attention packages are displayed in a concise format to avoid information overload.

[0091] For example, the adjustment of display attributes involves coordinated changes across multiple visual dimensions. The adjustment range for display area is set between 1.2 and 1.5 times the standard area, with 1.2 times applicable to medium-attention packages and 1.5 times for high-attention packages. Increased color saturation is achieved by adjusting the S component of the HSL color space, increasing the original saturation value by 20% to enhance the visual impact of the package card. The border width is increased from the standard 1 pixel to 3 pixels, using a gradient border that transitions from the package theme color to white, creating a soft visual boundary. Furthermore, the recommendation weight values ​​are sorted in descending order from left to right within the same row, conforming to user reading habits. When the recommendation weight values ​​of multiple packages are similar, secondary ranking factors are introduced, including the package's cost-effectiveness index and the probability of users historically purchasing similar packages.

[0092] In one embodiment, a 12-column grid system evenly divides the horizontal space of the page, with each column being 8.33% of the total page width. High-attention packages occupy 6 to 8 columns to ensure ample display space; medium-attention packages occupy 4 to 5 columns to maintain a moderate information density; and low-attention packages occupy 2 to 3 columns for a compact presentation. The responsive design of the grid system automatically adjusts column widths across different screen sizes, compressing the 12 columns to 4 on mobile devices and adjusting to 8 on tablets to ensure a consistent experience across devices.

[0093] For example, the gradient background uses a linear gradient, transitioning from 10% transparency of the package theme color to complete transparency, with the gradient angle set to 135 degrees, creating a visual flow from the upper left to the lower right.

[0094] Understandably, the entrance animation design takes into account the user's cognitive load, employing a gradual loading strategy. The high-attention package features an animation duration of 0.5 seconds, including a fade-in effect and a slight upward movement of 20 pixels. The medium-attention package's animation time is shortened to 0.3 seconds, containing only a fade-in effect. The low-attention package has no animation and is displayed directly. The animation timing uses staggered triggering, with a 0.1-second delay between each package, creating a wave-like visual rhythm.

[0095] Specifically, the personalized page layout comprises three layers: a visual focus layer showcasing 1 to 3 highly relevant packages, occupying a prime position at the top of the page; a transition layer displaying 4 to 6 moderately relevant packages, providing supplementary options; and a background layer displaying the remaining packages in a list format, maintaining information completeness. This layered layout guides the user's decision-making path through visual hierarchy, improving the efficiency and accuracy of package selection.

[0096] S106. Obtain the user's package click-through rate from the personalized package traffic plan, and identify the trend of user preference changes by continuously incorporating newly collected viewing time to determine the recommendation weight configuration.

[0097] The system extracts user click and purchase counts for each personalized data package from the user-defined traffic delivery plan. The click-through rate (CTR) for each package is calculated by dividing the purchase count by the click count. Simultaneously, it acquires the user's short drama viewing time data for the latest week and appends this new data to the end of the historical viewing records to form a time-series dataset. Preference change detection is performed on this time-series dataset. A 7-day sliding window is used, and the difference between the average daily viewing time in the current window and the previous window is calculated. If the absolute value of the difference exceeds 30 minutes, it is marked as a preference change point. The number of preference change points within 30 days is counted. When the number exceeds 3 and all differences are positive, it is considered an upward trend; when all differences are negative, it is considered a downward trend. Recommendation weights are adjusted based on preference change trends and CTRs. If there is an upward trend, the recommendation weight of the high-traffic package is multiplied by an enhancement factor of 1.2; if there is a downward trend, the recommendation weight of the lightweight package is multiplied by an enhancement factor of 1.3. Packages with a CTR below 0.1 have their recommendation weights multiplied by a decay factor of 0.8 to determine the updated recommendation weight configuration.

[0098] In one embodiment, the click conversion rate is calculated based on the user's actual purchase behavior.

[0099] Specifically, the total number of clicks for each personalized package during the display period is extracted from the execution logs of the personalized package delivery plan, and the purchase records for the corresponding packages are retrieved from the order database. If a 30GB monthly package is clicked 50 times and results in 5 purchases within a week, the click-through rate (CTR) for that package is 0.1. The CTR directly reflects the effectiveness of the package recommendation and becomes an important basis for subsequent weight adjustments. The time-series dataset is constructed using an incremental update method, retaining users' historical viewing records from the past 60 days, with newly added viewing time data appended to the end of the sequence each day to form a continuous time series. During the implementation of the sliding window method, the 7-day window size is selected based on the periodic characteristics of user behavior.

[0100] In one possible implementation, the first window covers data from days 1 to 7, calculating the average daily viewing time as 120 minutes; the second window covers days 2 to 8, with an average daily viewing time of 135 minutes, and the difference between the two windows is 15 minutes. When the sliding step is set to 1 day, a new difference data point is generated each day. The 30-minute threshold takes into account the normal fluctuation range of user viewing habits; exceeding this threshold indicates a substantial change in the user's viewing behavior.

[0101] For example, the judgment of preference change trends follows the principle of continuity. When four preference change points appear within a 30-day observation period, with three differences of positive 40 minutes and one difference of positive 25 minutes, it is judged as a clear upward trend, indicating that users' demand for short drama content is increasing. Conversely, if most differences are negative, it is judged as a downward trend. Furthermore, the adjustment of recommendation weights adopts a differentiated coefficient strategy. The 1.2 enhancement coefficient corresponding to an upward trend increases the recommendation weight of high-data-volume packages from the original 80 to 96, improving the priority of such packages in the recommendation list. The 1.3 enhancement coefficient triggered by a downward trend strengthens the recommendation of lightweight packages, meeting users' needs for reduced data consumption. Packages with a click-through conversion rate below 0.1 are considered unattractive, and a decay coefficient of 0.8 reduces their display frequency, freeing up display space for high-conversion packages.

[0102] Understandably, this dynamic adjustment mechanism enables real-time synchronization between recommendation strategies and user behavior, improving the accuracy and conversion efficiency of package recommendations.

[0103] S107. For the recommendation weight configuration, a feedback mechanism is used to reprocess the behavior pattern recognition. If the matching degree between dynamic demand changes and real-time interactive behavior improves, the final data plan recommendation output is obtained, including plan type, data capacity, price tier, and recommendation reason.

[0104] A feedback mechanism is employed for the recommendation weight configuration. User click and purchase behavior data for recommended packages are acquired, and the actual purchased packages are compared with the recommended packages. The recommendation accuracy is calculated as a deviation value. This deviation value is multiplied by an adjustment coefficient to update the threshold parameters for behavior pattern recognition. User behavior pattern recognition is then re-executed to obtain the adjusted behavior pattern identifier. Dynamic demand features are extracted based on the adjusted behavior pattern identifier. The daily average data consumption change rate is calculated from the user's data usage records over the past 7 days. The number of times the user views the details pages of different packages is statistically analyzed to obtain the user's current page browsing and click behavior sequence. The cosine similarity between the dynamic demand feature vector and the behavior sequence vector is calculated as the matching degree value. If the matching degree value exceeds a preset threshold of 0.7, package information corresponding to the user's behavior pattern is extracted from the package library, including package type, data capacity, and price tier. A recommendation reason description is generated based on the user's viewing time growth trend and data usage characteristics. If the matching degree value does not exceed the threshold, user behavior data is collected again. The package information and recommendation reason description are integrated, and the final data SIM card package recommendation result is output in a structured format according to package type, data capacity, price tier, and recommendation reason.

[0105] In one embodiment, the feedback mechanism employs a closed-loop control principle, continuously monitoring user responses to recommendations to optimize the recommendation strategy. After a user receives a package recommendation, their behavioral feedback within 24 hours is recorded, including whether they clicked to view details, added it to the comparison list, ultimately purchased it, and whether they purchased the recommended package or another package. Recommendation accuracy is calculated based on a multi-level scoring mechanism: if a user purchases the top-ranked recommended package, the accuracy is recorded as 1.0; if they purchase the second-ranked package, the accuracy is recorded as 0.8; if they purchase the third-ranked package, the accuracy is recorded as 0.6; and if they purchase a non-recommended package, the accuracy is recorded as 0. This gradient scoring method more accurately reflects recommendation quality, rather than a simple binary judgment. The behavioral pattern recognition parameters are updated using an adaptive adjustment strategy. The adjustment coefficient is dynamically set according to different ranges of the recommended accuracy: when the accuracy is below 0.3, the adjustment coefficient is set to 1.5, indicating that a large correction is needed; when the accuracy is between 0.3 and 0.7, the adjustment coefficient is 1.2, and moderate adjustment is made; when the accuracy is above 0.7, the adjustment coefficient is 1.05, and only a fine adjustment is made.

[0106] Preferably, the extraction of dynamic demand features covers multiple dimensions of user behavior changes.

[0107] In one possible implementation, the rate of change in data consumption is calculated using a moving average method, comparing the average daily data consumption over the past 7 days with the average daily data consumption over the previous 7 days to obtain the percentage increase or decrease. Changes in package viewing frequency are identified by statistically analyzing the distribution of the number of times users view package details pages at different times. If the number of views increases from an average of 2 times to 5 times in the evening, it indicates increased user interest in the package. The construction of behavioral sequences records the user's page access path, including the complete trajectory from entering the member page from the homepage, browsing the package list, clicking on a specific package to view details, and returning to the list to continue browsing. These multi-dimensional features collectively constitute a panoramic profile of the user's current needs, providing a rich data foundation for subsequent matching degree calculations.

[0108] For example, cosine similarity calculation maps dynamic demand feature vectors and behavioral sequence vectors to the same vector space. The dynamic demand feature vector includes dimensions such as traffic growth rate, peak usage percentage, and number of times large-capacity plans are viewed, with each dimension assigned different weights based on its importance. The behavioral sequence vector converts the user's action sequence into a numerical representation; for example, the sequence "view 30GB plan → view 50GB plan → return to 30GB plan" is encoded as a preference strength value for large-capacity plans. The matching threshold of 0.7 is set based on statistical analysis of historical data. When the matching degree exceeds this threshold, the recommended plan purchase conversion rate reaches over 35%, demonstrating high commercial value.

[0109] In one embodiment, the recommendation reason is generated using a combination of templates and dynamic content. The basic template includes guiding phrases such as "based on your viewing habits" and "taking into account your data usage needs," while the dynamic content is populated based on the user's specific characteristics.

[0110] For example, if a user's viewing time shows an upward trend and they frequently watch high-definition content, the recommendation reason might be described as, "Your recent short drama viewing time has increased by 40%, and high-definition viewing accounts for 70% of your viewing. This high-data-volume package is recommended to meet your viewing needs." When the matching score does not exceed the threshold, a data re-collection mechanism is triggered instead of forced recommendation. This design avoids the negative impact of low-quality recommendations on user experience.

[0111] Specifically, the structured output of the final recommendation results adopts a standardized format. The package type field includes classification identifiers for monthly, quarterly, and annual cards; data usage is represented as an integer in GB; price tiers are divided into three levels: Economy, Standard, and Deluxe; and recommendation reasons are limited to 100 characters to ensure conciseness and readability. This structured output format facilitates front-end display and user understanding, improving the efficiency of conveying recommendation information.

[0112] For example, a complete recommendation output includes complete information such as "30GB monthly card, 30, standard type, based on your usage habit of watching 3 hours of short dramas per day and preferring high-definition picture quality, this package can meet your monthly data needs", thus realizing a closed loop of personalized recommendations.

[0113] This invention provides a cloud-based real-time big data streaming system based on intelligent analysis, mainly comprising:

[0114] The behavior data acquisition module is used to collect users' real-time interactive behaviors, including short drama viewing time and member page interaction records, and process them into a standardized behavior dataset.

[0115] The user preference recognition module is used to identify viewing duration and image quality preferences through a standardized behavioral dataset, determine the user's data consumption type based on viewing duration and image quality preferences, and obtain the user's package matching type.

[0116] The initial recommendation generation module is used to extract statistics on the duration of the member page and the frequency of clicks on the package card from the user's package adaptation type. At the same time, it obtains the member page browsing path, matches the data card package preferences of similar user groups, and obtains an initial package recommendation list.

[0117] The behavior pattern analysis module is used to analyze user page jump behavior based on the member page browsing path, identify the current user behavior pattern, and if the user behavior pattern is a high-frequency interaction, evaluate the matching degree between the real-time interaction behavior and the user package adaptation type, prioritize the preliminary package recommendation list, and determine the optimized package combination scheme.

[0118] The personalized data traffic allocation module is used to adjust the display order of data SIM card packages on the member page based on the optimized package combination scheme, and to place the package scheme at the top based on the click frequency to obtain a personalized data traffic allocation scheme;

[0119] The weight configuration update module is used to obtain the user package click conversion rate from the personalized package traffic plan, and to identify the trend of user preference changes by continuously incorporating newly collected viewing time to determine the recommendation weight configuration.

[0120] The dynamic recommendation output module is used to reprocess behavioral pattern recognition based on the configuration of recommendation weights and a feedback mechanism. If the matching degree between dynamic demand changes and real-time interactive behavior improves, the final data plan recommendation output is obtained, including plan type, data capacity, price tier, and recommendation reason.

[0121] The above description is merely a specific implementation of this specification. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the scope of protection of this specification is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this specification, and these modifications or substitutions should all be covered within the scope of protection of this specification.

Claims

1. A cloud-based real-time big data delivery method based on intelligent analysis, characterized in that, The method includes: By collecting real-time user interaction behavior and formatting it, a standardized behavior dataset is obtained. This dataset includes short drama viewing time and member page interaction records. The standardized behavior dataset is used to identify user viewing time and image quality preferences, and based on these preferences, the user's data plan suitability type is determined. From the user's data plan suitability type, member page dwell time and data plan card click frequency are extracted to obtain the member page browsing path. This is then matched with the data plan preferences of similar user groups to generate a preliminary data plan recommendation list. Based on the member page browsing path, user page navigation behavior is analyzed to identify user behavior patterns, and real-time interaction is evaluated based on these patterns. The interaction is used to match the user's package type, sort the initial package recommendation list, and determine the optimized package combination scheme. Based on the optimized package combination scheme, the display order of data SIM card packages on the member page is adjusted, and packages are prioritized based on click frequency to generate a personalized package traffic distribution scheme. The user package click-through rate is extracted from the personalized package traffic distribution scheme, and combined with newly collected viewing time to identify user preference trends and determine the recommendation weight configuration. Based on the recommendation weight configuration, behavior pattern recognition is reprocessed, and based on the matching degree between dynamic demand changes and real-time interactive behavior, the final data SIM card package recommendation result is output, including package type, data capacity, and price tier.

2. The cloud-based big data real-time streaming method based on intelligent analysis according to claim 1, characterized in that, The process of collecting real-time user interaction behavior and formatting it to obtain a standardized behavior dataset includes: The system acquires real-time operation records of users on the short drama player and interaction records on the membership page, extracts short drama viewing duration, number of pauses, video quality switching records, package card click coordinates, and page dwell time, and aligns them according to timestamps to form an original behavior sequence; abnormal records are removed from the original behavior sequence, the data format is converted, and the numerical range is unified to the 0 to 1 interval through a normalization method to generate the standardized behavior dataset.

3. The cloud-based big data real-time streaming method based on intelligent analysis according to claim 1, characterized in that, The process of identifying user viewing time and image quality preferences using a standardized behavioral dataset, and determining the user's plan suitability type based on viewing time and image quality preferences, includes: Viewing duration sequences and image quality selection records are extracted from the standardized behavior dataset. The ratio of the daily cumulative viewing duration to the preset duration is calculated as a viewing density index. The percentage of standard definition, high definition, and ultra-high definition selections is statistically analyzed to form an image quality preference distribution. These are combined to form a user viewing behavior feature vector. Based on the user viewing behavior feature vector, the unit time data consumption is calculated by combining the image quality selection frequency and bitrate standard. Based on the viewing density index and image quality preference distribution, the high data consumption type, high-speed data consumption type, or low data consumption type is determined, and the corresponding package capacity is matched to generate the user package adaptation type.

4. The cloud-based big data real-time streaming method based on intelligent analysis according to claim 1, characterized in that, The process involves extracting the dwell time on the member page and the click frequency of the package card from the user's package adaptation type, obtaining the member page browsing path, matching the data card package preferences of similar user groups, and generating a preliminary package recommendation list, including: Extract the dwell time on the member page and the click frequency of the package card from the user package adaptation type, record the page jump path, and generate a user page interaction behavior dataset; calculate the package attention value based on the user page interaction behavior dataset, combine the historical package purchase records of similar user groups to statistically calculate the group preference value, calculate the comprehensive recommendation score, and sort to generate the preliminary package recommendation list.

5. The cloud-based big data real-time streaming method based on intelligent analysis according to claim 1, characterized in that, The process of analyzing user page navigation behavior based on member page browsing paths, identifying user behavior patterns, evaluating the matching degree between real-time interaction behavior and user package suitability based on user behavior patterns, sorting the initial package recommendation list, and determining the optimized package combination scheme includes: Extract page jump sequences from the member page browsing path, calculate interaction frequency indicators and repeated access counts of package cards, and generate a user page interaction feature set; determine high-frequency interaction patterns based on the user page interaction feature set, and record package access sequences and dwell times; calculate the overlap ratio between the real-time access package type and the user package adaptation type as a consistency score, assign package priority values ​​based on the consistency score, sort the preliminary package recommendation list, and generate the optimized package combination scheme containing the main and alternative packages.

6. The cloud-based big data real-time streaming method based on intelligent analysis according to claim 1, characterized in that, The process of adjusting the display order of data plan packages on the member page according to the optimized package combination scheme, prioritizing packages based on click frequency, and generating personalized data traffic allocation schemes includes: The recommendation priority value is extracted from the optimized package combination scheme, and the comprehensive display weight is calculated by combining the user's historical click count to determine the package display order; the page display parameters are configured according to the comprehensive display weight, and the display area and visual effects of the top-weighted package, medium-weighted package and low-weighted package are set to generate the personalized package traffic delivery scheme that includes position sorting and style.

7. The cloud-based big data real-time streaming method based on intelligent analysis according to claim 1, characterized in that, The process of adjusting the display order of data plan packages on the member page according to the optimized package combination scheme, prioritizing packages based on click frequency, and generating personalized data traffic allocation schemes includes: The recommended weight value and user matching score are extracted from the optimized package combination scheme. The visual prominence coefficient is calculated by combining the user's historical click count and dwell time to generate a package visual configuration parameter set. The package display area, color saturation and border style are adjusted according to the package visual configuration parameter set, the number of page grids is allocated, dynamic effects are added, and a personalized page display layout including position, size and style is generated.

8. The cloud-based big data real-time streaming method based on intelligent analysis according to claim 1, characterized in that, The process of extracting user subscription click-through rates from personalized subscription plans, combining this with newly collected viewing time data to identify trends in user preferences, and determining recommendation weight configurations includes: Click counts and purchase counts are extracted from the personalized package traffic delivery plan, the click-through rate is calculated, and a time-series dataset is formed by combining the latest viewing time data; based on the time-series dataset, preference change points are detected, change trends are statistically analyzed, the recommendation weights of high-traffic packages or low-traffic packages are adjusted, and the recommendation weight configuration is generated.

9. The cloud-based big data real-time streaming method based on intelligent analysis according to claim 1, characterized in that, The process of reprocessing behavioral pattern recognition based on recommendation weights, and outputting the final data plan recommendation result based on the matching degree between dynamic demand changes and real-time interaction behavior, includes plan type, data capacity, and price tier. Based on the recommended weight configuration, the behavior pattern recognition threshold is updated by combining user click and purchase behavior data, and a new behavior pattern identifier is generated; dynamic demand features are extracted based on the behavior pattern identifier, the traffic consumption change rate and the number of visits to the package details page are calculated, and the matching degree value is determined; based on the matching degree value, the package type, traffic capacity value and price tier are extracted to generate the structured final data card package recommendation result.

10. A cloud-based real-time big data streaming system based on intelligent analysis, characterized in that, The system includes: The behavior data acquisition module is used to collect users' real-time interactive behaviors, including short drama viewing time and member page interaction records, and process them into a standardized behavior dataset. The user preference recognition module is used to identify viewing duration and image quality preferences through a standardized behavioral dataset, determine the user's data consumption type based on viewing duration and image quality preferences, and obtain the user's package matching type. The initial recommendation generation module is used to extract statistics on the duration of the member page and the frequency of clicks on the package card from the user's package adaptation type. At the same time, it obtains the member page browsing path, matches the data card package preferences of similar user groups, and obtains an initial package recommendation list. The behavior pattern analysis module is used to analyze user page jump behavior based on the member page browsing path, identify the current user behavior pattern, and if the user behavior pattern is a high-frequency interaction, evaluate the matching degree between the real-time interaction behavior and the user package adaptation type, prioritize the preliminary package recommendation list, and determine the optimized package combination scheme. The personalized data traffic allocation module is used to adjust the display order of data SIM card packages on the member page based on the optimized package combination scheme, and to place the package scheme at the top based on the click frequency to obtain a personalized data traffic allocation scheme; The weight configuration update module is used to obtain the user package click conversion rate from the personalized package traffic plan, and to identify the trend of user preference changes by continuously incorporating newly collected viewing time to determine the recommendation weight configuration. The dynamic recommendation output module is used to reprocess behavioral pattern recognition based on the configuration of recommendation weights and a feedback mechanism. If the matching degree between dynamic demand changes and real-time interactive behavior improves, the final data plan recommendation output is obtained, including plan type, data capacity, price tier, and recommendation reason.

Images