Intelligent television personalized content recommendation method based on deep learning

Abstract

The application relates to the technical field of intelligent television application, in particular to an intelligent television personalized content recommendation method based on deep learning; the method comprises data collection, interaction, storage, recommendation processing, service grading, resource arrangement and state agent modules; the system collects multi-dimensional state data such as terminals, users and content, and carries out pretreatment to obtain corresponding features; the system inputs the features into an analysis model to extract long-term and current conversation preferences, determines instant intention strength in combination with identity recognition credibility, and defines task service grades; when preset conditions are met, the system will perform recommendation calculation and pre-fetching according to the service grades, or adaptively control the recommendation task; the application solves the defect that a unified portrait is easy to deviate from a target in a sharing scene, can effectively distinguish between identity switching and instant intention, and improves recommendation accuracy.

Description

Technical Field

[0001] This invention relates to the field of smart TV application technology, specifically a method for personalized content recommendation on smart TVs based on deep learning. Background Technology

[0002] In the current smart TV recommendation scenario shared in the home, the control terminal needs to collect and process multi-dimensional heterogeneous time-series data such as playback status, underlying resources, and user behavior in real time in order to generate personalized recommendation tasks.

[0003] To achieve content recommendation, existing solutions generally rely on a unified long-term user profile and execute a fixed recommendation calculation and media resource pre-fetching process. Although this solution has a certain processing capacity when there is a single user and sufficient terminal performance, it does not distinguish between family member identity switching and real-time intent, and the underlying hardware clock domain is not unified and requires complex synchronization processing, resulting in a lengthy calculation process and recommendation results that are prone to deviating from the current session goal.

[0004] Meanwhile, the fixed computing mechanism and video decoding and network buffer directly compete for intelligent processor and bandwidth resources. The lack of dynamic hierarchical and degradation control based on resource status and playback risk will lead to playback stuttering and response time exceeding the preset threshold, reducing the stability of multiple concurrent tasks.

[0005] Therefore, how to improve the real-time performance of multi-source time-series data processing in a home-shared viewing environment, while ensuring the continuity of video playback and the accuracy of intent recommendation, has become an urgent technical problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a personalized content recommendation method for smart TVs based on deep learning, addressing the following technical problems: Existing smart TV recommendation technologies have significant shortcomings in intent recognition when multiple users share content and in recommendation degradation control based on terminal resources. There is an urgent need to propose a personalized content recommendation method that can more accurately handle the correlation between user preferences and playback status and dynamically optimize the recommendation task. The purpose of this invention can be achieved through the following technical solutions:

[0007] A control terminal for smart TVs, comprising a data acquisition module, an interaction module, a storage module, a recommendation processing module, a service hierarchy module, a resource orchestration module, and a state proxy module, including:

[0008] The data acquisition module collects playback status, terminal resource status, user behavior status, user identity status, content status, network status, and viewing environment status according to a preset sampling period; it performs time alignment, anomaly removal, normalization, and short-term statistics on the collected status data to obtain playback features, user features, content features, network features, and viewing environment features.

[0009] The user characteristics, content characteristics, playback characteristics, and network characteristics are input into a feature analysis model to obtain long-term user preference characteristics, current session preference characteristics, and updated content characteristic data; playback risk indicators are determined based on the playback characteristics; the intensity of immediate intent is determined based on the user characteristics and user behavior status; resource availability is determined based on the terminal resource status; and content timeliness is determined based on the content status.

[0010] The service level of the recommendation task is determined according to the priority order of playback continuity protection service, instant intent response service, regular personalized recommendation service and background update service; according to the service level and the timeliness of the content, recommendation calculation, candidate content sorting, content prefetching, cache loading or background update are performed, and the sorting results of candidate content are filtered to generate recommendation results; according to the playback characteristics, the terminal resource status and the network status, the recommendation task is paused, downgraded or resumed when preset conditions are met.

[0011] Preferably, the playback status includes the remaining duration of the playback buffer, the video decoder occupancy rate, and the current video bitrate, and the network status includes the effective network throughput and the network packet loss rate;

[0012] The buffer risk value is determined based on the relationship between the remaining playback buffer duration and the low buffer threshold; the decoding risk value is determined based on the relationship between the video decoder occupancy rate and the decoding occupancy threshold; the bitrate risk value is determined based on the matching relationship between the current video bitrate and the effective network throughput; and the packet loss risk value is determined based on the relationship between the network packet loss rate and the packet loss threshold. Each risk value is then normalized and truncated to [a specific value]. The playback risk index is determined by weighting according to preset non-negative weights.

[0013] Preferably, the interaction module collects voice operations, remote control operations, page dwell times, content operations, and the correspondence between the currently playing content and the operation target.

[0014] The degree of voice matching is determined based on the similarity between the voice-recognized text and the content title, content category, or search target. The concentration of remote control behavior is determined based on the ratio of the number of operations on the same channel, the same content category, or the same function within a preset time window to the total number of remote control operations. The degree of dwell preference is determined based on the ratio of the page dwell time to the average dwell time within the preset time window. The consistency of content is determined based on the matching result between the current playback content and the identifier or category of the operation target.

[0015] The above indicators are weighted according to preset non-negative weights to determine the intensity of the immediate intent; when the intensity of the immediate intent reaches the immediate intent threshold, and a voice operation, search operation, channel selection operation or content confirmation operation with a recognition confidence level reaching the voice confidence threshold is detected, it is determined that there is an immediate viewing intent.

[0016] Preferably, the feature analysis model includes a long-term preference analysis sub-model and a session preference analysis sub-model; the user's historical playback, search, favorite, click and skip behavior is input into the long-term preference analysis sub-model to obtain the user's long-term preference features, and the behavior in the current boot or viewing session is input into the session preference analysis sub-model to obtain the current session preference features;

[0017] The credibility of user identification is determined based on account login status, voiceprint matching score, remote control usage history matching score, usage time period matching score, and home mode. The credibility level is set to a value of [value missing]. When the long-term preference feature and the current session preference feature have the same feature dimension, the long-term preference feature and the current session preference feature are weighted and combined according to the degree of credibility to obtain a combined user preference feature; when the long-term preference feature and the current session preference feature have different feature dimensions, the current session preference feature is projected to the same dimension as the long-term preference feature through a linear mapping matrix, and then weighted and combined.

[0018] When a user's identity is not confirmed, they are in guest mode, or different voiceprints, accounts, or remote control usage characteristics are detected alternating within a preset time window, the credibility level is reduced or the credibility level is reset to a preset low credibility value.

[0019] Preferably, the determination of the service level includes:

[0020] When the playback risk indicator reaches the playback protection threshold, the remaining duration of the playback buffer in the playback state is lower than the protection entry threshold, or the player status flag indicates that it is in the first frame loading, dragging, channel switching, or resolution switching stage, it is determined to be a playback continuity protection service; when it does not belong to the playback continuity protection service and the instant intent strength reaches the instant intent threshold, it is determined to be an instant intent response service; when it does not belong to the aforementioned two types of services, the resource availability reaches the recommended threshold, and the remaining duration of the playback buffer reaches the recommended buffer threshold, it is determined to be a regular personalized recommendation service.

[0021] If the number of interactive events within a preset time window is lower than the threshold for the number of interactions, or if there are no voice, remote control, search, or content confirmation operations for a preset duration, it is determined to be a background update service. When multiple conditions are met at the same time, the service level is determined in the following order: playback continuity protection service, instant intent response service, regular personalized recommendation service, and background update service.

[0022] Preferably, the terminal resource status includes intelligent processor utilization, dynamic random access memory bandwidth, network throughput, and recommended cache availability;

[0023] Based on the normalized values ​​of the intelligent processor idle rate, dynamic random access memory bandwidth margin, network throughput margin, and cache margin, the resource availability is determined according to the preset weights; in the playback continuity protection service, the complete recommendation calculation, media resource prefetching, model parameter download, and batch content feature generation are suspended, the video stream transmission priority is set, and the network bandwidth and dynamic random access memory usage of the recommendation task are limited;

[0024] In the instant intent response business, local content metadata and updated content feature data are read, and a simplified content sorting model is invoked, requesting only the cover and metadata; in the regular personalized recommendation business, recommendation calculations are performed in batches according to the idle rate of the intelligent processor; in the background update business, long-term preference updates, log uploads, model parameter downloads, or batch content feature generation are performed.

[0025] Preferably, the combined user preference features, updated content feature data, and viewing environment features are input into the content ranking model to obtain an initial score for the candidate content; a timeliness decay coefficient is determined based on the time difference between the publication time or event start time of the candidate content and the current time, wherein the publication time adopts exponential decay, the event start time adopts segmented decay before and after the event start, and the initial score of the candidate content is corrected using the timeliness decay coefficient.

[0026] When the intensity of the instant intent reaches the instant intent threshold and a voice operation, search operation, channel selection operation, or content confirmation operation is detected, a matching score is determined based on the similarity between the candidate content and the voice operation, search term, or operation target. The matching score is then weighted with the corrected initial score of the candidate content according to a preset weight to obtain a candidate content ranking score.

[0027] The filtering is based on copyright region, subscription permissions, terminal encoding capabilities, age rating, children's mode, playback buffer, and intelligent processor, dynamic random access memory bandwidth, and remaining network resources. In children's mode, the age rating of candidate content must not exceed a preset level. The resource budget includes at least one of the following: intelligent processor budget, dynamic random access memory bandwidth budget, network bandwidth budget, and cache space budget.

[0028] Preferably, the storage module includes a result buffer, a metadata buffer, and a media resource buffer; it obtains the metadata and media resources of the candidate content from the content server, the local content index, or the cache management module, saves the recommendation results, content identifiers, and display priorities in the result buffer, saves the title, category, duration, rating, playback address, age rating, content status, and cover address in the metadata buffer, and saves the cover, trailer, low bitrate preview clips, and content segments in the media resource buffer;

[0029] When the smart TV is powered on, woken up, resumed from standby, the remaining playback time of the current video is less than a preset end threshold, or the channel list or recommendation entry is opened a preset number of times within a preset time window, a simplified recommendation program is loaded, which includes local recommendation result reading, content index reading, and a simplified content sorting model execution program. The recommended content is first displayed using the result buffer and metadata buffer. When the remaining network bandwidth is higher than the preset loading threshold, media resources are loaded asynchronously according to the display priority, and the loading bitrate and concurrency are limited. When the remaining network bandwidth is less than or equal to the preset loading threshold, the loading of low display priority media resources is paused.

[0030] Preferably, the system continuously collects data on the intelligent processor utilization rate, dynamic random access memory bandwidth, system temperature, network throughput, playback buffer and cache utilization rate, and determines the recommended cache availability based on the cache utilization rate; when the service level upgrade / downgrade conditions are met for a preset number of sampling periods, the service level is switched, wherein the playback continuity protection service adopts a protection entry threshold and a protection exit threshold, and the protection entry threshold is less than the protection exit threshold;

[0031] When the intelligent processor utilization rate reaches the intelligent processor protection threshold or the available computing time is lower than the computing time threshold, switch to the simplified content sorting model or rule sorting method. When the network throughput is lower than the network availability threshold or continuous packet loss occurs, read the result buffer and metadata buffer. When the available dynamic random access memory is lower than the dynamic random access memory protection threshold, release the preview segments in the media resource buffer and reduce the number of candidates according to the available dynamic random access memory, but retain the preset minimum number of candidates. When the system temperature reaches the temperature protection threshold, reduce the recommended computing batch and extend the background task polling cycle.

[0032] The calculation batch and polling cycle are restored when the system temperature is below the preset recovery threshold and the preset sampling period is met continuously.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. This method extracts long-term and current session preferences through a feature analysis model, combines features with the credibility of identity recognition, and determines the intensity of immediate intent. This solves the problem that a unified profile is prone to deviating from the target in shared scenarios. This method effectively distinguishes between identity switching and immediate intent, improving the accuracy of intent recommendation.

[0035] 2. This method determines the service level based on playback risk indicators and resource availability, and dynamically executes the pause, downgrade, or resumption of recommended tasks; this solves the problem of playback stuttering and response delay caused by the fixed mechanism preempting resources; this method realizes dynamic hierarchical management and control, which ensures playback continuity while supporting the stable operation of multiple tasks. Attached Figure Description

[0036] The above and other objects, features and advantages of the present invention will become clearer from the following description of embodiments of the invention with reference to the accompanying drawings, in which:

[0037] Figure 1 A flowchart illustrating the steps of a deep learning-based personalized content recommendation method for smart TVs provided in this application embodiment. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0039] This embodiment applies to a smart TV shared in a home; the smart TV performs recommendation processing through a control terminal, and the data acquisition module in the control terminal is connected to the player, the terminal performance counter, the network interface, and the interaction module. The interaction module receives remote control operations, voice operations, search operations, and content operations from the recommendation page.

[0040] The recommendation processing module generates recommendation tasks based on the terminal operation data obtained during the sampling period, and determines the processing method of the recommendation tasks according to the current playback status and user operation status.

[0041] like Figure 1 As shown, a deep learning-based personalized content recommendation method for smart TVs is applied to the control terminal of a smart TV. The control terminal includes a data acquisition module, an interaction module, a storage module, a recommendation processing module, a service hierarchy module, a resource orchestration module, and a state proxy module. The method includes: collecting playback status, terminal resource status, user behavior status, user identity status, content status, network status, and viewing environment status through the data acquisition module according to a preset sampling period.

[0042] The collected state data is time-aligned, anomaly removed, normalized, and subjected to short-term statistics to obtain playback features, user features, content features, network features, and viewing environment features. The user features, content features, playback features, and network features are then input into the feature analysis model to obtain long-term user preference features, current session preference features, and updated content feature data.

[0043] Playback risk indicators are determined based on playback characteristics; the intensity of immediate intent is determined based on user characteristics and user behavior status; resource availability is determined based on terminal resource status; and content timeliness is determined based on content status. The service level of recommendation tasks is determined according to the priority order of playback continuity protection service, immediate intent response service, regular personalized recommendation service, and background update service.

[0044] Based on service level and content timeliness, the system performs recommendation calculations, candidate content sorting, content prefetching, cache loading, or background updates, and filters the sorting results of candidate content to generate recommendation results. Based on playback characteristics, terminal resource status, and network status, the system pauses, downgrades, or resumes recommendation tasks when preset conditions are met.

[0045] In this embodiment, the data acquisition module reads the remaining duration of the playback buffer, video decoder occupancy rate, current video bitrate and player stage flag of the player according to a unified sampling period, reads the intelligent processor occupancy rate, dynamic random access memory bandwidth, network throughput, recommended cache availability and system temperature in the terminal resource status, and obtains voice operation, remote control operation, page dwell and content operation records from the interaction module.

[0046] Content status is provided by the content server, local content index, or cache management module, and includes at least the content category, publication time or event start time, content bitrate, copyright region, subscription permissions, age rating, encoding format, and current playability; viewing environment status is used to indicate the current time period, screen resolution, current channel, and terminal environment information related to the viewing process.

[0047] To enable data from different sources to be used in the same recommendation request, the recommendation processing module aligns the time of various types of data according to the sampling period identifier; to address the conflict caused by the inconsistent underlying hardware clock domains, the data acquisition module sets an independent underlying polling frequency based on an asynchronous multi-threaded queue.

[0048] High-frequency state flow uses less than 1000 The microsecond-level interrupt drive uses a sliding window mean filter for the low-frequency state stream; when the current unified preset sampling clock arrives, the low-frequency data is latched by a zero-order hold to complete the time alignment of heterogeneous timing.

[0049] For data that is not successfully acquired in the current sampling period, the corresponding stable record in the previous valid sampling period is read as the reference value for the current period, and the validity flag of the record is retained; for data that exceeds the preset range, whose deviation from the adjacent sampling period does not conform to the acquisition pattern, or that cannot pass the interface verification, anomaly rejection is performed;

[0050] When a valid alternative record cannot be obtained, the corresponding feature is marked as unavailable, and the subsequent service level determines whether to use a local result buffer or a simplified sorting method.

[0051] Normalization converts playback buffer, decoder utilization, smart processor utilization, dynamic random access memory bandwidth utilization, network throughput, and cache utilization into comparable feature values.

[0052] In this embodiment of the invention, the intelligent processor utilization rate refers to the utilization rate of the terminal intelligent processor, the dynamic random access memory bandwidth refers to the bandwidth usage of the dynamic random access memory, the network throughput rate is the effective network throughput rate in the playback risk calculation, and the network throughput margin after deducting the playback bitstream occupation is used in the resource availability calculation. The recommended cache availability rate is calculated based on the cache utilization rate.

[0053] Short-term statistics are used to form the average value, direction of change, and number of consecutive sampling periods that meet the conditions within the current sampling period; the processed playback features, user features, content features, network features, and viewing environment features are written into the status record in the storage module and read by the recommendation processing module with the recommendation request identifier to avoid re-accessing the original collected data in subsequent recommendation calculations;

[0054] The feature analysis models include a long-term preference deep learning model using a transformer encoder architecture based on a deep self-attention mechanism, a session preference deep learning model using a gated recurrent unit architecture, and a content feature model using a multilayer perceptron. User history playback, search, favorite, click, and skip behaviors are encoded one-hot and dimensionality reduced by the embedding layer to form a 64-dimensional historical behavior sequence vector. Operations in the current boot or viewing session are also embedded to form a 64-dimensional current session behavior sequence vector.

[0055] The recommendation processing module feeds the two types of behavior sequences into the corresponding models. After forward propagation calculation, it outputs a 128-dimensional long-term user preference feature vector and a 128-dimensional current session preference feature vector, respectively. It also feeds the content title, description, category, actor, channel, tag, duration and program information into the content feature model to generate updated content feature data.

[0056] Playback risk indicators, real-time intent strength, resource availability, and content timeliness are written into the task record of the current recommendation request by the recommendation processing module; content timeliness is determined by the relationship between the publication time or event start time in the content status and the current time;

[0057] For content with an event start time, timeliness is determined based on the stages before the event starts, during the event, and after the event ends. For content with only a publication time, timeliness is determined based on the duration after the content is published. After reading the task record, the service classification module determines the business category to which the recommendation task belongs, and the resource orchestration module sets the execution order of recommendation calculation, candidate content sorting, content prefetching, cache loading, and background update tasks accordingly.

[0058] After the candidate content is sorted, the recommendation processing module filters it according to content availability, user safety, playback stability and resource budget. The filtered content identifiers, sorting results and display priorities are written to the storage module for the interaction module to read and display.

[0059] If the playback buffer decreases, the video decoder utilization rate increases, the network throughput is insufficient, or the network packet loss rate increases, causing the playback risk indicators to deviate from the playback safety conditions, the recommendation processing module will suspend the full recommendation calculation and media resource prefetching.

[0060] If the bandwidth of the intelligent processor, dynamic random access memory, network, or cache resources cannot meet the requirements of regular recommendations, the recommendation calculation batch, the number of candidates, or the model complexity will be reduced. If the terminal resources and playback buffer are restored to the corresponding recovery conditions and the preset sampling period is met continuously, the task will be restored in the order of simplified recommendation, regular recommendation, and background update. If the network request times out or the content server does not return candidate content, the current recommendation request will be completed using the local content index and the saved recommendation results.

[0061] Playback status includes the remaining duration of the playback buffer, video decoder occupancy rate, and current video bitrate; network status includes effective network throughput and network packet loss rate.

[0062] The buffer risk value is determined based on the relationship between the remaining playback buffer duration and the low buffer threshold; the decoding risk value is determined based on the relationship between the video decoder occupancy rate and the decoding occupancy threshold; the bitrate risk value is determined based on the matching relationship between the current video bitrate and the effective network throughput; and the packet loss risk value is determined based on the relationship between the network packet loss rate and the packet loss threshold. Each risk value is then normalized and truncated to [a specific value]. Playback risk indicators are determined by weighting according to preset non-negative weights.

[0063] In this embodiment, the recommendation processing module reads the remaining duration of the playback buffer and the current video bitrate from the player status record, reads the decoder occupancy rate from the video decoder interface, and reads the effective network throughput and network packet loss rate from the network statistics interface.

[0064] When the remaining playback buffer time is below the low buffer threshold, the buffer risk value increases as the remaining buffer time decreases; when the decoder occupancy rate is above the decoding occupancy threshold, the decoding risk value increases as the occupancy rate increases; when the current video bitrate is close to or higher than the network's effective throughput, the bitrate risk value... Through the formula:

[0065] The calculation shows that, A default constant penalty term, used to prevent division-by-zero crashes, has the same dimensions as the current video bitrate. The current video bitrate, For network effective throughput, The preset trigger threshold coefficient, This represents the operation of finding the maximum value between x and y; network packet loss rate. Above the packet loss threshold At that time, packet loss risk value Through the formula:

[0066]

[0067] The calculation shows that, Indicates taking and Minimum value operation;

[0068] Each risk value is normalized and limited to the range of zero to one. The recommended processing module configures risk initiation conditions and risk reference upper limits for each risk item. The risk initiation conditions are determined based on the historical percentile values ​​of the corresponding indicators of the smart TV under normal continuous playback conditions. The risk reference upper limit is determined based on the acceptable maximum value reported by the player or network interface. The configuration is saved when the terminal model, video encoding format and current playback resolution change.

[0069] When the value does not meet the risk initiation condition, the corresponding risk value is determined to be zero; when the value is between the risk initiation condition and the risk reference upper limit, the risk value is determined according to the relative position of the value between the two boundaries; when the value reaches or exceeds the risk reference upper limit, the corresponding risk value is determined to be one.

[0070] For indicators such as the remaining buffer time, which decrease as risk decreases, they are first converted into gap values ​​that are positively correlated with risk according to the relative position of the low buffer threshold and the protection reference time, and then processed according to the same rules; where the low buffer threshold, decoding occupancy threshold and packet loss threshold are the risk initiation conditions corresponding to buffer risk, decoding risk and packet loss risk, respectively.

[0071] The non-negative weights used for weighting are determined according to the historical impact of each risk item on playback stuttering, decoding failure, and bitrate switching, and are pre-normalized to a sum of weights of one. When historical impact data is lacking, the default weights for buffer risk, decoding risk, bitrate risk, and packet loss risk are used. The recommendation processing module performs weighting according to the preset non-negative weights corresponding to buffer risk, decoding risk, bitrate risk, and packet loss risk to form a playback risk index for the current sampling period, and writes the index into the playback status record.

[0072] When the playback risk index reaches the playback protection threshold, the playback continuity protection service is activated. The playback protection threshold is determined based on the distribution of risk indicators between the normal continuous playback state and the stuttering state, and is configured separately from the protection exit threshold to avoid the risk index from repeatedly switching near the critical value.

[0073] If a playback data item is missing within the current sampling period, the recommendation processing module prioritizes reading the previous valid record of that data item and retains the data age and validity flag. If the player reports an abnormal playback status, the data age exceeds the preset validity period, or the previous valid record does not exist, the risk item is set to protection value one, and unnecessary recommendation tasks that would consume network bandwidth or dynamic random access memory bandwidth are prohibited. After the player and network interface return valid data, the playback risk index is recalculated according to the recovery conditions.

[0074] The interactive module collects voice operations, remote control operations, page dwell times, content operations, and the correspondence between the currently playing content and the operation target. It determines the degree of voice matching based on the similarity between the voice-recognized text and the content title, content category, or search target; it determines the concentration of remote control behavior based on the ratio of the number of operations on the same channel, the same content category, or the same function within a preset time window to the total number of remote control operations; it determines the degree of dwell preference based on the ratio of the page dwell time to the average dwell time within a preset time window; and it determines the consistency of content based on the matching results of the current playing content and the identifier or category of the operation target.

[0075] The above indicators are weighted according to preset non-negative weights to determine the intensity of real-time intent. When the intensity of real-time intent reaches the real-time intent threshold, and a voice operation, search operation, channel selection operation, or content confirmation operation with a recognition confidence level reaching the voice confidence threshold is detected, it is determined that there is a real-time viewing intent.

[0076] The interaction module receives speech-recognized text, remote control button events, page entry and exit events, content details viewing events, card expansion events, playback confirmation events, and channel selection events. The recommendation processing module converts the speech-recognized text and the content title, content category, or search target into word feature vectors using the term frequency-inverse document frequency algorithm, and calculates the cosine similarity between the two feature vectors. This cosine similarity value is directly used as the degree of speech matching. ;

[0077] For speech recognition text containing explicit content names, channel names, or content categories, the speech matching degree is determined according to the corresponding text similarity; for text that does not form a valid content target, the speech matching degree is set to the first preset lower limit value; the speech recognition confidence is only used to determine whether the speech recognition result is valid, and the validity of the search operation, channel selection operation, and content confirmation operation is determined according to whether the search term is not empty, the channel identifier is valid, or the confirmation event is complete.

[0078] The concentration of remote control behavior is calculated from the remote control operation records within a preset time window; the recommendation processing module merges operations of the same channel, the same content category, or the same function, and uses the ratio of the number of merged operations to the total number of remote control operations to form the concentration of remote control behavior.

[0079] When there is no remote control operation within the preset time window, the concentration of remote control behavior is set to the first preset lower limit value to avoid using zero as the denominator; the page dwell time is compared with the average dwell time within the preset time window to form the dwell preference degree;

[0080] When there are no valid page dwell records within the time window, or the average dwell time is zero, the dwell preference level is set to the first preset lower limit value; when the current playback content has the same identifier as the operation target, or the two categories match, the content consistency takes the corresponding matching value; when neither the identifier nor the category matches, the content consistency takes the first preset lower limit value.

[0081] The above indicators are written into the user behavior status record and weighted according to the preset non-negative weights to form the instant intent intensity; when no remote control operation, voice operation, search operation or content confirmation operation occurs within the preset time window, the corresponding indicator adopts the first preset lower limit value, and the instant intent intensity calculation is not interrupted due to the lack of operation records.

[0082] When speech recognition fails, the recognition confidence level is lower than the speech confidence level threshold, or the search text is empty, the event will not be considered a valid instant viewing intent; similarly, when the channel selection operation lacks a valid channel identifier, or the content confirmation operation lacks a valid operation target, the event will also not be considered a valid instant viewing intent.

[0083] When the intensity of the instant intent reaches the instant intent threshold, and there are voice operations, valid search operations, channel selection operations, or content confirmation operations that reach the voice confidence threshold, the recommendation processing module writes an instant viewing intent flag to the service classification module; after reading the flag, the service classification module determines the corresponding recommendation task as an instant intent response service if it does not currently belong to the playback continuity protection service.

[0084] The system configures the credibility of user identity recognition between long-term user preference characteristics and current session preference characteristics, and stores the identity recognition results in isolation from the service level determination and candidate content ranking process;

[0085] The feature analysis model includes a long-term preference analysis sub-model and a session preference analysis sub-model. The user's historical playback, search, favorite, click and skip behavior is input into the long-term preference analysis sub-model to obtain the user's long-term preference features, and the behavior in the current boot or viewing session is input into the session preference analysis sub-model to obtain the current session preference features.

[0086] The credibility of user identification is determined based on account login status, voiceprint matching score, remote control usage history matching score, usage time period matching score, and home mode. The credibility level is set to a value of [value missing]. When long-term preference features and current session preference features have the same feature dimensions, the long-term preference features and current session preference features are weighted and combined according to the degree of credibility to obtain combined user preference features.

[0087] When the feature dimensions of long-term preference features and current session preference features are inconsistent, the current session preference features are projected to the same dimension as the long-term preference features through a linear mapping matrix, and then weighted combination is performed.

[0088] When a user's identity is not confirmed, they are in guest mode, or different voiceprints, accounts, or remote control usage characteristics are detected alternating within a preset time window, the trust level is reduced or reset to a preset low trust level value.

[0089] In this embodiment, the long-term preference analysis sub-model reads the historical playback, search, favorite, click and skip behaviors associated with the account or family mode from the storage module, and forms a historical behavior sequence according to the behavior type and the sampling period of the behavior occurrence;

[0090] The conversation preference analysis sub-model reads voice, remote control, search, channel selection, page dwell, and content confirmation behaviors generated during the current power-on or viewing conversation to form the current conversation behavior sequence; the two sub-models output long-term preference features and current conversation preference features, respectively.

[0091] The recommendation processing module obtains identity judgment criteria from account login records, voiceprint recognition results, remote control usage history and usage time records, and determines the credibility of user identity recognition in combination with the home mode; when the account login status is consistent with the current session account, the voiceprint matching score reaches the preset voiceprint threshold, the remote control usage history matching degree reaches the preset history matching threshold and the usage time matching degree matches the existing records, the credibility is improved;

[0092] When corresponding information is missing or inconsistent, the level of trust is reduced; the level of trust is limited to the range of zero to one and is written into the user identity status record along with the current session state.

[0093] Only in long-term preference eigenvectors and current session preference feature vector When features have the same dimensions, the recommendation processing module will then prioritize based on the credibility of the identity recognition. The linear weighted combination of the two types of features is calculated using the following formula:

[0094] This yields a combined user preference feature vector. ;

[0095] If the system's model heterogeneity leads to inconsistent dimensions between the two types of features, then feature dimensioning is introduced. and By constructing a learnable linear mapping matrix With bias term Using the formula After projecting the current session preference features onto the same dimension as the long-term preference features, a linear weighted combination is then performed;

[0096] When a unified dimensional transformation cannot be completed, a simplified sorting is performed using the current session preference features and the local content index.

[0097] When the account is not logged in, the user's identity is not confirmed, or the TV is in guest mode, the recommendation processing module will reduce the trust level to a preset low trust level; when different voiceprints, different accounts, or different remote control usage characteristics are detected to alternate within a preset time window, the recommendation processing module will reduce the trust level.

[0098] When the alternation operation reaches the identity switching judgment condition, the credibility level is reset to the preset low credibility value, and the session preference features are reconstructed based on the current session behavior; when the identity recognition interface is unresponsive or the voiceprint matching fails, the failure result is not used as a definite identity information.

[0099] The service level is determined as follows: when the playback risk indicator reaches the playback protection threshold, the remaining duration of the playback buffer in the playback state is lower than the protection entry threshold, or the player status flag indicates that it is in the first frame loading, dragging, channel switching or resolution switching stage, it is determined to be a playback continuity protection service.

[0100] When the service does not fall under the category of playback continuity protection and the intensity of the instant intent reaches the instant intent threshold, it is identified as an instant intent response service; when the service does not fall under either of the aforementioned two categories, the resource availability reaches the recommended threshold, and the remaining duration of the playback buffer reaches the recommended buffer threshold, it is identified as a regular personalized recommendation service; when the number of interaction events within the preset time window is lower than the interaction count threshold, or when there are no voice, remote control, search, or content confirmation operations for a preset duration, it is identified as a background update service.

[0101] When multiple conditions are met simultaneously, the service level is determined in the following order: playback continuity protection service, real-time intent response service, regular personalized recommendation service, and background update service.

[0102] In this embodiment, the service classification module reads the playback risk index, remaining playback buffer duration, player status flag, real-time intent strength, resource availability, and number of interaction events from the status record of the current recommendation request; the player status flag is written by the player during the first frame loading, dragging, channel switching, or resolution switching stages, and the recommendation processing module does not infer the above player stages based on a single user operation;

[0103] When the playback risk indicator reaches the playback protection threshold, the remaining duration of the playback buffer in the playback state is lower than the protection entry threshold, or the player status flag indicates that it is in a specified playback stage, the service classification module will mark the task record as a playback continuity protection service; this judgment takes precedence over other service judgments; even if there are search, channel selection or content confirmation operations at the same time, as long as the playback continuity protection condition is met, the current task will still be processed as a playback continuity protection service.

[0104] When the instant intent intensity reaches the instant intent threshold and the service classification module does not belong to the playback continuity protection service, the service classification module checks whether the interaction module provides valid voice, search, channel selection or content confirmation operations; if there are valid operations, the task is marked as an instant intent response service; if the instant intent intensity reaches the threshold but there are no valid operation records, it will not enter the instant intent response service.

[0105] If neither of the first two business conditions is met, the service classification module checks resource availability and the remaining playback buffer time. When both reach the recommendation threshold and the recommendation buffer threshold respectively, the task is marked as a regular personalized recommendation business. If the number of interaction events is lower than the interaction count threshold, or there are no voice, remote control, search, or content confirmation operations for a preset duration, and the task does not belong to the aforementioned business, it is marked as a background update business.

[0106] When multiple conditions are met simultaneously in the same sampling period, the service classification module selects the service level in the order of playback continuity protection, instant intent response, regular personalized recommendation, and background update, and writes the selection result into the task status record. If the status data of the current sampling period is missing, the service level of the previous valid period is used, and it is restricted to only executing the allowed tasks. When the continuous sampling period regains a valid status, the service level is switched.

[0107] Terminal resource status includes intelligent processor utilization, dynamic random access memory bandwidth, network throughput, and recommended cache availability. Based on the normalized values ​​of intelligent processor idle rate, dynamic random access memory bandwidth margin, network throughput margin, and cache margin, resource availability is determined according to preset weights.

[0108] In the continuous playback protection service, the full recommendation calculation, media resource prefetching, model parameter download and batch content feature generation are paused, the video stream transmission priority is set and the network bandwidth and dynamic random access memory usage of the recommendation task are limited; in the instant intent response service, local content metadata and updated content feature data are read, the simplified content sorting model is called, and only the cover and metadata are requested.

[0109] In regular personalized recommendation operations, recommendation calculations are performed in batches according to the idle rate of the intelligent processor; in background update operations, long-term preference updates, log uploads, model parameter downloads, or batch content feature generation are performed.

[0110] The system converts the status of various terminal resources into resource availability indicators and sets the resource availability indicators as the conditions for triggering the execution of regular recommendation services, so as to constrain the resources of intelligent processor, dynamic random access memory bandwidth, network throughput and recommendation cache availability.

[0111] The recommended processing module normalizes the intelligent processor idle rate, dynamic random access memory bandwidth margin, network throughput margin, and cache margin respectively, and then weights them according to preset weights to form resource availability; each resource quantity is configured with a lower limit of resource availability and a higher limit of resource reference, and the lower limit of resource availability is used as the zero point of normalization and the higher limit of resource reference is used as the point of normalization.

[0112] When the resource quantity is below the available lower limit, it is treated as zero; when it is above the reference upper limit, it is treated as one; when it is between the two, the normalized value is determined according to its relative position between the two boundaries; the intelligent processor idle rate is the remaining proportion of the intelligent processor utilization rate; the dynamic random access memory bandwidth margin is the available bandwidth of the dynamic random access memory; the network throughput margin is the remaining throughput capacity after deducting the actual occupation of the current playback bitstream from the effective network throughput rate; and the cache margin is the available space corresponding to the recommended cache availability.

[0113] The available lower limit and reference upper limit of each resource are determined based on the minimum playback resource requirements of the terminal model, the minimum execution resource requirements of the recommended task, and the range of the resource monitoring interface, and are stored in the terminal resource configuration record respectively.

[0114] When the configuration is missing or the two boundaries do not meet the requirement that the lower available limit is less than the upper reference limit, the default boundary of the corresponding resource item is adopted; the preset weight of each resource is determined according to its impact on playback continuity and recommendation response latency. When there is a lack of terminal running records, the default weight is used, and the sum of the weights is one.

[0115] When the intelligent processor utilization rate, dynamic random access memory bandwidth utilization rate, or cache utilization rate lacks valid sample values, the previous valid record is read and its validity period is checked; if no data is obtained after the validity period, the normalized value of the corresponding resource item is set to zero; if the network throughput rate cannot be obtained, the network throughput margin is not determined to be available, but media resource prefetching is restricted and local cache is read first.

[0116] Regular personalized recommendation services are only allowed when resource availability reaches the recommended threshold and the playback buffer reaches the recommended buffer threshold. The recommended threshold is determined based on the resource availability corresponding to the minimum resource requirement of the recommendation task, and a protection range below the recommended threshold is set to avoid the recommendation task being repeatedly started and stopped near the critical value of resource availability.

[0117] In the continuous playback protection service, the resource orchestration module suspends the full recommendation calculation, media resource prefetching, model parameter downloading and batch content feature generation, sets the video bitstream transmission to a higher transmission priority than the recommendation task, and sets the network bandwidth and dynamic random access memory usage limit for the recommendation task.

[0118] Low-priority tasks that have entered the execution queue but have not yet started remain in a waiting state; tasks that are currently being executed stop at the current interruptible node and save their task state for later retrieval during resumption.

[0119] In the real-time intent response business, the recommendation processing module prioritizes reading local content metadata and updated content feature data, calls a simplified content sorting model to complete candidate sorting, and only requests cover and metadata from the content server. It does not trigger large-scale media resource pre-fetching due to interactive operations. If the local content metadata lacks playback address, age rating or content status, the candidate content will not be included in the current recommendation results, or it will only be retained in the local index as unplayable content information.

[0120] In regular personalized recommendation services, the recommendation processing module determines the number of candidates and execution frequency for a single batch of recommendation calculations based on the idle rate of the intelligent processor; when the idle rate of the intelligent processor decreases, the number of candidates in a single batch is reduced and the calculation of the next batch is postponed.

[0121] When the intelligent processor idle rate recovers and the playback buffer reaches the recommended buffer threshold, the unfinished candidate content continues to be processed; the number of candidates in a single batch and the execution frequency are determined according to the pre-configured intelligent processor idle rate segmentation table, and the number of candidates in each segment is not less than the preset minimum number of candidates. When the intelligent processor idle rate is lower than the minimum execution condition, the calculation of that batch is paused.

[0122] In the background update process, the recommendation processing module performs long-term preference updates, log uploads, model parameter downloads, or batch content feature generation; background tasks utilize limited intelligent processor time slices, network bandwidth, and dynamic random access memory bandwidth.

[0123] When voice operation, remote control operation, search operation, channel selection or content confirmation operation occurs, the background task releases the currently occupied resources and writes the intermediate progress of unfinished tasks into the task status record; when the model parameter download is incomplete or the integrity verification fails, the currently valid model continues to be used and does not switch to unverified model parameters.

[0124] The combined user preference features, updated content feature data, and viewing environment features are input into the content ranking model to obtain the initial score of the candidate content. The time-effect decay coefficient is determined based on the time difference between the publication time or event start time of the candidate content and the current time. The publication time adopts exponential decay, and the event start time adopts segmented decay before and after the event start. The initial score of the candidate content is corrected using the time-effect decay coefficient.

[0125] When the intensity of the instant intent reaches the instant intent threshold and a voice operation, search operation, channel selection operation, or content confirmation operation is detected, the matching score is determined based on the similarity between the candidate content and the voice operation, search term, or operation target. The matching score is then weighted with the corrected initial score of the candidate content according to a preset weight to obtain the candidate content ranking score.

[0126] The filtering is based on copyright region, subscription permissions, terminal encoding capabilities, age rating, children's mode, playback buffer, and intelligent processor, dynamic random access memory bandwidth, and remaining network resources. In children's mode, the age rating of candidate content must not exceed a preset level. The resource budget includes at least one of the following: intelligent processor budget, dynamic random access memory bandwidth budget, network bandwidth budget, and cache space budget.

[0127] The recommendation processing module reads the combined user preference features formed by the aforementioned user identification and feature combination process, the updated content feature data output by the content feature model, and the viewing environment features, and inputs them into the content ranking model; the content ranking model calculates the initial score of the candidate content based on user preferences, content semantics, and the current viewing environment;

[0128] The viewing environment features include the current time period, current channel, playback stage, screen resolution, network status, and real-time intent tags; the content ranking model adopts a deep ranking model corresponding to the unified feature dimension. The model parameters are trained by using supervised labels formed from historical behavior samples according to playback completion, favorite, click, play after search, and skip behavior, and the model version that can simultaneously meet the requirements of ranking accuracy and inference latency is selected through the validation set.

[0129] The model version, input feature order, feature dimensions, and missing value handling method are written into the model configuration record, and features are read according to the configuration record during inference.

[0130] For candidate content with a publication time, the recommendation processing module calculates the exponential decay coefficient based on the time difference between the publication time and the current time; for candidate content with an event start time, the timeliness decay coefficient is calculated using a segmented decay method based on the time difference before the event starts, during the event, and after the event ends.

[0131] The first, second, and third advance phases are defined as follows: the closer the time is to the start time of the event, the smaller the time reduction; during the event, the set upper limit of time is maintained; after the event ends, the time reduction value is gradually reduced according to the time interval after the event ends.

[0132] The boundaries of each stage are determined based on the program cycle, event duration, and historical peak viewing periods of the content type, and are saved in the content category configuration record; the exponential decay parameter of the release time is determined based on the historical clicks and playback duration of the corresponding content category; when the historical records of the category are missing, the default parameters of the general content category are used; the recommendation processing module uses the time-effect decay coefficient to correct the initial score of candidate content;

[0133] When the content type is missing or neither the publication time nor the event start time exists, the basic timeliness processing method corresponding to the content category is used, and the time decay without a definite basis is not applied to the candidate content; for candidate content that has both a publication time and an event start time, the event timeliness processing is prioritized according to the event start time to avoid applying two timeliness decay methods to the same candidate content repeatedly.

[0134] When the intensity of the instant intent reaches the instant intent threshold and a voice operation, search operation, channel selection operation, or content confirmation operation is detected, the recommendation processing module calculates the similarity between the candidate content and the voice operation, search term, or operation target to form a matching score.

[0135] Similarity is calculated using the same text or category feature encoding method as when the content ranking model was trained, and matching scores that lack effective operation targets are marked as unusable; for candidate content that does not match the explicit operation target, forced removal is performed according to the content category or operation target corresponding to the immediate intent; for candidate content that passes the hard filtering, the matching score and the time-corrected score are weighted according to the preset weight to obtain the candidate content ranking score;

[0136] The preset weights are determined based on the click, play confirmation, and skip actions of candidate content after an instant action. When there are no records of instant action actions, the default weights are used, and the matching scores and the corrected scores are weighted using the same scoring range. When there are no valid instant actions, the instant intent matching score is not used to change the regular sorting results.

[0137] After the candidate content is sorted, the recommendation processing module checks the copyright area and subscription permissions in turn to exclude content that exceeds the current copyright area or that the current account has not obtained the subscription permission; then it checks the terminal encoding capability and the playable bitrate of the content to exclude content that the terminal cannot decode or that the current network cannot meet the playback conditions.

[0138] Filtering is performed based on age rating, children's mode, and playback stability; in children's mode, the age rating of candidate content must not exceed the preset level; when the playback buffer is lower than the recommended display threshold, or when the player is in the first frame loading, dragging, channel switching, or resolution switching stage, candidate content is prohibited from triggering new media requests.

[0139] For candidate content that is missing any of the following: copyright area, subscription permissions, terminal encoding capabilities, age rating, playback buffer, or resource status, the reason for the missing content will be recorded and the content will be handled as unplayable or unable to add new media requests. No other scoring items will be used to compensate for the missing condition.

[0140] The recommendation processing module writes at least one of the intelligent processor budget, dynamic random access memory bandwidth budget, network bandwidth budget, and cache space budget into the current recommendation task; candidate content only enters the recommendation list or media resource loading queue when the corresponding resource budget allows it; when the filtering conditions lack copyright region, subscription permission, age rating, or encoding capability data, it is treated as not having playable confirmation conditions and the candidate content is not directly written into the playable recommendation results;

[0141] The storage module includes a result buffer, a metadata buffer, and a media resource buffer. It obtains the metadata and media resources of candidate content from the content server, local content index, or cache management module, saves the recommendation results, content identifiers, and display priorities in the result buffer, saves the title, category, duration, rating, playback address, age rating, content status, and cover address in the metadata buffer, and saves the cover, trailer, low bitrate preview clips, and content segments in the media resource buffer.

[0142] When the smart TV is powered on, woke up, resumed from standby, the remaining playback time of the current video is less than the preset end threshold, or the channel list or recommendation entry is opened a preset number of times within the preset time window, a simplified recommendation program is loaded, which includes local recommendation result reading, content index reading and simplified content sorting model execution program. The recommended content is first displayed using the result buffer and metadata buffer.

[0143] When the remaining network bandwidth is higher than the preset loading threshold, media resources are loaded asynchronously according to the display priority, and the loading bitrate and concurrency are limited; when the remaining network bandwidth is less than or equal to the preset loading threshold, the loading of low display priority media resources is paused.

[0144] The storage module divides the recommendation data into a result buffer, a metadata buffer, and a media resource buffer. After obtaining the metadata and media resources of the candidate content from the content server, local content index, or cache management module, the recommendation processing module writes the recommendation results, content identifiers, and display priorities into the result buffer, writes the title, category, duration, rating, playback address, age rating, content status, and cover address into the metadata buffer, and writes the cover, trailer, low bitrate preview clips, and content segments into the media resource buffer.

[0145] The three buffers record content identifiers respectively, so that the result buffer and metadata buffer can still be used to complete the display of the recommendation page when media resources are missing;

[0146] When a smart TV is powered on, woken up, resumes from standby mode, or when the remaining playback time of the current video is less than a preset end threshold, or when the channel list or recommendation entry is opened a preset number of times within a preset time window, the recommendation processing module loads a simplified recommendation program; the simplified recommendation program includes local recommendation result reading, content index reading, and a simplified content sorting model execution program;

[0147] Once loaded, the recommendation processing module first reads the recommendation results from the result buffer, and then reads the title, category, duration, rating, playback address, age rating, content status and cover address from the metadata buffer. The interaction module then displays the recommended content based on this information.

[0148] When the result buffer contains valid recommendations but the metadata buffer is missing some fields, the interaction module only displays the content information that has been obtained and keeps the missing metadata as a record to be supplemented; when the result buffer is empty, the recommendation processing module reads the local content index and calls the simplified content sorting model to generate candidate results; when the simplified sorting model cannot be executed, the basic display is completed using the local recommendation results and the display priority already in the content index.

[0149] In this embodiment, the remaining network bandwidth is determined by deducting the actual bandwidth occupied by the current playback bitstream and the bandwidth occupied by other approved foreground network tasks from the effective network throughput. The recommendation processing module confirms that the remaining network bandwidth is higher than the preset loading threshold within the continuous sampling period, and at the same time confirms that the availability of the playback buffer, the recommendation cache and the network bandwidth budget meet the loading conditions before establishing a media resource loading task.

[0150] Only when the threshold is reached in a single sampling period, the current loading quantity is maintained and no new concurrent tasks are added; when the remaining network bandwidth is lower than the preset loading threshold, or when an effective network throughput cannot be obtained, the loading of low-priority media resources is paused, and the current video stream and front-end recommendation results are prioritized.

[0151] When the remaining network bandwidth meets the loading conditions, the recommendation processing module asynchronously loads media resources in descending order of display priority, and limits the loading bitrate of a single resource, the number of resources loaded at the same time, and the available space of the media resource buffer.

[0152] If the remaining network bandwidth decreases, the playback buffer falls below the recommended display threshold, or the availability of the recommended cache decreases during the media resource loading process, the loading of low-priority display resources will be paused, and the currently playing video stream and the completed front-end recommendation results will be retained first. If the media resource request times out or the content server does not return the resource, the existing cover, trailer, or low-bitrate preview clip in the media resource buffer will be used. If the corresponding resources do not exist, only the metadata will be retained for display.

[0153] The system continuously collects data on intelligent processor utilization, dynamic random access memory bandwidth, system temperature, network throughput, playback buffer and cache utilization, and determines recommended cache availability based on cache utilization. When the service level upgrade / downgrade conditions are met for a preset number of sampling periods, the service level is switched. The playback continuity protection service uses protection entry threshold and protection exit threshold, with the protection entry threshold being less than the protection exit threshold.

[0154] When the intelligent processor utilization reaches the intelligent processor protection threshold or the available computing time is lower than the computing time threshold, the simplified content sorting model or rule-based sorting method is switched. When the network throughput is lower than the network availability threshold or continuous packet loss occurs, the result buffer and metadata buffer are read. When the available dynamic random access memory is lower than the dynamic random access memory protection threshold, the preview segments in the media resource buffer are released and the number of candidates is reduced according to the available dynamic random access memory, but the preset minimum number of candidates is retained. When the system temperature reaches the temperature protection threshold, the recommended computing batch is reduced and the background task polling cycle is extended. When the system temperature is lower than the preset recovery threshold and the preset sampling cycle is met continuously, the computing batch and polling cycle are restored.

[0155] The status proxy module continuously collects terminal resources, network and playback data according to a preset period, and jointly determines the service level switching operation based on the status data of multiple consecutive sampling periods.

[0156] The status proxy module reads the intelligent processor utilization rate, dynamic random access memory bandwidth, system temperature, network throughput, playback buffer and cache utilization rate in each sampling period. In this embodiment of the invention, the intelligent processor utilization rate read by the status proxy module refers to the intelligent processor utilization rate, the dynamic random access memory bandwidth read refers to the dynamic random access memory bandwidth, and the network throughput rate read adopts the effective network throughput rate when playback risk or network availability threshold is involved, and the network throughput margin is formed after deducting the current playback bitstream occupation when resource availability is involved.

[0157] The higher the cache utilization rate, the lower the recommended cache availability. When the cache utilization rate cannot be obtained, the state proxy module will not allow the increase of media resource loading quantity and will use the saved recommended cache availability as the current judgment basis. The service level upgrade and downgrade conditions are set with the number of sampling periods that must be met continuously. The task status record will only be updated when the corresponding number is reached.

[0158] The playback continuity protection service uses a protection entry threshold and a protection exit threshold. The protection entry threshold is lower than the protection exit threshold. When the playback buffer drops below the protection entry threshold and continuously meets the entry conditions, the status agent module will switch the service level to the playback continuity protection service.

[0159] Playback continuity protection service is only allowed to exit when the playback buffer recovers to above the protection exit threshold and continuously meets the exit conditions; when it is between two thresholds, the current service level is maintained and it is not switched based on changes in a single sampling period.

[0160] When the intelligent processor utilization rate reaches the intelligent processor protection threshold, or the available computing time is lower than the computing time threshold, it is recommended that the processing module switch to the simplified content sorting model or the rule-based sorting method. When the simplified content sorting model cannot be loaded, the runtime context initialization fails, or the intelligent processor is unavailable, the local content index and result buffer are read, and the rule-based sorting is completed according to the program type, time period, channel continuity, and historical click records.

[0161] The sorting results still need to be filtered based on copyright region, subscription permissions, age rating, children's mode, terminal encoding capabilities, and playback stability.

[0162] When network throughput is lower than the network availability threshold or continuous packet loss occurs, the recommendation processing module stops real-time media prefetching and remote model parameter updates, reads the result buffer and metadata buffer to display the recommended content, and temporarily stores the new user behavior in the local log queue. After the network recovers, the logs are uploaded in batches according to the resource limits of the background update business. If the upload fails, the unuploaded records are retained, which does not affect the reading of the current recommendation results.

[0163] When the available dynamic random access memory (DRAM) is lower than the DRAM protection threshold, the recommended processing module first releases the preview segments in the media resource buffer, then stops the prefetching of low-priority media resources, and reduces the number of candidates and the inference batch size according to the remaining available DRAM, but retains the preset minimum number of candidates.

[0164] The current playback buffer, foreground recommendation results, and necessary metadata are not included in the release; if the available dynamic random access memory is still insufficient after releasing media resources, the system will switch to a quantized version of the deep sorting model or a rule-based sorting method.

[0165] When the system temperature reaches the temperature protection threshold, the recommendation processing module reduces the recommendation calculation batch, extends the background task polling cycle, and restricts the simultaneous execution of model updates and media prefetching; after the temperature is lower than the preset recovery threshold and the preset sampling cycle is met continuously, the state proxy module updates the recovery flag, and the recommendation processing module restores the corresponding calculation batch and background task polling cycle.

[0166] The recovery process is still constrained by the utilization rate of the intelligent processor, the bandwidth of the dynamic random access memory, the network throughput, the playback buffer, and the availability of the recommended cache; if any constraint condition is met again, the corresponding degradation process will be re-executed.

[0167] The protection entry threshold, protection exit threshold, instant intent threshold, voice confidence threshold, interaction count threshold, preset loading threshold, preset end threshold, smart processor protection threshold, calculation time threshold, network availability threshold, dynamic random access memory protection threshold, temperature protection threshold, and preset recovery threshold involved in this embodiment are all pre-configured based on the historical operating experience data of the smart TV or hardware specifications, or dynamically updated through the background system configuration interface during terminal operation.

[0168] The above specific embodiments are merely specific implementations of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based personalized content recommendation method for smart TVs, applied to the control terminal of a smart TV. The control terminal includes a data acquisition module, an interaction module, a storage module, a recommendation processing module, a service hierarchy module, a resource orchestration module, and a state proxy module. Its characteristic is that... The methods include: The data acquisition module collects playback status, terminal resource status, user behavior status, user identity status, content status, network status, and viewing environment status according to a preset sampling period; it performs time alignment, anomaly removal, normalization, and short-term statistics on the collected status data to obtain playback features, user features, content features, network features, and viewing environment features. The user characteristics, content characteristics, playback characteristics, and network characteristics are input into a feature analysis model to obtain long-term user preference characteristics, current session preference characteristics, and updated content characteristic data; playback risk indicators are determined based on the playback characteristics; the intensity of immediate intent is determined based on the user characteristics and user behavior status; resource availability is determined based on the terminal resource status; and content timeliness is determined based on the content status. The service level of the recommendation task is determined according to the priority order of playback continuity protection service, instant intent response service, regular personalized recommendation service and background update service; according to the service level and the timeliness of the content, recommendation calculation, candidate content sorting, content prefetching, cache loading or background update are performed, and the sorting results of candidate content are filtered to generate recommendation results; according to the playback characteristics, the terminal resource status and the network status, the recommendation task is paused, downgraded or resumed when preset conditions are met.

2. The method for personalized content recommendation of smart TVs based on deep learning according to claim 1, characterized in that, The playback status includes the remaining duration of the playback buffer, the video decoder occupancy rate, and the current video bitrate; the network status includes the effective network throughput and the network packet loss rate. The buffer risk value is determined based on the relationship between the remaining duration of the playback buffer and the low buffer threshold; the decoding risk value is determined based on the relationship between the video decoder occupancy rate and the decoding occupancy threshold; the bitrate risk value is determined based on the matching relationship between the current video bitrate and the effective network throughput; and the packet loss risk value is determined based on the relationship between the network packet loss rate and the packet loss threshold. Normalize and truncate each risk value. The playback risk index is determined by weighting according to preset non-negative weights.

3. The method for personalized content recommendation of smart TVs based on deep learning according to claim 1, characterized in that, The interaction module collects data on voice operations, remote control operations, page dwell times, content operations, and the correspondence between the currently playing content and the operation target. The degree of voice matching is determined based on the similarity between the voice-recognized text and the content title, content category, or search target. The concentration of remote control behavior is determined based on the ratio of the number of operations on the same channel, the same content category, or the same function within a preset time window to the total number of remote control operations. The degree of dwell preference is determined based on the ratio of the page dwell time to the average dwell time within the preset time window. The consistency of content is determined based on the matching result between the current playback content and the identifier or category of the operation target. The above indicators are weighted according to preset non-negative weights to determine the intensity of immediate intent; When the intensity of the instant intent reaches the instant intent threshold, and a voice operation, search operation, channel selection operation, or content confirmation operation with a recognition confidence level reaching the voice confidence threshold is detected, it is determined that there is an instant viewing intent.

4. The method for personalized content recommendation of smart TVs based on deep learning according to claim 1, characterized in that, The feature analysis model includes a long-term preference analysis sub-model and a session preference analysis sub-model. The user's historical playback, search, favorite, click and skip behavior is input into the long-term preference analysis sub-model to obtain the user's long-term preference features, and the behavior in the current boot or viewing session is input into the session preference analysis sub-model to obtain the current session preference features. The credibility of user identification is determined based on account login status, voiceprint matching score, remote control usage history matching score, usage time period matching score, and home mode. The credibility level is set to a value of [value missing]. When the long-term preference features and the current session preference features have the same feature dimension, the long-term preference features and the current session preference features are weighted and combined according to the degree of credibility to obtain combined user preference features. When the feature dimensions of the long-term preference features and the current session preference features are inconsistent, the current session preference features are projected to the same dimension as the long-term preference features through a linear mapping matrix, and then a weighted combination is performed. When a user's identity is not confirmed, they are in guest mode, or different voiceprints, accounts, or remote control usage characteristics are detected alternating within a preset time window, the credibility level is reduced or the credibility level is reset to a preset low credibility value.

5. The method for personalized content recommendation of smart TVs based on deep learning according to claim 1, characterized in that, The determination of the service level includes: When the playback risk indicator reaches the playback protection threshold, the remaining duration of the playback buffer in the playback state is lower than the protection entry threshold, or the player status flag indicates that it is in the first frame loading, dragging, channel switching, or resolution switching stage, it is determined to be a playback continuity protection service; when it does not belong to the playback continuity protection service and the instant intent strength reaches the instant intent threshold, it is determined to be an instant intent response service; when it does not belong to the aforementioned two types of services, the resource availability reaches the recommended threshold, and the remaining duration of the playback buffer reaches the recommended buffer threshold, it is determined to be a regular personalized recommendation service. If the number of interactive events within a preset time window is lower than the threshold for the number of interactions, or if there are no voice, remote control, search, or content confirmation operations for a preset duration, it is determined to be a background update service. When multiple conditions are met at the same time, the service level is determined in the following order: playback continuity protection service, instant intent response service, regular personalized recommendation service, and background update service.

6. The method for personalized content recommendation of smart TVs based on deep learning according to claim 5, characterized in that, The terminal resource status includes intelligent processor utilization, dynamic random access memory bandwidth, network throughput, and recommended cache availability. Resource availability is determined based on the normalized values ​​of intelligent processor idle rate, dynamic random access memory bandwidth margin, network throughput margin, and cache margin, according to preset weights. During continuous playback protection services, pause full recommendation calculation, media resource prefetching, model parameter downloading, and batch content feature generation; set video stream transmission priority and limit network bandwidth and dynamic random access memory usage for recommendation tasks. In the instant intent response business, read local content metadata and updated content feature data, call the simplified content sorting model, and only request the cover and metadata; In regular personalized recommendation operations, recommendation calculations are performed in batches according to the idle rate of the intelligent processor; in background update operations, long-term preference updates, log uploads, model parameter downloads, or batch content feature generation are performed.

7. The method for personalized content recommendation of smart TVs based on deep learning according to claim 4, characterized in that, The combined user preference features, updated content feature data, and viewing environment features are input into the content ranking model to obtain an initial score for candidate content. A time-effect decay coefficient is determined based on the time difference between the publication time or event start time of the candidate content and the current time, wherein the publication time adopts exponential decay and the event start time adopts segmented decay before and after the event start. The initial score of the candidate content is then corrected using the time-effect decay coefficient. When the intensity of the instant intent reaches the instant intent threshold and a voice operation, search operation, channel selection operation, or content confirmation operation is detected, a matching score is determined based on the similarity between the candidate content and the voice operation, search term, or operation target. The matching score is then weighted with the corrected initial score of the candidate content according to a preset weight to obtain a candidate content ranking score. The filtering is based on copyright region, subscription permissions, terminal encoding capabilities, age rating, children's mode, playback buffer, and intelligent processor, dynamic random access memory bandwidth, and remaining network resources. In children's mode, the age rating of candidate content must not exceed a preset level. The resource budget includes at least one of the following: intelligent processor budget, dynamic random access memory bandwidth budget, network bandwidth budget, and cache space budget.

8. The method for personalized content recommendation of smart TVs based on deep learning according to claim 1, characterized in that, The storage module includes a result buffer, a metadata buffer, and a media resource buffer. It obtains the metadata and media resources of candidate content from the content server, local content index, or cache management module, saves the recommendation results, content identifiers, and display priorities in the result buffer, saves the title, category, duration, rating, playback address, age rating, content status, and cover address in the metadata buffer, and saves the cover, trailer, low bitrate preview clips, and content segments in the media resource buffer. When the smart TV is powered on, woken up, resumed from standby, the remaining playback time of the current video is less than a preset end threshold, or the channel list or recommendation entry is opened a preset number of times within a preset time window, a simplified recommendation program is loaded, which includes local recommendation result reading, content index reading, and a simplified content sorting model execution program. The recommended content is first displayed using the result buffer and metadata buffer. When the remaining network bandwidth is higher than the preset loading threshold, media resources are loaded asynchronously according to the display priority, and the loading bitrate and concurrency are limited. When the remaining network bandwidth is less than or equal to the preset loading threshold, the loading of low display priority media resources is paused.

9. The method for personalized content recommendation of smart TVs based on deep learning according to claim 8, characterized in that, The system continuously collects data on intelligent processor utilization, dynamic random access memory bandwidth, system temperature, network throughput, playback buffer and cache utilization, and determines recommended cache availability based on cache utilization. When the service level upgrade / downgrade conditions are met for a preset number of sampling periods, the service level is switched. The playback continuity protection service uses a protection entry threshold and a protection exit threshold, where the protection entry threshold is less than the protection exit threshold. When the intelligent processor utilization rate reaches the intelligent processor protection threshold or the available computing time is lower than the computing time threshold, switch to the simplified content sorting model or rule sorting method. When the network throughput is lower than the network availability threshold or continuous packet loss occurs, read the result buffer and metadata buffer. When the available dynamic random access memory is lower than the dynamic random access memory protection threshold, release the preview segments in the media resource buffer and reduce the number of candidates according to the available dynamic random access memory, but retain the preset minimum number of candidates. When the system temperature reaches the temperature protection threshold, reduce the recommended computing batch and extend the background task polling cycle. The calculation batch and polling cycle are restored when the system temperature is below the preset recovery threshold and the preset sampling period is met continuously.

Images