A multi-granularity time period feature-based intelligent television recommendation method and system

Abstract

The application discloses a kind of intelligent television recommendation method and system based on multi-granularity time period feature.The present application updates family member user portrait in real time by multi-modal data fusion, and is based on multi-granularity time period feature modeling system, adopts multi-level time period division and dynamic behavior density cutting technology to generate personalized recommendation strategy.It has the following advantages: fine modeling in time dimension;constructing family user portrait matrix, integrating face recognition, speech characteristics and behavior pattern data, distinguishing family member identity in real time and updating individual preference;designing time period-scene-device three-dimensional adaptation algorithm based on Transformer attention mechanism, by dynamically adjusting recommendation feature weight, achieving high coordination of content recommendation and hardware resources.

Description

Technical Field

[0001] This invention relates to the field of content recommendation algorithms, and more specifically, to a smart TV recommendation method and system based on multi-granularity time-segment features. Background Technology

[0002] Current smart TV recommendation systems primarily rely on collaborative filtering algorithms to analyze users' historical behavior data to generate recommendation lists, and on content recommendation algorithms to match video metadata with user interest tags. Existing data collection methods are mostly based on single behavioral metrics such as remote control clicks and playback completion rates, while model training employs offline batch updates, such as daily or weekly updates.

[0003] The aforementioned recommendation system has the following problems:

[0004] Recommendation models rely on offline batch training, such as daily batch runs using Spark MLlib, which cannot respond to the high-frequency interactive behaviors of TV users, such as continuous channel switching and voice search. They also lack online learning mechanisms, making it impossible to capture real-time changes in user interests, such as switching to related topics after watching breaking news.

[0005] Coarse-grained modeling of home scenes: The ineffective use of multimodal data from the TV, such as cameras, microphones, and ambient light sensors, makes it impossible to accurately identify family members and scene features such as multiple viewers and night mode.

[0006] Inefficient advertising and value-added service delivery: Ad placement relies solely on static user profile tags, failing to consider time-of-day and context-specific characteristics. Value-added service promotion lacks a dynamic trigger mechanism. Recommendation strategies fail to take into account the characteristics of large TV screens, resulting in a mismatch between content format and device performance. Summary of the Invention

[0007] This invention overcomes the technical problems of existing technologies in real-time capture of interest changes, insufficient multimodal data fusion, and inefficient push notifications. It provides a smart TV recommendation method and system based on multi-granularity time-segment features, which is expected to solve the problems existing in the prior art.

[0008] To address the aforementioned technical problems, one aspect of the present invention provides a smart TV recommendation method based on multi-granularity time-segment features.

[0009] A smart TV recommendation method based on multi-granularity time-segment features includes the following steps:

[0010] S1 divides time into multi-level time periods and implements dynamic time period segmentation based on a neural network model;

[0011] S2 integrates data collected from television and uses a real-time data stream processing framework to extract data, achieving feature extraction and multimodal data fusion.

[0012] S3 uses the Transformer attention mechanism to dynamically allocate recommendation feature weights and combines them with adaptation strategies to achieve precise coordination between content recommendation and hardware resources, and then provides recommendation results.

[0013] The core technology of this invention lies in updating the user profile of family members in real time through multimodal data fusion, and generating personalized recommendation strategies based on a multi-granular time period feature modeling system and using multi-time period division and dynamic behavior density cutting technology.

[0014] A further technical solution is that the multi-level time segmentation specifically includes the following steps:

[0015] S11 Level 1 Classification: Based on the calendar API, weekday and weekend tags are obtained and dynamically calibrated by combining user historical behavior data;

[0016] S12 Level 2 Division: Preset fixed time intervals;

[0017] S13 Three-Level Segmentation: By analyzing dense areas in user behavior sequences, dynamic time-segmentation at the minute level is achieved.

[0018] By employing a multi-granularity time-segmentation strategy, combining a primary segmentation of weekday and weekend tags dynamically calibrated using calendar APIs, a secondary segmentation of predefined fixed intervals covering morning, midday, and prime time periods, and a tertiary segmentation of dynamic time periods at the minute level based on an LSTM neural network model to analyze dense user behavior areas, refined time-dimensional modeling can be achieved. This improves recommendation granularity from the traditional hourly level to the minute level. Time-segment feature coverage, the overlap between recommended content and actual user viewing time, and user dwell time are all enhanced.

[0019] A further technical solution involves integrating television data collection, specifically including the following steps:

[0020] S21 integrates multiple heterogeneous data sources:

[0021] S22 employs a hierarchical stream processing architecture to handle various types of heterogeneous data;

[0022] The processed multimodal features of S23 are input into the recommendation model via multiple paths.

[0023] A further technical solution is that the integration of multiple heterogeneous data sources specifically includes one or more of the following data:

[0024] TV interaction logs include user channel switching frequency, single content viewing duration, search keyword records, and playback completion rate;

[0025] Environmental sensor data: The TV camera captures ambient light intensity and user facial features, the microphone collects ambient volume fluctuations and voice commands, and the infrared sensor identifies the viewing distance.

[0026] It integrates with the calendar API to obtain information on holidays and work schedule adjustments, combines live broadcast times and participating team data from the sports event database, and synchronizes the popularity index and series update times from the film and television content library.

[0027] By constructing a family user profile matrix and integrating facial recognition, voice features, and behavioral pattern data, it is possible to distinguish the identities of family members and update individual preferences in real time, thus achieving more accurate recommendations.

[0028] A further technical solution is that the use of a hierarchical stream processing architecture to process multiple types of heterogeneous data specifically includes the following steps;

[0029] The S221 data acquisition layer uses the Flume distributed log collection system to collect multimodal data in parallel with high throughput.

[0030] The S222 data buffer layer uses Kafka message queues to partition and asynchronously distribute the raw data.

[0031] The S223 streaming computation layer performs windowed processing, including data cleaning, feature encoding, and time-series aggregation, to generate a structured feature matrix.

[0032] The S224 feature storage layer writes the processed features into a distributed database, supporting low-latency queries and incremental model training.

[0033] A further technical solution involves inputting the processed multimodal features into the recommendation model via multiple paths, specifically including the following steps:

[0034] The processed multimodal features are input into the recommendation model through the time-period prediction model and / or recommendation strategy engine;

[0035] The time period prediction model specifically involves inputting time feature labels into an LSTM network and outputting minute-level dynamic time period division results.

[0036] The recommendation strategy engine specifically embeds time period labels, device status, and user profile features into the Transformer model, and calculates the three-dimensional weights of time period, scene, and device through a multi-head attention mechanism.

[0037] A further technical solution involves dynamically allocating recommendation feature weights based on the Transformer attention mechanism, specifically including the following steps:

[0038] S31 Input current time period characteristics, user profile, and device status;

[0039] S32 transforms multi-level time period labels into one-hot vectors, which are then mapped to a low-dimensional space through an embedding layer.

[0040] S33 integrates demographic characteristics and interest tags;

[0041] S34 device status embedding;

[0042] S35 outputs the weights of each recommended feature.

[0043] The Transformer model increases the weight of time-period features, ultimately improving the click-through rate of ads during prime time, the accuracy of contextual ad matching, and the efficiency of ad display.

[0044] A further technical solution is that the combined adaptation strategy specifically combines a three-dimensional adaptation strategy of time period-scene-device.

[0045] The combined time-scene-device 3D adaptation strategy specifically includes:

[0046] One or more of the following: time-segment-scenario collaboration strategy, time-segment-device collaboration strategy, and scenario-device collaboration strategy.

[0047] Based on the Transformer attention mechanism, a three-dimensional adaptation algorithm for time period, scene, and device is designed. By dynamically adjusting the weights of recommendation features, it accurately matches time period features, scene status, and device performance parameters, thereby achieving a high degree of synergy between content recommendation and hardware resources.

[0048] The present invention also provides a smart TV recommendation system based on multi-granularity time period features.

[0049] A smart TV recommendation system based on multi-granularity time period features includes a multi-granularity time period segmentation module, a multi-modal data fusion module, and a time period-sensitive recommendation strategy engine.

[0050] The multi-granularity time period segmentation module is used to divide time into multi-level time periods and model the time dimension of user behavior.

[0051] The multimodal data fusion module is used to integrate television-collected data;

[0052] The time-sensitive recommendation strategy engine is used to dynamically allocate recommendation feature weights based on the Transformer attention mechanism, and combined with the adaptation strategy, to achieve precise coordination between content recommendation and hardware resources, and to provide recommendation results.

[0053] Compared with existing technologies, this invention has at least the following beneficial effects: The core technology of this invention lies in updating family member user profiles in real time through multimodal data fusion, and generating personalized recommendation strategies based on a multi-granularity time period feature modeling system using multi-level time period division and dynamic behavior density segmentation technology. It has the following advantages: First, through a multi-granularity time period division strategy, combined with a first-level division of weekday and weekend tags dynamically calibrated by calendar API, a second-level predefined fixed interval division covering morning, noon, and prime time periods, and a third-level minute-level dynamic time period segmentation based on an LSTM neural network model to analyze dense user behavior areas, refined modeling of the time dimension is achieved; second, a family user profile matrix is ​​constructed, integrating facial recognition, voice features, and behavioral pattern data to distinguish family member identities and update individual preferences in real time; finally, a three-dimensional adaptation algorithm based on the Transformer attention mechanism is designed for time period-scene-device, dynamically adjusting recommendation feature weights to accurately match time period features, scene states, and device performance parameters, thereby achieving a high degree of synergy between content recommendation and hardware resources. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating Embodiment 1 of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0056] Example 1

[0057] A smart TV recommendation method based on multi-granularity time-segment features, see [link to relevant documentation]. Figure 1 This includes the following steps:

[0058] S1 divides time into multi-level time periods and implements dynamic time period segmentation based on a neural network model;

[0059] In a preferred embodiment, the multi-level time segmentation specifically includes the following steps:

[0060] S11 Level 1 Classification: Based on the calendar API, weekday and weekend tags are obtained and dynamically calibrated by combining user historical behavior data;

[0061] S12 Level 2 Division: Preset fixed time intervals;

[0062] For example, the division of fixed time periods can refer to the following scheme:

[0063] Early morning 5:00-7:00 1. Morning News 2. Health and Wellness Programs Breakfast time 7:00-9:00 1. Educational content (K-12 courses, vocational skills) 2. Lifestyle services (food tutorials, healthy recipes) Morning period 9:00-12:00 1. Online vocational training courses 2. In-depth industry reports (such as technology and finance) lunch break 12:00-14:00 1. Short video entertainment 2. Lightweight learning (knowledge quizzes, short science videos) Afternoon 14:00-18:00 1. On-demand movies and TV shows 2. Live sports broadcasts Golden Hour 18:00-22:00 1. Family entertainment dramas 2. Parent-child interactive games Late night 22:00-24:00 1. Documentaries (history, nature) 2. Meditation and relaxation courses early morning hours 0:00-5:00 1. System maintenance and upgrades 2. Security monitoring footage (requires an external camera) 3. Emergency notification push notifications (earthquakes, extreme weather) 4. Sleep data monitoring (requires an external smart wearable device)

[0064] S13 Three-Level Segmentation: By analyzing dense areas in user behavior sequences, dynamic time-segmentation at the minute level is achieved.

[0065] For example, the three-level division can be constructed by integrating user behavior sequences, environmental sensor data, and external event labels, such as holidays and live sports events, to build a multi-dimensional temporal feature matrix, which is then input into a two-layer LSTM network to train a behavior-dense probability prediction model.

[0066] The model uses 15-minute time slices as units to analyze user viewing time, channel switching frequency and other behavioral characteristics. Combined with environmental data such as light intensity and volume, it outputs the probability value of dense time periods. Through dynamic threshold judgment and expert rule calibration, continuous high-probability time slices are merged into minute-level dynamic time periods, such as the weekday evening film and television period from 19:30 to 21:15.

[0067] Specifically, the three-level division includes the following steps:

[0068] Input data: User behavior sequences include viewing duration and channel switching frequency; environmental data includes light intensity and volume; external events include live sports broadcasts and holidays.

[0069] Model Construction: A two-layer LSTM network is used, with 15-minute time slices as the unit, to output dense probability values ​​for time periods. Continuous high-probability time slices are dynamically merged, such as those with a probability > 0.8, to generate minute-level time periods, such as "movie and television time period 19:30-21:15".

[0070] Calibration mechanism: By using an expert rule base, such as extending holiday periods by 30%, the model output is corrected to ensure that the time period division conforms to the actual scenario.

[0071] S2 integrates data collected from television and uses a real-time data stream processing framework to extract data, achieving feature extraction and multimodal data fusion.

[0072] In a preferred embodiment, integrating television-collected data specifically includes the following steps:

[0073] S21 integrates multiple heterogeneous data sources:

[0074] For example, integrating multiple heterogeneous data sources, specifically including one or more of the following data:

[0075] TV interaction logs include user channel switching frequency, single content viewing duration, search keyword records, and playback completion rate;

[0076] Environmental sensor data: The TV camera captures ambient light intensity and user facial features, the microphone collects ambient volume fluctuations and voice commands, and the infrared sensor identifies the viewing distance.

[0077] It integrates with the calendar API to obtain information on holidays and work schedule adjustments, combines live broadcast times and participating team data from the sports event database, and synchronizes the popularity index and series update times from the film and television content library.

[0078] S22 employs a hierarchical stream processing architecture to handle various types of heterogeneous data;

[0079] In a preferred embodiment, the process of using a hierarchical stream processing architecture to process multiple types of heterogeneous data specifically includes the following steps;

[0080] The S221 data acquisition layer uses the Flume distributed log collection system to collect multimodal data in parallel with high throughput.

[0081] The S222 data buffer layer uses Kafka message queues to partition and asynchronously distribute the raw data.

[0082] The S223 streaming computation layer performs windowed processing, including data cleaning, feature encoding, and time-series aggregation, to generate a structured feature matrix.

[0083] The S224 feature storage layer writes the processed features into a distributed database, supporting low-latency queries and incremental model training.

[0084] The processed multimodal features of S23 are input into the recommendation model via multiple paths.

[0085] The processed multimodal features are input into the recommendation model via multiple paths, specifically including the following steps:

[0086] The processed multimodal features are input into the recommendation model through the time-period prediction model and / or recommendation strategy engine;

[0087] For example, the time period prediction model specifically involves inputting time features into an LSTM network and outputting minute-level dynamic time period segmentation results.

[0088] The time features include historical behavior sequences, trends in ambient light intensity, and holiday labels.

[0089] For example, the recommendation strategy engine specifically embeds time period labels, device status, and user profile features into the Transformer model, and calculates the three-dimensional weights of time period, scene, and device through a multi-head attention mechanism.

[0090] S3 uses the Transformer attention mechanism to dynamically allocate recommendation feature weights and combines them with adaptation strategies to achieve precise coordination between content recommendation and hardware resources, and then provides recommendation results.

[0091] In a preferred embodiment, the dynamic allocation of recommendation feature weights based on the Transformer attention mechanism specifically includes the following steps:

[0092] S31 Input current time period characteristics, user profile, and device status;

[0093] S32 transforms multi-level time period labels into one-hot vectors, which are then mapped to a low-dimensional space through an embedding layer.

[0094] S33 integrates demographic characteristics and interest tags;

[0095] S34 device status embedding;

[0096] For example, the device status includes screen brightness (0-100) and external device type (HDMI1 / HDMI2 / USB).

[0097] S35 outputs the weights of each recommended feature.

[0098] In a preferred embodiment, the combined adaptation strategy specifically refers to a three-dimensional adaptation strategy combining time period, scene, and device.

[0099] The combined time-scene-device 3D adaptation strategy specifically includes:

[0100] One or more of the following: time-segment-scenario collaboration strategy, time-segment-device collaboration strategy, and scenario-device collaboration strategy.

[0101] For example, the time-scene adaptation strategy is as follows:

[0102] Based on user behavior patterns, recommendation weights are predefined for three typical time periods: Morning: News content accounts for 70% of the weight, and educational content accounts for 30%. Evening: Entertainment content accounts for 60%, and family interaction content accounts for 40%. Night: Documentaries account for 80%, and relaxation and sleep aid content accounts for 20%. The weight allocation is then dynamically adjusted and optimized based on user profiles. When users are under 18 years old (i.e., minors), the weight of educational content across all time periods is uniformly increased by 20%, with the weight of educational content in the morning period increasing from 30% to 50%.

[0103] Example of applying the time-scene adaptation strategy:

[0104] Recommended content for a household user during prime time (7:00 PM):

[0105] Basic weighting: Entertainment 60% + Family 40%;

[0106] If there are minor children, the weight of educational content will be increased by an additional 20%.

[0107] Final recommended allocation: Entertainment 50% + Family 30% + Education 20%;

[0108] The specific time-time-device coordination strategy is as follows:

[0109] By sensing the status of devices, intelligent linkage between time periods and hardware resources can be achieved.

[0110] The time-based-device collaboration strategy is applied in the following scenarios:

[0111] (1) Night mode optimization: When the current time period is detected to be night (22:00-24:00) and the screen brightness exceeds 50%, the brightness will be automatically adjusted to 30%, and the low volume mode (volume 30%) will be forcibly switched, and sleep-aiding content (white noise, meditation videos) will be recommended first.

[0112] (2) External device response: When an external device is detected as a game console, immediately switch to game mode (reduce latency and improve image quality), increase the proportion of game live streaming in the recommended content to 40% priority, and automatically close other irrelevant applications to free up system resources.

[0113] (3) Multi-device collaboration: When a home theater device is detected, the surround sound mode is automatically turned on, and the recommended content is prioritized to match cinema-grade audio-visual resources. When a fitness device is detected, exercise courses are pushed and heart rate data is displayed simultaneously, and the screen display is adjusted to eye protection mode.

[0114] This strategy is deeply coupled with time-of-day characteristics and the Transformer attention mechanism to form a closed-loop optimization of "scene-driven device, device feed-back scene", ultimately achieving full-dimensional collaboration of the smart TV system.

[0115] The specific scenario-device collaboration strategy is as follows:

[0116] Based on multimodal data fusion and dynamic weight adjustment of the Transformer model, the scene-device adaptation strategy identifies specific home scenes and links them with device status to achieve precise coordination between content recommendation and hardware resources.

[0117] The scenario-device collaboration strategy is applied in the following scenarios:

[0118] (1) Family Gathering Scene: The scene is identified as a family gathering scene by detecting facial features of multiple people through a camera and capturing ambient volume fluctuations (such as laughter and conversation) through a microphone, combined with time period characteristics (such as prime weekend hours). Third-party data (such as holiday markers) is used to further calibrate the scene labels. Device linkage includes: automatically switching to living room surround sound mode and recommending multi-person interactive content.

[0119] (2) Fitness Scene Adaptation: Detects fitness devices (such as smart treadmills and heart rate bracelets) via Bluetooth / USB connection, combined with user behavior data (such as usage records of fitness apps). Environmental sensors (such as cameras dynamically capturing user movement postures) assist in confirming the scene. Device linkage includes: switching the screen to eye protection mode, recommending fitness courses (such as HIIT training and yoga tutorials), split-screen display of real-time heart rate and calorie consumption data, automatic playback of dynamic music from external speakers, and dynamic volume adjustment based on exercise intensity.

[0120] (3) Child-friendly viewing scenarios: The child's identity (age <12 years old) is determined by facial recognition technology, combined with time period characteristics (e.g., after school, 14:00-18:00). The microphone captures high-pitched speech (e.g., cartoon character dialogue) to further verify the scenario. Device linkage includes: enabling child content filtering mode, restricting access to inappropriate content, automatically reducing screen brightness to 40%, limiting volume to 50%, and enabling blue light protection mode.

[0121] We recommend educational animations and programs. If we detect that you are watching for more than 30 minutes, we will pop up an eye protection reminder and pause the playback.

[0122] By using multi-granular time-segment feature modeling and a three-dimensional adaptation strategy, a three-in-one model of "precise time-segment reach + deep scenario integration + device value conversion" for advertising and value-added services has been achieved.

[0123] Example 2

[0124] A smart TV recommendation system based on multi-granularity time period features includes a multi-granularity time period segmentation module, a multi-modal data fusion module, and a time period-sensitive recommendation strategy engine.

[0125] The multi-granularity time period segmentation module is used to divide time into multi-level time periods and model the time dimension of user behavior.

[0126] The multimodal data fusion module is used to integrate television-collected data;

[0127] The time-sensitive recommendation strategy engine is used to dynamically allocate recommendation feature weights based on the Transformer attention mechanism, and combined with the adaptation strategy, to achieve precise coordination between content recommendation and hardware resources, and to provide recommendation results.

[0128] Although the invention has been described herein with reference to illustrative embodiments, it should be understood that many other modifications and implementations can be devised by those skilled in the art, which will fall within the scope and spirit of the principles disclosed herein. More specifically, various variations and modifications can be made to the components and / or layout of the subject matter combination within the scope of this disclosure. Besides variations and modifications to the components and / or layout, other uses will be apparent to those skilled in the art.

Claims

1. A smart TV recommendation method based on multi-granularity time-segment features, characterized in that, Includes the following steps: S1 divides time into multiple time periods and implements dynamic time period segmentation based on a neural network model; S2 integrates television-collected data and employs a real-time data stream processing framework to extract data, achieving feature extraction and multimodal data fusion. The S2 integration of television-collected data specifically includes the following steps: S21 integrates multiple heterogeneous data sources: S22 employs a hierarchical stream processing architecture to handle various types of heterogeneous data; S23 processed multimodal features are input into the recommendation model via multiple paths; S22 specifically includes the following steps; The S221 data acquisition layer uses the Flume distributed log collection system to collect multimodal data in parallel with high throughput. The S222 data buffer layer uses Kafka message queues to partition and asynchronously distribute the raw data. The S223 streaming computation layer performs windowed processing, including data cleaning, feature encoding, and time-series aggregation, to generate a structured feature matrix. The S224 feature storage layer writes the processed features into a distributed database, supporting low-latency queries and incremental model training. S3 uses the Transformer attention mechanism to dynamically allocate recommendation feature weights and combines them with adaptation strategies to achieve precise coordination between content recommendation and hardware resources, and then provides recommendation results.

2. The smart TV recommendation method based on multi-granularity time-segment features as described in claim 1, characterized in that, The S1 process for dividing time into multi-level time periods specifically includes the following steps: S11 Level 1 Classification: Based on the calendar API, weekday and weekend tags are obtained and dynamically calibrated by combining user historical behavior data; S12 Level 2 Division: Preset fixed time intervals; S13 Three-Level Segmentation: By analyzing dense areas in user behavior sequences, dynamic time-segmentation at the minute level is achieved.

3. The smart TV recommendation method based on multi-granularity time-segment features as described in claim 1, characterized in that, The integration of multiple heterogeneous data sources specifically includes one or more of the following data types: TV interaction logs include user channel switching frequency, single content viewing duration, search keyword records, and playback completion rate; Environmental sensor data: The TV camera captures ambient light intensity and user facial features, the microphone collects ambient volume fluctuations and voice commands, and the infrared sensor identifies the viewing distance. It integrates with the calendar API to obtain information on holidays and work schedule adjustments, combines live broadcast times and participating team data from the sports event database, and synchronizes the popularity index and series update times from the film and television content library.

4. The smart TV recommendation method based on multi-granularity time-segment features as described in claim 1, characterized in that, The processed multimodal features are input into the recommendation model via multiple paths, specifically including the following steps: The processed multimodal features are input into the recommendation model through the time-period prediction model and / or recommendation strategy engine; The time period prediction model specifically involves inputting time feature labels into an LSTM network and outputting minute-level dynamic time period division results. The recommendation strategy engine specifically embeds time period labels, device status, and user profile features into the Transformer model, and calculates the three-dimensional weights of time period, scene, and device through a multi-head attention mechanism.

5. The smart TV recommendation method based on multi-granularity time-segment features as described in claim 1, characterized in that, The dynamic allocation of recommendation feature weights based on the Transformer attention mechanism specifically includes the following steps: S31 Input current time period characteristics, user profile, and device status; S32 transforms multi-level time period labels into one-hot vectors, which are then mapped to a low-dimensional space through an embedding layer. S33 integrates demographic characteristics and interest tags; S34 device status embedding; S35 outputs the weights of each recommended feature.

6. The smart TV recommendation method based on multi-granularity time-segment features as described in claim 1, characterized in that, The specific adaptation strategy is to combine a three-dimensional adaptation strategy of time period-scene-device. The combined time-scene-device 3D adaptation strategy specifically includes: One or more of the following: time-segment-scenario collaboration strategy, time-segment-device collaboration strategy, and scenario-device collaboration strategy.

7. A smart TV recommendation system based on multi-granularity time-segment features, implementing the smart TV recommendation method based on multi-granularity time-segment features as described in any one of claims 1-6, characterized in that, This includes a multi-granularity time period segmentation module, a multi-modal data fusion module, and a time-sensitive recommendation strategy engine; The multi-granularity time period segmentation module is used to divide time into multi-level time periods and model the time dimension of user behavior. The multimodal data fusion module is used to integrate television-collected data; The time-sensitive recommendation strategy engine is used to dynamically allocate recommendation feature weights based on the Transformer attention mechanism, and combined with the adaptation strategy, to achieve precise coordination between content recommendation and hardware resources, and to provide recommendation results.

Images