An IPTV set-top box video content intelligent recommendation method based on deep learning

Abstract

The application relates to the technical field of video on demand, and discloses an IPTV set top box video content intelligent recommendation method based on deep learning, which comprises the following steps: collecting user identification and program identification at fixed time intervals to obtain K piecewise matrices; loading a dilution disc, mapping the i-th program to the dilution disc to obtain the coordinates of the i-th program in the dilution disc; determining the mapping radius corresponding to the k-th piecewise matrix, and setting a geometric dilution factor; determining the interaction weight of the k-th piecewise matrix, and generating the equivalent quality of the corresponding i-th program in combination with the geometric dilution factor; collecting and extracting S program identifications of the target user time sequence at fixed time intervals, and mapping to the dilution disc to obtain user potential energy; calculating the potential energy score of the i-th program according to the user potential energy and the coordinates of the i-th program; and performing intelligent recommendation processing according to the potential energy score of each program.

Description

Technical Field

[0001] This invention relates to the field of video-on-demand technology, and more specifically, to a method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning. Background Technology

[0002] With the widespread deployment of IPTV (Interactive Network Television) set-top boxes, user viewing behavior data is continuously collected and used to drive personalized recommendation systems. Because IPTV recommendation systems operate in environments with large screens, multi-user sharing, and remote control, users lack the means to actively express their interests. Therefore, they primarily rely on the collection and analysis of implicit behavioral data such as user engagement with programs, switching between programs, and playback at different times. To ensure real-time recommendations and user experience, most current systems employ a sliding time window mechanism for structured management of user behavior. This involves segmenting time-series behavior into minute-level or hour-level segments and then constructing an interaction matrix between users and content for recommendation calculations. This allows the system to dynamically incorporate new behaviors over time and supports both offline and online collaborative updates.

[0003] Under the sliding time window mechanism, users' latest viewing behaviors are concentrated in the current time slice, forming a natural hierarchical structure with the accumulated behavior distribution in previous time windows. When a new playback behavior occurs, a densely active block forms in the corresponding region of the current time window in the matrix. The user and content interaction values ​​(such as playback duration) in this region are much higher than in historical regions. Due to the high density and frequency of new behaviors, their weight in the overall matrix rises rapidly. At the same time, the interaction information accumulated in older time windows gradually loses its explicit expressive power. This structural difference in the distribution of new and old data causes recommendation algorithms to heavily favor recent content when performing similarity calculations, content ranking, or interest inference. Specifically: 1. Popular content takes center stage due to the abundance of new content, while programs with historical interest are squeezed out of the candidate pool.

[0004] 2. The system frequently responds to short-term behaviors, causing the recommendation list to change drastically in a short period of time.

[0005] In summary, existing IPTV recommendation systems suffer from technical problems, including matrix weight imbalances caused by uneven distribution of behavior over time periods. This continuously impacts the system's response stability, personalized expression capabilities, and the efficiency of system resource scheduling. Summary of the Invention

[0006] This invention provides a deep learning-based intelligent recommendation method for IPTV set-top box video content, addressing the problem of matrix weight imbalance caused by uneven distribution of behavior over time periods. This imbalance continuously affects the response stability, personalized expression capabilities, and resource scheduling efficiency of the recommendation system.

[0007] This invention provides a deep learning-based intelligent recommendation method for video content on IPTV set-top boxes, comprising: S1, collect user identifiers and program identifiers at fixed time intervals to obtain K segmentation matrices; S2, Load the dilution disk, map the i-th program to the dilution disk, and obtain the coordinates of the i-th program within the dilution disk; S3, determine the mapping radius corresponding to the k-th piecewise matrix and set the geometric dilution factor; where 1≤k≤K, and k is a positive integer; S4, determine the interaction weights of the k-th slice matrix, and combine them with the geometric dilution factor to generate the equivalent quality of the i-th program. S5: Collect and extract S types of program identifiers of the target user's time sequence at fixed time intervals, and map them to the dilution disk to obtain the user's potential energy. S6. Calculate the potential energy score of the i-th program based on the user's potential energy and the coordinates of the i-th program. S7 intelligently recommends programs based on their potential scores.

[0008] Furthermore, user identifiers and program identifiers are collected at fixed time intervals to obtain K segment matrices, including: Within a preset time period, user identifiers and program identifiers are collected at fixed time intervals; Set the number of segments K to divide the preset time period into K sub-time periods; Determine the cumulative dwell time of the u-th user on the i-th program within each sub-time period to construct K segmentation matrices; where the element in the u-th row and i-th column of the k-th segmentation matrix represents the cumulative dwell time of the u-th user on the i-th program within the k-th sub-time period; 1≤k≤K, where k is an integer.

[0009] Furthermore, a dilution disk is added, mapping the i-th program to the dilution disk, and the coordinates of the i-th program within the dilution disk are obtained, including: Load dilution disk D, ;in, This represents the coordinates of any point within the dilution disk; Calculate the mapping radius for the i-th type of program in the k-th segment matrix. , in, Indicates the preset radial step size; Processed using the i-th program mapping via a consistent hash function: The mapping angle of the i-th program is obtained. , ; Generate the coordinates of the i-th program within the dilution disk. .

[0010] Furthermore, determine the mapping radius corresponding to the k-th piecewise matrix and set the geometric dilution factor, including: Based on mapping radius Calculate the mapped area in the dilution disk. , ;in, Represents the hyperbolic cosine function; Based on the mapped area Set the geometric dilution factor for the k-th fragment matrix. , ;in, It is a constant greater than 0.

[0011] Furthermore, the interaction weights of the k-th fragment matrix are determined, and the equivalent quality of the corresponding i-th program is generated by combining the geometric dilution factor, including: Extract the interaction weights of the i-th type of program in the k-th segment matrix, including: Get the cumulative dwell time of each user on the i-th type of program within the k-th sub-time period; The sum of the cumulative dwell time is used as the interaction weight of the i-th type of program in the k-th segment matrix. ; The interaction weights are diluted using the geometric dilution factor of the k-th partition matrix, as follows: in, This represents the dilution interaction weight of the i-th type of program in the k-th segment matrix; The sum of the diluted interaction weights corresponding to the K fragment matrices is used as the equivalent quality of the i-th type of program.

[0012] Furthermore, the user's potential energy includes the potential energy source strength and potential energy coordinates; Within a preset time period, determine the cumulative dwell time of the target user on the j-th type of program within each sub-time period to extract the corresponding F segment matrices; 1≤j≤S, 1≤F≤K, where j and F are both integers; Extract the geometric dilution factor of the f-th fragment matrix, as well as the cumulative dwell time of the j-th program in the f-th fragment matrix and the coordinates of the j-th program on the dilution disk; The product of the geometric dilution factor of the f-th segment matrix and the cumulative dwell time of the j-th program in the f-th segment matrix is ​​used as the coordinate weighting coefficient of the potential energy coordinate; where 1≤f≤F, and f is a positive integer; The coordinates of the j-th program in the dilution disk are weighted based on the coordinate weighting coefficients to obtain the weighted coordinates of the j-th program in the dilution disk in the f-th segment matrix. The sum of the weighted coefficients of the coordinates of the target user for each program is used as the potential energy source strength; The ratio of the sum of the weighted coordinates of the target user for each program to the potential energy source strength is used as the potential energy coordinate.

[0013] Furthermore, based on the user's potential energy and the coordinates of the i-th type of program, the potential energy score of the i-th type of program is calculated, including: Calculate the geodesic distance between the potential energy coordinates and the coordinates of the i-th program; The denominator for distance attenuation is determined based on the geodetic distance; Based on the potential energy source strength, the equivalent quality of the i-th program, and the preset potential constant, the potential energy score of the i-th program is calculated as follows: in, Let represent the potential score of the i-th type of program. This indicates that the i-th type of program has a strong potential energy source. This represents the equivalent quality of the i-th type of program. The denominator represents the distance attenuation, and G represents a preset constant used for unifying dimensions.

[0014] Furthermore, intelligent recommendation processing is performed based on the potential score of each program, including: H programs are selected from high to low based on their potential scores to form a candidate set; For each program in the candidate set, calculate the corresponding diversity penalty coefficient based on the prior playback volume; The product of the potential score and the diversity penalty coefficient for each program is used as the final ranking score for the corresponding program. Based on the final ranking scores, select Q programs from high to low to form a recommendation list for the target user.

[0015] The beneficial effects of this invention are as follows: by introducing a dilution disk structure and a time-layered high-dimensional interaction matrix mapping mechanism, a unified expression and quantification of IPTV set-top box user behavior in the time, space and intensity dimensions are achieved; and by integrating user potential energy and program equivalent quality to construct a potential energy score model, the recommendation bias problem caused by uneven distribution of new and old behaviors is effectively solved, and the comprehensive performance of the recommendation system in terms of responsiveness, expression of user interest diversity and content exposure balance is improved. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0017] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0018] like Figure 1 As shown, a method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning includes: S1, collect user identifiers and program identifiers at fixed time intervals to obtain K segmentation matrices; S2, Load the dilution disk, map the i-th program to the dilution disk, and obtain the coordinates of the i-th program within the dilution disk; S3, determine the mapping radius corresponding to the k-th piecewise matrix and set the geometric dilution factor; where 1≤k≤K, and k is a positive integer; S4, determine the interaction weights of the k-th slice matrix, and combine them with the geometric dilution factor to generate the equivalent quality of the i-th program. S5: Collect and extract S types of program identifiers of the target user's time sequence at fixed time intervals, and map them to the dilution disk to obtain the user's potential energy. S6. Calculate the potential energy score of the i-th program based on the user's potential energy and the coordinates of the i-th program. S7 intelligently recommends programs based on their potential scores.

[0019] In one embodiment of the present invention, user identifiers and program identifiers are collected at fixed time intervals to obtain K segmentation matrices, including: Within a preset time period, user identifiers and program identifiers are collected at fixed time intervals; Set the number of segments K to divide the preset time period into K sub-time periods; Determine the cumulative dwell time of the u-th user on the i-th program within each sub-time period to construct K segmentation matrices; where the element in the u-th row and i-th column of the k-th segmentation matrix represents the cumulative dwell time of the u-th user on the i-th program within the k-th sub-time period; 1≤k≤K, where k is an integer.

[0020] It should be noted that user and program identifiers are continuously collected at fixed time intervals within a preset time period. The preset time period can be set according to the actual application scenario, such as 24 hours or 12 hours; the fixed time interval is the basic unit for dividing the preset time period, such as 30 minutes. Its purpose is to refine the recording of user viewing behavior along the time dimension, ensuring accurate differentiation of user interactions at different times. A number of segments, K, is set to evenly divide the entire preset time period into K consecutive sub-time periods. The duration of each sub-time period is equal to the total duration of the preset time period divided by K. For example, if the preset time period is 3 hours and K=6, each sub-time period is 30 minutes. This separates user behavior data at different times, forming a hierarchical structure along the time dimension, facilitating separate processing of new and old data. For each sub-time period, the cumulative dwell time of the u-th user on the i-th program is calculated. Cumulative dwell time refers to the total time user u spends watching program i within a given sub-time period. For example, if user u watches program i twice within a sub-time period, spending 10 minutes and 15 minutes respectively, the cumulative dwell time is 25 minutes. K segmented matrices are constructed based on this cumulative dwell time. Each segmented matrix corresponds to a sub-time period. The rows of the matrix represent users, and the columns represent programs. The element value in the u-th row and i-th column of the k-th segmented matrix represents the cumulative dwell time user u spends on program i within the k-th sub-time period. When user u does not watch program i within the k-th sub-time period, the value of this element is 0. The K segmented matrices constructed in this way clearly present the interaction intensity between users and programs within different time periods, providing a foundation for subsequent time-based weight balancing processing, enabling the system to selectively process user behavior data from the latest and historical time periods.

[0021] In one embodiment of the present invention, loading a dilution disk, mapping the i-th program to the dilution disk, and obtaining the coordinates of the i-th program within the dilution disk include: Load dilution disk D, ;in, This represents the coordinates of any point within the dilution disk; Calculate the mapping radius for the i-th type of program in the k-th segment matrix. , in, Indicates the preset radial step size; Processed using the i-th program mapping via a consistent hash function: The mapping angle of the i-th program is obtained. , ; Generate the coordinates of the i-th program within the dilution disk. .

[0022] It should be noted that the number k=1 is the number of the partition matrix closest to the current time. Then, the K partition matrices are sorted in order of time to obtain the sorting of each partition matrix.

[0023] It should be noted that loading the dilution disk is a fundamental step in mapping programs to a specific spatial structure. The loaded dilution disk is a Poincaré disk in negative curvature hyperbend space, mathematically defined as all disks satisfying x 2 Add y 2 Points less than 1 The Poincaré disk's core characteristic is that its area expands exponentially with its radius, providing a geometric basis for balancing program weights across different segment matrices. This means that the natural difference in spatial area dilutes the dense program weights in the newest segment matrix while preserving the inherent influence of programs from historical segment matrices. Calculating the mapping radius for the i-th program in the k-th segment matrix is ​​crucial for achieving time-layered mapping. Mapping radius. The calculation formula is: Here, k is the number of the segment matrix, corresponding to different sub-time periods; a smaller k value indicates a newer segment matrix. σ is a preset radial step size, a fixed spatial scale parameter. This results in programs from different segment matrices forming a layered concentric circle structure on the disk: the newest program (k=1) corresponds to the smallest radius and is located in the innermost layer of the disk; as the k value increases, the corresponding program's segment matrix gradually becomes older, and its radius increases accordingly, distributing it in the outer layers of the disk. Mapping the i-th program using a consistent hash function is the core operation for determining the program's spatial angle. The consistent hash function maps the program identifier i to the interval between 0 and 1, and then obtains the corresponding mapping angle of the program through proportional conversion. , The value range is from 0 to 2π. Using a consistent hash function, for the same program identifier i, the mapped angle obtained is... The coordinates remain stable and do not change due to time or other factors, thus ensuring that the angular position of the same program on the dilution disk is fixed. This avoids spatial mapping chaos caused by angular fluctuations, allowing the relative positional relationships between programs to stably reflect their inherent attribute associations. Generating the coordinates of the i-th program within the dilution disk is the final result of spatial mapping. Coordinate calculation uses a polar coordinate to rectangular coordinate conversion method, through... The i-th program is uniquely represented on the dilution disk as a spatial point: its radial position is determined by the segmentation matrix number k, reflecting its temporal attribute; its angular position is determined by the program identifier i, reflecting its own attribute. Ultimately, all programs naturally form a concentric circular band structure on the disk. Programs from the latest segmentation matrix are concentrated in the inner layer with a smaller radius, while programs from historical segmentation matrices are distributed in the outer layer with a larger radius. Furthermore, programs within the same layer are evenly distributed angularly, providing a clear spatial framework for subsequent weight balancing and potential score calculation based on spatial geometric characteristics.

[0024] In one embodiment of the present invention, determining the mapping radius corresponding to the k-th piecewise matrix and setting the geometric dilution factor includes: Based on mapping radius Calculate the mapped area in the dilution disk. , ;in, Represents the hyperbolic cosine function; Based on the mapped area Set the geometric dilution factor for the k-th fragment matrix. , ;in, It is a constant greater than 0.

[0025] It should be noted that calculating the mapped area based on the mapped radius utilizes the negative curvature space property. In the diluted disk, the mapped area of ​​the k-th piecewise matrix... pass The calculations show that the hyperbolic cosine function *cosh* is a core mathematical tool for negative curvature spaces. This function makes the mapped area... With radius This growth exhibits exponential growth, meaning a small increase in radius corresponds to a significant expansion in area. This area growth characteristic is a natural weight dilution mechanism. Programs in the newest segmented matrix, located in the inner layer with a small radius and small area, and dense data, can avoid excessive weight accumulation due to the area characteristic; programs in the historical segmented matrix, located in the outer layer with a large radius and large area, and sparse data, can have their weight proportion compensated for through area expansion, providing a geometric basis for balancing the influence of new and old data. Setting the geometric dilution factor based on the mapped area is the process of converting spatial characteristics into weight adjustment coefficients. Geometric Dilution Factor The calculation formula is: ,in, A constant greater than 0 is used for standardization. This setting makes... With the mapped area There is an inverse relationship: the smaller the value of the latest piecewise matrix k, the larger the corresponding mapping area. Small, therefore Larger; a larger historical piecewise matrix k value corresponds to a larger mapping area. Large, therefore The weights are relatively small. This inverse relationship essentially involves adjusting the coefficients to further balance the weights of different shard matrices; that is, although the latest data... Larger but smaller area results in diluted overall weight; outdated piecewise matrices, although... Smaller but larger areas are compensated for by their overall weight. Ultimately... It serves as a bridge connecting spatial structure and data weights, providing a quantitative basis for time decay processing when calculating the equivalent quality of programs, ensuring that old and new programs are reasonably considered in weight calculations, and avoiding bias in recommendation results towards recent programs.

[0026] In one embodiment of the present invention, determining the interaction weights of the k-th fragment matrix and generating the equivalent quality of the corresponding i-th program by combining the geometric dilution factor includes: Extract the interaction weights of the i-th type of program in the k-th segment matrix, including: Get the cumulative dwell time of each user on the i-th type of program within the k-th sub-time period; The sum of the cumulative dwell time is used as the interaction weight of the i-th type of program in the k-th segment matrix. ; The interaction weights are diluted using the geometric dilution factor of the k-th partition matrix, as follows: in, This represents the dilution interaction weight of the i-th type of program in the k-th segment matrix; The sum of the diluted interaction weights corresponding to the K fragment matrices is used as the equivalent quality of the i-th type of program.

[0027] It should be noted that determining the interaction weights of the k-th segment matrix and combining them with the geometric dilution factor to generate the equivalent quality of the i-th program combines the behavioral intensity of the time dimension with the weight balancing mechanism of the spatial dimension. This quantitative processing ensures that the program's influence reflects both real-time popularity and historical accumulation. The specific process is as follows: Extracting the interaction weight of the i-th program in the k-th segment matrix is ​​an operation to quantify the program's popularity within a specific time period. First, it's necessary to obtain the cumulative dwell time of each user on the i-th program within the k-th sub-time period. The cumulative dwell time directly reflects the user's engagement with program i; a longer dwell time indicates higher user interest. Then, these cumulative dwell times are summed, and the sum is the interaction weight of the i-th program in the k-th segment matrix. The interaction weight is essentially a quantified value of the overall popularity of program i within the k-th sub-time period, comprehensively reflecting the overall level of attention paid to program i by all users during that time period.

[0028] Geometric dilution factor The coefficient is calculated based on the mapping area corresponding to the k-th piecewise matrix. Its value decreases as k increases, reflecting the principle that the weight decays more over time. Dilution processing is performed using the formula... Implementation, in which Let represent the dilution interaction weight for the i-th type of program in the k-th segment matrix. The essence of dilution is to adjust the original popularity of program i in the k-th time period according to the time decay law, that is, the popularity of the most recent time period (smaller k value) is adjusted accordingly. Larger values ​​indicate greater retention of original heat; older periods (larger k-values) The smaller the initial buzz, the more diluted it becomes, thus preventing the high buzz of the latest time slot from overly dominating the overall influence of the program.

[0029] The equivalent quality of program i is obtained by summing the diluted interaction weights corresponding to the K segment matrices, which is the final quantification of the comprehensive influence across multiple time periods. The equivalent quality integrates the diluted popularity of program i across all sub-time periods, including both the real-time popularity contribution of the latest time period and the attenuated popularity contribution of historical time periods. This ensures that the overall influence of program i is neither excessively amplified by recent surges in popularity nor excessively weakened by historically accumulated popularity, forming a quantitative indicator that balances temporal characteristics and spatial balance. This provides a scientific quality basis for subsequent recommendation ranking based on the potential energy model, ensuring that the recommendation results balance new and old interests and avoid bias towards recent content or neglect of historical preferences.

[0030] In one embodiment of the present invention, the user potential energy includes the potential energy source strength and the potential energy coordinates; Within a preset time period, determine the cumulative dwell time of the target user on the j-th type of program within each sub-time period to extract the corresponding F segment matrices; 1≤j≤S, 1≤F≤K, where j and F are both integers; Extract the geometric dilution factor of the f-th fragment matrix, as well as the cumulative dwell time of the j-th program in the f-th fragment matrix and the coordinates of the j-th program on the dilution disk; The product of the geometric dilution factor of the f-th segment matrix and the cumulative dwell time of the j-th program in the f-th segment matrix is ​​used as the coordinate weighting coefficient of the potential energy coordinate; where 1≤f≤F, and f is a positive integer; The coordinates of the j-th program in the dilution disk are weighted based on the coordinate weighting coefficients to obtain the weighted coordinates of the j-th program in the dilution disk in the f-th segment matrix. The sum of the weighted coefficients of the coordinates of the target user for each program is used as the potential energy source strength; The ratio of the sum of the weighted coordinates of the target user for each program to the potential energy source strength is used as the potential energy coordinate.

[0031] For example, within a 2-hour period, there are four 30-minute sub-time periods. The target user watched two programs (program A and program B, i.e., S=2) within these 2 hours, and had viewing records in the 1st, 2nd, 3rd, and 4th sub-time periods (i.e., extracting the corresponding 4 segment matrices, f=1, 2, 3, 4, corresponding to the preset time periods k=1, 2, 3, 4). First, determine the target user's cumulative viewing time for the two programs within each sub-time period: in f=1 (the first 30-minute period), the user watched program A for a cumulative 600 seconds; in f=2 (the second 30-minute period), the user watched program A for a cumulative 300 seconds; in f=4 (the fourth 30-minute period), the user watched program B for a cumulative 400 seconds (f=3 has no viewing record and is not included in the calculation). Next, the geometric dilution factor of the segmented matrices corresponding to k=1, 2, and 4 is extracted (the older the sub-time period, the smaller the factor): the geometric dilution factor for k=1 is 0.8, for k=2 it is 0.5, and for k=4 it is 0.2. Simultaneously, the dilution disk coordinates for program A at k=1 and k=2 are extracted as (0.2, 0.3) and (0.4, 0.5), respectively, and the coordinates for program B at k=4 are (0.6, 0.7). Then, the coordinate weighting coefficients (geometric dilution factor × duration of stay in that time period) are calculated: the coefficient for program A at k=1 is 0.8 × 600 = 480; the coefficient for program A at k=2 is 0.5 × 300 = 150; and the coefficient for program B at k=4 is 0.2 × 400 = 80. The weighted coordinates are calculated based on coefficients: For program A with k=1, the weighted coordinate is 480, multiplied by its coordinates (0.2, 0.3), resulting in (96, 144); for program A with k=2, the weighted coordinate is 150, multiplied by its coordinates (0.4, 0.5), resulting in (60, 75); for program B with k=4, the weighted coordinate is 80, multiplied by its coordinates (0.6, 0.7), resulting in (48, 56). The potential energy source strength (the sum of all weighted coefficients) is calculated as follows: 480 + 150 + 80 = 710. Finally, the potential energy coordinates are calculated: first, the x-coordinate and y-coordinate values ​​of all weighted coordinates are added together, resulting in a total x-coordinate value of 96 + 60 + 48 = 204 and a total y-coordinate value of 144 + 75 + 56 = 275; then, these are divided by the source strength of 710, yielding the potential energy coordinates as (0.287, 0.387).

[0032] It should be noted that, within the preset time period, the cumulative dwell time of the target user on the j-th program in each sub-time period must first be determined. Based on this, F relevant segment matrices are extracted from the K segment matrices. Here, j represents the program identifier number watched by the target user, ranging from 1 to S, where S is the total number of programs watched by the user; F represents the number of segment matrices related to the target user's behavior, ranging from 1 to K, where K is the total number of segment matrices. This process filters out the segment matrices containing the target user's behavior data.

[0033] Next, the geometric dilution factor of the f-th slice matrix, the cumulative dwell time of the j-th program in the matrix, and its coordinates on the dilution disk are extracted. The geometric dilution factor is the time decay coefficient corresponding to the slice matrix, the cumulative dwell time reflects the user's interest in the program during that time period, and the coordinates are the spatial position of the program in negative curvature space.

[0034] The geometric dilution factor of the f-th segment matrix is ​​multiplied by the cumulative dwell time of the j-th program in that matrix, and the result is used as the coordinate weighting coefficient of the potential energy coordinate. The coordinate weighting coefficient essentially adjusts for time decay in user interest intensity; that is, the longer the target user dwells on program j, the higher the interest intensity. The newer the corresponding segment matrix (the smaller f), the larger the geometric dilution factor, and the more the adjusted interest intensity is retained. Through the coordinate weighting coefficient, the original behavioral intensity of the target user can be transformed into an effective interest weight that takes into account time decay.

[0035] The coordinates of program j in the dilution disk are weighted using a coordinate weighting coefficient to obtain the weighted coordinates of that program in the f-th segment matrix. This maps the effective interest weight of the target user for program j to the spatial dimension, so that programs that the target user is more interested in (stay longer and are newer in the time slot) have a higher proportion of spatial coordinates in the construction of the user's interest center, thus making the spatial location more consistent with the actual interest distribution of the target user.

[0036] The sum of the weighted coefficients of the target user's coordinates for all programs is the potential energy source strength. The source strength quantifies the overall intensity of the target user's interest. The larger the value, the more active the target user's overall interaction and the more significant the expression of interest within the preset time period. This provides a quantitative benchmark for the strength of the target user's attraction to the programs in subsequent potential energy calculations.

[0037] The weighted coordinates of the target user across all programs are summed, and then this sum is divided by the potential energy source strength. The result is the potential energy coordinates. The potential energy coordinates represent the center position of the target user's interest in negative curvature space. They integrate spatial information from all historically viewed programs by the target user and are weighted according to interest intensity. Therefore, they accurately reflect the core spatial distribution of the user's interests and provide a benchmark for subsequent calculations of the spatial distance between programs and the target user's interests.

[0038] In one embodiment of the present invention, the potential energy score of the i-th program is calculated based on the user's potential energy and the coordinates of the i-th program, including: Calculate the geodesic distance between the potential energy coordinates and the coordinates of the i-th program; The denominator for distance attenuation is determined based on the geodetic distance; Based on the potential energy source strength, the equivalent quality of the i-th program, and the preset potential constant, the potential energy score of the i-th program is calculated as follows: in, Let represent the potential score of the i-th type of program. This indicates that the i-th type of program has a strong potential energy source. This represents the equivalent quality of the i-th type of program. The denominator represents the distance attenuation, and G represents a preset constant used for unifying dimensions.

[0039] It should be noted that the user potential energy includes potential energy source strength and potential energy coordinates, which together constitute a quantitative representation of the target user's interest in negative curvature space. The potential energy source strength reflects the total intensity of the target user's interest, while the potential energy coordinates reflect the core position of the target user's interest in space. Both are generated based on the target user's historical behavior after time decay adjustment, providing a benchmark for subsequent matching calculations.

[0040] Calculating the geodesic distance between the potential energy coordinates and the coordinates of the i-th program is crucial for measuring the correlation between the target user's center of interest and the program's spatial location. The geodesic distance is the true distance between two points in negative curvature space, its value determined by the coordinates of both points. The program coordinates are jointly determined by the time window number (determining the radius) and the program identifier (determining the angle). Therefore, the smaller the distance, the more likely the program is either close to the target user's historical interests in time (closer radius) or similar to the program attributes preferred by the target user (closer angle), meaning a higher degree of match with the user's interests.

[0041] Determining the denominator for distance attenuation based on geodesic distance is the process of converting spatial distance into a quantifiable attenuation coefficient. The value of the denominator increases with increasing geodesic distance. This variation ensures that programs located closer to the target user have a weaker attenuation effect and a higher weight in the final score; conversely, programs located farther away have a stronger attenuation effect and a lower weight in the score. This allows programs whose spatial location more closely aligns with the target user's center of interest to receive a higher base score, consistent with the intuitive logic of interest matching.

[0042] The potential energy score is calculated as the product of the potential energy strength, the equivalent quality of the i-th type of program, and a preset constant G, divided by the distance attenuation denominator. The preset constant G is used to unify the dimensions; the potential energy strength reflects the total intensity of the target user's interest; the larger the value, the more active the target user's interaction within the preset time period, the more significant their interest expression, and the higher the basic strength of their attraction to the program; the equivalent quality is the program's comprehensive influence after time attenuation adjustment; the larger the value, the higher the overall value of the program; the distance attenuation denominator adjusts the matching degree through spatial distance; the closer the distance, the smaller the denominator, and the higher the score. The synergistic effect of these three factors ensures that the score reflects not only the strength of the target user's interest, but also the value of the program itself, and takes into account the matching degree of spatial location, achieving a multi-dimensional quantification of the program's suitability for the target user.

[0043] The system precisely matches the target user's historical interests (after time decay) with the program's overall quality (after time balance) in a negative curvature space. Since both user potential energy and program equivalent quality have been weighted and balanced over time using a geometric dilution factor, the potential energy score naturally avoids the bias problem caused by uneven distribution of new and old data. This ensures that the score reflects both the sensitivity of the latest interests and the influence of historical preferences, providing a scientific quantitative basis for subsequent recommendation ranking.

[0044] In one embodiment of the present invention, intelligent recommendation processing is performed based on the potential score of each program, including: H programs are selected from high to low based on their potential scores to form a candidate set; For each program in the candidate set, the corresponding diversity penalty coefficient is calculated based on the prior playback volume, as follows: in, Let represent the diversity penalty coefficient for the h-th type of program in the candidate set, and C represent the number of program types in the candidate set. This represents the prior play count of the h-th program in the candidate set.

[0045] The product of the potential score and the diversity penalty coefficient for each program is used as the final ranking score for the corresponding program. Based on the final ranking scores, select Q programs from high to low to form a recommendation list for the target user.

[0046] It should be noted that selecting H programs from each program's potential score, ranked from highest to lowest, to form a candidate set is the initial screening step in the recommendation process. The potential score comprehensively considers the target user's interest intensity, program equivalent quality, and spatial matching degree; a higher score indicates a stronger fit between the program and the target user. Selecting H programs as candidates retains highly suitable core programs while controlling the candidate size through a fixed number, providing an operational range for subsequent diversity optimization and avoiding computational redundancy due to too many candidates or the loss of potentially high-quality programs due to too few.

[0047] Calculating a diversity penalty coefficient for each program in the candidate set based on prior view counts is a key adjustment method for balancing popular and long-tail content. Prior view counts reflect a program's popularity among all users; higher view counts indicate a more popular program. The diversity penalty coefficient decreases as prior view counts increase, its function being to appropriately de-weight over overly popular programs. While popular programs may be highly relevant to the target users, excessive concentration can lead to a homogenized recommendation list, reducing the user experience. By using the diversity penalty coefficient, the recommendation results can be prevented from being monopolized by a few top-tier programs, providing reasonable exposure opportunities for long-tail programs.

[0048] The potential score ensures the accuracy of matching programs with user interests, while the diversity penalty coefficient adjusts the balance of content distribution. The product of the two ensures that the score reflects both the program's suitability to users and its contribution to diversity in the recommendation list. This ensures that the final ranking not only prioritizes highly suitable programs but also suppresses overly popular content through the penalty coefficient, allowing moderately popular but highly suitable programs to have a chance to enter the top recommendation list, achieving a balance between accuracy and diversity.

[0049] The output stage of the recommendation process involves selecting Q programs from high to low based on the final ranking scores to form a recommendation list for the target user. Q represents a fixed length of the recommendation list, aligning with the actual display needs of IPTV set-top boxes. The programs in the final list undergo both potential score filtering to ensure a high match with user interests and diversity adjustment to avoid content homogenization. This achieves a balance between accurate recommendations and a rich user experience, resolving the issues of over-concentration of popular content or neglect of historical interests in traditional recommendations, and improving the overall performance of the recommendation system.

[0050] Fragmentation Matrix Hierarchical Weight Table Among them, the equivalent quality of the program represents the diluted weight of the program watched by users in each segment matrix (reflecting the comprehensive intensity of interest across multiple time periods). Geodesic distance represents the geometric distance between the user's potential energy coordinates and the program on the dilution disk (quantifying the degree of interest space matching). The distance decay denominator represents the far-field decay of the simulated physical potential energy; the greater the distance, the larger the denominator. Potential energy score is expressed as To integrate user interests, program quality, and spatial distance; Prior play counts represent the program's historical play data, providing a basis for subsequent diversity penalties (to prevent popular programs from dominating the screen).

[0051] Drama (Program 1): The program has the highest equivalent quality (184.7354) and the closest geodesic distance (0.3194) → the smallest distance attenuation denominator (0.0257) → the highest potential score (92167.124), reflecting the best interest matching; News (Program 7): The program's equivalent quality (198.7266) is the second highest, but the geodesic distance (0.4213) is relatively far, resulting in a large distance attenuation denominator (0.045) and a slightly low potential energy score (56616.635). It is necessary to combine the prior playback volume (90, which is relatively high) to determine whether it is suppressed by the diversity mechanism. Music (Program 4): Lowest equivalent quality (93.8907) + farthest geodesic distance (0.4969) → lowest potential score (19114.953), but its prior play count (40, lowest) makes it easier to get a recommendation opportunity through diversity penalty.

[0052] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning, characterized in that, include: S1. Collect user identifiers and program identifiers at fixed time intervals to obtain K segment matrices; where k is the number of the segment matrix, 1≤k≤K and k is an integer, the smaller the value of k, the newer the sub-time period corresponding to the segment matrix; S2, load the dilution disk, map the i-th program onto the dilution disk, and obtain the coordinates of the i-th program within the dilution disk; the dilution disk D satisfies: ;in, Let represent the coordinates of any point within the dilution disk; for the ith type of program in the k-th segment matrix, calculate the mapping radius. , ;in, Indicates the preset radial step size; the i-th program is mapped using a consistent hash function: The mapping angle of the i-th program is obtained. , Generate the coordinates of the i-th program within the dilution disk. ; S3, determine the mapping radius corresponding to the k-th piecewise matrix and set the geometric dilution factor; based on the mapping radius Calculate the mapped area in the dilution disk. , ;in, Represent the hyperbolic cosine function; based on the area of ​​the mapping Set the geometric dilution factor for the k-th fragment matrix. , ;in, A constant greater than 0; S4, determine the interaction weights of the k-th slice matrix, and combine them with the geometric dilution factor to generate the equivalent quality of the i-th program. S5. Collect and extract S program identifiers of the target user's time sequence at fixed time intervals, and map them to a dilution disk to obtain the user's potential energy; the user potential energy includes the potential energy source strength and potential energy coordinates; within a preset time period, determine the cumulative dwell time of the target user on the j-th program in each sub-time period to extract the corresponding F segment matrices; 1≤j≤S, 1≤F≤K, j and F are both integers; extract the geometric dilution factor of the f-th segment matrix, as well as the cumulative dwell time of the j-th program in the f-th segment matrix and the position of the j-th program on the dilution disk. The coordinates of the disk are determined; the product of the geometric dilution factor of the f-th fragment matrix and the cumulative dwell time of the j-th program in the f-th fragment matrix is ​​used as the coordinate weighting coefficient of the potential energy coordinates; the coordinates of the j-th program in the dilution disk are weighted based on the coordinate weighting coefficient to obtain the weighted coordinates of the j-th program in the f-th fragment matrix in the dilution disk; the sum of the coordinate weighting coefficients of the target user for each program is used as the potential energy source strength; the ratio of the sum of the weighted coordinates of the target user for each program to the potential energy source strength is used as the potential energy coordinates. S6. Based on the user's potential energy and the coordinates of the i-th program, calculate the potential energy score of the i-th program; calculate the geodesic distance between the potential energy coordinates and the coordinates of the i-th program; determine the distance attenuation denominator based on the geodesic distance; calculate the potential energy score of the i-th program based on the potential energy source strength, the equivalent quality of the i-th program, and the preset potential constant, as follows: ； in, Let represent the potential score of the i-th type of program. This indicates that the i-th type of program has a strong potential energy source. This represents the equivalent quality of the i-th type of program. The denominator represents the distance attenuation. These are preset constants used for unifying dimensions; S7 intelligently recommends programs based on their potential scores.

2. The method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning according to claim 1, characterized in that, User identifiers and program identifiers are collected at fixed time intervals to obtain K segment matrices, including: Within a preset time period, user identifiers and program identifiers are collected at fixed time intervals; Set the number of segments K to divide the preset time period into K sub-time periods; Determine the cumulative dwell time of the u-th user on the i-th program within each sub-time period to construct K segmentation matrices; where the element in the u-th row and i-th column of the k-th segmentation matrix represents the cumulative dwell time of the u-th user on the i-th program within the k-th sub-time period.

3. The method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning according to claim 1, characterized in that, Determine the interaction weights of the k-th fragment matrix and combine them with the geometric dilution factor to generate the equivalent quality of the corresponding i-th program, including: Extract the interaction weights of the i-th type of program in the k-th segment matrix, including: Get the cumulative dwell time of each user on the i-th type of program within the k-th sub-time period; The sum of the cumulative dwell time is used as the interaction weight of the i-th type of program in the k-th segment matrix. ; The interaction weights are diluted using the geometric dilution factor of the k-th partition matrix, as follows: ； in, This represents the dilution interaction weight of the i-th type of program in the k-th segment matrix; The sum of the diluted interaction weights corresponding to the K fragment matrices is used as the equivalent quality of the i-th type of program.

4. The method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning according to claim 1, characterized in that, Intelligent recommendation processing is performed based on the potential score of each program, including: H programs are selected from high to low based on their potential scores to form a candidate set; For each program in the candidate set, the corresponding diversity penalty coefficient is calculated based on the prior playback volume, as follows: ； in, This represents the diversity penalty coefficient for the h-th type of program in the candidate set. This indicates the types of programs in the candidate set. This represents the prior play count of the h-th program in the candidate set; The product of the potential score and the diversity penalty coefficient for each program is used as the final ranking score for the corresponding program. Based on the final ranking scores, select Q programs from high to low to form a recommendation list for the target user.

Images