Intelligent content recommendation method and system based on AI visual algorithm

Abstract

The invention provides an intelligent content recommendation method and system based on an AI visual algorithm, and relates to the technical field of data processing, and the method comprises the steps: calculating each interest weight of a current user for multi-dimensional visual features based on the historical interaction content of the current user; generating a dynamic threshold in combination with the interest weight and a pre-acquired user activeness; determining anchor point content in combination with the historical interaction content and all interest weights, and determining a primary candidate content set from the plurality of to-be-recommended contents based on the lightweight visual association index table; for each primary candidate content in the primary candidate content set, performing multi-dimensional weighting calculation according to all interest weights of the current user to obtain a visual similarity score; according to the method, the primary candidate content with the visual similarity score exceeding the dynamic threshold value is screened out, the content recommendation list is generated according to the screened primary candidate content, and the accuracy of visual content recommendation is improved through dynamic interest modeling, lightweight index construction and self-adaptive threshold value optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an intelligent content recommendation method and system based on an AI vision algorithm. Background Art

[0002] With the booming development of the digital content industry, the problem of information overload faced by users is becoming increasingly severe. Traditional content recommendation systems mainly rely on users' historical behavior data (such as clicks, browsing duration, collections, etc.) for collaborative filtering or content-based recommendation, but these methods have significant limitations. On the one hand, traditional methods often ignore the multi-dimensional features of visual content (such as object structure, scene semantics, style aesthetics, and dynamic behavior), resulting in insufficient matching of the recommended results with the user's current visual preferences; on the other hand, the dynamic changes of user interests (such as short-term interest drift or long-term preference evolution) are difficult to capture by static models, further reducing the timeliness and accuracy of recommendations. In addition, when dealing with large-scale visual content, existing systems often face performance bottlenecks due to excessively high computational complexity and are difficult to meet the requirements of real-time recommendation.

[0003] Therefore, it is necessary to provide an intelligent content recommendation method and system based on an AI vision algorithm to solve the above technical problems. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides an intelligent content recommendation method and system based on an AI vision algorithm, which improves the accuracy of visual content recommendation through dynamic interest modeling, lightweight index construction, and adaptive threshold optimization.

[0005] The present invention provides an intelligent content recommendation method based on an AI vision algorithm, and the method includes the following steps: Based on the historical interaction content of the current user, use the exponential decay function to calculate the respective interest weights of the current user for multi-dimensional visual features, where the interest weight is the preference intensity of the current user for each dimension of visual features in the multi-dimensional visual features; Based on all the interest weights, generate a dynamic threshold in combination with the pre-acquired user activity; Determine the anchor content in combination with the historical interaction content and all the interest weights, and determine a primary candidate content set from multiple content to be recommended based on a pre-constructed lightweight visual association index table, where the pre-constructed lightweight visual association index table represents a visual relationship matrix between two content to be recommended; For each primary candidate content in the primary candidate content set, perform multi-dimensional weighted calculation according to all the interest weights of the current user to obtain a visual similarity score; Filter out the primary candidate content whose visual similarity score exceeds the dynamic threshold, and generate a content recommendation list based on the filtered primary candidate content.

[0006] Preferably, the calculation process of the interest weight includes: Extract the object detection feature, scene semantic feature, style aesthetic feature, and dynamic behavior feature of each content in the historical interaction content of the current user; For each dimension of visual features, according to the interaction timestamp between the current user and the historical interaction content, calculate the preference intensity at the current moment through an exponential decay function, and use it as the interest weight. The calculation formula for the preference intensity is: Where represents the current moment For the preference intensity of the dimension visual feature, represents the th interaction timestamp, represents the decay coefficient, represents the th historical interaction content's dimension visual feature's significance score, represents the base of the natural logarithm.

[0007] Preferably, the generation process of the dynamic threshold includes: According to the historical interaction frequency of the current user within a preset time window and the number of valid operations triggered in the current session, calculate the user activity score, where the historical interaction frequency and the number of valid operations are fused through a weighted sum; Based on the logarithmic function relationship between the sum of preference intensities and the user activity score, generate a dynamic threshold through a normalization scaling factor.

[0008] Preferably, the construction process of the pre-constructed lightweight visual association index table includes: For all content to be recommended, extract the set of visual features to be recommended with the same dimension as the multi-dimensional visual features; Based on the set of visual features to be recommended, calculate the visual similarity between every two content to be recommended, and construct a visual similarity matrix, where each matrix element represents the visual similarity between the corresponding two content to be recommended, and the visual similarity is calculated using the weighted cosine similarity of each dimension of visual features; Perform sparsification processing on the visual similarity matrix and store it using a hash table structure to form a lightweight visual association index table.

[0009] Preferably, the determination process of the primary candidate content set includes: Select at least one piece of historical interaction content with the highest matching degree with the normalized preference intensity of the current user from the historical interaction content, where the matching degree is determined by calculating the weighted dot product of the multi-dimensional visual features of the historical interaction content and the normalized preference intensity; Based on the visual similarity matrix stored in the lightweight visual association index table, retrieve all the content to be recommended with a visual similarity exceeding the preset association threshold with the anchor content, and form a primary candidate content set.

[0010] Preferably, the process of obtaining the visual similarity score includes: For each primary candidate content, calculate the feature similarity between the visual feature set to be recommended and the multi-dimensional visual features of the anchor content in each dimension; Weight-multiply the feature similarity in each dimension with the normalized preference intensity of the current user to obtain the weighted similarity component in each dimension; Sum up the weighted similarity components in all dimensions to obtain the visual similarity score of this primary candidate content relative to the current user, and perform normalization.

[0011] Preferably, the process of generating the content recommendation list includes: Compare the visual similarity score of each primary candidate content with the generated dynamic threshold in real time, and screen out the primary candidate content with a score exceeding the dynamic threshold; For the screened primary candidate content, sort them in descending order according to their visual similarity scores to generate an ordered recommended content queue; Based on the ordered recommended content queue, intercept a preset number of candidate content to generate the final content recommendation list.

[0012] The present invention also provides an intelligent content recommendation system based on an AI visual algorithm for executing the intelligent content recommendation method based on an AI visual algorithm, and the system includes: A weight calculation module, configured to calculate the respective interest weights of the current user for multi-dimensional visual features by using an exponential decay function based on the historical interaction content of the current user, where the interest weight is the preference intensity of the current user for each dimension of visual features in the multi-dimensional visual features; A threshold dynamic generation module, configured to generate a dynamic threshold based on all the interest weights and in combination with the pre-acquired user activity; A content primary screening module, configured to determine the anchor content in combination with the historical interaction content and all the interest weights, and determine a primary candidate content set from multiple pieces of content to be recommended based on a pre-constructed lightweight visual association index table, where the pre-constructed lightweight visual association index table represents the visual relationship matrix between two pieces of content to be recommended; a multi-dimensional weighting module for performing a multi-dimensional weighted calculation on each primary candidate content in the primary candidate content set according to all interest weights of the current user to obtain a visual similarity score; The recommended content generation module is configured to screen out primary candidate content whose visual similarity scores exceed the dynamic threshold, and generate a content recommendation list based on the screened primary candidate content.

[0013] Compared with related technologies, the intelligent content recommendation method and system based on AI visual algorithm provided by the present invention have the following beneficial effects: Through the collaborative optimization of dynamic interest modeling (combining multi-dimensional visual features with exponential decay functions), lightweight visual association indexing, and adaptive dynamic thresholds (adjusted based on the linkage between user activity and interest weights), this invention systematically addresses the key deficiencies of existing recommendation systems in terms of dynamic interest capture, computational efficiency, and adaptability, providing an efficient solution for large-scale visual content recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flowchart of an intelligent content recommendation method based on AI visual algorithm provided by the present invention; Figure 2 This is a module structure diagram of an intelligent content recommendation system based on AI visual algorithm provided by the present invention. DETAILED DESCRIPTION

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It will be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all of the structures. Furthermore, the embodiments of the present invention and the features of the embodiments may be combined with one another unless there is a conflict.

[0016] It should also be noted that, for ease of description, only portions relevant to the present invention are shown in the accompanying drawings, rather than all of the contents. Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the various operations (or steps) as sequential processes, many of the operations can be performed in parallel, concurrently, or simultaneously. In addition, the order of the various operations can be rearranged. The process can be terminated when its operations are completed, but may also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0017] Example 1 The present invention provides an intelligent content recommendation method based on an AI vision algorithm, referring to Figure 1 as shown, the method includes the following steps: S1: Based on the historical interaction content of the current user, use the exponential decay function to calculate the respective interest weights of the current user for multi-dimensional visual features, where the interest weight is the preference intensity of the current user for each dimension of the multi-dimensional visual features.

[0018] Specifically, the calculation process of the interest weight includes: S11: Extract the object detection feature, scene semantic feature, style aesthetic feature, and dynamic behavior feature of each content in the historical interaction content of the current user.

[0019] In this embodiment, four types of visual features are extracted from the historical interaction content of the user, namely object detection feature, scene semantic feature, style aesthetic feature, and dynamic behavior feature, and each type of feature corresponds to a visual dimension.

[0020] For the object detection feature, its extraction process is as follows: Use the pre-trained YOLOv7 model to detect the objects in the content. The input images are uniformly scaled to a resolution of 640×640, maintaining the original aspect ratio, and non-square images are processed by padding with gray edges. The model outputs the coordinate information of the detection boxes and the object category confidence of the 80-class COCO dataset. For each content, select the top 5 object categories with the highest confidence, count their occurrence frequencies and encode them into a 256-dimensional feature vector, and use principal component analysis (PCA) to reduce the dimension to 64 dimensions, retaining 90% of the variance information.

[0021] For the scene semantic feature, its extraction process is as follows: Adopt the ResNet-50 model fine-tuned on the Places365 dataset. The input images are centrally cropped to 224×224 pixels, and a 365-dimensional scene category probability distribution is output. Select the top 3 scene categories with the highest probability values and splice their probability values into a 3-dimensional scene feature vector.

[0022] For the style aesthetic feature, its extraction process is as follows: Extract the style feature through the StyleGAN-2 model. Specifically, use the 6th layer style modulation parameter of the generator as a 512-dimensional style vector. At the same time, the aesthetic scoring model trained based on the AVA dataset outputs an aesthetic score from 0 to 1, and splice the style vector and the aesthetic score into a 513-dimensional composite feature.

[0023] For the dynamic behavior feature, its extraction process is as follows: Use the SlowFast network to extract spatio-temporal features for video content. The input video segment is sampled into 32 frames, and the RGB frame and optical flow frame features are extracted respectively, and a 2048-dimensional dynamic feature vector is obtained after fusion.

[0024] S12: For the visual features of each dimension, calculate the preference intensity at the current moment according to the interaction timestamp between the current user and the historical interaction content through an exponential decay function, and use it as the interest weight. The calculation formula for the preference intensity is: Where represents the current moment For the preference intensity of the dimensional visual features, represents the timestamp of the th interaction, represents the decay coefficient, represents the th historical interaction content of the dimensional visual feature significance score, represents the base of the natural logarithm.

[0025] In this embodiment, the interest weight is dynamically adjusted according to the time distribution of the user's historical interaction behavior.

[0026] The calculation formula for the preference intensity calculates the user's interest weight for multi-dimensional visual features through an exponential decay function, and its specific form is a comprehensive evaluation of the user's each historical interaction behavior in the time dimension and the content significance dimension. The base of the natural logarithm in the formula is a mathematical constant used to construct a standard exponential decay relationship. The decay coefficient is an adjustable parameter, and its value range is limited between 0.1 and 0.5. This range has been experimentally verified to effectively balance the influence of recent behavior and long-term interest. The time difference calculation uses the timestamp difference accurate to seconds to ensure the accuracy of the timeliness evaluation.

[0027] Significance score is a quantitative index to measure the influence intensity of a certain visual feature dimension in the user's historical interaction content on the user's interest. Its definition is: under the visual feature dimension, the significant degree of the user's feature performance of a certain historical interaction content.

[0028] The significance score needs to be calculated separately for each visual feature dimension (object, scene, style, dynamics) and quantified in combination with the key parameters in the feature extraction process.

[0029] Specifically, the calculation process of the significance score is: The saliency score of object detection features, which measures the degree of user attention to object recognition in a certain piece of historical interaction content. First, use the YOLOv7 model to detect objects in the content and output each detected object category and its confidence level. Select the top five object categories with the highest confidence levels from these detection results, and calculate the average confidence level of these five object categories as the saliency score of the content in the object dimension. If there are fewer than five detected objects, take the average confidence level of the actually detected objects as the saliency score. The final obtained score ranges from 0 to 1, representing the saliency of the content in terms of object detection features.

[0030] The saliency score of scene semantic features, which is used to evaluate the intensity of user interest in scene classification in a certain piece of historical interaction content. Analyze the content through the fine-tuned ResNet-50 model and output the probability distribution of 365 possible scene categories. Select the top three scene categories with the highest probability values, and add up the probability values of these three scenes to obtain a total. To ensure that this total can be represented within the standard range of 0 to 1, we further divide it by 3 to complete the normalization process, thereby obtaining the saliency score of the content in terms of scene semantic features. The purpose of doing this is to make the comparison between different contents more fair and reasonable.

[0031] The saliency score of style aesthetic features, which reflects the strength of user interest in the style and aesthetics aspects of a certain piece of historical interaction content. Directly use the aesthetic score given by the aesthetic scoring model trained based on the AVA dataset as the saliency score of style aesthetic features, and the range of the aesthetic score is from 0 to 1. For those contents without aesthetic scores (such as certain videos), calculate the norm of the style vector of the 6th layer of the StyleGAN-2 generator and divide it by the preset maximum norm threshold to replace the aesthetic score, and use this as the saliency score of style aesthetic features.

[0032] The saliency score of dynamic behavior features, which is applicable to video content and reflects the degree of user attention to the dynamic changes in the video. For video content, use the SlowFast network to extract its spatio-temporal feature vector and calculate the variance of this vector to measure the change intensity of video features. Set the weight according to the duration of the video: if the video length exceeds 30 seconds, the weight is 1.2; if the video length is less than 30 seconds, the weight is 0.8. Obtain the saliency score of dynamic behavior features by dividing the feature vector variance by the global feature variance mean and multiplying by the video duration weight. For non-video content, the saliency score of dynamic behavior features is set to 0 and this dimension is ignored in subsequent calculations.

[0033] S2: Based on all interest weights, generate a dynamic threshold by combining the pre-acquired user activity.

[0034] Specifically, the generation process of the dynamic threshold includes: S21: Calculate the user activity score based on the historical interaction frequency of the current user within a preset time window and the number of valid operations triggered in the current session, where the historical interaction frequency and the number of valid operations are fused through weighted summation.

[0035] In this embodiment, the real-time activity of the user is calculated through multi-dimensional behavior analysis, and the specific implementation process is as follows: First, perform historical interaction frequency statistics: Count the number of valid interaction behaviors of the user within a preset time window (default set to the last 7 days), including but not limited to core operations such as clicks, favorites, shares, and complete plays.

[0036] Assign different weights to different types of interaction behaviors (click weight 0.5, favorite 1.0, share 1.2, complete play 1.5), and calculate the historical interaction intensity value through weighted summation.

[0037] The time window supports dynamic adjustment. For example, for high-frequency users (with more than 50 daily interactions), the window is automatically shortened to 3 days to more sensitively capture interest changes.

[0038] Then, perform current session behavior analysis: Real-time record all operations of the user from login to the current moment, including page view depth (access path length), single-content stay duration (seconds), and action frequencies such as refreshes / swipes.

[0039] Calculate the session activity through a time decay factor: The behavior weight within the last 10 minutes is 1.0, and the behavior weight exceeding 30 minutes decays by 20% per hour.

[0040] Set a gain coefficient for key operations (such as favorites, repeat plays). When repeating the same type of operation within a single day, the contribution value of each operation increases by 10%.

[0041] Finally, score fusion and normalization: The historical interaction intensity and the current session activity are weighted and fused according to a set ratio, such as 6 to 4.

[0042] S22: Generate a dynamic threshold through a normalization scaling factor based on the logarithmic function relationship between the sum of preference intensities and the user activity score.

[0043] In this embodiment, an adaptive threshold is generated based on the dynamic relationship between user activity and preference intensity, and the specific implementation process is as follows: First, calculate the sum of preference intensities (S), specifically by adding up the preference intensities of each dimension output in step S13 to obtain an overall user interest intensity index.

[0044] Subsequently, calculate the user activity score (A), which is derived from the calculation result of step S21 and has a value range of 0 to 100 (after percentage mapping).

[0045] The dynamic threshold (T) is inversely proportional to the logarithm of the user activity score, and the calculation formula is: where, is the normalization scaling factor with a default value of 0.75, which is optimized and determined within the range of 0.6 to 0.9 through A / B testing.

[0046] For low-active users with a user activity score A less than or equal to 5, enable the protection mechanism: force the dynamic threshold T to be greater than or equal to 0.3 to avoid the decline in experience caused by an overly short recommendation list.

[0047] In addition, it is also necessary to constrain the range of the dynamic threshold and hard-limit the dynamic threshold within the reasonable range of When the calculated value exceeds the upper limit, enable the threshold decay strategy: automatically reduce it by 0.02 every 30 minutes until it returns below the threshold upper limit.

[0048] For a sudden surge in activity (such as A increasing by more than 50% within 5 minutes), start smoothing filtering: take the moving average of the most recent 3 activity scores to suppress noise fluctuations.

[0049] At the same time, the dynamic threshold has a real-time update mechanism: the dynamic threshold is recalculated every 5 minutes and persistently stored in the distributed cache.

[0050] When it is detected that the interval between two consecutive user requests is less than 1 minute, trigger an immediate threshold update to ensure the real-time nature of the recommendation results.

[0051] S3: Determine the anchor content by combining the historical interaction content and all interest weights, and determine the primary candidate content set from multiple content to be recommended based on the pre-constructed lightweight visual association index table, where the pre-constructed lightweight visual association index table represents the visual relationship matrix between pairwise content to be recommended.

[0052] In step S3, the construction process of the lightweight visual association index table includes: First, for all content to be recommended, extract the set of visual features to be recommended with the same dimension as the multi-dimensional visual features.

[0053] In this embodiment, multi-dimensional visual feature extraction is performed on all content to be recommended to ensure strict alignment with the feature dimensions of the user's historical interaction content.

[0054] Object detection feature extraction: Using the same YOLOv7 model configuration as in step S11, detect the objects in the candidate content. The input images are uniformly scaled to a resolution of 640×640, and bilinear interpolation is used to maintain the image quality.

[0055] Post-process the detection results: Filter out objects with a confidence level lower than 0.4, and count the class distribution of the remaining objects. Generate a 64-dimensional feature vector (after PCA dimensionality reduction) for each candidate content and store it in the feature database.

[0056] Scene semantic feature extraction: Use the fine-tuned ResNet-50 model in step S11, and the input images go through the same preprocessing process (center cropped to 224×224 pixels).

[0057] After outputting the 365-dimensional scene probability distribution, retain the probability values of the top-3 scene categories to form a 3-dimensional scene feature vector.

[0058] Style aesthetic feature extraction: Adopt the style parameters of the 6th layer of the StyleGAN-2 generator to extract a 512-dimensional style vector; combine the scalar value (0 to 1 point) output by the aesthetic scoring model, and concatenate the style vector and the aesthetic score into a 513-dimensional composite feature.

[0059] Dynamic behavior feature extraction (for video content): For video candidate content, use the SlowFast network to extract spatio-temporal features. The input video clip is sampled into 32 frames, the TV-L1 algorithm is used for optical flow calculation, and a 2048-dimensional vector is output after fusing RGB and optical flow features; for ultra-long videos (longer than 5 minutes), extract features in segments of every 30 seconds and then take the average.

[0060] Secondly, based on the set of visual features to be recommended, calculate the visual similarity between every two contents to be recommended, and construct a visual similarity matrix, where each matrix element represents the visual similarity between the corresponding two contents to be recommended. The visual similarity is calculated using the weighted cosine similarity of each dimension of visual features.

[0061] In this embodiment, based on the extracted set of features to be recommended, calculate the multi-dimensional similarity relationship between contents.

[0062] It includes single-dimensional similarity calculation, which is specifically divided into the following steps: Object feature similarity: Use IoU-aware cosine similarity. First calculate the cosine similarity of the object class distribution vectors of the two contents, and then multiply it by the average overlap rate of the detection boxes (IoU).

[0063] Scene feature similarity: Use Jensen-Shannon divergence to measure the difference in scene probability distributions and convert it into a 0-1 similarity value.

[0064] Style feature similarity: Calculate the cosine similarity of the composite feature vectors, and apply a penalty factor of 0.8 when the aesthetic score difference exceeds 0.3.

[0065] Dynamic feature similarity: For video content, calculate the similarity after aligning the feature sequences using Dynamic Time Warping (DTW).

[0066] Then, perform weighted fusion according to the proportion of the dimension assignment of the user interest weight (including but not limited to object 0.4, scene 0.3, style 0.2, dynamic 0.1, which can be selected according to the actual situation).

[0067] For cross-modal content (such as images and videos), only calculate the similarity of the common dimensions (such as object, scene, style, dynamic), and distribute the weight of the missing dimension proportionally to other dimensions.

[0068] Finally, optimize the matrix. Adopt a block calculation strategy, divide the content library of tens of millions into 100×100 sub-matrix blocks, use GPU parallel acceleration for calculation, directly assign zero to the cells with similarity lower than 0.2, and reduce the overhead of invalid calculations.

[0069] Finally, sparsify the visual similarity matrix and store it using a hash table structure to form a lightweight visual association index table.

[0070] In this embodiment, converting the original similarity matrix into a lightweight retrieval structure specifically includes: Sparsification processing: Set a dynamic retention threshold: Only retain the top-50 adjacent contents with the highest similarity for each row (single content), and set the rest to zero.

[0071] For the popular content at the top (such as the click-through rate is greater than 100,000 times), additionally retain the long-tail associated items with similarity greater than 0.7 to improve the recall rate.

[0072] Adopt a two-layer hash index: The first layer uses the content ID as the key and stores the list of the top-50 adjacent content IDs; the second layer uses the content ID pair as the key (such as "ID1 to ID2") and stores the exact similarity value.

[0073] Perform differential coding compression on the adjacent list: Convert the content ID sequence into a difference array to reduce the storage space occupation, and finally use the Memory-Mapped File (MMAP) technology to load the index table into the memory to support high-speed random access.

[0074] Specifically, the process of determining the primary candidate content set includes: S31: Select at least one historical interaction content with the highest matching degree with the current user's normalized preference intensity from the historical interaction content as the anchor content, where the matching degree is determined by calculating the weighted dot product of the multi-dimensional visual features of the historical interaction content and the normalized preference intensity.

[0075] In this embodiment, the anchor content that best matches the current preference is filtered out from the user's historical interaction content.

[0076] First, for each historical interaction, the weighted dot product of its multidimensional visual feature vector and the user's normalized preference strength is calculated. For object detection features, the 64-dimensional feature vector is multiplied by the user's object-dimension preference strength and the sum is calculated. Scene semantic features are directly dot-producted using a 3-dimensional vector with the scene preference strength. Style and dynamic features are processed similarly, using 513-dimensional and 2048-dimensional feature vectors, respectively. The top three historical interactions with the highest weighted dot product scores are automatically selected as anchor points. If the difference in the highest scores is less than 0.1, the top five anchor points are expanded to avoid omissions.

[0077] For new or low-activity users, when available historical content is insufficient, a fallback strategy is activated: 10 pieces of content are randomly selected from a pre-set popular content library as temporary anchors, and replaced in real time based on first-time interactions. All anchor content must meet time validity requirements, and outdated content older than 30 days is automatically filtered to ensure timely recommendations.

[0078] S32: Based on the visual similarity matrix stored in the lightweight visual association index table, all to-be-recommended content whose visual similarity with the anchor content exceeds a preset association threshold is retrieved to form a primary candidate content set.

[0079] In this embodiment, multi-anchor association retrieval is performed based on a lightweight visual association index table. For each selected anchor, the top 50 similar contents associated with it are extracted from the hash table index, and entries with a similarity exceeding a dynamic association threshold (initial value 0.6) are screened.

[0080] Cross-source merging of multi-anchor search results: When the same content is associated with multiple anchors, only the highest similarity value is retained, and the number of associated anchors is recorded as an auxiliary ranking factor. A two-stage filtering mechanism is used to first eliminate candidate content with a similarity below 0.4, and then exclude content that the user has interacted with in the past 7 days. Diversity control is implemented on the merged candidate set, and categories with a share exceeding 30% are statistically distributed according to content categories, with a downgrade (score × 0.7). The final candidate set size is controlled between 200 and 500 items. When insufficient results are found, the association threshold is automatically lowered (by 0.05 each time) and the search is repeated. The retrieval process uses batch parallel computing, compressing the response time of a single request to less than 50ms, and supporting tens of thousands of concurrent queries per second.

[0081] S4: For each primary candidate content in the primary candidate content set, perform multi-dimensional weighted calculation based on all interest weights of the current user to obtain a visual similarity score.

[0082] Specifically, the process of obtaining the visual similarity score includes: S41: For each primary candidate content, calculate the feature similarity of its visual feature set to be recommended and the multi-dimensional visual features of the anchor content in each dimension.

[0083] In this embodiment, specifically, first extract the multi-dimensional visual features of the candidate content and the anchor content, including object detection features, scene semantic features, style aesthetic features, and dynamic behavior features.

[0084] For object detection features, calculate the cosine similarity between the category distribution vectors of the two, and adjust the result by combining the average overlap rate (IoU) of the detection boxes to ensure that both the category matching degree and the consistency of the spatial position are considered.

[0085] For scene semantic features, use Jensen-Shannon divergence to measure the difference in the scene probability distributions of the two, and map the result to the similarity interval from 0 to 1.

[0086] For style aesthetic features, calculate the cosine similarity of the composite feature vectors. If the aesthetic score difference exceeds a certain threshold (such as 0.3), apply a penalty factor (such as 0.8) to the similarity to reduce the impact of overly obvious aesthetic differences on the similarity.

[0087] For dynamic behavior features, use the dynamic time warping (DTW) algorithm to align the feature sequences and then calculate the similarity, so as to handle the time offset problem of the inter-frame features in video content. Finally, obtain the feature similarity of the candidate content and the anchor content in each visual dimension.

[0088] S42: Multiply the feature similarities of each dimension by the normalized preference intensity of the current user to obtain the weighted similarity components of each dimension.

[0089] In step S42, based on the feature similarities of each dimension calculated in step S41, combined with the normalized preference intensity of the current user, perform weighted multiplication to obtain the weighted similarity components of each dimension.

[0090] Specifically, first read the normalized preference intensity of the current user from the user profile database, which includes the interest weights of the user in the four dimensions of object, scene, style, and dynamic. Then, multiply the feature similarity of each dimension by the corresponding interest weight to obtain the weighted similarity component of that dimension.

[0091] For example, if the user's interest weight for the object dimension is 0.4, and the feature similarity between the candidate content and the anchor content in the object dimension is 0.8, then the weighted similarity component for this dimension is 0.4 multiplied by 0.8, which is 0.32. This process is applied to all visual dimensions respectively to ensure that the similarity contribution of each dimension can be differentially adjusted according to the user's personalized preferences, thereby reflecting the user's attention degree to different visual features.

[0092] S43: Sum up the weighted similarity components of all dimensions to obtain the visual similarity score of this primary candidate content relative to the current user, and perform normalization.

[0093] In step S43, sum up the weighted similarity components of all dimensions to obtain the visual similarity score of this primary candidate content relative to the current user, and perform normalization processing on it.

[0094] Specifically, first accumulate the weighted similarity components of all dimensions calculated in step S42 to form a comprehensive score, indicating the overall visual similarity between the candidate content and the anchor content. Then, for the convenience of subsequent comparison and screening, perform normalization processing on this comprehensive score. The normalization process usually adopts the smooth L1 normalization method, that is, divide the comprehensive score of each candidate content by the sum of the comprehensive scores of all candidate contents, and perform logarithmic transformation on the result to enhance the discrimination.

[0095] In addition, outliers also need to be processed during the normalization process. For example, truncate the scores that exceed the reasonable range to avoid the influence of extreme values on the overall sorting effect.

[0096] Finally, each candidate content will obtain a normalized visual similarity score, which is used for subsequent dynamic threshold screening and content recommendation list generation.

[0097] S5: Screen out the primary candidate contents whose visual similarity scores exceed the dynamic threshold, and generate a content recommendation list based on the screened primary candidate contents.

[0098] Specifically, the process of generating the content recommendation list includes: S51: Compare the visual similarity score of each primary candidate content with the generated dynamic threshold in real time, and screen out the primary candidate contents whose scores exceed the dynamic threshold.

[0099] For each candidate content in the primary candidate content set, it is necessary to compare its visual similarity score with the dynamic threshold in real time, and screen out the candidate contents whose scores exceed the dynamic threshold.

[0100] Specifically, first read the latest dynamic threshold (T) of the current user from the distributed cache. This threshold is generated and updated in real time in step S2. Then, sequentially obtain the visual similarity scores of each candidate content (the calculation results from step S43) and compare them with the dynamic threshold. If the score of a certain candidate content is higher than the dynamic threshold, it is retained as a valid recommendation candidate; otherwise, it is excluded.

[0101] To ensure the screening efficiency, in actual operation, a batch processing method is adopted to compare the scores of multiple candidate contents with the threshold in parallel, so as to support the real-time response requirements in high-concurrency scenarios.

[0102] In addition, to avoid the recommendation list being too short or too long due to abnormal situations, automatically record the distribution of screening results and trigger the immediate adjustment mechanism of the dynamic threshold when necessary.

[0103] S52: For the primary candidate contents after screening, sort them in descending order according to their visual similarity scores to generate an ordered recommended content queue.

[0104] In step S52, specifically, first load all the candidate contents that pass the dynamic threshold screening and their corresponding visual similarity scores into memory to form a list to be sorted. Then, use an efficient sorting algorithm (including but not limited to quicksort or heapsort) to sort the list in descending order to ensure that the content with the highest score is displayed first.

[0105] To further improve the user experience, diversity control and time decay factors are also introduced: for the situation where the proportion of the same type of content is too high, downweight the low-ranked content; at the same time, set the time weight according to the freshness of the content and give priority to recommending the recently updated content. After sorting, a structured recommendation queue will be generated and its intermediate results will be stored in the cache for subsequent steps to use.

[0106] S53: Based on the ordered recommended content queue, intercept a preset number of candidate contents to generate the final content recommendation list.

[0107] In step S53, specifically, first read the target length of the recommendation list (such as 10 or 20) from the configuration file. This parameter can be dynamically adjusted according to the user device type, network condition or personalized needs.

[0108] Then, sequentially extract the target number of contents from the sorted recommendation queue and format them to ensure that each recommended content contains complete meta-information (such as title, cover image, description, etc.).

[0109] To further optimize the recommendation effect, the candidate content will be further filtered to eliminate content that may cause duplicate recommendations or is irrelevant (such as content that the user has clearly marked as "not interested"). Finally, the generated recommendation list will be encapsulated in JSON or other standardized formats and returned to the front-end display layer through an interface. During the entire process, the system will record the generation time and version number of the recommendation list for subsequent analysis and optimization.

[0110] Embodiment 2 The present invention also provides an intelligent content recommendation system based on an AI vision algorithm for implementing the method for intelligent content recommendation based on an AI vision algorithm as described above. Referring to Figure 2 as shown, the system includes: A weight calculation module 100 for calculating, based on the historical interaction content of the current user, the respective interest weights of the current user for multi-dimensional visual features by using an exponential decay function, where the interest weights are the preference intensities of the current user for each dimension of the multi-dimensional visual features.

[0111] A threshold dynamic generation module 200 for generating a dynamic threshold by combining the pre-acquired user activity based on all the interest weights.

[0112] A content primary screening module 300 for determining anchor content by combining the historical interaction content and all the interest weights, and determining a primary candidate content set from multiple content to be recommended based on a pre-constructed lightweight visual association index table, where the pre-constructed lightweight visual association index table represents a visual relationship matrix between pairs of content to be recommended.

[0113] A multi-dimensional weighting module 400 for performing multi-dimensional weighting calculations for each primary candidate content in the primary candidate content set according to all the interest weights of the current user to obtain a visual similarity score.

[0114] A recommended content generation module 500 for screening out the primary candidate content whose visual similarity score exceeds the dynamic threshold and generating a content recommendation list based on the screened primary candidate content.

[0115] This application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of this application. It should be understood that each process and / or block in the flowcharts and / or block diagrams can be implemented by computer program instructions, and the combination of the processes and / or blocks in the flowcharts and / or block diagrams can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate for implementing the processes Figure 1apparatus for the functions specified in one or more processes and / or boxes Figure 1 or multiple boxes

[0116] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program, and this program can be stored in a computer-readable storage medium. The storage medium includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc memories, magnetic disc memories, tape memories, or any other medium that can be used to carry or store data and is computer-readable.

[0117] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, commodity or device including the element.

Claims

1. An intelligent content recommendation method based on AI visual algorithm, characterized in that: The method includes the following steps: Based on the historical interaction content of the current user, use an exponential decay function to calculate the respective interest weights of the current user for multi-dimensional visual features, where the interest weight is the preference intensity of the current user for each dimension of the multi-dimensional visual features; Based on all the interest weights, generate a dynamic threshold in combination with the pre-acquired user activity; Determine the anchor content in combination with the historical interaction content and all the interest weights, and determine a primary candidate content set from multiple content to be recommended based on a pre-constructed lightweight visual association index table, where the pre-constructed lightweight visual association index table represents a visual relationship matrix between every two content to be recommended; For each primary candidate content in the primary candidate content set, perform multi-dimensional weighted calculation according to all the interest weights of the current user to obtain a visual similarity score; Filter out the primary candidate content whose visual similarity score exceeds the dynamic threshold, and generate a content recommendation list according to the filtered primary candidate content.

2. The intelligent content recommendation method based on the AI vision algorithm according to claim 1, wherein, The calculation process of the interest weight includes: Extract the object detection features, scene semantic features, style aesthetic features, and dynamic behavior features of each content in the historical interaction content of the current user; For each dimension of visual features, calculate the preference intensity at the current moment through an exponential decay function according to the interaction timestamp of the current user and the historical interaction content, and use it as the interest weight. The calculation formula of the preference intensity is: where represents the current moment the preference intensity for the dimensional visual feature, represents the timestamp of the th interaction, represents the attenuation coefficient, represents the significance score of the dimensional visual feature in the th historical interaction content, represents the base of the natural logarithm.

3. The intelligent content recommendation method based on AI visual algorithm according to claim 2, characterized in that: The generation process of the dynamic threshold includes: Calculate the user activity score according to the historical interaction frequency of the current user within a preset time window and the number of effective operations triggered in the current session, where the historical interaction frequency and the number of effective operations are fused through a weighted sum; Based on the logarithmic function relationship between the sum of preference intensities and the user activity score, generate a dynamic threshold through a normalization scaling factor.

4. An intelligent content recommendation method based on AI vision algorithm according to claim 3, characterized in that, The construction process of the pre-constructed lightweight visual association index table includes: For all content to be recommended, extract a set of visual features to be recommended with the same dimension as the multi-dimensional visual features; Based on the set of visual features to be recommended, calculate the visual similarity between every two content to be recommended, and construct a visual similarity matrix, where each matrix element represents the visual similarity between the corresponding two content to be recommended, and the visual similarity is calculated using the weighted cosine similarity of each dimension of visual features; Perform sparsification processing on the visual similarity matrix and store it using a hash table structure to form a lightweight visual association index table.

5. The intelligent content recommendation method based on AI visual algorithm according to claim 4, characterized in that: The determination process of the primary candidate content set includes: Select at least one historical interaction content with the highest matching degree with the normalized preference intensity of the current user from the historical interaction content, where the matching degree is determined by calculating the weighted dot product of the multi-dimensional visual features of the historical interaction content and the normalized preference intensity; Based on the visual similarity matrix stored in the lightweight visual association index table, retrieve all content to be recommended whose visual similarity with the anchor content exceeds a preset association threshold to form a primary candidate content set.

6. The intelligent content recommendation method based on AI visual algorithm according to claim 5, characterized in that: The process of obtaining the visual similarity score includes: For each primary candidate content, calculate the feature similarity between its visual feature set to be recommended and the multi-dimensional visual features of the anchor content in each dimension; Weight-multiply the feature similarity in each dimension with the normalized preference intensity of the current user to obtain the weighted similarity component in each dimension; Sum up the weighted similarity components in all dimensions to obtain the visual similarity score of the primary candidate content relative to the current user, and perform normalization.

7. The intelligent content recommendation method based on AI visual algorithm according to claim 6, characterized in that: The generation process of the content recommendation list includes: Compare the visual similarity score of each primary candidate content with the generated dynamic threshold in real time, and filter out the primary candidate contents whose scores exceed the dynamic threshold; Sort the filtered primary candidate contents in descending order according to their visual similarity scores to generate an ordered recommended content queue; Based on the ordered recommended content queue, intercept a preset number of candidate contents to generate the final content recommendation list.

8. An intelligent content recommendation system based on AI vision algorithm, which is used to execute an intelligent content recommendation method based on AI vision algorithm as described in any one of claims 1 to 7, characterized in that, The system includes: A weight calculation module, configured to calculate the respective interest weights of the current user for the multi-dimensional visual features by using an exponential decay function based on the historical interaction content of the current user, where the interest weight is the preference intensity of the current user for each dimension of the multi-dimensional visual features; A threshold dynamic generation module, configured to generate a dynamic threshold based on all the interest weights and in combination with the pre-acquired user activity; A content primary screening module, configured to determine the anchor content by combining the historical interaction content and all the interest weights, and determine a set of primary candidate contents from multiple contents to be recommended based on a pre-constructed lightweight visual association index table, where the pre-constructed lightweight visual association index table represents a visual relationship matrix between any two contents to be recommended; A multi-dimensional weighting module, configured to perform multi-dimensional weighting calculation for each primary candidate content in the set of primary candidate contents according to all the interest weights of the current user to obtain a visual similarity score; A recommended content generation module, configured to filter out the primary candidate contents whose visual similarity scores exceed the dynamic threshold, and generate a content recommendation list according to the filtered primary candidate contents.

Images