Digital media content recommendation optimization method based on behavior analysis

Abstract

The invention discloses a digital media content recommendation optimization method based on behavior analysis, and particularly relates to the field of digital media recommendation, comprising the following steps: step S1, collecting interactive behavior data of a user and digital media content, including explicit behaviors and implicit behaviors; s2, the user requests to open a camera of the user during browsing, the user collects the face of the user through the camera after agreement, and facial behavior data is obtained after the facial data of the user is analyzed and transmitted to the interactive behavior data; s3, under the condition that the data point set is obtained, the original frequency of the data point set is analyzed, and the dynamic frequency of the data point set is calculated; and S4, constructing a short-term interest and long-term interest fused dynamic model. The subconscious emotional tendency of the user can be captured, the problem of poor recommendation timeliness is effectively solved, interest modeling is closer to the real preference of the user, and the situation that a recommendation system deviates from the core interest of the user due to useless data is avoided.

Description

Technical Field

[0001] The present invention relates to the field of digital media recommendation, and more particularly to a digital media content recommendation optimization method based on behavior analysis. Background Art

[0002] With the rapid development of internet technology, digital media content has exploded, and users are facing the challenge of "information overload." Personalized recommendation systems, as a core technology to address this problem, analyze user behavior data to predict interests and preferences and have been widely used on major digital media platforms. However, existing recommendation methods still have the following key flaws: 1. Traditional recommendation methods primarily rely on explicit user behavior and simple implicit behavior modeling, but ignore the emotional feedback data users display when viewing content. For example, facial expressions like a smile when watching a funny video, a focused frown when watching suspenseful content, and a serious expression when reading negative news all hint at a user's true emotional inclinations toward the content. However, existing methods fail to incorporate these behaviors into their behavioral analysis, resulting in interest modeling remaining at the behavioral level and failing to address the emotional layer. This leads to a misalignment between recommendation results and users' true needs.

[0003] 2. User behavior data and facial behavior data are easily interfered by accidental factors. Existing methods often directly use the original data or perform simple normalization processing without establishing an effective data removal mechanism, which reduces the accuracy of the recommendation model.

[0004] 3. User interests are significantly dynamic over time. Existing methods mostly use static or slowly updated models, which make it difficult to capture sudden changes in user interests. However, traditional models may still recommend scientific content due to the excessive weighting of long-term interests, resulting in poor timeliness of recommendations.

[0005] 4. New users can only generate initial interests through registration information, resulting in highly generalized but poorly targeted content recommendations. New content struggles to reach target users through collaborative filtering, resulting in low exposure. Existing methods overly rely on users' historical interests, easily creating "information cocoons," which reduces users' exposure to new content and lowers long-term retention rates.

[0006] In summary, existing digital media recommendation methods have significant deficiencies in terms of behavioral data dimensions, data quality, dynamic adaptability, and cold start / diversity balance. There is an urgent need for digital media content recommendation optimization methods based on behavioral analysis to improve recommendation accuracy and user experience.

[0007] In order to solve the above problems, a technical solution is now provided. Summary of the Invention

[0008] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a digital media content recommendation optimization method based on behavior analysis to solve the problems raised in the above-mentioned background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions: The digital media content recommendation optimization method based on behavioral analysis includes the following steps: Step S1, collecting interactive behavior data between users and digital media content, including explicit behavior and implicit behavior; Step S2: When the user is browsing, the user requests to open the user's camera. After the user agrees, the user's face is captured through the camera, and the user's facial data is analyzed to obtain facial behavior data and transmit it to the interactive behavior data; Step S3: when a data point collection is obtained, analyzing the original frequency of the data point collection and calculating the dynamic frequency of the data point collection; Step S4, constructing a dynamic model integrating short-term interests and long-term interests; Step S5: extract multimodal features from digital media content and recommend digital media content.

[0010] In a preferred embodiment, explicit behaviors include clicks, favorites, likes, comments, and shares, and implicit behaviors include content dwell time, playback progress, scroll depth, fast forward and rewind times, and interaction frequency; Collect the user's facial expression feature dataset and extract the local features of the expression feature dataset, including the duration of each local feature and the residence time of two adjacent local features ; For every two adjacent local features, calculate the duration of each local feature and the residence time of two adjacent local features The sum of ; The duration of each local feature and the residence time of two adjacent local features Add the sum of the values to get the total value, that is ; The duration of each local feature Add together to get the total local feature time, that is ; Calculate facial behavior data, the calculation formula is: ; Where MB represents facial behavior data.

[0011] In a preferred embodiment, all data in the interactive behavior data, including clicks, favorites, likes, comments, shares, content dwell time, playback progress, scroll depth, fast forward and rewind times, and interaction frequency data, as well as facial behavior data, are marked as visited; Randomly select a visited data point S and query the neighborhood value of the data point S marked as , defining the range of points required to form a core neighborhood as ,like Less than or equal to or Greater than or equal to , then S is marked as a useless point; like Greater than and less than , then S is a key data, create a new data set with S as the key data, and add S to the created data set. All points in are added to a processing queue; Data points are extracted from the processing queue in sequence. If a data point is marked as useless, the useless data point is removed until all data points are extracted and a data point collection is obtained.

[0012] In a preferred embodiment, when a data point collection is obtained, the original frequency of the data point collection is analyzed and the dynamic frequency of the data point collection is calculated using the following formula: ; Where, They are dynamic frequency and original frequency respectively. is the instability coefficient, is the second controllable threshold of instability, for The dynamic frequency is used to replace the original frequency. If the result of the dynamic frequency is a decimal, determine whether the first decimal place of the decimal is greater than or equal to 5. If it is greater than or equal to 5, round it up; otherwise, round it down.

[0013] In a preferred embodiment, the dynamic frequency in the past 7 days is captured and the short-term interest vector is output. ; Through matrix decomposition, dynamic frequency modeling over 30 days is performed to capture users' stable preferences and output long-term interest vectors. ; The short-term interest vector and the long-term interest vector are fused to obtain the dynamic interest vector , calculated as: ; in , is the attenuation coefficient, The interval between the time when the behavior occurs and the current time.

[0014] In a preferred embodiment, content titles and descriptions are semantically encoded to generate text feature vectors, video keyframes or images are extracted to generate visual feature vectors, audio rhythm and pitch are extracted through Mel spectrum analysis to generate audio feature vectors, and content tags, release time, and author influence are combined to generate metadata feature vectors. The facial behavior data MB is taken as an independent feature vector and concatenated with the text feature vector, visual feature vector, audio feature vector, and metadata feature vector to form a content feature vector.

[0015] In a preferred embodiment, a deep fusion neural network is constructed, which inputs the user's dynamic interest vector and the content feature vector and outputs the user's interaction probability with the content; A multi-head attention mechanism is used to calculate the interaction weights between user interests and content features, and the user interaction probability is output through a fully connected layer and a sigmoid function.

[0016] In a preferred embodiment, initial interest tags are generated based on facial expression data, and mixed recommendations are made by combining high-profile content with content associated with facial features; Calculate the cosine similarity between the visual text features of new content and historical high facial attention content, and give priority to recommending it to users who have shown high facial attention to similar content.

[0017] The technical effects and advantages of the digital media content recommendation optimization method based on behavior analysis of the present invention are as follows: 1. By collecting the duration and adjacent dwell time of local facial expression features, this method quantifies facial behavior data, transforming users' emotional engagement in viewing content into calculable feature dimensions for the first time. Compared to existing technologies that rely solely on behavioral data such as clicks and dwell time, this method achieves a leap from the "behavioral layer" to the "emotional layer," enabling recommendation systems to capture users' subconscious emotional tendencies. 2. By dynamically calculating frequency and integrating short-term and long-term interest vectors, the system can adjust its recommendation strategy when user interests suddenly change. This improves response speed compared to traditional monthly update models and effectively addresses the issue of poor recommendation timeliness. 3. Dynamic frequency calculation uses instability coefficients and thresholds to filter out accidental behaviors such as false touches and temporary interruptions. It also combines a data point neighborhood filtering mechanism to remove useless data, making interest modeling more closely aligned with users' true preferences and preventing the recommendation system from deviating from users' core interests due to useless data. 4. New users generate initial interest tags based on facial expressions, and a mixed recommendation system combining highly popular content with content associated with facial features is used for recommendations. This significantly improves the click-through rate of initial recommendations compared to traditional registration information recommendations, solving the problem of generalized recommendations for new users with "zero behavioral data." New content is matched with historically high MB value content through visual / textual features, prioritizing users who have shown high facial interest in similar content. This significantly improves the click-through rate of new content exposure, alleviating the cold start dilemma of "difficulty in recommending without interactive data." BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the flow of the digital media content recommendation optimization method based on behavior analysis of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] Example 1

[0021] Figure 1 The present invention provides a method for optimizing digital media content recommendation based on behavior analysis, which specifically includes the following steps: Step S1, collecting interactive behavior data between users and digital media content, including explicit behavior and implicit behavior; Step S2: When the user is browsing, the user requests to open the user's camera. After the user agrees, the user's face is captured through the camera, and the user's facial data is analyzed to obtain facial behavior data and transmit it to the interactive behavior data; Step S3: when a data point collection is obtained, analyzing the original frequency of the data point collection and calculating the dynamic frequency of the data point collection; Step S4, constructing a dynamic model integrating short-term interests and long-term interests; Step S5: extract multimodal features from digital media content and recommend digital media content.

[0022] Explicit behaviors include clicks, favorites, likes, comments, and shares, while implicit behaviors include content dwell time, playback progress, scroll depth, fast-forward and rewind times, and interaction frequency. Collect the user's facial expression feature dataset and extract the local features of the expression feature dataset, including the duration of each local feature and the residence time of two adjacent local features ; For every two adjacent local features, calculate the duration of each local feature and the residence time of two adjacent local features The sum of ; The duration of each local feature and the residence time of two adjacent local features Add the sum of the values to get the total value, that is ; The duration of each local feature Add together to get the total local feature time, that is ; Calculate facial behavior data, the calculation formula is: ; Where MB represents facial behavior data.

[0023] This analysis not only covers explicit actions like clicks and favorites, but also explores users' subconscious emotional reactions and attention distribution through temporal analysis of local facial expression features (such as duration and adjacent dwell time). This avoids the one-sidedness of explicit behavioral data and provides a comprehensive portrayal of user behavior and psychological state. By calculating the sum and total of local feature time and adjacent dwell time, abstract facial expression changes are converted into quantifiable MB data, providing objective numerical evidence for assessing user emotional engagement and content appeal. This more accurately reflects the depth of user interaction with content than traditional behavioral analysis.

[0024] A comprehensive model that combines explicit and implicit behavioral data can more comprehensively capture user behavior patterns. For example, time series analysis of facial expression features can identify changes in user interests or fatigue in advance, providing more forward-looking decision support for personalized recommendations, content optimization, and more targeted user experience optimization.

[0025] Mark all data in the interactive behavior data, including clicks, favorites, likes, comments, shares, content dwell time, playback progress, scroll depth, fast forward and rewind times, interaction frequency, and facial behavior data, as visited; Randomly select a visited data point S and query the neighborhood value of the data point S marked as , defining the range of points required to form a core neighborhood as ,like Less than or equal to or Greater than or equal to , then S is marked as a useless point; like Greater than and less than , then S is a key data, create a new data set with S as the key data, and add S to the created data set. All points in are added to a processing queue; Data points are extracted from the processing queue in sequence. If a data point is marked as useless, the useless data point is removed until all data points are extracted and a data point collection is obtained.

[0026] In the process of collecting user behavior data, useless data such as accidental clicks and abnormal dwell time caused by device anomalies are inevitably generated. By defining the range of points required for the core neighborhood, useless points that deviate from the normal data distribution can be identified and marked, such as an occasional excessively long dwell time on content (perhaps caused by the user forgetting to close the page), and removed from the dataset, thereby improving the purity and reliability of the data and preventing useless data from interfering with subsequent analysis and modeling. Filter out key data and create a new dataset to ensure that the retained data points are representative and valuable. These key data can more accurately reflect the user's true behavior patterns and interest preferences. For example, data points where users frequently click, stay for a long time, and have positive facial expressions, when included in the key data, can provide a more accurate basis for the subsequent construction of user interest models, improving the recommendation system's ability to capture the user's true interests.

[0027] After forming a processing queue, data points are processed sequentially, avoiding blind processing of all data and reducing unnecessary computing resource consumption. By quickly eliminating useless points and focusing on key data, the data processing process becomes more efficient. This is particularly suitable for processing large-scale user behavior data, enabling data removal and integration to be completed in a relatively short period of time, supporting the real-time calculation of subsequent recommendation algorithms. The collection of data points after removal has higher data quality, and its distribution is more consistent with actual user behavior patterns. Inputting this data into the recommendation model can reduce errors and fluctuations during model training, making the model more stable and reliable. For example, when training a user interest prediction model, stable data can help the model learn more accurate user behavior patterns, thereby reducing the randomness and inaccuracy of recommendation results and improving the consistency and credibility of recommendations.

[0028] When a data point collection is obtained, the original frequency of the data point collection is analyzed and the dynamic frequency of the data point collection is calculated. The calculation formula is: ; Where, They are dynamic frequency and original frequency respectively. is the instability coefficient, is the second controllable threshold of instability, for The dynamic frequency is used to replace the original frequency. If the result of the dynamic frequency is a decimal, determine whether the first decimal place of the decimal is greater than or equal to 5. If it is greater than or equal to 5, round it up; otherwise, round it down.

[0029] User interests can change dynamically over time, due to factors such as hot events, and traditional statistical methods based on raw frequency cannot capture these changes. This formula introduces an instability coefficient and a threshold to dynamically adjust frequency based on fluctuations in user behavior data. When a user's interaction with a certain type of content suddenly increases or decreases (a change in the instability coefficient), dynamic frequency can quickly respond and more accurately capture short-term fluctuations in user interest. For example, if a user frequently browses a specific news item within a short period of time due to a hot event, dynamic frequency can promptly increase the weight of that content. User behavior data often contains accidental behaviors due to misuse or temporary needs, which can distort the assessment of a user's true interests. By setting a controllable instability threshold and a correction factor, this method can filter out abnormal frequency fluctuations caused by accidental behavior. For example, if a user accidentally clicks on a certain type of content, the system will not significantly increase the recommendation weight of that content due to this single action. Instead, by calculating dynamic frequency, the impact of this accidental behavior is minimized, thereby uncovering a more stable and authentic user's interests.

[0030] In the digital media landscape, content popularity and user interest are highly time-sensitive. Dynamic frequency calculation enables recommendation systems to promptly detect changing trends in user interests and prioritize content that aligns with current user interests. For example, during the release of a popular film or TV show, user interest in related film and television information and videos increases. With rapid dynamic frequency adjustments, the recommendation system can deliver relevant content in a timely manner, improving the timeliness of recommendations and user satisfaction. This, compared to recommendations based on fixed raw frequencies, can better meet users' current needs.

[0031] Capture the dynamic frequency in the past 7 days and output short-term interest vector ; Through matrix decomposition, dynamic frequency modeling over 30 days is performed to capture users' stable preferences and output long-term interest vectors. ; The short-term interest vector and the long-term interest vector are fused to obtain the dynamic interest vector , calculated as: ; in , is the attenuation coefficient, The interval between the time when the behavior occurs and the current time.

[0032] Focusing on activity frequency over the past seven days allows real-time responses to sudden changes in interest (e.g., users frequently viewing a certain type of content due to a hot topic), avoiding the dilution of short-term interests caused by traditional fixed-window statistics (e.g., 30 days). For example, if a user suddenly watches "concert" videos for three consecutive days, the short-term interest vector will rapidly increase the weight of this type of content, allowing the recommendation system to promptly push concert information. Using matrix decomposition to process activity frequency over 30 days, we filter out short-term unhelpful content and capture users' persistent interests (e.g., a long-term interest in "technology"). This prevents recommendations from deviating too far from the user's core preferences due to short-term fluctuations. For example, a user may occasionally watch entertainment content, but their long-term interest vector remains primarily focused on technology, ensuring a stable recommendation baseline. Recent activity (small t) has a large attenuation coefficient, indicating a high proportion of short-term interest; long-term activity (large t) has a small attenuation coefficient, indicating a dominant long-term preference, thus achieving a dynamic balance between "hotspot tracking" and "long-term interest." The high-frequency update of activity frequency over the past seven days enables the interest model to respond immediately to changes in user interests. For example, when a user's interest shifts from "watching TV dramas" to "fitness," the short-term interest vector updates rapidly, allowing the recommendation system to adjust recommendations within one to two days, a 4-5 times faster response than traditional monthly update models. Long-term interest vectors are modeled based on at least 30 days of data. Using matrix factorization, dimensionality reduction, and feature extraction, they reduce the impact of accidental behavior (such as accidental clicks and temporary activities) on interest representation. For example, if a user accidentally watches a "food" video, the long-term interest vector is minimally affected, preventing the recommendation system from being significantly skewed by useless data.

[0033] For users who are sensitive to hot topics, short-term interest vectors can promptly capture their high-frequency interactive behaviors, allowing the recommendation system to quickly switch to relevant content and increase user engagement. For example, if users frequently watch football content during the World Cup, short-term interest vectors will dominate recommendations, matching real-time needs. Users with stable interests: For users with clear long-term preferences, long-term interest vectors will dominate, ensuring that recommended content consistently matches core interests and avoiding irrelevant content recommendations due to short-term hot topics. For example, if a user has a long-term interest in "artificial intelligence," even if they occasionally watch entertainment videos, recommendations will still focus on technology content. By dynamically adjusting weights through attenuation factors, we can capture users' temporary interests without losing long-term preferences, achieving "elastic adaptation" of personalized recommendations.

[0034] The "hotspot exploration" mechanism introduced by short-term interest vectors enables the recommendation system to proactively recommend content in new areas as user interests shift, preventing them from being confined to a single interest category. For example, if a user who has long focused on "sports" shifts their short-term interest to "music," the recommendation system can promptly expand its recommendation scope and increase content diversity. Dynamic interest modeling ensures that recommended content is more closely aligned with the intersection of a user's current needs and historical preferences, reducing ineffective recommendations. For example, if a user has recently focused on "Graduate Entrance Exam" (short-term) and has a long-term preference for "English learning," the combined interest vector will prioritize content related to "Graduate Entrance Exam English," increasing click-through rate and user time. The complementarity of short-term and long-term interests enables the interest vector to both adapt to new content (short-term interests quickly connect) and leverage historical data (long-term interests provide prior knowledge), optimizing recommendations in cold-start scenarios (for example, new content can be matched with similar users based on their long-term interests).

[0035] The decay coefficient can be flexibly adjusted based on the business scenario (e.g., λ=0.1 for short video platforms and λ=0.05 for news platforms), balancing the speed of responsiveness to hot topics with the stability of long-term interests. For example, entertainment platforms increase the λ value to emphasize short-term interests and adapt to users' rapidly changing entertainment preferences. The separation of short-term and long-term interest vectors allows recommendation results to be traced back to users' recent behavior or historical preferences.

[0036] Semantically encode content titles and descriptions to generate text feature vectors. Extract video keyframes or images to generate visual feature vectors. Use Mel spectrum analysis to extract audio rhythm and pitch to generate audio feature vectors. Finally, combine content tags, release time, and author influence into metadata feature vectors. The facial behavior data MB is taken as an independent feature vector and concatenated with the text feature vector, visual feature vector, audio feature vector, and metadata feature vector to form a content feature vector.

[0037] Construct a deep fusion neural network, input the user's dynamic interest vector and content feature vector, and output the user's interaction probability with the content; A multi-head attention mechanism is used to calculate the interaction weights between user interests and content features, and the user interaction probability is output through a fully connected layer and a sigmoid function.

[0038] Generate initial interest tags based on facial expression data, and combine popular content with content related to facial features for mixed recommendations; Calculate the cosine similarity between the visual text features of new content and historical high facial attention content, and give priority to recommending it to users who have shown high facial attention to similar content.

[0039] Initial interest tags are generated based on facial expressions during the first visit, and mixed recommendations are made based on high-profile content and content associated with facial features. New users smile frequently when watching test videos, so the system generates an initial tag of "entertainment". 50% of the recommended list is high-profile entertainment videos, 30% is funny content with a historical high MB value, and 20% is exploratory content, quickly accumulating effective behavioral data.

[0040] Calculate the cosine similarity between the visual / text features of new content and historical high MB value content, and give priority to pushing it to users who have shown high facial attention to similar content (for example, if the visual features of a new technology video are similar to those of a historical high MB technology video, it will be pushed to users who have MB>0.7 when watching technology content).

[0041] The standard deviation of the MB value of the content in the recommended list must be ≥ a preset threshold (e.g., 0.3) to avoid recommending a high concentration of similar emotional content (e.g., continuously pushing negative news that triggers anxiety). For example, if a user's recent high MB value content is mostly "thrilling and suspenseful," the system will forcibly insert some "lighthearted and entertaining" content (low MB value) to balance the user's emotional experience and reduce fatigue.

[0042] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0043] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0044] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0045] In the several embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units described is merely a logical functional division, and actual implementation may have other division methods. In addition, the coupling or direct coupling or communication connection shown or discussed may be through some interface, indirect coupling or communication connection of devices or units, and may be electrical, mechanical, or other forms.

[0046] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0047] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0048] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0049] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0050] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A digital media content recommendation optimization method based on behavioral analysis, characterized in that: The steps include: Step S1, collecting interactive behavior data between users and digital media content, including explicit behavior and implicit behavior; Step S2: When the user is browsing, the user requests to open the user's camera. After the user agrees, the user's face is captured through the camera, and the user's facial data is analyzed to obtain facial behavior data and transmit it to the interactive behavior data; Step S3: when a data point collection is obtained, analyzing the original frequency of the data point collection and calculating the dynamic frequency of the data point collection; Step S4, constructing a dynamic model integrating short-term interests and long-term interests; Step S5: extract multimodal features from digital media content and recommend digital media content.

2. The method for optimizing digital media content recommendation based on behavioral analysis according to claim 1, characterized in that: Explicit behaviors include clicks, favorites, likes, comments, and shares, while implicit behaviors include content dwell time, playback progress, scroll depth, fast-forward and rewind times, and interaction frequency. Collect the user's facial expression feature dataset and extract the local features of the expression feature dataset, including the duration of each local feature and the residence time of two adjacent local features ; For every two adjacent local features, calculate the duration of each local feature and the residence time of two adjacent local features The sum of ; The duration of each local feature and the residence time of two adjacent local features Add the sum of the values to get the total value, that is ; The duration of each local feature Add together to get the total local feature time, that is ; Calculate facial behavior data, the calculation formula is: ； Where MB represents facial behavior data.

3. The method for optimizing digital media content recommendation based on behavioral analysis according to claim 2, characterized in that: Mark all data in the interactive behavior data, including clicks, favorites, likes, comments, shares, content dwell time, playback progress, scroll depth, fast forward and rewind times, interaction frequency, and facial behavior data, as visited; Randomly select a visited data point S and query the neighborhood value of the data point S marked as , defining the range of points required to form a core neighborhood as ,like Less than or equal to or Greater than or equal to , then S is marked as a useless point; like Greater than and less than , then S is a key data, create a new data set with S as the key data, and add S to the created data set. All points in are added to a processing queue; Data points are extracted from the processing queue in sequence. If a data point is marked as useless, the useless data point is removed until all data points are extracted and a data point collection is obtained.

4. The method for optimizing digital media content recommendation based on behavioral analysis according to claim 3, characterized in that: When a collection of data points is obtained, the original frequency of the collection of data points is analyzed and the dynamic frequency of the collection of data points is calculated. The calculation formula is: ; Where, They are dynamic frequency and original frequency respectively. is the instability coefficient, is the second controllable threshold of instability, for The dynamic frequency is used to replace the original frequency. If the result of the dynamic frequency is a decimal, determine whether the first decimal place of the decimal is greater than or equal to 5. If it is greater than or equal to 5, round it up; otherwise, round it down.

5. The method for optimizing digital media content recommendation based on behavioral analysis according to claim 4, characterized in that: Capture the dynamic frequency in the past 7 days and output short-term interest vector ; Through matrix decomposition, dynamic frequency modeling over 30 days is performed to capture users' stable preferences and output long-term interest vectors. ; The short-term interest vector and the long-term interest vector are fused to obtain the dynamic interest vector , calculated as: ； in , is the attenuation coefficient, The interval between the time when the behavior occurs and the current time.

6. The method for optimizing digital media content recommendation based on behavioral analysis according to claim 5, characterized in that: Semantically encode content titles and descriptions to generate text feature vectors. Extract video keyframes or images to generate visual feature vectors. Use Mel spectrum analysis to extract audio rhythm and pitch to generate audio feature vectors. Finally, combine content tags, release time, and author influence into metadata feature vectors. The facial behavior data MB is taken as an independent feature vector and concatenated with the text feature vector, visual feature vector, audio feature vector, and metadata feature vector to form a content feature vector.

7. The method for optimizing digital media content recommendation based on behavioral analysis according to claim 6, characterized in that: Construct a deep fusion neural network, input the user's dynamic interest vector and content feature vector, and output the user's interaction probability with the content; A multi-head attention mechanism is used to calculate the interaction weights between user interests and content features, and the user interaction probability is output through a fully connected layer and a sigmoid function.

8. The method for optimizing digital media content recommendation based on behavioral analysis according to claim 7, characterized in that: Generate initial interest tags based on facial expression data, and combine popular content with content related to facial features for mixed recommendations; Calculate the cosine similarity between the visual text features of new content and historical high facial attention content, and give priority to recommending it to users who have shown high facial attention to similar content.

Images