Multimedia content accurate prediction and push method based on user portrait

Abstract

The invention belongs to the technical field of multimedia content pushing, and discloses a multimedia content accurate prediction and pushing method based on a user portrait. Comprising the following steps: collecting data information of a user on a login content platform, and constructing a user portrait; obtaining content data of the multimedia content, and constructing a multimedia content portrait; matching the user portrait with the multimedia content portrait, and determining to-be-pushed content; based on feedback data analysis of the user, judging whether the pushed content to be pushed is accurate or not; and when the to-be-pushed content is inaccurate, correspondingly adjusting the to-be-pushed content. According to the method, the hidden core content preference characteristics of the user are mined according to the dominant core content preference characteristics of the user, the potential preference requirements of the user can be mined, the subsequently recommended multimedia content fits the real interest of the user and covers the preference of the user which is not directly exposed, recommendation can accurately reach the deep preference of the user, and the user experience is improved. Therefore, the user experience and the stickiness are improved.

Description

A method for accurate prediction and delivery of multimedia content based on user profiles Technical Field

[0001] This invention relates to the field of multimedia content delivery technology, and more specifically, to a method for accurate prediction and delivery of multimedia content based on user profiles. Background Technology

[0002] With the widespread adoption of the internet and mobile smart devices, multimedia content such as video, audio, and text has experienced explosive growth, leading to information overload for users. The cost of acquiring effective content that meets their needs continues to rise, while content platforms urgently need to improve content dissemination efficiency and user engagement. Against this backdrop, multimedia content recommendation technology based on user profiles has gradually become a core solution in the industry.

[0003] Existing content platforms primarily recommend multimedia content to users based on their historical browsing history. This approach yields limited results, potentially trapping users in their existing information cocoons and failing to effectively uncover their potential interests, leading to a lack of interest exploration and ultimately user churn. Furthermore, current technologies rely heavily on historical user behavior data to build user profiles, while new user registration only provides static basic data, lacking behavioral data support. This results in delayed new user profile building, with initial pushes consisting mostly of random content or broadly tagged content, leading to extremely low content relevance to user needs. This can result in a poor first-screen experience for new users and a high short-term churn rate.

[0004] In view of this, the present invention proposes a method for accurate prediction and push of multimedia content based on user profiles to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the existing technology and achieve the above objectives, this invention provides the following technical solution: a method for accurate prediction and push of multimedia content based on user profiles, comprising: Step 1, collecting user data information on the content platform they log in to, constructing a user profile, the data information including static basic data, viewing record data, and content preference data; Step 2, acquiring content data of multimedia content, and analyzing the content data of multimedia content to obtain multimedia content profiles for each multimedia content; Step 3, matching the user profiles with the multimedia content profiles to determine the predicted multimedia content to be pushed; Step 4, selecting multimedia content of the same type as the predicted multimedia content to be pushed as the content to be pushed, pushing it to users according to the content to be pushed, obtaining user feedback data, and analyzing the feedback data to determine whether the content to be pushed is accurate; Step 5, when the content to be pushed is inaccurate, adjusting the content to be pushed accordingly.

[0006] Furthermore, the method for constructing user profiles is as follows: When a user is a new user, based on the acquired static basic data of the new user, it is input into a pre-built matching coefficient generation model to obtain the matching coefficients of each prominent content preference feature. These prominent content preference features are then sorted in descending order according to their matching coefficients, and the top-ranked prominent content preference feature is selected as the new user's content preference feature, thus constructing the new user's user profile. When a user is an existing user, the existing user's content preference data is acquired and analyzed to obtain a set of explicit core content preference features and a set of implicit core content preference features. The explicit core content preference features within the explicit core content preference feature set are then... The implicit core content preference features in the explicit and implicit core content preference feature sets are sorted in descending order. The top Z explicit core content preference features in the explicit core content preference feature set and the top X implicit core content preference features in the implicit core content preference feature set are taken as the content preference features of old users, and user profiles of old users are constructed. The descending order sorting method is as follows: the explicit core content preference features in the explicit core content preference feature set are sorted in descending order according to the frequency of occurrence of each explicit core content preference feature, and the implicit core content preference features in the implicit core content preference feature set are sorted in descending order according to the frequency of occurrence of each implicit core content preference feature.

[0007] Furthermore, the method for constructing the matching coefficient generation model is as follows: Static basic data from different users is collected, and static basic data of the same data type are grouped together to obtain multiple sets of static basic data combinations; data clustering is performed on each set of static basic data combinations to obtain multiple grouping categories for each set, and corresponding identification labels are set for each grouping category; the static basic data of each user is converted into corresponding identification labels to obtain a combination of identification labels for all users; the identification label combinations are clustered to obtain multiple subgroups, and a corresponding set of group identification labels is constructed based on the identification labels of each subgroup; based on the static basic data of the input new user, it is converted into corresponding identification labels to construct a set of user identification labels; the Jaccard similarity between the set of user identification labels and the set of group identification labels is calculated to obtain the Jaccard similarity of each subgroup, and the maximum Jaccard similarity is selected as the overlap coefficient; the preference coefficients of each prominent content preference feature are obtained, and the preference coefficients are multiplied by the overlap coefficients to obtain the matching coefficients of each prominent content preference feature.

[0008] Furthermore, the method for obtaining the preference coefficient is as follows: Content preference data of users in each segmented group is obtained; features of the content preference data are extracted to obtain multiple content preference features; the number of occurrences of each content preference feature in each segmented group is counted and recorded as the content preference feature quantity; the content preference feature quantity is compared with a preset quantity threshold, and content preference features whose quantity is greater than the threshold are selected and recorded as prominent content preference features; the proportion of each prominent content preference feature quantity in the corresponding segmented group is calculated and recorded as the proportion degree; the mutual information between each prominent content preference feature and all segmented groups is calculated and recorded as the specificity degree; the proportion degree and specificity degree of each prominent content preference feature are multiplied to obtain the preference coefficient of each prominent content preference feature.

[0009] Furthermore, the method for obtaining the explicit core content preference feature set is as follows: Obtain content preference data of existing users in different scenarios; based on content preference data in the same scenario, obtain multiple content preference feature tags; count the number of times each content preference feature tag appears in the same scenario, denoted as the tag occurrence count; designate content preference feature tags whose tag occurrence count is greater than the tag occurrence count threshold in the same scenario as explicit core content preference features, and designate content preference feature tags whose tag occurrence count is less than or equal to the tag occurrence count threshold in the same scenario as explicit non-core content preference features; integrate the multiple explicit core content preference features obtained in each scenario to obtain a set of explicit core content preference features; count the number of times the same explicit core content preference feature appears in the set of explicit core content preference features, denoted as the same occurrence count; filter out explicit core content preference features in the set of explicit core content preference features whose same occurrence count is greater than the same occurrence count threshold, and construct the explicit core content preference feature set.

[0010] Furthermore, the method for obtaining the implicit core content preference feature set is as follows: Obtain user viewing history data, extract multimedia content from the viewing history data that contains both explicit core content preference features and explicit non-core content preference features, and construct a multimedia content library; combine a single explicit core content preference feature with a single explicit non-core content preference feature to obtain multiple feature combinations; in the multimedia content library, count the frequency of each feature combination, the frequency of each explicit core content preference feature, and the frequency of each explicit non-core content preference feature, and calculate the support, confidence, and lift of each feature combination; then compare the support, confidence, and lift with the set of implicit core content preference features. The support threshold, confidence threshold, and lift threshold are compared to select feature combinations that meet the following conditions, which are denoted as candidate feature combinations. The conditions are: support greater than the support threshold, confidence greater than the confidence threshold, and lift greater than the lift threshold. It is then determined whether the explicit core content preference features corresponding to each candidate feature combination belong to the explicit core content preference feature set. If not, the corresponding candidate feature combination is removed. If it does, the corresponding candidate feature combination is retained, and the explicit non-core content preference features in the corresponding candidate feature combination are extracted and denoted as implicit core content preference features. Based on each implicit core content preference feature, an implicit core content preference feature set is constructed.

[0011] Furthermore, the method for determining the predicted multimedia content to be pushed is as follows: Content data for each multimedia content is acquired; features are extracted from the multimedia content to obtain content preference features for each multimedia content; and a multimedia content profile for each multimedia content is constructed. Each multimedia content profile is compared with a user profile, and the number of content preference features corresponding to the user profile in each multimedia content profile is counted, denoted as the first number. The total number of content preference features in each multimedia content profile is counted, denoted as the second number. The ratio of the first number to the second number is recorded as the conformity rate, and the conformity rate of each multimedia content is obtained. The conformity rate of each multimedia content is compared with a preset conformity rate judgment threshold, and multimedia content with a conformity rate greater than the conformity rate judgment threshold is selected and recorded as the predicted multimedia content to be pushed.

[0012] Furthermore, the method for determining the accuracy of the content to be pushed is as follows: Multimedia content of the same type as the predicted multimedia content to be pushed is taken as the content to be pushed. Based on the different types of multimedia content within the content to be pushed, the content to be pushed is divided into multiple types. Feedback data is obtained each time a user watches the same type of content to be pushed; the feedback data is quantified to obtain the quantified values ​​of each feedback indicator; the quantified values ​​of each feedback indicator are weighted and accumulated to obtain the corresponding feedback score; the feedback score of the user after watching the same type of content to be pushed is obtained at time series t. In the XY coordinate system, with time as the X-axis and the feedback score as the Y-axis, a [data structure is constructed]. Construct a feedback rating change function; obtain the area enclosed by the feedback rating change function and the X-axis at time series t, normalize the area and record it as the first judgment indicator; obtain the average growth rate of the feedback rating at time series t, obtain the rate of change of the feedback rating at adjacent time points, take the maximum rate of change as the instantaneous rate of change, calculate the mean of the average growth rate and the instantaneous rate of change, record it as the growth value, normalize the growth value and record it as the second judgment indicator; add the first judgment indicator and the second judgment indicator to obtain the judgment value; when the judgment value is greater than the preset judgment threshold, the content to be pushed is judged to be accurate; otherwise, the content to be pushed is judged to be inaccurate.

[0013] Furthermore, the method for adjusting the content to be pushed is as follows: when it is determined that the content to be pushed is inaccurate, the same type of content to be pushed is removed and no further push is made; when it is determined that the content to be pushed is accurate, the same type of content to be pushed is retained, and push is made based on the established push strategy.

[0014] Furthermore, the method for formulating the push strategy is as follows: set K push time periods, obtain the number of times and duration of playback of each type of content to be pushed by users in each push time period, normalize the playback number and playback duration and then add them together to obtain the suitability of each type of content to be pushed in each push time period; sort the content to be pushed in each type in descending order according to the suitability in each push time period, and push the content of the same type that ranks higher first.

[0015] The technical effects and advantages of the multimedia content accurate prediction and push method based on user profiles in this invention are as follows: When a user is a new user, this invention can quickly generate a profile using only the user's static basic data, without waiting for the user to generate behavioral data, which can significantly shorten the new user profile construction cycle and effectively solve the problem of cold start for new users; at the same time, it adopts multi-level quantitative calculation of proportion, exclusivity and overlap coefficient, so that the generated new user profile fits the common preferences of the segmented group, and can also filter out the differentiated features exclusive to the group, thereby improving the preference accuracy of the new user profile.

[0016] This invention discovers users' implicit core content preference characteristics based on their explicit core content preference characteristics. It can uncover users' potential preference needs, so that the recommended multimedia content not only matches users' real interests, but also covers preferences that users have not directly revealed. This allows recommendations to accurately reach users' deep preferences, thereby improving user experience and stickiness. Attached Figure Description

[0017] Figure 1 is a flowchart of the steps of the multimedia content accurate prediction and push method based on user profile in Embodiment 1 of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Embodiment 1:

[0019] This embodiment discloses a method for accurate prediction and push of multimedia content based on user profiles. Please refer to Figure 1. The method mainly includes: Step 1: Collecting user data information from the content platform they log into to construct a user profile. The data information includes static basic data, viewing history data, and content preference data. Step 2: Obtaining the content data of multimedia content and analyzing it to obtain multimedia content profiles for each multimedia content. Step 3: Matching the user profiles with the multimedia content profiles to determine the predicted multimedia content to be pushed. Step 4: Selecting multimedia content of the same type as the predicted multimedia content to be pushed as the content to be pushed, and pushing it to users based on the content to be pushed, obtaining user feedback data, and analyzing the feedback data to determine whether the pushed content is accurate. Step 5: When the content to be pushed is inaccurate, adjusting it accordingly.

[0020] Through the above technical solution, this embodiment first collects user data information on the content platform they log in to, and then constructs a user profile based on this data. This data includes static basic data, viewing history data, and content preference data. Static basic data consists of basic information filled in by the user when logging into the content platform, such as age, gender, city, and interest tags. Content preference data reflects the user's liking and selection tendencies for content during interaction with multimedia content, such as preferred content types (workplace short videos, healing audio, etc.) and preferred content styles (funny, science fiction, workplace, etc.). Viewing history data is the user's historical data on watching multimedia content on the content platform, including the content and type of the multimedia content watched. Then, based on the content data of the multimedia content, which includes content duration, content tags, keywords, and theme style, feature extraction is performed to obtain corresponding content preference features, thereby constructing a multimedia content profile. The process involves creating a multimedia content profile for each user. This profile is then matched with the user profile to select multimedia content that matches the user profile, which is then used as predicted multimedia content to be pushed to users. Multimedia content of the same type as the predicted multimedia content is then used as the content to be pushed to users. Based on the different types of multimedia content within the content to be pushed, this content is categorized into multiple types. The categorized content is then pushed to users, and user feedback data is obtained. This feedback data includes user completion rate, click-through rate, number of comments, number of favorites, and dwell time. Based on the feedback data analysis, the accuracy of the pushed content of the same type is determined. If the determination is accurate, it means that the type of multimedia content matches the user's preferences, and a corresponding push strategy is developed to push the same type of content. If the determination is inaccurate, it means that the type of multimedia content does not match the user's preferences, and the same type of content is removed and will not be pushed again, thus completing the multimedia content recommendation process.

[0021] The method for constructing user profiles is as follows: When a user is a new user, based on the acquired static basic data of the new user, it is input into a pre-built matching coefficient generation model to obtain the matching coefficients of each prominent content preference feature. These prominent content preference features are then sorted in descending order according to their matching coefficients, and the top-ranked prominent content preference feature is selected as the new user's content preference feature, thus constructing the new user's user profile. The method for constructing the matching coefficient generation model is as follows: Static basic data from different users is collected, and static basic data of the same data type are grouped together to obtain multiple sets of static basic data combinations. Data clustering is then performed on each set of static basic data combinations to obtain the data for each group. The system combines static basic data into multiple group categories, assigning corresponding identification labels to each group category. Each user's static basic data is converted into corresponding identification labels, resulting in a combination of identification labels for all users. These identification label combinations are then clustered to obtain multiple subgroups, and a corresponding group identification label set is constructed based on the identification labels of each subgroup. Based on the input static basic data of new users, it is converted into corresponding identification labels, forming a user identification label set. The Jaccard similarity between the user identification label set and the group identification label set is calculated to obtain the Jaccard similarity for each subgroup, and the highest Jaccard similarity is selected as the... The overlap coefficient is used to obtain content preference data of users in each segmented group. Content preference data features are extracted to obtain multiple content preference features. The number of occurrences of each content preference feature in each segmented group is counted and recorded as the content preference feature quantity. The content preference feature quantity is compared with a preset quantity threshold, and content preference features with a quantity greater than the threshold are selected and recorded as prominent content preference features. The proportion of each prominent content preference feature in the corresponding segmented group is calculated and recorded as the proportion degree. The mutual information between each prominent content preference feature and all segmented groups is calculated and recorded as the specificity degree. The proportion degree and specificity degree of each prominent content preference feature are multiplied to obtain the value of each prominent content preference feature. The preference coefficients of content preference features are obtained; the preference coefficients are multiplied by the overlap coefficients to obtain the matching coefficients of each prominent content preference feature; when the user is an old user, the content preference data of the old user is obtained and analyzed to obtain the explicit core content preference feature set and the implicit core content preference feature set. The explicit core content preference features in the explicit core content preference feature set and the implicit core content preference features in the implicit core content preference feature set are sorted in descending order respectively. The top Z explicit core content preference features in the explicit core content preference feature set and the top X implicit core content preference features in the implicit core content preference feature set are taken as the content preference features of the old user, and a user profile of the old user is constructed.The descending order sorting method is as follows: Within the explicit core content preference feature set, the explicit core content preference features are sorted in descending order based on the frequency of occurrence of each explicit core content preference feature; similarly, within the implicit core content preference feature set, the implicit core content preference features are sorted in descending order based on the frequency of occurrence of each implicit core content preference feature.

[0022] The method for obtaining the explicit core content preference feature set is as follows: Obtain content preference data from existing users in different scenarios; based on content preference data in the same scenario, obtain multiple content preference feature tags; count the frequency of each content preference feature tag in the same scenario, denoted as the tag occurrence count; designate content preference feature tags whose occurrence count in the same scenario is greater than a tag occurrence count threshold as explicit core content preference features, and designate content preference feature tags whose occurrence count in the same scenario is less than or equal to the tag occurrence count threshold as explicit non-core content preference features; integrate the multiple explicit core content preference features obtained from each scenario to obtain a set of explicit core content preference features; count the frequency of identical explicit core content preference features in the set of explicit core content preference features, denoted as the "identical frequency"; filter out explicit core content preference features in the set of explicit core content preference features whose identical frequency is greater than the identical frequency threshold, and construct the explicit core content preference feature set. The method for obtaining the implicit core content preference feature set is as follows: Obtain user viewing history data, extract data from the viewing history data containing both explicit core content preference features and explicit non-core content preference features. Multimedia content is used to build a multimedia content library. A single explicit core content preference feature is combined with a single explicit non-core content preference feature to obtain multiple feature combinations. In the multimedia content library, the frequency of occurrence of each feature combination, each explicit core content preference feature, and each explicit non-core content preference feature is counted, and the support, confidence, and lift of each feature combination are calculated. The support, confidence, and lift are compared with set support, confidence, and lift thresholds, respectively, to select feature combinations that meet the criteria. Let each of these be a candidate feature combination. The conditions for selection are: simultaneously satisfying a support greater than the support threshold, a confidence threshold greater than the confidence threshold, and a lift greater than the lift threshold. Determine whether the explicit core content preference features corresponding to each candidate feature combination belong to the explicit core content preference feature set. If not, remove the corresponding candidate feature combination. If it does, retain the corresponding candidate feature combination and extract the explicit non-core content preference features from the corresponding candidate feature combination, denoted as implicit core content preference features. Based on each implicit core content preference feature, construct the implicit core content preference feature set.

[0023] The above technical solution provides a specific method for constructing user profiles. Since users logging into the content platform may be new users who have never used the platform before or returning users who have used it multiple times, users are divided into new users and returning users, and separate profiles are constructed for each category. Specifically, when a user is a new user, static basic data of different users is first collected, such as age, gender, city, and interest tags. Then, static basic data of the same data type are grouped together, resulting in multiple sets of static basic data combinations, such as age groups, gender groups, city groups, and interest tag groups. Data clustering is then performed on each set of static basic data combinations to obtain multiple grouping categories for each set. Each group category is assigned a corresponding identification label; for example, within an age group, the age group is divided into multiple subgroups based on age range, with those aged 18-23 in one group and those aged 24-30 in another, each group assigned an identification label; then, each user's static basic data is converted into a corresponding identification label, resulting in a combination of identification labels for all users. These identification label combinations are then clustered to obtain multiple subgroups, and a corresponding group identification label set is constructed based on the identification labels of each subgroup; based on the static basic data of the new user input, it is converted into a corresponding identification label, and a user identification label set is constructed; the Jaccard ratio between the user identification label set and the group identification label set is calculated. Similarity was calculated by obtaining the Jaccard similarity for each subgroup, and the highest Jaccard similarity was selected as the overlap coefficient. Jaccard similarity reflects the degree of similarity between two sets; therefore, a higher overlap coefficient indicates a closer resemblance between the new user and that subgroup. Content preference data was obtained from users in each subgroup, and features were extracted to obtain multiple content preference features. Content preference data refers to the degree of liking and selection tendency of users during their interaction with multimedia content, while content preference features are the labeled representations of each content preference data point. For example, in a certain subgroup, 120 instances of short workplace videos appeared during historical multimedia content interactions. 70 instances of food-related images and text, 85 instances of commuting audio, 95 instances of long-form entertainment videos, and 90 instances of science fiction movies yield the following content preference characteristics: short workplace videos, food-related images and text, commuting audio, long-form entertainment videos, and science fiction movies. The number of occurrences of each content preference characteristic in each segment is counted and recorded as the content preference characteristic count. This count is compared to a preset threshold, which can be determined by those skilled in the art based on experience and professional knowledge. Content preference characteristics with a count greater than the threshold are selected and recorded as prominent content preference characteristics. For example, if the threshold is set to 80, the prominent content preference characteristics would be long-form entertainment videos, science fiction movies, short workplace videos, and commuting audio.Calculate the proportion of each prominent content preference feature in the corresponding subgroup, and record it as the proportion. For example, if the total number of subgroups is 200, then the proportion of workplace short videos is 120 / 200=0.6. Calculate the mutual information between each prominent content preference feature and all subgroups, and record it as the exclusivity. For example, if a certain subgroup accounts for 20% of all groups, then the exclusivity of the corresponding workplace short video is . 1.1; Multiply the proportion and specificity of each prominent content preference feature to obtain the preference coefficient of each prominent content preference feature; the proportion represents the high-frequency preference of the prominent content preference feature in the corresponding group, and the specificity represents the differentiated preference of the prominent content preference feature in the group; finally, multiply the preference coefficient and overlap coefficient to obtain the matching coefficient of each prominent content preference feature, thus completing the construction of the matching coefficient generation model; then, input the obtained static basic data of new users into the pre-built matching coefficient generation model to obtain the matching coefficient of each prominent content preference feature, sort the prominent content preference features in descending order according to the matching coefficient, and select the top C prominent content preference feature as the content preference feature of the new user, as the user profile of the new user; this method can quickly generate profiles with only the static basic data of users, without waiting for users to generate behavioral data, which can significantly shorten the new user profile construction cycle and effectively solve the cold start problem of new users; at the same time, the multi-level quantitative calculation of proportion, specificity and overlap coefficient makes the generated new user profile fit the common preferences of the segmented group, and can also filter out the differentiated features specific to the group, thereby improving the preference accuracy of the new user profile.

[0024] When a user is a returning user, content preference data for that user in different scenarios is obtained. Based on content preference data within the same scenario, multiple content preference feature tags are derived. The frequency of each content preference feature tag appearing within the same scenario is counted and recorded as the tag occurrence count. Content preference feature tags whose frequency in the same scenario exceeds a certain threshold are designated as explicit core content preference features, while those whose frequency is less than or equal to the threshold are designated as explicit non-core content preference features. The tag occurrence count threshold can be determined by those skilled in the art based on experience and professional knowledge. The higher the frequency of a content preference feature tag, the higher the user's preference level. Therefore, the same scenario... Content preference features whose frequency of occurrence in a given scene exceeds a certain threshold are denoted as explicit core content preference features. Then, multiple explicit core content preference features obtained from various scenes are integrated to form a set of explicit core content preference features. The frequency of identical explicit core content preference features within this set is counted and denoted as the "same frequency." Explicit core content preference features with the same frequency exceeding a certain threshold are selected to construct the explicit core content preference feature set. The "same frequency" threshold is determined by those skilled in the art based on experience and professional knowledge. The higher the frequency of a particular explicit core content preference feature across all scenes, the higher the user's preference level. Therefore, we select high-frequency explicit core content preference features appearing in all scenarios to construct an explicit core content preference feature set, which can accurately reflect the user's explicit needs. We acquire user viewing history data, which is the historical data of users watching multimedia content on content platforms, including the content and type of the viewed multimedia content. Based on the content of each multimedia content, we determine the corresponding content preference features. We extract multimedia content from the viewing history data that contains both explicit core content preference features and explicit non-core content preference features to build a multimedia content library. Finally, we combine a single explicit core content preference feature with a single explicit non-core content preference feature to obtain multiple feature combinations. In the multimedia content library, the frequency of occurrence of each feature combination, each explicit core content preference feature, and each explicit non-core content preference feature is counted. The support, confidence, and lift of each feature combination are then calculated. The support, confidence, and lift are compared with set support, confidence, and lift thresholds, respectively, to filter out feature combinations that meet the following criteria, which are then recorded as candidate feature combinations. The criteria are: simultaneously, support is greater than the support threshold, confidence is greater than the confidence threshold, and lift is greater than the lift threshold. Higher support, confidence, and lift indicate a stronger correlation between the explicit core content preference features and the explicit non-core content preference features in the feature combination.It should be noted that the methods for calculating support, confidence, and lift are all existing technologies and will not be described in detail here. The support threshold, confidence threshold, and lift threshold are determined based on the professional knowledge of those skilled in the art. The process involves determining whether the explicit core content preference features corresponding to each candidate feature combination belong to the explicit core content preference feature set. If not, the corresponding candidate feature combination is removed. If it does, the corresponding candidate feature combination is retained, and the explicit non-core content preference features in the corresponding candidate feature combination are extracted and denoted as implicit core content preference features. Based on each implicit core content preference feature, an implicit core content preference feature set is constructed. By using this method to discover the user's implicit core content preference features based on the user's explicit core content preference features, the user's potential needs can be uncovered. This ensures that the subsequently recommended multimedia content not only matches the user's true interests but also covers preferences that the user has not directly expressed, enabling recommendations to accurately reach the user's deep preferences and thus improve the user experience.

[0025] Finally, based on the acquired explicit core content preference feature set and implicit core content preference feature set, the explicit core content preference features in the explicit core content preference feature set and the implicit core content preference features in the implicit core content preference feature set are sorted in descending order. The top Z explicit core content preference features in the explicit core content preference feature set and the top X implicit core content preference features in the implicit core content preference feature set are taken as the content preference features of existing users, forming the user profile of existing users. The descending order sorting method is as follows: the explicit core content preference features in the explicit core content preference feature set are sorted in descending order according to the frequency of occurrence of each explicit core content preference feature, and the implicit core content preference features in the implicit core content preference feature set are sorted in descending order according to the frequency of occurrence of each implicit core content preference feature. In this way, the user profile of existing users is constructed based on the user's explicit preferences (the user's clearly expressed needs) and implicit related preferences (potential related needs), making the constructed profile more closely match the real needs of existing users, thereby effectively improving the stickiness of existing users.

[0026] The method for determining the predicted multimedia content to be pushed is as follows: Content data for each multimedia content is acquired; features are extracted from the multimedia content to obtain content preference features for each multimedia content, and a multimedia content profile is constructed for each multimedia content; each multimedia content profile is compared with a user profile, and the number of content preference features corresponding to the user profile in each multimedia content profile is counted, denoted as the first quantity; the total number of content preference features in each multimedia content profile is counted, denoted as the second quantity; the ratio of the first quantity to the second quantity is denoted as the conformity rate, and the conformity rate of each multimedia content is obtained; the conformity rate of each multimedia content is compared with a preset conformity rate judgment threshold, and multimedia content with a conformity rate greater than the conformity rate judgment threshold is selected and recorded as the predicted multimedia content to be pushed.

[0027] The above technical solution provides a specific method for determining the predicted multimedia content to be pushed. First, content data for each multimedia content is acquired, and features are extracted from the multimedia content to obtain content preference features for each multimedia content, denoted as a multimedia content profile. Then, the multimedia content profile is compared with the user profile, and the number of content preference features corresponding to the user profile in each multimedia content profile is counted, denoted as the first quantity. The larger the first quantity, the more content preference features the multimedia content profile contains, indicating that the multimedia content is more relevant to the user. Finally, the total number of content preference features contained in each multimedia profile is counted, denoted as the second quantity. The second quantity represents the total number of content preference features contained in the multimedia content. The system calculates the matching rate by considering all the data collected. The ratio of the first data collection to the second data collection is recorded as the matching rate. This yields the matching rate for each multimedia content item. A higher matching rate indicates that the multimedia content is more relevant to the user. Therefore, based on the experience and expertise of those in the field, a matching rate threshold is set. The matching rate of each multimedia content item is compared with the preset matching rate threshold. Multimedia content items with matching rates greater than the threshold are selected. A higher matching rate indicates that the multimedia content matches the user's needs more closely, and these are recorded as predicted multimedia content to be pushed. In this way, the matching rate is calculated based on the number of user profile feature matches, directly anchoring the user's explicit preferences and avoiding the recommendation of multimedia content unrelated to the user profile, thus ensuring the accuracy of the recommendation.

[0028] The method for determining the accuracy of the content to be pushed is as follows: Multimedia content of the same type as the predicted multimedia content to be pushed is used as the content to be pushed. Based on the different types of multimedia content within the content to be pushed, the content is divided into multiple types. Feedback data is obtained each time a user watches the same type of content to be pushed. This feedback data is then quantified to obtain quantified values ​​for each feedback indicator. The quantified values ​​of each feedback indicator are weighted and summed to obtain the corresponding feedback score. The feedback score after a user watches the same type of content to be pushed is obtained at time series t. In an XY coordinate system, with time as the X-axis and the feedback score as the Y-axis, a reverse... The feedback score change function is used to calculate the area enclosed by the feedback score change function and the X-axis at time series t. This area is normalized and recorded as the first judgment indicator. The average growth rate of the feedback score at time series t is obtained, along with the rate of change of the feedback score at adjacent time points. The maximum rate of change is taken as the instantaneous rate of change. The average growth rate and the instantaneous rate of change are calculated and recorded as the growth value. This growth value is normalized and recorded as the second judgment indicator. The first and second judgment indicators are added together to obtain the judgment value. If the judgment value is greater than a preset judgment threshold, the content to be pushed is judged to be accurate; otherwise, the content to be pushed is judged to be inaccurate.

[0029] The above technical solution provides a specific method for determining the accuracy of the content to be pushed. First, multimedia content of the same type as the predicted multimedia content to be pushed is considered as the content to be pushed. Based on the different types of multimedia content within the content to be pushed, the content to be pushed is divided into multiple types, for example, based on the different content preference characteristics of each multimedia content. For content to be pushed of the same type, feedback data from each user viewing is obtained. This feedback data includes completion rate, click-through rate, number of comments, number of favorites, and dwell time. The feedback data is then quantified to obtain the quantified values ​​of each feedback indicator, which are the corresponding completion rate, click-through rate, etc. The system considers factors such as user behavior, comments, favorites, and dwell time. The quantified values ​​of each feedback metric are determined by mapping their actual performance to a pre-defined range; higher values ​​indicate better performance. The quantified values ​​of each metric are weighted and summed to obtain a corresponding feedback score. Higher scores indicate a stronger user preference for the multimedia content. Feedback scores are obtained for each user's viewing of similar content at time series t. A feedback score variation function is constructed in an XY coordinate system, with time as the X-axis and the feedback score as the Y-axis. The area enclosed by this function and the X-axis at time series t is then normalized. After normalization, the first judgment indicator is defined as the overall cumulative performance of user feedback ratings for the same type of content to be pushed within the time series t. The larger the area, the higher the overall level of feedback ratings and the better the sustained performance, reflecting the user's long-term acceptance and stable preference for this type of content. The average growth rate of feedback ratings over time series t is obtained, as are the rates of change of feedback ratings at adjacent time points. The maximum rate of change is taken as the instantaneous rate of change. The average growth rate and the instantaneous rate of change are calculated and denoted as the growth value. After normalization, the growth value is defined as the second judgment indicator. The second judgment indicator indicates the intensity of dynamic changes in user feedback ratings and can comprehensively reflect the user's... The overall trend and sudden fluctuations in user preferences reflect the potential for improvement in content delivery. Finally, the first and second judgment indicators are added together to obtain the judgment value. The preset judgment threshold can be determined by those skilled in the art based on their experience and professional knowledge. When the judgment value is greater than the preset judgment threshold, the content to be pushed is judged to be accurate; otherwise, the content to be pushed is judged to be inaccurate. By combining the first and second judgment indicators in this way, the limitations of only looking at the overall feedback level and ignoring changes in user interests are avoided, as is the misjudgment of the long-term adaptability of content based on short-term fluctuations. This allows for a more comprehensive and accurate assessment of the accuracy of pushing similar multimedia content.

[0030] The method for adjusting the content to be pushed is as follows: when it is determined that the content to be pushed is inaccurate, the same type of content to be pushed is removed and no longer pushed; when it is determined that the content to be pushed is accurate, the same type of content to be pushed is retained, and it is pushed according to the established push strategy. The method for establishing the push strategy is as follows: set K push time periods, obtain the number of times users watch each type of content to be pushed and the playback duration in each push time period, normalize the number of times and playback duration and then add them together to obtain the suitability of each type of content to be pushed in each push time period; within each push time period, sort the content to be pushed in each type of content in descending order according to the suitability, and push the same type of content that ranks higher first.

[0031] The above technical solution provides a specific method for adjusting the content to be pushed. When it is determined that the content to be pushed is inaccurate, it means that the same type of content does not meet the user's preferences, so the inaccurate content of the same type is removed and will not be pushed again. When it is determined that the content to be pushed is accurate, the same type of content is retained, and a corresponding push strategy is formulated for pushing. Specifically, based on the user's viewing habits, K push time periods are set, and the number of times and duration of each type of content to be pushed are obtained within each push time period. The number of times and duration of viewing are normalized and then added together to obtain the adaptability of each type of content to be pushed in each push time period. The higher the adaptability, the more popular the content of that type is with users during that time period. Within each push time period, the content of each type is sorted in descending order according to the adaptability, and the content of the same type that ranks higher is pushed first. The higher the ranking, the higher the user's preference, and the higher the push content of that type will be pushed in the corresponding push time period. By dividing the push into different time periods and combining the number of times and duration of playback by users within each push period to calculate the suitability, it is possible to accurately capture users' content preferences in different time periods, so that the push of multimedia content can be accurately adapted to users' time-based preferences.

[0032] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0033] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0034] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for accurate prediction and push of multimedia content based on user profiles, characterized in that, The method includes: Step 1, collecting user data on content platforms to build user profiles, including static basic data, viewing history data, and content preference data; Step 2, acquiring multimedia content data and analyzing it to obtain multimedia content profiles for each type of multimedia content; Step 3, matching user profiles with multimedia content profiles to determine predicted multimedia content to be pushed; Step 4, selecting multimedia content of the same type as the predicted multimedia content to be pushed as the content to be pushed, pushing it to users, obtaining user feedback data, and analyzing the feedback data to determine the accuracy of the pushed content; Step 5, adjusting the pushed content accordingly when it is inaccurate.

2. The method for accurate prediction and push of multimedia content based on user profiles according to claim 1, characterized in that, The method for constructing user profiles is as follows: When a user is a new user, the static basic data of the new user is obtained and input into a pre-constructed matching coefficient generation model to obtain the matching coefficients of each prominent content preference feature. The prominent content preference features are sorted in descending order according to the matching coefficients, and the top C prominent content preference features are selected as the content preference features of the new user and the user profile of the new user is constructed. When a user is a returning user, their content preference data is acquired and analyzed to obtain explicit core content preference feature sets and implicit core content preference feature sets. The explicit core content preference features in the explicit core content preference feature set and the implicit core content preference features in the implicit core content preference feature set are then sorted in descending order. The top Z explicit core content preference features in the explicit core content preference feature set and the top X implicit core content preference features in the implicit core content preference feature set are taken as the returning user's content preference features, and a user profile of the returning user is constructed. The descending order sorting method is as follows: the explicit core content preference features in the explicit core content preference feature set are sorted in descending order based on the frequency of occurrence of each explicit core content preference feature, and the implicit core content preference features in the implicit core content preference feature set are sorted in descending order based on the frequency of occurrence of each implicit core content preference feature.

3. The method for accurate prediction and push of multimedia content based on user profiles according to claim 2, characterized in that, The matching coefficient generation model is constructed as follows: Static basic data from different users is collected, and static basic data of the same data type are grouped together to obtain multiple sets of static basic data combinations. Data clustering is performed on each set of static basic data combinations to obtain multiple grouping categories, and a corresponding identification label is assigned to each grouping category. The static basic data of each user is converted into corresponding identification labels to obtain a combination of identification labels for all users. The identification label combinations are clustered to obtain multiple subgroups, and a corresponding set of group identification labels is constructed based on the identification labels of each subgroup. Based on the static basic data of the new user input, it is converted into corresponding identification labels to construct a user identification label set. Calculate the Jaccard similarity between the set of user identification tags and the set of group identification tags to obtain the Jaccard similarity of each subgroup, and select the maximum Jaccard similarity as the overlap coefficient. Obtain the preference coefficients for each prominent content preference feature, and multiply the preference coefficients by the overlap coefficients to obtain the matching coefficients for each prominent content preference feature.

4. The method for accurate prediction and push of multimedia content based on user profiles according to claim 3, characterized in that, The method for obtaining the preference coefficient is as follows: obtain the content preference data of users in each segmented group, extract the features of the content preference data, and obtain multiple content preference features; The number of occurrences of each content preference feature in each subgroup is counted and denoted as the content preference feature count. The number of content preference features is compared with a preset threshold, and content preference features whose number of content preference features is greater than the threshold are selected and recorded as prominent content preference features. Calculate the proportion of each prominent content preference feature in the corresponding subgroup, and record it as the proportion. Calculate the mutual information between each prominent content preference feature and all subgroups, and denote it as the specificity; multiply the proportion of each prominent content preference feature by the specificity to obtain the preference coefficient of each prominent content preference feature.

5. The method for accurate prediction and push of multimedia content based on user profiles according to claim 2, characterized in that, The method for obtaining explicit core content preference feature set is as follows: obtain content preference data of old users in different scenarios, and obtain multiple content preference feature labels based on content preference data in the same scenario; The frequency of occurrence of each content preference feature tag in the same scenario is counted and recorded as the tag occurrence count. Content preference features whose frequency of occurrence in the same scene is greater than the threshold of the frequency of occurrence of the label are recorded as explicit core content preference features, and content preference features whose frequency of occurrence in the same scene is less than or equal to the threshold of the frequency of occurrence of the label are recorded as explicit non-core content preference features. The multiple explicit core content preference features obtained from various scenarios are integrated to obtain a set of explicit core content preference features; The number of times the same explicit core content preference feature appears in the set of explicit core content preference features is recorded as the number of times they appear. Filter out explicit core content preference features that have the same frequency greater than the threshold from the set of explicit core content preference features, and construct an explicit core content preference feature set.

6. The method for accurate prediction and push of multimedia content based on user profiles according to claim 2, characterized in that, The method for obtaining the implicit core content preference feature set is as follows: obtain the user's viewing record data, extract the multimedia content in the viewing record data that contains both explicit core content preference features and explicit non-core content preference features, and build a multimedia content library. By combining a single explicit core content preference feature with a single explicit non-core content preference feature, multiple feature combinations can be obtained. In the multimedia content library, the frequency of occurrence of each feature combination, each explicit core content preference feature, and each explicit non-core content preference feature is counted. The support, confidence, and lift of each feature combination are calculated. The support, confidence, and lift are compared with set support, confidence, and lift thresholds, respectively, to filter out feature combinations that meet the following conditions, which are recorded as candidate feature combinations. The conditions for meeting these conditions are: simultaneously satisfying a support threshold greater than the support threshold, a confidence threshold greater than the confidence threshold, and a lift threshold greater than the lift threshold. It is then determined whether the explicit core content preference feature corresponding to each candidate feature combination belongs to the explicit core content preference feature set; if not, the corresponding candidate feature combination is removed. If it belongs to the category, the corresponding candidate feature combination is retained, and the explicit non-core content preference features in the corresponding candidate feature combination are extracted and denoted as implicit core content preference features. Based on the implicit core content preference features, a set of implicit core content preference features is constructed.

7. The method for accurate prediction and push of multimedia content based on user profiles according to claim 1, characterized in that, The method for determining the predicted multimedia content to be pushed is as follows: obtain the content data of each multimedia content, extract features from the multimedia content, obtain the content preference features of each multimedia content, and construct a multimedia content profile for each multimedia content. Compare each multimedia content profile with the user profile, and count the number of content preference features corresponding to the user profile in each multimedia content profile, which is recorded as the first number. The total number of content preference features contained in each multimedia content profile is counted and denoted as the second number; the ratio of the first number to the second number is denoted as the conformity rate, and the conformity rate of each multimedia content is obtained; the conformity rate of each multimedia content is compared with the preset conformity rate judgment threshold, and the multimedia content with a conformity rate greater than the conformity rate judgment threshold is selected and denoted as the predicted multimedia content to be pushed.

8. The method for accurate prediction and push of multimedia content based on user profiles according to claim 1, characterized in that, The method for determining the accuracy of the content to be pushed is as follows: Multimedia content of the same type as the predicted multimedia content to be pushed is used as the content to be pushed. Based on the different types of multimedia content within the content to be pushed, the content is divided into multiple types. Feedback data is obtained each time a user watches the same type of content to be pushed. This feedback data is then quantified to obtain quantified values ​​for each feedback indicator. The quantified values ​​of each feedback indicator are weighted and summed to obtain the corresponding feedback score. The feedback score after a user watches the same type of content to be pushed is obtained at time series t. In an XY coordinate system, with time as the X-axis and the feedback score as the Y-axis, a reverse... The feedback score change function is used to calculate the area enclosed by the feedback score change function and the X-axis at time series t. This area is normalized and recorded as the first judgment indicator. The average growth rate of the feedback score at time series t is obtained, along with the rate of change of the feedback score at adjacent time points. The maximum rate of change is taken as the instantaneous rate of change. The average growth rate and the instantaneous rate of change are calculated and recorded as the growth value. This growth value is normalized and recorded as the second judgment indicator. The first and second judgment indicators are added together to obtain the judgment value. If the judgment value is greater than a preset judgment threshold, the content to be pushed is judged to be accurate; otherwise, the content to be pushed is judged to be inaccurate.

9. The method for accurate prediction and push of multimedia content based on user profiles according to claim 1, characterized in that, The method for adjusting the content to be pushed is as follows: if it is determined that the content to be pushed is inaccurate, remove the same type of content to be pushed and stop pushing it; if it is determined that the content to be pushed is accurate, retain the same type of content to be pushed and push it based on the established push strategy.

10. The method for accurate prediction and push of multimedia content based on user profiles according to claim 9, characterized in that, The method for formulating the push strategy is as follows: set K push time periods, obtain the number of times and duration of playback of each type of content to be pushed by users in each push time period, normalize the playback number and playback duration and then add them together to obtain the suitability of each type of content to be pushed in each push time period; sort the content to be pushed in each type in descending order according to the suitability in each push time period, and push the content of the same type that ranks first.

Images