Disease knowledge popular science video recommendation method and system

Abstract

This invention relates to the field of reinforcement learning technology and discloses a method and system for recommending popular science videos on disease knowledge. The method includes: acquiring user account information, medical records, and user habits as recommendation information; storing and recording address tags based on multi-level classification tags; calculating the correlation between users using a hierarchical weighted fusion mechanism based on account information to generate a correlation distribution map; responding to a logged-in user's video push request, matching the recommendation information with the stored address tags for coarse selection to obtain a coarse selection range of videos; calculating the push value for each popular science video based on the coarse selection range and the correlation distribution map, and generating a video recommendation list by sorting the videos in descending order of the push values; and sending popular science videos to terminal devices sequentially according to the order of the recommendation list. This invention prioritizes pushing videos commented on by users with low social relevance to the user, promoting community interaction while protecting health privacy and improving the effectiveness of popular science dissemination.

Description

Technical Field

[0001] This invention relates to the field of reinforcement learning technology, and in particular to a method and system for recommending popular science videos on disease knowledge. Background Technology

[0002] Disease knowledge dissemination, as a crucial component of public health intervention, plays an irreplaceable role in improving public health literacy, achieving early disease prevention, and reducing the social burden of healthcare. Through the effective dissemination of authoritative medical information, the public can establish a scientific framework for understanding diseases, proactively identify high-risk factors, improve unhealthy lifestyle habits, and seek professional medical attention promptly when suspected symptoms appear, thereby significantly reducing the incidence and misdiagnosis rates of preventable diseases. Particularly in the field of chronic disease prevention and control, continuous science education has been proven to effectively improve patients' self-management abilities and long-term health outcomes.

[0003] With the deep penetration of mobile internet, short videos, with their advantages of high information density, rapid dissemination, and strong audience acceptance, have become the core form of disease knowledge dissemination. Medical and health institutions, science popularization creators, and platforms use algorithm-driven push mechanisms to transform professional medical content into easy-to-understand short videos, precisely reaching the target audience. This model breaks down traditional barriers to accessing health information while also building a real-time science popularization network covering hundreds of millions of users.

[0004] However, the current short video recommendation logic of mainstream platforms relies heavily on the correlation matching of users' multi-dimensional attributes. This makes it highly likely that users with similar health concerns or social attributes will be recommended disease-related science content on the same topic. When users realize that the disease-related science content they are watching may be recognized by their social circles, they will actively suppress their interaction due to concerns about the leakage of their personal health privacy, generally avoiding public actions such as commenting or liking, thus forming a widespread phenomenon of "silent viewing." While this mechanism improves the efficiency of content delivery, it severely weakens users' willingness to engage in public discussion, resulting in significantly lower community interaction activity for this type of science content compared to other types of content. This makes it difficult to form a healthy information exchange ecosystem and ultimately restricts the spread of the social influence of science content.

[0005] Therefore, a disease knowledge popularization video recommendation scheme is needed to solve the problem of limited dissemination effectiveness of popular science videos. Summary of the Invention

[0006] The purpose of this invention is to address the technical problem that existing disease knowledge popularization video recommendation mechanisms rely too heavily on user attribute correlation matching, leading to users suppressing their interaction due to privacy concerns, resulting in low community activity and limited dissemination of popular science videos. This invention provides a method and system for recommending disease knowledge popularization videos that, while protecting users' health and privacy sensitivity, balances accurate content matching with community interaction activity. By prioritizing popular science videos commented on by users with low social relevance to the user, it reduces users' inhibition of interaction due to concerns about being recognized by acquaintances, promotes the natural occurrence of public interactions such as comments and likes, and thus enhances the dissemination effect and social impact of popular science content.

[0007] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0008] Recommended methods for watching disease knowledge videos include:

[0009] S1. Obtain user account information, medical record information, and operation habit information; use the medical record information and operation habit information as recommendation information; store the recommendation information based on multi-level classification tags and record the storage address tags; calculate the correlation between users based on the account information using a hierarchical weighted fusion mechanism, and generate a correlation distribution map for each user according to the correlation gradient;

[0010] S2. Responding to the video push request from the logged-in user, a coarse selection is performed based on the matching of the logged-in user's recommendation information and storage address tags to obtain a coarse selection range of videos;

[0011] S3. Based on the coarse selection range of the videos and the correlation distribution map, calculate the push value of each science video within the coarse selection range of the videos, sort them in descending order according to the push value of the science videos, and generate the final video recommendation list.

[0012] S4. Send science videos to the logged-in user's terminal device in the order of the video recommendation list.

[0013] Furthermore, the multi-level classification labels include regional level, age level, gender level, disease type level, and treatment method level.

[0014] Further, in step S1, the step of calculating the correlation between users based on the account information using a hierarchical weighted fusion mechanism, and generating a correlation distribution map for each user according to the correlation gradient, includes:

[0015] Define the elements for calculating the degree of association;

[0016] The weighted results of each relationship are combined to generate the correlation between the two users;

[0017] Several correlation levels are generated with the correlation gradient as the center. The correlation calculation results between users and other users are entered into the corresponding levels to generate a correlation distribution map, which is then dynamically updated.

[0018] Furthermore, the correlation calculation elements include direct relationships, indirect relationships, and similar relationships; among which, direct relationships include IP address relationships, family relationships, work unit relationships, and friend relationships; indirect relationships include secondary direct relationships and comment interaction relationships; and similar relationships include user active geographical location relationships, age similarity relationships, and gender similarity relationships.

[0019] Further, in step S3, the step of calculating the push value for each science video within the coarse selection range includes:

[0020] S31. Extract the commenters of the first Q science videos from the coarse selection range corresponding to the logged-in user;

[0021] S32. Select N anchoring correlation calculation elements from K types of correlation calculation elements; with the N anchoring correlation calculation elements converging and the remaining non-anchoring correlation calculation elements differing as the guide, calculate the matching degree between the logged-in user and all comment users under each science popularization video; rank the science popularization videos based on the matching degree to obtain the initial list of video recommendations;

[0022] S33. Calculate the association risk value between the logged-in user and each science video in the initial video recommendation list in sequence;

[0023] S34. Obtain the overlap between each science video in the initial video recommendation list and the logged-in user;

[0024] S35. Sum the associated risk value and overlap with weights, and use the result as the push value for the science video; sort the science videos in descending order of their push values ​​to generate the final video recommendation list.

[0025] Furthermore, the method introduces a reinforcement learning adaptive optimization mechanism to replace or supplement the push value calculation process in S3, specifically including:

[0026] Z1. Establish a reinforcement learning model and map each logged-in user to an agent i. The calculation process of the push value of any popular science video in the initial video recommendation list is formalized as a triple H=(S,A,R), where S represents the state space, A represents the action space, and R represents the reward.

[0027] Z2. Encode all possible candidate actions that agent i can choose at the current time step t into a candidate action dictionary;

[0028] Z3. Use a greedy algorithm to select the best action from the candidate action dictionary and update the state of the next time step;

[0029] Z4: The selected action to be actually executed. Applied to the current state s t And update the state s for the next time step. t+1 .

[0030] Furthermore, the triple H=(S,A,R) in Z1 is specifically:

[0031] State-space function: ;

[0032] Where t represents the index of the time step, and T represents the total number of time steps; m t s represents the comment depth at time step t; t b represents the state at time step t; t This represents the vector of elements selected for anchoring correlation calculation.

[0033] Action space function: ;

[0034] in, This represents the action of agent i at time step t; Choose the amount of change in the vector for the comment depth at time step t; Calculate the element operation vector for the anchoring correlation degree at time step t;

[0035] Reward function:

[0036] ;

[0037] in, This represents the reward value of agent i at time step t; Indicates the weighting factor; This represents the risk reward value for anchoring correlation. Indicates the sensitivity coefficient; Indicates risk cost; This represents the efficiency cost reward value.

[0038] Furthermore, the Z3 specifically includes:

[0039] Z31. Obtain the action with the highest cumulative reward value among all past actions of agent i, and use it as the selection position in the candidate action dictionary;

[0040] Z32, calculate the average viewing time T of the three most recent science popularization videos for the logged-in user corresponding to agent i. max The average viewing time T maxThe linear mapping function is used to transform the algorithm into radius RGD, and the selection position of the candidate action dictionary is determined by Z31. Using RGD as the radius and center, a selection space is generated from the candidate action dictionary;

[0041] Z33. Select the best action from the selection space.

[0042] Furthermore, the Z33 specifically includes:

[0043] Z331. Pre-set the exploration probability ε; select the generation path of the action at the current time step based on the exploration probability ε;

[0044] Z332. When the exploration probability ε is used as the probability, it is based on the reward value. and cumulative reward value Choose the best action Calculate the candidate actions to be actually executed. The reward value of the new state after combining with the current state and the reward value Add to the candidate action dictionary;

[0045] Z333. When the probability is 1-ε, randomly select an action from the selection space as the optimal action.

[0046] A disease knowledge popularization video recommendation system, comprising a correlation distribution calculation module, a video coarse selection module, a video push value calculation module, and a video recommendation module:

[0047] The correlation distribution calculation module is used to obtain user account information, medical record information, and operation habit information, and use the medical record information and operation habit information as recommendation information; store the recommendation information based on multi-level classification tags and record the storage address tags; calculate the correlation between users based on the account information using a hierarchical weighted fusion mechanism, and generate a correlation distribution map for each user according to the correlation gradient;

[0048] The video coarse selection module is used to respond to the video push request of the logged-in user, and to perform coarse selection based on the matching of the login user's recommendation information and storage address tags to obtain the video coarse selection range;

[0049] The video push value calculation module is used to calculate the push value of each science video within the coarse selection range of the video and the correlation distribution map, sort the science videos in descending order of their push values, and generate the final video recommendation list.

[0050] The video recommendation module is used to send popular science videos to the logged-in user's terminal device in the order of the video recommendation list.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] (1) Privacy protection and interaction promotion: By prioritizing the recommendation of videos with low correlation (such as weak correlation features such as geographical location / age / gender), the system reduces users' concerns about health and privacy leaks, eliminates the psychological barrier of "silent viewing", and thus naturally increases the frequency of public interaction behaviors such as liking and commenting.

[0053] (2) Precise content matching: Based on a dual matching mechanism of multi-level classification tags (disease type / treatment method, etc.) and user case information (medical history / medication, etc.) and operating habits (viewing rate / attention behavior, etc.), the content pushed is highly matched with individual health management needs, and the effectiveness of knowledge absorption is enhanced.

[0054] (3) System efficiency adaptive optimization: By utilizing the dynamic optimization mechanism of the agent in the state space (anchor element selection, comment filtering dimension), the reward function (positive correlation of efficiency, negative correlation of association risk) autonomously balances the recommendation accuracy and privacy protection strength, so as to achieve continuous optimization of resource consumption and push timeliness. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a schematic flowchart illustrating the method for recommending disease knowledge popular science videos provided in an embodiment of the present invention.

[0057] Figure 2 The structural block diagram of the disease knowledge popular science video recommendation system provided in the embodiments of the present invention. Detailed Implementation

[0058] 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0059] It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations. Additionally, the terms "connected," "linked," etc., can refer to a direct connection between elements, components, modules, etc., or an indirect connection via other elements, components, modules, etc.

[0060] Example 1:

[0061] This invention is achieved through the following technical solutions, such as... Figure 1 As shown, this invention provides a method for recommending disease knowledge popular science videos, the method comprising:

[0062] S1. Obtain user account information, medical record information, and operation habit information, and use the medical record information and operation habit information as recommendation information; store the recommendation information based on multi-level classification tags and record the storage address tags; calculate the correlation degree between users based on the account information using a hierarchical weighted fusion mechanism, and generate a correlation degree distribution map for each user according to the correlation degree gradient.

[0063] Users initiate registration requests through their terminal devices, and the system collects the user's publicly disclosed content registration information on social platforms. This content registration information includes, but is not limited to, nicknames, avatars, and publicly set geographic information.

[0064] The system uses the aforementioned registration information to generate a unique account for the user within the system. This account information includes anonymized internal ID codes and attribute fields (such as geographic region text) extracted from a public dataset. Account information refers to a user's unique identifier on a third-party communication platform and the associated public dataset, which includes basic information such as nickname, avatar, and region.

[0065] Account information is a de-identified data set generated by parsing non-sensitive public data from social media accounts. It is used for internal system-wide anchoring correlation calculations and does not contain the original social media account's private content, including chat logs and friend lists. Based on this account information, the system obtains the user's medical records by accessing the hospital's HIS system and monitors interaction logs to obtain the user's operational habits. This medical record information and operational habits information are then used as recommendation information. Specifically, after obtaining explicit authorization from the user, the system accesses the hospital's HIS system to obtain the user's medical records. Simultaneously, it collects the user's operational habits in real time by monitoring the interaction logs between the terminal device and the system. The medical record information and operational habits information are then used as recommendation information for that user.

[0066] For example, the system retrieves user A's social media account, which includes their public information: nickname "Health Guru," location set to "XX City XX District," and profile picture a sunflower. By directly copying this information, account information is generated for one-click registration. The registered account information includes the nickname "Health Guru," the geographical location text "XX City XX District," and the profile picture a sunflower. Furthermore, the account information also includes a unique internal ID code used to identify the user.

[0067] The case information refers to digital medical data generated during hospital visits that reflects a user's health status and disease management requirements. In some embodiments, the case information includes the user's past medical history, treatment information, nursing information, and medication information. Past medical history refers to the name of the disease the user has had, the date of diagnosis, and the stage of the disease; treatment information refers to the treatment plans, surgical procedures, and physical therapies the user has received; nursing information refers to the rehabilitation measures the user has received; and medication information refers to the name, dosage, and method of use of medications the user is currently using or has previously used.

[0068] The user behavior information includes the number of times a user likes or comments on science videos, the number of views, the number of follow actions, the number of interactive actions, and the skip rate. Specifically, the view rate is the percentage of a single video's viewing time out of the total video duration; follow actions refer to the frequency and nature of a user's active follow of videos tagged with a specific disease; interactive actions refer to the trigger timestamps and corresponding video IDs for likes, comments, and favorites; and the skip rate is the percentage of times a user closes a video before reaching the required viewing time out of the total number of actions.

[0069] The recommendation information is stored based on multi-level classification tags, and the quality value is updated according to user operation feedback within the same storage address tag.

[0070] The multi-level classification tags refer to a hierarchical set of classification identifiers used for structured storage and retrieval of science popularization videos. In some embodiments, the multi-level classification tags include: regional level, age level, gender level, disease type level, and treatment method level.

[0071] In some embodiments, the system generates a corresponding storage address tag for each science video. The storage address tag is formed by concatenating the aforementioned multi-level classification tags in a preset order. For science videos with the same storage address tag, the system updates the quality value of each science video in real time based on user operation feedback and arranges them in descending order of quality value within the storage address tag.

[0072] Specifically, taking science video A as an example, the initial quality value of science video A was set to 0.8; after one week of operation, the following user feedback data was collected:

[0073] The viewing completion rate was 92%, corresponding to a weighting coefficient of 0.5;

[0074] The like rate is 15%, corresponding to a weighting coefficient of 0.3;

[0075] The collection rate is 8%, corresponding to a weighting coefficient of 0.2;

[0076] The updated quality score for science video A is as follows:

[0077]

[0078] The updated quality value of 1.321 is higher than the initial quality value, indicating that the science video A has received good user feedback. The system has promoted its ranking within the same storage address tag to the top, thus giving it a priority recommendation opportunity.

[0079] The system calculates the anchoring correlation degree between each user and any other user based on each user's account information. Specifically, firstly, correlation degree calculation elements are defined. As a possible implementation, these elements include direct relationships, indirect relationships, and similarity relationships. Direct relationships include IP address relationships, family relationships, work unit relationships, and friend relationships; indirect relationships include secondary direct relationships and comment interaction relationships; similarity relationships include user activity location geographical location relationships, age similarity relationships, and gender similarity relationships. It should be noted that this embodiment does not limit the types and number of correlation degree calculation elements; this embodiment is merely an example.

[0080] Secondary direct relationships refer to the number of third-party users who have a direct relationship with the target user (such as the number of mutual friends), with weight decreasing at each level of relationship (e.g., the weight of a friend of a friend = 0.3 × the weight of a direct relationship). Comment interaction relationships refer to the frequency of comment interactions between the two parties on the same video; the higher the frequency, the higher the weight.

[0081] The geographical similarity of users' active regions is weighted by the degree of overlap of administrative regions (e.g., the same community = 0.4, the same block = 0.3, the same street = 0.2); the age similarity is weighted by the absolute value of the age difference (e.g., ≤5 = 0.2, decreasing by 0.05 for every 5 years); the gender similarity is weighted by 0.1 for the same gender and 0 for different genders.

[0082] In some embodiments, the correlation calculation employs a hierarchical weighted fusion mechanism, specifically:

[0083] (1) Prioritize direct relationships: If any direct relationship exists, the anchoring correlation degree is directly assigned the highest weight (e.g., 1.0), and other relationships are not calculated.

[0084] (2) Indirect relationships are secondary: If there is no direct relationship, then secondary direct relationships and comment interaction relationships are calculated;

[0085] (3) Supplementing similarity relations: When both direct and indirect relations are missing, similarity relations are calculated.

[0086] Then, the weighted results of each relationship are combined to generate the correlation between the two users.

[0087] Next, the system generates a correlation distribution map for each user based on the correlation degree: several correlation degree levels are generated with the correlation degree gradient as the center, the correlation degree calculation results of the user and other users are entered into the corresponding level, a correlation distribution map is generated, and the correlation distribution map is dynamically updated.

[0088] S2. In response to the video push request from the logged-in user, a coarse selection is performed based on the matching of the logged-in user's recommendation information and storage address tags to obtain a coarse selection range of videos.

[0089] In response to a logged-in user's video push request, similar to S1, the system obtains the logged-in user's account information, medical records, and operating habits information. The medical records and operating habits information are used as recommendation information, and the recommendation information is stored based on multi-level classification tags, with the storage address tag recorded.

[0090] Based on the matching of the logged-in user's recommendation information and storage address tags, a coarse selection is performed to obtain the coarse selection range of videos. Specifically, when a logged-in user sends a video push request through a terminal device, the system parses the logged-in user's recommendation information, performs a step-by-step logical mapping between the recommendation information and storage address tags, and filters out all storage address tag sets that match the tags, forming the coarse selection range of videos for the logged-in user (i.e., the initial video pool).

[0091] For example, based on the recommendation information of logged-in user C, the system recommends 3 storage address tags. The first storage address tag contains 10 science videos, the second storage address tag contains 13 science videos, and the third storage address tag contains 4 science videos. Therefore, the coarse selection range of videos for logged-in user C is a total of 27 science videos.

[0092] S3. Based on the coarse selection range of videos and the correlation distribution map, calculate the push value of each popular science video within the coarse selection range of videos, and generate a video recommendation list by sorting the push values ​​in descending order.

[0093] The push score is a quantitative metric used to measure the limited extent to which a science video should be recommended to logged-in users. The higher the push score, the higher the ranking of the science video in the video recommendation list.

[0094] The higher the correlation between a logged-in user and the commenters of a science video, the lower the push value of that science video (inversely proportional); the higher the overlap between the logged-in user's recommendation information and the interest tags of that science video, the higher the push value of that science video (directly proportional).

[0095] The specific steps the system takes to calculate the push value for each science video include:

[0096] S31. Extract the comment users of the first Q science videos from the coarse selection range corresponding to the logged-in user.

[0097] S32. Select N anchoring correlation calculation elements from K types of correlation calculation elements (prioritizing those with similar relationships); calculate the matching degree between the logged-in user and all comment users under each science popularization video, guided by the convergence of the N anchoring correlation calculation elements and the dissimilarity of the remaining non-anchoring correlation calculation elements; rank the science popularization videos based on the matching degree to obtain the initial list of video recommendations.

[0098] It should be noted that the criteria for determining direct relationships are too precise. If direct relationships are used as the anchor element for calculating correlation, then a certain type of similarity is required among the direct relationships. For example, similarity in friendships, family relationships, or IP address relationships may lead to a "familiarity effect."

[0099] Furthermore, given the current abundance of video types and users, directly describing interpersonal relationships is too straightforward and makes it difficult to segment users into sufficiently large groups. In contrast, choosing less strongly related relationships like similarity as anchoring elements for correlation calculations is preferable. Firstly, the principle of convergence in anchoring correlation calculation elements reduces the likelihood of acquaintances meeting in the comments section. Secondly, the similarity criteria are broader, encompassing more elements and allowing for the segmentation of more user groups. Subsequent users integrating into these segmented comment sections will have even fewer opportunities to encounter familiar faces.

[0100] For example, if IP address relationships are used as the anchor element for calculating relevance, then people in the comment section of the same video would all have the same IP address, but different information elsewhere. On the one hand, people with the same IP address might already know each other, but the system's information is unclear.

[0101] On the other hand, there are relatively few people with the same IP address but different other information. This will limit the spread of the video.

[0102] Therefore, using similarity as the anchoring element for calculating relevance is the most cost-effective approach. For example, assuming there are 9 relevance calculation elements (K=9), S32 selects 4 anchoring elements (N=4), and calculates the matching degree (K=N+V) based on the convergence of these 4 anchoring elements and the dissimilarity of the remaining 5 non-anchoring elements (V=5).

[0103] The formula for calculating the matching degree is as follows:

[0104] ;

[0105] Where q represents the index of the science video; i represents the index of the logged-in user; This represents the match degree between logged-in user i and the q-th science video; , These represent the first and second weight parameters, respectively; n represents the index of the anchoring correlation calculation element, N represents the total number of anchoring correlation calculation elements, n=1,2,...,N; j represents the index of the commenting user in the popular science video, J represents the total number of commenting users in the popular science video, j=1,2,...,J; This represents the nth anchoring correlation calculation element for the j-th commenting user; This represents the nth anchored correlation calculation element for logged-in user i; v represents the index of the non-anchored correlation calculation element, V represents the total number of non-anchored correlation calculation elements, v=1,2,...,V; K represents the total number of correlation calculation elements, K=N+V; This represents the v-th non-anchored correlation element of the j-th commenting user. This represents the v-th non-anchored association degree calculation element of logged-in user i;

[0106] This indicates a positive number indicator function, which is defined when the following conditions are met. hour, Select 1 if the value is 1, otherwise select 0. This indicates that in the science video q, all commenting users and logged-in user i have the same number of elements in the anchoring correlation calculation;

[0107] This indicates a negative number indicator function, which is defined when the following conditions are met. hour, Select 1 if the value is 1, otherwise select 0. This indicates the number of commenting users and logged-in user i that are different in the non-anchored association degree calculation elements in the popular science video q.

[0108] The above provides the formula for calculating the matching degree. In a further implementation, the non-anchored correlation calculation elements are assigned corresponding weights according to direct, indirect, and similar relationships. For example, direct relationships are assigned the highest weight, indirect relationships the second highest weight, and similar relationships the lowest weight. If logged-in user i and commenting user j have the same non-anchored correlation calculation element (belonging to a direct relationship), their matching degree will be sharply increased due to the high weight. The value of this item directly leads to a decrease in the matching degree.

[0109] Furthermore, in the matching degree calculation formula The specific calculation formula for this item is:

[0110]

[0111] ;

[0112] Where U1 represents the total number of non-anchored correlation elements in the direct relationship, v=1,2,...,U1; U2 represents the total number of non-anchored correlation elements in the indirect relationship, v=U1+1,...,U1+U2; U3 represents the total number of non-anchored correlation elements in the similarity relationship, v=U1+U2+1,...,U1+U2+U3; , , These represent the first non-anchored weight parameter, the second non-anchored weight parameter, and the third non-anchored weight parameter, respectively.

[0113] Optionally, different weight coefficients can be set for non-anchored association degree calculation elements according to direct relationship, indirect relationship, and similarity relationship.

[0114] S33. Calculate the association risk value between the logged-in user and each science video in the initial video recommendation list in sequence.

[0115] The association risk value is the maximum correlation between the top m comments of each science video in the initial video recommendation list and the logged-in user, where m is a preset comment depth (e.g., m=3). Calculating only the maximum correlation of the top m comments is essentially a precise balance between privacy risk control and computational efficiency, as users typically pay the most attention to the first few comments.

[0116] S34. Obtain the overlap between each science video in the initial video recommendation list and the logged-in user.

[0117] The overlap refers to the degree of matching between the logged-in user's recommended information and the interest tags of the science videos (which can be calculated using existing technology).

[0118] S35. Sum the associated risk value and overlap with weights, and use the result as the push value for the science video; sort the science videos in descending order of their push values ​​to generate the final video recommendation list.

[0119] Specifically, the higher the association risk value of a science video, the lower its push value (inversely proportional); the higher the overlap of science videos, the higher their push value (directly proportional). Optionally, during the calculation of the push value, if the push value of a certain science video exceeds a preset threshold, the science video is immediately added to the video recommendation list and pushed to accelerate the push efficiency.

[0120] Furthermore, to enable the system to dynamically generate a video recommendation list that meets the criteria of "high overlap and low association risk" under limited computing resources and maintain rapid response, this embodiment further introduces a reinforcement learning adaptive optimization mechanism. This mechanism can replace or supplement the push value calculation process in S3, realizing dynamic optimization of the push value calculation process. Specifically, it includes the following sub-steps Z1 to Z4:

[0121] Z1. Establish a reinforcement learning model and map each logged-in user to an agent i. Formalize the calculation process of the push value of any popular science video in the initial video recommendation list into a triple H=(S,A,R), where S represents the state space, A represents the action space, and R represents the reward.

[0122] The state-space function is: ;

[0123] Where t represents the index of the time step, T represents the total number of time steps, t=1,2,...,T; m t s represents the comment depth at time step t; tb represents the state at time step t; t This represents the vector of elements selected for anchoring correlation calculation. K is the total number of elements used for correlation calculation, and k represents the index of the element used for anchoring correlation calculation, k=1,2,...,K. We have:

[0124] ;

[0125] For example, This indicates that at time step t, the first and third relevance calculation elements are selected as anchor relevance calculation elements, with a comment depth of 3.

[0126] The action space function is: ;

[0127] in, This represents the action of agent i at time step t; Choose the amount of change in the vector for the comment depth at time step t; Calculate the element operation vector for the anchoring correlation degree at time step t. ,have:

[0128] ;

[0129] For example, in performing an action Previously (i.e., time step t-1), m t-1 =3, In Then, when performing the action m t =4 indicates that the comment depth has increased by 1.

[0130] For example, suppose This means keeping the first element, adding the second element, deleting the third element, and increasing the comment depth by 1.

[0131] The reward function is:

[0132] ;

[0133] in, This represents the reward value of agent i at time step t; Indicates the weighting factor; This represents the risk reward value for anchoring correlation; β represents the sensitivity coefficient, β > 0; The risk cost is positively correlated with the total number of elements K in the correlation calculation and negatively correlated with the comment depth. This represents the efficiency-cost reward value, which is negatively correlated with comment depth; the greater the comment depth, the greater the computational load. The smaller the value; e represents the natural logarithm.

[0134] The more anchoring relevance calculation element N selected, the more complex the user overlap becomes when pushing science videos to logged-in users, and the more likely it is to lead to an increase in the similarity of user groups under the science videos. Conversely, the fewer anchoring relevance calculation element N selected, the lower the risk cost. Reduce the risk reward value of anchored correlation. Increase;

[0135] Risk cost: ;

[0136] in, This represents the number of vectors selected for anchor correlation calculation at time step t, such as the number of anchor correlation calculation elements selected at time step t; m t This represents the comment depth at time step t.

[0137] It should be noted that the reward value It describes the action taken by agent i at time step t. The rewards that can be obtained are divided into two parts: the anchored correlation risk reward value. And the negative value of efficiency cost reward ;

[0138] when Unchanged, m t When increasing, reduce, Increase, Reduce, reward value Subject to weighting factors The impact;

[0139] when Increase, m t When unchanged, Increase, reduce, Reward value remains unchanged. reduce.

[0140] Therefore, it can be concluded that if you want a reward value Increase, need to decrease At the same time, increase m t And increase m t It may lead to Reduce, therefore the algorithm will reduce the weight factor Under the control of [the system], there is a tendency to choose a reasonable [m]. t and .

[0141] Z2. Encode all possible candidate actions that agent i can choose at the current time step t into a candidate action dictionary.

[0142] For each agent i, generate all possible candidate actions for the current time step t. Since the action only involves the operation vector of anchoring correlation degree calculation elements. The change in the comment depth selection vector Therefore, the total number of possible candidate actions is finite and enumerable.

[0143] Encode each possible candidate action as an entry in the candidate action dictionary, for example:

[0144] Suppose the initial action of agent i is The initial state is Then, after the initial action is performed... After that, the state becomes .

[0145] It's important to note that each element in the state space has a maximum and a minimum value. When a value exceeds this range, a loop or boundary checks are applied. For example, in 1+1=2, 1 represents adding an element, and 2 has no real meaning, so we can simply set 2 to 0, keeping it unchanged. Similarly, -1-1=-2, where -2 has no real meaning, so we simply set -2 to -1. The minimum value for the number of comments scanned is set to 0 to avoid negative numbers that could affect program operation.

[0146] During the time step iteration, each element in the candidate action dictionary does not change, but each time a candidate action is selected for updating, the reward value obtained by agent i for each candidate action is recorded, and the cumulative reward value is calculated.

[0147] The movement (or diffusion) of each agent i in the state space S is as follows:

[0148] Initialize the state space to obtain s1, and in each time step t, cyclically decide the action to be chosen by agent i. Then update the state of the next time step. s t+1 For the updated state, s t This is the state before the update. The action selected for the agent;

[0149] Calculate each selected action Reward value at time:

[0150] ;

[0151] Calculate the cumulative reward value for all current time steps T' at each time step. :

[0152] ;

[0153] Among them, P t-1 This represents the discount factor, used to describe the impact of historical time step reward values ​​on cumulative reward values; T' represents all current time steps. It should be noted that T' refers to the total number of current time steps from the start time step to the current time step; it is not a fixed value but rather a value that increases continuously as time steps are executed.

[0154] Z3. Use a greedy algorithm to select the best action from the candidate action dictionary and update the state of the next time step.

[0155] Furthermore, the Z3 specifically includes the following steps:

[0156] Z31. Obtain the action with the highest cumulative reward value among all past actions of agent i, and use it as the selection position in the candidate action dictionary.

[0157] It should be noted that the candidate action dictionary will not change after it is constructed, and the candidate action dictionary can represent the operation methods that agent i can choose. The operation methods specifically include the direction of movement and the distance of movement, and the operation methods are independent of the location of agent i.

[0158] Suppose that the candidate action dictionary includes the following four elements:

[0159] Action 1: Move 1m in the positive Y-axis direction;

[0160] Action 2: Move 1m in the negative Y-axis direction;

[0161] Action 3: Move 1m in the positive X-axis direction;

[0162] Action 4: Move 1m in the negative X-axis direction.

[0163] At the first time step, agent i performed action 1 at position (0, 0). Therefore, agent i's initial position at the second time step is (0, 1), and a reward value is calculated at this time. .

[0164] Therefore, the current reward value of action 1 is 1, and the cumulative reward value of action 1 is 1;

[0165] The current reward value for action 2 is 0, and the cumulative reward value for action 2 is 0.

[0166] The current reward value for action 3 is 0, and the cumulative reward value for action 3 is 0.

[0167] The current reward value for action 4 is 0, and the cumulative reward value for action 4 is 0.

[0168] It should be noted that the reward value for agent i for actions that are not performed is 0.

[0169] Then, agent i needs to select a new action at the second time step. At this point, action 1 has the highest cumulative reward value. Therefore, action 1 is selected as the choice position.

[0170] It should be noted that the elements in the candidate action dictionary will not change, so the system can accumulate the reward values ​​of the same action in all time steps after each time step is completed to obtain the cumulative reward value.

[0171] For example, in 10 time steps, action 1 was selected to be executed 3 times, action 2 was selected to be executed 4 times, action 3 was selected to be executed 2 times, and action 4 was selected to be executed 1 time:

[0172] If agent i receives reward values ​​of 1, 2, and 3 for performing action 1 3 times, then the cumulative reward value for action 1 is 6.

[0173] If agent i receives reward values ​​of 1, 1, 1, and 1 respectively for performing action 2 4 times, then the cumulative reward value for action 2 is 4.

[0174] If agent i receives a reward of 1 and 1 respectively for performing action 3 twice, then the cumulative reward value for action 3 is 2.

[0175] If agent i receives a reward value of 1 for performing action 4 once, then the cumulative reward value for action 4 is 1.

[0176] Therefore, when selecting the action for the 11th time step, the action 1 (i.e., the action with the highest cumulative reward value) should be selected.

[0177] Z32, calculate the average viewing time T of the three most recent science popularization videos for the logged-in user corresponding to agent i. max The average viewing time T max The linear mapping function is used to transform the algorithm into radius RGD, and the selection position of the candidate action dictionary is determined by Z31. Using RGD as the radius and center, a selection space is generated from the candidate action dictionary.

[0178] It should be noted that when executing the greedy algorithm, there is a traversal range. The larger the traversal range, the more candidate actions the greedy algorithm needs to calculate. Reward value and cumulative reward value The more actions there are, the longer the computation time of the greedy algorithm can be reduced by changing the candidate actions. The possible values ​​can be controlled. For example, candidate actions within time step t. With only 10 possibilities, compared to 100 possible candidate actions... The greedy algorithm's computation time is reduced. Therefore, the purpose of step Z32 is to control the traversal range of the greedy algorithm and reduce the computation time.

[0179] Assuming a plane, candidate actions Moving forward 10 meters represents the first dimension; moving 5 meters to the right represents the second dimension. Therefore, agent i will execute the candidate action from its current position. Afterwards, a new position will be obtained in the plane.

[0180] Therefore, candidate actions can be... Enumerate all possible values, if the candidate action It is two-dimensional, meaning that the possible solutions can form a plane; if the candidate action The fact that it is three-dimensional indicates that this possible solution can form a cube.

[0181] The candidate action dictionary is fixed and does not change in each time step. The position that can be selected is independent of the position of agent i. Therefore, the action with the highest weighted sum of the obtained reward value and the cumulative reward value can be compared in each time step.

[0182] All candidate actions After enumeration, an action space, or candidate action dictionary, is obtained. This action space represents the range that the greedy algorithm needs to traverse. Determining a selection space within the action space reduces the traversal range of the greedy algorithm. This selection space is determined by the selection position in the candidate action dictionary. Let be the center and RGD be the radius.

[0183] For agent i, the actions of acquiring all historical time steps The action with the highest reward value and the highest cumulative reward value is selected as the position in the candidate action dictionary.

[0184] radius RGD and T max Related, using a linear mapping function to transform T max Transform it. For example, T max =10s, then RGD=1. As T max As the value increases, the radius RGD will also increase.

[0185] Thus, T max The change in T will affect the size of the selection space. max The larger the value of T, the larger the choice space, the more actions the greedy algorithm needs to consider, the more accurate the algorithm, and the longer the computation time; conversely, the smaller the value of T, the larger the choice space, the more actions the greedy algorithm needs to consider, the more accurate the algorithm, and the longer the computation time. max The smaller the value, the smaller the selection space, the fewer actions the greedy algorithm needs to consider, resulting in lower algorithm accuracy and reduced computation time.

[0186] Z33. Select the best action from the selection space.

[0187] Z331. Pre-set the exploration probability ε; select the generation path of the action at the current time step based on the exploration probability ε.

[0188] After obtaining the selection space, two schemes are used to select the action of agent i, with each scheme having an exploration probability ε. Specifically, at each time step, the first scheme is selected with an exploration probability ε, and the second scheme is selected with a probability of 1-ε. The first scheme involves using a greedy algorithm to select the optimal action within the selection space; the second scheme involves directly and randomly selecting the optimal action within the selection space.

[0189] Z332. When the exploration probability ε is used as the probability, it is based on the reward value. and cumulative reward value Choose the best action Calculate the candidate actions to be actually executed. The reward value of the new state after combining with the current state and the reward value Add to the candidate action dictionary.

[0190] In agent algorithms, agent i will determine the reward value. and cumulative reward value Choose the best action Calculate the reward value for all actions in the next time step t+1. and cumulative reward value Choose to let the reward value and cumulative reward value Weighted sum of the largest candidate actions As the best action to be actually executed in the next time step , will reward value Add it to the candidate action dictionary for subsequent time step selection.

[0191] Z333. When the probability is 1-ε, randomly select an action from the selection space as the optimal action.

[0192] Z4. Select the actions to be actually executed. Applied to the current state s t And update the state s for the next time step. t+1 .

[0193] In this embodiment, through the aforementioned reinforcement learning adaptive optimization mechanism, the system can dynamically select the optimal combination of anchoring correlation calculation elements and comment depth when allocating limited computing resources to each agent. This significantly improves the system's response efficiency and resource utilization while ensuring push quality (high overlap and low correlation risk).

[0194] S4. Send science videos to the logged-in user's terminal device in the order of the video recommendation list.

[0195] The system sends science videos to the logged-in user's terminal device sequentially through an encrypted channel, according to the order of the video recommendation list.

[0196] The above description is merely a preferred implementation of this embodiment and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

[0197] Example 2:

[0198] This embodiment is based on Embodiment 1, such as Figure 2 As shown in the figure, this embodiment provides a disease knowledge popular science video recommendation system. The system includes a correlation distribution calculation module, a video coarse selection module, a video push value calculation module, and a video recommendation module.

[0199] The correlation distribution calculation module is used to obtain user account information, medical record information, and operation habit information, and use the medical record information and operation habit information as recommendation information; store the recommendation information based on multi-level classification tags and record the storage address tags; calculate the correlation between users based on the account information using a hierarchical weighted fusion mechanism, and generate a correlation distribution map for each user according to the correlation gradient;

[0200] The video coarse selection module is used to respond to the video push request of the logged-in user, and to perform coarse selection based on the matching of the login user's recommendation information and storage address tags to obtain the video coarse selection range;

[0201] The video push value calculation module is used to calculate the push value of each science video within the coarse selection range of the video and the correlation distribution map, sort the science videos in descending order of their push values, and generate the final video recommendation list.

[0202] The video recommendation module is used to send popular science videos to the logged-in user's terminal device in the order of the video recommendation list.

[0203] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for recommending disease knowledge popular science videos, characterized in that, The method includes: S1. Obtain user account information, medical record information, and operation habit information; use the medical record information and operation habit information as recommendation information; store the recommendation information based on multi-level classification tags and record the storage address tags; calculate the correlation between users based on the account information using a hierarchical weighted fusion mechanism, and generate a correlation distribution map for each user according to the correlation gradient; The steps of calculating the correlation between users based on the account information using a hierarchical weighted fusion mechanism, and generating a correlation distribution map for each user according to the correlation gradient, include: Define relevance calculation elements, which include direct relationships, indirect relationships, and similarity relationships; direct relationships include IP address relationships, family relationships, work unit relationships, and friend relationships; indirect relationships include secondary direct relationships and comment interaction relationships; similarity relationships include user active geographical location relationships, age similarity relationships, and gender similarity relationships. The weighted results of each relationship are then combined to obtain the degree of association between the two users. Several correlation levels are generated with the correlation gradient as the center. The correlation calculation results between users and other users are entered into the corresponding levels to generate a correlation distribution map, and the correlation distribution map is dynamically updated. S2. Responding to the video push request from the logged-in user, a coarse selection is performed based on the matching of the logged-in user's recommendation information and storage address tags to obtain a coarse selection range of videos; S3. Based on the coarse selection range of the videos and the correlation distribution map, calculate the push value of each science video within the coarse selection range of the videos, sort them in descending order according to the push value of the science videos, and generate the final video recommendation list. The calculation of the push value for each science video within the coarse selection range includes: S31. Extract the commenters of the first Q science videos from the coarse selection range corresponding to the logged-in user; S32. Select N anchoring correlation calculation elements from K types of correlation calculation elements; with the N anchoring correlation calculation elements converging and the remaining non-anchoring correlation calculation elements differing as the guide, calculate the matching degree between the logged-in user and all comment users under each science popularization video; rank the science popularization videos based on the matching degree to obtain the initial list of video recommendations; S33. Calculate the association risk value between the logged-in user and each science video in the initial video recommendation list in sequence; S34. Obtain the overlap between each science video in the initial video recommendation list and the logged-in user; S35. The association risk value and overlap are weighted and summed, and the result is used as the push value of the science video. The videos are sorted in descending order of their push values ​​to generate the final video recommendation list. The higher the highest association between the logged-in user and the commenters of a science video, the lower the push value of the science video. The higher the overlap between the logged-in user's recommendation information and the interest tags of the science video, the higher the push value of the science video. S4. Send science videos to the logged-in user's terminal device in the order of the video recommendation list.

2. The method for recommending disease knowledge popular science videos as described in claim 1, characterized in that, The multi-level classification labels include regional level, age level, gender level, disease type level, and treatment method level.

3. The method for recommending disease knowledge popular science videos according to claim 1, characterized in that, The method introduces a reinforcement learning adaptive optimization mechanism to replace or supplement the push value calculation process in S3, specifically including: Z1. Establish a reinforcement learning model and map each logged-in user to an agent i. The calculation process of the push value of any popular science video in the initial video recommendation list is formalized as a triple H=(S,A,R), where S represents the state space, A represents the action space, and R represents the reward. Specifically, the triple H=(S,A,R) is: State-space function: ; Where t represents the index of the time step, and T represents the total number of time steps; m t s represents the comment depth at time step t; t b represents the state at time step t; t This represents the vector of elements selected for anchoring correlation calculation. Action space function: ; in, This represents the action of agent i at time step t; Choose the amount of change in the vector for the comment depth at time step t; Calculate the element operation vector for the anchoring correlation degree at time step t; Reward function: ； in, This represents the reward value of agent i at time step t; Indicates the weighting factor; This represents the risk reward value for anchoring correlation. Indicates the sensitivity coefficient; Indicates risk cost; This represents the efficiency cost reward value; Z2. Encode all possible candidate actions that agent i can choose at the current time step t into a candidate action dictionary; Z3. Use a greedy algorithm to select the best action from the candidate action dictionary and update the state of the next time step; Z4. Select the actions to be actually executed. Applied to the current state s t And update the state s for the next time step. t+1 .

4. The method for recommending disease knowledge popular science videos according to claim 3, characterized in that, The Z3 specifically includes: Z31. Obtain the action with the highest cumulative reward value among all past actions of agent i, and use it as the selection position in the candidate action dictionary; Z32, calculate the average viewing time T of the three most recent science popularization videos for the logged-in user corresponding to agent i. max The average viewing time T max The linear mapping function is used to transform the algorithm into radius RGD, and the selection position of the candidate action dictionary is determined by Z31. Using RGD as the radius and center, a selection space is generated from the candidate action dictionary; Z33. Select the best action from the selection space.

5. The method for recommending disease knowledge popular science videos according to claim 4, characterized in that, The Z33 specifically includes: Z331. Pre-set the exploration probability ε; select the generation path of the action at the current time step based on the exploration probability ε; Z332. When the exploration probability ε is used as the probability, it is based on the reward value. and cumulative reward value Choose the best action Calculate the candidate actions to be actually executed. The reward value of the new state after combining with the current state and the reward value Add to the candidate action dictionary; Z333. When the probability is 1-ε, randomly select an action from the selection space as the optimal action.

6. A disease knowledge popular science video recommendation system, used to implement the disease knowledge popular science video recommendation method according to any one of claims 1-5, characterized in that, The system includes a correlation distribution calculation module, a video coarse selection module, a video push value calculation module, and a video recommendation module; The correlation distribution calculation module is used to obtain the user's account information, medical record information, and operation habit information, and to use the medical record information and the operation habit information as recommendation information. The recommendation information is stored based on multi-level classification tags, and the storage address tags are recorded. Based on the account information, a hierarchical weighted fusion mechanism is used to calculate the correlation between users, and a correlation distribution map for each user is generated according to the correlation gradient. The video coarse selection module is used to respond to the video push request of the logged-in user, and to perform coarse selection based on the matching of the login user's recommendation information and storage address tags to obtain the video coarse selection range; The video push value calculation module is used to calculate the push value of each science video within the coarse selection range of the video and the correlation distribution map, sort the science videos in descending order of their push values, and generate the final video recommendation list. The video recommendation module is used to send popular science videos to the logged-in user's terminal device in the order of the video recommendation list.

Images