Short video advertisement putting optimization evaluation method based on deep learning

Abstract

The invention discloses a short video advertisement putting optimization evaluation method based on deep learning, and relates to the technical field of advertisement putting, and the method collects user behavior data in real time, constructs a dynamic user portrait, enables an advertisement system to adjust the advertisement content and putting strategy in time according to the interest change of a user, improves the correlation of advertisements, and improves the advertisement putting efficiency. Meanwhile, user behavior mode data and advertisement content labels are subjected to multi-dimensional matching, and the matching degree of the advertisements and user interests is calculated by combining cosine similarity with a weighting algorithm, so that advertisement recommendation is more personalized and accurate, the advertisement click rate and conversion rate are greatly improved, and the advertisement recommendation efficiency is improved. Meanwhile, the optimal opportunity of advertisement display is defined by analyzing the exit tendency of the user and the advertisement clicking tendency, it is guaranteed that the advertisement can be displayed at the moment when the user is most likely to interact and click, and the problem of low conversion rate caused by poor opportunity in a traditional method is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of advertisement delivery, and in particular to a short video advertisement delivery optimization evaluation method based on deep learning. Background Art

[0002] With the rapid development of short video platforms, short video advertising has become more and more common. This type of advertising is favored by advertisers because of its concise, diverse forms and strong interactivity. The characteristics of the advertisement are short time, flexible form and strong interactivity, which can quickly attract user attention and promote marketing.

[0003] In the current short video advertising optimization evaluation method, dynamic user portrait and dynamic advertising insertion strategy are

[0004] Dynamic user portraits are designed to build real-time updated user models through multi-dimensional data (such as user behavior, interests, emotions, etc.) to achieve accurate advertising push. However, traditional user portraits mainly rely on users' basic attributes and historical behaviors, and cannot timely reflect the dynamic changes of users' interests. They lack real-time dynamics. Users' focus will change constantly, and traditional portraits fail to be updated in real time, resulting in advertising push unable to capture users' latest interests, reducing the accuracy of advertising. In addition, short video ads contain complex interactive data, such as barrage and likes. Traditional portrait models cannot effectively process such data, resulting in incomplete portrayal of user portraits.

[0005] Some solutions build user portraits based on users' basic attributes (such as age, gender) and historical behaviors (such as viewing history and click behavior), which can predict user preferences to a certain extent, but the accuracy of the portrait is still limited. Or they adopt an advertising insertion strategy to insert advertisements at fixed time nodes or video lengths, but do not consider changes in the plot of the user's video and the current user's emotional fluctuations. For example, inserting advertisements at the climax of the plot is very likely to cause disgust. In order to reduce users' advertising fatigue, some platforms use exposure frequency control to avoid frequently pushing the same advertisements in a short period of time, but they only avoid pushing the same ones. Therefore, there is an urgent need for a short video advertising optimization evaluation method based on deep learning to solve such problems. Summary of the invention

[0006] In view of the above existing problems, the present invention is proposed.

[0007] The present invention provides a short video advertising delivery optimization evaluation method based on deep learning to solve the problems that user portrait accuracy is not enough, the advertising insertion strategy does not take video plot changes into consideration, and only controls the frequency and ignores the type.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0009] The present invention provides a short video advertisement delivery optimization evaluation method based on deep learning, which comprises:

[0010] Step S1, constructing a user dynamic portrait,

[0011] Monitor user behavior on short video platforms in real time, identify user behavior patterns and points of interest, and build dynamic user portraits. The monitoring content includes: viewing time, interaction frequency, video selection and skipping behavior, tendency to click on ads, and viewing content preferences.

[0012] Step S2: Matching advertising creativity with delivery strategy.

[0013] Based on the dynamic user portrait constructed in step S1, the advertising creative is matched with the user's interests, and an advertising creative suitability score is generated for each user;

[0014] Step S3: Dynamically adjust the timing and frequency of advertising.

[0015] During the advertisement delivery process, the advertisement display timing and frequency are dynamically adjusted in combination with the viewing behavior pattern data of the user in step S1 and the advertisement response rate, thereby forming an adjustment rule for dynamically adjusting the advertisement display timing and frequency.

[0016] Furthermore, in step S1, the user behavior pattern and the point of interest are identified as follows:

[0017] Analyze when users are more likely to exit the video and when they are more willing to click on ads, identify user behavior patterns,

[0018] Capture users’ interests based on their viewing time, interaction frequency, and video type preferences.

[0019] Combine the user's social network data, historical consumption records and device data to improve the user portrait and build a dynamic user portrait, namely the user's interests, content preferences and behavior patterns.

[0020] Furthermore, in step S1, the step of constructing a user dynamic portrait includes:

[0021] Real-time collection of user behavior data, including: viewing time T w , Interaction frequency I f , video selection and skipping behavior S v , ad click tendency C a and viewing content preference P v ,

[0022] Perform feature extraction on the collected data: Among them, W t is the weight of the user's viewing time, which is used to measure the relative value of the user's viewing time and the average viewing time.w is the actual viewing time of a video by the user, T avg is the average viewing time of all users on the platform, W t It is used to measure whether users have a longer viewing time for a specific type of video, that is, preference.

[0023] Interaction frequency weight: Among them, W i is the weight of user interaction frequency, I f The number of user interactions in the short video, including likes, comments, and shares. avg The average number of interactions among all users on the platform, used to determine the user's participation in a certain type of content.

[0024] Ad click tendency: Among them, P c is the proportion of advertisement click tendency, C a is the number of times users click on the ad, C total The total number of times a user views an ad, used to determine the user's preference for a certain type of advertising content.

[0025] Furthermore, in step S1, the step of constructing a user dynamic portrait also includes:

[0026] Combine the comprehensive data of user viewing behavior and interactive behavior to calculate the user's behavior pattern.

[0027] Comprehensive user behavior weight: W u =αW t +βW i +γP c , where W u is the user's comprehensive behavior weight, which comprehensively measures the user's interest and interactive behavior. α, β, and γ are adjustment coefficients corresponding to viewing time, interaction frequency, and ad click tendency;

[0028] During the dynamic update of user portraits, new behavior data is collected in real time and dynamically adjusted based on past weights and current behavior data:

[0029] in, is the updated user portrait, that is, the user's current interest point. is the previous user portrait, representing the user’s past interests. λ is the update rate, which controls the update speed of the user portrait. The range is 0≤λ≤1, so as to achieve real-time monitoring of user behavior, capture user behavior patterns and interests, and build and dynamically update user portraits.

[0030] Furthermore, the matching method of step S2 is:

[0031] Label the advertising creatives according to the type of advertising content (such as entertainment, education, shopping, etc.).

[0032] By utilizing the interest points in the dynamic user portrait, a compatibility score between each user and the advertising creative is generated, and based on the collaborative filtering recommendation algorithm, advertising creatives are recommended according to the user's preferences.

[0033] Furthermore, in step S2, the advertising creativity and delivery strategy are matched in the following manner:

[0034] Extract and label the features of the advertising creativity and generate the feature vector of the advertising creativity. Assume that each advertising creativity A i With a specific set of content tags T a , each label represents a certain content attribute of the advertising creative, including entertainment, education and shopping, using the feature vector Indicates advertising creativity A i Tag features:

[0035] in, For advertising i The feature vector of T contains n labels. a1 ,T a2 ,…,T an For advertising creativity A i The content tags are generated based on the content type and target audience of the advertisement.

[0036] Based on the dynamic user portrait in step S1, construct each user U j Interest vector The user's interest tags are also used to represent the user's interest tags. The interest tags are generated based on the analysis of the user's viewing history, interaction frequency, and preferred content. The user's interest vector It is expressed as:

[0037] in, For user U j The interest feature vector contains n interest tags, T u1 ,T u2 ,…,T un For user U j Interest tags indicate the user's interest level in certain categories.

[0038] Furthermore, in step S2, the advertising creativity and delivery strategy matching method further includes:

[0039] Use cosine similarity to calculate ad creative A i With user U j The matching degree:

[0040] Among them, S ij For advertising i With user U j Interest matching score, is the inner product of the advertisement feature vector and the user interest vector, For advertising i The modulus of the eigenvector, i.e. the size of the vector, For user U j The modulus of the interest vector,

[0041] Set weights for different tags to generate weighted matching degrees. The weighted cosine similarity formula is:

[0042] in, Score the matching degree between advertisement and user interests, w k is the weight of the advertising tag k on the overall matching degree, reflecting the impact of the tag on the matching degree of user interests, T ak and T uk Advertisement A i and user U j The value on the kth label, weight w k Adjust according to advertiser's goals;

[0043] Combine advertising delivery strategies to provide each user j Generate Ad A i The comprehensive suitability score of the advertisement is integrated with the interest matching degree of the user and the advertiser's goal. The final advertisement suitability score is expressed as:

[0044] Among them, F ij For advertising i For User U j The comprehensive fitness score of the advertiser is α, and β are the weight coefficients of the matching score and the advertiser's target score. Score the weighted ad's match with the user's interests. For advertisers to advertise A i The placement priority is determined to generate a suitability score for each user for the advertisement. Advertisers adjust α and β to control the impact of both on the score.

[0045] Furthermore, in step S3, the timing and frequency of advertisement display are adjusted as follows:

[0046] The timing of delivery is adjusted based on user viewing behavior, that is, when they are likely to exit and when they click on the ads, and the timing of user acceptance of the ads is analyzed.

[0047] Frequency adjustment is based on user response data to advertisements, using A / B testing and multi-arm gambling algorithms to dynamically adjust the frequency of ad display.

[0048] Furthermore, in step S3, the method of dynamically adjusting the timing and frequency of advertisement delivery is as follows:

[0049] The user's viewing behavior pattern data is obtained from step S1, including when the user tends to exit the video and when the user is more likely to click on the advertisement, and the advertisement response rate data includes the click-through rate and conversion rate generated after the user sees the advertisement.

[0050] The data involved in user behavior patterns include:

[0051] Watch time T w , which is the user's patience with the content. The larger the value, the higher the interest.

[0052] Click on the C a , that is, the user's tendency to click on the advertisement,

[0053] Exit probability P q , which is the probability that a user will exit at a specific point in time,

[0054] Ad response rate data:

[0055] Ad click rate R c , that is, the probability of being clicked after the ad is displayed,

[0056] Conversion rate R t , which is the percentage of clicks on an ad that are converted into purchases.

[0057] Combine the user's exit probability and ad click tendency to define the point of ad display timing:

[0058] Among them, S t is the display fitness of the advertisement at time t, P q (t) is the probability that the user exits the video at time t, C a (t) is the user's tendency to click on the advertisement at time t, T w (t) is the cumulative viewing time of the user at time t, here T w (t) as the denominator to avoid ads being displayed too long.

[0059] The multi-arm gambling algorithm is used to balance the display frequency of advertisements, avoiding over-display. The reward function is defined after each display of advertisements. The size of the reward depends on the click-through rate R of the advertisement. c and conversion rate R t :

[0060] r(t)=R c (t)·Rt (t), where r(t) is the reward value for displaying an ad at time t, R c (t) is the advertisement click rate at time t, R t (t) is the conversion rate at time t,

[0061] The multi-arm gambling algorithm decides whether to continue displaying ads based on historical rewards. Specifically:

[0062] Among them, P d (t) is the probability of continuing to display the ad at time t, r(i) is the reward value of the past N ad displays, and N is the number of ad displays in the past. When the past ad display effect is good, that is, r(i) is high, the ad display frequency increases accordingly. When the ad effect deteriorates, the display frequency decreases.

[0063] Furthermore, in step S3, the method of dynamically adjusting the timing and frequency of advertisement delivery also includes:

[0064] Combining the timing and frequency of ad display, a comprehensive ad display adjustment rule is formed. The final ad display timing and frequency optimization formula is:

[0065] F d (t) = αS t +βP d (t), where F d (t) is the comprehensive adjustment score of the advertisement display at time t, α and β are the weight coefficients of adjustment timing and frequency respectively, S t is the suitability of the advertisement display timing, indicating the suitability of displaying the advertisement at time t, P d (t) is the probability of ad display frequency, which indicates the probability of continuing to display ads at time t. d The larger the (t) value is, the more appropriate it is to display the ad at time t. Advertisers adjust their display strategies based on the settings of α and β, and dynamically adjust the display timing and frequency of ads through comprehensive analysis of user viewing behavior pattern data and ad response rate. The display timing is determined by the user's exit tendency and ad click tendency, while the display frequency is dynamically adjusted using a multi-armed gambling algorithm to make ad display more accurate and the frequency more reasonable, maximizing advertising effectiveness while reducing interference to users.

[0066] The beneficial effects of the present invention are:

[0067] The present invention collects user behavior data in real time and constructs a dynamic user portrait, so that the advertising system can timely adjust the advertising content and delivery strategy according to the changes in user interests, thereby improving the relevance of advertising and reducing the interference of advertising on users.

[0068] The present invention performs multi-dimensional matching of user behavior pattern data and advertising content labels, and uses cosine similarity combined with a weighted algorithm to calculate the matching degree between advertisements and user interests, thereby making advertisement recommendations more personalized and accurate, greatly improving advertisement click-through rates and conversion rates, and solving the problem of mismatch between advertisement content and user interests in traditional methods.

[0069] The present invention defines the optimal time for advertising display by analyzing the user's exit tendency and advertisement clicking tendency, thereby ensuring that the advertisement can be displayed at the moment when the user is most likely to interact and click, avoiding the low conversion rate problem caused by poor timing in traditional methods, and combining the analysis of user viewing behavior and advertisement response rate, the advertisement display timing can be dynamically adjusted to maximize the advertising effect.

[0070] The present invention introduces a multi-arm gambling algorithm to dynamically adjust the display frequency of advertisements based on the historical display effects of advertisements. Through a reward mechanism, the display frequency of advertisements will be automatically optimized according to the effects to avoid over-display, effectively solving the problem of fixed frequency of advertisements in traditional evaluations and reducing the negative impact on user experience.

[0071] The present invention introduces the target weight of the advertiser into the advertisement suitability score, combines the placement priority of the advertiser, and adopts a double matching mechanism to better meet the needs of the advertiser, while ensuring the user experience and improving the return on investment.

[0072] The present invention integrates the adjustment rules of advertisement display timing and frequency to form a unified advertisement delivery scoring mechanism, so that advertisers can flexibly adjust delivery strategies according to user behavior and market feedback. It not only solves the problem of lack of overall optimization in traditional advertisement evaluation, but also enables the advertisement system to perform adaptive adjustment in a complex delivery environment, further improving the effectiveness of advertisements and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0074] Figure 1 This is a flow chart of the short video advertising delivery optimization evaluation method based on deep learning of the present invention. DETAILED DESCRIPTION

[0075] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0076] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0077] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0078] Example 1, reference Figure 1 This embodiment provides a short video advertisement delivery optimization evaluation method based on deep learning, comprising the following steps:

[0079] Step S1, constructing a user dynamic portrait,

[0080] Monitor user behavior on short video platforms in real time, identify user behavior patterns and points of interest, and build dynamic user portraits. The monitoring content includes: viewing time, interaction frequency, video selection and skipping behavior, tendency to click on ads, and viewing content preferences.

[0081] In step S1, the user behavior pattern and points of interest are identified as follows:

[0082] Analyze when users are more likely to exit the video and when they are more willing to click on ads, identify user behavior patterns,

[0083] Capture users’ interests based on their viewing time, interaction frequency, and video type preferences.

[0084] Combine the user's social network data, historical consumption records and device data to improve the user portrait and build a dynamic user portrait, namely the user's interests, content preferences and behavior patterns.

[0085] In step S1, the steps of constructing a user dynamic portrait include:

[0086] Real-time collection of user behavior data, including: viewing time T w , Interaction frequency I f , video selection and skipping behavior S v , ad click tendency C a and viewing content preference P v ,

[0087] Perform feature extraction on the collected data: Among them, W tis the weight of the user's viewing time, which is used to measure the relative value of the user's viewing time and the average viewing time. w is the actual viewing time of a video by the user, T avg is the average viewing time of all users on the platform, W t It is used to measure whether users have a longer viewing time for a specific type of video, that is, preference.

[0088] Interaction frequency weight: Among them, W i is the weight of user interaction frequency, I f The number of user interactions in the short video, including likes, comments, and shares. avg The average number of interactions among all users on the platform, used to determine the user's participation in a certain type of content.

[0089] Ad click tendency: Among them, P c is the proportion of advertisement click tendency, C a is the number of times users click on the ad, C total The total number of times a user views an ad, used to determine the user's preference for a certain type of ad content.

[0090] Combine the comprehensive data of user viewing behavior and interactive behavior to calculate the user's behavior pattern.

[0091] Comprehensive user behavior weight: W u =αW t +βW i +γP c , where W u is the user's comprehensive behavior weight, which comprehensively measures the user's interest and interactive behavior. α, β, and γ are adjustment coefficients corresponding to viewing time, interaction frequency, and ad click tendency;

[0092] During the dynamic update of user portraits, new behavior data is collected in real time and dynamically adjusted based on past weights and current behavior data:

[0093] in, is the updated user portrait, that is, the user's current interest point. is the previous user portrait, representing the user's past interests, and λ is the update rate, which controls the update speed of the user portrait, ranging from 0≤λ≤1, so as to achieve real-time monitoring of user behavior, capture user behavior patterns and interests, and build and dynamically update user portraits.

[0094] Specifically, based on traditional viewing time and interaction frequency, as well as ad click tendencies, video selection and skipping behavior data, a dynamic user profile is constructed so that the profile can dynamically reflect changes in user interests.

[0095] Step S2: Matching advertising creativity with delivery strategy.

[0096] Based on the dynamic user portrait constructed in step S1, the advertising creative is matched with the user's interests, and an advertising creative suitability score is generated for each user;

[0097] The matching method of step S2 is:

[0098] Label the advertising creatives according to the type of advertising content (such as entertainment, education, shopping, etc.).

[0099] By using the interest points in the dynamic user portrait, we can generate a compatibility score between each user and the advertising creative. Based on the collaborative filtering recommendation algorithm, we can recommend advertising creatives according to the user's preferences.

[0100] In step S2, the advertising creative and delivery strategy are matched in the following manner:

[0101] Extract and label the features of the advertising creativity and generate the feature vector of the advertising creativity. Assume that each advertising creativity A i With a specific set of content tags T a , each label represents a certain content attribute of the advertising creative, including entertainment, education and shopping, using the feature vector Indicates advertising creativity A i Tag features:

[0102] in, For advertising i The feature vector of T contains n labels. a1 ,T a2 ,…,T an For advertising creativity A i The content tags are generated based on the content type and target audience of the advertisement.

[0103] Based on the dynamic user portrait in step S1, construct each user U j Interest vector The user's interest tags are also used to represent the user's interest tags. The interest tags are generated based on the analysis of the user's viewing history, interaction frequency, and preferred content. The user's interest vector It is expressed as:

[0104] in, For user U j The interest feature vector contains n interest tags, T u1 ,T u2 ,…,T un For user U jInterest tags indicate the user's interest in certain categories.

[0105] Use cosine similarity to calculate ad creative A i With user U j The matching degree:

[0106] Among them, S ij For advertising i With user U j Interest matching score, is the inner product of the advertisement feature vector and the user interest vector, For advertising i The modulus of the eigenvector, i.e. the size of the vector, For user U j The modulus of the interest vector,

[0107] Set weights for different tags to generate weighted matching degrees. The weighted cosine similarity formula is:

[0108] in, Score the matching degree between advertisement and user interests, w k is the weight of the advertising tag k on the overall matching degree, reflecting the impact of the tag on the matching degree of user interests, T ak and T uk Advertisement A i and user U j The value on the kth label, weight w k Adjust according to advertiser's goals;

[0109] Combine advertising delivery strategies to provide each user j Generate Ad A i The comprehensive suitability score of the advertisement is integrated with the interest matching degree of the user and the advertiser's goal. The final advertisement suitability score is expressed as:

[0110] Among them, F ij For advertising i For User U j The comprehensive fitness score of the advertiser is α, and β are the weight coefficients of the matching score and the advertiser's target score. Score the weighted ad's match with the user's interests. For advertisers to advertise A i The placement priority is determined to generate a score of each user's suitability for the advertisement. Advertisers adjust α and β to control the impact of both on the score.

[0111] Specifically, in the matching process of advertising creativity and user interests, the characteristics of user interests and advertising creativity are represented by vectorization, and the cosine similarity is combined with weight analysis to effectively improve the matching degree between advertising content and user preferences. The advertiser's target weight is introduced to make the advertising creativity meet user needs and maximize the advertiser's business goals. In addition, the recommendation algorithm based on collaborative filtering enhances the personalized recommendation effect of advertising creativity, so that the advertisements seen by each user are highly relevant to their interests, effectively reducing invalid displays and reasonably increasing advertising click-through rate and conversion rate.

[0112] Step S3: Dynamically adjust the timing and frequency of advertising.

[0113] During the advertisement delivery process, the advertisement display timing and frequency are dynamically adjusted in combination with the viewing behavior pattern data of the user in step S1 and the advertisement response rate, so as to form an adjustment rule for dynamically adjusting the advertisement display timing and frequency;

[0114] Specifically, in terms of adjusting the timing and frequency of advertising display, the best display time is determined based on the user's exit tendency and ad click tendency, and the advertisement is delivered at a time when the user still maintains a high interest. The frequency of advertising display is dynamically adjusted in combination with the multi-arm gambling algorithm. This can automatically adjust the frequency of advertising display to avoid excessive display while ensuring continuous optimization of advertising effects.

[0115] In step S3, the timing and frequency of advertisement display are adjusted as follows:

[0116] The timing of delivery is adjusted based on user viewing behavior, that is, when they are likely to exit and when they click on the ads, and the timing of user acceptance of the ads is analyzed.

[0117] Frequency regulation: Based on user response data to ads, A / B testing and multi-arm gambling algorithms are used to dynamically adjust the frequency of ad display.

[0118] In step S3, the method of dynamically adjusting the timing and frequency of advertisement delivery is as follows:

[0119] The user's viewing behavior pattern data is obtained from step S1, including when the user tends to exit the video and when the user is more likely to click on the advertisement, and the advertisement response rate data includes the click-through rate and conversion rate generated after the user sees the advertisement.

[0120] The data involved in user behavior patterns include:

[0121] Watch time T w , which is the user's patience with the content. The larger the value, the higher the interest.

[0122] Click on the C a , that is, the user's tendency to click on the advertisement,

[0123] Exit probability P q , which is the probability that a user will exit at a specific point in time,

[0124] Ad response rate data:

[0125] Ad click rate R c , that is, the probability of being clicked after the ad is displayed,

[0126] Conversion rate R t , which is the percentage of clicks on an ad that are converted into purchases.

[0127] Combine the user's exit probability and ad click tendency to define the point of ad display timing:

[0128] Among them, S t is the display fitness of the advertisement at time t, P q (t) is the probability that the user exits the video at time t, C a (t) is the user's tendency to click on the advertisement at time t, T w (t) is the cumulative viewing time of the user at time t, here T w (t) as the denominator to avoid ads being displayed too long.

[0129] The multi-arm gambling algorithm is used to balance the display frequency of advertisements, avoiding over-display. The reward function is defined after each display of advertisements. The size of the reward depends on the click-through rate R of the advertisement. c and conversion rate R t :

[0130] r(t)=R c (t)·R t (t), where r(t) is the reward value for displaying an ad at time t, R c (t) is the advertisement click rate at time t, R t (t) is the conversion rate at time t,

[0131] The multi-arm gambling algorithm decides whether to continue displaying ads based on historical rewards. Specifically:

[0132] Among them, P d (t) is the probability of continuing to display the ad at time t, r(i) is the reward value of the past N ad displays, and N is the number of ad displays in the past. When the past ad display effect is good, that is, r(i) is high, the ad display frequency increases accordingly. When the ad effect becomes worse, the display frequency decreases.

[0133] Combining the timing and frequency of ad display, a comprehensive ad display adjustment rule is formed. The final ad display timing and frequency optimization formula is:

[0134] F d (t) = αS t +βP d (t), where F d (t) is the comprehensive adjustment score of the advertisement display at time t, α and β are the weight coefficients of adjustment timing and frequency respectively, S t is the suitability of the advertisement display timing, indicating the suitability of displaying the advertisement at time t, P d (t) is the probability of ad display frequency, which indicates the probability of continuing to display ads at time t. d The larger the (t) value is, the more appropriate it is to display the ad at time t. Advertisers adjust their display strategies based on the settings of α and β. They dynamically adjust the display timing and frequency of ads through comprehensive analysis of user viewing behavior pattern data and ad response rate. The display timing is determined by the user's exit tendency and ad click tendency, while the display frequency is dynamically adjusted using a multi-arm gambling algorithm, making ad display more accurate and frequency more reasonable, maximizing advertising effects while reducing interference to users.

[0135] Specifically, the comprehensive advertising display timing and frequency optimization formula combines timing suitability and frequency adjustment to form a unified scoring system for advertising display. Advertisers can flexibly adjust their delivery strategies based on this score to achieve accurate advertising display. Through the intelligent adjustment mechanism, advertising delivery can not only be more in line with users' interests and behavior patterns in terms of timing, but can also flexibly respond to user feedback and market changes in terms of frequency, achieving a double optimization effect.

[0136] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A short video advertising delivery optimization evaluation method based on deep learning, characterized by: include, Step S1, constructing a user dynamic portrait, monitoring the user's behavior on the short video platform in real time, identifying user behavior patterns and points of interest, and constructing a dynamic user portrait. The monitoring content includes: viewing time, interaction frequency, video selection and skipping behavior, tendency to click on ads, and viewing content preferences. Step S2, matching the advertising creative with the delivery strategy, matching the advertising creative with the user's interests based on the dynamic user portrait constructed in step S1, and generating an advertising creative suitability score for each user; Step S3, dynamically adjust the timing and frequency of advertising. During the advertising delivery process, the timing and frequency of advertising display are dynamically adjusted in combination with the user's viewing behavior pattern data and the advertising response rate in step S1 to form an adjustment rule for dynamically adjusting the timing and frequency of advertising display.

2. According to claim 1, a short video advertising delivery optimization evaluation method based on deep learning is characterized in that: In step S1, the user behavior pattern and points of interest are identified as follows: Analyze when users are more likely to exit the video and when they are more willing to click on ads, identify user behavior patterns, Capture users’ interests based on their viewing time, interaction frequency, and video type preferences. Combine the user's social network data, historical consumption records and device data to improve the user portrait and build a dynamic user portrait, namely the user's interests, content preferences and behavior patterns.

3. According to claim 2, a short video advertising delivery optimization evaluation method based on deep learning is characterized in that: In step S1, the steps of constructing a user dynamic portrait include: Real-time collection of user behavior data, including: viewing time T w , Interaction frequency I f , video selection and skipping behavior S v , ad click tendency C a and viewing content preference P v , Perform feature extraction on the collected data: Among them, W t is the weight of the user's viewing time, which is used to measure the relative value of the user's viewing time and the average viewing time. w is the actual viewing time of a video by the user, T avg is the average viewing time of all users on the platform, Interaction frequency weight: Among them, W i is the weight of user interaction frequency, I f The number of user interactions in the short video, including likes, comments, and shares. avg The average number of interactions among all users on the platform, used to determine the user's participation in a certain type of content. Ad click tendency: Among them, P c is the proportion of advertisement click tendency, C a is the number of times users click on the ad, C total The total number of times a user views an ad, used to determine the user's preference for a certain type of advertising content.

4. According to claim 3, a short video advertising delivery optimization evaluation method based on deep learning is characterized in that: In step S1, the step of constructing a user dynamic portrait also includes: Combine the comprehensive data of user viewing behavior and interactive behavior to calculate the user's behavior pattern. Comprehensive user behavior weight: W u =αW t +βW i +γP c , where W u is the user's comprehensive behavior weight, which comprehensively measures the user's interest and interactive behavior. α, β, and γ are adjustment coefficients, corresponding to viewing time, interaction frequency, and ad click tendency; During the dynamic update of user portraits, new behavior data is collected in real time and dynamically adjusted based on past weights and current behavior data: in, is the updated user portrait, that is, the user's current interest point. is the previous user portrait, representing the user’s past interests, and λ is the update rate, which controls the update speed of the user portrait, ranging from 0≤λ≤1.

5. According to claim 4, a short video advertising delivery optimization evaluation method based on deep learning is characterized in that: The matching method of step S2 is: Label the advertising creatives according to the type of advertising content. By utilizing the interest points in the dynamic user portrait, a compatibility score between each user and the advertising creative is generated, and based on the collaborative filtering recommendation algorithm, advertising creatives are recommended according to user preferences.

6. A short video advertising delivery optimization evaluation method based on deep learning according to claim 5, characterized in that: In step S2, the advertising creative and delivery strategy are matched in the following manner: Extract and label the features of the advertising creativity and generate the feature vector of the advertising creativity. Assume that each advertising creativity A i With a specific set of content tags T a , each label represents a certain content attribute of the advertising creative, including entertainment, education and shopping, using the feature vector Indicates advertising creativity A i Tag features: in, For advertising i The feature vector of T contains n labels. a1 , T a2 , ..., T an For advertising creativity A i The content tags are generated based on the content type and target audience of the advertisement. Based on the dynamic user portrait in step S1, construct each user U j Interest vector The user's interest tags are also used to represent the user's interest tags. The interest tags are generated based on the analysis of the user's viewing history, interaction frequency, and preferred content. The user's interest vector It is expressed as: in, For user U j The interest feature vector contains n interest tags, T u1 , T u2 , ..., T un For user U j Interest tags indicate the user's interest level in certain categories.

7. The short video advertising delivery optimization evaluation method based on deep learning according to claim 6 is characterized in that: In step S2, the advertising creative and delivery strategy matching method also includes: Use cosine similarity to calculate ad creative A i With user U j The matching degree: Among them, Si j Score the interest matching degree between advertisement Ai and user Uj, is the inner product of the advertisement feature vector and the user interest vector, For advertising i The modulus of the eigenvector, i.e. the size of the vector, For user U j The modulus of the interest vector, Set weights for different tags to generate weighted matching degrees. The weighted cosine similarity formula is: in, Score the matching degree between advertisement and user interests, w k is the weight of the advertising tag k on the overall matching degree, reflecting the impact of the tag on the matching degree of user interests, T ak and T uk Advertisement A i and user U j The value on the kth label, weight w k Adjust according to advertiser's goals; Combine advertising delivery strategies to provide each user j Generate Ad A i The comprehensive suitability score of the advertisement is integrated with the interest matching degree of the user and the advertiser's goal. The final advertisement suitability score is expressed as: Among them, F ij For advertising i For User U j The comprehensive fitness score of the advertiser is α, and β are the weight coefficients of the matching score and the advertiser's target score. Score the weighted ad's match with the user's interests. For advertisers to advertise A i The delivery priority of .

8. The short video advertising delivery optimization evaluation method based on deep learning according to claim 7 is characterized in that: In step S3, the timing and frequency of advertisement display are adjusted as follows: The timing of delivery is adjusted based on user viewing behavior, that is, when they are likely to exit and when they click on the ads, and the timing of user acceptance of the ads is analyzed. Frequency adjustment is based on user response data to advertisements, using A / B testing and multi-arm gambling algorithms to dynamically adjust the frequency of ad display.

9. The short video advertising delivery optimization evaluation method based on deep learning according to claim 8 is characterized in that: In step S3, the method of dynamically adjusting the timing and frequency of advertisement delivery is as follows: The user's viewing behavior pattern data is obtained from step S1, including when the user tends to exit the video and when the user is more likely to click on the advertisement, and the advertisement response rate data includes the click-through rate and conversion rate generated after the user sees the advertisement. The data involved in user behavior patterns include: Watch time T w , which is the user's patience with the content. The larger the value, the higher the interest. Click on the C a , that is, the user's tendency to click on the advertisement, Exit probability P q , which is the probability that a user will exit at a specific point in time, Ad response rate data: Ad click rate R c , that is, the probability of being clicked after the ad is displayed, Conversion rate R t , which is the percentage of clicks on an ad that are converted into purchases. Combine the user's exit probability and ad click tendency to define the point of ad display timing: Among them, S t is the display fitness of the advertisement at time t, P q (t) is the probability that the user exits the video at time t, C a (t) is the user's tendency to click on the advertisement at time t, T w (t) is the cumulative viewing time of the user at time t, The multi-arm gambling algorithm is used to balance the display frequency of advertisements and define the reward function after each advertisement display. The size of the reward depends on the click rate R of the advertisement. c and conversion rate R t : r(t)=R c (t)·R t (t), where r(t) is the reward value for displaying an ad at time t, R c (t) is the advertisement click rate at time t, R t (t) is the conversion rate at time t, The multi-arm gambling algorithm decides whether to continue displaying ads based on historical rewards. Specifically: Among them, P d (t) is the probability of continuing to display the ad at time t, r(i) is the reward value of the past N ad displays, and N is the number of ad displays in the past. When the past ad display effect is good, that is, r(i) is high, the ad display frequency increases accordingly. When the ad effect deteriorates, the display frequency decreases.

10. The short video advertising delivery optimization evaluation method based on deep learning according to claim 9, characterized in that: In step S3, the method of dynamically adjusting the timing and frequency of advertisement delivery also includes: Combining the timing and frequency of ad display, a comprehensive ad display adjustment rule is formed. The final ad display timing and frequency optimization formula is: F d (t) = αS t +βP d (t), where F d (t) is the comprehensive adjustment score of the advertisement display at time t, α and β are the weight coefficients of adjustment timing and frequency respectively, S t is the suitability of the advertisement display timing, indicating the suitability of displaying the advertisement at time t, P d (t) is the probability of ad display frequency, which indicates the probability of continuing to display ads at time t. d The larger the (t) value is, the more appropriate it is to display the ad at time t, and advertisers adjust their display strategies based on the settings of α and β.

Images