A personalized content recommendation method and system for a short video platform

Abstract

The application relates to the technical field of big data, in particular to a personalized content recommendation method and system for a short video platform. The method comprises the following steps: screening opposite short videos of each to-be-recommended short video and determining the initial opposite recommendation degree of each opposite short video according to the actual feature difference between each to-be-recommended short video and a user preferred short video and the theoretical feature difference between the opposite short video of each to-be-recommended short video and the user preferred short video; selecting a test short video and the opposite short video thereof, generating a transition graphic text and pushing the transition graphic text to a user, recording the positive feedback of the user on the transition graphic text, and calculating a user preference evolution feedback index; updating the initial opposite recommendation degree to obtain an updated opposite recommendation degree; determining positive short videos and the opposite short videos thereof to be recommended to the user, and recommending each positive short video and the opposite short video thereof to the user in groups. The method of recommending positive short videos and opposite short videos in groups breaks through the information barrier.

Description

Technical Field

[0001] This invention relates to the field of big data technology, specifically to a personalized content recommendation method and system for short video platforms. Background Technology

[0002] In recent years, with the explosive growth of short video platforms, personalized recommendation systems have become the core link connecting users and content. Current mainstream short video content recommendation algorithms, such as deep learning, collaborative filtering, and weighted directed graphs, primarily determine users' potential interests and preferences by mining static statistical features such as historical browsing, likes, and favorites, and then push short videos that match those preferences. This method, constantly catering to user preferences and pushing short videos with singular, one-sided viewpoints, can lead users into localized information barriers, hindering the development of systematic thinking habits. Summary of the Invention

[0003] To address the technical problem of existing short video recommendation methods resulting in one-sided and limited viewpoints, which traps users in a partial information barrier, the present invention aims to provide a personalized content recommendation method and system for short video platforms. The specific technical solution adopted is as follows: In a first aspect, the present invention provides a personalized content recommendation method for a short video platform, comprising the following steps: 1) Obtain the short videos to be recommended and the user's preferred short videos from the user's historical short video browsing history; Based on the actual feature differences between each short video to be recommended and the short videos preferred by users, as well as the theoretical feature differences between each short video to be recommended and the short videos preferred by users, the short videos to be recommended are selected as the opposing short videos, and the initial degree of opposition recommendation for each opposing short video is determined. 2) Select test short videos from the short videos to be recommended; determine the opposing short videos based on the initial opposing recommendation degree of each opposing short video of the test short video; generate transitional text and images based on the test short videos and their opposing short videos and push them to users, record the positive feedback of users to the transitional text and images, and calculate the user preference evolution feedback index. 3) Based on the user preference evolution feedback index, update the theoretical feature differences between the opposing short videos and the user preferred short videos for each short video to be recommended, and update the initial opposing recommendation degree based on the update results to obtain the updated opposing recommendation degree; 4) Based on the actual feature differences between each short video to be recommended and the short videos preferred by the user, determine the positive short videos to be recommended to the user from the short videos to be recommended, and determine the opposing short videos of each positive short video based on the updated degree of each opposing recommendation, and recommend each positive short video and its opposing short video to the user in a group.

[0004] In conjunction with the first aspect mentioned above, among some possible implementation methods, the calculation methods for the actual feature differences between each short video to be recommended and the user's preferred short videos include: Construct the position feature vectors of user-preferred short videos and the position feature vectors of short videos to be recommended; calculate the projection distance of each position feature vector in the direction of the baseline dimension axis vector; The user preference distance is obtained by averaging the corresponding projection distances using the user preference ratings of each user's preferred short videos as weights. Based on the difference between the projection distance of the stance feature vector of each short video to be recommended in the direction of the baseline dimension axis vector and the user preference distance, the actual feature difference between each short video to be recommended and the user preference short video is determined.

[0005] In conjunction with the first aspect mentioned above, among some possible implementation methods, the methods for determining the opposing short videos for each short video to be recommended include: For any short video to be recommended, the short videos that have the smallest difference between the actual feature difference and the theoretical feature difference between the short video to be recommended and the short video preferred by the user are selected as the short videos to be recommended.

[0006] In conjunction with the first aspect mentioned above, among some possible implementation methods, the methods for determining the initial degree of opposition in recommendations for each opposing short video include: For any short video to be recommended, calculate the mean of the actual feature differences between each opposing short video and the user's preferred short video; Based on the differences in actual characteristics between each opposing short video and the user-preferred short video and the mean, the initial opposing recommendation degree of each opposing short video in the opposing short video cluster of the short video to be recommended is obtained.

[0007] In conjunction with the first aspect mentioned above, among some possible implementation methods, positive user feedback on transitional images and text includes: number of reads, dwell time, and number of shares; methods for calculating the user preference evolution feedback index include: The user preference evolution feedback index is obtained by combining ratings based on the number of reads, the duration of stay, and the number of shares.

[0008] In conjunction with the first aspect mentioned above, among some possible implementation methods, the methods for updating the theoretical feature differences between each recommended short video and the user-preferred short video include: Calculate the probability of steady-state maintenance of user preferences based on user-initiated search content in the user's historical short video browsing records; Calculate the damping coefficient based on the user preference evolution feedback index; By combining the damping coefficient and the probability of steady-state maintenance of user preferences, the degree of preference stubbornness can be obtained. By integrating the inverse mapping results of preference stubbornness and the theoretical feature differences between the opposing short videos and the user-preferred short videos for each short video to be recommended, the updated theoretical feature differences between the opposing short videos and the user-preferred short videos for each short video to be recommended are obtained.

[0009] In conjunction with the first aspect mentioned above, among some possible implementations, methods for calculating the steady-state maintenance probability of user preferences include: Based on the user's historical short video browsing records, feature vectors of each user's active search content are constructed, and the actual feature differences between each user's active search content and user preferences are calculated based on the feature vectors of each user's active search content. Calculate the absolute value of the difference between the actual feature difference between the user's active search content and the user's preference in each search and the mean of the actual feature difference between each opposing short video that best matches the user's preference. Then, calculate the weighted mean according to the time decay weight of each search to obtain the probability of steady-state maintenance of user preference.

[0010] In conjunction with the first aspect mentioned above, among some possible implementation methods, the methods for calculating the damping coefficient include: The damping coefficient is obtained by integrating the inverse mapping results of the user preference evolution feedback index and the actual feature differences between the recommended short videos that best match user preferences and the short videos that user preferences prefer.

[0011] In conjunction with the first aspect mentioned above, among some possible implementation methods, the methods for determining the opposing short videos to each positive short video include: For any positive short video, select the opposing short video with the highest updated opposing recommendation level from its opposing short videos, and use it as the opposing short video for that positive short video.

[0012] Secondly, the present invention also provides a personalized content recommendation system for a short video platform, including a client and a server. The client is used to collect the user's historical short video browsing records and send them to the server. The server is used to recommend short videos to the user using the personalized content recommendation method of the short video platform of the present invention.

[0013] The present invention has the following beneficial effects: The personalized content recommendation method and system for short video platforms of the present invention first determines the user's preferred short videos based on the user's historical browsing records, then calculates the actual feature differences between each short video to be recommended and the user's preferred short videos, and determines the opposing short videos for each short video to be recommended. When recommending short videos, positive short videos are first selected from the short videos to be recommended, and positive short videos and opposing short videos are recommended in groups, effectively breaking down information barriers. Furthermore, since the user's acceptance of opposing short videos is uncertain, blindly recommending opposing short videos may lead to user aversion. Therefore, before recommending short videos in groups, the present invention first generates transitional graphics between test short videos and their opposing short videos to test changes in user preferences, indirectly reflecting the user's acceptance of opposing short videos. Then, the opposing recommendation level of opposing short videos is updated, and opposing short videos are recommended based on the updated opposing recommendation level, recommending opposing short videos that users can accept as much as possible, breaking down information barriers without causing user aversion. Attached Figure Description

[0014] Figure 1 A flowchart illustrating a personalized content recommendation method for a short video platform according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating how the opposing recommendation levels of opposing short videos are updated according to an embodiment of the present invention. Detailed Implementation

[0015] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0016] The personalized content recommendation system for a short video platform of the present invention includes a client and a server. The client is used to collect the user's historical short video browsing records and send them to the server. The server is used to determine the short video recommendation ranking using the personalized content recommendation method of the short video platform of the present invention.

[0017] In this implementation, the client's short video recommendation page recommends multiple short videos at a time, and the recommended short videos and accompanying text / images appear in groups. Each group of short videos / text / images includes at least two short videos with opposing viewpoints (a positive short video and a contrasting short video). For example, recommended short video 1 is "The Secret to Earning a Million Yuan Annually at 30," and recommended short video 2 is "What Hinders Income Growth." The positive short video is one that aligns with the user's preferences, while the contrasting short video discusses the same content as the positive short video but presents opposing viewpoints.

[0018] The personalized content recommendation method for short video platforms of the present invention, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the user's historical short video browsing history, calculate the user preference score for each short video, then filter out the user's preferred short videos and group them according to short video tags. Specifically, this includes the following steps: 1.1) Obtain the user's full short video browsing history (i.e., historical short video browsing history) within a preset period, and at the same time obtain the association information of each short video in the browsing history.

[0019] In this embodiment, the preset period is taken as the past 30 days.

[0020] The associated information for each short video includes: user interaction behavior characteristics, multi-dimensional content characteristics, and search timestamps.

[0021] User interaction behavior characteristics include: short video tags, replay rate, likes, favorites, follows, and shares.

[0022] The method for determining multi-dimensional content features includes: using ASR (Automatic Speech Recognition) to separate audio from short videos, perform speech-to-text transcription, and segmentation to extract the full-text script; then using OCR (Optical Character Recognition) to extract frames with sudden scene changes or long dwell times as keyframes; performing region detection and recognition on subtitles, titles, or whiteboard text in keyframes to obtain visually significant information in the short video; aligning the ASR script and OCR subtitles chronologically, removing redundant interjections or repetitive background text; using an NLP (Natural Language Processing) model to extract keywords and entities, and outputting text semantic units to obtain multi-dimensional content features; the text semantic units include: a temporal script, visual emphasis points, a semantic summary, and keywords. The temporal script contains the full-text ASR content and its corresponding start and end times, reflecting the logical axis of the short video content; visual emphasis points include keyframe titles, emphasis phrases, or core content extracted by OCR, typically representing the focus of the short video content; the semantic summary and keywords are 3-5 short summaries generated based on the full-text content and position keywords reflecting the core arguments.

[0023] Search timestamps are used not only to record the time when a user actively searches by entering content in the search box, but also to record the sequence of keywords entered by the user.

[0024] 1.2) Based on the user interaction behavior characteristics of each short video, calculate the user preference score of each short video in the full short video browsing record, and then select the user preference short videos from the full short video browsing record to construct a user preference short video set.

[0025] In this embodiment, user interaction behavior features are divided into explicit behaviors and implicit behaviors. Explicit behaviors include: liking, saving, following, and sharing. For these explicit behaviors, this embodiment uses 0 or 1 for encoding to obtain explicit behavior data. 0 indicates that the user did not perform the action, and 1 indicates that the user performed the action. Implicit behavior includes replay rate (hereinafter also referred to as implicit behavior data). The replay rate increases by 1 every time the user watches a short video completely. In order to avoid multiple replays of a single short video due to network problems of the short video platform or when the user is idle, this embodiment performs amplitude limiting processing. When the replay rate reaches the preset replay rate upper limit, the replay rate will no longer increase. The preset replay rate upper limit in this embodiment is 5.

[0026] Based on explicit and implicit behavioral data for each short video, user preference scores are calculated for each short video across the entire short video viewing history. The calculation method for any given short video's user preference score includes: In the formula, Indicates user preference rating, , , and These indicate whether the user has performed actions such as liking, saving, following, and sharing. For replay rate, , , , and The rating weights for likes, favorites, follows, shares, and replay rate are preset values. Since the user's preference for the short video increases sequentially from likes, favorites, follows, to shares, in this implementation, the rating weights for these four explicit behaviors increase sequentially. , , , Since users watch the entire short video, and even watch it repeatedly, it indicates a high level of preference for that short video. Therefore, in this implementation, the replay rate rating weight is set to the maximum value among all rating weights. Specifically, it is set... .

[0027] After obtaining the user preference ratings for each short video in the full short video browsing history, all short videos are grouped according to the short video tags in the user interaction behavior characteristics. Short videos with the same short video tag are grouped together. After grouping, the short videos in each group are sorted in descending order of user preference ratings. The short videos with the highest user preference ratings in each group are selected as the user preference short video set for each group. In this embodiment, the set percentage is 10%. If the total number of short videos in a group multiplied by the set percentage results in a decimal, it is rounded up.

[0028] Step 2: Construct the stance feature vector of each short video to be recommended and the stance feature vector of each user-preferred short video. Calculate the relative distance difference between each short video to be recommended and the user-preferred short video to determine the opposing short video clusters of each short video to be recommended, and calculate the degree of opposing recommendation of each opposing short video in each opposing short video cluster.

[0029] In this embodiment, to improve the accuracy of calculating the relative distance difference between each short video to be recommended and the user's preferred short videos, each short video to be recommended is also grouped according to its short video tag. Subsequent steps are all performed under the constraint of the same short video tag. Specifically, the following steps are included: 2.1) Extract the arguments of each user preference short video in the user preference short video set of each group, construct the position feature vector of each short video, and then determine the baseline dimension axis vector of each group and the user preference distance in the direction of the baseline dimension axis vector of each group.

[0030] Taking any short video reflecting user preferences as an example, the method for extracting the arguments from that short video and then constructing its stance feature vector includes: Based on the text semantic units of this short video, an argument mining model is used to identify logical conjunctions and sentence-breaking predicates within the text semantic units, and these are used as segmentation points to divide the text semantic units into sets of argument segments containing multiple independent argument fragments. The argument mining model is a pre-trained logic extraction model; in this implementation, a logic extraction model based on the Transformer architecture is used as the argument mining model.

[0031] The individual argument segments of the short video reflecting user preferences are input into a pre-trained semantic model, which maps each argument segment to a high-dimensional space, forming multiple high-dimensional feature vectors. In this embodiment, the pre-trained semantic model is the RoBERTa-large model.

[0032] Based on the visual emphasis points in the text semantic units, the weight of each independent argument segment is determined, and the weighted mean of the high-dimensional feature vectors of all independent argument segments of the short video is calculated according to the weight to obtain the stance feature vector of the user preference short video.

[0033] Methods for determining the weight of each independent argument segment include assigning the highest weight to the independent argument segments that appear in keyframes. Independent argument passages appearing in emphasis or core content are assigned the second highest weight. Set the lowest weight for independent argument fragments that appear in the rest of the regular content. In this embodiment, it is set that... , , .

[0034] Repeat the above steps to obtain the position feature vector of all users' preferred short videos in a group. Then, determine the baseline dimension axis vector of this group using the following method: For the position feature vectors of all user-preferred short videos in this group, principal component analysis (PCA) is used to extract the first principal component, obtaining the baseline dimension vector of this group. This vector represents the direction with the greatest semantic difference and the most concentrated logic within the group. If the magnitude of the baseline dimension vector extracted by PCA is 0 (i.e., the preferred short video features within the group are completely consistent), the arithmetic mean vector of the position feature vectors of all user-preferred short videos in this group is directly calculated and used as the baseline dimension vector for subsequent projection calculations. The high-dimensional feature vectors mapped by the semantic model are usually non-zero; the mean vector, as a substitute axis, effectively preserves the absolute directional baseline in high-dimensional space, thereby distinguishing short videos to be recommended that exhibit differences.

[0035] Calculate the projection distance of the stance feature vector of each user's preferred short video in the group along the reference dimension axis vector direction of the group, and calculate the weighted mean of the projection distance with the user preference rating as the weight to obtain the user preference distance along the reference dimension axis vector direction of the group.

[0036] Repeat the above steps for all groups to obtain the baseline dimension axis vector for each group and the user preference distance along the baseline dimension axis vector direction for each group.

[0037] Methods for calculating the projected distance of the position feature vector along the reference dimension axis vector direction of this group include: In the formula, For the projection distance, This is the baseline dimension axis vector for this group. For vector magnitude operations, This is the stance feature vector of the short video. It can be the stance feature vector of each short video in the group, or it can be the stance feature vector of the short video to be recommended.

[0038] Taking any group as an example, methods for calculating the user preference distance along the reference dimension axis vector include: In the formula, This represents the user preference distance along the baseline dimension axis vector of this group. This indicates the first group. Users prefer short videos. The number of short videos preferred by users in this group. For the first group The projection distance of the user's preference for short videos' position feature vector onto the baseline dimension axis vector. For the first group The normalization result of the user preference ratings for short videos in this embodiment is obtained by using maximum-minimum value normalization. The maximum value is the maximum user preference rating of all short videos in the group, and the minimum value is the minimum user preference rating of all short videos in the group. If the extracted maximum and minimum values ​​are the same (i.e., the range is zero), the normalization result of all processed objects within the group is directly preset to an equal weight (e.g., 1). The purpose of normalization is to use the relative magnitude of the user preference ratings of all short videos in the group for weighting, to avoid the problem that the user preference ratings of the group are all concentrated in large values ​​(e.g., 8-9), with relatively similar absolute magnitudes, and the weighted average result is close to the arithmetic mean.

[0039] 2.2) Construct the stance feature vector of the short video to be recommended, calculate the projection distance of the stance feature vector of each short video to be recommended in the direction of the baseline dimension axis vector of the group with the same short video tag, and then calculate the relative distance difference (i.e., the actual feature difference) between each short video to be recommended and the user's preferred short video.

[0040] Specific methods include: Calculate the stance feature vector for each short video to be recommended, and then calculate the projection distance of the stance feature vector of each short video to be recommended in the direction of the reference dimension axis vector of the group with the same short video tag. The specific calculation method of the projection distance has been explained in step 2.1), and will not be repeated here.

[0041] Calculate the absolute value of the difference between the projection distance of the stance feature vector of the short video to be recommended in the direction of the baseline dimension axis vector of the same short video tag group and the user preference distance in the direction of the baseline dimension axis vector of the group, and then obtain the relative distance difference between each short video to be recommended and the user preferred short video.

[0042] Taking any short video to be recommended as an example, the specific formula is as follows: In the formula, This represents the relative distance between the short videos to be recommended and the short videos preferred by the user. This represents the projection distance of the stance feature vector of the short video to be recommended onto the reference dimension axis vector direction of the grouping of short videos with the same tag. User preference distance along the baseline dimension axis vector for groups of short videos with the same tag. This indicates normalization processing. In this embodiment, maximum and minimum value normalization processing is used. During the processing, the maximum and minimum values ​​are taken from all short videos to be recommended that share the same short video tag. The maximum and minimum values. If the projection distances of the stance feature vectors of all short videos to be recommended with the same short video tag are equal along the reference dimension axis vector direction of the grouping with the same short video tag, then... Since the maximum and minimum values ​​are equal, no normalization is needed. The relative distance difference between each short video to be recommended and the user's preferred short video can be set to a fixed constant of 0 to 1. The specific value can be arbitrarily selected within the range of 0 to 1. As long as the value is kept fixed, there will be no distinction in the relative distance difference between each short video to be recommended under this short video tag.

[0043] 2.3) Based on the relative distance difference between each short video to be recommended and the short videos preferred by the user, determine the opposing short video clusters of each short video to be recommended, and calculate the degree of opposing recommendation of each opposing short video in the cluster.

[0044] The following explanation uses any short video to be recommended from any group as an example. The specific method is as follows: Based on the relative distance difference between the short video to be recommended and the short videos preferred by the user (referred to as relative distance difference). Calculate the theoretical relative distance difference between the opposing short video and the user's preferred short video (i.e., the theoretical feature difference, hereinafter referred to as the theoretical relative distance difference between the opposing short video and the short video to be recommended). Based on this theory, relative distance difference K-nearest neighbor search is performed on the grouped short videos to be recommended to obtain the cluster of opposing short videos of the short videos to be recommended. The short videos in the opposing short video cluster are also short videos to be recommended in essence. However, for the purpose of differentiation, the short videos to be recommended in the opposing short video cluster are referred to as opposing short videos in this embodiment.

[0045] In this embodiment, the specific method for K-nearest neighbor search includes: determining the relative distance difference and the theoretical relative distance difference. The K short videos with the smallest differences are selected to form a cluster of opposing short videos for each video to be recommended. The mean of the relative distance differences among all opposing short videos in this cluster is then determined. K is a set number. If it is a decimal, it is rounded up. Since one positive short video is sufficient to determine one opposing short video, the set number does not need to be too large. In this embodiment, K is set to 10% of the total number of short videos to be recommended in this group.

[0046] For any cluster of opposing short videos to be recommended, the absolute value of the difference between the relative distance difference of each opposing short video (in this embodiment, the short videos in the cluster of opposing short videos to be recommended are referred to as opposing short videos) and the mean of the relative distance differences of all opposing short videos in the cluster (referred to as the mean of the relative distance differences of the opposing short video cluster) is calculated. After performing maximum and minimum value normalization processing, the initial opposing recommendation degree of each short video in the cluster is obtained. In this embodiment, the maximum and minimum values ​​in the maximum and minimum value normalization process are the maximum and minimum values ​​of the absolute values ​​of the difference between the relative distance difference of each opposing short video and the mean of the relative distance differences of the opposing short video cluster.

[0047] The smaller the absolute value of the difference between the relative distance difference of opposing short videos and the mean of the relative distance difference of the opposing short video cluster, the closer the arguments and positions of the opposing short video are to the arguments and positions of the opposing short video cluster, and the greater its initial degree of opposition recommendation. Conversely, the larger the absolute value of the difference between the relative distance difference of opposing short videos and the mean of the relative distance difference of the opposing short video cluster, the more opposed the arguments and positions of the opposing short video are to the arguments and positions of the opposing short video cluster, and the lower its initial degree of opposition recommendation.

[0048] Step 3: Construct a weighted directed graph.

[0049] Users and recommended short videos are treated as nodes in a graph structure. Based on the relative distance difference between each recommended short video and the user's preferred short video, as well as the degree of opposition recommendation of each opposing short video in the opposing short video cluster of each recommended short video, the edge weights between nodes are calculated to construct a weighted directed graph.

[0050] The graph structure of a weighted directed graph is represented as follows: ,in Represents a node. The edge weights are represented by the graph structure, which treats users and recommended short videos as nodes. The edge weights between nodes are calculated based on the relative distance between each recommended short video and the user's preferred short video, as well as the degree of opposition between each opposing short video in the opposing short video cluster of each recommended short video, thus constructing a weighted directed graph.

[0051] In this embodiment, the relative distance difference between each short video to be recommended and the user's preferred short video is reverse-mapped to obtain the similarity between each short video to be recommended and the user's preferred short video, and this similarity is used as the edge weight between the user node and the short video node to be recommended.

[0052] Taking any short video to be recommended as an example, the reverse mapping method in this implementation is as follows: In the formula, The similarity between the short video to be recommended and the short videos preferred by the user.

[0053] In this embodiment, a single user is used as an example. All instances of "user" in this embodiment refer to the same user; therefore, there is only one user node in this embodiment. In practical applications, there can be multiple user nodes in the graph structure. For different users, the above steps are repeated to obtain the edge weights between each user node and each short video node to be recommended.

[0054] The degree of opposition between each opposing short video (which also belongs to the short videos to be recommended) in each cluster of opposing short videos to be recommended is used as the edge weight between the nodes of the short videos to be recommended. This completes the construction of the weighted directed graph.

[0055] The weighted directed graph serves as the data carrier in this embodiment.

[0056] Step 4: Based on the constructed weighted directed graph, determine the test short videos and opposing short videos to recommend to users, generate transitional graphics and text to push to users, and record the positive feedback of users on the transitional graphics and text.

[0057] Since user preferences are not static, this invention, when recommending short videos to users, first determines the positive short videos and their opposing short videos that match the user's historical preferences for each group. Then, it generates transitional text and images based on the test short videos and opposing short videos, pushes the transitional text and images to users, tests whether user preferences are evolving towards a neutral direction, records the user's positive feedback to the text and images, and constructs the user preference evolution feedback index for that group.

[0058] The method for determining the positive short videos and their opposing short videos to recommend to users includes: selecting the short video that best matches the user's preferences from each group of short videos to be recommended under each short video tag as a test short video, and selecting the short video with the highest initial degree of opposing recommendation from the opposing short video cluster of the test short video as the opposing short video. The method for selecting the short video that best matches the user's preferences includes selecting the short video with the smallest relative distance difference among the short videos to be recommended in the group (i.e., the short video with the highest edge weight between the user node and the short video node to be recommended—and the highest similarity to the user's preferred short video) as the short video that best matches the user's preferences.

[0059] Transitional text is used to illustrate the evolution from test short videos to opposing short videos. For example, the test short video is "The Secret to Earning Millions at Age 30," the opposing short video is "What Happens to Income Growth," and the transitional text is "From Poverty to Wealth, What are the Most Critical Influencing Factors?" In this embodiment, the method for generating transitional text includes: extracting the core argument segments with the highest weight from the test short video and the opposing short video respectively, inputting them into a pre-trained large language model (such as a pre-trained GPT series model), and generating transitional text.

[0060] In this implementation, the positive user feedback on the transitional images and text is recorded, including the number of reads, dwell time, and number of shares. The user preference evolution feedback index is obtained by combining these three data points. The formula is as follows: In the formula, The user preference evolution feedback index Rate the number of reads. Rate the duration of stay. The three scores for the number of shares are obtained by normalizing the number of reads, the dwell time, and the number of shares, respectively. In this embodiment, the normalization method for all three scores is the maximum-minimum value normalization method, and the minimum value in the normalization process is 0. The maximum value for normalizing the number of reads is a preset maximum number of reads, such as 1 time. The maximum value for normalizing the dwell time is a preset maximum dwell time, which is the theoretical time required to fully read the transition text and image once. The maximum value for normalizing the number of shares is a preset maximum number of shares, such as 1 time. , and The scores for read count, dwell time, and share count are weighted separately, and their sum equals 1. In this implementation, since these three scores have equal importance and influence, they are set as follows: .

[0061] Since the purpose of normalization in this embodiment is only to eliminate dimensions, the maximum value of normalization is taken as the time required for one complete reading, and it is not necessary to completely map the normalization result to the range of 0-1. The normalization result can exceed 1.

[0062] The User Preference Evolution Feedback Index measures users' acceptance of transitional images and text, indirectly reflecting the degree to which user preferences have evolved towards opposing short videos. The higher the value, the greater the user's acceptance of transitional images and text, and the greater the degree to which user preferences have evolved towards opposing short videos.

[0063] It should be noted that, since users have different levels of preference persistence for different short video tag groups, the above method is required to generate transitional graphics and text for each group and construct the user preference evolution feedback index for that group.

[0064] Step 5: Based on the positive feedback from users regarding the transitional images and text for each group, update the theoretical relative distance difference between the opposing short videos of each group and the short videos to be recommended, and update the degree of opposition recommendation of each opposing short video in the short video cluster of each group and the weighted directed graph.

[0065] The method and process for updating the opposing recommendation levels of various opposing short videos are as follows: Figure 2 As shown, it includes the following steps: 5.1) Based on the user preference evolution feedback index of each group, construct the damping coefficient of each group.

[0066] For any group of short videos to be recommended under any short video tag, the methods for calculating the damping coefficient based on the user preference evolution feedback index include: In the formula, This indicates the first short video in the set of short videos to be recommended under this short video tag. In this embodiment, the short video is used as the one that best matches the user's preferences for the short video tag. The damping coefficient is... The relative distance difference for short videos that best match user preferences. The preset sensitivity coefficient ranges from 0.5 to 1.5. A larger value indicates a faster damping coefficient decay, while a smaller value indicates a slower decay. In this embodiment... The value is 1. It is a natural exponential function.

[0067] The damping coefficient reflects the recommendation resistance of opposing short videos; the larger the value, the greater the recommendation resistance. Since the relative distance difference reflects the difference in arguments between positive and user-preferred short videos, the larger the value, the less likely both positive and opposing short videos are to be accepted by users, resulting in a larger damping coefficient. The larger the value, the greater the evolution of user preferences towards opposing short videos, and the greater the user's acceptance of opposing short videos. Therefore, it is necessary to reduce the recommendation resistance of opposing short videos. This implementation method uses an inverse natural exponential function. right Perform a reverse mapping to the interval (0, 1) to reduce the damping coefficient.

[0068] In addition to positive user feedback on transitional images and text, the user's active search records within a preset period can also reflect the evolution of user preferences. Therefore, this invention calculates the probability of user preference steady-state maintenance through the following steps.

[0069] 5.2) Based on the search timestamps (user search content) of each user's active search for short videos within a preset period, calculate the steady-state maintenance probability of user preferences for each group.

[0070] Based on the search content entered by the user in the search box each time they actively search within a preset period, the feature vector of each search content is calculated, and the short video tag corresponding to each search content is determined. Then, a set of short videos to be recommended is determined for the grouping of the short video tags. The short video with the smallest relative distance difference is selected from the set of short videos to be recommended as the short video that best matches the user's preferences.

[0071] Based on the short video that best matches user preferences (which is also the video to be recommended), identify the opposing short video cluster that best matches user preferences, and determine the mean of their relative distance difference.

[0072] Based on the feature vectors of each search item, calculate the projection distance of the feature vectors of each search item onto the reference dimension axis vector direction of the corresponding group. Then, the distance difference between the projection distance and the user preference distance in the reference dimension axis vector direction of the same short video tag is calculated (also known as the relative distance difference between each search content and the user preference). Calculate the absolute value of the difference between the distance difference and the mean relative distance difference between the opposite short video cluster that best matches the user's preference. Then, weight the difference according to the time decay weight of each search to obtain the probability of the steady state maintenance of the user preference.

[0073] The first short video in the collection of short videos to be recommended with the same short video tag Taking the minimum relative distance difference among the short videos to be recommended as an example, the specific calculation method includes: In the formula, Indicates the first Second search, For the first The projection distance of the feature vector of the search content in the next search onto the reference dimension axis vector of the corresponding group. For the first The difference between the projection distance of the search content feature vector of the second search onto the baseline dimension vector of the corresponding group and the user preference distance on the baseline dimension vector of the same short video tag. The probability of user preference maintaining steady state. For the number of user searches, The mean relative distance difference between the short video clusters that best match user preferences and their opposing short videos (the relative distance difference between the short video to be recommended and the opposing short video). This indicates the degree of difference between user search content and opposing short videos that the user prefers. The smaller the value, the more the user's search preference is shifting towards opposing short videos; the larger the value, the more the user is actively searching for content that matches their preference, or the user's preference remains stable. For the current time, For the first The search time is relatively long in this implementation because user preference evolution is a lengthy process. and Use days as the unit of measurement, and remove the dimensions before participating in the calculation. For the first Time decay weight of each search The preset time sensitivity adjustment factor has a value range of [0.5, 1.5]. A larger value indicates a faster time decay weight, while a smaller value indicates a slower time decay weight. In this embodiment... , No. The further back in time the search is from the current time, the smaller the time decay weight. The larger the value, the more stable the user's preferences are; the smaller the value, the more unstable the user's preferences are, and the more the user's search content begins to shift in the opposite direction of the user's preferences.

[0074] It should be noted that if a user has never actively searched within the preset period, the probability of the user preference remaining stable is set to 1, and there is no need to calculate it using the above method.

[0075] The method for calculating the feature vectors of each search content is, in principle, the same as the method for calculating the high-dimensional feature vectors in step 2.1). Furthermore, when the system executes an active search for content that is a simple text fragment sequence rather than a short video content source carrying audiovisual information, before invoking the high-dimensional extraction action, the process of bypassing or ignoring the optical character recognition extraction frames and the weighted high-order distribution of keyframes is bypassed. Instead, the system directly uses a basic text language encoding model to extract the independent feature vectors of the natural semantic units of the pure text fragment sequence itself, thus obtaining the feature vectors of each search content.

[0076] Since the user's search content is not a short video and does not have a short video tag, the methods for determining the short video tag corresponding to the search content include: Calculate the feature vectors of each short video tag, and then calculate the cosine similarity between the feature vectors of each short video tag and the feature vectors of each search content. The short video tag with the highest cosine similarity to the feature vectors of each search content is selected as the corresponding short video tag for that search content. If the feature vectors of the search content and the short video tags have different dimensions, add several zero elements after all elements of the lower-dimensional feature vector to make it equal in dimension to the higher-dimensional feature vector, and then calculate the cosine similarity again.

[0077] 5.3) Based on the steady-state maintenance probability and damping coefficient of user preferences in each group, the theoretical relative distance difference between the short videos to be recommended and the opposing short videos in each group is updated to obtain the updated theoretical relative distance difference between the short videos to be recommended and the opposing short videos. Then, the initial opposing recommendation degree of each opposing short video in the opposing short video cluster is updated to obtain the updated recommendation degree of each opposing short video.

[0078] When user preferences change, the previously strictly opposed positive and opposing short videos are no longer strictly opposed, which causes the theoretical relative distance difference between the opposing short videos of the short videos to be recommended to also change. In order to reflect this change, the theoretical relative distance difference between the opposing short videos of each short video to be recommended needs to be updated.

[0079] Specific update methods include: In the formula, The theoretical relative distance difference between the updated recommended short videos and the opposing short videos. The theoretical relative distance difference between the short videos to be recommended and the opposing short videos. This indicates a minimum value operation, used to prevent the relative distance difference between the updated recommended short video and its opposing short video cluster from exceeding the limit (more than 1). The preset threshold for preference stubbornness. To determine the degree of stubbornness, when = hour, ,Keep The value remains unchanged when At that time, it was believed that users' preferences were too stubborn and they could not accept opposing short videos. Less than 1, will Shrink, when At that time, it was believed that users' preferences were less stubborn and more evolving, sufficient to make them accept opposing short videos. Greater than 1, will Zoom in. The preference persistence threshold is set to [0.5, 1.5]. In this embodiment, its initial value is set to 1. If the churn rate increases after pushing opposing content to a certain short video tag group, the preference persistence threshold is automatically increased to raise the push threshold; if the group has long-term monotonous content and insufficient diversity, the preference persistence threshold is automatically slightly decreased to relax the triggering conditions for opposing content; during the adjustment process, the preference persistence threshold always falls within the range of [0.5, 1.5].

[0080] Based on the theoretical relative distance difference between the opposing short video clusters of the short video to be recommended before and after the update, the initial opposing recommendation degree of each opposing short video in the opposing short video cluster of the short video to be recommended is updated to obtain the updated opposing recommendation degree.

[0081] For any single opposing short video within a cluster of opposing short videos to be recommended, the specific method for updating its degree of opposition recommendation includes: In the formula, The first in the opposing short video cluster to be recommended The updated recommendation level of opposing short videos. The first in the opposing short video cluster to be recommended The initial degree of opposition in the recommendation of opposing short videos. The first in the opposing short video cluster to be recommended The relative distance between the two opposing short videos This indicates normalization processing. In this embodiment, maximum and minimum value normalization processing is used. During the processing, the maximum and minimum values ​​are the relative distance differences of all opposing short videos in the opposing short video cluster of the short video to be recommended. The theoretical relative distance between the updated recommended short video and the opposing short video. If the maximum and minimum values ​​of the absolute values ​​of the difference are the same (i.e. the range result is zero), then the normalization result of all processed objects in the group is directly preset to an equal weight (e.g., 0.5, not 1). The smaller the value, the more... The closer the relative distance difference between the opposing short videos is to the theoretical relative distance difference after the update, the better. The larger the value, The higher the value, the greater the degree of opposing recommendations after the update.

[0082] Step 6: Update the weighted directed graph and determine the short videos to be recommended and their opposing short videos according to the weighted directed graph.

[0083] The degree of opposition recommendation of each opposing short video in the opposing short video cluster of the updated short video to be recommended is used as the edge weight between the short video node to be recommended and the short video node to be recommended (opposing short videos also belong to the short videos to be recommended), thereby updating the weighted directed graph and obtaining the updated weighted directed graph.

[0084] Based on the edge weights (the degree of similarity between each recommended short video and the user's preferred short videos) between user nodes and recommended short video nodes in the updated weighted directed graph, the order of recommending positive short videos is determined. Then, based on the recommended short video node with the largest edge weight (the degree of opposition to the recommended short video) between each positive short video and other recommended short video nodes, the opposing short videos for each positive short video are determined. Finally, each positive short video and its opposing short video are recommended to the user in the order of the positive short videos.

[0085] Thus, this invention achieves the ability to recommend opposing short videos to users based on the evolution of user preferences, preventing them from falling into information barriers.

Claims

1. A personalized content recommendation method for a short video platform, characterized in that, Includes the following steps: 1) Obtain the short videos to be recommended and the user's preferred short videos from the user's historical short video browsing history; Based on the actual feature differences between each short video to be recommended and the short videos preferred by users, as well as the theoretical feature differences between each short video to be recommended and the short videos preferred by users, the short videos to be recommended are selected as the opposing short videos, and the initial degree of opposition recommendation for each opposing short video is determined. 2) Select test short videos from the short videos to be recommended; determine the opposing short videos based on the initial opposing recommendation degree of each opposing short video of the test short video; Transitional images and texts are generated based on the test short video and its opposing short video and pushed to users. The positive feedback of users to the transitional images and texts is recorded, and the user preference evolution feedback index is calculated. 3) Based on the user preference evolution feedback index, update the theoretical feature differences between the opposing short videos and the user preferred short videos for each short video to be recommended, and update the initial opposing recommendation degree based on the update results to obtain the updated opposing recommendation degree; 4) Based on the actual feature differences between each short video to be recommended and the short videos preferred by the user, determine the positive short videos to be recommended to the user from the short videos to be recommended, and determine the opposing short videos of each positive short video based on the updated degree of each opposing recommendation, and recommend each positive short video and its opposing short video to the user in a group.

2. The personalized content recommendation method for short video platforms according to claim 1, characterized in that, The calculation methods for the actual feature differences between each short video to be recommended and the short videos preferred by the user include: Construct the position feature vectors of user-preferred short videos and the position feature vectors of short videos to be recommended; calculate the projection distance of each position feature vector in the direction of the baseline dimension axis vector; The user preference distance is obtained by averaging the corresponding projection distances using the user preference ratings of each user's preferred short videos as weights. Based on the difference between the projection distance of the stance feature vector of each short video to be recommended in the direction of the baseline dimension axis vector and the user preference distance, the actual feature difference between each short video to be recommended and the user preference short video is determined.

3. The personalized content recommendation method for short video platforms according to claim 1, characterized in that, The methods for identifying the opposing short videos to each short video to be recommended include: For any short video to be recommended, the short videos that have the smallest difference between the actual feature difference and the theoretical feature difference between the short video to be recommended and the short video preferred by the user are selected as the short videos to be recommended.

4. The personalized content recommendation method for a short video platform according to claim 1, characterized in that, Methods for determining the initial degree of opposition in recommendations for each opposing short video include: For any short video to be recommended, calculate the mean of the actual feature differences between each opposing short video and the user's preferred short video; Based on the differences in actual characteristics between each opposing short video and the user-preferred short video and the mean, the initial opposing recommendation degree of each opposing short video in the opposing short video cluster of the short video to be recommended is obtained.

5. The personalized content recommendation method for a short video platform according to claim 1, characterized in that, Positive user feedback on transitional images and text includes: number of reads, dwell time, and number of shares; methods for calculating the user preference evolution feedback index include: The user preference evolution feedback index is obtained by combining ratings based on the number of reads, the duration of stay, and the number of shares.

6. The personalized content recommendation method for a short video platform according to claim 1, characterized in that, Methods for updating the theoretical characteristics of opposing short videos and user-preferred short videos for each short video to be recommended include: Calculate the probability of steady-state maintenance of user preferences based on user-initiated search content in the user's historical short video browsing records; Calculate the damping coefficient based on the user preference evolution feedback index; By combining the damping coefficient and the probability of steady-state maintenance of user preferences, the degree of preference stubbornness can be obtained. By integrating the inverse mapping results of preference stubbornness and the theoretical feature differences between the opposing short videos and the user-preferred short videos for each short video to be recommended, the updated theoretical feature differences between the opposing short videos and the user-preferred short videos for each short video to be recommended are obtained.

7. The personalized content recommendation method for a short video platform according to claim 6, characterized in that, Methods for calculating the steady-state maintenance probability of user preferences include: Based on the user's historical short video browsing records, feature vectors of each user's active search content are constructed, and the actual feature differences between each user's active search content and user preferences are calculated based on the feature vectors of each user's active search content. Calculate the absolute value of the difference between the actual feature difference between the user's active search content and the user's preference in each search and the mean of the actual feature difference between each opposing short video that best matches the user's preference. Then, calculate the weighted mean according to the time decay weight of each search to obtain the probability of steady-state maintenance of user preference.

8. The personalized content recommendation method for a short video platform according to claim 6, characterized in that, Methods for calculating the damping coefficient include: The damping coefficient is obtained by integrating the inverse mapping results of the user preference evolution feedback index and the actual feature differences between the recommended short videos that best match user preferences and the short videos that user preferences prefer.

9. The personalized content recommendation method for a short video platform according to claim 1, characterized in that, Methods for identifying the opposing short videos to each positive short video include: For any positive short video, select the opposing short video with the highest updated opposing recommendation level from its opposing short videos, and use it as the opposing short video for that positive short video.

10. A personalized content recommendation system for a short video platform, comprising a client and a server, wherein the client is used to collect a user's historical short video browsing history and send it to the server, characterized in that, The server is used to recommend short videos to users using the personalized content recommendation method of the short video platform as described in any one of claims 1-9.

Images