Short video real-time dynamic recommendation method and system based on deep learning

Abstract

The invention discloses a short video real-time dynamic recommendation method and system based on deep learning, and relates to the technical field of Internet, through overall execution of the steps, a user behavior set Beh can be divided and modeled under different time period sets Tsg, a time period feature set Tfe and a training set Trn are generated, and the time period feature set Tfe and the training set Trn are subjected to real-time dynamic recommendation. And finally obtaining an interest model set Mod and a time period weight set Twg. When a user request arrives, the corresponding interest model set Mod can be called immediately according to the time period set Tsg, and the recommendation result Rlt is dynamically generated in combination with the time period weight set Twg, so that the recommendation process not only avoids the problems of fuzzy interest description and static stiffness of the recommendation result caused by modeling of a single time dimension in a traditional method, but also improves the user experience. And the recommendation result Rlt can be adaptively updated along with the switching of the time period set Tsg, so that the timeliness, the matching degree and the user experience of the real-time recommendation of the short video are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a method and system for real-time dynamic recommendation of short videos based on deep learning. Background Art

[0002] In China, deep learning is widely used in recommendation systems, a key area of ​​application for information retrieval and personalized content delivery. Short video platforms, as the fastest-growing content delivery platform in recent years, have become a central application scenario for deep learning recommendation algorithms. Short video platforms not only need to handle the real-time flow of massive amounts of video data, but also need to instantly capture user interests and deliver relevant content, thereby increasing user viewing time and platform activity.

[0003] In existing short video recommendation practices, most recommendation systems generally adopt a unified approach to modeling users' overall interests. This approach aggregates users' long-term historical behavior to generate fixed interest tags, which are then used as the basis for recommendations. However, this approach has significant shortcomings in practical applications: it ignores the fact that user interests tend to drift over time. For example, users may prefer to watch short, light-hearted content during their commute, but are more willing to invest time in longer videos during their evenings. If the recommendation system fails to recognize these time-related interest shifts and simply relies on overall interest tags for recommendations, the recommended content will be out of touch with the user's current needs, affecting the immediacy and accuracy of recommendations. Therefore, existing recommendation models lack a mechanism for modeling interests across time periods, making it impossible to switch weights between different time periods. This results in a poor user experience and, in turn, calls for new technical solutions to address this issue. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a real-time dynamic recommendation method and system for short videos based on deep learning, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a real-time dynamic recommendation method for short videos based on deep learning, comprising the following steps: S1. Collect basic user behavior data to form a user behavior set Beh, including click count Clk, stay duration Dwt, and completion ratio Fin; and divide it into time period set Tsg according to timestamp; S2. Divide the user behavior set Beh into a time period set Tsg to obtain a time period feature set Tfe; and form a training set Trn for modeling; S3. Based on the training set Trn, establish interest model sets Mod under different time period sets Tsg, and calculate the corresponding time period weight sets Twg to form a weight switching strategy Wps; S4. When a user request arrives, the interest model set Mod is called according to the time period set Tsg, and the recommendation result Rlt is generated in combination with the time period weight set Twg.

[0006] Preferably, said S1 includes S11 and S12; S11. When a user uses a short video platform, extract the user's unique identifier and record it as a user identifier Uid, monitor and record the multi-dimensional behavioral signals generated when interacting with the short video content, and form a user behavior set Beh; The user behavior set Beh includes the number of clicks Clk, the length of stay Dwt, the completion ratio Fin, the user identifier Uid and the video identifier Vid; The number of clicks Clk is recorded in real time by users clicking on video covers and play buttons in recommendation lists and play pages. Each click event is written into the interaction event log and is accompanied by a unique video identifier, marked as a video identifier Vid. The logs are then accumulated in units of intraday time windows to obtain the number of clicks Clk of users in the corresponding window. The dwell time Dwt is calculated by recording the timestamps of the video entry and exit times when the player is running, and calculating the difference to obtain the dwell time of a single viewing; The short video platform backend log service writes the duration of stay into the behavior record table and stores it with the video identifier and user identifier as the index. When merging multiple viewing records of the same user, the user's duration of stay on the video is generated as the average value Dwt; The completion ratio Fin is determined by real-time monitoring of the start and end points of the playback progress bar. When the playback progress reaches more than 95% of the total video length, it is determined to be a completion event. The ratio of completion events in the number of playbacks of the same user is counted and marked as the user's completion ratio Fin on the video.

[0007] Preferably, S12, after completing the collection of the user behavior set Beh, simultaneously obtain the timestamp information corresponding to each behavior data, the timestamp information is automatically written by the front-end interaction log and the playback log, and each log record contains the user identifier, the video identifier and the time when the behavior occurred; Parse and classify the timestamp data to form a time period set Tsg and a time period annotation result Tlb corresponding to each behavioral data. The specific process includes timestamp standardization and time period division and annotation generation; The timestamp standardization process converts the timestamps in logs from different sources into a standard time format and removes time zone differences. The time period division and labeling generation is based on predefined division rules, dividing a day into multiple fixed time periods, each fixed time period represents a time period label, forming a time period set Tsg, and then mapping the timestamps that have been standardized to the corresponding time period labels, and generating the corresponding time period labeling results Tlb for each behavioral data.

[0008] Preferably, said S2 includes S21 and S22; S21, based on the acquired user behavior set Beh, time period set Tsg and time period annotation result Tlb, group the user behavior set Beh according to the time period set Tsg to generate a structured time period feature set Tfe; The time period feature set Tfe is obtained by grouping processing in steps S211, S212 and S213; S211. Generate a grouping key Key = {user identifier Uid, video identifier Vid, time period labeling result Tlb}; locate all atomic behavior records of the same user, the same video, and the same time period in the user behavior set Beh; S212: For each grouping key Key, directly reference the three basic indicators obtained in step S1 from the located atomic behavior record, including the number of clicks Clk, the dwell time Dwt, and the completion ratio Fin, and write them into the time period feature set Tfe; when there are multiple atomic records, only one is selected for reference according to the selection rule: Selection rule 1: Select the record with the latest timestamp; Selection rule 2: If the timestamps are the same, the one written first is selected based on the order in which the logs were written. S213. Use the grouping key Key={user ID Uid, video ID Vid, time period labeling result Tlb} as the unique index, and the fields of number of clicks Clk, stay duration Dwt, and completion ratio Fin are all directly referenced from the selected atomic records of the user behavior set Beh to form a feature record. After integrating all feature records, obtain the structured time period feature set Tfe.

[0009] Preferably, S22, based on the acquired time period feature set Tfe, and using the grouping key Key={user identifier Uid, video identifier Vid, time period labeling result Tlb} as the unique index, the three fields from the time period feature set Tfe: number of clicks Clk, stay time Dwt and completion ratio Fin are aligned and organized in the order of the time period set Tsg to form a training set Trm.

[0010] Preferably, said S3 includes S31, S32 and S33; S31, based on the acquired training set Trn, performing feature subset division and index extraction on the training set Trn, specifically including steps S311 and S312; S311. Divide the training set Trn according to the time period labeling result Tlb according to the grouping key Key to form multiple training subsets, each training subset corresponding to the time period set Tsg; S312. In each training subset, extract the following three core indicators: Core indicator 1: Behavior data volume Dct, obtained by counting the number of user behavior records contained in the training subset. The number of user behavior records corresponds to the number of different grouping keys, reflecting the user behavior activity level of the training subset in the corresponding time period; Core indicator 2: User number Uct, obtained by counting the number of unique user identifiers Uid in the training subset, reflecting the user coverage of the training subset in the corresponding time period; Core indicator three: time period distribution breadth Tgd, which is obtained by calculating the range span of the time period annotation results Tlb corresponding to the behavioral data in the training subset.

[0011] Preferably, S32, after completing the division of the training subsets, establishing an interest model for each training subset, and integrating all training subsets to form an interest model set Mod; In the process of establishing the interest model, the number of clicks Clk, stay duration Dwt and completion ratio Fin contained in the user behavior set Beh are used as input features, and the amount of behavioral data Dct, number of users Uct and time period distribution breadth Tgd extracted from the training subset are used as constraints.

[0012] Preferably, S33, after the establishment of the interest model set Mod is completed, the input features and constraints are jointly learned based on the long short-term memory network LSTM, and the performance status of each time period is output, wherein the performance status includes the model prediction accuracy Acc, the coverage user ratio Cov, and the fit with the target optimization index Fit; After obtaining the performance status of each time period, the model prediction accuracy Acc, coverage user ratio Cov, and the fit of the target optimization index Fit of different interest models in their respective time periods are compared and normalized to the range of [0,1]. The normalized model prediction accuracy Nor(Acc), normalized coverage user ratio Nor(Cov), and normalized fit of the target optimization index Nor(Fit) are obtained, and then the geometric average is calculated to generate the comprehensive performance Per. The comprehensive performance Per of each time period is mapped to the weight in the time period weight set Twg, and the time period weight Twg of each time period is obtained; Then, the time period weight set Twg is matched with the interest model set Mod to form a weight switching strategy Wps, which is used to dynamically switch different interest models according to the time period during the user behavior prediction and recommendation process.

[0013] Preferably, said S4 includes S41; S41. When a user request arrives, time identification is performed based on the request time to obtain a time period labeling result Tlb. The time period labeling result Tlb is then indexed in the weight switching strategy Wps to obtain the binding relationship between the interest model set Mod corresponding to the time period labeling result Tlb and the time period weight set Twg. The weight values ​​corresponding to the time period labeling result Tlb in the time period weight set Twg are the comprehensive performance Per of each interest model in the time period. According to the obtained comprehensive performance Per, the interest model set Mod of the time period labeling result Tlb is weighted and sorted, and the interest model ranked first in the comprehensive performance Per from high to low is selected as the execution object; According to the mapping relationship maintained in the weight switching strategy Wps, locate the grouping key Key set bound to the interest model and time segment labeling result Tlb of the selected execution object; The grouping key Key is an index composed of the user identifier Uid, the video identifier Vid and the time period labeling result Tlb; in the grouping key Key set, select the record consistent with the user identifier Uid of the current request, and extract the corresponding video identifier Vid set from it; use the video identifier Vid set as the seed, perform similarity expansion based on the content feature set Cfe, obtain the candidate content set, then deduplicate and regularize the candidate set, and prioritize it to generate the recommendation result Rlt and return it to the user in real time.

[0014] A real-time dynamic recommendation system for short videos based on deep learning, including a user data collection module, a time segmentation and feature extraction module, a deep learning and modeling module, and a matching and push module; The user data collection module collects basic behavior data of users to form a user behavior set Beh, including the number of clicks Clk, the length of stay Dwt and the completion ratio Fin; and divides it into a time period set Tsg according to the timestamp; The time period division and feature extraction module divides the user behavior set Beh into the time period set Tsg to obtain the time period feature set Tfe; and forms a training set Trn for modeling; The deep learning and modeling module establishes interest model sets Mod under different time period sets Tsg based on the training set Trn, and calculates the corresponding time period weight sets Twg to form a weight switching strategy Wps; When a user request arrives, the matching push module calls the interest model set Mod according to the time period set Tsg, and generates a recommendation result Rlt in combination with the time period weight set Twg.

[0015] The present invention provides a real-time dynamic recommendation method and system for short videos based on deep learning, which has the following beneficial effects: (1) Through the overall execution of the above steps, the user behavior set Beh can be divided and modeled under different time period sets Tsg, generating a time period feature set Tfe and a training set Trn, and finally obtaining an interest model set Mod and a time period weight set Twg. When a user request arrives, the corresponding interest model set Mod can be called immediately based on the time period set Tsg, and the recommendation result Rlt can be dynamically generated in combination with the time period weight set Twg. In this way, the recommendation process not only avoids the problem of fuzzy interest characterization and static rigidity of recommendation results caused by modeling a single time dimension in traditional methods, but also enables the recommendation result Rlt to be adaptively updated with the switching of the time period set Tsg, ensuring the sensitivity and accuracy of the recommendation result Rlt to the differences in user behavior in different time periods, thereby significantly improving the timeliness, matching and user experience of real-time short video recommendations.

[0016] (2) By using the long short-term memory network LSTM through the interest model set Mod, the model prediction accuracy Acc, the coverage user ratio Cov, and the fit of the target optimization index Fit are output, and the comprehensive performance Per is generated through normalization and geometric averaging, and finally mapped to the weights in the time period weight set Twg. The special advantage of this processing is that the time period weight set Twg does not rely solely on a single indicator, but is based on a comprehensive calculation of multi-dimensional performance status, thereby avoiding the risk of a single indicator overly dominating the model effect. Unlike the static weight allocation method commonly used in existing technologies, this method realizes the flexibility and pertinence of the weight switching strategy Wps in the user behavior prediction and recommendation links through the dynamic normalization and weight generation mechanism of the comprehensive performance Per. In this way, it is possible to accurately balance the prediction accuracy, user coverage breadth, and target optimization fit in different time periods, thereby ensuring the overall stability and balance of the recommendation result Rlt under multi-dimensional objectives.

[0017] (3) In the real-time scenario where the user request arrives, time recognition is first completed based on the request time to obtain the time period labeling result Tlb. The time period labeling result Tlb is then indexed in the weight switching strategy Wps, so that the binding relationship between the interest model set Mod corresponding to the time period labeling result Tlb and the time period weight set Twg can be directly obtained. Furthermore, the comprehensive performance Per of each interest model in the time period weight set Twg is used to weight the interest model set Mod, ensuring that when multiple interest models coexist, the interest model with the best comprehensive performance Per is always selected as the execution object. Through this mechanism, the grouping key Key set bound to the interest model and the time period labeling result Tlb can be quickly inferred, and the video identifier Vid set consistent with the user identifier Uid can be further filtered out. Then, the video identifier Vid set is used as a seed, and similar expansion is performed based on the content feature set Cfe to generate a candidate content set. Finally, after deduplication, regularization and priority sorting, the output recommendation result Rlt can be returned to the user in real time. The special benefit of this process is that it no longer relies on static or preset recommendation logic, but can dynamically identify the optimal interest model based on the comprehensive performance Per at the moment the user actually requests it, and expand it in combination with the video identifier Vid set directly associated with the user identifier Uid. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of the steps of the real-time dynamic recommendation method for short videos based on deep learning of the present invention; Figure 2 This is a schematic block diagram of the short video real-time dynamic recommendation system based on deep learning of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0020] Example 1: The present invention provides a real-time dynamic recommendation method for short videos based on deep learning. Figure 1 , including the following steps: S1. Collect basic user behavior data to form a user behavior set Beh, including click count Clk, stay duration Dwt, and completion ratio Fin; and divide it into time period set Tsg according to timestamp; S2. Divide the user behavior set Beh into a time period set Tsg to obtain a time period feature set Tfe; and form a training set Trn for modeling; S3. Based on the training set Trn, establish interest model sets Mod under different time period sets Tsg, and calculate the corresponding time period weight sets Twg to form a weight switching strategy Wps; S4. When a user request arrives, the interest model set Mod is called according to the time period set Tsg, and the recommendation result Rlt is generated in combination with the time period weight set Twg.

[0021] In this embodiment, through the overall execution of the above steps, the user behavior set Beh (including the number of clicks Clk, the length of stay Dwt, and the completion ratio Fin) can be divided and modeled under different time period sets Tsg, and the time period feature set Tfe and the training set Trn can be generated, and finally the interest model set Mod and the time period weight set Twg can be obtained. When a user request arrives, the corresponding interest model set Mod can be called immediately based on the time period set Tsg, and the recommendation result Rlt can be dynamically generated in combination with the time period weight set Twg. In this way, the recommendation process not only avoids the problems of fuzzy interest characterization and static rigidity of recommendation results caused by modeling a single time dimension in traditional methods, but also enables the recommendation result Rlt to be adaptively updated as the time period set Tsg is switched, ensuring the sensitivity and accuracy of the recommendation result Rlt to the differences in user behavior in different time periods, thereby significantly improving the timeliness, matching degree and user experience of real-time short video recommendations.

[0022] Example 2: Specifically: S1 includes S11 and S12; S11. When a user uses a short video platform, extract the user's unique identifier and record it as a user identifier Uid, monitor and record the multi-dimensional behavioral signals generated when interacting with the short video content, and form a user behavior set Beh; The user behavior set Beh includes the number of clicks Clk, the length of stay Dwt, the completion ratio Fin, the user identifier Uid and the video identifier Vid; The number of clicks Clk is recorded in real time by users clicking on video covers and play buttons in recommendation lists and play pages. Each click event is written into the interaction event log and is accompanied by a unique video identifier, marked as a video identifier Vid. The logs are then accumulated in units of intraday time windows to obtain the number of clicks Clk of users in the corresponding window. The dwell time Dwt is calculated by recording the timestamps of the video entry and exit times when the player is running, and calculating the difference to obtain the dwell time of a single viewing; The short video platform backend log service writes the duration of stay into the behavior record table and stores it with the video identifier and user identifier as the index. When merging multiple viewing records of the same user, the user's duration of stay on the video is generated as the average value Dwt; The completion ratio Fin is determined by real-time monitoring of the start and end points of the playback progress bar. When the playback progress reaches more than 95% of the total video length, it is determined to be a completion event. The ratio of completion events in the number of playbacks of the same user is counted and marked as the user's completion ratio Fin on the video.

[0023] S12. After completing the collection of the user behavior set Beh, obtain the timestamp information corresponding to each behavior data at the same time. The timestamp information is automatically written by the front-end interaction log and the playback log. Each log record contains the user identifier, video identifier and the time when the behavior occurred; Parse and classify the timestamp data to form a time period set Tsg and a time period annotation result Tlb corresponding to each behavioral data. The specific process includes timestamp standardization and time period division and annotation generation; The timestamp standardization process converts the timestamps in logs from different sources into a standard time format (year, month, day, hour, minute) and removes time zone differences to ensure consistency in the time records of all behavioral data. The time period division and labeling generation is based on predefined division rules, dividing a day into multiple fixed time periods. Each fixed time period represents a time period label to form a time period set Tsg. For example, 06:00–09:00 is defined as "morning", 09:00–11:00 is defined as "commuting", and 19:00–23:00 is defined as "night". The timestamps that have undergone timestamp standardization are then mapped to the corresponding time period labels, and a corresponding time period labeling result Tlb is generated for each piece of behavioral data. The time period labeling result Tlb corresponds one-to-one with the user identifier and video identifier of the behavioral data to ensure logical traceability and consistency in subsequent feature processing and modeling links.

[0024] In this embodiment, the integrity and accuracy of the data source can be guaranteed in the collection of the user behavior set Beh (including the number of clicks Clk, the length of stay Dwt, the completion ratio Fin, the user identifier Uid and the video identifier Vid), and the time period set Tsg and the corresponding time period labeling result Tlb are formed through the standardization of timestamps and the division of time periods. Since each piece of behavior data corresponds one-to-one with the user identifier Uid, the video identifier Vid and the time period labeling result Tlb, the logical consistency and traceability of the feature data can be ensured in the subsequent modeling process. Unlike the common user behavior data in the prior art that lacks time granularity distinction and data confusion caused by inconsistent timestamps across log sources, this method realizes fine-grained time period management of multi-dimensional user behavior signals through a unified timestamp standardization and time period division strategy. It can accurately reflect the user's real behavior pattern in different time periods when generating the interest model set Mod in the future, avoid modeling distortion caused by data deviation, and thus improve the reliability and long-term stability of recommendations.

[0025] Example 3: Specifically: S2 includes S21 and S22; S21, based on the acquired user behavior set Beh, time period set Tsg and time period annotation result Tlb, group the user behavior set Beh according to the time period set Tsg to generate a structured time period feature set Tfe; The time period feature set Tfe is obtained by grouping processing in steps S211, S212 and S213; S211. Generate a grouping key Key = {user ID Uid, video ID Vid, time period labeling result Tlb}; locate all atomic behavior records of the same user, the same video, and the same time period in the user behavior set Beh. It should be noted that this is only used to locate the records and no statistical aggregation is performed; S212: For each grouping key Key, directly reference the three basic indicators obtained in step S1 from the located atomic behavior record, including the number of clicks Clk, the dwell time Dwt, and the completion ratio Fin, and write them into the time period feature set Tfe; when there are multiple atomic records, only one is selected for reference according to the selection rule: Selection rule 1: Select the record with the latest timestamp; Selection rule 2: If the timestamps are the same, the one written first is selected based on the order in which the logs were written. It should be noted that when writing the time period feature set Tfe, the number of clicks Clk, the dwell time Dwt, and the completion ratio Fin are not recalculated. Only record selection and field copying are performed to ensure consistency with the user behavior set Beh. S213. Use the grouping key Key={user ID Uid, video ID Vid, time period labeling result Tlb} as the unique index, and the fields of number of clicks Clk, stay duration Dwt, and completion ratio Fin are all directly referenced from the selected atomic records of the user behavior set Beh to form a feature record. After integrating all feature records, obtain the structured time period feature set Tfe.

[0026] S22. Based on the acquired time period feature set Tfe, and using the grouping key Key = {user ID Uid, video ID Vid, time period annotation result Tlb} as a unique index, the three fields from the time period feature set Tfe: click count Clk, dwell time Dwt, and completion rate Fin are aligned and organized in the order of the time period set Tsg to form a training set Trm; It should be noted that this process maintains a direct reference to the time period feature set Tfe and does not perform secondary processing or calculation.

[0027] In this embodiment, through the processing of S21 and S22, a structured time feature set Tfe can be generated based on the user behavior set Beh, the time period set Tsg, and the time period labeling result Tlb, and further organized to form a training set Trm. The special advantage of this process is that the grouping key Key={user identifier Uid, video identifier Vid, time period labeling result Tlb} is uniformly applied to all data processing links, ensuring the complete consistency of the behavior records, feature sets, and training sets in the index dimension. Since the number of clicks Clk, the length of stay Dwt, and the completion ratio Fin are all directly derived from the original records of the user behavior set Beh, and are accurately selected through accurate selection rules without any secondary processing or statistics, the numerical deviation that may be caused by repeated calculations or aggregation operations is avoided. Unlike the common problems of data distortion or time period misalignment in the aggregation process of behavioral features in existing technologies, this method uses a direct reference mechanism and a sequential alignment mechanism to enable the training set Trm to accurately reflect the user's real interaction behavior in each time period under the dimension of the time period set Tsg, thereby significantly improving the authenticity and stability of the training set Trm in the modeling stage, ensuring that subsequent model training can capture the subtle differences in user interests in the time dimension.

[0028] Example 4: Specifically: S3 includes S31, S32 and S33; S31, based on the acquired training set Trn, performing feature subset division and index extraction on the training set Trn, specifically including steps S311 and S312; S311. Based on the grouping key Key, the training set Trn is divided according to the time period labeling result Tlb to form multiple training subsets. Each training subset corresponds to a time period set Tsg one-to-one, ensuring that the source of the model training data under each time period set Tsg is clear; S312. In each training subset, extract the following three core indicators: Core indicator 1: Behavior data volume Dct, obtained by counting the number of user behavior records contained in the training subset. The number of user behavior records corresponds to the number of different grouping keys, reflecting the user behavior activity level of the training subset in the corresponding time period; Core indicator 2: User number Uct, obtained by counting the number of unique user identifiers Uid in the training subset, reflecting the user coverage of the training subset in the corresponding time period; Core indicator three: time period distribution breadth Tgd, obtained by calculating the range span of the time period labeling results Tlb corresponding to the behavioral data in the training subset. For example, if the behavioral data in a training subset covers the three different time period labels of "morning", "commuting" and "night" in the time period set Tsg, then the span indicates the distribution completeness of the training subset in different time periods.

[0029] S32. After completing the division of the training subsets, establish an interest model for each training subset, and integrate all training subsets to form an interest model set Mod; In the process of establishing the interest model, the number of clicks Clk, the length of stay Dwt and the completion ratio Fin contained in the user behavior set Beh are used as input features, and the amount of behavioral data Dct, the number of users Uct and the time period distribution breadth Tgd extracted from the training subset are used as constraints. The input features and constraints are used as the input data source of the long short-term memory network LSTM, which is used to provide unified data support when calculating the time period weight set Twg after training the interest model, ensuring that the training process of the interest model in different time periods is fully representative and stable.

[0030] S33. After the interest model set Mod is established, the input features and constraints are jointly learned based on the long short-term memory network LSTM, and the performance status of each time period is output. The performance status includes the model prediction accuracy Acc, the coverage user ratio Cov, and the fit with the target optimization indicator Fit; After obtaining the performance status of each time period, the model prediction accuracy Acc, coverage user ratio Cov, and the fit of the target optimization index Fit of different interest models in their respective time periods are compared and normalized to the range of [0,1]. The normalized model prediction accuracy Nor(Acc), normalized coverage user ratio Nor(Cov), and normalized fit of the target optimization index Nor(Fit) are obtained, and then the geometric average is calculated to generate the comprehensive performance Per. The comprehensive performance Per of each time period is mapped to the weight in the time period weight set Twg, and the time period weight Twg of each time period is obtained; The model prediction accuracy Acc is used to measure the accuracy of the interest model in predicting user behavior in the corresponding time period. The higher the model prediction accuracy Acc, the more accurately the model in that time period can reflect the changes in user interests, and the lower the Acc, the more accurately the model can reflect the changes in user interests. The coverage user ratio Cov represents the percentage of valid users that the statistical interest model can cover in this time period. When the coverage user ratio Cov is higher, it means that the interest model has explanatory power for more user behaviors and is not limited to a small group. The smaller the coverage user ratio Cov, the opposite is true. The Fit of the target optimization indicator is a measure of how well the recommendation results of the interest model within the time period fit the preset optimization goal (such as increased click-through rate, improved completion rate, or increased dwell time). The higher the Fit of the target optimization indicator, the stronger the optimization effect of the model in terms of goal orientation, and the smaller the Fit, the stronger the optimization effect. The comprehensive performance Per is obtained by the following calculation formula: ; Where, Represents the comprehensive performance of the i-th specific time period in the time period set Tsg; represents the normalized model prediction accuracy of the i-th specific time period; represents the normalized coverage user ratio in the i-th specific time period; It represents the fitting degree Nor (Fit) of the normalized target optimization index in the i-th specific time period; The time period weight Twg of each time period is obtained by the following calculation formula: ; Where, represents the weight of the i-th specific time period in the time period set Tsg, n represents the total number of time periods in the time period set Tsg, represents the comprehensive performance of the jth specific time period in the time period set Tsg; Then, the time period weight set Twg is matched with the interest model set Mod to form a weight switching strategy Wps, which is used to dynamically switch different interest models according to the time period during the user behavior prediction and recommendation process.

[0031] In this embodiment, by completing the decomposition and indicator extraction of the training set Trn, and based on this, introducing the behavioral data volume Dct, the number of users Uct, and the time period distribution breadth Tgd as constraints, the process of establishing the interest model set Mod ensures that it can truly reflect the user activity, coverage, and distribution integrity in different time period sets Tsg. Furthermore, using the long short-term memory network (LSTM) to jointly learn the input features (number of clicks Clk, dwell time Dwt, and completion rate Fin) with the above constraints, the model prediction accuracy Acc, the coverage user ratio Cov, and the fit of the target optimization indicator Fit are output. The comprehensive performance Per is generated through normalization and geometric averaging, and ultimately mapped to the weights in the time period weight set Twg. A special advantage of this process is that the time period weight set Twg does not rely solely on a single indicator, but is based on a comprehensive calculation of multi-dimensional performance states, thus avoiding the risk of a single indicator overly dominating the model effect. Unlike the static weight allocation method commonly used in the prior art, this method achieves flexibility and targeted weight switching strategy Wps in the user behavior prediction and recommendation process through dynamic normalization and weight generation mechanism of the comprehensive performance Per. In this way, it is possible to accurately balance prediction accuracy, user coverage breadth and target optimization fit in different time periods, thereby ensuring the overall stability and balance of the recommendation results Rlt under multi-dimensional goals.

[0032] Example 5: Specifically: S41, when a user request arrives, time identification is completed according to the request time to obtain the time period labeling result Tlb, and the time period labeling result Tlb is then indexed in the weight switching strategy Wps to obtain the binding relationship between the interest model set Mod corresponding to the time period labeling result Tlb and the time period weight set Twg; wherein, each weight value corresponding to the time period labeling result Tlb in the time period weight set Twg is the comprehensive performance Per of each interest model in the time period; According to the obtained comprehensive performance Per, the interest model set Mod of the time period labeling result Tlb is weighted and sorted, and the interest model ranked first in the comprehensive performance Per from high to low is selected as the execution object; According to the mapping relationship maintained in the weight switching strategy Wps, locate the grouping key Key set bound to the interest model and time segment labeling result Tlb of the selected execution object; The grouping key Key is an index composed of the user identifier Uid, the video identifier Vid and the time period labeling result Tlb; in the grouping key Key set, select the record consistent with the user identifier Uid of the current request, and extract the corresponding video identifier Vid set from it; use the video identifier Vid set as the seed, perform similarity expansion based on the content feature set Cfe, obtain the candidate content set, then deduplicate and regularize the candidate set, and prioritize it to generate the recommendation result Rlt and return it to the user in real time.

[0033] In this embodiment, in a real-time scenario where a user request arrives, time identification is first performed based on the request time to obtain the time period labeling result Tlb. The time period labeling result Tlb is then indexed in the weight switching strategy Wps, thereby directly obtaining the binding relationship between the interest model set Mod corresponding to the time period labeling result Tlb and the time period weight set Twg. Furthermore, the interest model set Mod is weighted using the comprehensive performance Per of each interest model in the time period weight set Twg, ensuring that when multiple interest models coexist, the interest model with the best comprehensive performance Per is always selected as the execution object. Through this mechanism, the grouping key Key set bound to the interest model and the time period labeling result Tlb can be quickly inferred, and the video identifier Vid set consistent with the user identifier Uid can be further filtered out. Then, using the video identifier Vid set as a seed, similarity expansion is performed based on the content feature set Cfe to generate a candidate content set. Finally, after deduplication, regularization, and priority sorting, the output recommendation result Rlt can be returned to the user in real time. The special benefit of this process is that it no longer relies on static or preset recommendation logic, but can dynamically identify the optimal interest model based on the comprehensive performance Per at the moment of the user's actual request, and expand it in combination with the set of video identifiers Vid directly associated with the user identifier Uid. For example, when a user initiates a request during the lunch break, the time period labeling result Tlb can be used to prioritize the interest model with the highest comprehensive performance Per during the lunch break, and the video identifiers Vid that the user watched and completed with a high rate during that time period can be inferred. Based on this, a set of candidate videos with similar styles, themes, and durations can be expanded, thereby more accurately fitting the user's immediate viewing habits and interests, achieving true "scenario-based instant recommendation" and avoiding recommendation results that are out of touch with the current time period's interests.

[0034] Example 6: Real-time dynamic recommendation system for short videos based on deep learning, please refer to Figure 2 ,Specifically: user data collection module, time period division and feature extraction module, deep learning and modeling module and matching push module; The user data collection module collects basic behavior data of users to form a user behavior set Beh, including the number of clicks Clk, the length of stay Dwt and the completion ratio Fin; and divides it into a time period set Tsg according to the timestamp; The time period division and feature extraction module divides the user behavior set Beh into the time period set Tsg to obtain the time period feature set Tfe; and forms a training set Trn for modeling; The deep learning and modeling module establishes interest model sets Mod under different time period sets Tsg based on the training set Trn, and calculates the corresponding time period weight sets Twg to form a weight switching strategy Wps; When a user request arrives, the matching push module calls the interest model set Mod according to the time period set Tsg, and generates a recommendation result Rlt in combination with the time period weight set Twg.

[0035] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A real-time dynamic recommendation method for short videos based on deep learning, characterized by: The following steps are involved: S1. Collect basic user behavior data to form a user behavior set Beh, including click count Clk, stay duration Dwt and completion ratio Fin; Divide into time period set Tsg according to timestamp; S2. Divide the user behavior set Beh into a time period set Tsg to obtain a time period feature set Tfe; and form a training set Trn for modeling; S3. Based on the training set Trn, establish interest model sets Mod under different time period sets Tsg, and calculate the corresponding time period weight sets Twg to form a weight switching strategy Wps; S4. When a user request arrives, the interest model set Mod is called according to the time period set Tsg, and the recommendation result Rlt is generated in combination with the time period weight set Twg.

2. The method for real-time dynamic recommendation of short videos based on deep learning according to claim 1, characterized in that: Said S1 includes S11 and S12; S11. When a user uses a short video platform, extract the user's unique identifier and record it as a user identifier Uid, monitor and record the multi-dimensional behavioral signals generated when interacting with the short video content, and form a user behavior set Beh; The user behavior set Beh includes the number of clicks Clk, the length of stay Dwt, the completion ratio Fin, the user identifier Uid and the video identifier Vid; The number of clicks Clk is recorded in real time by the user clicking on the video cover and the play button in the recommendation list and the play page; each click event is written into the interaction event log and is accompanied by a unique video identifier, which is marked as the video identifier Vid; Then, the logs are accumulated and calculated based on the intraday time window to obtain the number of clicks Clk of the user in the corresponding window; The dwell time Dwt is calculated by recording the timestamps of the video entry and exit times when the player is running, and calculating the difference to obtain the dwell time of a single viewing; The short video platform backend log service writes the duration of stay into the behavior record table and stores it with the video identifier and user identifier as the index. When merging multiple viewing records of the same user, the user's duration of stay on the video is generated as the average value Dwt; The completion ratio Fin is determined by real-time monitoring of the start and end points of the playback progress bar. When the playback progress reaches more than 95% of the total video length, it is determined to be a completion event. The ratio of completion events in the number of playbacks of the same user is counted and marked as the user's completion ratio Fin on the video.

3. The method for real-time dynamic recommendation of short videos based on deep learning according to claim 2, characterized in that: S12. After completing the collection of the user behavior set Beh, obtain the timestamp information corresponding to each behavior data at the same time. The timestamp information is automatically written by the front-end interaction log and the playback log. Each log record contains the user identifier, video identifier and the time when the behavior occurred; Parse and classify the timestamp data to form a time period set Tsg and a time period annotation result Tlb corresponding to each behavioral data. The specific process includes timestamp standardization and time period division and annotation generation; The timestamp standardization process converts the timestamps in logs from different sources into a standard time format and removes time zone differences. The time period division and labeling generation is based on predefined division rules, dividing a day into multiple fixed time periods, each fixed time period represents a time period label, forming a time period set Tsg, and then mapping the timestamps that have been standardized to the corresponding time period labels, and generating the corresponding time period labeling results Tlb for each behavioral data.

4. The method for real-time dynamic recommendation of short videos based on deep learning according to claim 3, characterized in that: Said S2 includes S21 and S22; S21, based on the acquired user behavior set Beh, time period set Tsg and time period annotation result Tlb, group the user behavior set Beh according to the time period set Tsg to generate a structured time period feature set Tfe; The time period feature set Tfe is obtained by grouping processing in steps S211, S212 and S213; S211. Generate a grouping key Key = {user identifier Uid, video identifier Vid, time period labeling result Tlb}; locate all atomic behavior records of the same user, the same video, and the same time period in the user behavior set Beh; S212: For each grouping key Key, directly reference the three basic indicators obtained in step S1 from the located atomic behavior record, including the number of clicks Clk, the dwell time Dwt, and the completion ratio Fin, and write them into the time period feature set Tfe; when there are multiple atomic records, only one is selected for reference according to the selection rule: Selection rule 1: Select the record with the latest timestamp; Selection rule 2: If the timestamps are the same, the one written first is selected based on the order in which the logs were written. S213. Use the grouping key Key={user ID Uid, video ID Vid, time period labeling result Tlb} as the unique index, and the fields of number of clicks Clk, stay duration Dwt, and completion ratio Fin are all directly referenced from the selected atomic records of the user behavior set Beh to form a feature record. After integrating all feature records, obtain the structured time period feature set Tfe.

5. The method for real-time dynamic recommendation of short videos based on deep learning according to claim 4, characterized in that: S22. Based on the acquired time period feature set Tfe, and using the grouping key Key = {user ID Uid, video ID Vid, time period annotation result Tlb} as the unique index, the three fields from the time period feature set Tfe: number of clicks Clk, dwell time Dwt, and completion ratio Fin are aligned and organized in the order of the time period set Tsg to form the training set Trm.

6. The method for real-time dynamic recommendation of short videos based on deep learning according to claim 5, characterized in that: Said S3 includes S31, S32 and S33; S31, based on the acquired training set Trn, performing feature subset division and index extraction on the training set Trn, specifically including steps S311 and S312; S311. Divide the training set Trn according to the time period labeling result Tlb according to the grouping key Key to form multiple training subsets, each training subset corresponding to the time period set Tsg; S312. In each training subset, extract the following three core indicators: Core indicator 1: Behavior data volume Dct, obtained by counting the number of user behavior records contained in the training subset. The number of user behavior records corresponds to the number of different grouping keys, reflecting the user behavior activity level of the training subset in the corresponding time period; Core indicator 2: User number Uct, obtained by counting the number of unique user identifiers Uid in the training subset, reflecting the user coverage of the training subset in the corresponding time period; Core indicator three: time period distribution breadth Tgd, which is obtained by calculating the range span of the time period annotation results Tlb corresponding to the behavioral data in the training subset.

7. The method for real-time dynamic recommendation of short videos based on deep learning according to claim 6, characterized in that: S32. After completing the division of the training subsets, establish an interest model for each training subset, and integrate all training subsets to form an interest model set Mod; In the process of establishing the interest model, the number of clicks Clk, stay duration Dwt and completion ratio Fin contained in the user behavior set Beh are used as input features, and the amount of behavioral data Dct, number of users Uct and time period distribution breadth Tgd extracted from the training subset are used as constraints.

8. The method for real-time dynamic recommendation of short videos based on deep learning according to claim 7, characterized in that: S33. After the interest model set Mod is established, the input features and constraints are jointly learned based on the long short-term memory network LSTM, and the performance status of each time period is output. The performance status includes the model prediction accuracy Acc, the coverage user ratio Cov, and the fit with the target optimization indicator Fit; After obtaining the performance status of each time period, the model prediction accuracy Acc, coverage user ratio Cov, and the fit of the target optimization index Fit of different interest models in their respective time periods are compared and normalized to the range of [0,1]. The normalized model prediction accuracy Nor(Acc), normalized coverage user ratio Nor(Cov), and normalized fit of the target optimization index Nor(Fit) are obtained, and then the geometric average is calculated to generate the comprehensive performance Per. The comprehensive performance Per of each time period is mapped to the weight in the time period weight set Twg, and the time period weight Twg of each time period is obtained; Then, the time period weight set Twg is matched with the interest model set Mod to form a weight switching strategy Wps, which is used to dynamically switch different interest models according to the time period during the user behavior prediction and recommendation process.

9. The method for real-time dynamic recommendation of short videos based on deep learning according to claim 8, characterized in that: Said S4 includes S41; S41. When a user request arrives, time identification is performed based on the request time to obtain a time period labeling result Tlb. The time period labeling result Tlb is then indexed in the weight switching strategy Wps to obtain the binding relationship between the interest model set Mod corresponding to the time period labeling result Tlb and the time period weight set Twg. The weight values ​​corresponding to the time period labeling result Tlb in the time period weight set Twg are the comprehensive performance Per of each interest model in the time period. According to the obtained comprehensive performance Per, the interest model set Mod of the time period labeling result Tlb is weighted and sorted, and the interest model ranked first in the comprehensive performance Per from high to low is selected as the execution object; According to the mapping relationship maintained in the weight switching strategy Wps, locate the grouping key Key set bound to the interest model and time segment labeling result Tlb of the selected execution object; The grouping key Key is an index composed of the user identifier Uid, the video identifier Vid and the time period labeling result Tlb; in the grouping key Key set, select the record consistent with the user identifier Uid of the current request, and extract the corresponding video identifier Vid set from it; use the video identifier Vid set as the seed, perform similarity expansion based on the content feature set Cfe, obtain the candidate content set, then deduplicate and regularize the candidate set, and prioritize it to generate the recommendation result Rlt and return it to the user in real time.

10. A short video real-time dynamic recommendation system based on deep learning, applied to the short video real-time dynamic recommendation method based on deep learning according to any one of claims 1 to 9, characterized in that: User data collection module, time segmentation and feature extraction module, deep learning and modeling module, and matching push module; The user data collection module collects basic behavior data of users to form a user behavior set Beh, including the number of clicks Clk, the length of stay Dwt and the completion ratio Fin; Divide into time period set Tsg according to timestamp; The time period division and feature extraction module divides the user behavior set Beh into the time period set Tsg to obtain the time period feature set Tfe; and forms a training set Trn for modeling; The deep learning and modeling module establishes interest model sets Mod under different time period sets Tsg based on the training set Trn, and calculates the corresponding time period weight sets Twg to form a weight switching strategy Wps; When a user request arrives, the matching push module calls the interest model set Mod according to the time period set Tsg, and generates a recommendation result Rlt in combination with the time period weight set Twg.

Images