Deep learning-based short video real-time dynamic recommendation method and system

Abstract

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

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet, in particular to a short video real-time dynamic recommendation method and system based on deep learning. BACKGROUND

[0002] In the application branch of artificial intelligence, deep learning is widely used in recommendation systems, and the recommendation system is an important direction of information retrieval and personalized content pushing. In the application form of the recommendation system, short video platforms have become the concentrated application scene of deep learning recommendation algorithms as the fastest growing content transmission carrier in recent years. Short video platforms not only need to cope with the real-time flow of massive video data, but also need to grasp the user interest in a moment and push appropriate content to improve the user's viewing time and platform activity.

[0003] In the existing short video recommendation practice, most recommendation systems generally adopt a unified modeling method for user overall interest, that is, by summarizing the user's long-term historical behavior, a fixed interest label is generated, and then the recommendation is made based on this. However, this method has significant shortcomings in actual application: it ignores the objective law that user interest will drift with time period. For example, users may prefer to watch light and short content during the commuting period, but prefer to invest time in watching long videos during the evening leisure period. If the recommendation system fails to identify this time period-related interest change and simply relies on the overall interest label for recommendation, it will cause the pushed content to be inconsistent with the user's actual needs at the moment, affecting the immediacy and accuracy of the recommendation. Therefore, the existing recommendation mode lacks a cross-time period interest modeling mechanism and cannot switch weights at different time periods, resulting in insufficient user experience, which leads to the need for a new technical solution to solve this problem. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a short video real-time dynamic recommendation method and system based on deep learning, which solves the problems mentioned in the background art.

[0005] To achieve the above purpose, the present application realizes the following technical scheme: a short video real-time dynamic recommendation method based on deep learning, comprising the following steps:

[0006] S1, collect the basic behavior data of the user to form a user behavior set Beh, including the number of clicks Clk, the dwell time Dwt and the completion ratio Fin; and divide according to the time stamp to form a time period set Tsg;

[0007] S2, divide the user behavior set Beh according to the time period set Tsg to obtain a time period feature set Tfe; and form a training set Trn for modeling;

[0008] S3, based on the training set Trn, respectively establish interest model set Mod under different time period set Tsg, and calculate the corresponding time period weight set Twg, form weight switching strategy Wps;

[0009] S4, when the user request arrives, according to the time period set Tsg, call the interest model set Mod, and combine the time period weight set Twg to generate the recommendation result Rlt.

[0010] Preferably, the S1 includes S11 and S12;

[0011] S11, when the user uses the short video platform, extract the user unique identification record as the user identification Uid, monitor and record the multi-dimensional behavior signal generated when interacting with the short video content, and form the user behavior set Beh;

[0012] The user behavior set Beh includes the click number Clk, the stay time Dwt, the complete play ratio Fin, the user identification Uid and the video identification Vid;

[0013] The click number Clk is recorded in real time by triggering the video cover click and the play button click in the recommendation list and the play page; Each click event is written into the interaction event log, and has a unique video identifier, which is marked as the video identification Vid; Then, the log is accumulated and calculated in the daily time window, and the click number Clk of the user in the corresponding window is obtained;

[0014] The stay time Dwt records the time stamp of the video entering time and the exit time when the player runs, and the stay time of single viewing is obtained by calculating the difference;

[0015] The short video platform background log service writes the stay time into the behavior record table, and stores it with the video identifier and the user identifier as the index, when merging multiple viewing records of the same user, in the form of average value, the user's stay time Dwt on the video is generated;

[0016] The complete play ratio Fin is obtained by real-time monitoring the start point and the end point of the play progress bar, when the play progress reaches more than 95% of the total time of the video, it is determined as a complete play event; The proportion of complete play events in the play times of the same user is marked as the complete play ratio Fin of the user on the video.

[0017] Preferably, S12, after completing the user behavior set Beh collection, the time stamp information corresponding to each piece of behavior data is obtained, the time stamp information is automatically written by the front-end interaction log and the play log, and each log record contains the user identifier, the video identifier and the behavior occurrence time;

[0018] The timestamp data is parsed and classified to form a time period set Tsg and a time period label result Tlb corresponding to each behavior data, and the specific process includes timestamp standardization processing and time period division and label generation;

[0019] The timestamp standardization processing converts the timestamps in different source logs into a standard time format and removes the time zone difference;

[0020] The time period division and label generation divides one day into a plurality of fixed time periods according to a predefined division rule, each fixed time period representing a time period label to form a time period set Tsg, and then maps the timestamp after the timestamp standardization processing to the corresponding time period label and generates a time period label result Tlb corresponding to each behavior data.

[0021] Preferably, the S2 includes S21 and S22;

[0022] S21, based on the obtained user behavior set Beh, time period set Tsg and time period label result Tlb, the user behavior set Beh is grouped according to the time period set Tsg to generate a structured time period feature set Tfe;

[0023] The time period feature set Tfe is obtained by grouping processing of steps S211, S212 and S213;

[0024] S211, a grouping key Key={user identification Uid, video identification Vid, time period label result Tlb} is generated, and all atomic behavior records under the same user, the same video and the same time period in the user behavior set Beh are located;

[0025] S212, for each grouping key Key, the three basic indicators including the click number Clk, the dwell time Dwt and the complete playback ratio Fin obtained in step S1 are directly referenced from the located atomic behavior records, and written into the time period feature set Tfe; when there are multiple atomic records, only one is selected according to the selection rule:

[0026] The first selection rule: selecting the record with the latest timestamp;

[0027] The second selection rule: if the timestamps are consistent, the first written record is selected according to the log writing order;

[0028] S213, taking the grouping key Key={user identification Uid, video identification Vid, time period label result Tlb} as a unique index, and the field click number Clk, the stay time Dwt and the complete play ratio Fin as direct references from the selected atomic records of the user behavior set Beh, a feature record is formed, and after integrating all feature records, a structured time period feature set Tfe is obtained.

[0029] Preferably, S22, based on the obtained time period feature set Tfe, the grouping key Key={user identification Uid, video identification Vid, time period label result Tlb} is taken as a unique index, and the three fields: click number Clk, stay time Dwt and complete play ratio Fin from the time period feature set Tfe are aligned and organized according to the order of the time period set Tsg to form a training set Trm.

[0030] Preferably, the S3 includes S31, S32 and S33.

[0031] S31, based on the obtained training set Trn, the training set Trn is divided into feature subsets and indexes are extracted, specifically including steps S311 and S312.

[0032] S311, according to the grouping key Key, the training set Trn is divided according to the time period label result Tlb to form multiple training subsets, each training subset corresponds to the time period set Tsg one by one.

[0033] S312, in each training subset, the following three core indexes are extracted:

[0034] Core index one: behavior data volume Dct, obtained by counting the number of user behavior records contained in the training subset, the number of user behavior records corresponding to different grouping keys Key, reflecting the user behavior activity degree of the training subset in the corresponding time period;

[0035] Core index two: user number Uct, obtained by counting the number of non-repeated user identification Uid in the training subset, reflecting the user coverage range of the training subset in the corresponding time period;

[0036] Core index three: time period distribution breadth Tgd, obtained by calculating the range span of the time period label result Tlb corresponding to the behavior data in the training subset.

[0037] Preferably, S32, after completing the training subset division, an interest model is established for each training subset respectively, and all training subsets are integrated to form an interest model set Mod.

[0038] In the process of establishing the interest model, the click times Clk, the dwell time Dwt and the complete play ratio Fin in the user behavior set Beh are taken as input features, and the behavior data amount Dct, the user amount Uct and the time period distribution width Tgd extracted from the sub-set are taken as constraint conditions.

[0039] Preferably, after the establishment of the interest model set Mod, joint learning is performed on the input features and the constraint conditions based on the long short-term memory network LSTM, and the performance state of each time period is output, wherein the performance state includes the model prediction accuracy Acc, the coverage user ratio Cov and the fitting degree Fit of the target optimization index.

[0040] After the performance state of the time period is obtained, the model prediction accuracy Acc, the coverage user ratio Cov and the fitting degree Fit of the target optimization index of different interest models in each time period are compared, normalized to the range of [0, 1], and the normalized model prediction accuracy Nor(Acc), the normalized coverage user ratio Nor(Cov) and the normalized fitting degree Nor(Fit) of the target optimization index are obtained. Then, the geometric mean is calculated to generate the comprehensive performance Per.

[0041] 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.

[0042] The time period weight set Twg and the interest model set Mod are then corresponded to form the weight switching strategy Wps, which is used to dynamically switch different interest models according to the belonging time period in the user behavior prediction and recommendation process.

[0043] Preferably, the S4 includes S41.

[0044] S41, when the user request arrives, the time period labeling result Tlb is obtained by completing time identification according to the request time, and the binding relationship between the interest model set Mod and the time period weight set Twg corresponding to the time period labeling result Tlb is obtained by indexing the time period labeling result Tlb in the weight switching strategy Wps, wherein the weight values corresponding to the time period labeling result Tlb in the time period weight set Twg are the comprehensive performances Per of the interest models in the time period.

[0045] The interest model set Mod of the time period labeling result Tlb is weighted and sorted according to the obtained comprehensive performance Per, and the interest model ranked first is selected as the execution object according to the comprehensive performance Per from high to low.

[0046] Based on the mapping relationship maintained in the weight switching strategy Wps, locate the set of grouping keys Key that are bound to the interest model and time segment labeling result Tlb of the selected execution object;

[0047] The grouping key Key is an index composed of user identifier Uid, video identifier Vid, and time period labeling result Tlb. In the grouping key Key set, the record that matches the user identifier Uid of the current request is selected, and the corresponding video identifier Vid set is extracted from it. Using the video identifier Vid set as a seed, similarity expansion is performed based on the content feature set Cfe to obtain a candidate content set. Then, the candidate set is deduplicated, normalized, and prioritized to generate the recommendation result Rlt, which is returned and pushed to the user in real time.

[0048] A real-time dynamic recommendation system for short videos based on deep learning, comprising a user data collection module, a time period segmentation and feature extraction module, a deep learning and modeling module, and a matching and push module;

[0049] The user data collection module collects basic user behavior data to form a user behavior set Beh, including the number of clicks Clk, the dwell time Dwt, and the completion rate Fin; and divides it into a time period set Tsg according to the timestamp.

[0050] The time period segmentation and feature extraction module divides the user behavior set Beh according to the time period set Tsg to obtain the time period feature set Tfe; and forms a training set Trn for modeling.

[0051] 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 set Twg to form a weight switching strategy Wps;

[0052] When a user request arrives, the matching and push module calls the interest model set Mod based on the time period set Tsg, and combines it with the time period weight set Twg to generate the recommendation result Rlt.

[0053] This invention provides a method and system for real-time dynamic recommendation of short videos based on deep learning, which has the following beneficial effects:

[0054] (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, the time period feature set Tfe and the training set Trn are generated, and finally the interest model set Mod and the time period weight set Twg are obtained. When the user request arrives, the corresponding interest model set Mod can be called according to the time period set Tsg in real time, and the recommendation result Rlt is dynamically generated in combination with the time period weight set Twg. In this way, the recommendation process not only avoids the problems of interest description ambiguity and recommendation result static rigidity caused by modeling of 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 behavior differences of users in different time periods, thereby significantly improving the timeliness, matching degree and user experience of real-time recommendation of short videos.

[0055] (2) Through the interest model set Mod using the long short-term memory network LSTM, the model prediction accuracy Acc, the coverage user ratio Cov and the fitting degree Fit of the target optimization index are output, and the comprehensive performance Per is generated through normalization and geometric average method, and finally mapped to the weight in the time period weight set Twg. The special advantage of this processing is that the time period weight set Twg not only depends on a single index, but is based on the comprehensive calculation of multi-dimensional performance state, thereby avoiding the risk of excessive dominance of a single index on model effect. Unlike the static weight allocation method commonly used in the prior art, the dynamic normalization and weight generation mechanism of the comprehensive performance Per are used in the method to realize the flexibility and pertinence of the weight switching strategy Wps in the user behavior prediction and recommendation link. In this way, the prediction accuracy, user coverage range and target optimization fitting degree can be accurately balanced under different time periods, thereby ensuring the overall stability and balance of the recommendation result Rlt under multi-dimensional targets.

[0056] (3) In the real-time scene of user request arrival, first, time identification is completed based on the request time to obtain the time period labeling result Tlb, and then the time period labeling result Tlb is 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. Further, the interest model set Mod is weighted and sorted by using the comprehensive performance Per of each interest model in the time period weight set Twg, so that in the case of multiple interest models coexisting, the interest model with the optimal 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 deduced, and the video identifier Vid set consistent with the user identifier Uid is further screened out, and then the video identifier Vid set is used as a seed to perform similar expansion based on the content feature set Cfe to generate a candidate content set. Finally, the recommended result Rlt output after deduplication, normalization and priority sorting processing 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 according to the comprehensive performance Per at the time of the actual request of the user, and expand in combination with the video identifier Vid set directly associated with the user identifier Uid. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 A step schematic diagram of the short video real-time dynamic recommendation method based on deep learning of the present application;

[0058] Figure 2 A block diagram schematic diagram of the short video real-time dynamic recommendation system based on deep learning of the present application. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0060] Embodiment 1: The present application provides a short video real-time dynamic recommendation method based on deep learning, please refer to Figure 1 , which includes the following steps:

[0061] S1, collect the basic behavior data of the user to form a user behavior set Beh, including the click frequency Clk, the stay time Dwt and the complete play ratio Fin; divide into a time period set Tsg according to the time stamp;

[0062] S2, divide the user behavior set Beh according to the time period set Tsg to obtain a time period feature set Tfe; and form a training set Trn for modeling;

[0063] S3, based on the training set Trn, respectively establish an interest model set Mod under different time period sets Tsg, and calculate a corresponding time period weight set Twg to form a weight switching strategy Wps;

[0064] S4, when a user request arrives, call the interest model set Mod according to the time period set Tsg, and generate a recommendation result Rlt in combination with the time period weight set Twg.

[0065] In this embodiment, through the overall execution of the above steps, the user behavior set Beh (including the number of clicks Clk, the dwell time Dwt and the complete playback ratio Fin) can be divided and modeled under different time period sets Tsg, the time period feature set Tfe and the training set Trn are generated, and finally the interest model set Mod and the time period weight set Twg are obtained. When a user request arrives, the corresponding interest model set Mod can be called in real time according to 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 description and static rigidity of the recommendation result caused by modeling of a single time dimension in the traditional method, 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 behavior differences of the user in different time periods, thereby significantly improving the timeliness, matching degree and user experience of real-time recommendation of short videos.

[0066] Embodiment 2: Specifically, the S1 includes S11 and S12;

[0067] S11, when a user uses a short video platform, extract a user unique identification record as a user identification Uid, monitor and record multi-dimensional behavior signals generated when the user interacts with short video content to form a user behavior set Beh;

[0068] The user behavior set Beh includes the number of clicks Clk, the dwell time Dwt, the complete playback ratio Fin, the user identification Uid and the video identification Vid;

[0069] The number of clicks Clk is recorded in real time by triggering the video cover clicks and the play button clicks in the recommendation list and the play page; each click event is written into an interaction event log and has a unique video identifier, which is marked as a video identification Vid; and the log is accumulated and calculated in a daily time window to obtain the number of clicks Clk of the user in the corresponding window;

[0070] The dwell time Dwt records the time stamp of the video entering time and the time stamp of the video exiting time when the player is running, and obtains the dwell time of a single viewing by calculating the difference value;

[0071] The short video platform background log service writes the dwell time into the behavior record table and stores it with the video identifier and the user identifier as indexes. When merging multiple viewing records of the same user, the dwell time of the user on the video is generated in the form of an average value.

[0072] The complete play ratio Fin is obtained by real-time monitoring of the start point and the end point of the play progress bar. When the play progress reaches more than 95% of the total time of the video, it is determined as a complete play event. The proportion of complete play events in the play times of the same user is marked as the complete play ratio Fin of the user on the video.

[0073] S12, after completing the user behavior set Beh collection, the time stamp information corresponding to each piece of behavior data is obtained, the time stamp information is automatically written by the front-end interaction log and the play log, and each log record contains the user identifier, the video identifier and the behavior occurrence time;

[0074] The time stamp data is analyzed and classified to form a time period set Tsg and a time period label result Tlb corresponding to each piece of behavior data. The specific process includes time stamp standardization processing and time period division and label generation;

[0075] The time stamp standardization processing unifies the time stamps in different source logs into a standard time format (year, month, day, hour, minute), and removes the time zone difference, so as to ensure the consistency of the time records of all behavior data;

[0076] The time period division and label generation divides one day into multiple fixed time periods according to the predefined division rule, each fixed time period represents a time period label, forms a time period set Tsg, for example, 06:00-09:00 is defined as “morning”, 09:00-11:00 is defined as “commute”, and 19:00-23:00 is defined as “night”. After the time stamp standardization processing, the time stamp is mapped to the corresponding time period label, and the time period label result Tlb corresponding to each piece of behavior data is generated. The time period label result Tlb corresponds to the user identifier and the video identifier of the behavior data one by one, so as to ensure the logical traceability and consistency in the subsequent feature processing and modeling links.

[0077] In this embodiment, the integrity and accuracy of the data source can be ensured in the collection of the user behavior set Beh (including the click count Clk, the dwell time 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 label result Tlb are formed through the standardized processing and time period division of the timestamp. Since each behavior data corresponds to the user identifier Uid, the video identifier Vid, and the time period label 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, which lacks time granularity division and causes data confusion due to inconsistent timestamps across log sources, the method realizes fine-grained periodization management of multi-dimensional user behavior signals through unified timestamp standardization and time period division strategies. When generating the interest model set Mod, the real behavior patterns of users in different time periods can be accurately reflected, modeling distortion caused by data bias can be avoided, and the reliability and long-term stability of the recommendation can be improved.

[0078] Embodiment 3: Specifically, the S2 includes S21 and S22;

[0079] S21, based on the acquired user behavior set Beh, time period set Tsg, and time period label result Tlb, grouping processing is performed on the user behavior set Beh according to the time period set Tsg, and a structured time period feature set Tfe is generated;

[0080] The time period feature set Tfe is obtained through grouping processing of steps S211, S212, and S213;

[0081] S211, a grouping key Key={user identifier Uid, video identifier Vid, time period label result Tlb} is generated; all atomic behavior records of the same user, the same video, and the same time period in the user behavior set Beh are located, and it should be noted that this is only used for positioning the record and does not perform any statistical aggregation;

[0082] S212, for each grouping key Key, the three basic indicators including the click count Clk, the dwell time Dwt, and the completion ratio Fin acquired in step S1 are directly referenced from the located atomic behavior records, and written into the time period feature set Tfe; when there are multiple atomic records, only one is selected according to the selection rule:

[0083] Selection rule one: selecting the record with the most recent timestamp;

[0084] Selection rule two: if the timestamps are consistent, the first written record is selected according to the log writing order;

[0085] It should be noted that: in the process of writing the time period feature set Tfe, the click count Clk, the dwell time Dwt and the complete playback ratio Fin are not re-calculated, only record selection and field copy are made, and consistency with the user behavior set Beh is ensured;

[0086] S213, with the grouping key Key={user identification Uid, video identification Vid, time period label result Tlb} as the unique index, and the fields of click count Clk, dwell time Dwt and complete playback ratio Fin being directly referenced from the selected atomic records of the user behavior set Beh, a feature record is formed, and after integrating all feature records, a structured time period feature set Tfe is obtained.

[0087] S22, based on the obtained time period feature set Tfe, and with the grouping key Key={user identification Uid, video identification Vid, time period label result Tlb} as the unique index, the three fields: click count Clk, dwell time Dwt and complete playback ratio Fin from the time period feature set Tfe are aligned and organized according to the order of the time period set Tsg to form a training set Trm;

[0088] It should be noted that: this process maintains direct reference to the time period feature set Tfe and does not perform secondary processing or calculation.

[0089] In this embodiment, through the processing of S21 and S22, a structured time period feature set Tfe can be generated based on the user behavior set Beh, the time period set Tsg and the time period label result Tlb, and further organized to form a training set Trm. The special advantage of this process is that the grouping key Key={user identification Uid, video identification Vid, time period label 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 click count Clk, the dwell time Dwt and the complete playback ratio Fin are all directly derived from the original records of the user behavior set Beh, and are accurately selected by the selection rule without any secondary processing or statistics, the numerical deviation caused by repeated calculation or aggregation operation is avoided. Unlike the common behavior features in the prior art, which may cause data distortion or time period misalignment in the aggregation process, the present method uses a direct reference mechanism and a sequential alignment mechanism to accurately reflect the real interactive behavior of users in each time period in the training set Trm in the time period set Tsg dimension, thereby significantly improving the authenticity and stability of the training set Trm in the modeling stage, and ensuring that the subsequent model training can capture the subtle differences of user interest in the time dimension.

[0090] Embodiment 4: Specifically, the S3 includes S31, S32 and S33;

[0091] S31, based on the obtained training set Trn, performing feature subset division and index extraction on the training set Trn, specifically including steps S311 and S312;

[0092] S311, according to the grouping key Key, dividing the training set Trn according to the time period label result Tlb to form a plurality of training subsets, each training subset corresponding to a time period set Tsg one by one, ensuring that the model training data under each time period set Tsg is clear in origin;

[0093] S312, in each training subset, the following three core indexes are extracted:

[0094] Core index one: behavior data amount Dct, obtained by counting the number of user behavior records contained in the training subset, the number of different grouping keys Key, reflecting the user behavior activity level of the training subset in the corresponding time period;

[0095] Core index two: user number Uct, obtained by counting the number of non-repeated user identifiers Uid in the training subset, reflecting the user coverage range of the training subset in the corresponding time period;

[0096] Core index three: time period distribution width Tgd, obtained by calculating the range span of the time period label result Tlb corresponding to the behavior data in the training subset, for example, if the behavior data in a training subset covers three different time period labels of "morning", "commute" and "night" in the time period set Tsg, then the span represents the distribution integrity of the training subset in different time periods.

[0097] S32, after completing the training subset division, an interest model is established for each training subset, and all training subsets are integrated to form an interest model set Mod;

[0098] In the process of establishing the interest model, the number of clicks Clk, the dwell time Dwt and the complete playback ratio Fin contained in the user behavior set Beh are used as input features, the behavior data amount Dct, the user number Uct and the time period distribution width Tgd extracted from the training subset are used as constraint conditions, and the input features and the constraint conditions are used as input data sources of the long short-term memory network LSTM, 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 has sufficient representativeness and stability.

[0099] S33, after the establishment of the interest model set Mod is completed, based on long short-term memory network LSTM, the input features and the constraint conditions are jointly learned, and the performance state of each time period is output, the performance state includes model prediction accuracy Acc, coverage user ratio Cov and fitting degree Fit of target optimization index;

[0100] After obtaining the performance state of the time period, the model prediction accuracy Acc, the coverage user ratio Cov and the fitting degree Fit of the target optimization index of different interest models in the respective time period are compared, and normalized processing is performed, the range is limited to [0, 1] range, the normalized model prediction accuracy Nor(Acc), the normalized coverage user ratio Nor(Cov) and the normalized fitting degree Nor(Fit) of the target optimization index are obtained, and then the geometric mean is calculated to generate the comprehensive performance Per;

[0101] 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;

[0102] 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 the model in the time period can reflect the change of user interest, and the smaller the opposite;

[0103] The coverage user ratio Cov represents the proportion of the number of effective users covered by the interest model in the time period, the higher the coverage user ratio Cov, the more the interest model has explanatory power for more user behavior, and the smaller the opposite;

[0104] The fitting degree Fit of the target optimization index represents the degree of fit between the recommendation result of the interest model in the time period and the preset optimization target (such as click rate improvement, complete play rate improvement or stay time increase), the higher the fitting degree Fit of the target optimization index, the stronger the optimization effect of the model in the target direction, and the smaller the opposite;

[0105] The comprehensive performance Per is obtained by the following calculation formula:

[0106] ;

[0107] In the formula, Indicates the comprehensive performance of the i-th specific time period in the time period set Tsg; Indicates the normalized model prediction accuracy of the i-th specific time period; Indicates the normalized coverage user ratio of the i-th specific time period; Indicates the normalized fitting degree Nor(Fit) of the target optimization index of the i-th specific time period;

[0108] The time period weight Twg of each time period is obtained by the following calculation formula:

[0109] ;

[0110] In the formula, represents the weight of the i-th specific time period in the time period set Tsg, and n represents the total number of the time period set Tsg, represents the comprehensive performance of the j-th specific time period in the time period set Tsg;

[0111] The time period weight set Twg is then corresponded to the interest model set Mod to form the weight switching strategy Wps, which is used to dynamically switch different interest models according to the corresponding time period in the user behavior prediction and recommendation process.

[0112] In the embodiment, the decomposition and index extraction of the training set Trn are completed, and on this basis, the behavior data volume Dct, the number of users Uct and the time period distribution breadth Tgd are introduced as constraint conditions, which ensures that the establishment process of the interest model set Mod can truly reflect the activity, coverage and distribution integrity of the user in different time period sets Tsg. Further, in the joint learning process of the input features (click times Clk, dwell time Dwt and complete playback ratio Fin) and the above constraint conditions, the long short-term memory network LSTM outputs the model prediction accuracy Acc, the coverage user ratio Cov and the fitting degree Fit of the target optimization index, and through normalization and geometric mean, the comprehensive performance Per is generated, which is finally mapped to the weight in the time period weight set Twg. The special advantage of this processing is that the time period weight set Twg not only depends on a single index, but is based on the comprehensive calculation of multi-dimensional performance state, thereby avoiding the risk of excessive dominance of a single index on the model effect. Unlike the common static weight allocation method in the prior art, the present method realizes the flexibility and pertinence of the weight switching strategy Wps in the user behavior prediction and recommendation link through the dynamic normalization and weight generation mechanism of the comprehensive performance Per. In this way, the prediction accuracy, user coverage breadth and target optimization fitting degree can be accurately balanced under different time periods, thereby ensuring the overall stability and balance of the recommendation result Rlt under multi-dimensional targets.

[0113] Embodiment 5: Specifically: S41, when the user request arrives, the time period labeling result Tlb is obtained according to the request time to complete time identification, and the time period labeling result Tlb is 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 in the time period weight set Twg corresponding to the time period labeling result Tlb is the comprehensive performance Per of each interest model in the time period;

[0114] According to the obtained comprehensive performance Per, the interest model set Mod of the time period labeling result Tlb is sorted by weight, and the interest model with the highest comprehensive performance Per is selected as the execution object;

[0115] According to the mapping relationship maintained in the weight switching strategy Wps, the grouping key Key set bound to the interest model of the selected execution object and the time period labeling result Tlb is located;

[0116] 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, the record consistent with the user identifier Uid of the current request is selected, and the corresponding video identifier Vid set is extracted therefrom; the video identifier Vid set is used as a seed to perform similar expansion based on the content feature set Cfe to obtain a candidate content set, and the candidate set is de-duplicated and regularized, and priority-ordered to generate a recommendation result Rlt and return to the user in real time.

[0117] In the real-time scenario of user request arrival in this embodiment, first, time recognition is completed based on the request time to obtain the time period labeling result Tlb, and then the time period labeling result Tlb is 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. Further, the interest model set Mod is weighted and sorted using the comprehensive performance Per of each interest model in the time period weight set Twg, so that in the case of multiple interest models coexisting, 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 deduced, and the video identifier Vid set consistent with the user identifier Uid is further filtered out, and then the video identifier Vid set is used as a seed to perform similar expansion based on the content feature set Cfe to generate a candidate content set. Finally, through deduplication, normalization and priority sorting processing, the recommended result Rlt can be output 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 according to the comprehensive performance Per at the time of the user's actual request, and expand it in combination with the video identifier Vid set directly associated with the user identifier Uid. For example, when the user initiates a request during the lunch break period, the interest model with the highest comprehensive performance Per during the lunch break period can be preferentially selected through the time period labeling result Tlb, the video identifier Vid with which the user watches and completes the playback at a higher rate during the period is deduced, and a candidate video set similar in style, theme and length is expanded accordingly, so that the user's immediate viewing habits and interest points can be more accurately matched, realizing truly "scene-based instant recommendation" and avoiding recommendation results that are out of line with the current period of interest.

[0118] Embodiment 6: Short video real-time dynamic recommendation system based on deep learning, please refer to Figure 2 , in particular: user data acquisition module, time period division and feature extraction module, deep learning and modeling module and matching and pushing module;

[0119] The user data acquisition module acquires the basic behavior data of the user to form a user behavior set Beh, including the number of clicks Clk, the dwell time Dwt and the complete playback ratio Fin; and divides the time stamp into a time period set Tsg;

[0120] The time period division and feature extraction module divides the user behavior set Beh according to the time period set Tsg to obtain a time period feature set Tfe; and forms a training set Trn for modeling;

[0121] The deep learning and modeling module respectively establishes a set of interest models Mod under a set of different time periods Tsg based on the training set Trn, and calculates a corresponding set of time period weights Twg to form a weight switching strategy Wps;

[0122] The matching and pushing module invokes the set of interest models Mod according to the set of time periods Tsg when a user request arrives, and generates a recommendation result Rlt in combination with the set of time period weights Twg.

[0123] Although the embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which 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: Includes the following steps: S1. Collect basic user behavior data to form a user behavior set Beh, including click count Clk, dwell time Dwt, and completion rate Fin; Divide the data into time period sets Tsg based on timestamps; S2. Divide the user behavior set Beh according to the time period set Tsg to obtain the time period feature set Tfe; and form a training set Trn for modeling. S2 includes S21 and S22; S21. Based on the obtained user behavior set Beh, time period set Tsg, and time period labeling result Tlb, the user behavior set Beh is grouped according to the time period set Tsg to generate a structured time period feature set Tfe. The time period feature set Tfe is obtained through 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 for the same user, the same video, and the same time period in the user behavior set Beh; S212. For each grouping key, directly reference the three basic metrics obtained in step S1 from the located atomic behavior record, including click count Clk, dwell time Dwt, and completion rate Fin, and write them into the time period feature set Tfe; when there are multiple atomic records, select only one for reference according to the selection rule: Selection Rule 1: Select the record with the most recent timestamp; Rule 2: If the timestamps are the same, the log entry written first will be selected based on the log writing order. S213. Using the grouping key Key={user identifier Uid, video identifier Vid, time period labeling result Tlb} as the unique index, and the fields click count Clk, dwell time Dwt, and completion ratio Fin are all directly referenced from the selected atomic records of the user behavior set Beh, forming a feature record. After integrating all feature records, a structured time period feature set Tfe is obtained. 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 set Twg to form a weight switching strategy Wps; S3 includes S31, S32 and S33; S31. Based on the acquired training set Trn, perform feature subset partitioning and index extraction on the training set Trn, specifically including steps S311 and S312. S311. Based on the grouping key Key, divide the training set Trn into multiple training subsets according to the time period labeling results Tlb, and each training subset corresponds one-to-one with the time period set Tsg. S312. In each training subset, extract the following three core metrics: Key Metric 1: Behavioral 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 activity level of user behavior in the training subset during the corresponding time period. Core metric 2: User count Uct, obtained by counting the number of unique user identifiers Uid in the training subset, reflects the user coverage of the training subset in the corresponding time period; Core metric 3: Time period distribution breadth (Tgd), which is obtained by calculating the range of the time period labeling results (Tlb) corresponding to the behavioral data in the training subset; S32. After completing the training subset division, establish an interest model for each training subset and integrate all training subsets to form an interest model set Mod. In the process of building the interest model, the number of clicks Clk, dwell time Dwt, and completion rate 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. S33. After completing the establishment of the interest model set Mod, 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 proportion of covered users Cov, and the fit with the target optimization index Fit. After obtaining the performance status of different interest models in their respective time periods, the results of model prediction accuracy Acc, coverage ratio Cov, and fit of the target optimization index Fit are compared. The results are then normalized and limited to the range of [0,1] to obtain the normalized model prediction accuracy Nor (Acc), normalized coverage ratio Nor (Cov), and normalized fit of the target optimization index Nor (Fit). The results are then calculated using geometric mean to generate the comprehensive performance Per. The overall performance Per of each time period is mapped to the weight in the time period weight set Twg, thus obtaining the time period weight Twg for each time period; Then, the time period weight set Twg is mapped to 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 user behavior prediction and recommendation. S4. When a user request arrives, the interest model set Mod is called based on the time period set Tsg, and the recommendation result Rlt is generated by combining the time period weight set Twg.

2. The real-time dynamic recommendation method for short videos based on deep learning according to claim 1, characterized in that: S1 includes S11 and S12; S11. When a user uses a short video platform, extract the user's unique identifier and record it as the user identifier Uid. Monitor and record the multi-dimensional behavioral signals generated when interacting with short video content to form a user behavior set Beh. The user behavior set Beh includes the number of clicks Clk, the dwell time Dwt, the completion rate Fin, the user identifier Uid, and the video identifier Vid; The click count Clk is recorded in real time by the user's clicks on video covers and play buttons triggered in the recommendation list and playback page; each click event is written to the interaction event log and is marked with a unique video identifier, which is called video identifier Vid; Then, using intraday time windows as units, the logs are cumulatively calculated to obtain the number of clicks Clk by the user in the corresponding window; The dwell time Dwt is recorded during player operation by recording the timestamps of the video entry and exit times respectively, and the dwell time for a single viewing is obtained by calculating the difference. The short video platform's backend log service writes the dwell time into the behavior record table and stores it using video identifiers and user identifiers as indexes. When merging multiple viewing records of the same user, the user's dwell time on the video is generated as an average value, Dwt. The completion rate 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 duration, it is considered a completion event. The proportion of completion events among the playback counts of the same user is counted and marked as the user's completion rate Fin on the video.

3. The real-time dynamic recommendation method for short videos based on deep learning according to claim 2, characterized in that: S12. After completing the collection of user behavior set Beh, the timestamp information corresponding to each behavior data is obtained. The timestamp information is automatically written by the front-end interaction log and playback log. Each log record contains the user identifier, video identifier and the time when the behavior occurred. The timestamp data is parsed and classified to form a time period set Tsg and a time period labeling result Tlb corresponding to each behavioral data. The specific process includes timestamp standardization processing and time period division and labeling generation. The timestamp standardization process converts timestamps from logs from different sources into a standard time format and removes time zone differences. The time period division and label generation are based on predefined division rules. A day is divided into multiple fixed time periods, each fixed time period represents a time period label, forming a time period set Tsg. Then, the timestamps that have undergone timestamp standardization are mapped to the corresponding time period labels, and a corresponding time period labeling result Tlb is generated for each piece of behavioral data.

4. The real-time dynamic recommendation method for short videos based on deep learning according to claim 1, characterized in that: S22. Based on the obtained 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, dwell time Dwt, and completion ratio Fin are aligned and organized according to the order of the time period set Tsg to form the training set Trm.

5. The real-time dynamic recommendation method for short videos based on deep learning according to claim 1, characterized in that: S4 includes S41; S41. When a user request arrives, time identification is performed based on the request time to obtain the time period labeling result Tlb. Then, the time period labeling result Tlb is indexed in the weight switching strategy Wps to obtain the binding relationship between the interest model set Mod and the time period weight set Twg corresponding to the time period labeling result Tlb. Among them, the weight values ​​of each item in the time period weight set Twg corresponding to the time period labeling result Tlb are the comprehensive performance Per of each interest model in that time period. Based on the obtained comprehensive performance Per, the interest model set Mod of the time period labeling result Tlb is sorted by weight, and the interest model ranked first in the comprehensive performance Per is selected as the execution object. Based on the mapping relationship maintained in the weight switching strategy Wps, locate the set of grouping keys Key that are 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 user identifier Uid, video identifier Vid, and time period labeling result Tlb. In the grouping key Key set, the record that matches the user identifier Uid of the current request is selected, and the corresponding video identifier Vid set is extracted from it. Using the video identifier Vid set as a seed, similarity expansion is performed based on the content feature set Cfe to obtain a candidate content set. Then, the candidate set is deduplicated, normalized, and prioritized to generate the recommendation result Rlt, which is returned and pushed to the user in real time.

6. A deep learning-based real-time dynamic recommendation system for short videos, applied to the deep learning-based real-time dynamic recommendation method for short videos as described in any one of claims 1 to 5, characterized in that: The module includes a user data collection module, a time period segmentation and feature extraction module, a deep learning and modeling module, and a matching and push module. The user data collection module collects basic user behavior data to form a user behavior set Beh, which includes the number of clicks Clk, the dwell time Dwt, and the completion rate Fin. Divide the data into time period sets Tsg based on timestamps; The time period segmentation and feature extraction module divides the user behavior set Beh according to 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 set Twg to form a weight switching strategy Wps; When a user request arrives, the matching and push module calls the interest model set Mod based on the time period set Tsg, and combines it with the time period weight set Twg to generate the recommendation result Rlt.

Images