A user development analysis method for video recommendation

Abstract

The application discloses a user development analysis method for video recommendation, which comprises the following steps: data collection, user development cycle stage transition, generation of stage transition graph, video abnormal recommendation detection and report generation. The application belongs to the technical field of user analysis, and specifically relates to a user development analysis method for video recommendation. In the scheme, a basic matrix decomposition model is used to learn the basic interest of a user and the core features of a video, a state vector is used to perceive the current state of the user, a short-term preference score is combined with a Q function for measuring long-term value to make a recommendation from candidate videos, a reward function is calculated to quantize the long-term value of the recommendation action to the user development cycle transition, a sliding window is used to generate a time sequence, an encoder is used to extract the potential feature distribution in the window to capture the system health mode, a weighted cumulative reconstruction error is calculated as an abnormal score function, a dynamic threshold is set, and accurate perception of the abnormal state of the recommendation system is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of user analysis, and specifically relates to a user development analysis method for video recommendation. BACKGROUND

[0002] The core of the user development analysis method for video recommendation is to realize the dual improvement of user scale and quality by optimizing the performance of the recommendation system at each development cycle stage. However, the existing user development analysis method for video recommendation has the technical problems that the traditional video recommendation method ignores the conversion of users, cannot bring sustainable user growth to the platform, can bring maximum short-term indicators, but damages the long-term health of the platform; the traditional recommendation system lacks fine-grained perception ability, leading to hidden faults difficult to find, cannot identify the slow degradation of the recommendation quality of the recommendation system, the rule threshold alarm is sensitive to traffic fluctuations and hot events, produces a large number of invalid warnings to submerge real abnormal signals, lacks potential pattern analysis ability when abnormality occurs, and has difficulty in root cause positioning. SUMMARY

[0003] In view of the above situation, in order to overcome the defects of the prior art, the application provides a user development analysis method for video recommendation, which is used to solve the technical problems that the traditional video recommendation method ignores the conversion of users, cannot bring sustainable user growth to the platform, can bring maximum short-term indicators, but damages the long-term health of the platform, adopts a basic matrix decomposition model to learn the basic interests of users and the core features of videos, makes recommendations from candidate videos by combining a state vector that perceives the current state of users with a short-term preference score and a Q function that measures long-term value, and calculates a reward function to quantify the long-term value of the recommendation action to the user development cycle transition; and the technical problems that the traditional recommendation system lacks fine-grained perception ability, leading to hidden faults difficult to find, cannot identify the slow degradation of the recommendation quality of the recommendation system, the rule threshold alarm is sensitive to traffic fluctuations and hot events, produces a large number of invalid warnings to submerge real abnormal signals, lacks potential pattern analysis ability when abnormality occurs, and has difficulty in root cause positioning, adopts a sliding window to generate a time series, uses an encoder to extract the potential feature distribution in the window to capture the system health pattern, predicts the future potential state through a self-attention model, calculates a weighted cumulative reconstruction error as an abnormal scoring function, sets a dynamic threshold, and realizes accurate perception of the abnormal state of the recommendation system.

[0004] The technical scheme adopted by the application is as follows: the application provides a user development analysis method for video recommendation, which comprises the following steps:

[0005] Step S1: data acquisition;

[0006] Step S2: user development cycle stage transition;

[0007] Step S3: generating a stage transition map;

[0008] Step S4: video anomaly recommendation detection;

[0009] Step S5: generating a report.

[0010] Further, in step S1, the data collection, specifically collecting user behavior data, the user behavior data including user behavior data, user attribute data and video content data, data cleaning and feature extraction are performed on the user behavior data, and a user matrix and a user feature vector are constructed;

[0011] The user behavior data includes video viewing time, video completion rate, click rate, collection, like, sharing, login frequency, and last login time.

[0012] The user attribute data includes age, region, package type and terminal device.

[0013] The video content data includes video ID, video tag and video category.

[0014] Further, in step S2, the user development cycle stage transition includes the following steps:

[0015] Step S21: defining a user development cycle, clearly dividing the user development cycle stage according to the user behavior data, the user development cycle including an introduction period, a growth period, a mature period, a loss period and a dormant period;

[0016] Step S22: training a basic matrix decomposition model for obtaining a static preference representation of users and videos, learning the basic interest of users and the core features of videos from historical data, defining a prediction score representing the short-term preference of users, and the formula is as follows:

[0017] ;

[0018] ;

[0019] In the formula, Q1 represents a target function, represents the sum of all known user and video interaction pairs, K represents a training data set, r ui represents the real feedback of user u to video i, p u represents the latent feature vector of user u, q i represents the latent feature vector of video, represents the preference score of user u to video i predicted by the basic model, i.e. the prediction score of the short-term preference of users, μ represents a weight decay term, and T represents a transpose operation.

[0020] Step S23: Construct a state vector for perceiving the user's current stage in the development cycle, using the following formula:

[0021] ;

[0022] where s t represents the state vector at time t, H(L t ) represents the one-hot encoding of the user's current stage in the development cycle, f t represents the user feature vector at time t, c t represents the context feature vector at time t, which describes the environmental situation at the time of the video recommendation request;

[0023] Step S24: Action execution;

[0024] Recall and pre-sorting, based on the user's latent vector, using approximate nearest neighbor search and basic matrix decomposition model scoring, to generate a refined action space, using the following formula:

[0025] ;

[0026] where A t represents the action space at time t, a i represents the candidate video, R t represents the recall set at time t, which is a set of candidate videos selected from the recall stage in the millions of full video library, represents the sorting selection function, which is used to sort and truncate the videos in the recall set according to the basic preference score;

[0027] Select the video with the highest final score for recommendation, output the recommended action sequence, using the following formula:

[0028] ;

[0029] where a t represents the video that the system finally decides to recommend, argmax represents the maximum parameter solver, represents the action value function, θ represents the parameters of the Q network, which is used to measure whether the long-term value of the recommended action is beneficial to the user's optimization to a better development cycle stage;

[0030] Step S25: Calculate the reward function to quantify the long-term value of the recommended action, observe the user's immediate feedback on the video recommendation, calculate the immediate reward and stage transition reward, and store the experience tuples back to the buffer, using the following formula:

[0031] ;

[0032] ;

[0033] wherein r t represents the total reward obtained after performing the action, represents the immediate reward, i.e. the immediate feedback of the user to the recommended video, β1 represents the weight coefficient of the transition reward, M represents the stage transition reward matrix, L t represents the user development cycle stage before recommendation, L t+1 represents the user development cycle stage after recommendation, D represents the experience replay buffer, represents the updated experience replay buffer, represents the assignment operation, represents the set union operation, s t+1 represents the next state to which the environment transitions after performing the action, represents the experience tuple;

[0034] Step S26: updating the Q network, by continuously recommending videos and collecting user feedback, using gradient descent to update the parameters θ to minimize the loss function of the model, the formula used is as follows:

[0035] ;

[0036] wherein, represents the loss function of the reinforcement learning model, by adjusting the parameters to minimize the loss function, N represents the total number of experience tuples randomly sampled from the experience replay buffer, j represents the index of the experience tuple, represents the target Q value, which is used to represent a better cumulative return estimate for sample j.

[0037] Further, in step S3, the generation of the stage transition map, specifically, the system automatically measures the actual development cycle stage transition after each recommendation, and dynamically generates a stage transition map based on the transition process. The stage transition map is highly dependent on the stability and health of the recommendation system. If the system is abnormal, not only can it not achieve the goal of user growth, but it can also accelerate user loss.

[0038] Further, in step S4, the video abnormal recommendation detection includes the following steps:

[0039] Step S41: video recommendation index monitoring, to ensure the reliable execution of the user development cycle stage transition, the running state of the recommendation system is monitored in real time. The core indicators of the video recommendation system mainly include recommendation exposure click rate, user watching completion rate, and negative feedback rate.

[0040] Step S42: Generate a sliding window, input the original time series data of the video recommendation index, and generate a continuous sliding window sequence, the formula is as follows:

[0041] ;

[0042] In the formula, W t represents the recommendation index data segment in the time window, x t represents the index value at time t, L represents the window length, S represents the sliding step, and N represents the total length of the index sequence;

[0043] Step S43: Learn the latent features of the recommendation model, map the window to a low-dimensional latent space through an encoder, output the feature distribution parameter mean and standard deviation of the recommendation pattern in the window, and generate latent features z t for capturing the abstract pattern of the healthy operation of the recommendation system in the window;

[0044] Step S44: Generate a latent sequence, concatenate the latent feature representations of all windows in time sequence to form a latent representation sequence , input the latent representation sequence into a multi-head self-attention mechanism model, learn the context information, and predict the latent representation of the next time window;

[0045] Step S45: Calculate the anomaly score function, for each window, calculate the cumulative reconstruction error of the predicted value and the true value to obtain the anomaly score function, avoid misjudgment caused by instantaneous noise fluctuation of a single hot video, and introduce an exponential decay weight, the formula is as follows:

[0046] ;

[0047] In the formula, d t represents the anomaly score function at time t, which is used to measure the degree of video recommendation anomaly of the current window, represents the predicted value of the model for the window at time t-v, represents the true value of the window at time t-v, s represents the number of backtracking windows, v represents the step number sequence when calculating the cumulative error, and h represents the decay factor, which is used to give higher weight to recent errors;

[0048] Step S46: Set a dynamic threshold, when the anomaly score function exceeds the threshold, mark the window as a recommendation anomaly window, the formula is as follows:

[0049] ;

[0050] In the formula, represents the dynamic threshold, when the value of the anomaly score function is greater than the dynamic threshold, the current window is marked as a recommendation anomaly window, g represents the initial high quantile, and Ng a number of windows exceeding the initial high quantile, a scale parameter of the GPD, q represents a confidence level, a shape parameter of the GPD, n represents an empirical value.

[0051] Further, in step S5, the report is generated, specifically, the health condition of the recommended system in the detection period is summarized, each labeled abnormal window is described in detail, the reason for the abnormality is explained, and whether the abnormal recommendation interrupts the positive development cycle stage transition process of the user is judged.

[0052] The beneficial results achieved by the present application using the above scheme are as follows:

[0053] (1) In view of the technical problem that the traditional video recommendation method ignores the conversion of users, cannot bring sustained user growth to the platform, although it can bring maximum short-term indicators, but damages the long-term health of the platform, the basic matrix decomposition model is used to learn the basic interest of the user and the core features of the video, the state vector is used to perceive the current state of the user, the short-term preference score is combined with the Q function for measuring long-term value to make recommendations from the candidate video, and the reward function is calculated to quantify the long-term value of the recommended action to the user development cycle transition;

[0054] (2) In view of the technical problem that the traditional recommendation system lacks fine-grained perception ability, leading to difficulty in discovering implicit faults, being unable to identify the slow degradation of the recommendation quality of the recommendation system, the rule threshold alarm being sensitive to traffic fluctuations and hot events, generating a large number of invalid warnings to submerge real abnormal signals, and lacking potential pattern analysis ability when the abnormality occurs, making it difficult to locate the root cause, the sliding window is used to generate time series, the encoder is used to extract the potential feature distribution in the window to capture the system health pattern, the self-attention model is used to predict the future potential state, the weighted cumulative reconstruction error is calculated as an abnormal scoring function, and a dynamic threshold is set to realize accurate perception of the abnormal state of the recommendation system. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 A flowchart of a user development analysis method for video recommendation provided by the present application;

[0056] Figure 2 A flowchart of step S2;

[0057] Figure 3 A flowchart of step S4.

[0058] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0060] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0061] Embodiment one, refer to Figure 1 The present application provides a user development analysis method for video recommendation, which comprises the following steps:

[0062] Step S1: data acquisition, collecting user behavior data;

[0063] Step S2: user development cycle stage transfer, defining the user development cycle, learning the basic interests of the user and the core features of the video by using the basic matrix decomposition model, making recommendations from the candidate videos by combining the short-term preference score with the Q function measuring the long-term value through the state vector perception, calculating the reward function to quantify the long-term value of the recommended action sequence to the user development cycle transfer, and outputting the recommended action sequence;

[0064] Step S3: generating a stage transfer map for reflecting the user stage transfer driven by the recommendation effect;

[0065] Step S4: video abnormal recommendation detection, monitoring video recommendation indicators, generating time series by using a sliding window, extracting potential feature distribution within the window by using an encoder to capture the system health pattern, predicting future potential state by using a self-attention model, calculating weighted cumulative reconstruction error as an abnormal score function, setting a dynamic threshold, and realizing accurate perception of the abnormal state of the recommendation system;

[0066] Step S5: generating a report, analyzing the reasons for video recommendation abnormalities and judging the influence on the positive development cycle stage transfer of the user.

[0067] Embodiment two, refer to Figure 1The embodiment is based on the above embodiment, and in step S1, the data collection, specifically collecting user behavior data, the user behavior data including user behavior data, user attribute data and video content data, data cleaning and feature extraction are performed on the user behavior data, a user matrix and a user feature vector are constructed;

[0068] The user behavior data includes video watching time, video completion rate, click rate, collection, like, sharing, login frequency and last login time.

[0069] The user attribute data includes age, region, package type and terminal device.

[0070] The video content data includes video ID, video tag and video category.

[0071] Embodiment three, refer to Figure 1 and Figure 2 The embodiment is based on the above embodiment, and in step S2, the user development cycle stage transition includes the following steps:

[0072] Step S21: defining the user development cycle, the user development cycle stage is determined according to the user behavior data, and the user development cycle includes the introduction period, the growth period, the mature period, the loss period and the dormancy period;

[0073] Step S22: training a basic matrix decomposition model for obtaining a static preference representation of users and videos, learning the basic interest of users and the core features of videos from historical data, defining a prediction score representing the short-term preference of users, and the formula is as follows:

[0074] ;

[0075] ;

[0076] In the formula, Q1 represents the target function, represents the summation of all known user and video interaction pairs, K represents the training data set, r ui represents the real feedback of user u to video i, p u represents the latent feature vector of user u, q i represents the latent feature vector of video, represents the preference score of user u to video i predicted by the basic model, i.e. the prediction score of the short-term preference of users, μ represents the weight decay term, and T represents the transpose operation.

[0077] Step S23: constructing a state vector for sensing the development cycle stage in which the user is currently located, and the formula is as follows:

[0078] ;

[0079] where s t denotes the state vector at time t, H(L t ) denotes the one-hot encoding of the current stage of the development cycle the user is in, f t denotes the user feature vector at time t, c t denotes the context feature vector at time t, which describes the environmental context at the time of the video recommendation request;

[0080] Step S24: Action execution;

[0081] Recall and pre-ranking, based on the user latent vector, using approximate nearest neighbor search and a base matrix factorization model to score, generating a refined action space, the formula used is as follows:

[0082] ;

[0083] where A t denotes the action space at time t, a i denotes the candidate video, R t denotes the recall set at time t, from the recall stage, a set of candidate videos is selected from the millions of full video library, denotes the ranking selection function, which is used to sort and truncate the selection of videos in the recall set according to the base preference score;

[0084] Select the video with the highest final score for recommendation, output the recommended action sequence, the formula used is as follows:

[0085] ;

[0086] where a t denotes the video that the system finally decides to recommend, argmax denotes the maximum parameter solver, denotes the action value function, and θ denotes the parameters of the Q network, which is used to measure whether the long-term value of the recommended action is beneficial to the user optimizing to a better development cycle stage;

[0087] Step S25: Calculate the reward function, which is used to quantify the long-term value of the recommended action, observe the immediate feedback of the user to the video recommendation, calculate the immediate reward and stage transition reward, and store the experience tuple back to the buffer, the formula used is as follows:

[0088] ;

[0089] ;

[0090] where r t denotes the total reward obtained after executing the action, represents the immediate reward, i.e., the immediate feedback of the user to the recommended video, β1represents the weight coefficient of the transition reward, M represents the stage transition reward matrix, L t represents the user development cycle stage before recommendation, L t+1 represents the user development cycle stage after recommendation, D represents the experience replay buffer, represents the updated experience replay buffer, represents the assignment operation, represents the set union operation, s t+1 represents the next state to which the environment transitions after the execution of the action, represents the experience tuple;

[0091] Step S26: updating the Q network, by continuously recommending videos and collecting user feedback, using gradient descent to update the parameters θ to minimize the loss function of the model, the formula is as follows:

[0092] ;

[0093] In the formula, represents the loss function of the reinforcement learning model, by adjusting the parameters to minimize the loss function, N represents the total number of experience tuples randomly sampled from the experience replay buffer, j represents the index of the experience tuple, represents the target Q value, which is used to represent a better cumulative return estimate for sample j.

[0094] By performing the above operations, the basic matrix decomposition model is used to learn the basic interests of the user and the core features of the video, the state vector is used to perceive the current state of the user, the short-term preference score is combined with the Q function that measures the long-term value to make recommendations from the candidate videos, the reward function is used to quantify the long-term value of the user development cycle transition caused by the recommended action, and the technical problem that the traditional video recommendation method ignores the user conversion and cannot bring sustainable user growth to the platform is solved. Although it can maximize the short-term indicators, it damages the long-term health of the platform.

[0095] Embodiment four, see Figure 1 This embodiment is based on the above embodiment, in step S3, the generation of the stage transition map, specifically, the system automatically measures the actual development cycle stage transition of the user after each recommendation, and dynamically generates a stage transition map based on the transition process. The stage transition map is highly dependent on the stability and health of the recommendation system itself. If the system is abnormal, not only can it not achieve the goal of user growth, but it will also accelerate the loss of users.

[0096] Embodiment five, see Figure 1 and Figure 3 This embodiment is based on the above embodiment, in step S4, the video abnormal recommendation detection, including the following steps:

[0097] Step S41: Video recommendation index monitoring. In order to ensure the reliable execution of the user development cycle stage transition, the running state of the recommendation system is monitored in real time, and the core indicators of the video recommendation system mainly include the recommendation exposure click rate, the user watching completion rate and the negative feedback rate.

[0098] Step S42: Generating a sliding window. The original time series data of the video recommendation index is input, and a continuous sliding window sequence is generated. The formula is as follows:

[0099] ;

[0100] In the formula, W t represents the recommended index data segment in the time window, x t represents the index value at time t, L represents the window length, S represents the sliding step, and N represents the total length of the index sequence.

[0101] Step S43: Learning the latent features of the recommendation model. The window is mapped to a low-dimensional latent space through an encoder, and the feature distribution parameter mean and standard deviation of the recommendation mode in the window are output, generating latent features z t , which are used to capture the abstract mode of the healthy operation of the recommendation system.

[0102] Step S44: Generating a latent sequence. The latent feature representations of all windows are spliced into a latent representation sequence in time sequence. The latent representation sequence is input into a multi-head self-attention mechanism model, and the latent representation of the next time window is predicted by learning the context information.

[0103] Step S45: Calculating the anomaly score function. For each window, the cumulative reconstruction error of the predicted value and the true value is calculated to obtain the anomaly score function, which avoids the misjudgment caused by the instantaneous noise fluctuation of a single hot video. An exponential decay weight is introduced, and the formula is as follows:

[0104] ;

[0105] In the formula, d t represents the anomaly score function at time t, which is used to measure the video recommendation anomaly degree of the current window, represents the predicted value of the model for the window at time t-v, represents the true value of the window at time t-v, s represents the backtracking window number, v represents the step number sequence when calculating the cumulative error, and h represents the decay factor, which is used to give higher weight to the recent error.

[0106] Step S46: set a dynamic threshold, when the anomaly score function exceeds the threshold, mark the window as a recommended abnormal window, the formula is as follows:

[0107] ;

[0108] wherein, indicates a dynamic threshold, when the value of the anomaly score function is greater than the dynamic threshold, the current window is marked as a recommended abnormal window, g indicates an initial high quantile, N g indicates the number of windows exceeding the initial high quantile, indicates the scale parameter of the GPD, q indicates the confidence level, indicates the shape parameter of the GPD, n indicates the empirical value.

[0109] By performing the above operation, a time series is generated by using a sliding window, a latent feature distribution in the window is extracted by using an encoder to capture the system health pattern, a future latent state is predicted by using a self-attention model, a weighted cumulative reconstruction error is calculated as an anomaly score function, a dynamic threshold is set, precise perception of a recommended system abnormal state is realized, and the technical problems that implicit faults are difficult to find due to lack of fine-grained perception ability of a traditional recommended system, slow degradation of recommended quality of the recommended system cannot be identified, a rule threshold alarm is sensitive to traffic fluctuations and hot events, a large number of invalid warnings are generated to submerge real abnormal signals, and potential pattern analysis ability is lacked when an anomaly occurs, and root cause positioning is difficult are solved.

[0110] Embodiment six, refer to Figure 1 This embodiment is based on the above-mentioned embodiments, in step S5, the report is generated, specifically, the health status of the recommended system in the detection period is summarized, each labeled abnormal window is described in detail, the cause of the anomaly is explained, and it is judged whether the abnormal recommendation interrupts the user's positive development cycle phase transition process.

[0111] It should be noted that, in this document, relational terms such as first and second and the like can merely be used to distinguish one entity or action from another, without necessarily requiring or implying that the entities or actions are in any way mutually exclusive or in any way directly or indirectly related to each other. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0112] Although the embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made without departing from the principles and spirit of the present application.

[0113] The above describes the present application and its embodiments, which are not limited, and the drawings only show one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired thereby, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution are not creative, and should belong to the protection scope of the present application.

Claims

1. A user development analysis method for video recommendation, characterized in that: The method comprises the following steps: Step S1: data collection, collecting user behavior data; Step S2: user development cycle stage transition, defining the user development cycle according to the user behavior data, learning the basic interests of the user and the core features of the video by using a basic matrix decomposition model, making recommendations from candidate videos by combining the predicted score representing the user's short-term preference with the Q function measuring the long-term value through the state vector, calculating the reward function to quantify the long-term value of the recommended action to the user development cycle transition, and outputting the recommended action sequence; Step S3: generating a stage transition atlas, generating a transition atlas according to the user development cycle stage transition path; Step S4: video abnormal recommendation detection, monitoring video recommendation indicators, generating a time series using a sliding window, extracting the latent feature distribution within the window using an encoder to capture the system health pattern, predicting the future latent state through a self-attention model, calculating the weighted cumulative reconstruction error as an anomaly scoring function, and setting a dynamic threshold; Step S5: generating a report, analyzing the reasons for video recommendation anomalies and judging the impact on the user's positive development cycle stage transition; In step S2, the user development cycle stage transition comprises the following steps: Step S21: defining the user development cycle, clearly dividing the user development cycle stages according to the user behavior data; Step S22: training a basic matrix decomposition model, defining a predicted score representing the user's short-term preference, and using the following formula: ； ； where Q1represents the objective function, represents all known user and video interaction pairs, K represents the training dataset, r ui represents the real feedback of user u to video i, p u represents the user latent vector, q i represents the latent feature vector of the video, represents the predicted score of the user short-term preference, μ represents the weight decay term, and T represents the transpose operation; Step S23: constructing a state vector, using the following formula: ； where s t represents the state vector at time t, H(L t ) represents the one-hot encoding of the development stage the user is currently in, f t represents the user feature vector at time t, c t represents the context feature vector at time t; Step S24: action execution; Recall and pre-sorting, based on the user latent vector, using approximate nearest neighbor search and basic matrix decomposition model scoring to generate an action space, using the following formula: ； In the formula, A t denotes the action space at time t, a i denotes the candidate video, R t denotes the candidate video set filtered by the recall link at time t, denotes the ranking filtering function, which is used to sort and select the videos in the recall set according to the predicted score of the user's short-term preference; Select the video with the highest final score for recommendation, output the recommended action sequence, and use the following formula: ； In the formula, a t argmax represents the final decision of the system to select the recommended video, and argmax represents the maximum parameter solver, The action value function is represented by θ, which represents the parameters of the Q network, and is used to measure the long-term value of the recommended action. Step S25: calculating the reward function and storing the experience tuples back to the buffer, using the following formula: ； ； where r t represents the total reward obtained after performing an action, represents the immediate reward, β1represents a weight coefficient of the transition reward, M represents a stage transition reward matrix, L t represents the user development cycle stage before recommendation, L t+1 represents the user development cycle stage after recommendation, D represents an experience replay buffer, represents the updated experience replay buffer, represents an assignment operation, represents a set union operation, s t+1 represents the next state to which the environment transitions after performing an action, represents an experience tuple; Step S26: updating the Q network, continuously updating the parameters θ using gradient descent to minimize the loss function of the model through continuous video recommendation and collection of user feedback, using the following formula: ； wherein represents a loss function of the reinforcement learning model, the loss function is minimized by adjusting parameters , N represents a total number of experience tuples randomly sampled from the experience replay buffer, j represents an index of the experience tuple, represents a target Q value.

2. The user development analysis method for video recommendation according to claim 1, wherein: In step S4, the video abnormal recommendation detection comprises the following steps: Step S41: video recommendation indicator monitoring, real-time monitoring of the running state of the recommendation system, monitoring the core indicators of the video recommendation system including recommendation exposure click rate, user watch completion rate, and negative feedback rate; Step S42: generating a sliding window, inputting the original time series data of the video recommendation indicators to generate a continuous sliding window sequence, using the following formula: ； In the formula, W t represents the recommended indicator data segment within the time window, x t represents the indicator value at time t, L represents the window length, and N represents the total length of the indicator sequence. Step S43: learn the latent feature of the recommendation model, map the window to a low-dimensional latent space through the encoder, output the feature distribution parameter mean and standard deviation of the recommendation pattern in the window, and generate the latent feature z t , for capturing the abstract pattern of the healthy operation of the recommendation system; Step S44: generating a potential sequence, splicing the potential feature representations of all windows into a potential representation sequence in chronological order inputting the potential representation sequence into a multi-head self-attention mechanism model, learning context information, and predicting a potential representation of a next time window; Step S45: calculating the anomaly scoring function, for each window, calculating the cumulative reconstruction error of the predicted value and the true value to obtain the anomaly scoring function, introducing an exponential decay weight, using the following formula: ； In the formula, d t represents an anomaly score function at time t, used to measure the degree of video recommendation anomaly of the current window, represents a prediction value of the model for the window at time t-v, represents a true value of the window at time t-v, s represents a number of backtracking windows, v represents a step number sequence number traversed when calculating the cumulative error, and h represents a decay factor, used to give higher weight to recent errors. Step S46: setting a dynamic threshold, when the anomaly scoring function exceeds the threshold, marking the window as a recommendation anomaly window, using the following formula: ； wherein, represents a dynamic threshold, when the value of the anomaly score function is greater than the dynamic threshold, the current window is marked as a recommended anomaly window, g represents an initial high quantile, N g represents the number of windows that exceed the initial high quantile, represents a scale parameter of the GPD, q represents a confidence level, represents a shape parameter of the GPD, n represents an experience value.

3. The user development analysis method for video recommendation of claim 1, wherein: In step S1, the data collection specifically collects user behavior data, and the user behavior data includes user video watching data, user attribute data, and video content data. The user behavior data is subjected to data cleaning and feature extraction.

4. The user development analysis method for video recommendation of claim 1, wherein: In step S3, the generation phase transition map, specifically, the system automatically measures the actual development cycle phase transition of the user after each recommendation, and dynamically generates a phase transition map based on the transition process.

5. The user development analysis method for video recommendation of claim 1, wherein: In step S5, the generation report, specifically, summarizes the health status of the recommendation system within the detection cycle, describes each marked abnormal window in detail, explains the cause of the abnormality, and judges whether the abnormal recommendation interrupts the positive development cycle phase transition process of the user.

Images