Online live broadcast behavior analysis method and system based on neuro-hox process

Abstract

The invention discloses an on-line live broadcast behavior analysis method and system based on a neuro-hox process, and relates to the technical field of neural networks, and the method comprises the steps: obtaining on-line live broadcast data, setting fixed time periods, segmenting the on-line live broadcast data based on each fixed time period, obtaining an event corresponding to each fixed time period, and storing the event corresponding to each fixed time period; generating an interaction event behavior sequence based on the event corresponding to each fixed time period; generating an online coding structure based on the intensity function of each event type of the interaction event behavior sequence, obtaining a watching behavior sequence based on the interaction event behavior sequence, inputting the watching behavior sequence into the online coding structure, and obtaining a predicted next behavior type, and predicting the corresponding watching behavior occurrence time based on the predicted next behavior type.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, and in particular to an online live broadcast behavior analysis method and system based on a neural Hawkes process. Background Art

[0002] In today's wave of e-commerce, online live streaming sales have become a popular marketing method. Although live streaming can attract a large number of viewers, the actual purchase rate of viewers is not always high, so it is necessary to deeply explore consumers' viewing behavior. Summary of the invention

[0003] In order to solve the deficiencies mentioned in the above background technology, the purpose of the present invention is to provide an online live broadcast behavior analysis method and system based on neural Hawkes process.

[0004] In a first aspect, the purpose of the present invention can be achieved by the following technical solution: a method for analyzing online live broadcast behavior based on a neural Hawkes process, the method comprising the following steps:

[0005] Acquire online live broadcast data, set fixed time periods, segment the online live broadcast data based on each fixed time period, obtain events corresponding to each fixed time period, and generate interactive event behavior sequences based on the events corresponding to each fixed time period;

[0006] An online coding structure is generated based on the intensity function of each event type of the interactive event behavior sequence, and a viewing behavior sequence is obtained based on the interactive event behavior sequence. The viewing behavior sequence is input into the online coding structure to obtain the predicted next behavior type, and the occurrence time of the corresponding viewing behavior is predicted based on the predicted next behavior type.

[0007] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the process of segmenting the online live broadcast data based on each fixed time period:

[0008] Given a fixed time period, by dividing the time period into non-overlapping time periods, observing the events that occur in each time period, the events belong to the category in, Indicates that the event did not occur, k indicates the event category, K indicates the total number of event categories, and the set of interaction events is Indicates time period A sequence of events occurs, where an event is represented by k represents the category of the event, t i Indicates the exact time point when the event occurs, [t i-1 ,t i ]The division of time periods is determined by the time when the event occurs. After the event occurs, the time periods are automatically divided.

[0009] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the process of generating an interactive event behavior sequence based on the event corresponding to each fixed time period:

[0010] The generative probability model of event sequences is called a multivariate point process, and the category At time O ij occurs, in an infinitesimal time interval [t i ,t i +dt) in The probability of occurrence depends on historical information The probability of occurrence depends on the mutual information z'(l1,l2,t i ), and the mutual information z'(l1,l2,t i ) is composed as follows:

[0011] z′(l1,l2,t i )=R(γ(O ij ),f E (l u ,l w ,t i ),f l (l u ,t i ′),f l (l w ,t i ′),z(l u ,t i ′),z(l w ,t i ′))

[0012] Among them, γ(t i ) calculates the function of the time-varying representation vector of the node, and transforms the interaction time t i The function that is converted into the corresponding time vector; f E (l u ,l w ,t i ) represents node l u With node l w The interactive features of l (l u ,t i '),f l (l w ,t i ') represent node l respectively u and l w At time point O ij Previous features; z(lu ,t i '),z(l w ,t i ') represents node l u and l w The representation vector before the interaction occurs; z'(l1,l2,t i ) represents the calculation of time-varying vectors.

[0013] In combination with the first aspect, in some implementations of the first aspect, the method further includes: when When The limit when p k (t i ), p k (t i ) is the intensity function, and the accumulation of all types of intensity functions is expressed as

[0014] By introducing a nonlinear transfer function, the total activation function is ensured to remain positive and allowed to be negative. The following is the modified formula:

[0015]

[0016] As time t i Growth, events occur step by step, intensity function p k It will rise or fall, but will eventually tend to the basic rate R k (ρ k +0).

[0017] In combination with the first aspect, in some implementations of the first aspect, the method further includes: a process of generating an online coding structure based on the intensity function of each event type of the interactive event behavior sequence:

[0018] The intensity function p for each event type k (t i ) changes dynamically over time; the hidden state vector h is closely related to the storage unit vector c; each storage unit c decays exponentially according to a certain decay rate δ; the definition of the ODE-LSTM network strength function is as follows:

[0019] p k (t i )=R k (w k h(t i ))

[0020]

[0021] The symbol h(t i ) is abbreviated as h i , c(ti ), Abbreviated as c i , h i The ODE-LSTM encoder is a LSTM in which the transformation between two observations is given by a neural ODE; at each time interval i, the LSTM unit is defined as follows:

[0022] I i =σ(w i-1 k i-1 +w hin h i +b in )

[0023] F i =σ(w xf k i-1 +w hf h i +b f )

[0024] o i =σ(w xo k i-1 +w ho h i +b o )

[0025]

[0026]

[0027] h i =o i ⊙tanh(c i )

[0028] Among them, h i is the intermediate hidden state, with the initial state h i-1 and the time interval [t i-1 ,t i ] is the solution to the ODE initial value problem.

[0029] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the viewing behavior sequence is defined at a timestamp t i The behavior that occurs as an event The behaviors included are watching, commenting, placing orders, and exiting the live broadcast room; the ordered event sequence is represented as N is the total number of events;

[0030] The live streaming ordering situation is formally defined as the time from watching to placing an order T X Total online time TO The ratio of , the quality of live order is expressed as The live order quality is divided into four labels, L∈0,1,2,3,5}; when L=0, it means no order, L=1, it means Q∈(0.75,1], when L=2, it means Q∈(0.5,0.75], when L=3, it means Q∈(0.25,0.5], when L=4, it means Q∈(0,0.25], and the quality detection is classified.

[0031] In combination with the first aspect, in some implementations of the first aspect, the method further includes: a prediction process of the predicted next behavior type:

[0032] Modeling the time series of events based on time point t i Previous historical live sequence Using intensity functions to predict future events The intensity function is defined as:

[0033]

[0034] in Represents Previous viewing behavior, Represents the event The viewing behavior embedding vector, h i-1 represents the encoded d-dimensional latent feature, w k represents the weight matrix of the fully connected layer, f p Represents the activation function, which uses the following formula:

[0035]

[0036] Where s is the scale hyperparameter; watch behavior As a historical viewing sequence The probability P of the occurrence of the i-th condition in the generation is expressed as:

[0037]

[0038] is the lifetime function, defined as follows:

[0039]

[0040] An exponential loss function is used to predict the type of next viewing behavior:

[0041]

[0042] L type is the loss function for the viewing behavior type.

[0043] In combination with the first aspect, in some implementations of the first aspect, the method further includes: a process of predicting the occurrence time of the corresponding viewing behavior based on the predicted next behavior type:

[0044] For the prediction of the occurrence time of viewing behavior, the occurrence time of the viewing behavior of the next event is predicted based on the event and historical viewing behavior sequence event The intensity function occurring at a certain time in the future is defined as follows:

[0045]

[0046] Then, given the viewing behavior and historical viewing behavior sequence The probability of the i-th event occurrence time under the condition is expressed as:

[0047]

[0048] The expected time of the next viewing behavior is calculated as follows:

[0049]

[0050] Use mean squared error as the loss function for time prediction:

[0051]

[0052] The loss function of the event-level decoder consists of two parts: one is the error L generated when predicting the viewing behavior type type , and the other part is the error L in predicting the time when the viewing behavior occurs time , the loss function is as follows L E :

[0053] L E =L type +L time .

[0054] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the decoder first integrates the event-level representation using an average pooling technique to extract a comprehensive representation Z at the live broadcast room level:

[0055]

[0056] The live broadcast room level representation is used as the input data of the decoder. The prediction is based on the pre-defined online viewing quality label, and the following exponential loss function is used to measure the prediction performance:

[0057]

[0058] Then, the total loss function is constructed. The total loss function L is constructed as the event-level loss L E and the live broadcast room level loss L c The accumulation of:

[0059] L=L E +L c .

[0060] In a second aspect, in order to achieve the above-mentioned purpose, the present invention discloses an online live broadcast behavior analysis system based on a neural Hawkes process, comprising:

[0061] A sequence generation module is used to obtain online live broadcast data, set fixed time periods, segment the online live broadcast data based on each fixed time period, obtain events corresponding to each fixed time period, and generate interactive event behavior sequences based on the events corresponding to each fixed time period;

[0062] The behavior analysis module is used to generate an online coding structure based on the intensity function of each event type in the interactive event behavior sequence, obtain a viewing behavior sequence based on the interactive event behavior sequence, input the viewing behavior sequence into the online coding structure, obtain the predicted next behavior type, and predict the occurrence time of the corresponding viewing behavior based on the predicted next behavior type.

[0063] Beneficial effects of the present invention:

[0064] The present invention evaluates the quality of live broadcast by the duration of user participation in live broadcast. Then, a live broadcast behavior sequence is defined, which can accurately track every step of user interaction in live broadcast. In addition, the layered encoding and decoding model proposed in the present invention aims to improve the prediction accuracy of user behavior and live broadcast quality. Finally, we apply this model to the actual shopping platform dataset and prove the superiority of this framework in predicting live broadcast viewing behavior by comparing with the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0066] Figure 1 It is a schematic flow chart of the method of the present invention;

[0067] Figure 2 It is a schematic diagram of the overall framework of the online viewing behavior modeling method of the present invention;

[0068] Figure 3 It is a schematic diagram of the core framework of the present invention;

[0069] Figure 4 It is a schematic diagram of experimental comparison results of the present invention;

[0070] Figure 5 It is a schematic diagram of the visualization result of the present invention;

[0071] Figure 6 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0072] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] Embodiment 1:

[0074] like Figure 1 As shown, a method for analyzing online live broadcast behavior based on a neural Hawkes process comprises the following steps:

[0075] S101: Acquire online live broadcast data, set fixed time periods, segment the online live broadcast data based on each fixed time period, obtain events corresponding to each fixed time period, and generate interactive event behavior sequences based on the events corresponding to each fixed time period;

[0076] The process of segmenting online live data based on each fixed time period:

[0077] Given a fixed time period, by dividing the time period into non-overlapping time periods, observing the events that occur in each time period, the events belong to the category in, Indicates that the event did not occur, k indicates the event category, K indicates the total number of event categories, and the set of interaction events is Indicates time period A sequence of events occurs, where an event is represented by k represents the category of the event, t i Indicates the exact time point when the event occurs, [t i-1 ,t i ]The division of time periods is determined by the time when the event occurs. After the event occurs, the time periods are automatically divided.

[0078] The process of generating interactive event behavior sequences based on the corresponding events in each fixed time period:

[0079] The generative probability model of event sequences is called a multivariate point process, and the category At time O ij occurs, in an infinitesimal time interval [t i ,t i +dt) in The probability of occurrence depends on historical information The probability of occurrence depends on the mutual information z'(l1,l2,t i ), and the mutual information z'(l1,l2,t i ) is composed as follows:

[0080] z'(l1,l2,t i )=R(γ(O ij ),f E (l u ,l w ,t i ),f l (l u ,t i '),f l (l w ,t i '),z(l u ,t i '),z(l w ,t i '))

[0081] Among them, γ(t i ) calculates the function of the time-varying representation vector of the node, and transforms the interaction time t i The function that is converted into the corresponding time vector; f E (l u ,l w ,t i ) represents node l u With node l w The interactive features of l (l u ,t i '),f l (l w ,t i ') represent node l respectively u and l w At time point O ij Previous features; z(l u ,t i '),z(l w ,t i ') represents node l u and l wThe representation vector before the interaction occurs; z'(l1,l2,t i ) represents the calculation of time-varying vectors.

[0082] when When The limit when p k (t i ), p k (t i ) is the intensity function, and the accumulation of all types of intensity functions is expressed as

[0083] By introducing a nonlinear transfer function, the total activation function is ensured to remain positive and allowed to be negative. The following is the modified formula:

[0084]

[0085] As time t i Growth, events occur step by step, intensity function p k It will rise or fall, but will eventually tend to the basic rate R k (ρ k +0).

[0086] S102: Generate an online coding structure based on the intensity function of each event type of the interactive event behavior sequence, obtain a viewing behavior sequence based on the interactive event behavior sequence, input the viewing behavior sequence into the online coding structure to obtain a predicted next behavior type, and predict the occurrence time of the corresponding viewing behavior based on the predicted next behavior type.

[0087] The process of generating online encoding structure based on the intensity function of each event type of interactive event behavior sequence:

[0088] The intensity function p for each event type k (t i ) changes dynamically over time; the hidden state vector h is closely related to the storage unit vector c; each storage unit c decays exponentially according to a certain decay rate δ; the definition of the ODE-LSTM network strength function is as follows:

[0089] p k (t i )=R k (w k h(t i ))

[0090]

[0091] The symbol h(t i ) is abbreviated as h i , c(t i ), Abbreviated as c i , h i The ODE-LSTM encoder is a LSTM in which the transformation between two observations is given by a neural ODE; at each time interval i, the LSTM unit is defined as follows:

[0092] I i =σ(w i-1 k i-1 +w hin h i +b in )

[0093] F i =σ(w xf k i-1 +w hf h i +b f )

[0094] o i =σ(w xo k i-1 +w ho h i +b o )

[0095]

[0096]

[0097] h i =o i ⊙tanh(c i )

[0098] Among them, h i is the intermediate hidden state, with the initial state h i-1 and the time interval [t i-1 ,t i ] is the solution to the ODE initial value problem.

[0099] The viewing behavior sequence is defined at timestamp t i The behavior that occurs as an event The behaviors included are watching, commenting, placing orders, and exiting the live broadcast room; the ordered event sequence is represented as N is the total number of events;

[0100] The live streaming ordering situation is formally defined as the time from watching to placing an order T X Total online time T O The ratio of , the quality of live order is expressed as The live order quality is divided into four labels, L∈0,1,2,3,5}; when L=0, it means no order, L=1, it means Q∈(0.75,1], when L=2, it means Q∈(0.5,0.75], when L=3, it means Q∈(0.25,0.5], when L=4, it means Q∈(0,0.25], and the quality detection is classified.

[0101] like Figure 2 As described, the study proposed an overall architecture for modeling online viewing behavior that includes an encoder and an event-level decoder. The input of this architecture includes live broadcast behavior sequences, user feature vectors, and live broadcast room feature vectors, aiming to optimize the prediction of the quality of user live broadcast experience and extract the feature representation that best represents the online viewing behavior pattern. With this approach, the study can not only capture the diversity of user behavior, but also perform personalized modeling based on the specific goals of each user. This helps to more accurately predict and understand user behavior patterns in live broadcast environments, thereby improving the quality of live broadcast content and the audience's viewing experience.

[0102] The core architecture of the present invention is a representation model, which is specially designed for modeling online live broadcast behaviors, such as Figure 3 As shown in the figure. The model adopts a method based on the neural Hawkes process. The technical framework mainly includes a one-level encoder and two-level decoders. The encoder part is composed of an ODE-LSTM network, which is responsible for converting the live broadcast behavior sequence into a vector representation on the time series. The decoder part consists of two levels: event level and live broadcast room level.

[0103] The event-level decoder receives the vectorized representation output by the encoder and predicts the live broadcast behavior event that may occur at the next time point. The live broadcast room-level decoder further processes these vector representations, generates feature vectors on timestamps through average pooling operations, and predicts the live broadcast quality of the current live broadcast room based on them.

[0104] In order to improve the accuracy of prediction, the present invention not only optimizes the prediction loss function of live broadcast behavior events, but also optimizes the prediction loss function of live broadcast quality. In this way, we can find the best vector representation set across timestamps, which will eventually be transformed into a deep understanding of online live broadcast behavior patterns and provide a higher quality live broadcast experience for the live broadcast room.

[0105] The main function of the encoder is to convert the online viewing behavior in the live broadcast into a vector form at each time point. In the present invention, we use the ODE-LSTM network technology described in the present invention to achieve this goal. The interactive behaviors of users in the live broadcast can be summarized into three main types: watching, commenting and ordering, and exiting the live broadcast room, which are marked by black, red and blue cells respectively. We assume that at any given time point, only one type of interactive behavior will occur.

[0106] Through this encoding method, we are able to capture the nuances of user behavior and provide a rich data foundation for further analysis and prediction. Such vectorized representation not only helps to understand the user's behavior patterns during the live broadcast, but also provides the necessary input for the model to make more accurate event predictions and quality assessments.

[0107] The event-level decoder focuses on modeling the granularity of a single event to construct online viewing behavior patterns. It does this by completing two key tasks: first, predicting the type of live streaming behavior that is about to occur, and second, predicting the specific time point at which the behavior will occur.

[0108] (1) Predicting the next behavior type: This paper first explores how to predict the type of viewing behavior that the user will take next. We use the concept of neural Hawkes process to model the time series of events. i Previous historical live sequence We define an intensity function to predict the occurrence of future events. The intensity function is defined as:

[0109]

[0110] in Represents Previous viewing behavior, Represents the event The viewing behavior embedding vector, h i-1 represents the encoded d-dimensional latent feature, w k represents the weight matrix of the fully connected layer, f p represents the activation function, and the activation function of the present invention adopts the following formula:

[0111]

[0112] Here, s is a scale hyperparameter that controls the curvature of the function. Then, we observe the behavior As a historical viewing sequence The probability P of the occurrence of the i-th condition in the generation can be expressed as:

[0113]

[0114] is the lifetime function, defined as follows:

[0115]

[0116] An exponential loss function is used to predict the type of next viewing behavior:

[0117]

[0118] L type is the loss function for the viewing behavior type.

[0119] (2) Prediction of the time when the next viewing behavior will occur:

[0120] For the prediction of the occurrence time of viewing behavior, the goal of this invention is to predict the occurrence time of the next event (viewing behavior).

[0121] Time. Event-based and historical viewing behavior sequence event The intensity function that occurs at a certain time in the future

[0122] The number is defined as follows:

[0123]

[0124] Then, given the viewing behavior and historical viewing behavior sequence The probability of the i-th event occurring time under the condition can be expressed as:

[0125]

[0126] The expected time of the next viewing behavior is calculated as follows:

[0127]

[0128] The present invention uses mean square error as the loss function for time prediction:

[0129]

[0130] Finally, the loss function of the event-level decoder consists of two parts: one is the error in predicting the type of viewing behavior, and the other is the error in predicting the time when the viewing behavior occurs. The sum of these two constitutes the loss function of the entire event-level decoder, which is used to measure the accuracy of the model prediction and guide the model optimization process.

[0131] L E=L type +L time

[0132] The event-level decoder focuses on modeling the viewing behavior patterns of a single event, while the live broadcast room decoder analyzes how the viewing behaviors of individuals are related to each other from a more macro perspective. To achieve this goal, the present invention first uses the average pooling technique to integrate the event-level representations, thereby extracting a comprehensive representation Z at the live broadcast room level. This method not only captures the details of local behaviors, but also reveals the intrinsic connections of global behavior patterns.

[0133]

[0134] Next, the present invention uses the live broadcast room level representation as the input data of the decoder to predict the viewing quality, which is based on the predefined online viewing quality label. In order to evaluate the accuracy of the predicted viewing quality, the present invention uses the following exponential loss function to measure the prediction performance:

[0135]

[0136] Finally, the algorithm proposed in this paper comprehensively optimizes the prediction of viewing behavior and the evaluation of viewing quality, aiming to capture the viewing behavior characteristics of online users at both the local and global levels. In mathematical expression, the total loss function is constructed as the accumulation of event-level loss and live broadcast room-level loss:

[0137] L=L E +L c

[0138] After training, the algorithm can capture event-level and live broadcast room-level features.

[0139] Specifically, the scheme of the present invention is further described below through experiments:

[0140] Experimental analysis

[0141] In this paper, we will introduce several core parts step by step to show our workflow and results:

[0142] 1) Dataset Overview:

[0143] This paper uses the online live broadcast function data from the "User Shopping Platform". We first conduct a basic statistical analysis of the data set, then explain the configuration details of the experiment in detail, and finally compare the performance of our proposed model with several existing algorithms on the same data set.

[0144] 2) Dataset and preprocessing:

[0145] User logs record the behavior of viewers in the live broadcast room in detail, including user ID, video ID, behavior type, timestamp, and the start and end time of the video. We paid special attention to records in which the user's viewing time exceeded half of the total length of the video to improve the quality and representativeness of the data. After preprocessing, the dataset contains 12,899 viewing records of 2,312 users and 5,968 live broadcast rooms, spanning from August 2020 to December 2022. In addition, the dataset also includes text information of users and live broadcast rooms, which provides important context for analyzing user behavior patterns.

[0146] 3) Experimental configuration:

[0147] In terms of live broadcast quality assessment, we divide live broadcast quality into five levels, forming a multi-category classification task. Our goal is to predict the quality of experience of users when watching live broadcasts by analyzing the embedding vectors of users, the embedding vectors of live broadcast rooms, and the behavioral characteristics of event sequences. User statistical features such as gender and age are encoded using One-hot, while live broadcast room information such as ID, title, and introduction are directly used for analysis. Text features are embedded using the Word2Vec model to capture the deep semantics of the text.

[0148] With this approach, our model can not only predict live broadcast quality in detail, but also identify key factors that affect user experience, providing a reference for personalized recommendations and quality improvements for live broadcast platforms.

[0149] 4) Model evaluation and comparison:

[0150] Our proposed model has shown its advantages in the live broadcast viewing prediction task in experiments. By comparing with existing methods, we have demonstrated the effectiveness and superiority of our framework. These experimental results not only verify the prediction accuracy of our algorithm, but also demonstrate its potential in practical applications.

[0151] 3) Evaluation indicators

[0152] The present invention uses the following indicators to measure the performance of the model in predicting live broadcast quality: The judgment results of the model are divided into positive examples (P) and negative examples (N), and the correctness of the judgment is represented by true examples (T) and false examples (F).

[0153] 1) Accuracy (ACC):

[0154]

[0155] Among them, TP, TN, FP and FN represent the number of true positives, true negatives, false positives and false negatives, respectively.

[0156] 2) Matthews Correlation Coefficient (MCC):

[0157]

[0158] MCC takes into account the values ​​of all four confusion matrices and is a balanced indicator for measuring the performance of a binary classification model. The higher the values ​​of these two indicators, the better the performance of the model. In the live broadcast quality prediction task, the present invention mainly uses ACC and MCC as evaluation criteria.

[0159] 3) Comparison algorithm:

[0160] The proposed technique will be compared with the following baseline algorithms:

[0161] -Logistic Regression (LR): A static model that maps prediction results to probability space through a logistic function and is suitable for classification tasks.

[0162] -Recurrent Neural Network (RNN): A standard model that can process sequence data.

[0163] -Long Short Term Memory (LSTM): LSTM is an extension of RNN and is particularly suitable for modeling sequential data.

[0164] -Gated Recurrent Unit (GRU): GRU is a variant of LSTM that optimizes the transfer of information between time steps.

[0165] 4) LHawkes-*:

[0166] The present invention first uses the proposed method to generate representations of live broadcasting behaviors, and then inputs these representations into the above classification model for prediction. The proposed solution is named LHawkes-*.

[0167] In the experimental setting, the implicit spatial representation dimension D of the model is set to 32 and 64 respectively. Stochastic gradient descent is used in the model training phase, using ADAM as the optimizer, the batch size is set to 128, the learning rate ranges from 0.001 to 0.01, and the dropout rate is set between 0.3 and 0.5 to avoid overfitting. All experiments are conducted on a machine equipped with an Intel i712700K 3.6GHz CPU (12 cores) and an AMD Radeon RX Vega 64 graphics card, and the operating system is Windows 10.

[0168] In this example, we will conduct an in-depth comparative analysis of the performance of the model proposed by the present invention and all the comparison methods. Figure 4 As shown in the figure, the performance of each model in different implicit space dimensions in the live broadcast quality prediction task is summarized based on the two indicators of accuracy (ACC) and Matthews correlation coefficient (MCC). Among all the compared methods, logistic regression (LR) is the only model that does not utilize the temporal characteristics of user behavior. Recurrent neural network (RNN), gated recurrent unit (GRU) and long short-term memory network (LSTM) all consider the temporal information of user behavior.

[0169] However, from Figure 4 From the overall trend, when the latent space embedding in the technical architecture of the present invention is not taken into account, the performance of RNN, GRU and LSTM is not significantly different from that of LR. This shows that it is difficult for these algorithms alone to effectively capture the timing and dynamic changes of user behavior. When we integrate the representation scheme proposed by the present invention into these comparison algorithms, we can observe a significant performance improvement over all baseline algorithms.

[0170] In addition, the experimental results also revealed an interesting phenomenon: as the dimension of the implicit representation space increases, for example, from 32 to 64, the accuracy of the model does not change significantly, but a more obvious improvement is observed on MCC. This finding shows that the method proposed in this invention can effectively preserve the pattern of users' online viewing behavior in the viewing representation, and surpasses traditional baseline algorithms such as LR, RNN, GRU and LSTM in terms of stability.

[0171] In this example, we will use visualization to analyze the experimental results of the representation model proposed in this invention in depth to intuitively demonstrate the effectiveness of the algorithm. To achieve this, we use t-SNE technology, which is a method that effectively reduces the dimensionality of high-dimensional data to two-dimensional space. Figure 5 As shown, the method of the present invention maps data points onto a two-dimensional plane, where different colors represent different viewing quality levels.

[0172] observe Figure 5 A striking phenomenon can be found: when the viewing quality reaches the highest level (level 4, marked in dark blue) or the lowest level (level 0, marked in light blue), the learned features show a high concentration in the two-dimensional space. This shows that our model is able to capture obvious features related to viewing quality, and these features have good discrimination in the feature space. This visualization result not only verifies the effectiveness of our algorithm, but also provides an intuitive view for further understanding how the model characterizes different viewing qualities.

[0173] The innovation of this study is to propose a novel online live streaming behavior modeling method based on the neural Hawkes process. The following are the specific steps of the research implementation:

[0174] 1. Live broadcast quality evaluation: We have designed an innovative evaluation mechanism to measure the attractiveness of live broadcasts based on the length of time users participate in the live broadcast.

[0175] 2. Behavior sequence definition: We have built a live broadcast behavior sequence that can record every interaction of users in the live broadcast in detail, thereby capturing their behavior patterns.

[0176] 3. Innovation of encoding and decoding model: This study proposed a hierarchical encoding and decoding model, which significantly improved the prediction accuracy of user behavior and live broadcast quality by combining user behavior sequences and live broadcast characteristics.

[0177] 4. Empirical study: We applied the developed model to a real shopping platform dataset and compared it with the existing technology. The experimental results show that our model has significant advantages in predicting user live broadcast viewing behavior.

[0178] Through these steps, the present invention not only improves the understanding of user online behavior, but also provides a powerful tool for the live broadcast platform to optimize user experience and improve live broadcast effect.

[0179] Embodiment 2: In the second aspect, in order to achieve the above-mentioned purpose, the present invention discloses an online live broadcast behavior analysis system based on a neural Hawkes process, comprising:

[0180] The sequence generation module 11 is used to obtain online live broadcast data, set fixed time periods, segment the online live broadcast data based on each fixed time period, obtain events corresponding to each fixed time period, and generate interactive event behavior sequences based on the events corresponding to each fixed time period;

[0181] The behavior analysis module 12 is used to generate an online coding structure based on the intensity function of each event type in the interactive event behavior sequence, obtain a viewing behavior sequence based on the interactive event behavior sequence, input the viewing behavior sequence into the online coding structure, obtain the predicted next behavior type, and predict the occurrence time of the corresponding viewing behavior based on the predicted next behavior type.

[0182] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.

[0183] It needs to be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium, on which a computer program is stored, and the computer program is executed by a processor to execute the above method. The storage medium can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0184] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0185] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, the present disclosure may have various changes and improvements, and these changes and improvements fall within the scope of the present disclosure to be protected.

Claims

1. A method for analyzing online live broadcast behavior based on a neural Hawkes process, characterized in that: The method comprises the following steps: Acquire online live broadcast data, set fixed time periods, segment the online live broadcast data based on each fixed time period, obtain events corresponding to each fixed time period, and generate interactive event behavior sequences based on the events corresponding to each fixed time period; An online coding structure is generated based on the intensity function of each event type in the interactive event behavior sequence, and a viewing behavior sequence is obtained based on the interactive event behavior sequence. The viewing behavior sequence is input into the online coding structure to obtain the predicted next behavior type, and the occurrence time of the corresponding viewing behavior is predicted based on the predicted next behavior type.

2. According to claim 1, a method for analyzing online live broadcast behavior based on a neural Hawkes process is characterized in that: The process of segmenting the online live broadcast data based on each fixed time period: Given a fixed time period, by dividing the time period into non-overlapping time periods, observing the events that occur in each time period, the events belong to the category in, Indicates that the event did not occur, k indicates the event category, K indicates the total number of event categories, and the set of interaction events is Indicates time period A sequence of events occurs, where an event is represented by k represents the category of the event, t i Indicates the exact time point when the event occurs, [t i-1 ,t i ]The division of time periods is determined by the time when the event occurs. After the event occurs, the time periods are automatically divided.

3. According to claim 1, a method for analyzing online live broadcast behavior based on a neural Hawkes process is characterized in that: The process of generating an interactive event behavior sequence based on the events corresponding to each fixed time period: The generative probability model of event sequences is called a multivariate point process, and the category At time O ij occurs, in an infinitesimal time interval [t i ,t i +dt) in The probability of occurrence depends on historical information p k (t i |z'(l1,l2,t i The probability of occurrence of )) depends on the interaction information z'(l1,l2,t i ), and the mutual information z'(l1,l2,t i ) is composed as follows: z'(l1,l2,t i )=R(γ(O ij ),f E (l u ,l w ,t i ),f l (l u ,t i '),f l (l w ,t i '),z(l u ,t i '),z(l w ,t i ')) Among them, γ(t i ) calculates the function of the time-varying representation vector of the node, and transforms the interaction time t i The function that is converted into the corresponding time vector; f E (l u ,l w ,t i ) represents node l u With node l w The interactive features of l (l u ,t i '),f l (l w ,t i ') represent node l respectively u and l w At time point O ij Previous features; z(l u ,t i '),z(l w ,t i ') represents node l u and l w The representation vector before the interaction occurs; z'(l1,l2,t i ) represents the calculation of time-varying vectors.

4. The online live broadcast behavior analysis method based on the neural Hawkes process according to claim 3 is characterized in that: when When The limit when p k (t i ), p k (t i ) is the intensity function, and the accumulation of all types of intensity functions is expressed as By introducing a nonlinear transfer function, the total activation function is ensured to remain positive and allowed to be negative. The following is the modified formula: As time t i Growth, events occur step by step, intensity function p k It will rise or fall, but will eventually tend to the basic rate R k (ρ k +0).

5. The online live broadcast behavior analysis method based on the neural Hawkes process according to claim 1 is characterized in that: The process of generating an online coding structure based on the intensity function of each event type of the interactive event behavior sequence: The intensity function p for each event type k (t i ) changes dynamically over time; the hidden state vector h is closely related to the storage unit vector c; each storage unit c decays exponentially according to a certain decay rate δ; the definition of the ODE-LSTM network strength function is as follows: p k (t i )=R k (w k h(t i )) The symbol h(t i ) is abbreviated as h i , c(t i ), Abbreviated as c i , h i The ODE-LSTM encoder is a LSTM in which the transformation between two observations is given by a neural ODE; at each time interval i, the LSTM unit is defined as follows: I i =σ(w i-1 k i-1 +w hin h i '+b in ) F i =σ(w xf k i-1 +w hf h i '+b f ) o i =σ(w xo k i-1 +w ho h i '+b o ) h i =o i ⊙tanh(c i ) Among them, h i is the intermediate hidden state, with the initial state h i-1 and the time interval [t i-1 ,t i ] is the solution to the ODE initial value problem.

6. The method for analyzing online live broadcast behavior based on the neural Hawkes process according to claim 1, characterized in that: The viewing behavior sequence is defined at timestamp t i The behavior that occurs as an event The behaviors included are watching, commenting, placing orders, and exiting the live broadcast room; the ordered event sequence is represented as N is the total number of events; The live streaming ordering situation is formally defined as the time from watching to placing an order T X Total online time T O The ratio of , the quality of live order is expressed as The live order quality is divided into four labels, L∈{0,1,2,3,5}; when L=0, it means no order, L=1, it means Q∈(0.75,1], when L=2, it means Q∈(0.5,0.75], when L=3, it means Q∈(0.25,0.5], when L=4, it means Q∈(0,0.25], and the quality detection is classified.

7. The method for analyzing online live broadcast behavior based on the neural Hawkes process according to claim 1, characterized in that: The prediction process of the next behavior type is as follows: Modeling the time series of events based on time point t i Previous historical live sequence Using intensity functions to predict future events The intensity function is defined as: in Represents Previous viewing behavior, Represents the event The viewing behavior embedding vector, h i-1 represents the encoded d-dimensional latent feature, w k represents the weight matrix of the fully connected layer, f p Represents the activation function, which uses the following formula: Where s is the scale hyperparameter; watch behavior As a historical viewing sequence The probability P of the i-th condition occurring in a generation is expressed as: is the lifetime function, defined as follows: An exponential loss function is used to predict the type of next viewing behavior: L type is the loss function for the viewing behavior type.

8. The method for analyzing online live broadcast behavior based on the neural Hawkes process according to claim 7, characterized in that: The process of predicting the occurrence time of the corresponding viewing behavior based on the predicted next behavior type: For the prediction of the occurrence time of viewing behavior, the occurrence time of the viewing behavior of the next event is predicted based on the event and historical viewing behavior sequence event The intensity function occurring at a certain time in the future is defined as follows: Then, given the viewing behavior and historical viewing behavior sequence The probability of the i-th event occurring time under the condition is expressed as: The expected time of the next viewing behavior is calculated as follows: Use mean squared error as the loss function for time prediction: The loss function of the event-level decoder consists of two parts: one is the error L generated when predicting the viewing behavior type type , and the other part is the error L in predicting the time when the viewing behavior occurs time , the loss function is as follows L E : L E =L type +L time .

9. The online live broadcast behavior analysis method based on the neural Hawkes process according to claim 8 is characterized in that: The decoder first uses the average pooling technique to integrate the event-level representations to extract the comprehensive representation Z at the live broadcast room level: The live broadcast room level representation is used as the input data of the decoder. The prediction is based on the pre-defined online viewing quality label, and the following exponential loss function is used to measure the prediction performance: Then, the total loss function is constructed. The total loss function L is constructed as the event-level loss L E and the live broadcast room level loss L c The accumulation of: L=L E +L c .

10. An online live broadcast behavior analysis system based on neural Hawkes process, characterized in that: include: A sequence generation module is used to obtain online live broadcast data, set fixed time periods, segment the online live broadcast data based on each fixed time period, obtain events corresponding to each fixed time period, and generate interactive event behavior sequences based on the events corresponding to each fixed time period; The behavior analysis module is used to generate an online coding structure based on the intensity function of each event type in the interactive event behavior sequence, obtain a viewing behavior sequence based on the interactive event behavior sequence, input the viewing behavior sequence into the online coding structure, obtain the predicted next behavior type, and predict the occurrence time of the corresponding viewing behavior based on the predicted next behavior type.

Images