Reservation service order returning prediction method and device based on multi-scale time sequence feature fusion
Through multi-scale timing feature fusion and multi-task learning prediction network, a order rejection and order damage prediction model is built, which solves the accuracy of order damage prediction for appointment service orders, and improves the accuracy and comprehensiveness of prediction results.
Patent Information
- Application Number
- CN202411120291.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2024-12-06
AI Technical Summary
The prior art is difficult to accurately predict the order-breaking of reservation service orders, resulting in reduced user satisfaction and damaged service provider reputation.
Adopt the reservation service order refund prediction method based on multi-scale timing feature fusion, and by obtaining the characteristic information of service providers and users, statistical timing data and non-time sequence data, combined with the multi-task learning prediction network, the order rejection prediction auxiliary model and order damage are constructed. Predict the main model and make predictions.
It improves the accuracy of order damage prediction for appointment service orders, ensures the comprehensiveness and accuracy of prediction results, and reduces the inaccuracy of order damage prediction.
Smart Images

Figure CN119090060A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of prediction of cancellation of reservation orders, and in particular to a method and device for predicting cancellation of reservation services based on multi-scale time series feature fusion. Background Art
[0002] In many service industries, orders are often cancelled unexpectedly after booking. For example, customers make reservations for car rental services on platforms such as Meituan and Airbnb, but are told that there are no cars when using them, or customers make reservations for restaurants, but are told that there are no tables available when they arrive. These cancellations will reduce user satisfaction, damage the reputation of service providers, and cause customer loss. Therefore, there is an urgent need for a method that can accurately predict cancellations of reservation service orders. Summary of the invention
[0003] The purpose of this application is to provide a reservation service return prediction method and device based on multi-scale time series feature fusion, which can accurately predict whether the reservation service order will be cancelled during the service.
[0004] To achieve the above objectives, this application provides the following solutions:
[0005] In a first aspect, the present application provides a reservation service return prediction method based on multi-scale time series feature fusion, comprising:
[0006] Acquire reservation feature data of the reservation service order to be predicted; the reservation feature data includes feature information of the service provider, reservation feature information of the user, statistical time series data and statistical non-time series data; the statistical time series data includes a time series of the number of orders received, a time series of orders destroyed during service, a time series of orders rejected, a time series of reservation platform attention and a time attribute series within a first preset time period before the reservation service time point of the reservation service order to be predicted; the statistical non-time series data includes a total number of orders received, a total number of orders destroyed during service, a total number of orders rejected and a total number of reservation platform attention within a second preset time period before the reservation service time point of the reservation service order to be predicted;
[0007] The multi-scale time series features of the order quantity, service order cancellation time series, rejected order quantity, reservation platform attention time series and time attribute series are fused to obtain a comprehensive representation of time series features.
[0008] Input the service provider feature information, user reservation feature information and statistical non-time series data into the order rejection prediction auxiliary model of the trained multi-task learning prediction network to obtain the order rejection prediction result of the reservation service order to be predicted;
[0009] The comprehensive representation of time series features, service provider feature information, user reservation feature information, statistical non-time series data and order rejection prediction results are input into the main order destruction prediction model of the trained multi-task learning prediction network to obtain the order destruction prediction results of the reservation service orders to be predicted that are destroyed during the service; the order destruction prediction main model and the order rejection prediction auxiliary model are machine learning models.
[0010] In a second aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for predicting a refund of an appointment service based on multi-scale time series feature fusion.
[0011] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting order returns for appointment services based on multi-scale time series feature fusion.
[0012] In a fourth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the method for predicting cancellation of an appointment service based on multi-scale time series feature fusion.
[0013] According to the specific embodiments provided in this application, this application discloses the following technical effects:
[0014] The present application provides a reservation service return order prediction method and device based on multi-scale time series feature fusion. On the first aspect, reservation feature data of reservation service orders are used to predict cancellation of orders, wherein the reservation feature data includes service provider feature information, user reservation feature information, statistical time series data and statistical non-time series data, wherein the statistical non-time series data and statistical time series data include indicators such as the number of orders received, the number of cancelled orders during service, the number of rejected orders and the attention of the reservation platform; using various indicator information of reservation service orders to participate in the prediction of cancelled orders, the feature data is more comprehensive and can better ensure the accuracy of the prediction results. On the second aspect, the present invention further performs multi-scale time series feature fusion on the time series corresponding to each indicator in the statistical time series data, and highlights the characteristics of the data through feature fusion to further improve the accuracy of the prediction. On the third aspect, the present invention also introduces a multi-task learning prediction network, and uses a rejection prediction auxiliary model for auxiliary prediction to effectively improve the accuracy of the cancellation prediction main model. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0016] Figure 1 A flowchart of a reservation service return prediction method based on multi-scale time series feature fusion provided in an embodiment of the present application;
[0017] Figure 2 A schematic diagram of the technical concept of a reservation service return prediction method based on multi-scale time series feature fusion provided in an embodiment of the present application;
[0018] Figure 3 A structural diagram of a main order cancellation prediction model provided in an embodiment of the present application;
[0019] Figure 4 A schematic diagram of a multi-scale time series feature fusion process for each time series in statistical time series data provided in an embodiment of the present application. DETAILED DESCRIPTION
[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with 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 of the embodiments.
[0021] Example 1
[0022] like Figure 1 and Figure 2 As shown, a reservation service return prediction method based on multi-scale time series feature fusion is provided, including the following steps 201 to 204.
[0023] Step 201, obtaining reservation feature data of the reservation service order to be predicted; the reservation feature data includes service provider feature information, user reservation feature information, statistical time series data and statistical non-time series data.
[0024] The reservation feature data of the reservation service order can be obtained by integrating the relevant data of the online reservation service platform and the service provider management system as well as the external data source (local activity arrangement events).
[0025] Among them, the service provider characteristic information includes the service price of the service provider, the star rating of the service provider, the brand of the service provider, the type of service provider (such as whether it is a direct sales / self-operated service provider), the city where the service provider is located and the regional business district, the scale of the service provider and other information. User reservation characteristic information includes the number of days in advance for reservation, the reservation service time (if the service provider is a hotel service provider, it is in hours), the service week, whether it is a holiday, and whether it is a local activity day. Among them, the statistical time series data includes the time series of the number of orders received in the first preset time period before the reservation service time point of the reservation service order to be predicted, the time series of the number of orders destroyed during service, the time series of the number of rejected orders, the time series of the reservation platform attention, and the time attribute series; the statistical non-time series data includes the total number of orders received in the second preset time period before the reservation service time point of the reservation service order to be predicted, the total number of orders destroyed during service, the total number of rejected orders, and the total number of reservation platform attention. The first preset time period and the second preset time period can be determined according to actual needs. As an example, statistical time series data includes the total number of orders received by service providers, the total number of orders cancelled during service, the total number of rejected orders, the total number of reservation platform attention, and the corresponding time attribute series collected for each day within 28 consecutive days within 1 / 7 / 30 / 90 days before that day. Taking the number of orders as an example, 1 / 7 / 30 / 90 days correspond to four order time series. As an example, statistical non-time series data includes the total number of orders received by service providers, the total number of orders cancelled during service, the total number of rejected orders, and the total number of reservation platform attention within 1 / 7 / 30 / 90 days before the reservation service time point of the reservation service order. Taking the number of orders as an example, 1 / 7 / 30 / 90 days correspond to four order totals. The specific day values of 28 days, 1 / 7 / 30 / 90 days involved here are set according to demand. There is no specific restriction here. It is just an example. Among them, the number of orders received refers to the number of appointment service orders agreed to be booked by the service provider; the number of orders cancelled during service refers to the number of appointment service orders agreed to be booked by the service provider that were rejected by the service provider when waiting for service; the number of historical rejected orders refers to the number of appointment service orders booked without the consent of the service provider; the attention to the appointment platform refers to the frequency of service providers maintaining the appointment inventory of the appointment platform.
[0026] Step 202, multi-scale time series feature fusion is performed on the time series of received orders, the time series of cancelled orders during service, the time series of rejected orders, the time series of attention to the reservation platform, and the time attribute series to obtain a comprehensive representation of the time series features.
[0027] Step 203, input the service provider characteristic information, the user reservation characteristic information and the statistical non-time series data into the trained order rejection prediction auxiliary model to obtain the order rejection prediction result of the reservation service order to be predicted.
[0028] Step 204: Comprehensively represent the time series features, service provider feature information, user reservation feature information, statistical non-time series data, and order rejection prediction results (corresponding to Figure 2 The rejection vector in ) is input into the trained order cancellation prediction main model to obtain the order cancellation prediction result (corresponding to Figure 2 Among them, the main model for order cancellation prediction and the auxiliary model for order rejection prediction are machine learning models, such as perceptron MLP; Figure 3 , shows the structure of a main model for order rejection prediction (i.e., a main model without prediction), which mainly includes 3 layers of ReLU layers. The input data of the auxiliary model for order rejection prediction and the main model for order rejection prediction need to be concatenated before being input into the model.
[0029] In another exemplary embodiment of the present application, the multi-scale temporal feature fusion in step 202 includes three processes, such as Figure 4 As shown, it includes single-scale cross-sequence fusion, multi-scale representation fusion, and time-based Attention representation processing.
[0030] (1) Single-scale cross-sequence fusion
[0031] Single-scale cross-sequence fusion first uses J different convolution kernels to extract features from each time series, and then fuses the output matrices of the same convolution kernel for different time series. The time series is represented by a curve, such as Figure 4 As shown in the figure, the horizontal axis of the curve corresponding to the time series is the time step, and the vertical axis represents the indicator values of different types of time series, such as the number of orders received, the number of cancelled orders, and the number of rejected orders. Figure 4 Indicated by "Value".
[0032] Specifically, for the order quantity time series, service time series, order rejection time series, reservation platform attention time series and time attribute series, we use S = (s1, s2, ..., s i ,...,s K ) is represented by the following processing:
[0033] A. Multi-scale sequence representation extraction. For each time series s i Perform a one-dimensional convolution of a given scale and use the one-dimensional convolution of a given scale to kernel the time series s i Perform feature extraction on multiple time scales to capture the characteristics of each scale. Suppose that for the time series s i Using J one-dimensional convolution kernels of different sizes, CNN ij For the time series s i The convolution kernel size is a one-dimensional convolution kernel of the jth size, and the convolution kernel size is several predetermined integer values, such as 3, 7, 14, and 28. Each time series s iThe same number of convolution kernels are used, and the same size of convolution kernels are used between different time series. The formal expression is as follows:
[0034] C ij =CNN ij (s i );
[0035] M(CNN 1j )=M(CNN 2j )=…=M(CNN Kj );
[0036]
[0037] Among them, C ij For the time series s i The feature matrix under the jth convolution kernel (i.e., the convolution result matrix) has Rows and columns, C is the convolution output dimension (that is, the number of output channels, all convolution operations set the same output dimension), is the j-th size convolution kernel in the time series s i Each row of the feature matrix output corresponds to a time step / time interval, and a row in the matrix represents the convolution output vector of a certain time series feature in a certain time interval when a convolution kernel slides. M(·) represents the convolution kernel size (an integer value for one-dimensional convolution).
[0038] B. Fusion of representations across sequences at the same scale. Obtain the convolution result matrix of each time series under the same convolution kernel size. For all convolution result matrices of the j-th convolution kernel size, record them as:
[0039]
[0040] Then, a fusion expression is performed at the same scale, that is, a single-scale fusion matrix G of each scale (convolution kernel) is obtained. j The specific fusion operation uses the MLP network. Each scale corresponds to an independent fusion network. The fusion expression at the same scale is:
[0041]
[0042] (2) Multi-scale sequence representation fusion
[0043] Multi-scale sequence representation fusion is mainly to further fuse the single-scale fusion matrices obtained in (1) single-scale cross-sequence fusion. Since the above CNN convolution kernels have different sizes, G under different convolution kernels j The number of rows of (the number of steps of the jth convolution kernel sliding, that is, the number of time intervals) are all different, so a padding operation is required first to fill them all to the maximum number of rows LM , the completed fusion matrix It is expressed as:
[0044]
[0045] Then, the Attention mechanism is used to fuse the sequence fusion representations of multiple scales to obtain the temporal comprehensive representation of the multi-dimensional time series.
[0046] Among them, the completed fusion matrix It is expressed as: in R C , represents the feature vector in the fusion matrix corresponding to the jth convolution kernel at time step / time interval t. According to the matrix Get the matrix matrix By each The cross-scale combination matrix obtained by combining the t-th row vector of , which represents the multi-scale features in a certain time interval.
[0047] Based on the Attention mechanism, multiple cross-scale combination matrices are cross-scale fused to obtain a cross-scale fusion matrix Its formal expression is as follows:
[0048]
[0049] α t =softmax(score t );
[0050]
[0051] Among them, α t =(α t1 ,α t2 ,…,α tJ ),W1∈R 1×J ,B1∈R 1×C ,W2∈R C×J ,B2∈R 1×J , W1, B1, W2, B2 represent weight coefficients. f1(·) and f2(·) are activation functions, such as relu activation function or tanh activation function.
[0052] (3) Comprehensive representation of temporal features based on time dimension Attention
[0053] In the above (2), we get the cross-scale fusion matrix After that, the Attention mechanism is applied from the time dimension to capture the time period that can reflect the characteristics of the time trend, and finally the comprehensive representation of the time series features V is obtained. When doing the Attention in the time dimension, the reservation feature information such as the number of days in advance, the order time (in hours), the service week, and whether it is used on holidays is used as the query. The query is represented by Q∈R q The specific calculation method is as follows:
[0054]
[0055]
[0056]
[0057]
[0058] In the formula, concat() means concatenation; MLP means perceptron. When calculating the Concat operation in the formula of z, are matrices with the same number of rows, It is an L M The matrix is a C-row matrix, but Q itself is just a 1-row, q-column vector, so we need to copy this vector to L M times, get an L M A matrix with q rows and columns
[0059] In another exemplary embodiment of the present application, before executing step 203 "inputting the service provider characteristic information, the user reservation characteristic information and the statistical non-time series data into the trained order rejection prediction auxiliary model", it also includes: training the multi-task learning prediction network; the multi-task learning prediction network includes a main model for order rejection prediction and an auxiliary model for order rejection prediction; the training process specifically includes:
[0060] (b1) Obtain a sample set of historical reservation service orders and label data corresponding to each sample of historical reservation service orders; the label data includes reservation labels and service labels: reservation labels include successful reservation and unsuccessful reservation; service labels include successful reservation and normal service and successful reservation but unsuccessful service; successful reservation but unsuccessful service means that the reservation is successful but the order is destroyed during the service.
[0061] (b2) Obtaining reservation feature sample data of each historical reservation service order sample; the reservation feature sample data is the same as the index included in the reservation feature data in the aforementioned step 201, and is only sample data here.
[0062] (b3) The sample order quantity time series, the sample order cancellation time series, the sample order rejection time series, the sample reservation platform attention time series and the sample time attribute series are subjected to multi-scale time series feature fusion to obtain a comprehensive representation of the sample time series feature. The multi-scale time series feature fusion process here is the same as the multi-scale time series feature fusion process described above, and the aforementioned fusion process can be referred to, which will not be repeated here.
[0063] (b4) Using the service provider feature sample information, user reservation feature sample information and statistical non-time series sample data of each historical reservation service order sample as input, and the reservation label corresponding to each historical reservation service order sample as label data, train the order rejection prediction auxiliary model.
[0064] (b5) Taking the comprehensive representation of sample time series features of each historical appointment service order sample, service provider feature sample information, user appointment feature sample information, statistical non-time series sample data and sample rejection prediction results of the rejection prediction auxiliary model as input, and the service label corresponding to each historical appointment service order sample as label data, the main model for rejection prediction is trained.
[0065] (b6) Calculate the overall loss error of the multi-task learning prediction network; the overall loss error includes the loss error of the rejection auxiliary task and the loss error of the destruction main task.
[0066] Among them, the expression of the overall loss error is:
[0067] Loss = L nr +α*L rf ;
[0068] in,
[0069]
[0070] Where Loss represents the overall loss error; L rf represents the loss error of the rejection auxiliary task; L nr Indicates the loss error of the destroyed main task (corresponding to Figure 2 α is the weight of the auxiliary task, which can be 0.1; y ri is the actual result of rejection of the i-th historical reservation service order sample; y' ri is the rejection prediction result of the i-th historical reservation service order sample; y ni is the actual result of order cancellation of the i-th historical reservation service order sample; y' ni is the prediction result of cancellation of the i-th historical reservation service order sample; ri Indicates whether the i-th historical reservation service order sample is a sample of unsuccessful reservation, I riCan be 0 or 1; I ni Indicates whether the i-th historical reservation service order sample is a sample that was successfully booked but not served normally, I ni It can be 0 or 1; N represents the number of samples.
[0071] (b7) When the overall loss error converges or reaches the preset maximum number of training iterations, the trained multi-task learning prediction network is obtained.
[0072] In order to solve the problem of inaccurate order destruction prediction results caused by the sparsity of destroyed order data during service, the present invention obtains statistical time series data and statistical non-time series data as input data for order destruction prediction, and performs multi-scale time series feature fusion on each time series of statistical time series data to capture features at each scale to improve the accuracy of order destruction prediction. On the other hand, based on the main model for order destruction prediction, an auxiliary model for order rejection prediction is introduced, and the output of the auxiliary model for order rejection prediction is used as the input of the main model for order destruction prediction. During training, the rejection loss error and the destroyed order loss error are used to jointly constitute the overall loss error, and the model parameters of the order destruction prediction model are adjusted according to the overall loss error, which can ensure the prediction accuracy of the trained main model for order destruction prediction.
[0073] The following is a verification result of the accuracy of the prediction of order cancellation by the method of the present invention. The reservation service order data of 3 months from an online reservation platform was collected as a sample. The total number of samples (orders) was about 10 million, and the number of orders that were successfully booked but cancelled during service was 10,000.
[0074] Among them, the models involved in the comparison include: M1: a simple MLP model, in which the features do not use any time series features; M2: an LSTM+MLP model, in which the time series features are processed using a basic LSTM model; M3: a LAXCAT+MLP model, in which the time series features are processed using a LAXCAT model; M4: a DARNN+MLP model, in which the time series features are processed using a DARNN model.
[0075] The comparison indicator is AUC. AUC is defined as the area under the ROC curve. The larger the AUC, the better the model effect.
[0076] Table 1 Comparison of the effects of different models in single task and multi-task
[0077]
[0078] Table 2 Comparison of the effects of time series features at each single scale and fusion of multi-scale time series
[0079] Convolution scale AUC Only use scale Scale-3 (i.e. convolution kernel = 3) 0.925 Only use scale Scale-7 (i.e. convolution kernel = 7) 0.926 Only use scale Scale-14 (i.e. convolution kernel = 14) 0.924 Only use scale Scale-28 (i.e. convolution kernel = 28) 0.924 The present invention uses the above four scales for fusion 0.929
[0080] Table 3 Comparison of the effects of different TimeAttention methods
[0081]
[0082] In order to prove the superiority of the method of the present invention, the present invention has carried out comparative experiments with other advanced methods on real data sets, and the experimental results are shown in Table 1, Table 2 and Table 3. From Table 1, after introducing the rejection auxiliary task on the basis of each single-task model, there is a significant improvement, and the multi-task learning prediction network proposed by the present invention has the best effect. From Table 2, although the multi-scale fusion representation method is not very outstanding and has only a slight advantage over the single-scale fusion representation, it can reduce the workload of exploring and optimizing at multiple scales when adjusting the model parameters, because if a single scale is used, it is necessary to choose which one of 3, 7, 14, and 18 needs to be tested, but if it is multi-scale, there is no need to adjust the parameter of the convolution kernel size, and it can be directly fused and used, so the multi-scale fusion representation method still has certain practical significance. From Table 3, the Attention method in the LAXCAT paper (i.e., Hsieh, Tsung-Yu, et al. "Explainable multivariate timeseries classification: a deep neural network which learns to attend to important variables as well as time intervals." Proceedings of the 14th ACM international conference on web search and data mining. 2021) adopts a pure parameter learning approach, that is, the Attention Query is obtained by pure parameterized learning, while the Attention Query of the present invention uses appointment feature information, such as the implicit relationship between the usage date of the order sample and the corresponding date of each time step in the time series, and therefore has better performance.
[0083] The present application also provides an application scenario, which applies the above-mentioned reservation service refund prediction method based on multi-scale time series feature fusion. Specifically: the reservation service refund prediction method based on multi-scale time series feature fusion provided in this embodiment can be applied to various service reservation scenarios in life. For example, restaurant reservations in Meituan and Dianping, homestay reservations in Airbnb, etc. These scenarios include service message publishing links, service reservation links, and order cancellation prediction links. The service provider publishes services on the platform, and users make service reservations through the platform. When the service provider confirms that the user's reservation is accepted, the present invention will predict the probability of the user's order being cancelled on the reservation date, so as to take measures in advance to protect the user's rights and interests. The reservation service refund prediction method based on multi-scale time series feature fusion provided in this embodiment belongs to the order cancellation prediction link.
[0084] Example 2
[0085] This embodiment provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the reservation service return prediction method based on multi-scale time series feature fusion in the above-mentioned embodiment 1 is implemented.
[0086] Example 3
[0087] This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the method for predicting cancellation of an appointment service based on multi-scale time series feature fusion in Embodiment 1 is implemented.
[0088] Example 4
[0089] This embodiment provides a computer program product, including a computer program, which, when executed by a processor, implements the reservation service return prediction method based on multi-scale time series feature fusion in Embodiment 1.
[0090] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.
Claims
1. A reservation service return prediction method based on multi-scale time series feature fusion, characterized in that: include: Obtaining reservation feature data of the reservation service order to be predicted; The reservation feature data includes feature information of the service provider, user reservation feature information, statistical time series data and statistical non-time series data; the statistical time series data includes a time series of orders received, a time series of orders destroyed during service, a time series of rejected orders, a time series of reservation platform attention and a time attribute series within a first preset time period before the reservation service time point of the reservation service order to be predicted; the statistical non-time series data includes a total number of orders received, a total number of orders destroyed during service, a total number of rejected orders and a total number of reservation platform attention within a second preset time period before the reservation service time point of the reservation service order to be predicted; Perform multi-scale time series feature fusion on the time series of received orders, the time series of cancelled orders during service, the time series of rejected orders, the time series of attention to the reservation platform, and the time attribute series to obtain a comprehensive representation of the time series features; Inputting the service provider characteristic information, the user reservation characteristic information and the statistical non-time series data into the trained order rejection prediction auxiliary model to obtain the order rejection prediction result of the reservation service order to be predicted; The comprehensive representation of time series features, the service provider feature information, the user reservation feature information, the statistical non-time series data and the order rejection prediction result are input into the trained order destruction prediction main model to obtain the order destruction prediction result of the reservation service order to be predicted being destroyed during the service; the order destruction prediction main model and the order rejection prediction auxiliary model are machine learning models.
2. The reservation service return prediction method based on multi-scale time series feature fusion according to claim 1 is characterized in that: The order volume refers to the number of appointment service orders agreed to be booked by the service provider; the order volume cancelled during service refers to the number of appointment service orders agreed to be booked by the service provider but rejected by the service provider when waiting for service; the rejected order volume refers to the number of appointment service orders booked without the consent of the service provider; the appointment platform attention refers to the frequency of appointment inventory maintenance by the service provider on the appointment platform.
3. The reservation service return prediction method based on multi-scale time series feature fusion according to claim 1 is characterized in that: The multi-scale time series feature fusion of the order quantity time series, the service time series of the cancelled orders, the rejected order quantity time series, the reservation platform attention time series and the time attribute series is performed to obtain a comprehensive representation of the time series features, specifically including: The time series of received orders, the time series of cancelled orders during service, the time series of rejected orders, the time series of reservation platform attention, and the time attribute series are all convolved using J convolution kernels of different sizes to obtain J convolution result matrices for each time series; The convolution result matrices corresponding to the same convolution kernel are fused to obtain a single-scale fusion matrix corresponding to each convolution kernel; Fill the single-scale fusion matrices corresponding to J convolution kernels into matrices of equal size; Combine the vectors in the same row of the J padded fusion matrices to obtain several cross-scale combination matrices; According to each of the cross-scale combination matrices, an Attention mechanism is applied to perform cross-scale fusion to obtain a cross-scale fusion matrix; The Attention mechanism is applied to the cross-scale fusion matrix to perform weighted fusion in the time dimension to obtain a comprehensive representation of temporal features.
4. The reservation service return prediction method based on multi-scale time series feature fusion according to claim 3 is characterized in that: The expression of the cross-scale fusion matrix is: in, In the formula, represents the cross-scale fusion matrix; j represents the jth convolution kernel; L M Indicates the maximum number of time steps after padding; Represents the padded fusion matrix corresponding to the jth convolution kernel The tth row in The t-th row in represents the feature vector at the t-th time step; α tj is α t The jth element of t =softmax(score t ), Indicates each The cross-scale combination matrix is obtained by combining the t-th row vector of ; f1(·) and f2(·) are activation functions; W1, W2, B1, B2 represent weight coefficients; J represents the number of convolution kernel types.
5. The reservation service return prediction method based on multi-scale time series feature fusion according to claim 4 is characterized in that: The expression for the comprehensive representation of the timing characteristics is: In the formula, V is the comprehensive representation of the time series characteristics; β t is the tth element in β, β=softmax(z), Q represents the Query in the Attention mechanism, which is represented by the user's reservation feature information; concat() represents concatenation; MLP represents the perceptron.
6. The reservation service return prediction method based on multi-scale time series feature fusion according to claim 1 is characterized in that: Before executing the step of "inputting the service provider characteristic information, the user reservation characteristic information and the statistical non-time series data into the trained order rejection prediction auxiliary model", the method further includes: training a multi-task learning prediction network, wherein the multi-task learning prediction network includes an order rejection prediction main model and an order rejection prediction auxiliary model; the training process specifically includes: Obtain a sample set of historical reservation service orders and label data corresponding to each sample of historical reservation service orders; the label data includes reservation labels and service labels: the reservation labels include successful reservation and unsuccessful reservation; the service labels include successful reservation and normal service and successful reservation and unsuccessful service; successful reservation and unsuccessful service means that the reservation is successful but the order is destroyed during the service; Obtaining reservation feature sample data for each historical reservation service order sample; The multi-scale time series features of the sample order quantity time series, the sample order cancellation time series during service, the sample order rejection time series, the sample reservation platform attention time series and the sample time attribute series are fused to obtain a comprehensive representation of the sample time series features; Taking the service provider feature sample information, the user reservation feature sample information and the statistical non-time series sample data of each historical reservation service order sample as input, and taking the reservation label corresponding to each historical reservation service order sample as label data, training the rejection prediction auxiliary model; Taking the sample time series feature comprehensive representation of each historical appointment service order sample, the service provider feature sample information, the user appointment feature sample information, the statistical non-time series sample data and the sample rejection prediction result of the rejection prediction auxiliary model as input, and taking the service label corresponding to each historical appointment service order sample as label data, the rejection prediction main model is trained; Calculating the overall loss error of the multi-task learning prediction network; the overall loss error includes the loss error of the order rejection auxiliary task and the loss error of the order destruction main task; When the overall loss error converges or reaches a preset maximum number of training iterations, a trained multi-task learning prediction network is obtained.
7. The reservation service return prediction method based on multi-scale time series feature fusion according to claim 6 is characterized in that: The expression of the overall loss error is: Loss=L nr +a*L rf ; in, Where Loss represents the overall loss error; l rf represents the loss error of the rejection auxiliary task; L nr represents the loss error of the main task; α is the weight of the auxiliary task; y ri is the actual result of rejection of the i-th historical reservation service order sample; y' ri is the rejection prediction result of the i-th historical reservation service order sample; y ni is the actual result of order cancellation of the i-th historical reservation service order sample; y' ni is the prediction result of order cancellation for the i-th historical reservation service order sample; ri Indicates whether the i-th historical reservation service order sample is a sample of unsuccessful reservation; I ni Indicates whether the i-th historical reservation service order sample is a sample of successful booking but unserviced; N represents the number of samples.
8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the reservation service return prediction method based on multi-scale time series feature fusion as described in any one of claims 1 to 7.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the reservation service return prediction method based on multi-scale time series feature fusion described in any one of claims 1 to 7 is implemented.
10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the reservation service return prediction method based on multi-scale time series feature fusion described in any one of claims 1 to 7 is implemented.
