TWI872368B - A training method of tour group registration prediction model and a system of tour group registration prediction - Google Patents

Abstract

This invention is clustering to tour group data which are collected to judge registration popularity, according to registration number, and clustering tour group data to groups in which all tours in one group have similar popularities. Because all tour group data have a lot of advertisement features, such as price, date, travel country, scenic spot, travel period, accommodation hotel and other information. A tour group registration prediction model which is trained by tour group advertisement of tour group cluster is able to judge registration popularity of tour group advertisement. Then, a CPU trains to enrolled number data of tour group cluster and obtains a tour group registration prediction number model which is able to predict registration number.

Description

Training method of tour group registration prediction model and tour group registration prediction system

The present invention is applied to the registration-related business of the tourism industry, and particularly refers to a training method for a tour group registration prediction model and a tour group registration prediction system.

The planning department or product department of each travel agency needs to design travel itineraries that attract tourists, and publish advertisements or arrange personnel to attract customers. When there are no customers participating in the travel itinerary or the number of people in the group does not reach their preset number, the travel agency will change the price, departure date, etc. of the travel itinerary, or even cancel the travel itinerary, resulting in losses for the travel agency and tourists. These losses include the loss of hotels, meals, tickets, human resources, etc. in foreign countries of the travel agency, and cause confusion in the arrangements of customers in their itineraries and company work.

In order to avoid the above losses, it is necessary to predict the tour group formation rate and the number of registrations in advance. However, this has always been a difficult issue. It is difficult for staff to make predictions based on experience. Therefore, it is necessary to find a way to use computers to assist in improving the accuracy of tour group predictions.

In view of the shortcomings of the prior art, the inventors propose the following solutions:

A tour group registration prediction system includes: a registration popularity prediction model and a processor. The processor determines the registration popularity of the tour group advertisement data according to the tour group advertisement data and the registration popularity prediction model to obtain a popularity judgment result.

A method for training a tour group registration prediction model includes: a processor groups a plurality of tour group data based on registration number data to obtain a plurality of tour group groups; wherein the tour group data respectively include registration number data and tour group advertising data; and the processor trains a first prediction model based on the tour group groups and the tour group advertising data to obtain a registration popularity prediction model.

The invention mainly performs grouping based on the collected tour group data. The grouping method is to judge the registration popularity of the tour group based on the number of registrations, and classify tour groups with similar registration popularity into the same tour group group. Since each tour group advertising data contains advertising features of various tour groups, such as price, date, travel country, scenic spot, travel period, accommodation hotel and other information, after training the tour group advertising data in each tour group group, the registration popularity prediction model obtained can be used to judge the registration popularity of tour group advertisements. This allows travel agencies to evaluate in advance whether the travel itinerary planning meets customer expectations.

In addition, the number of registrations in each tour group group can be further trained, and the resulting registration number prediction model can be used to predict the number of registrations. For example, the number of registrations on the nth day can be used to predict the number of registrations on the n+1th day, and it can even be predicted whether the tour group can be formed before the registration deadline, so that the travel agency can take corresponding measures in advance, thereby reducing its losses.

Please refer to FIG. 1 , according to one embodiment, the training method of the tour group registration prediction model includes:

Step S01: The processor 10 groups a plurality of tour group data based on the number of registered people data to obtain a plurality of tour group groups; the tour group data respectively include a number of registered people data, a registration period, and a tour group advertisement data. Among them, after the tour group advertisement is published, whether or not the tour group is finally formed, it can be used as the tour group data.

Among them, the registration number data includes the number of daily registrations, the cumulative number of daily registrations, the number of registration growth, the number of registration growth rate, etc. The number of registration growth can be the number of registration growth in the next two days, or the number of registration growth in the next week. Similarly, the number of registration growth rate can be the number of registration growth rate in the next two days, or the number of registration growth rate in the next week. In some embodiments, the purpose of grouping the tour group data based on the registration number data is to group the tour group data with similar registration popularity. Therefore, the registration number data can be used as a basis for judging the popularity of the tour group data. Therefore, all those that can be used as a basis for judging the popularity of registration are the registration number data referred to in this manual.

The tour group advertising materials include multiple advertising features. Advertising features are the characteristic information published in the tour group advertisement, and the characteristic information is the information that can affect consumers' decision to sign up for the tour group. Therefore, advertising features include: price, tour date, travel agency reputation, itinerary, number of days, weather, and transportation. The itinerary includes: sightseeing itinerary, consumption itinerary, accommodation, etc. All information that can influence consumers' willingness to sign up for the tour group is the advertising features referred to in this manual. For example, if the ad feature is price, the price can be set as low price (0~20000), medium price (20001~40000), and high price (40001~60000) according to the cost range; if the ad feature is tour date, the tour date can be divided into peak season, summer vacation, winter vacation, and Chinese New Year; if the ad feature is tourist itinerary, the tourist itinerary can be divided into Europe, Japan, and Australia; if the ad feature is weather, the Weather can be subdivided into rainy season and typhoon; when the advertisement feature is the number of travel days, the travel days can be subdivided into short-term (1-3 days), medium-term (4-5 days), and long-term (6-9 days); when the advertisement feature is a consumption itinerary, the consumption itinerary can be subdivided into shopping and self-funded; when the advertisement feature is transportation, the transportation can be subdivided into whether there is a transfer, early or late flights, and low-cost airlines; when the advertisement feature is transportation, the accommodation can be subdivided into the city center, five-star hotels, etc. The above-mentioned various subdivisions are for reference only and are not limited to them in actual implementation. They can also be adjusted, reset, or changed according to user needs. The above tour group data is expressed in mathematical form as: D={ , ,…, },in Represents a certain tour group data. Each tour group data includes the number of registered persons and tour group advertising data. ,in For tour group advertising materials, The registration start date, The registration deadline is For the number of applicants, Indicates whether the tour group is formed. In addition, assume that the registration data of each group On day j there are Family groups register, the information is: =( .like On day j there are Individual registration: = .

The clustering method is based on the number of registration data. The clustering method can be: k-means, hierarchical, DB-scan, Gaussian mixture model GMM, Fuzzy, k-means ++ (low-dimensional), kernal k means (high-dimensional). In some embodiments, k-means is used for clustering. In the clustering process, the number of clusters of the tour group data is first determined, and then the basis for clustering is determined, such as Euclidean distance, cosine distance, etc. In some embodiments, cosine distance is used for clustering. In this way, after k-means clustering, multiple tour group clusters will be obtained, each tour group cluster includes multiple tour group data, and the number of registration data between each tour group data is similar, indicating that the popularity of registration between each other is similar. In addition, these tour group groups are sorted to show the different popularity of each tour group group. For example, these tour group groups may include: first tour group group, second tour group group, third tour group group, etc. Therefore, the registration popularity of the tour group data belonging to the first tour group group must be higher than that of the tour group data belonging to the second tour group group. Therefore, the registration popularity of each tour group data can be understood through the tour group group. In addition, it is worth mentioning that the tour group data that has not been formed will also be classified as a tour group group to indicate that the registration popularity is extremely low and it is difficult to form a group.

In some embodiments, step S02 is performed before step S01: the processor 10 pre-processes the data of a plurality of tour groups. After the tour group data is pre-processed, the grouping effect of step S01 can be improved.

In some embodiments, the pre-processing includes one or both of registration period normalization and registration number normalization. First, registration period normalization is to unify the registration period of each tour group, and the unification method is to reduce or extend the registration period of tour group data with different registration periods based on the registration period of one tour group data. In some embodiments, the registration period of the tour group data with the longest registration period is extended in equal proportion to the registration period of the remaining tour group data. The extension method is to extend the registration period in equal proportion according to the number of daily progressive registrations. For example, in Table 1 below, when the registration period of one of the tour group data is seven days, the number of daily progressive registrations is 0 on the first and second days, 18 on the third to fifth days, and 41 on the sixth and seventh days. Now, when the registration period is uniformly extended to 40 days, since the number of progressive registrations on the first and second days is the same and is the same as the registration period, ,due to Therefore, the first and second days will be extended to 11 days, and the number of progressive registrations per day will be 0 as shown in Table 2 below; similarly, the number of progressive registrations from the third to the fifth day will be the same, and they are also the registration period. ,due to , so the first and second days will be extended to 17 days, and the progressive number of registrations per day is 18 as shown in Table 2 below; the progressive number of registrations on the seventh and eighth days is the same, and the remaining days are Therefore, the seventh and eighth days will be extended to 12 days, and the progressive number of registrants per day is 41 as shown in Table 2 below. Registration days Day 1 the next day Day 3 Day 4 Day 5 Day 6 Day 7 Number of applicants 0 0 18 18 18 41 41 Table 1 Registration days Day 1 the next day … Day 11 Day 12 Day 13 Number of applicants 0 0 0 0 18 18 Registration days … Day 28 Day 29 Day 30 … Day 40 Number of applicants 18 18 41 41 41 41 Table 2

Next, the number of applicants is standardized to reduce the difference in the number of applicants, thereby improving the effect of grouping. In some embodiments, this is accomplished using a statistical standardization formula, which is: ,in for The standardized results, is the average value of the population, is the standard deviation of the population. In some embodiments, the Gaussian function is used for standardization, and the Gaussian function used is , where a, b, c are constants, and a＞0, and x is the number of applicants.

Step S03: The processor 10 trains a first prediction model according to the tour group groups and the tour group advertising data to obtain a registration popularity prediction model 20. Since the registration popularity of each tour group data in the same tour group group is similar, it also means that the tour group advertising data between the tour group data will also be similar. Therefore, after training each tour group group with the tour group advertising data, the characteristics of each tour group group can be obtained, and these characteristics can be used to predict the registration popularity of the travel itinerary designed by the tour group. For example, when a travel agency launches a new tour itinerary, the processor 10 can determine which tour group the tour itinerary is similar to based on the advertising features of the tour itinerary, such as price, departure date, travel agency popularity, itinerary, number of days, weather, traffic, etc., and then determine the popularity of the tour itinerary based on the tour group group.

In some embodiments, the processor uses named entity recognition (NER) to extract advertising features from the travel itinerary, such as price, departure date, travel agency popularity, itinerary, number of days, weather, traffic, etc., and then uses fuzzy matching to delete similar advertising features from the advertising features proposed by the named entity recognition method, and then uses slot filling to match the advertising features. Filling) generates feature sentences. Among them, named entity recognition is to identify entities with specific meanings in the travel itinerary. Taking this manual as an example, the specific and meaningful entities are price, departure date, travel agency popularity, itinerary, number of days, weather, traffic, etc.; string fuzzy matching is a method used to identify the similarity between strings and match the most similar strings. When used in conjunction with named entity recognition, it can more effectively extract advertising features in the travel itinerary; as for the feature sentence, for example: the departure date is April 6, 2022; the destination (country or city) is D1, D2 and D3; the attractions are A1, A2 and A3; the travel fee is about US$2,000; the flight company is F1; the flight takes off at 5 pm; the travel agency is T; the number of days in the itinerary is 12 days; the hotels are H1, H2, H3, H4 and H5.

After obtaining the advertisement features, the processor 10 trains the first prediction model using the advertisement features as input and the tour group groups as output. In some embodiments, the training process is performed using a bidirectional encoder representation technique based on transformers (BERT) in combination with a convolutional neural network (CNN). The bidirectional encoder representation technique based on transformers is used for the conversion of word vectors. In some embodiments, the convolutional neural network sequentially includes two convolution layers, a pooling layer, a flatten layer, and two fully connected layers. In addition, in some embodiments, when the processor 10 trains the first prediction model, the number of neurons in its hidden layer is 256.

The above is an introduction to the training method of the registration popularity prediction model 20, which is used to predict the registration popularity of a new travel itinerary. The next is an introduction to the training method of the registration number prediction model 30. The registration number prediction model 30 allows travel agencies to directly predict the number of registrations for each subsequent day based on the number of registrations on the nth day. In some embodiments, it can even predict the number of registrations on the day when the registration deadline ends, or predict that the number of registrations on the nth day will reach the threshold of the number of people for the group. Therefore, it can be predicted whether the travel itinerary can be formed, so that the travel agency can take corresponding measures in advance, thereby reducing its losses.

Step S04: The processor 10 trains a second prediction model according to the grouping of the tour groups and the registration number data to obtain a registration number prediction model 30. Using the registration number prediction model 30, the processor 10 can predict the registration number of the n+1 day according to the registration number of the previous n days, and can even predict in advance whether a tour group can be formed. The training method is to use the daily number of people data of any two days in the registration number data as input and output respectively, and the daily number of people data includes the daily number of people and the daily number of people cumulative. In some embodiments, the daily number of people cumulative for two consecutive days is used for training, such as using the daily number of people cumulative for the nth day as input and the daily number of people cumulative for the n+1th day as output for training, and the number of neurons in the hidden layer is 50. In some embodiments, the second prediction model is trained using Long Short-Term Memory (LSTM).

The above introduces the training methods of the registration popularity prediction model 20 and the registration number prediction model 30. Next, the related application systems will be introduced.

Please refer to FIG. 2 , which is a tour group registration prediction system, comprising: a registration popularity prediction model 20, a registration number prediction model 30, and a processor 10. The processor 10 determines the registration popularity of the tour group advertisement data according to the tour group advertisement data and the registration popularity prediction model 20, and obtains a popularity determination result.

Therefore, when a travel agency wants to launch a travel itinerary, as long as the feature data is input, the processor 10 will use the registration popularity prediction model to determine which tour group the travel itinerary belongs to based on the feature data, and the popularity of the travel itinerary can be predicted based on the tour group. Even when the travel itinerary is classified into a tour group that cannot be formed, it can be determined that the travel itinerary will not be formed, and the travel agency can modify the itinerary as soon as possible.

Then, when the travel itinerary is launched, a registration number data is generated based on the number of applicants. In some embodiments, the daily registration number accumulation is used as the registration number data. Then, the processor 10 predicts the number of applicants based on the tour group advertising data, the registration number data and the registration number prediction model 30 to obtain a registration number prediction value. For example, when the daily registration number accumulation of the nth day is input, the processor 10 can use the registration number prediction model 30 to predict the daily registration number accumulation of the n+1th day, and even the processor 10 can predict the daily registration number accumulation of the registration deadline. This allows the travel agency to know in advance whether the travel itinerary can be formed, so as to carry out corresponding contingency processing methods in advance.

In some embodiments, the processor 10 obtains a group formation rate prediction value based on a minimum number of group members and the progressive registration number prediction value. The group formation rate prediction value is used to concretely present the probability of group formation in a numerical value, and can also specifically let travel agency personnel understand the probability of group formation of the travel itinerary.

Although the present invention has been disclosed as above by way of embodiments, it is not intended to limit the present invention. Any person with ordinary knowledge in the relevant technical field may make some modifications and changes without departing from the spirit and scope of the present invention. Therefore, as long as these modifications and changes are within the scope of the attached patent application and its equivalent, the present invention will also cover these modifications and changes.

S01~S04: Steps 10: Processor 20: Registration popularity prediction model 30: Registration number prediction model

FIG1 shows a flow chart of the steps of the tour group registration prediction model training method of the present invention. FIG2 shows a functional block diagram of the tour group registration prediction system of the present invention.

S01~S04: Steps

Claims (5)

A method for training a tour group registration prediction model comprises: a processor groups a plurality of tour group data based on registration number data to obtain a plurality of tour group groups; wherein the tour group data respectively include the registration number data and tour group advertising data; the processor trains a first prediction model based on the tour group groups and the tour group advertising data to obtain a registration popularity prediction model; in the step of the processor grouping the plurality of tour group data based on the registration number data, the processor trains a first prediction model based on the tour group groups and the tour group advertising data to obtain a registration popularity prediction model; The tour group data are grouped based on the registration number data and whether the tour group is formed; the tour group advertising data includes multiple advertising features; in the step of the processor training a prediction model based on the tour group groups and the tour group advertising data, the processor trains the first prediction model using the advertising features as input and the tour group groups as output; and the processor trains a second prediction model based on the tour group groups and the registration number data to obtain a registration number prediction model. The method for training a tour group registration prediction model as described in claim 1, wherein the registration number data includes daily number data; in the step of the processor training the second prediction model according to the tour group groups and the registration number data, the daily number data of two days are taken as input and output respectively to train the second prediction model. A method for training a tour group registration prediction model as described in claim 2, wherein before the processor performs a step of grouping multiple tour group data based on the number of registration data, the number of registrations of the tour group data is standardized. A method for training a tour group registration prediction model as described in claim 2, wherein the processor proportionally shrinks or expands the registration days of the tour group data before grouping the tour group data based on the registration number data. A tour group registration prediction system includes: a registration popularity prediction model; a processor, judging the registration popularity of the tour group advertising data and the registration popularity prediction model, and obtaining a popularity judgment result; a registration number prediction model, the processor obtaining a registration number prediction value based on a registration number data, the tour group advertising data and the registration number prediction model; and the processor obtaining a group formation rate prediction value based on a minimum number of group members and the registration number prediction value.

Images