JP7037708B2 - Withdrawal prospect evaluation device - Google Patents

Description

The present invention relates to a withdrawal prospect device for evaluating the prospect of a customer withdrawing from a service that is continuously visited and enjoyed at a service providing place.

There are services such as fitness clubs that provide services to customers who continuously visit the facilities. Customers continue to visit for service, but stop visiting for various reasons. It is important for service providers to identify customers who are likely to withdraw (stop continuous visits) and take measures such as encouraging such customers to continue visiting. The system of Patent Document 1 is used as an example in a membership-based fitness club. This system extracts members whose visit frequency is decreasing as the target audience, and informs the staff of the information about the extracted target audience. It is expected that the staff will speak to such a target person to prevent the target person from withdrawing from the membership. In Patent Document 1, as a specific evaluation method of the visit frequency of members, whether or not they have visited the venue for 14 days or more since the previous visit, the evaluation of the number of times of facility use in the previous month, and the comparison of the number of times of facility use in the previous month and the month before last are compared. (See FIG. 6 of Patent Document 1).

Patent No. 5399164

In the method of evaluating the frequency of visits in Patent Document 1, only the number of days since the previous visit date and the number of times the facility is used each month are evaluated. It is not always possible to fully grasp the tendency of customers to withdraw (withdraw) from the evaluation of the most recent visit date as in the former case. Moreover, in the evaluation of the number of times of facility use per month as in the latter case, only a rough evaluation is made about the visit tendency of customers. There is a need for a means to better understand the withdrawal of customers.

An object of the present invention is to provide a withdrawal prospect evaluation device capable of appropriately evaluating a customer's withdrawal prospect.

The withdrawal prospect evaluation device of the present invention is a device for evaluating the prospect of a customer withdrawing from a service that is continuously visited and enjoyed at a service providing place, and is set to at least one of a plurality of visit intervals and a difference thereof. It is provided with a trained neural network that outputs a prospect value indicating the prospect based on the above.

According to the withdrawal probability evaluation device of the present invention, the expected value indicating the withdrawal probability is calculated based on at least the visit intervals for a plurality of times or the difference thereof. Then, a trained neural network is used to calculate the expected value. Therefore, it is possible to appropriately evaluate the tendency of the visit frequency in the visit history for a plurality of times. Therefore, the prospect of customer withdrawal is appropriately evaluated as compared with the conventional technology in which only the most recent visit date and the number of times of facility use per month are evaluated.

Further, in the present invention, the input value to the neural network may not include the visit interval but include the difference thereof. According to this, the estimated value can be calculated based only on the visit interval and the latter of the differences. Therefore, the expected value can be calculated with a smaller input value than when both the visit interval and the difference are used.

Further, in the present invention, it is preferable that the neural network includes a recurrent neural network whose input value is at least one of the visit interval and the difference thereof. According to this, a neural network suitable for handling time series data is used. Therefore, the expected value can be calculated with high accuracy.

It is a functional block diagram which shows the schematic structure of the withdrawal prediction apparatus which concerns on one Embodiment of this invention. It is a figure explaining the visit interval and the difference thereof input to the withdrawal prediction apparatus of FIG. It is a graph which shows the derivation result of the predicted value by the withdrawal prediction apparatus which concerns on one Example of this invention. It is a graph which shows the derivation result of the predicted value by the withdrawal prediction apparatus which concerns on another Embodiment of this invention. It is a graph which shows the derivation result of the predicted value by the withdrawal prediction apparatus which concerns on still another Embodiment of this invention.

Hereinafter, the withdrawal prediction device 1 (withdrawal probability evaluation device in the present invention) according to the embodiment of the present invention will be described with reference to FIGS. 1 and 2. The withdrawal prediction device 1 is a device that predicts that a member of a membership service such as a fitness club (hereinafter referred to as the service) will withdraw from the service. The withdrawal prediction device 1 is constructed by hardware such as a computer and software such as program data stored in a memory device (hereinafter referred to as a memory) in the computer. These software can be recorded and distributed on various recording media. The computer includes a memory such as a CPU (Central Processing Unit), a ROM (Read-Only Memory) and a RAM (Random Access Memory), and various interfaces such as an input / output interface. Further, various input devices such as a keyboard and a mouse, various output devices such as a display, and other external devices are connected to the computer. In the withdrawal prediction device 1, the hardware executes various information processing such as arithmetic processing and input / output processing according to software. As a result, the following functions of the withdrawal prediction device 1 are realized.

The withdrawal prediction device 1 has a neural network 10 composed of a trained neural network and an output layer 20. The learning of the neural network 10 will be described later. The neural network 10 includes a GRU 11 made of a Gated recurrent unit (hereinafter referred to as GRU), which is a kind of recurrent neural network (hereinafter referred to as RNN), and a fully connected layer 12.

The GRU used for the GRU 11 calculates n input values y ₁ , y ₂ , ..., Y _n from n input values x ₁ , x ₂ , ..., X _n , respectively. It has n intermediate layers to be output. Each intermediate layer performs such an operation using predetermined weights set corresponding to each of the input values x ₁ to x _n . In an RNN such as GRU, it is assumed that the input values are values that are arranged in the order of x ₁ → x ₂ →… → x _n from the past to the future (that is, time series data). Will be done. The output value of the immediately preceding intermediate layer is input to each intermediate layer of the RNN. For example, when i is a natural number of 2 or more and n or less, the i-th intermediate layer has the i-th input value x _i and the output value of the previous intermediate layer y _i-1 . Both are entered. The i-th intermediate layer calculates y _i from these input values and outputs it. In this way, the RNN is a neural network that can transmit the influence of past input values to the future by sequentially outputting the calculation results of the previous intermediate layer to the later intermediate layer in time. It is a network. Further, the GRU has an update gate in each intermediate layer, so that the influence of the past input value can be transmitted to the future intermediate layer over a long period of time.

In GRU11, a plurality of input values indicating member information regarding each member who is the target of withdrawal prediction are input. These input values are input to the withdrawal prediction device 1 via a recording medium such as a USB (Universal Serial Bus) memory or a communication line such as the Internet, and are associated with the member identification information such as the membership number to be associated with the withdrawal prediction device 1. It is stored in the memory etc. in the computer. After that, the input value is taken out from the memory or the like and input to the GRU 11. It should be noted that these input values may be input to the withdrawal prediction device 1 by other methods such as direct input by the user using an input device connected to a computer.

The input value to GRU11 consists of a value indicating the gender, age, number of consecutive days, purpose of admission, visit interval column, and difference column of visit interval for the member. Input values related to gender indicate gender. The input value related to the age indicates, for example, the age of teens, 20s, 30s, 40s, and so on. The number of continuous days indicates the number of days from enrollment to the present time (when the withdrawal is predicted; the same applies hereinafter). The purpose of membership indicates, for example, the items selected by the member as the purpose of membership from items such as physical fitness improvement, health maintenance, and stress relief. The visit interval column consists of a plurality of numerical values indicating the visit intervals corresponding to the most recent four or more visits from the present time. The visit interval is based on the day when the member visits the facility that provides the service (visit date), and indicates the number of days from one visit date to the next visit date. For example, as shown in FIG. 2, when the last five visit dates are September 2, September 10, September 15, September 24, and September 29, the visit interval column is 8, 5, 9, 5 (September 10-September 2 = 8, September 15-September 10 = 5, September 24-September 15 = 9, September 29-9 24th of the month = 5). The difference column of the visit interval consists of a plurality of numerical values indicating the difference between the visit intervals corresponding to the latest four or more visits. The difference between visit intervals is a value obtained by subtracting the immediately preceding visit interval from a certain visit interval. According to the example of the last 5 visit days, the difference column of the visit interval is -3, 4, -4 (5-8 = -3, 9-5 = 4, 5-9) as shown in FIG. = -4). In GRU11, it is input in the order of gender → age → number of consecutive days → purpose of admission → visit interval column → difference column of visit interval. These correspond to the columns of the above input values x ₁ → x ₂ →… → x _n . Of these, the visit interval column is input to the GRU 11 in the order from the visit interval corresponding to the old visit date to the visit interval corresponding to the new visit date. According to the above example of the last 5 visits, the input is in the order of 8 → 5 → 9 → 5. Further, the difference column of the visit interval is input to the GRU 11 in the order from the difference of the visit interval corresponding to the old visit date to the difference of the visit interval corresponding to the new visit date. According to the example of the last 5 visit days, the input is in the order of -3 → 4 → -4. The GRU 11 performs an operation using a predetermined weight on the input value, and as a result, outputs the same number of output values as the input value to the fully connected layer 12. These output values correspond to the above y ₁ , y ₂ , ..., Y _n .

The fully coupled layer 12 multiplies each of the plurality of output values from the GRU 11 by a predetermined weight, and adds up all the multiplication results to derive one output value and output it to the output layer 20. The output layer 20 performs a predetermined operation on the output value from the fully connected layer 12, and as a result, derives a predicted value indicating the possibility of withdrawal of the member. The operation by the output layer 20 is performed using the activation function. In this embodiment, the sigmoid function is used as the activation function. In this case, assuming that x is the output value from the fully connected layer 12, the output value y of the output layer 20 is represented by y = 1 / (1 + exp (−x)). According to this, the predicted value is in the range of 0 to 1. Note that exp ( ^z ) is equivalent to the exponential function ez. In the present embodiment, the closer the predicted value is to 1, the higher the possibility of withdrawal.

The withdrawal prediction device 1 controls an output device such as a display so as to output the predicted value derived by the output layer 20 together with the member identification information such as the member number. The withdrawal prediction device 1 may be configured so as to extract members who are likely to withdraw based on the predicted value and output the predicted value together with the member identification information to the output device only for the extracted members. .. In this case, for example, a member corresponding to a predicted value higher than a predetermined threshold value may be extracted.

Next, learning of the neural network 10 will be described. The training data used in this learning consists of member information data and withdrawal information data generated based on actual information about a plurality of members in a certain service. The member information data shows the gender, age, number of consecutive days, purpose of admission, visit interval column, and difference column of visit interval of each member. The withdrawal information data is correct answer data indicating withdrawal / continuation of each member. For example, 1 is taken in the case of withdrawal and 0 is taken in the case of continuation. The learning is done by optimizing the weights contained in the GRU 11 and the weights contained in the fully connected layer 12 so that the loss function is minimized. The loss function is a variable between the predicted value output from the output layer 20 as a result of inputting the member information data into the neural network 10 and the correct answer value indicated by the withdrawal information data, and the larger the difference between these values, the larger the value. take. Specifically, a method of updating the weight by the gradient descent method may be used. Also, batch learning or backpropagation may be used.

Hereinafter, the first to third embodiments according to the present invention will be described.

<First Example>
In the first embodiment, the learning of the neural network 10 according to the above-described embodiment and the derivation of the predicted value by the withdrawal prediction device 1 using the learned neural network 10 were carried out. In the learning, as the training data, the member information data and the withdrawal information data generated based on the actual information were used as in the above-described embodiment. In addition, RMSprop was used as an optimization method. The conditions such as hyperparameters used are as follows.
[Learning coefficient] 0.001
[Number of epochs] 100
[Number of GRU11 units] 125
[RMSprop parameters]
-Decay: 0.000001
-Gradient moving average attenuation coefficient (rho): 0.9
[Loss function] Binary cross entropy

Next, using the learned neural network 10, the withdrawal of the member was predicted based on the member information data regarding the actual member. Table 1 shows the results. In Table 1, the section corresponds to the range of predicted values output from the output layer 20. For example, the section of the first line corresponds to a range in which the predicted value exceeds 0.9 and becomes 1 or less. Further, the section on the fourth line corresponds to the range in which the predicted value exceeds 0.6 and is 0.7 or less. The number of members in each section indicates the total number of members whose predicted value falls within the range corresponding to that section. The number of withdrawals indicates the number of members who actually withdrew from the members corresponding to each section. The withdrawal rate indicates the result of dividing the number of withdrawals by the number of members. FIG. 3 is a graph showing the withdrawal rate for the section of Table 1. As shown in Table 1 and FIG. 3, it is shown that the predicted value has a size that largely reflects the actual withdrawal rate.

[Table 1]

<Second Example>
In the second embodiment, a neural network different from that of the first embodiment was used to learn the neural network and to derive the predicted value using the trained neural network. The neural network of the second embodiment is different from the neural network 10 in that the input values consist of values indicating gender, age, number of consecutive days, purpose of admission, and visit interval sequence. That is, in the neural network according to the second embodiment, only the former visit interval column is used among the visit interval sequence and the difference sequence of the visit interval. Therefore, the learning was performed using the data of each member's gender, age, number of consecutive days, purpose of admission, and visit interval column as member information data. The other implementation conditions in the second embodiment are the same as those in the first embodiment.

Then, using the learned neural network, the withdrawal of the member was predicted based on the same member information data as in the first embodiment. Table 2 shows the results. Each item in Table 2 corresponds to each item in Table 1. FIG. 4 is a graph showing the withdrawal rate for the section of Table 2. As shown in Table 2 and FIG. 4, it is shown that the predicted value has a size that largely reflects the actual withdrawal rate.

[Table 2]

<Third Example>
In the third embodiment, a neural network different from that of the first embodiment was used to learn the neural network and to derive the predicted value using the trained neural network. The neural network of the third embodiment is different from the neural network 10 in that the input value consists of a value indicating a difference column of gender, age, number of consecutive days, purpose of admission, and visit interval. That is, in the neural network according to the third embodiment, only the latter difference column of the visit interval is used among the visit interval sequence and the difference sequence of the visit interval. Therefore, the learning was performed using the data of the difference column of each member's gender, age, number of consecutive days, purpose of admission, and visit interval as member information data. The other implementation conditions in the third embodiment are the same as those in the first embodiment.

Then, using the learned neural network, the withdrawal of the member was predicted based on the same member information data as in the first embodiment. Table 3 shows the results. Each item in Table 3 corresponds to each item in Table 1. FIG. 5 is a graph showing the withdrawal rate for the section of Table 3. As shown in Table 3 and FIG. 5, it is shown that the predicted value has a size that largely reflects the actual withdrawal rate.

[Table 3]

According to the present embodiment described above, a predicted value indicating the possibility of withdrawal of a member is calculated based on at least one of the latest plurality of visit intervals and the difference thereof. Then, the trained neural network 10 is used to calculate the predicted value. Therefore, it is possible to appropriately evaluate the tendency of the frequency of visits in the history of visits for a plurality of times. Therefore, as shown in the above-mentioned embodiment, the possibility of withdrawal of the member is appropriately evaluated as compared with the conventional technique in which only the most recent visit date and the number of times of facility use per month are evaluated.

As shown in the above-described embodiment, both the visit interval and the difference thereof may be used for the calculation of the predicted value, or only one of them, for example, only the difference may be used. When only one of the visit interval and the difference is used, the predicted value can be calculated with a smaller input value than when both the visit interval and the difference are used.

In this embodiment, a recurrent neural network, specifically GRU, is used as the neural network. Recurrent neural networks are suitable for handling time series data. Therefore, as shown in the above-mentioned embodiment, the predicted value can be calculated with high accuracy.

<Modification example>
The above is a description of a preferred embodiment of the present invention, but the present invention is not limited to the above-described embodiment, and various changes can be made to the extent described in the means for solving the problem. It is possible.

For example, in the above-described embodiment, a recurrent neural network, specifically a GRU, is used as the neural network. However, other recurrent neural networks such as LSTM (Long short-term memory) may be used, or neural networks other than recurrent neural networks such as convolutional neural networks may be used. Further, in the above-described embodiment, RMSprop is used as the optimization method, but other optimization methods such as Adagrad and Adadelta may be used.

Further, in the above-described embodiment, it is assumed that the present invention is used for a membership service such as a fitness club. However, the invention may be used for other services. For example, the present invention is applied to various services such as elementary and junior high schools, various schools such as cram schools and preparatory schools for entrance examinations, hairdressing and beauty, medical care, etc., in which customers continuously visit the service providing place and enjoy the service. You can do it. That is, in such various services, the device according to the present invention can be applied when evaluating the possibility that the customer will stop the continuous visit (withdraw from the service).

1 Withdrawal prediction device 10 Neural network 11 GRU
12 Fully coupled layer 20 Output layer

Claims (3)

It is a device that evaluates the possibility that customers will withdraw from the services that they enjoy by continuously visiting the place where the service is provided.
A withdrawal probability evaluation device comprising a trained neural network that outputs a prospect value indicating the prospect based on at least one of a plurality of visit intervals and a difference thereof. The withdrawal probability evaluation device according to claim 1, wherein the input value to the neural network does not include the visit interval but includes the difference thereof. The neural network
The withdrawal probability evaluation device according to claim 1 or 2, wherein the recurrent neural network having at least one of the visit interval and the difference thereof as an input value is included.

Images