TWI735932B - Real estate potential customer forecasting system and method thereof - Google Patents

Abstract

A real estate potential customer forecasting system and a method thereof are provided. The real estate potential customer forecasting system includes a potential customer database, a customer audio and video interactive module, a deep learning engine, a maturity level multi-level array module and an object search module. The potential customer database stores a plurality of potential customer information. The customer audio and video interactive module generates a plurality of customer interaction information. The deep learning engine is connected to the potential customer database and the customer audio and video interactive module, and receives the plurality of potential customer information and the plurality of customer interaction information and correspondingly generates a plurality of potential customer prediction information. The maturity multi-level array module connects the deep learning engine and receives a plurality of potential customer prediction information, and generates a list of potential customer desires corresponding to each of potential customer prediction information. The object search module is connected to the maturity level multi-level array module and receives a plurality of potential customer desire lists, and compares the object information corresponding to each potential customer desire list.

Description

Real estate potential customer prediction system and method

The present invention relates to a forecasting system and method, in particular to a real estate potential customer forecasting system and method.

Traditional real estate potential customer forecasting methods mostly use manual methods to analyze traffic and customer consumption behavior, or use some simple tools to assist the analysis. This method is a very labor-intensive and time-consuming task. Generally speaking, it takes a lot of cost to collect a large amount of potential customer information. Therefore, salespersons usually make guesses based on the accumulated professional knowledge over the years or the list of customers that may be obtained in the past. However, after all, it is difficult to grasp the timing of changes in real estate market prices and buying and selling demand, and consumers' behaviors from planning to actual purchase vary in length. It is often seen that potential buyers are pre-planned through the analysis of real-price login data, and they find that the customer has already passed. The original desire to buy has caused a gap between real demand and expectations, causing salesmen or intermediaries to face the following situations:

1. Affected by the increasing importance of the concept of personal data protection, customer information is not easy to obtain, let alone the information of customers who have the ability to buy and sell real estate in real estate, so there is not enough information on hand to accurately predict potential customers.

2. Salespersons generally speculate based on past experience and statistical analysis of existing customer data, and cannot know the potential demand in detail.

The disadvantages of conventional technology are:

1. The cost of data sources required to establish a potential customer model is not only very high, and the identification of data characteristics may not be universally applicable to every demand for mining potential customers.

2. Traditional analysis and prediction methods are not only susceptible to the distribution of data characteristics, but also easy to cause overfitting. Often, a large gap occurs when a trained model is put into another set of data for verification.

3. The salesperson cannot accurately predict the buyer's willingness to buy a house while also taking into account the timing of the seller's bidding for the building.

4. The potential customer model is difficult to define the true meaning of the potential. In other words, it is impossible to precisely define the customers who may buy and sell real estate in the near future. Therefore, it is impossible to accurately deliver to the right customers and often waste a lot of marketing budget costs.

In view of the above-mentioned conventional problems, the purpose of the present invention is to provide a real estate potential customer prediction system and method to solve the problems faced by the conventional technology.

Based on the foregoing objectives, the present invention provides a real estate potential customer prediction system including a potential customer database, a customer audio-visual interaction module, a deep learning engine, a maturity style multi-level queue module, and an object search module. The potential customer database stores multiple potential customer information. The customer audio-visual interaction module generates a plurality of customer interaction information. The deep learning engine links the potential customer database and the customer audio-visual interaction module, and receives multiple potential customer information and multiple customer interaction information and correspondingly generates multiple potential customer prediction information. The maturity style multi-level queue module is connected to the deep learning engine and receives multiple potential customer forecast information, and generates a potential customer desire list corresponding to each potential customer forecast information. The object search module connects the maturity style multi-level queue module and receives a plurality of potential customer desire lists, and compares the object information corresponding to each potential customer desire list.

Preferably, the potential customer database can be linked to the regional smart distribution module. The regional smart distribution module generates control parameters based on performance responsibility distribution and transmits the control parameters to the retractable range locking module. The retractable range locking module adjusts according to the control parameters. In the mining cycle, the real estate registration transaction comparison module compares the real estate registration information corresponding to the mining cycle to generate multiple potential customer information, and transmits multiple potential customer information to the potential customer database.

Preferably, the deep learning engine may include a two-way sliding window low-level feature extraction module and a high-level feature fuzzy label module. The two-way sliding window low-level feature extraction module generates multiple sample information based on multiple potential customer information. The feature fuzzy label module generates multiple pieces of real information based on multiple pieces of customer interaction information.

Preferably, the deep learning engine may include an optimized training module, which connects the two-way sliding window low-level feature extraction module and the high-level feature fuzzy label module, and receives a plurality of sample information and a plurality of real information, and determines Each sample information or each real information received comes from a two-way sliding window low-level feature extraction module or a high-level feature fuzzy label module.

Preferably, when the optimization training module cannot determine that the received sample information is from the low-level feature extraction module of the two-way sliding window, it means that the training of the low-level feature extraction module of the two-way sliding window is completed.

Based on the above objective, the present invention further provides a real estate potential customer prediction method, which includes the following steps: storing a plurality of potential customer information in a potential customer database. A plurality of customer interaction information is generated by the customer audio-visual interaction module. A plurality of potential customer information and a plurality of customer interaction information are received by the deep learning engine, and a plurality of potential customer prediction information is generated correspondingly. The maturity style multi-level queue module receives multiple potential customer forecast information, and generates a potential customer desire list corresponding to each potential customer forecast information. A plurality of potential customer desire lists are received by the object search module, and the object information corresponding to each potential customer desire list is compared.

Preferably, the potential customer database can be linked to the regional smart distribution module. The regional smart distribution module generates control parameters based on performance responsibility distribution and transmits the control parameters to the retractable range locking module. The retractable range locking module adjusts according to the control parameters. In the mining cycle, the real estate registration transaction comparison module compares the real estate registration information corresponding to the mining cycle to generate multiple potential customer information, and transmits multiple potential customer information to the potential customer database.

Preferably, the deep learning engine may include a two-way sliding window low-level feature extraction module and a high-level feature fuzzy label module. The two-way sliding window low-level feature extraction module generates multiple sample information based on multiple potential customer information. The feature fuzzy label module generates multiple pieces of real information based on multiple pieces of customer interaction information.

Preferably, the deep learning engine may include an optimized training module, which connects the two-way sliding window low-level feature extraction module and the high-level feature fuzzy label module, and receives a plurality of sample information and a plurality of real information, and determines Each sample information or each real information received comes from a two-way sliding window low-level feature extraction module or a high-level feature fuzzy label module.

Preferably, when the optimization training module cannot determine that the received sample information is from the low-level feature extraction module of the two-way sliding window, it means that the training of the low-level feature extraction module of the two-way sliding window is completed.

Based on the above, the real estate potential customer prediction system and method of the present invention have the following advantages:

1. The present invention establishes the data sources required for automatic online parallel retrieval and analysis to establish the initial model of potential customers, without manual intervention to compare real estate registration transaction data, which saves a lot of personnel costs, and the range of data is very broad and may cover every type of mining The needs of potential customers.

2. With deep learning methods, changes in the distribution of data characteristics can be quickly adapted, and overfitting can be avoided.

3. Through mature operation and processing, the brewing process of the wish list provides a more accurate and correct timing for house purchases, and the salesperson can accurately predict the buyer's willingness to buy a house while also taking into account the timing of the seller's bidding for the building.

4. The present invention can use the deep learning technology to accurately target the right customers, which can not precisely define the customers who may buy and sell real estate in the near future with the conventional technology. In addition to narrowing the scope of objects to reduce advertising costs, it can more effectively promote marketing. Performance.

In order to understand the features, content and advantages of the present invention and its achievable effects, the present invention is combined with the figures and described in detail in the form of an embodiment as follows. The figures used therein are only intended to The schematic and auxiliary instructions are not necessarily the true proportions and precise configurations after the implementation of the present invention. Therefore, the proportions and configuration relationships of the attached drawings should not be interpreted as to limit the scope of rights of the present invention in actual implementation.

As shown in Figures 1 to 4; Figure 1 is a block diagram of the real estate potential customer prediction system of the present invention. Figure 2 is a schematic diagram of the potential customer maturation method of the real estate potential customer prediction system of the present invention. Figure 3 is a schematic diagram of the fuzzy labeling method of the real estate potential customer prediction system of the present invention. Figure 4 is a schematic diagram of the optimized training framework of the real estate potential customer prediction system of the present invention. As shown in the figure, the real estate potential customer prediction system 100 of the present invention includes a potential customer database 110, a customer audio-visual interaction module 120, a deep learning engine 130, a maturity style multi-level queue module 140, and an object search module 150 .

In addition, the potential customer database 110 stores a plurality of potential customer information. The customer audio-visual interaction module 120 generates a plurality of customer interaction information.

However, the deep learning engine 130 connects the potential customer database 110 and the customer audio-visual interaction module 120, and receives a plurality of potential customer information and a plurality of customer interaction information and correspondingly generates a plurality of potential customer prediction information.

However, the maturity style multi-level queue module 140 is connected to the deep learning engine 130 and receives a plurality of potential customer prediction information, and generates a potential customer desire list corresponding to each potential customer prediction information.

Among them, as shown in Figure 2, according to the potential customer maturity method, the present invention designs a data structure of a potential customer's maturity pattern multi-level queue and records its process, and uses the fuzzy competitive price level as the maturity method multi-level queue. The price competition level channel of the data input listed. According to the potential customer maturity method, the quality of the input data is sorted in a queue, and other price levels are added or reduced according to desires or other price levels are input at the same time. Finally, the mature potential customer input is compared with the actual announcement to find suitable objects, so that the salesperson can instantly Recommend real estate and complete the transaction according to your personal situation.

For example, according to the level of competition, the price levels are more than 50 million, more than 20 million, more than 10 million, more than 700 million, and less than 5 million, and there are a total of 5 price competition levels. Each price competition level pipeline adopts the first in first out (FIFO) method. According to the rules of the quality queue sorting maturity method, continue to wait for the mature output. When the maturity has reached the stage where the output is necessary, it means that the actual announcement of the suitable object has been found and the actual transaction has been found. At this time, the output is potential The customer list contains prices and categories.

Each value represents an individual of a potential customer, and the content includes items that potential customers desire, which can be single or multiple. The mature method of sorting each individual according to the quality column must be ranked in the last quality column, and multiple fuzzy price competition channels can be entered at the same time. For example, in the dotted box in the second figure, the potential individual with the value 2 has 5 thousand at the same time More than 10,000 "Tootian" and more than 20 million "Villa" desire items are waiting to be output.

Sometimes potential individuals may increase or decrease their desires, and link up with other fuzzy price levels, as shown by the dotted arrow in Figure 2. For example, a potential individual with a value of 6 originally had a "storefront" item of more than 10 million as the third in the price hierarchy. During the period, the item "triangular window and home" was added due to desire, and the potential individual with a value of 6 increased at the same time. It has more than 20 million "storefront + triangular windows + residential" items.

In addition, taking the example of a reduction in desire items, a potential individual with a value of 23 originally owned an "apartment + parking space" item of more than 7 million, and the item "parking space" was reduced due to desire. After the reduction, a potential individual with a value of 23 also has 5 "Apartment" items below one million. Although each individual must be sorted, the only thing that can surpass the previous ranking is when the transaction price increases. For example, a potential individual with a value of 12 is a potential individual that surpasses the value of 17 after an increase in the transaction price.

However, the object search module 150 is connected to the maturity-style multi-level queue module 140 and receives a plurality of potential customer desire lists, and compares the object information corresponding to each potential customer desire list.

Further, the potential customer database 110 can be connected to the regional smart distribution module 111. The regional smart distribution module 111 generates control parameters based on performance responsibility distribution and transmits the control parameters to the telescopic range locking module 112, and the telescopic range locking module 112 Adjusting the mining cycle according to the control parameters, the real estate registration transaction comparison module 113 compares the real estate registration information corresponding to the mining cycle to generate multiple potential customer information, and transmits the multiple potential customer information to the potential customer database 110.

Among them, the real estate registration transaction comparison module 113 analyzes the key content of public data from various sources, and compares the serial number of the real estate registration object according to the basic customer database collected over the years to determine whether it is necessary to further explore the customer's object Change the details. Simultaneously launch multiple web page capture and analysis tasks to speed up the search speed, and filter out meaningless interactive messages, so that the search process can be batch searched according to the entire section, and if there is an error in the process, it is recorded in the error list. It also regularly launches sales and purchase registration inquiries based on the segments with more frequent changes. After cross-comparing the internal customer database with external data, a potential customer database 110 is formed, which is the source of the potential customer base. After further and sales After personnel interaction and program data exploration process, the process characteristic records are revised in the potential customer database 110.

In addition, the deep learning engine 130 may include a two-way sliding window low-level feature extraction module 131 and a high-level feature fuzzy label module 132. The two-way sliding window low-level feature extraction module 131 generates multiple sample information based on multiple potential customer information. , The high-level feature fuzzy label module 132 generates a plurality of real information according to a plurality of customer interaction information.

Among them, in the low-level feature extraction stage of the two-way sliding window:

After the specific customer data in the potential customer database 110 is intercepted, according to different orderly attributes, such as address, administrative area, age, registration date, etc., the regional characteristics are generated based on these attributes, and then each passing The results of the specific attribute arrangement are combined to form a two-dimensional low-level training set. Then use a convolutional neural network to scan a two-dimensional low-level training set back and forth with a fixed-size sliding window range filter (matrix), and the result is a convolutional layer, which is processed by multiple convolutional layers to maximize pooling The method forms a plurality of pooling layers. At this stage, the final pooling layer matrix is flattened, and finally a specific number of input tensors are formed.

However, the high-order feature fuzzy labeling method stage:

The convolutional neural network generated in the above stage lacks a fully connected layer. The final output tensor is used to record the behavior of the customer's desire change through the manual interaction of the salesperson. As shown in Figure 3, the classification and labeling operations are carried out in a fuzzy manner. , The category of potential customers' desire to purchase objects is fuzzy to determine the price level of their desires in the manner of regional adjacency. Compared with the traditional labeling method (Traditional Labeling Method; TLM), the use of single sequence feature labeling requires a longer calculation time and cost. The present invention uses the Fuzzy Labeling Method (TLM), which uses the method of region adjacent random feature labeling to fuzzy label [1,1,1] in a group of 3 strokes, and uses the Kappa consistency coefficient (K coefficient of agreement) measurement reliability, including Cohen's Kappa or Fleiss's Kappa to verify the consistency of coders, as the basis for changing the category of fuzzy tags when collisions of fuzzy tags occur. Automatically fill up adjacent fuzzy labels. As shown in the dashed box in Figure 3, when the first area adjacent feature 2 numerical cluster [2,2,2] collides with the second domain adjacent feature 1 numerical cluster [1,1,1], and the third domain adjacent feature 2 Numerical cluster [2,2,2], which automatically fills up the adjacent fuzzy labels and generates the consistency evaluation condition, then it will automatically fill up the adjacent features of the first region. 2 Numerical clusters [2,2,2,2] And the second area adjacent feature 1 numerical cluster [1,1,1,1], and then do the consistency evaluation test, when the evaluation is simple collision, use Cohen's Kappa, and use Fleiss's Kappa on the contrary, if the Kappa value is greater than 0.4, It remains the same and does not change, if it is less than or equal to 0.4, it will be changed to the last category. The fuzzy feature labeling method is used to achieve the purpose of accelerating feature classification.

For example, some information such as real price registration is not a certain real estate price, but only the average price of a road section area. You can imagine the actual price of each building on the floor of each house. If you know it for sure, you can form some training sets, but usually this is very difficult. Training with real-priced registered data sets should also be the test performance of real-priced registrations, so it cannot completely solve the problem of excessive prediction differences, and it is difficult to solve this problem even with in-depth analysis of data mining. But from another point of view, if we already have part of the real building floor transaction data, will the predicted performance result be the real transaction data, that is, the final result of the so-called real estate appraisal.

In addition, the deep learning engine 130 may include an optimized training module 133, which connects the two-way sliding window low-level feature extraction module 131 and the high-level feature fuzzy label module 132, and receives a plurality of sample information and a plurality of real data. It is determined that each sample information or each real information received comes from the low-level feature extraction module 131 or the high-level feature fuzzy label module 132 of the two-way sliding window.

Among them, as shown in Figure 4, the optimized training framework is composed of a two-way sliding window low-level feature extraction module 131 and an optimized training module 133. The purpose of the two-way sliding window low-level feature extraction module 131 It is to learn the data distribution of the high-level feature fuzzy label module 132 as much as possible, and the purpose of the optimization training module 133 is to correctly determine whether the input data is from the real data of the high-level feature fuzzy label module 132 or from the two-way sliding window. The sample data of the first-order feature extraction module 131; in order to obtain more accurate prediction results, the two need to be continuously optimized. A Nash equilibrium. Therefore, the differential function M and the differential function L can be used to represent the optimization training module 133 and the two-way sliding window low-order feature extraction module 131, respectively. The inputs of the two are real data h and potential customer data d. L(d) is the sample data generated by the differential function L that follows the real data distribution pdata as much as possible. If the input of the optimization training module 133 is real data h, it is marked as 1, and if the input is sample data L(d), it is marked as 0. Here, the goal of the differential function M is to realize the binary classification of the data source, true (the distribution of real data h) or not (the sample data L(d) of low-order feature extraction), and the goal of the differential function L is to make The performance M(L(d)) of the sample data L(d) of the low-order feature extraction on the differential function M is consistent with the performance M(h) of the real data h on the differential function M. These two mutual optimization processes , The differential function M and the differential function L are continuously improved. When the optimal training ability of the final differential function M is improved to a certain level and the data source cannot be correctly identified, it can be considered that the two-way sliding window low-order feature extraction module 131 has been Learn the distribution of real data.

Therefore, when the optimization training module 133 cannot determine that the received sample information comes from the two-way sliding window low-level feature extraction module 131, it means that the training of the two-way sliding window low-level feature extraction module 131 is completed.

Although the foregoing description of the real estate potential customer prediction system of the present invention has also explained the concept of the real estate potential customer prediction method of the present invention, for the sake of clarity, a flowchart is shown below for detailed description.

Please refer to Figure 5, which is a flowchart of the method for predicting potential real estate customers of the present invention. As shown in the figure, the real estate potential customer prediction method of the present invention is suitable for the aforementioned real estate potential customer prediction system. The real estate potential customer prediction method includes the following steps:

In step S51: store a plurality of potential customer information in the potential customer database.

In step S52, a plurality of customer interaction information is generated by the customer audio-visual interaction module.

In step S53, a plurality of potential customer information and a plurality of customer interaction information are received by the deep learning engine, and a plurality of potential customer prediction information is correspondingly generated.

In step S54: the maturity style multi-level queue module receives a plurality of potential customer forecast information, and generates a potential customer desire list corresponding to each potential customer forecast information.

In step S55: receiving a plurality of desire lists of potential customers through the object searching module, and comparing the object information corresponding to the desire lists of each potential customer.

Furthermore, the potential customer database can be linked to regional smart distribution modules. The regional smart distribution modules generate control parameters based on performance responsibility distribution and send the control parameters to the telescopic range locking module. The telescopic range locking module adjusts according to the control parameters. In the mining cycle, the real estate registration transaction comparison module compares the real estate registration information corresponding to the mining cycle to generate multiple potential customer information, and transmits multiple potential customer information to the potential customer database.

In addition, the deep learning engine can include a two-way sliding window low-level feature extraction module and a high-level feature fuzzy label module. The two-way sliding window low-level feature extraction module generates multiple sample information based on multiple potential customer information. The feature fuzzy label module generates multiple pieces of real information based on multiple pieces of customer interaction information.

On the other hand, the deep learning engine can include an optimized training module, which connects the two-way sliding window low-level feature extraction module and the high-level feature fuzzy label module, and receives a plurality of sample information and a plurality of real information, and judges Each sample information or each real information received comes from a two-way sliding window low-level feature extraction module or a high-level feature fuzzy label module.

In addition, when the optimization training module cannot determine that the received sample information is from the low-level feature extraction module of the two-way sliding window, it means that the training of the low-level feature extraction module of the two-way sliding window is completed.

The detailed description and implementation of the real estate potential customer prediction method of the present invention have been described in the foregoing description of the real estate potential customer prediction system of the present invention, and will not be repeated here for the sake of brief description.

Based on the above, the real estate potential customer prediction system and method of the present invention have the following advantages:

1. The present invention establishes the data sources required for automatic online parallel retrieval and analysis to establish the initial model of potential customers, without manual intervention to compare real estate registration transaction data, which saves a lot of personnel costs, and the range of data is very broad and may cover every type of mining The needs of potential customers.

2. With deep learning methods, changes in the distribution of data characteristics can be quickly adapted, and overfitting can be avoided.

3. Through mature operation and processing, the brewing process of the wish list provides a more accurate and correct timing for house purchases, and the salesperson can accurately predict the buyer's willingness to buy a house while also taking into account the timing of the seller's bidding for the building.

4. The present invention can use the deep learning technology to accurately target the right customers, which can not precisely define the customers who may buy and sell real estate in the near future with the conventional technology. In addition to narrowing the scope of objects to reduce advertising costs, it can more effectively promote marketing. Performance.

The above-mentioned embodiments are only to illustrate the technical ideas and features of the present invention, and their purpose is to enable those who are familiar with the art to understand the content of the present invention and implement them accordingly. When they cannot be used to limit the patent scope of the present invention, That is, all equal changes or modifications made in accordance with the spirit of the present invention should still be covered by the patent scope of the present invention.

100: Real estate potential customer prediction system 110: Potential customer database 111: Regional Smart Distribution Module 112: Telescopic range locking module 113: Real estate registration transaction comparison module 120: Customer audio-visual interactive module 130: Deep Learning Engine 131: Two-way sliding window low-level feature extraction module 132: High-level feature fuzzy label module 133: Optimized training module 140: Maturity style multi-level queue module 150: Object Finding Module S51 to S55: steps

Figure 1 is a block diagram of the real estate potential customer prediction system of the present invention. Figure 2 is a schematic diagram of the potential customer maturation method of the real estate potential customer prediction system of the present invention. Figure 3 is a schematic diagram of the fuzzy labeling method of the real estate potential customer prediction system of the present invention. Figure 4 is a schematic diagram of the optimized training framework of the real estate potential customer prediction system of the present invention. Figure 5 is a flowchart of the method for predicting potential real estate customers of the present invention.

100: Real estate potential customer prediction system

110: Potential customer database

111: Regional Smart Distribution Module

112: Telescopic range locking module

113: Real estate registration transaction comparison module

120: Customer audio-visual interactive module

130: Deep Learning Engine

131: Two-way sliding window low-level feature extraction module

132: High-level feature fuzzy label module

133: Optimized training module

140: Maturity style multi-level queue module

150: Object Finding Module

Claims (8)

A real estate potential customer prediction system, including: a potential customer database, which stores multiple potential customer information; a customer audio-visual interaction module, which generates multiple customer interaction information; a deep learning engine, which links the potential customer information The library and the customer audio-visual interaction module, and receive the plurality of potential customer information and the plurality of customer interaction information, and correspondingly generate a plurality of potential customer forecast information; a maturity-like multi-level queue module that links the depth The learning engine receives the plurality of potential customer prediction information, and generates a potential customer desire list corresponding to each of the potential customer prediction information; and an object search module that links the maturity pattern multi-level queue module and receives the A plurality of potential customer desire lists are compared, and the object information corresponding to each of the potential customer desire lists is compared; the deep learning engine includes a two-way sliding window low-level feature extraction module and a high-level feature fuzzy label module, The two-way sliding window low-level feature extraction module generates a plurality of sample information based on the plurality of potential customer information, and the high-level feature fuzzy label module generates a plurality of real information based on the plurality of customer interaction information. For example, the real estate potential customer prediction system described in the first item of the scope of patent application, wherein the potential customer database is linked to a regional smart distribution module, and the regional smart distribution module generates a control parameter based on performance responsibility distribution and transmits it The control parameter is to a telescopic range locking module, the telescopic range locking module adjusts the mining cycle according to the control parameter, and a real estate registration transaction comparison module compares the real estate registration information corresponding to the mining cycle to generate the plurality of potential customers News, And send the multiple potential customer information to the potential customer database. Such as the real estate potential customer prediction system described in the scope of patent application 1, wherein the deep learning engine includes an optimization training module that links the two-way sliding window low-level feature extraction module and the high-level feature fuzzy label Module, and receives the plurality of sample information and the plurality of real information, and determines that each of the received sample information or each of the real information comes from the two-way sliding window low-level feature extraction module or the high-level feature fuzzy label Module. For example, the real estate potential customer prediction system described in item 3 of the scope of patent application, wherein when the optimization training module cannot determine that the sample information received is from the two-way sliding window low-level feature extraction module, it means The training of the low-level feature extraction module of the two-way sliding window is completed. A real estate potential customer prediction method includes the following steps: store multiple potential customer information through a potential customer database; generate multiple customer interactive information through a customer video and audio interactive module; receive the multiple information through a deep learning engine A plurality of potential customer information and the plurality of customer interaction information and correspondingly generate a plurality of potential customer prediction information; a maturity-like multi-level queue module receives the plurality of potential customer prediction information and corresponds to each of the potential customer predictions The information generates a potential customer desire list; and through an object search module to receive the plurality of potential customer desire lists, and compare the corresponding object information of each of the potential customer desire lists; wherein the deep learning engine includes a two-way Sliding window low-level feature extraction module and a high-level feature fuzzy label module, the two-way sliding window low-level feature extraction module The group generates a plurality of sample information based on the plurality of potential customer information, and the high-level feature fuzzy label module generates a plurality of real information based on the plurality of customer interaction information. For example, the real estate potential customer prediction method described in item 5 of the scope of patent application, wherein the potential customer database is linked to a regional smart distribution module, and the regional smart distribution module generates a control parameter based on performance responsibility distribution and transmits it The control parameter is to a telescopic range locking module, the telescopic range locking module adjusts the mining cycle according to the control parameter, and a real estate registration transaction comparison module compares the real estate registration information corresponding to the mining cycle to generate the plurality of potential customers Information, and send the plurality of potential customer information to the potential customer database. The real estate potential customer prediction method described in item 5 of the scope of patent application, wherein the deep learning engine includes an optimization training module that links the two-way sliding window low-level feature extraction module and the high-level feature fuzzy label Module, and receives the plurality of sample information and the plurality of real information, and determines that each of the received sample information or each of the real information comes from the two-way sliding window low-level feature extraction module or the high-level feature fuzzy label Module. For example, the real estate potential customer prediction method described in item 7 of the scope of patent application, wherein when the optimization training module cannot determine that the sample information received is from the two-way sliding window low-level feature extraction module, it means The training of the low-level feature extraction module of the two-way sliding window is completed.

Images