CN108648023A - A kind of businessman's passenger flow forecast method of fusion history mean value and boosted tree - Google Patents

Abstract

The present invention relates to a kind of businessman's passenger flow forecast methods of fusion history mean value and boosted tree, which is characterized in that includes the following steps：The complete behavioral data of the businessman of certain time period is pre-processed；To passing through pretreated data construction feature；Passenger flow forecast model is built based on history mean value and boosted tree；Carry out passenger flow forecast.The present invention proposes internet businessman's passenger flow forecast model that history mean value is merged with boosted tree.The essence of the model is to promote tree-model and history mean value model, according to the weight coefficient calculated by calculation formula, the weighted sum that merges according to a certain percentage.The present invention not only allows for how improving the precision of prediction of model, and also contemplates the dependence of the prediction and time of the volume of the flow of passengers, and is made that comparative analysis to the prediction result of different models.

Description

A Merchant Customer Flow Prediction Method Combining Historical Mean and Boosting Tree

technical field

The invention relates to a passenger flow forecasting model that integrates historical mean values and boosted trees, and belongs to the fields of intelligent information processing and machine learning.

Background technique

The development of mobile positioning services has led to a sharp increase in the "online and offline" transaction data of Internet merchants. Compared with the traditional retail industry, the marketing of Internet merchants pays more attention to user consumption, and is committed to bringing users a better consumption experience in terms of product details page introduction, customer service, and convenient mobile payment. For example, some business intelligence service platforms can provide sales forecasts for each merchant. Based on the prediction results, merchants can establish a trust relationship with users, attract more loyal users, optimize operational decisions, reduce costs, and improve user experience.

Existing sales forecasting techniques generally use historical data and simply use time-weighted series methods for forecasting. However, in real life, users' consumption behavior is often affected by factors such as holidays and weather. At this time, the existing technology cannot predict the customer flow of the merchant in time, which may lead to unsatisfactory prediction accuracy. To a large extent, it deviates from the actual passenger flow of the merchant.

Contents of the invention

The purpose of the present invention is to provide a method capable of predicting passenger flow more accurately.

In order to achieve the above object, the technical solution of the present invention is to provide a method for predicting customer flow of merchants that integrates historical mean values and boosted trees, which is characterized in that it includes the following steps:

Step 1. Preprocess the complete merchant behavior data for a certain period of time. The complete merchant behavior data includes merchant characteristic data, user payment behavior data and user browsing behavior data;

Step 2. Build features on the preprocessed data, adding holiday data and weather feature data;

Step 3, constructing a passenger flow forecasting model based on the historical average value and the boosted tree, including the following steps:

Step 301: Construct two learning models for XGBoost and GBDT respectively, adjust the parameters of tree depth, learning rate and number of iterations for the two learning models, and introduce the XGBoost learning model when determining the learning rate of the XGBoost learning model and the maximum depth of the tree The built-in cv function in;

Step 302, using the data obtained in step 2 to train the XGBoost learning model and the GBDT learning model, setting a forecast date, and calculating the average passenger flow and sales increment from before the forecast date to a certain day;

Step 4. Use the correlation matrix of the historical sales volume of a certain time period in the past as the input of the passenger flow prediction model trained in step 3, and combine the sales volume of a certain time period in the future and the weight of the model fusion of the XGBoost learning model and the GBDT learning model The coefficient Credit is output as:

In the formula, is the average sales volume of a certain period of time in the past; Fus _last is the sales volume of a certain period of time in the past, thus, the average sales volume and sales value of a certain period of time in the past obtained by the XGBoost learning model, GBDT learning model and historical mean model, Substitute into the weight coefficient Credit formula to find the corresponding weight coefficient. Finally, the different results of the two groups of XGBoost learning models and two groups of GBDT learning models obtained by training are respectively compared with the historical average model according to the obtained corresponding weight coefficients. Proportional fusion is used to predict the passenger flow for a certain time period in the future.

Preferably, the pretreatment described in step 1 comprises the following steps:

Step 101. Eliminate the data of the 7 days before the business opened and the data of the 3 days before and after the sales interruption in the complete behavior data of the business, and divide the remaining data into a training set and a test set;

Step 102, remove the duplicate data in the training set and the test set, and use a rule-based method to normalize the deduplicated data in the training set and the test set, thereby eliminating abnormal data caused by a single user buying in large quantities in a short period of time;

For abnormal data caused by special time nodes and abnormal data caused by unpredictable large fluctuations, the model pre-training method is used to eliminate, that is, the under-fitting algorithm is used to pre-train the passenger flow prediction model, and the residual in the cleared data is 10% and 25% of the data.

Preferably, said step 2 includes the following steps:

Step 201, collecting weather data of various provinces and cities across the country;

Step 202, simply convert the weather conditions into two indicators of precipitation index and fineness index, and generate a human comfort index as an important feature of passenger flow forecasting model training;

Step 203 , collecting holiday data in the current time period, marking weekdays as 0, weekends as 1, and holidays as 2.

The invention proposes a forecasting model of traffic flow of Internet merchants in which the historical mean value and the boosting tree are fused. The essence of the model is the weighted sum of the promotion tree model and the historical average model, which are fused according to a certain proportion according to the weight coefficient obtained by the calculation formula. The invention not only considers how to improve the prediction accuracy of the model, but also considers the dependence of the passenger flow prediction on time, and compares and analyzes the prediction results of different models.

Description of drawings

Figure 1 is the prediction chart of the fusion model of historical mean and boosted tree:

Figure 2 is the prediction chart of time series weighted regression model.

Detailed ways

Below in conjunction with specific embodiment, further illustrate the present invention. It should be understood that these examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the teachings of the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of the present application.

The present invention provides a method for predicting customer flow of merchants by integrating historical mean value and boosting tree, comprising the following steps:

Step 1: Preprocessing the complete behavior data of merchants

The data used in the present invention comes from the Tianchi big data platform, including the complete behavior data of merchants from July 1 of a certain year to October 31 of the following year. It includes "merchant characteristics" data, "user payment behavior" data and "user browsing behavior" data. Because directly using the original data to train the model will not only generate errors, but also consume a lot of computing resources. Therefore, the abnormal values existing in the original data set are eliminated, deduplicated, and normalized. On the one hand, since there is a certain start-up time for merchants from entering the platform to increasing sales, and sales may be interrupted for a certain period of time, the data of 7 days before the opening of the merchant and the data of 3 days before and after the interruption of sales are not used as training data ; On the other hand, due to the fact that a single user purchases a large amount in a short period of time in the original data, in order to eliminate the impact of this abnormal consumption on the prediction, a rule-based method is used to normalize the original data. In addition, there are some special time nodes and unpredictable large fluctuations in the original data: such as large-scale holidays (such as Mid-Autumn Festival, National Day, etc.), business closures, and the situation of a single user buying in large quantities when merchants carry out promotional activities. For these outliers that are difficult to deal with by rule-based methods, the present invention adopts a model pre-training method. That is, firstly, the underfitting algorithm is used to pre-train the passenger flow forecasting model, and the data with residual error of 10% and 25% in the original data are cleared. Since the prediction target is the daily sales of merchants, the data used for training after preprocessing is the total sales of merchants counted by hour.

Step 2: Construct features on the preprocessed data.

In order to improve the accuracy of model prediction, the present invention collects weather data of various provinces and cities across the country and weather data of holidays as a supplement to the original data. In the additional collected data such as temperature, humidity, air pressure, etc., according to experience, the weather conditions are simply converted into two indicators: precipitation index and sunny index. Since the human body’s perception of meteorological parameters is not in a linear relationship, the human comfort index ( Comfort Index of Human Body, SSD) as an important feature of model training. Finally, the features and labels used for model training and prediction are shown in Table 1.

Table 1 Features used for model training and prediction

Step 3: Construct a passenger flow forecasting model based on the historical mean value and the boosted tree.

In order to obtain a high-precision passenger flow prediction model, the present invention adopts a two-stage training method. In the first phase of training, XGBoost (eXtreme Gradient Boost) and GBDT (Gradient Boosting Decision Tree) models were used. The parameters of model training are shown in Table 2 and Table 3. Each model uses 2 sets of parameters for training, and a total of 4 models are obtained.

Table 2 Different parameters of XGBoost algorithm

XGBoost number 1 number 2 objective function linear regression model linear regression model the maximum depth of the tree 3 5 learning rate 0.1 0.03 Number of boosted trees 500 1600 L1 regularization parameter 0 1 L2 regularization parameter 1 0

Table 3 Different parameters of GBDT algorithm

GBDT the maximum depth of the tree learning rate Number of boosted trees training sampling ratio number 1 3 0.1 500 0.95 number 2 5 0.1 500 0.95

The invention adjusts the parameters of tree depth, learning rate and iteration number in the XGBoost and GBDT algorithms. In the No. 1 model of the XGBoost algorithm, in general, the default value of the learning rate is 0.1, and the default maximum depth of the tree is 3. However, for different problems, the ideal learning rate sometimes fluctuates between some specific intervals. The deeper the tree, the better it fits the data. Therefore, the present invention introduces the built-in cv function in the XGBoost algorithm when determining the learning rate of the No. 2 model of the XGBoost algorithm and the maximum depth of the tree, and the cv function uses cross-validation in each round of iteration, according to the adjustment of the algorithm parameters, and Returns the ideal number of decision trees. Therefore, through the more accurate calculation of the cv function, the learning rate of the No. 2 model is adjusted to 0.03, and the maximum depth of the tree is 5. The second stage of training uses the historical mean model. Based on the forecast date, the historical average model first calculates the average sales volume of the 21 days before the forecast date to obtain the daily average sales volume; secondly, counts the median and average sales volume of each week in units of weeks, and passes A linear fit yields weekly sales increments.

Step 4: Carry out multi-model weighted fusion of the trained learner to predict the traffic flow of merchants; use the correlation matrix of historical sales in the past 21 days as input; combine the sales and historical average models of the next two weeks with the first-stage The weight coefficients of the model fusion are output. The mean model has a fusion scale of up to 0.75. The fusion weight coefficient Credit is calculated as the formula:

In the formula, is the average sales volume of the past three weeks, and Fus _last is the sales volume of the past three weeks. Therefore, XGBoost, GBDT and historical average model (historical average model is a kind of information based on the forecast date to obtain the average passenger flow, sales increment and other information from before the forecast date to a certain day, and then use the weight coefficient as the fusion model The average sales volume and sales volume value obtained in the past three weeks are respectively substituted into the weight coefficient formula, and the corresponding weight coefficients can be obtained as follows: 0.47, 0.34, and 0.19. Finally, the different results of the 2 sets of XGBoost models and 2 sets of GBDT obtained from training are fused with the historical average model at a ratio of 0.47, 0.34, and 0.19, respectively, and the passenger flow forecast for the next 14 days is obtained.

By optimizing the algorithm parameters and using the test set samples to predict the modeling results, the running results and accuracy tests of the algorithm are shown in Table 4.

Table 4 Historical mean and boosted tree fusion model accuracy test

In the experiment, the performance evaluation of the proposed model is carried out by using the evaluation function customized by XGBoost. When calling the evaluation function, the validation set and the predicted value on the validation set are passed in as function parameters, and an evaluation value fevalerror of floating point type is returned. The larger the value of fevalerror, the lower the prediction accuracy of the model. Conversely, the smaller the value of fevalerror, the higher the prediction accuracy of the model. The results show that with the increase of the sample size of the training set, the calculation time increases, the fevalerror value gradually decreases, but the accuracy gradually increases. Therefore, the fusion model of historical mean value and boosted tree has the advantages of higher prediction accuracy and faster operation speed.

Since the time series reflects the characteristics of entity attributes in time order, the time series weighted regression algorithm is implemented. After analyzing the prediction results of the two algorithms, the top 500 Internet merchants in the next 14 years are obtained as shown in Figure 1 and Figure 2. The daily passenger flow development trend. Wherein, the horizontal axis is the ID number of the merchant, and the vertical axis represents the predicted value of the passenger flow. Analyzing the development trend of passenger flow shows that:

1) Variables related to browsing actions contribute the most to the model, because browsing is the most important way of user interaction, and its information richness is much higher than other features;

2) Some merchants may have higher product evaluations, and the return rate of customers has led to a steady increase in the passenger flow of some merchants.

3) The total customer flow of most merchants in 14 days has exceeded 5,000, and a few even reached the level of about 25,000. This is most likely caused by some kind of promotional activity by the merchant recently. For example, different levels of coupons, cash red envelopes, and discounts for purchases of a certain amount are distributed through the platform. But how to adjust one's own operation strategy to attract more passenger flow is very important.

Claims (3)

1. a kind of businessman's passenger flow forecast method of fusion history mean value and boosted tree, which is characterized in that include the following steps：

Step 1 pre-processes the complete behavioral data of the businessman of certain time period, and the complete behavior data packet of businessman includes businessman spy Levy data, user's payment behavior data and user browsing behavior data；

Step 2, to passing through pretreated data construction feature, increase festivals or holidays data and weather characteristics data；

Step 3 builds passenger flow forecast model based on history mean value and boosted tree, includes the following steps：

Step 301 builds 2 learning models to XGBoost and GBDT respectively, depth, study to 2 learning model adjustment trees The parameter of rate and iterations, determine XGBoost learning models learning rate and tree depth capacity when, introduce Built-in cv functions in XGBoost learning models；

Step 302, the data obtained using step 2 are trained XGBoost learning models and GBDT learning models, and setting is pre- It surveys day, the average volume of the flow of passengers, sales volume increment before calculating prediction day to some day；.

Step 4, the passenger flow forecast mould that the correlation matrix of the history sales volume of section in those years has been trained as step 3 The input of type, by the power of the sales volume of the following certain time period and XGBoost learning models and the Model Fusion of GBDT learning models Weight coefficient Credit is as output：

In formula,It is the average sales volume of section in those years；Fus_lastIt is the sales volume of section in those years, as a result, will The average sales volume and pin for the section in those years that XGBoost learning models, GBDT learning models and history mean value model obtain Magnitude substitutes into weight coefficient Credit formula, finds out corresponding weight coefficient respectively, finally, 2 groups that training is obtained The Different Results of XGBoost learning models and 2 groups of GBDT learning models are corresponding by what is found out respectively to history mean value model respectively Weight coefficient ratio fusion, obtain the volume of the flow of passengers for predicting the following certain time period.

2. businessman's passenger flow forecast method of a kind of fusion history mean value and boosted tree according to claim 1, feature It is, pretreatment described in step 1 includes the following steps：

Step 101 rejects the number that the preceding 7 days data of businessman's opening and sales volume in the complete behavioral data of businessman interrupt front and back 3 days According to remaining data is divided into training set and test set；

Duplicate data in step 102, removal training set and test set, using rule-based method to training set and test set Data after middle duplicate removal are normalized, to eliminate single user in the short time make a big purchase in large quantities and caused by abnormal number According to；

For due to special timing node and caused by abnormal data and unpredicted fluctuation and caused by abnormal data, It is rejected using model pre-training method, that is, uses poor fitting algorithm to carry out pre-training to passenger flow forecast model, in clearing data The data that residual error is 10% and 25%.

3. businessman's passenger flow forecast method of a kind of fusion history mean value and boosted tree according to claim 1, feature It is, the step 2 includes the following steps：

Step 201, the weather data for acquiring national each province and city；

Weather conditions simple conversion for Precipitation Index and is become a fine day index two indices, and generates human comfort and refer to by step 202 An important feature of the number as passenger flow forecast model training；

Step 203, the festivals or holidays data for acquiring current slot, will be labeled as 0 working day, weekend is labeled as 1, and vacation is labeled as 2。