CN111639978A - Event-driven demand forecasting method for e-commerce based on Prophet-random forest - Google Patents

Abstract

The invention discloses an event-driven demand forecasting method for e-commerce based on Prophet-random forest, comprising the steps of: acquiring historical sales data of an e-commerce platform, the sales data including time series data and user data for purchasing related products; cleaning the historical sales data , improve data quality; compress time series data to reduce data volatility; conduct event modeling based on festival activities and e-commerce platform promotion events; add event regressors to the Prophet model, and perform Prophet prediction, eliminate and interpolate as events Outliers within the effect range, and perform random forest prediction; combine Prophet prediction and random forest prediction results; perform accuracy evaluation to observe model generalization ability and prediction effect. Aiming at the characteristics of e-commerce event demand and non-event demand, the present invention uses Prophet and random forest for prediction respectively, and improves the prediction accuracy of the event-driven demand of e-commerce.

Description

Event-driven demand forecasting method for e-commerce based on Prophet-random forest

technical field

The invention belongs to the field of e-commerce demand forecasting, and in particular relates to an event-driven demand forecast method for e-commerce based on Prophet-random forest.

Background technique

The date interval between festival activities and e-commerce platform promotion events is not fixed, and does not show a standard cyclical movement law. The short-term consumer demand driven by it shows a surge in the short term of the event and a small fluctuation in the non-event effect period. state. However, the current common method for e-commerce demand forecasting is the autoregressive moving average model. The premise of time series data forecasting is that the data is stable, and the law of unstable data cannot be well captured. Under the circumstance that platform promotion events are driven and have the characteristics of emergencies, the existing models cannot represent the needs of the real business environment.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an event-driven demand forecasting method for e-commerce based on Prophet-random forest. According to the characteristics of e-commerce event demand and non-event demand, the present invention uses Prophet and random forest for prediction respectively, which improves the efficiency of e-commerce. Forecast accuracy for event-driven demand.

The technical scheme adopted in the present invention is:

An event-driven demand forecasting method for e-commerce based on Prophet-random forest, including steps:

Step 1. Obtain historical sales data of the e-commerce platform, including time series data and user data for purchasing related products;

Time series data includes sales data at different times to describe the changes in demand over time; user data for purchasing related products should include product category descriptions and customer information.

Step 2. Clean historical sales data to improve data quality;

Data cleaning operations include invalid value processing, missing value and duplicate value processing, consistency processing, data subset filtering and data sorting. Among them, invalid value processing is to identify extreme values with statistical analysis methods and delete or replace them. Duplicate value processing is to replace missing values with mean, maximum, minimum or probability estimates and remove duplicate data. Consistency processing is to remove data that are outside the normal range according to the legal rules and logic of variables. Data subset screening is to filter Useful information and reduce field redundancy. Data sorting is to sort the overall data in time series according to the time series data in the sales data to observe the time series change law of the data.

Step 3. Scale the time series data to reduce data volatility;

The time series data is the extracted timestamp and order demand; the timestamp data column contains the date (YYYY-MM-DD) or the specific time point (YYYY-MM-DD HH:MM:SS); the e-commerce order demand data column is numeric variable;

Scale compression is a multiplication model y(t)=T*S*H*I based on time series decomposition, that is, the time series is decomposed into trend items T, seasonal items S, festival activities and e-commerce platform promotion event items H, and random fluctuation items I; The way the data is scaled compressed:

y(t) ⁽¹⁾ = ln(y(t))

Step 4. Conduct event modeling based on festival activities and e-commerce platform promotion events;

The event modeling is to set the time window values before and after different festival activities and e-commerce platform promotional events Hi, _{where Hi = (H 1 , H 2 ,..., H n ) , and assume} festival activities and e-commerce platform promotional events The influence on the order demand in the front and back window values is Gaussian distribution, ie k～N(0, γ ² ), where γ represents the degree of influence of the event on the model. For the window day t in a specific event, its event effect E(t) is represented by a dummy variable:

Step 5,

1. Attach an event regressor to the Prophet model and perform Prophet prediction. Prophet prediction includes steps:

1) Create an event additional regressor, assign the same dummy variable assignment to the event that positively affects the order demand in step 4, but assign different γ to different events to further distinguish the degree of influence of different events;

2) By creating an event list, import the event additional regressor and the event effect in step 3;

3) Cut the training set for model training, and use the test set to preliminarily evaluate the model score. And adjust the parameters based on the model training and testing results to determine the final trend item model, the location, number and growth rate of mutation points, the degree of seasonal fit and related seasonal factors;

4) Use the evaluation indicators mean absolute error, root mean square error, and mean absolute percentage error to evaluate the model prediction effect.

2. Eliminate and interpolate outliers belonging to event effects, and perform random forest prediction;

The outliers of the event effect are the outliers in the time series data within the scope of the event effect screened out by the Box Plot plot and the modeling event, and they are eliminated and interpolated before random forest prediction to make up for the elimination of outliers. missing item after point;

Random forest prediction consists of steps:

1) Combine time series data, event modeling results and other information in the original data to construct time series features, product and user features;

Time series features include historical features, translation features and sliding windows; time series historical features include related weeks, months, and years; week feature values include {0, 1, 2, 3, 4, 5, 6}, which in turn correspond to the current order date. Monday to Sunday; the month feature value includes {1, 2, 3, 4...12}, corresponding to the order date month January to December in turn; the year feature value corresponds to the corresponding year in the order date in turn; the translation feature refers to the time The correlation characteristics between the corresponding order demand and the original data after the sequence is shifted forward; the sliding window feature is the embodiment of the same feature in different time dimensions, which is obtained from the statistical analysis of the corresponding order demand in the fixed interval time period. out;

The product and user features are constructed based on the product category, user information, and commodity attribute information in the original data;

2) Based on the Gini index to evaluate the importance of each feature for feature ranking and screening;

Among them, feature screening refers to eliminating features whose importance is less than a certain degree (such as 1%) to avoid model overfitting;

3) Carry out model training and parameter adjustment;

The premise of model training is to cut the training set, and the cutting ratio should be the same as that of the Prophet algorithm;

Parameter adjustment is to use the test set to initially evaluate the random forest model score, and adjust the parameters based on the model score, and adjust the model fitting parameters such as the maximum number of iterations of the weak learner and the depth of the decision tree;

4) Use the mean absolute error, root mean square error, and mean absolute percentage error evaluation indicators to evaluate the model prediction effect.

Step 6. Combine Prophet prediction and random forest prediction results;

The combined results include the predicted value of event effects in the prediction results of Prophet and the predicted value of non-event effects in the prediction results of random forest.

Step 7. Carry out accuracy evaluation to observe the generalization ability and prediction effect of the model;

Calculate the evaluation indicators of the mean absolute error, root mean square error, and mean absolute percentage error of the Prophet-random forest combination model, and compare with the prediction accuracy of the Prophet and random forest models respectively to evaluate the prediction effect of the combined model.

The beneficial effects of the present invention are:

Aiming at the characteristics of e-commerce event requirements and non-event requirements, the present invention uses Prophet and random forest for prediction respectively, which improves the prediction accuracy of e-commerce event-driven requirements. The Prophet model is mainly used for e-commerce event modeling. By establishing additional events Regressor, sensitively captures e-commerce events, and speculates on corresponding future events, effectively solving the problem that conventional demand forecasting methods cannot model e-commerce holidays and promotional events according to a stable periodic pattern. The forest model combines product and user characteristics in the original data to perform regression prediction on non-event time series, which makes up for the problem that the non-event time series prediction interval of the Prophet model is too wide due to extreme outliers, and at the same time plays a rational use. The role of other raw data other than non-sequential features.

Description of drawings

FIG. 1 is a flow chart of main steps of an embodiment of the present invention.

FIG. 2 is a comparison diagram of the actual demand and the predicted amount of the e-commerce in the embodiment of the present invention.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and embodiments.

Taking the demand for maternal and infant products as an example, as shown in Figure 1, an event-driven demand forecasting method for e-commerce based on Prophet-Random Forest is disclosed, including the steps:

Step 1. Obtain the historical sales data of the e-commerce platform, and the sales data includes time series data and user data for purchasing related products;

The historical data uses the Taobao and Tmall maternal and child sales data sets uploaded on the Tianchi Data Lab website (https://tianchi.aliyun.com). The dataset includes two data tables, namely the infant information table and the transaction information table. The data fields of the baby information table include the date of birth, gender, and user ID of the baby; the data fields of the transaction information table include user ID, user behavior description, product serial number, product secondary category, item attribute description, user purchase quantity, and purchase date. Among them, the two data tables are linked together by the user ID data column, and one row of data represents the transaction information of a user on a purchase date.

Step 2. Clean historical sales data to improve data quality;

Data cleaning operations include invalid value processing, missing and duplicate value processing, consistency processing, data subset filtering, and data sorting:

1) Invalid value processing;

Handling invalid baby age data. After combining the two data tables, use the baby's birth date and the user's purchase date to calculate the baby's age when the user made a purchase. Considering that the baby information entered by the customer may be inaccurate, delete the invalid baby age data, that is, delete the data that is older than 28 years old and less than 0 years old and are not in the pregnancy cycle, and delete the baby who is less than 0 years old but still in the pregnancy cycle. Age changed to 0; also handle extreme value of purchase quantity. Perform descriptive statistical analysis on the purchase quantity data, and delete the data row where the maximum value is 10000;

2) Handling of missing values and duplicate values: After searching and filtering, the data set has no blank values and duplicate values;

3) Consistency processing;

Handling abnormal gender data. Baby gender is 0 or 1, representing male or female, respectively. Remove illogical baby gender 2 data; also process purchase date data. Change the purchase date uniformly to the standard date format;

4) Data subset screening;

Since the method used in the present invention is based on time series analysis and product feature analysis, data columns of infant age, infant gender, secondary category of commodities, user purchase quantity, and purchase date are selected from the original data set. Among them, the item attribute description data column is too complicated and its meaning cannot be known, so this data is not considered;

5) Data sorting;

Overall ordering of the dataset based on time series. And combine the purchase quantities of the same purchase date.

Step 3. Scale the time series data to reduce data volatility;

The multiplication model y(t)=T*S*H*I based on time series decomposition, that is, the time series is decomposed into trend item T, seasonal item S, holiday and promotion event item H, random fluctuation item I, and the data is processed by logarithmic transformation. Scale compression.

Step 4. Conduct event modeling based on festival activities and e-commerce platform promotion events;

Based on the time series contained in the dataset, the main e-commerce maternal and child events were modeled. Set the before and after time window values for the important e-commerce maternal and child events Hi within the scope of the data event, where _Hi =( _{H 1} , H ₂ ,..., H ₆ ), and assume that the e-commerce event has a significant impact on The influence of the order demand is Gaussian distribution, namely k～N(0, γ ² ), where γ is 10; for the window day t in a specific event, the event effect E(t) is represented by a dummy variable;

It should be noted that, in other feasible embodiments, the specific window period before and after the event is not limited, and different products, time series, and e-commerce platforms also have different event lists.

Table 1 List of e-commerce maternal and child events

Step 5,

1. Attach the event regressor to the Prophet model, and perform the Prophet model prediction;

Prophet prediction consists of steps:

1) Create an event additional regressor, assign the same dummy variable assignment to the event that positively affects the order demand in step 4, but assign different γ to different events to further distinguish the degree of influence of different events. In this embodiment, the additional events are Women's Day, Mother's Day, 99 Promotion, Double Eleven, Double Twelve; the additional events are respectively set with specific γ values: 10, 20, 20, 40, 20;

2) By creating an event list, import the event additional regressor and the event effect in step 3;

3) Cut the training set for model training, and set the proportion of the training set to the total data set to 0.75. The results of adjusting parameters based on the model training and testing results are as follows: the final trend term model is a piecewise linear model, the growth rate of the mutation point is 0.1, and the confidence interval is 0.95;

4) Use the evaluation indicators Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) to evaluate the prediction effect of the model.

2. Eliminate and interpolate outliers belonging to event effects, and perform random forest prediction;

According to the Box plot, there are 55 outliers in the time series data in this example. The outliers in this embodiment are the driving results of small e-commerce events such as individual store promotions, and belong to normal observation data. Therefore, in order to predict the non-event sales data, only 11 outliers belonging to the event effect are eliminated, and the Prophet algorithm is used to interpolate the missing values generated by the proposed data column.

First-order difference processing is performed on the data after processing outliers to obtain more stable data;

Random forest prediction consists of steps:

1) Combine time series data, event modeling results and other information in the original data to construct time series features, product and user features;

This embodiment mainly creates time-series features based on time-series historical features and translation features, and selects subsets based on the results of step 2.

Create product and user features, and add maternal and infant event features based on the event modeling results in step 4.

Table 2 Feature Creation List

2) Based on the Gini index to evaluate the importance of each feature for feature ranking and screening;

Features less than 1% importance are removed to avoid model overfitting. The remaining features after removal are: Day, Weekday, Month, Year, Event, Good_28, Good_50008168, Good_50014815, Good_38, Good_122650008, Good_50022520, Baby_age;

3) Carry out model training and parameter adjustment;

The training set is cut for model training, and the cutting ratio is the same as that of the Prophet algorithm, accounting for 0.75 of the total data set. Use the test set to initially evaluate random forest model scores. Adjust the parameters based on the model score, adjust the maximum number of iterations of the weak learner to 200, the maximum depth of the decision tree to 20, and use the minimum square mean error (MSE) for the segmentation strategy;

4) Use the evaluation indicators Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) to evaluate the prediction effect of the model.

Table 3 Comparison of model evaluation indicators

Step 6. Combine Prophet prediction and random forest prediction results;

The combined results include the predicted value of event effects in the prediction results of Prophet and the predicted value of non-event effects in the prediction results of random forest.

Step 7. Carry out accuracy evaluation to observe the generalization ability and prediction effect of the model;

Calculate the evaluation indicators of the mean absolute error, root mean square error, and mean absolute percentage error of the Prophet-random forest combination model, and compare with the prediction accuracy of the Prophet and random forest models respectively to evaluate the prediction effect of the combined model. As shown in FIG. 2 , the model evaluation result of the present invention shows that the Prophet-Random Forest algorithm of the present invention has better prediction performance for the event-driven demand of e-commerce compared with the Prophet algorithm and the random forest algorithm.

It should be understood that, for those skilled in the art, improvements or changes can be made according to the above description, and all these improvements and changes should fall within the protection scope of the appended claims of the present invention.

Claims (10)

1. an e-commerce event-driven demand forecasting method based on Prophet-random forest, is characterized in that: comprise steps, Step 1. Obtain historical sales data of the e-commerce platform, including time series data and user data for purchasing related products; Step 2. Clean historical sales data to improve data quality; Step 3. Scale the time series data to reduce data volatility; Step 4. Conduct event modeling based on festival activities and e-commerce platform promotion events; Step 5. Add an event regressor to the Prophet model, and perform Prophet prediction; remove and interpolate outliers that belong to the range of event effects, and perform random forest prediction; Step 6. Combine Prophet prediction and random forest prediction results; Step 7: Carry out an accuracy evaluation to observe the generalization ability and prediction effect of the model. 2. The event-driven demand forecasting method for e-commerce based on Prophet-random forest as claimed in claim 1, wherein in step 1, the time series data includes sales data at different times to describe demand over time Changes; purchase-related product user data includes product category descriptions and customer information. 3. The prophet-random forest-based e-commerce event-driven demand forecasting method according to claim 1, wherein in step 2, the data cleaning operation includes invalid value processing, missing value and repeated value processing, consistent Sexual processing, data subset screening and data sorting, among which, invalid value processing is to identify extreme values by statistical analysis and delete or replace them; Replace missing values and delete duplicate data. Consistency processing is to delete data that exceeds the normal range according to the legal rules and logic of variables. Data subset filtering is to filter useful information and reduce field redundancy. Data sorting is based on sales data. Time series data sorts the overall data in time series to observe the time series variation of the data. 4. The prophet-random forest-based e-commerce event-driven demand forecasting method as claimed in claim 1, wherein in step 3, the time series data is the extracted timestamp and order demand, and the timestamp data The column contains dates or specific time points, and the order demand data column is a numerical variable. The multiplication model y(t)=T*S*H*I based on time series decomposition decomposes the time series into trend items T, seasonal items S, festivals Activity and e-commerce platform promotion event item H and random fluctuation item I, the data is scaled compressed by logarithmic transformation. 5. The prophet-random forest-based e-commerce event-driven demand forecasting method as claimed in claim 4, wherein in step 4, the event modeling is set as different festival activities and e-commerce platform promotional events H _i Before and after the time window value, and it is assumed that the influence of festival activities and e-commerce platform promotion events Hi on the order demand in the before and after window values is Gaussian distribution, that is, _k ～N(0, γ ² ), where γ represents the degree of influence; For the window day t in a specific event, its event effect E(t) is represented by a dummy variable:

6. The event-driven demand forecasting method for e-commerce based on Prophet-random forest as claimed in claim 5, characterized in that: in step 5, adding an event regressor to the Prophet model, and performing the Prophet prediction, comprising the steps of: : 1) Create an event additional regressor, assign the same dummy variable assignment to the event that positively affects the order demand in step 4, but assign different events to γ to further distinguish the degree of influence of different events; 2) By creating an event list, import the event additional regressor and the event effect in step 3; 3) Cut the training set for model training, use the test set to preliminarily evaluate the model score, and adjust the parameters based on the model training and test results to determine the final trend item model, the location, number and growth rate of mutation points, and the degree of seasonality fitting and related seasonal factors; 4) Use the evaluation indicators mean absolute error, root mean square error, and mean absolute percentage error to evaluate the model prediction effect. 7. The event-driven demand forecasting method for e-commerce based on Prophet-random forest as claimed in claim 5, characterized in that: in step 5, when eliminating and interpolating outliers belonging to the event effect range, the Cluster processing is to use Box Plot and modeling events to screen out outliers in the time series data within the scope of the event effect, remove them and perform interpolation before random forest prediction, to make up for the missing after removing outliers. item. 8. The prophet-random forest-based e-commerce event-driven demand forecasting method as claimed in claim 7, wherein in step 5, when performing random forest forecasting, the method comprises the steps: 1) Combine time series data, event modeling results and other information in the original data to construct time series features, product and user features; Time series features include historical features, translation features, and sliding windows; time series historical features include related weeks, months, and years, and week feature values include {0, 1, 2, 3, 4, 5, 6}, which correspond to the current order date. From Monday to Sunday, the month feature value includes {1, 2, 3, 4...12}, which corresponds to the order date month from January to December, and the year feature value corresponds to the corresponding year in the order date in turn; the translation feature refers to the time The correlation characteristics between the corresponding order demand and the original data after the sequence is shifted forward; the sliding window feature is the embodiment of the same feature in different time dimensions, which is obtained from the statistical analysis of the corresponding order demand in the fixed interval time period. out; The product and user features are constructed based on the product category, user information, and commodity attribute information in the original data; 2) Based on the Gini index to evaluate the importance of each feature for feature ranking and screening; Feature screening is to remove features whose importance is less than a certain level to avoid model overfitting; 3) Carry out model training and parameter adjustment; The premise of model training is to cut the training set, and the cutting ratio is the same as that of the Prophet algorithm; Parameter adjustment is to use the test set to initially evaluate the random forest model score, and adjust the parameters based on the model score, and adjust the model fitting parameters such as the maximum number of iterations of the weak learner and the depth of the decision tree; 4) Use the evaluation indicators mean absolute error, root mean square error, and mean absolute percentage error to evaluate the model prediction effect. 9. The event-driven demand forecast method for e-commerce based on Prophet-random forest as claimed in claim 1, characterized in that: in step 5, in step 6, the combined result includes the event effect prediction in the Prophet prediction result value and the non-event effect predicted value in the random forest prediction results. 10. The prophet-random forest-based e-commerce event-driven demand forecasting method as claimed in claim 1, wherein in step 7, the mean absolute error and the root mean square error of the Prophet-random forest combined model are calculated , the evaluation index of the mean absolute percentage error, and compared with the prediction accuracy of the Prophet and random forest models respectively to evaluate the prediction effect of the combined model.

Images