CN110599234A - Product sales prediction method - Google Patents

Abstract

The present invention relates to the field of artificial intelligence, and more specifically, to a product sales forecasting method. The present invention utilizes external data such as historical sales data and crawled weather. Compared with the traditional sales forecasting method, the present invention is based on a wide and deep model, and can use the powerful time-space sequence prediction of the convolutional long-short-term memory network Ability to fully tap the correlation between products and obtain structural information of products. By integrating external features such as regional characteristics and weather characteristics, this method well combines the influence characteristics of external factors on consumer consumption patterns, and realizes sales forecasts for different regions.

Description

A Product Sales Forecasting Method

technical field

The present invention relates to the field of artificial intelligence, and more specifically, to a product sales forecasting method.

Background technique

In recent years, with the continuous rise of new retail, the progress and changes of enterprises have accelerated and concentrated, becoming faster and more explosive. In the process of realizing social informatization and digitalization by the Internet, the retail industry has achieved rapid development relying on the development and changes of science and technology. At the same time, it has brought more challenges. Its characteristics can be summarized as digitalization, omni-channel and more flexible supply. chain. In the rapid development and change, the cost of the industry is reduced and the profit is increasing. At the same time, it also brings new challenges to the new retail industry. The impact of dimensionality reduction from multiple industries and platforms, and the diversity of consumer demand changes all affect enterprises. decisions have had a great impact. Whether it can predict the needs of consumers more accurately will play an important role in the new retail industry.

Traditional sales forecasting methods mainly include qualitative forecasting and quantitative forecasting. Qualitative forecasting mainly relies on the personal experience of managers to make decisions, and uses their judgments on the nature and degree of future development of things as the main basis for predicting the future. It has greater flexibility, but the qualitative prediction method is poor in portability and has strong subjective limitations; quantitative prediction is to reveal the objective law of a certain phenomenon quantitatively, which is divided into onlooker level and macro level, revealing from a certain angle The relationship between change and time, discover the law in essence, dig out its internal information, mainly use time series analysis methods to model historical sales data, including moving average method, ARIMA, Kalman filter and gray theory, machine Learning methods include support vector regression (SVR), tree model (XGBoot) and other algorithms. In addition, with the increasing amount of data and the development of neural network models, some time series data have also begun to be processed using deep learning techniques. With the help of the powerful learning ability of machine learning, these models can achieve better prediction accuracy. However, the existing forecasting methods are usually applicable to a single commodity, and a sample can only consider one commodity, ignoring the mutual influence between commodities. Even complex feature engineering and model fusion of historical data cannot improve the prediction performance.

Defects of existing methods and inventions: 1) Existing methods are usually applicable to a single commodity, and a sample can only consider one commodity, ignoring the interaction between commodities; 2) Existing methods only focus on the memory ability of the model, In order to enhance the generalization ability of the model and mine the potential correlation within the data, more artificial feature engineering is required; 3) The existing methods do not deal well with the influence of external factors on people's consumption patterns, just like the impact of heavy rain on people's consumption patterns. Consumers in different regions have completely opposite effects, and the portability of the model is poor.

Contents of the invention

In order to solve the prior art:

1) Existing methods are usually applicable to a single commodity, and a sample can only consider one commodity, ignoring the mutual influence between commodities;

2) Existing methods only focus on the memory ability of the model. In order to enhance the generalization ability of the model and mine potential correlations within the data, more artificial feature engineering is required;

3) The existing methods do not deal well with the impact of external factors on people's consumption patterns, such as heavy rains have completely opposite effects on consumers in different regions, and the portability of the model is poor.

In order to solve the problems of the technologies described above, the technical solution of the present invention is as follows:

A method for forecasting product sales, comprising the following steps:

Step S1: Obtain external data and internal data that affect product sales;

Step S2: Process the acquired data, remove outliers and fill missing values, and divide the processed data into training set and verification set;

Step S3: building a wide and deep model;

Step S4: Input the training set into the wide and deep model to train the model to obtain the optimized wide and deep model;

Step S5: Input the verification set into the optimized wide and deep model to verify the accuracy of the model.

Preferably, the external data described in step S1 includes external data including weather attributes, time features and regional features corresponding to the sales time period, the weather attributes include temperature, humidity, wind speed, rain and snow; time features include holidays, single and double breaks, Weekly, monthly; regional characteristics include surrounding residents, number of communities, number of competitors and surrounding housing prices.

Preferably, the internal data described in step S1 is the historical sales data of the product, including sales information, discount information and price information;

Preferably, the specific process of removing outliers in step S2 is as follows:

The method of outlier elimination is to use the lowess method to smooth the sales data. The specific process of lowess smoothing is as follows:

Take a point x as the center, intercept a piece of data with a length of frac forward and backward, and use the weight function w to do a weighted linear regression for this piece of data, record is the central value of the regression line, where For the corresponding value of the curve after fitting, n weighted regression lines can be made for all n data points, and the connection line of the central value of each regression line is the Lowess curve of this data.

Preferably, the specific process of filling the missing values in step S2 is as follows: Take the average value (x ₁ +x ₂ +...+x _n )/n of the sales data x of n days in history to fill in, and the external data needs to be filled according to The model needs to convert the data format accordingly. The conversion methods include OneHotEncoder one-hot encoding, LabelEncoder digital encoding, etc. At the same time, equal-frequency binning encoding is performed on numerical features such as temperature and housing prices.

Preferably, the specific steps of constructing the wide and deep model in step S3 are as follows:

The wide and deep models include two parts: the wide model and the deep model. The wide model satisfies the Memorization feature, and the deep model satisfies the Generalization feature. The Memorization feature is the rules learned by the model from historical data; The combination of new features that appear, and memorize the historical data, and then generalize these historical data to features that have not appeared before to improve the generalization ability of the model; integrate the Memorization characteristics of the wide model and the Generalization characteristics of the deep model, and play memorization at the same time and the role of generalization;

The wide model adopts a linear regression linear model. The input features include continuous features such as sales information and discount information, as well as sparse discrete features such as weather features and regional features. After the discrete features are crossed, a higher-dimensional discrete feature can be formed. Features, by introducing L1 regularization in linear model training, discrete features can converge to effective feature combinations;

The Deep model uses convLSTM to build a deep network model. By introducing the convLSTM model, the Deep model can realize the function of LSTM to consider the time information of sales, and can also realize the function of convolutional neural network to learn the spatial information of data, that is, different commodity The interaction between them; the core essence of the ConvLSTM model is the same as that of the LSTM. By adding convolution operations on the basis of the LSTM, the LSTM can not only obtain the temporal relationship, but also extract spatial features like a convolutional layer. By fully integrating the product Between the features, the spatio-temporal features are obtained, and the switching between states is converted into convolution calculations.

Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

The present invention uses external data such as historical sales data and crawled weather. Compared with the traditional sales forecasting method, the present invention can fully tap the correlation between products by utilizing the powerful time-space sequence prediction ability of the convolutional long-short-term memory network. Obtain the structural information of the product. By integrating external features such as regional characteristics and weather characteristics, this method well combines the influence characteristics of external factors on consumer consumption patterns, and realizes sales forecasts for different regions.

Description of drawings

Fig. 1 is a flow chart of the method of the present invention.

Figure 2 shows the framework of the Wide and Deep model.

Figure 3 shows individual convLSTM units.

Figure 4 shows the output of the deep terminal fully connected.

Detailed ways

The accompanying drawings are for illustrative purposes only and cannot be construed as limiting the patent;

In order to better illustrate this embodiment, some parts in the drawings will be omitted, enlarged or reduced, and do not represent the size of the actual product;

For those skilled in the art, it is understandable that some well-known structures and descriptions thereof may be omitted in the drawings.

The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

Example 1

As shown in Figure 1, a product sales forecasting method includes the following steps:

Step S1: Obtain external data and internal data that affect product sales;

Step S2: Process the acquired data, remove outliers and fill missing values, and divide the processed data into training set and verification set;

Step S3: building a wide and deep model;

Step S4: Input the training set into the wide and deep model to train the model to obtain the optimized wide and deep model;

Step S5: Input the verification set into the optimized wide and deep model to verify the accuracy of the model.

As a preferred embodiment, the external data described in step S1 includes external data including weather attributes, time features, and regional features corresponding to the sales period. Weather attributes include temperature, humidity, wind speed, rain and snow; time features include holidays, Single weekend, week, month; regional characteristics include surrounding residents, number of communities, number of competitors and surrounding housing prices.

As a preferred embodiment, the internal data described in step S1 is the historical sales data of the product, including sales information, discount information and price information;

As a preferred embodiment, the specific process of removing outliers in step S2 is as follows:

The method of outlier elimination is to use the lowess method to smooth the sales data. The specific process of lowess smoothing is as follows:

Take a point x as the center, intercept a piece of data with a length of frac forward and backward, and use the weight function w to do a weighted linear regression for this piece of data, record is the central value of the regression line, where For the corresponding value of the curve after fitting, n weighted regression lines can be made for all n data points, and the connection line of the central value of each regression line is the Lowess curve of this data.

As a preferred embodiment, the specific process of filling the missing values in step S2 is as follows: Take the mean value (x ₁ +x ₂ +...+x _n )/n of the sales data x of n days in history to fill in, and the external The data needs to be converted to the data format according to the model requirements. The conversion methods include OneHotEncoder one-hot encoding, LabelEncoder digital encoding, etc. At the same time, the numerical features such as temperature and house price are encoded in equal frequency bins.

As a preferred embodiment, the specific steps of constructing the wide and deep model in step S3 are as follows:

The wide and deep models include two parts: the wide model and the deep model. The wide model satisfies the Memorization feature, and the deep model satisfies the Generalization feature. The Memorization feature is the rules learned by the model from historical data; The combination of new features that appear, and memorize the historical data, and then generalize these historical data to features that have not appeared before to improve the generalization ability of the model; integrate the Memorization characteristics of the wide model and the Generalization characteristics of the deep model, and play memorization at the same time and the role of generalization;

The wide model adopts a linear regression linear model. The input features include continuous features such as sales information and discount information, as well as sparse discrete features such as weather features and regional features. After the discrete features are crossed, a higher-dimensional discrete feature can be formed. Features, by introducing L1 regularization in linear model training, discrete features can converge to effective feature combinations;

The Deep model uses convLSTM to build a deep network model. By introducing the convLSTM model, the Deep model can realize the function of LSTM to consider the time information of sales, and can also realize the function of convolutional neural network to learn the spatial information of data, that is, different commodity The interaction between them; the core essence of the ConvLSTM model is the same as that of the LSTM. By adding convolution operations on the basis of the LSTM, the LSTM can not only obtain the temporal relationship, but also extract spatial features like a convolutional layer. By fully integrating the product Between the features, the spatio-temporal features are obtained, and the switching between states is converted into convolution calculations.

Example 2

As shown in Figures 1 to 4, the specific process of this embodiment is as follows:

data processing:

Obtain the historical sales data of the product to be predicted, and the external data of the time period corresponding to the historical sales data. The external data includes weather attributes, time characteristics and regional characteristics.

Feature engineering:

The acquired data is processed into a format that can be applied to the model input through feature engineering, and then the data is divided into a training set and a verification set for model training.

specific:

1: For numerical features such as sales data and price features, outliers are eliminated and missing values are filled. At the same time, they need to be normalized, which speeds up the speed of model gradient descent to find the optimal solution and helps the model converge.

The normalization method is:

2: Carry out OneHotEncoder one-hot encoding, LabelEncoder digital encoding, etc. for the categorical features of weather attributes and time features, and perform equal-frequency binning encoding for numerical features such as temperature and housing prices;

3: In order to ensure the prediction of the time series by the model, it is necessary to sort the processed data according to time, and use a certain time point as the segmentation point to divide the data into a training set and a verification set, so as to ensure that there is no training set in the verification set. For example, the time span of the data set is 2018.01.01~2018.12.30, 2018.01.01~2018.10.01 can be used as the training set, and 2018.10.01~2018.12.30 can be used as the verification set.

wide end linear regression model

1. Use x ₁ , x ₂ , x _n to describe the components in the feature, such as x ₁ = historical sales of the product, x ₂ = temperature, x _n = whether it is a weekend, and get an estimation function:

h(x)=h _θ (x)=θ ₀ +θ ₁ x ₁ +θ ₂ x ₂

θ is a parameter that indicates the importance of the sales volume of each feature in the adjustment feature, for example, weather or holidays have a greater impact on sales.

2. For the convenience of calculation, set x ₀ =1, and express the estimation function in the form of vector:

h _θ (x) = θ ^T X

3. Involving a mechanism to evaluate the value of θ works best, and whether the model prediction result is close to the target, so it is necessary to evaluate the h function made. Generally, this function is called loss function (loss function) or error The error function is used to describe the degree of excellence of the h function. The defined loss function is as follows:

4. Use the gradient descent method to solve the loss function of the model, and find out the importance of each sales feature when it best meets the prediction target.

5. Each feature of the wide-end model training corresponds to the importance, and the output value of the estimation function during prediction is the predicted value.

Construct the deep-end convLSTM depth model:

The deep end uses the convLSTM depth model, and the input and output elements of the convLSTM depth model are 3D tensors that retain all spatial information. Since the network has multiple stacked ConvLSTM layers, it has strong representational power, making it suitable for making predictions in complex dynamic systems.

1. Construct the input feature data into a 3D tensor in time order,

The specific method is: regard the sales characteristics of each product every day as a 1D tensor, arrange the products according to a certain relationship order, find out the correlation between products through clustering, and at the same time splice the corresponding position according to the position of the sorted products 1D tensor, after splicing the sales features of all products every day, it corresponds to a 2D tensor, and splicing the obtained 2D tensors in chronological order to obtain the 3D tensor required by the input end of the convLSTM depth model

2. Build a convLSTM deep model network,

A single convLSTM unit is shown in Figure 3. ConvLSTM can learn the long-term dependence information of product sales in time through the gating unit, and through the convolution operation of the input sales subdivision matrix Xt, it can learn the potential relationship between the product sales data, that is, through convolution. Learn the relationship between products, and its mathematical model is as follows:

In the figure, * indicates convolution calculation, o indicates Hadamard product, Xt is the input variable at time t, Ot indicates the final output variable, it, ft and ot are the input gate, forget gate and output gate respectively, and Ct is the hidden layer memory unit The updated state at time t, Ht is the final state of the hidden layer memory unit at time t, Wi, Wf, Wo, Wc, Ui, Uf, Uo and Uc are connection weights, bi, bf, bo and bc are activation biases, The activation function selects sigmoid and tanh activation functions.

It is worth noting that X, C, H, i, f, and o here are all three-dimensional tensors, and their last two dimensions represent the spatial information of rows and columns, that is, imagine ConvLSTM as processing two-dimensional grids. A model of the eigenvectors of , which can predict the characteristics of the central grid from the characteristics of the surrounding points in the grid.

3. Deep end output 3D tensor result fully connected conversion

The output 3D tensor of ConvLSTM is shown in Figure 4. After fully connected transformation, the predicted sales volume of the corresponding product is output, the loss function of the predicted sales volume is calculated, and the connection weights Wi, Wf, Wo, Wc, Ui, Uf, Uo and Uc perform backpropagation training.

Fusion of wide-end and deep-end model prediction results:

The predicted values of the above two basic models are fused, and the weight fusion method is used to obtain the final model. The formula is as follows:

prediction＝w ₁ f ₁ +w ₂ f ₂

Among them, w ₁ is the weight of the output value of the wide end, and w ₂ is the weight of the output value of the deep end. The weights of the two basic models can be calculated according to the model evaluation index mape. The formula is as follows:

Among them, mape ₁ is the mape evaluation value of the wide end, and mape ₂ is the mape evaluation value of the deep end.

The same or similar reference numerals correspond to the same or similar components;

The terms describing the positional relationship in the drawings are only for illustrative purposes and cannot be interpreted as limitations on this patent;

Apparently, the above-mentioned embodiments of the present invention are only examples for clearly illustrating the present invention, rather than limiting the implementation of the present invention. For those of ordinary skill in the art, other changes or changes in different forms can be made on the basis of the above description. It is not necessary and impossible to exhaustively list all the implementation manners here. All modifications, equivalent replacements and improvements made within the spirit and principles of the present invention shall be included within the protection scope of the claims of the present invention.

Claims (6)

1. A product sales forecasting method, comprising the steps of:

step S1: acquiring external data and internal data which influence the product sales volume;

step S2: processing the acquired data, removing abnormal values and filling missing values, and dividing the processed data into a training set and a verification set;

step S3: constructing a wide and deep model;

step S4: inputting the training set into a wide and deep model to train the model to obtain an optimized wide and deep model;

step S5: and inputting the verification set into the optimized wide and deep model to verify the accuracy of the model.

2. The product sales forecasting method according to claim 1, wherein the external data in step S1 includes weather attributes, time characteristics, and regional characteristics corresponding to the sales time period, the weather attributes including temperature, humidity, wind speed, rain and snow; the time characteristics comprise holidays, single and double break, weeks and months; regional characteristics include surrounding households, number of cells, number of competitors, and surrounding rates.

3. The method of claim 1, wherein the internal data in step S1 is historical sales data of the product, including sales information, discount information and price information.

4. The product sales prediction method of claim 3, wherein the step S2 of eliminating the abnormal value comprises the following steps:

the method for eliminating the abnormal value is to smooth the sales data by using a lowss method, and the specific process of lowss smoothing is as follows:

taking a point x as a center, intercepting a section of data with the length of frac forward and backward, performing weighted linear regression on the section of data by using a weight function w, and recordingIs the central value of the regression line, whereinFor the fitted curve corresponding values, n weighted regression lines can be made for all n data points, and the connecting line of the central value of each regression line is the Lowess curve of the data.

5. The product sales prediction method of claim 3, wherein the missing values in step S2 are filled as follows: taking the average value (x) of historical n-day sales data x₁+x₂+...+x_n) Filling is carried out according to/n, and external data needs to be according to the modeThe data format is correspondingly converted according to the type requirement, the conversion method comprises one HotEncoder one-hot coding, LabelEncoder digital coding and the like, and meanwhile, the equal-frequency box-dividing coding is carried out on the numerical characteristics of temperature, room price and the like.

6. The product sales forecasting method according to claim 3, wherein the step S3 of constructing the wide and deep model comprises the following steps:

the wide and deep model comprises a wide model and a deep model, wherein the wide model meets the memorialization characteristic, the deep model meets the Generalization characteristic, and the memorialization characteristic is a rule learned by the model into historical data; the Generalization characteristic is based on historical data, a new feature combination which never appears in the past is explored, the historical data is memorized, and then the historical data is generalized to the features which do not appear before, so that the Generalization capability of the model is improved; fusing the memorisation characteristic of the wide model and the generalisation characteristic of the deep model, and simultaneously playing the roles of the memorisation and the generalisation;

the wide model adopts a linear regression linear model, input features comprise continuous features such as sales information and discount information and sparse discrete features such as weather features and regional features, the discrete features can form higher-dimensional discrete features after being crossed, and the discrete features can be converged into an effective feature combination by introducing L1 regularization in linear model training;

the Deep network model is constructed by the Deep model in the convLSTM mode, and the Deep model can realize the function of time information of the LSTM considering the sales volume and the function of learning the space information of the data by the convolutional neural network by introducing the convLSTM mode, namely the mutual influence among different commodities; the ConvLSTM model has the same core essence as the LSTM, and by adding convolution operation on the basis of the LSTM, the LSTM can not only obtain a time sequence relation, but also extract spatial features like a convolution layer, obtain space-time features by fully fusing the features among commodities, and convert the switching between states into convolution calculation.