CN1828306A - Realization of sensory index prediction method for formula products based on M5' model tree - Google Patents

Abstract

To overcome the defects in prior art, in this invention, constructing a basic decision tree; then, in pruning stage, building LRM for every node, while eliminating some sub-trees to prevent over fitting; finally, reducing the nonlinearity on model fragmenting points as small sample with smoothing process to obtain the correlation model and the express prediction system. Wherein, the key lies on that it introduces the M5' model into the prediction process.

Description

Realization of sensory index prediction method for formula products based on M5' model tree

technical field

The invention relates to a method flow for data prediction and analysis, in particular to establishing a method for predicting sensory quality indicators in the production process of formula products.

Background technique

In the manufacturing industry of existing formula products, such as cigarettes for daily consumption, food, spices, food additives, etc., the formula components and their proportions used are related to the quality and grade assessment of raw materials and finished products. For example, cigarette products are usually evaluated by indicators such as fragrance style, irritation, strength, etc., so as to indicate their different grades to consumers. To study the relationship between the formula components of formula products and physical, chemical and sensory indicators, and then improve the quality of product manufacturing and the evaluation of grade products is a task with a large amount of data processing. It has always been researched by industry experts and strives to obtain internal Regular process improvement direction.

The previous evaluation of existing formula products relied on industry experts to classify the grades, advantages and disadvantages through on-site tasting and personal sensory experience. Although the manufacturers of formula products have accumulated a certain amount of expert evaluation data through long-term production management, since the quality evaluation is performed by individuals, there are inevitably many human factors in these evaluation data. For example, in the process of quality evaluation, experts will be interfered by factors such as their own emotions, physical conditions, personal sensory preferences, and the degree of fatigue, and there will be sensory errors objectively, which will eventually be reflected in the inaccurate classification of formula products and the difficulty To further improve and optimize the production process. Moreover, organizing experts to conduct quality assessment also requires high costs and a lot of time.

The existing improvement scheme is to use the artificial neural network (BP network) to predict the sensory index. However, this kind of BP network needs to modify many parameters to predict sensory indicators. For different indicators, different parameters should be selected according to their characteristics, such as the number of hidden layer units, momentum coefficient, learning rate, etc. In practical applications, how to estimate the number of neurons in the hidden layer has always been the difficulty and key to determine the structure of BP network, and there is no strict theoretical basis yet. In addition, the data composition conditions for the formation of sensory indicators are relatively complex, such as the place of origin, climate, soil and other factors have a greater impact on the indicator data. Different forecasting models should be established for different data, which will lead to many problems such as heavy workload and difficulty in adjusting parameters.

Contents of the invention

The sensory index prediction method of formula products based on the M5' model tree of the present invention is aimed at solving the above-mentioned problems and deficiencies by constructing a basic decision tree, and then establishing a linear regression model for each node in the pruning stage, and subtracting Part of the subtree is used to prevent overfitting, and finally the smoothing process is used to reduce the nonlinearity at the model segmentation point caused by the small sample size, so as to establish a correlation model that can more accurately describe the physical and chemical data and various sensory indicators. In order to establish a quick prediction system that reflects its internal laws.

The core of the sensory index prediction method described in the present invention is to introduce the M5' model tree into the prediction process, so as to realize the combination of knowledge data provided by formula product evaluation experts and equipment machine learning technology.

The so-called decision tree is a widely used machine learning technique (References, WITTEN, I.H., FRANK, E., 1999. DATA MINING: PRACTICAL MACHINE LEARNING TOOLS AND TECHNIQUES WITH JAVA IMPLEMENTATIONS. MORGAN KAUFMANN, SAN FRANCISCO.).

Decision trees can be applied to data classification and prediction of data. A decision tree consists of leaf nodes representing classes and internal nodes representing classification conditions. Inducing a decision tree from top to bottom is a common processing method, which can make the classification process start from a root node, and continuously generate subtrees until a leaf node is generated.

Since the current basic decision tree cannot be applied to the problem of numerical prediction (prediction for continuous values), the present invention combines the decision tree with linear regression to generate the M5' model tree.

The key to applying the M5' model tree is:

First, the basic decision tree is generated according to the principle of information gain maximization, and the split attributes and corresponding split values are found according to the significance of the impact on the output;

Then, pruning the basic decision tree to prevent overfitting;

Finally, smooth the pruned model; smoothing can greatly improve prediction accuracy, especially for model trees generated from a small amount of training sample data.

The M5' model tree is actually a piecewise linear function. The M5' model tree is like a typical regression equation, which predicts the value of a variable (called a class) through a series of independent variables (called attributes).

The training data represented in the form of a table can be directly used to construct a decision tree. In the data table, each row (sample) is expressed as (x ₁ , x ₂ , ... x _N , y), where x _i represents the value of the Nth attribute and y is the class value (target value).

For a given data set, a typical linear regression algorithm can only give a single regression equation, but the M5' model tree can divide the sample space into rectangular areas with parallel edges, and determine a corresponding regression model for each partition .

The M5' model tree tests the value of a specific attribute at each internal node and predicts a class value at each leaf node. When given a new data sample, which can be used to predict its class value, the tree is interpreted starting from the root node. At each internal node, the left branch or the right branch is selected according to a specific attribute value of the sample. When the selected node is a leaf node, the output is predicted by the model of the leaf node.

The structure of the M5' model tree is generated recursively, starting from the entire training sample set. At each level of the model tree, the most discriminative attribute is selected as the root node of the subtree, and the samples arriving at this node are divided into several subsets according to the value of the node attribute.

Statistically, the attribute that minimizes the variance of the target attribute set is the most discriminative. The M5' model tree uses variance (VARIANCE) induction as a heuristic method, and fills constant values in leaf nodes as models. For discrete attributes, each branch of an internal node represents a possible value of the attribute of the parent node. For continuous attributes, the algorithm will determine a segmentation point, so as to generate two branches according to this segmentation point. This constructor is called recursively for each subtree of the model tree.

When the variance or the number of samples of the class attribute set of samples reaching a certain node is small enough, the tree construction method stops, and this node is a leaf node.

Pruning (PRUNING) is an important method to prevent the tree from over-learning the training samples. Pruning can be performed during tree construction (PRE-PRUNING), or after the basic tree is constructed (POST-PRUNING).

The M5' model tree adopts the method of post-pruning. In the pruning stage, if the performance of the linear model of the internal node is not lower than the performance of the subtree of this node, the internal node will be turned into a leaf node containing the linear model. The only properties that a node's linear model may contain are all properties of its subtrees, resulting from linear regression on the subset of samples that reach this node.

For the smoothing process, the M5' model tree is directly smoothed after pruning. That is, the linear model of the internal node is merged into the model of the leaf node. When predicting, when a sample arrives at a leaf node from the root node of the tree, only the linear model of the leaf node is used to predict the output.

Relates the current predicted value of the sample with the predicted value of the linear model for the nodes reached until the root node is reached. The smooth point expression is: $p^{\′}=\frac{\mathrm{np}+\mathrm{kq}}{\mathrm{no}+k}.$

Among them, p' is the predicted value passed from the current node to the parent node,

p is the predicted value passed from the child node to the current node,

q is the predicted value of the linear model of the current node,

n is the number of samples arriving at the child node,

k is a smoothing constant.

Smooth the leaf nodes of the tree according to the number, and set the current leaf node as the current node. If the parent node of the current node is non-empty, use the linear regression model of the parent node to smooth the linear model of the current leaf node. The properties of the model after smoothing are:

The attribute Y of the current model of the current leaf node is the attribute of the parent node model of the current node, and the correlation coefficient expression corresponding to the i-th attribute is: $\mathrm{newcoeff}\left[i\right]=\frac{\mathrm{np}+\mathrm{kq}}{\mathrm{no}+k},$

Among them, n is the number of samples arriving at the current node,

k is a smoothing constant (usually k=15).

Set the parent node of the current node as the current node, and continue smoothing; if the parent node of the current node is empty, the smoothing ends, and the smoothing model of the current leaf node is obtained.

The M5' model tree is a global model combined by a series of segmented linear models, which realizes the non-linearity required by the correlation prediction method between complex data existing in formula products and sensory indicators.

According to the present invention, based on the M5' model tree, the sensory index prediction method of the formula product is realized, and its flow process is:

Detect various physical and chemical data and sensory indicators of raw materials and finished products of formula products, organize industry experts to evaluate their single materials and finished products, and record the obtained data as the sample data set of this method;

Eliminate erroneous or idiosyncratic sample data based on experts' industry experience;

According to the index parameters such as place of origin, grade, style, etc., the sorted data samples are divided into several groups of sample sets;

Perform data preprocessing on a set of sample sets, including eliminating samples with missing target values, filling samples with missing input attribute values, and converting discrete attribute values into continuous attribute values;

According to the principle of maximum information gain, select split attributes and split values, and recursively build a basic decision tree from the root node;

The basic decision tree is pruned recursively from the leaf node from bottom to top until it reaches the root node; if the performance of the linear model of the internal node is not lower than the performance of the subtree of this node, then this internal node becomes a A leaf node containing a linear model; the linear model of a node may contain only all the attributes of its subtrees, which are generated by linear regression on the subset of samples arriving at this node;

Smooth directly after pruning, and merge the linear model of the internal node into the model of the leaf node; when predicting, when the sample reaches a leaf node from the root node of the tree, only the linear model of the leaf node is used to predict the output;

The piecewise linear model formed between the physical and chemical data of all raw materials and sensory indicators is obtained, and the whole process ends.

In summary, the method for predicting sensory indicators of formula products based on the M5' model tree of the present invention has the advantages and beneficial effects of:

1. The forecasting system established by applying this kind of forecasting method can solve the artificial influence caused by the subjective factors of the existing experts when evaluating.

2. The application of this type of method is simpler, the data prediction speed is faster, and the efficiency is higher.

3. The correlation model established by this method is intuitive and clear, and can directly solve the single material and finished product quality control and grade delineation of formula products.

Description of drawings

Fig. 1 is the flow chart of the method for predicting the sensory index of the formula product based on the M5' model tree.

Fig. 2 is a flow chart of modeling and predicting cigarette flavor type using the flow shown in Fig. 1 .

Detailed ways

Example 1, as shown in Figure 1, the method for predicting sensory indicators of formula products based on the M5' model tree is applied. For the physical and chemical data prediction process related to the sensory indicators of cigarette flavor type, it is:

Detect the physical and chemical indicators and smoke analysis indicators of single-material cigarettes and finished cigarettes, organize industry experts to evaluate single-material cigarettes and finished cigarettes, and record the obtained data as a sample set for the algorithm;

Eliminate erroneous or idiosyncratic samples based on experts' industry experience;

According to the origin, grade, style and other indicators, the sorted data samples are divided into several groups of sample sets;

Perform data preprocessing on a set of sample sets, including eliminating samples with missing target values, filling samples with missing input attribute values, and converting discrete attribute values into continuous attribute values;

According to the principle of maximum information gain, select split attributes and split values, and recursively build a basic decision tree from the root node;

The basic decision tree is pruned from the leaf node recursively from bottom to top until reaching the root node. If the performance of the linear model of an internal node is not lower than the performance of this node's subtree, then this internal node becomes a leaf node containing a linear model. The attributes that may be included in the linear model of a node are only all attributes of its subtrees, which are generated by linear regression on the subset of samples reaching this node;

Smoothing directly after pruning merges the linear models of internal nodes into the models of leaf nodes. When predicting, when the sample reaches a leaf node from the root node of the tree, only the linear model of the leaf node is used to predict the output;

A piecewise linear model of all tobacco leaf physical and chemical indicators, sensory and smoke is obtained.

The mission is over.

As shown in Figure 2, the M5' model tree is used to predict the correlation between the flavor type in the sensory index of cigarettes and the CO in smoke as an example.

The M5' model of flavor type is: total sugar <= 26.1:

K<=2.19: LM1 (88/70.575%)

K>2.19:

K<=3.035:

Cl<=0.39:

Total nitrogen <= 1.85: LM2 (3/78.187%)

Total nitrogen > 1.85: LM3 (9/60.543%)

Cl>0.39: LM4 (34/98.289%)

K>3.035: LM5 (16/105.789%)

Total sugar > 26.1: LM6 (94/106.778%), of which,

LM1, flavor type=-0.0131*total sugar-0.644*total nicotine+0.0629*Shimuke value-0.1972*sugar-alkaline ratio

+7.5537;

LM2, flavor type=0.0648*total sugar-0.3288*total nicotine-0.0671*reducing sugar+1.4019*total nitrogen-

      1.3315*Cl+1.6809*K+0.0629*Shimuk value-0.0806*sugar-alkaline ratio-

      0.1932*potassium-to-chloride ratio+0.6876;

LM3, flavor type=0.0648*total sugar-0.3288*total nicotine-0.0671*reducing sugar+1.2669*total nitrogen-

      1.3315*Cl+2.1067*K+0.0629*Shimuk value-0.0806*sugar-alkaline ratio-

      0.1932*potassium-to-chloride ratio+0.0757;

LM4, flavor type=0.1171*total sugar-0.4038*total nicotine-0.0671*reducing sugar+1.5779*total nitrogen-

     0.7337*Cl+0.3629*K+0.0629*Schmuck value-0.0578*sugar-alkaline ratio-

     0.1208*potassium-to-chloride ratio+2.4177;

LM5, flavor type=0.1402*total sugar-0.156*total nicotine-0.132*reducing sugar+0.3752*total nitrogen-

        1.8351*Cl-0.3795*K+0.0629*Shimuk value-0.0522*sugar-alkaline ratio-

     0.1156*potassium-to-chloride ratio+6.4475;

LM6, flavor type=-0.0198*total sugar+0.4856*total nicotine-0.8497*total nitrogen+0.0953*Shimug value-

      0.0099*sugar-alkaline ratio + 3.9963.

The M5' model tree of fragrance type is shown in Figure 2.

Predicted by the fragrance type M5' model tree, the fragrance type score is divided into different intervals by the four attribute values of total sugar, K, Cl, and total nitrogen. The impact of the four indicators on the fragrance type in different regions is either positive or negative .

In general, the impact of total sugar on flavor type is the largest among the 9 input attributes, showing a negative correlation when the total sugar value is small and large (flavor type changes from strong fragrance to light fragrance), and the middle area is positive correlation (The fragrance type changes from light fragrance to strong fragrance). K and total nitrogen are basically positively correlated with fragrance type, and Cl is negatively correlated, which can be explained that K promotes combustion and Cl inhibits combustion. The more complete the combustion, the stronger the fragrance.

As mentioned above, it is the method for predicting sensory indicators of formula products based on the M5' model tree.

Claims (3)

1. A sensory index prediction method for realizing a formula product based on an M5' model tree is characterized by comprising the following steps: the flow of the method is that,

detecting various physical and chemical data and sensory indexes of raw materials and finished products of the formula product, evaluating single materials and finished products by an organization industry expert, and recording obtained data as a sample data set of the method;

removing wrong or specific sample data according to the industry experience of experts;

dividing the sorted data samples into a plurality of groups of sample sets according to index parameters such as producing areas, grades and styles;

performing data preprocessing on a group of sample sets, wherein the data preprocessing comprises the steps of eliminating samples with missing target values, filling samples with missing input attribute values and converting discrete attribute values into continuous attribute values;

selecting splitting attributes and splitting values according to the principle of maximum information gain, and recursively establishing a basic decision tree by a root node;

recursively pruning the basic decision tree from leaf nodes from bottom to top until the root node is reached; if the performance of the linear model of the internal node is not lower than that of the subtree of the node, changing the internal node into a leaf node containing the linear model; the linear model of a node may contain only all the attributes of its subtrees, resulting from linear regression on a subset of samples that reach the node;

after pruning, directly smoothing, and merging the linear models of the internal nodes into the models of the leaf nodes; in prediction, when a sample reaches a certain leaf node from a root node of the tree, only the linear model of the leaf node is used for predicting output;

and obtaining a piecewise linear model formed between all the raw material physicochemical data and the sensory indexes.

2. The method of claim 1, wherein the method comprises the steps of: the prediction method is to combine decision tree and linear regression to generate M5' model tree;

when M5' model tree modeling is applied, a post-pruning mode is adopted, and in the pruning stage, if the performance of the linear model of an internal node is not lower than that of the subtree of the node, the internal node is changed into a leaf node containing the linear model; the linear model of a node may contain only all the attributes of its subtrees, resulting from linear regression on a subset of samples that reach that node.

3. The method of claim 2, wherein the method comprises the steps of: said is toThe current predicted value and the smooth point predicted value p' of the current node satisfy the following expression <math> <mrow> <msup> <mi>p</mi> <mo>&prime;</mo> </msup> <mo>=</mo> <mfrac> <mrow> <mi>np</mi> <mo>+</mo> <mi>kq</mi> </mrow> <mrow> <mi>n</mi> <mo>+</mo> <mi>k</mi> </mrow> </mfrac> <mo>,</mo> </mrow> </math> Wherein,

p is the predicted value passed from the child node to the current node, q is the predicted value of the linear model of the current node, n is the number of samples to reach the child node, and k is a smoothing constant.

Images