WO2021232856A1 - Big data-based online sales commodity sampling and testing method - Google Patents

Abstract

A big data-based online sales commodity sampling and testing method. The method for calculating an emotion score of a certain kind of commodities comprises the following steps: B1) collecting online sales commodity links belonging to the same kind of commodities on a network platform, and data information corresponding to the commodity links and including brand data, comment data, and sales volume data; B2) by means of an emotional tendency analysis method based on a comment analysis dictionary, performing emotional analysis on each comment of each online sales commodity link collected in step B1), and calculating the emotion score of each comment under each online sales commodity link; and B3) based on the emotion score of each comment under each online sales commodity link, calculating a commodity emotion score of each online sales commodity link. Under the condition that the types of the online sales commodities are complex, and particularly under the condition that the total number of all the commodities is uncertain, the method can obtain a suitable sampling probability and a sampling amount, so that a quality monitoring or sampling and testing work is more targeted, has higher efficiency, and is more scientific and reasonable.

Description

Sampling Method for Online Product Inspection Based on Big Data Technical field

The invention relates to an inspection and sampling method, in particular to an inspection and sampling method for online merchandise.

Background technique

In the quality management process, products must be tested to determine the overall quality of the product. In testing, the number of random inspections not only closely affects the accuracy of the overall quality judgment, but is also closely related to time cost, capital cost, and personnel cost.

With the continuous development of social life, the online sales model is also constantly evolving and evolving. Online shopping has gradually penetrated into thousands of households, and more and more people accept and choose to purchase required commodities and materials through various online sales platforms. However, for online platform products, there are many problems such as multiple brands, complex product classification, uncertainty in the total number of products, and unfixed sales channels, especially the uncertainty in the total number of products, making it difficult for traditional sampling methods to perform well. It is suitable for online marketing products to determine the appropriate sampling probability or sampling quantity.

The quality control of online sales of goods has always been a core issue that quality supervision departments and online marketing platforms pay attention to. How to obtain a suitable sample number when the total number of goods is uncertain, sales channels are not fixed, there are many brands, and the classification of goods is complicated, or How to determine a suitable or smaller sample size to obtain higher judgment accuracy is a problem that needs to be solved urgently.

Summary of the invention

Objective of the invention: In order to solve the deficiencies in the prior art, the objective of the present invention is to provide a big data-based inspection and sampling method for online sales of commodities.

Technical solution: In order to solve the above technical problems, the present invention provides a big data-based online commodity inspection sampling method, which includes an emotion score calculation step and a sampling data calculation step;

The steps for calculating the emotional score for a similar product include:

B1) Collect the online merchandise links of the same kind of merchandise on the web platform and their corresponding data information including brand data, review data, and sales volume data;

B2) Through the sentiment analysis method based on the comment analysis dictionary, perform sentiment analysis on each comment of each online product link collected in step B1), and calculate the sentiment score of each comment under each online product link;

B3) Based on the emotional score of each comment under each online product link, calculate the product emotional score of each online product link;

Among them, the calculation steps of sampling data for a similar product include:

C1) According to the product sentiment score of each online product link of the same product, combined with brand data, calculate the prior probability of each online product link of each brand of the same product, and the priori probability of each brand of the same product. Probability

C2) Combining the brand data under the similar product to determine the sampling probability of each brand under the similar product;

C4) Combining the total number of products to be sampled for the same type of product, determine the number of products to be sampled for each brand of the same type of product.

As a further preference, the step B1) is: collecting online merchandise links belonging to the same kind of merchandise on the online platform and their corresponding data information including brand data, online store data, review data, and sales volume data;

After the step C2), it also includes step C3): combining the online store data of the similar product and the sampling probability of each brand under the similar product to determine the sample entry of each brand under the similar product in each online sales store Probability

After the step C4), it also includes step C5): according to the total number of products to be sampled for the similar product, determine the sample quantity of each brand of the same product in each online sales store.

Preferably, the method further includes an initialization step; wherein the initialization step includes: A0) a step of constructing and/or updating a product review analysis dictionary based on the user's multi-source review data on the product on the network platform;

The comment analysis dictionary includes an emotional word dictionary, a negative word dictionary, a degree word dictionary, and/or a stop word dictionary;

The emotional word dictionary includes a number of emotional words and the corresponding emotional word score of each emotional word;

The negative word dictionary includes several negative words;

The degree word dictionary includes several degree words and the corresponding degree word scores of each degree word;

The stop word dictionary includes several stop words.

As a further preference, said B2) performs sentiment analysis on each comment of each online product link collected in step B1) through the emotional tendency analysis method based on the comment analysis dictionary, and calculates each comment under each online product link _{In the step of sentiment score, the sentiment analysis calculation of a certain comment under a certain online marketing product link b ij} based on the sentiment tendency analysis method of the comment analysis dictionary to obtain the sentiment score of the comment includes the following steps:

B21) Clause segmentation: According to punctuation, the comment text c corresponding to a comment is divided into several clauses

B22) Modification relationship analysis: According to the comment analysis dictionary, for each clause, identify the emotional words (a ₁ ,a ₂ ,...), degree words (d ₁ ,d ₂ ,...), negative words (h ₁ ,h ₂ ,...) and stop words, and record their positions; combine the stop words to determine the target sentiment words modified by each degree word and negative word, and analyze the corresponding degree word scores and sentiment words in the dictionary with comments The score and the number of negative words determine the modification relationship between each degree word, negative word and target emotional word in the clause;

B23) Calculation of the emotional score of each clause: Determine the emotional score of each clause according to the obtained modification relationship, where the emotional score of the clause c _i is:

Among them, |H| represents the number of times the negative word appears, D represents the score of the degree word,

Represents the emotional word score of the emotional word w _{k , n w} represents the number of occurrences of the emotional word in the _{clause c i} ; among them, the emotional score of the clause c _{i where s i} is a positive value uses the clause positive emotion score

Indicates that _{the emotional score of the clause c i whose s i} is negative is used for the negative emotional score of the clause

Express;

B24) Calculation of the sentiment score of the comment: For the comment text c corresponding to the comment, add up the positive sentiment scores of all clauses to obtain the positive sentiment score s ^{+ of} the comment, and all The negative sentiment scores of the clauses in the clause are accumulated to obtain the negative sentiment score s ^-of the comment:

Among them, m _c represents the number of clauses of comment c.

Preferably, the step B3) is: based on the emotional score of each comment under each online marketing product link, combined with standardization and entropy-based weighting processing methods, calculating the product emotional score of each online marketing product link;

Among them, the product sentiment score of a certain online product link b _{ij is:}

in,

with

These are the positive and negative emotional standard scores of the online merchandise link b _ij:

in,

n _c is the total number of comments under the online merchandise link b _ij;

with

Respectively are the positive and negative sentiment scores of the kth comment c _k of the online merchandise link b _ij;

with

Are the average values of the positive and negative sentiment scores of all comments on the online merchandise link b _ij;

with

Are the standard deviations of the positive and negative sentiment scores of all comments on the online merchandise link b _ij;

Among them, α ⁺ and α ^- are the positive weight and the negative weight, respectively.

Further preferably, for the online merchandise link b _ij , the positive weight α ⁺ and the negative weight α ^- are obtained through the following steps:

K1) Perform min-max standardization on the positive and negative sentiment scores of each comment under the online merchandise link b _{ij to map the results to the [0,1] interval, including:}

Convert the positive sentiment scores of all comments under the online merchandise link b _ij into a positive indicator, where the online merchandise link b _{ij is} the positive indicator of the uth comment

for:

And, transform the negative sentiment scores of all comments under the online marketing product link b _ij into a negative indicator, where the online marketing product link b _{ij is} the negative indicator of the uth comment

for:

Among them, u=1,2,...n _c ;

Is the positive sentiment score of the uth comment of the online product link b _ij ^{; Max(S +} ) and Min(S ⁺ ) are the maximum and minimum values of the positive sentiment score of all reviews of the online product link b _ij;

Score of net sales product link b _ij u review article negative ^{emotions; Max (S -), Min} (S -) were linked network of goods sold b _ij review all negative emotion score maximum and minimum;

Positive, negative gravity to index K2) calculated network link b _ij of each pin product reviews, wherein the web link pin product u b _ij article reviews the positive, negative gravity indicator are:

K3) positive indicators calculated net sales in all product link b _ij comment entropy e ⁺ and negative indicators of entropy e ^-:

K4) the calculation of net sales product link b _ij difference coefficient under all positive indicators of the comments of g ⁺ and negative indicators of differences in coefficient g ^-:

g ⁺ ＝1-e ⁺

g ^- = 1-e ^-

K5) Calculate the positive weight a ⁺ and the negative weight a ^{-of the} online merchandise link b _ij :

Preferably, the step C1) calculates the prior probability of each online product link of each online product link of each brand of the same product based on the product sentiment score of each online product link of the same product and the brand data, and the product of the same category In the prior probability of each brand under the product,

The prior probability of the online product link b _ij of the brand B _{i is:}

P(b _ij )=(Max(x)-x _ij )/(Max(x)-Min(x))×100%

Among them, x _ij is the product sentiment score of the online product link b _ij , and Max(x) and Min(x) are the maximum and minimum product sentiment scores of all the online product links under the _{brand B i in the similar product;}

Priori probability brand B _i is:

Among them, w _j is the proportion of the product sales of the online product link b _ij of the brand B _{i under the similar product,}

n _{i is} the number of links to all online _{merchandise of the brand B i} under the same kind of merchandise.

Preferably, in the step C2) determining the sampling probability of each brand under the similar commodity,

The sampling probability _{of brand B i} under similar product G is:

Among them, P(G|B _i ) is the proportion of the sales volume _{of the brand B i} under the same product G _{, and n b} is the number of all brands under the same product G;

Said step C4) combined with the total number of products to be sampled of the same kind of commodities, determine the sample quantity of each brand of the same kind of commodities,

The sample quantity of products _{of the brand B i} under the similar product G to be inspected is:

Among them, M is the total number of products to be sampled of similar products G to be inspected, where the symbol

Indicates that the calculated number in this symbol is rounded down.

Preferably, the step C3) combines the online store data of the similar product and the sampling probability of each brand under the similar product to determine the sampling probability of each brand under the same product in each online sales store,

The sampling probability _{of the brand B i} under the similar product G in the online store T _{k is:}

_{Wherein, P (B i | T k} ) in the network for the brand B _i T _k pin proportion of sales in shops similar products G, n _t is the number of net sales shops selling the same brand of product G to B _i;

Said step C5) according to the total number of products to be sampled for the same product, determine the sample number of products of each brand under the same product in each online sales store,

The sample quantity of _{products of the brand B i} in the online sales store T _j under the similar product G to be inspected is:

Where the symbol

Indicates that the calculated number in this symbol is rounded down.

Further preferably, after determining the sampling probability of each brand of the similar product through the step C2), the following steps are further included:

C2P1) Combining the historical data of quality inspection, use the priority number R5 as the ratio to strictly deal with the sampling probability _{of selected brands B i} , B _i+1 ,..., B _{i+h under similar product G, and then treat similar products} The sampling probability of all brands under G is normalized:

in,

Here, n _b is the number of all brands under the same product G;

C2P2) Update the sampling probability of all brands under similar product G as:

P(B _k |G)=P′(B _k |G)

Among them, k∈[1,n _b ].

Preferably, the multi-source review data includes review data of several online sales platforms.

Preferably, the step B2) further includes an outlier removal step, which includes: performing an sentiment analysis calculation on each comment under a certain online marketing product link by an emotional tendency analysis method based on a comment analysis dictionary to obtain each comment After the sentiment scores, the box-and-plot method is used to remove the outliers in the positive and negative sentiment scores of each comment under the online merchandise link.

Preferably, step B4) is included after step B3): combining the product emotional score and brand data of each online product link obtained in step B3) to calculate the emotional score of each brand under the same product;

Among them, the emotional score of a certain brand B _{i under the similar product is:}

in,

x _ij is the product sentiment score of the online product link b _ij;

w _j is the proportion of the product sales of the online product link b _ij of the brand B _{i under the similar product,}

n _{i is} the number of links to all online _{merchandise of the brand B i} under the same kind of merchandise.

Preferably, the priority number

Beneficial effects: The online product inspection sampling method based on big data provided by the present invention converts the user's qualitative comments on the online product links into emotional scores indicating the quality of products and brands, and then into similar products. The prior probability of each online product link and the prior probability of each brand under the same product, so as to determine the sampling probability and the number of product samples for each brand under the same product, and the number of samples for each brand under the same product in each online sales Compared with the existing technology, the sampling probability and product sampling quantity in the store include the following advantages:

1. Based on big data, convert user reviews of online products into emotional scores that indicate the quality of products and brands, and use this as a prior probability to provide a calculation basis for subsequent inspection sampling. In the case of unfixed, multiple brands and complex product classification, especially when the total number of products is uncertain, relatively suitable sampling probability and sampling quantity can be obtained, which makes the quality monitoring or sampling work more targeted. With limited resources, the efficiency of sampling and testing has been greatly improved.

2. When performing the sampling work before the online product sampling and testing through the present invention, the sentiment score/emotional orientation analysis obtained based on user comments is first transformed into the probability of the quality of the product or brand corresponding to the online product link. When the product user evaluation under each online product link is worse and there are more dissatisfaction with the quality problem, the lower the product sentiment score of the online product link, the greater the prior probability, and the corresponding sample number of the product The higher the value, the more focused attention should be paid to the intensity of sampling; on the contrary, when the product user evaluation under the online product link is relatively better, the corresponding emotional product sentiment score is higher, and the prior probability is lower, then the corresponding product sampling The lower the number, in order to relatively reduce the intensity of sampling; the sampled data are affected by user comments and are different and have their own emphasis, which is more scientific and reasonable.

3. By transforming users' qualitative comments on online merchandise links into emotional scores that indicate the quality of the product and brand, amplifying the effect of negative reviews (ie negative emotional scores) and highlighting more problematic products and brands.

4. Further provide statistics on the sentiment scoring tendency with brand as the dimension, which can avoid the adverse interference caused by the complicated and scattered names of the online commodity links corresponding to the online commodity.

5. Further, it can greatly reduce the data impact and interference caused by useless evaluations (such as repeated evaluations, cliché evaluations, etc.) caused by phenomenon such as brushing orders. Compared with the existing praise rate or praise rating, it provides more reference significance The emotional tendency score is provided as a reference for subsequent inspection sampling.

6. Further, combined with historical quality inspection data, the introduction of a method of tightening the sampling probability of certain selected brands under the same commodity based on the priority number, makes the inspection sampling combined with real-time data and historical data, which is more reasonable More targeted.

7. User comments continue to increase in actual application scenarios. This method does not require model training in advance. It can quickly and quickly adapt to changes in the number of comments in real time. It can realize real-time collection, real-time calculation, or rolling collection, cumulative calculation and other methods, with strong real-time performance. High flexibility.

Description of the drawings

Figure 1 shows the emotional score calculation model of a similar product sold online;

Figure 2 is a schematic diagram of the emotional word dictionary in the comment analysis dictionary;

Figure 3 is a schematic diagram of the negative word dictionary in the comment analysis dictionary;

Figure 4 is a schematic diagram of the degree word dictionary in the comment analysis dictionary;

Figure 5 is a schematic diagram of the stop word dictionary in the comment analysis dictionary;

Figure 6 is a schematic diagram of the emotional score calculation process of each comment under the online merchandise link;

Figure 7 is a schematic diagram of box plots to remove outliers;

Figure 8 shows a stratified sampling model of similar products sold online;

FIG. 9 is a comparison diagram of the emotional score of each brand and the prior probability of each brand under similar products (air conditioners) in the example given in the embodiment;

FIG. 10 is a comparison diagram of the emotional score, prior probability, and sample probability of each brand under similar products (air conditioners) in the example given in the embodiment;

FIG. 11 is a comparison diagram of the sampling probability of each brand under the same product (air-conditioning category) in the example given in the embodiment before and after the strict normalization process;

FIG. 12 is a data comparison diagram of positive emotion original score (ScorePositive) and negative emotion original score (ScoreNegative) data of a number of online marketing product links b _{ij that are not standardized in the example of the embodiment;}

FIG. 13 is a data comparison diagram of positive sentiment raw scores (ScorePositive) and positive sentiment standard scores (z_ScorePositive) after normalization processing for several online marketing product links b _{ij in the example of the embodiment;}

FIG. 14 is a data comparison diagram of negative emotion original scores (ScoreNegative) and negative emotion standard scores (z_ScoreNegative) after normalization processing of several online marketing product links b _{ij in the example of the embodiment;}

15 is the plurality of net sales product link b _ij praise of (Rate) and the data comparison chart standardized and based on the respective net sales product link weighted entropy process b _ij commodity emotion score (z_Score) exemplified embodiment .

Detailed ways

The present invention will be further described in detail below in conjunction with the embodiments and the drawings, and the following embodiments do not limit the present invention.

The big data-based online commodity inspection sampling method provided in this embodiment includes an emotional score calculation step and a sampling data calculation step.

The comment analysis dictionary described in the article includes an emotional word dictionary, a negative word dictionary, a degree word dictionary, and/or a stop word dictionary. The comment analysis dictionary can be directly formed by using the sentiment word dictionary, the negative word dictionary, the degree word dictionary and/or the stop word dictionary in the prior art.

Of course, in some embodiments, the comment analysis dictionary can also be constructed and/or updated based on the user's multi-source comment data on the product on the network platform. That is to say, the big data-based online commodity inspection sampling method provided in this embodiment further includes an initialization step. Wherein, the initialization step includes: A0) The step of constructing and/or updating the comment analysis dictionary of the commodity based on the user's multi-source comment data on the commodity on the network platform.

Among them, the updated product review analysis dictionary can be updated on the basis of the review analysis dictionary constructed from the various dictionaries in the prior art, or it can be constructed based on the user's multi-source review data on the product on the network platform. Updated on the basis of the comment analysis dictionary.

As shown in Figure 2, the emotional word dictionary includes several emotional words and the emotional word score corresponding to each emotional word. As shown in Figure 3, the negative word dictionary includes several negative words. The appearance of negative words will directly turn the sentiment of the sentence in the opposite direction, and the utility is usually superimposed.

As shown in Figure 4, the degree word dictionary includes several degree words and the corresponding degree word scores of each degree word. The degree word score is a numerical value indicating the strength of the degree adverb. The data format in the degree word dictionary is shown in Figure 4. There are two columns. The first column is the degree word (also called the degree adverb), and the second column is Degree word score (also called degree value), the value>1 indicates that the emotion is strengthened, and the value<1 indicates that the emotion is weakened.

As shown in Figure 5, the stop word dictionary includes several stop words.

The above-mentioned multi-source comment data includes comment data of several online sales platforms. Several online sales platforms such as Taobao, Tmall, JD, Suning, etc.

The steps for calculating the emotional score for a similar product include (Figure 1 is the emotional score calculation level model for a similar product sold online):

B1) Collect online merchandise links belonging to the same kind of merchandise on the web platform and their corresponding data information including merchandise data, brand data, review data, and sales volume data.

B2) Through the sentiment analysis method based on the comment analysis dictionary, perform sentiment analysis on each comment of each online product link collected in step B1), and calculate the sentiment score of each comment under each online product link:

The sentiment score calculation described in the article can also be called sentiment analysis, sentiment calculation, sentiment orientation analysis and opinion mining, etc. It is the process of analyzing, processing, inducing and reasoning about subjective text with emotional color. Since the sentence structure of the product review itself is relatively simple and has strong emotional color, the emotional tendency analysis method based on the review analysis dictionary can effectively calculate the emotional tendency of the review.

As shown in Figure 6, the sentiment analysis method based on the sentiment analysis dictionary of the comment analysis dictionary _{performs sentiment analysis on a certain comment under a certain online marketing product link b ij} to obtain the sentiment score of the comment including the following steps:

B21) Clause segmentation: According to punctuation, the comment text c corresponding to a comment is divided into several clauses

B22) Modification relationship analysis: According to the comment analysis dictionary, for each clause, identify the emotional words (a ₁ ,a ₂ ,...), degree words (d ₁ ,d ₂ ,...), negative words (h ₁ ,h ₂ ,...) and stop words, and record their positions; combine the stop words to determine the target sentiment words modified by each degree word and negative word, and analyze the corresponding degree word scores and sentiment words in the dictionary with comments The score and the number of negative words determine the modification relationship between each degree word, negative word and target emotional word in the clause;

B23) Calculation of the emotional score of each clause: Determine the emotional score of each clause according to the obtained modification relationship, where the emotional score of the clause c _i is:

Among them, |H| represents the number of times the negative word appears, D represents the score of the degree word,

Represents the emotional word score of the emotional word w _{k , n w} represents the number of occurrences of the emotional word in the _{clause c i} ; among them, the emotional score of the clause c _{i where s i} is a positive value uses the clause positive emotion score

Indicates that _{the emotional score of the clause c i whose s i} is negative is used for the negative emotional score of the clause

Express;

B24) Calculation of the sentiment score of the comment (also called the summary of the comment clause tendency): For the comment text c corresponding to the comment, the positive sentiment scores of the clauses in all its clauses are accumulated to obtain the comment The positive sentiment score s ⁺ , and the negative sentiment scores in all clauses of the clause are accumulated to obtain the negative sentiment score s ^-of the comment:

Among them, m _c represents the number of clauses of comment c.

In some preferred embodiments, step B24) here also includes an outlier removal step, which includes: performing sentiment analysis on each comment under a certain online product link through a sentiment analysis method based on a comment analysis dictionary After calculating the sentiment score of each review, the box plot method is used to remove the outliers in the positive and negative sentiment scores of each review under the online merchandise link. Specifically, in this embodiment, as shown in Figure 7, the following operations are performed _{on the online sales commodity link b ij:}

1) Remove all the outliers in the positive sentiment scores of each comment under the online merchandise link b _ij

Step11. Arrange all the positive sentiment scores of each comment under the online merchandise link b _{ij in descending order to form a set}

n _c is the total number of comments under the online merchandise link b _{ij, and}

Step12. Calculate the median of ^S+

Step13. Calculate the upper quartile

Collection

When m is an even number, k=m/2, when m is an odd number, k=(m+1)/2;

Step14. Calculate the lower quartile

Collection

Median

Step15. Calculate the interquartile range

Step16. Calculate the upper edge value

Step17. Calculate the lower edge value

Step18. Determine the outliers in the positive sentiment score

And remove.

2) Remove all the outliers in the negative sentiment scores of each comment under the online merchandise link b _ij

Step21. Arrange all negative sentiment scores of each comment under the online merchandise link b _{ij in descending order to form a set}

and

. Step22 calculated S ^- median

Step23. Calculate the upper quartile

Collection

When m is an even number, k=m/2, when m is an odd number, k=(m+1)/2;

Step24. Calculate the lower quartile

Collection

Median

Step25. Calculate the interquartile range

Step26. Calculate the upper edge value

Step27. Calculate the lower edge value

Step28. Determine the outliers in the negative sentiment score

And remove.

Of course, the above-mentioned abnormal value removal can also be achieved by other methods of removing abnormal values adopted in the prior art or conventional technical means in the art.

B3) Based on the emotional score of each comment under each online product link, calculate the product emotional score of each online product link:

In this preferred embodiment, the step B3) is: based on the emotional score of each comment under each online marketing product link, combined with standardization and entropy-based weighting processing methods, calculate the product emotional score of each online marketing product link. In some preferred embodiments, it can also be said that the step B3) is: based on the sentiment score of each comment under each online product link, combined with the z-score standardization method to calculate the positive and negative values of each online product link Based on the emotional standard score, and then based on the entropy weighted processing method, the product emotional score of each online sales product link is calculated. specific:

Among them, the product sentiment score of a certain online product link b _{ij is:}

in,

with

They are the positive and negative emotional standard scores of the online merchandise link b _ij , which are calculated using the z-score standardization method in this embodiment:

in,

n _c is the total number of comments under the online merchandise link b _ij;

with

Respectively are the positive and negative sentiment scores of the kth comment c _k of the online merchandise link b _ij;

with

Are the average values of the positive and negative sentiment scores of all comments on the online merchandise link b _ij;

with

Are the standard deviations of the positive and negative sentiment scores of all comments on the online merchandise link b _ij;

Among them, α ⁺ and α ^- are the positive weight and the negative weight, respectively. In this embodiment, α ⁺ and α ^- are respectively the positive weight and the negative weight calculated based on the entropy method (it can also be said that the weight is calculated by the entropy method).

For the online merchandise link b _ij , the positive weight α ⁺ and the negative weight α ^- are obtained through the following steps:

K1) Perform min-max standardization on the positive and negative sentiment scores of each comment under the online merchandise link b _{ij to map the results to the [0,1] interval, including:}

Convert the positive sentiment scores of all comments under the online merchandise link b _ij into a positive indicator, where the online merchandise link b _{ij is} the positive indicator of the uth comment

for:

And, transform the negative sentiment scores of all comments under the online marketing product link b _ij into a negative indicator, where the online marketing product link b _{ij is} the negative indicator of the uth comment

for:

Among them, u=1,2,...n _c ;

Is the positive sentiment score of the uth comment of the online product link b _ij ^{; Max(S +} ) and Min(S ⁺ ) are the maximum and minimum values of the positive sentiment score of all reviews of the online product link b _ij;

Score of net sales product link b _ij u review article negative ^{emotions; Max (S -), Min} (S -) were linked network of goods sold b _ij review all negative emotion score maximum and minimum;

Positive, negative gravity to index K2) calculated network link b _ij of each pin product reviews, wherein the web link pin product u b _ij article reviews the positive, negative gravity indicator are:

K3) positive indicators calculated net sales in all product link b _ij comment entropy e ⁺ and negative indicators of entropy e ^-:

K4) the calculation of net sales product link b _ij difference coefficient under all positive indicators of the comments of g ⁺ and negative indicators of differences in coefficient g ^-:

g ⁺ ＝1-e ⁺

g ^- = 1-e ^-

K5) Calculate the positive weight a ⁺ and the negative weight a ^{-of the} online merchandise link b _ij :

At present, in the review data of online sales products, most of the negative reviews accounted for less, the positive or default praise accounted for more, and there were also more frequent reviews of the order. Therefore, on the one hand, the negative reviews that often act as reminders are mostly buried in Among the large number of positive reviews, on the other hand, the positive rates given by different sales links based on the proportion of positive and negative reviews are also very different. It is difficult to get close or the same positive rate (or praise) such as 98% and 99%. In this way, in the distinction of 1 to 2 points, the difference of products or services can be perceived. For some scoring systems, it is difficult to distinguish products or services from similar or the same scoring value such as 4.8 and 4.9. Differences in services.

Here, through the introduction and calculation of entropy and standard scores, the data impact and interference caused by useless evaluations (such as repeated evaluations, routine evaluations, deliberate praise or malicious bad reviews, etc.) caused by phenomena such as scalping can be greatly reduced, and the praise and The negative reviews are considered comprehensively, and the weights of positive and negative reviews are dynamically adjusted according to the situation of the negative reviews. Compared with the existing favorable ratings, favorable ratings, ratings (such as scoring ratings, star ratings, or comprehensive ratings, etc.), it provides a more differentiated and reference-meaning emotional tendency score, which is more in line with people's intuitive experience and can be selected by customers It provides a reference for the commodity. In the application of this embodiment, it is used as the basic data for inspection and sampling to provide a reference for quality inspection sampling.

Here, a simulation experiment is used to demonstrate: a number of online sales commodity links b _ij are randomly selected to simulate the above method steps. In the example of this simulation experiment, the original online sales data comes from Tmall.

The abscissas of Fig. 12, Fig. 13, Fig. 14 and Fig. 15 are all selected links b _{ij of} online sales products. In the figure, the original positive sentiment score ScorePositive of each online sales product link is the average score of the positive sentiment scores of all reviews under the corresponding online sales product link, and the negative sentiment raw score ScoreNegative of each online sales product link is the corresponding online sales The average of the negative sentiment scores of all reviews of the product link.

Take the online merchandise link b _{ij as} an example, the positive emotion standard score z_ScorePositive of each online merchandise link in the figure after standardized processing, that is, the positive emotional standard score corresponding to the online merchandise link b _ij in the article

In the figure, the negative emotional standard score z_ScoreNegative of each online product link after standardized processing is the negative emotional standard score corresponding to the online product link b _ij in the text

_{In the figure, the product sentiment score z_Score of each online sales product link b ij} after normalization and entropy-based weighting processing (also called standardized and entropy-based weighting), which corresponds to the online sales product link b _ij in the text Commodity sentiment score x _ij .

Figure 12 is a comparison of the raw scores of positive emotions ScorePositive and ScoreNegative of these selected online merchandise links b _{ij that have not been standardized} The calculated original score of positive emotion ScorePositive and the original score of negative emotion scorenegative are quite different, and it is difficult to aggregate directly.

FIG. 13 is a data comparison diagram of the raw positive emotion scores ScorePositive of these selected online marketing product links b _ij without normalization processing and the positive emotion standard score z_ScorePositive after normalization processing. It can be seen from Figure 13 that the standard score of positive emotion z_ScorePositive after normalization is consistent with the original score of positive emotion ScorePositive without normalization. While maintaining the difference and trend, the difference between the scores is reduced. At the same time, the positive score and the negative score are controlled within a similar range of magnitude, which reduces the degree of influence of the positive emotional score, and facilitates the aggregation of the positive and negative emotional scores, so as to facilitate the link between different online marketing products Compare.

FIG. 14 is a data comparison diagram of the negative emotion original scores ScoreNegative without standardization processing for these selected online marketing product links b _{ij and the negative emotion standard score z_ScoreNegative after standardization processing.} It can be seen from Figure 14 that the negative sentiment standard score z_ScoreNegative after the standardization process has the same trend as the original negative sentiment score ScoreNegative that has not been standardized, but it magnifies the difference in the link scores of different online marketing products, making it negative The positive effect is more prominent, with the same order of magnitude as the positive emotion standard score z_ScorePositive, which is convenient for summarizing with the positive emotion score to facilitate comparison between different online sales product links.

FIG 15 is a positive feedback of the selected network link b _ij of goods sold and Rate-based post-standardized and weighted entropy (and may also be referred to as a standardized method based on the weighted entropy) of each network link pin product b _ij Data comparison chart of z_Score of product sentiment score. It can be clearly seen from Figure 15 that the standard scores of different online sales product links, that is, the product sentiment score z_Score, are significantly different, which is more distinguishing than the praise rate Rate, especially the value of the praise rate of many online sales product links is the same (as shown in the figure) Some of the peaks are flat), but the value of the product sentiment score after normalization and entropy-based weighting is significantly different, which better reflects the difference in product quality under different online product links. Therefore, the original data of the simulation comes from Tmall, so the favorable rate here is the score obtained by scoring (for example, a comprehensive score with a full score of 5).

In addition, different websites/online sales platforms may have different scoring systems for the favorable rating (favorable rating), some with a full score of 5 (such as Tmall), and some with a full score of 100% (such as JD). Therefore, it is difficult to directly compare the positive ratings of different websites/online sales platforms. This method uses user reviews to calculate the product sentiment score z_Score of each online sales product link. Even if different websites, the product sentiment score z_Score of each online sales product link has the same meaning and order of magnitude, and it can be directly used between different online sales platforms. Compare.

B4) Combining the product emotional score and brand data of each online sales product link obtained in step B3), calculate the emotional score of each brand under the same product:

Among them, the emotional score of a certain brand B _{i under the similar product is:}

Among them, x _ij is the product sentiment score of the online merchandise link b _ij;

w _j is the proportion of sales of goods of the same product under the brand B _i in the net sales product link b _ij of (for brand B _i in the similar products, net sales product link b _ij cent of all the net of the similar products of the brand B _i of The proportion of the sales of the merchandise link),

n _{i is} the number of links to all online _{merchandise of the brand B i} under the same kind of merchandise.

Using the sentiment score/ sentiment orientation analysis based on user reviews, it is transformed into a probability indicating the quality of the product or brand corresponding to the online product link. This probability is based on the review history data of the user's purchase or selection or experience, which is called a priori Probability. Further use the prior probability to calculate the probability of product sampling inspection (called the sampling probability) and the number of product samples during subsequent quality monitoring (such as online product sampling). The sampling probability described in this article can also be called the posterior probability or sampling probability. The sample number of commodities mentioned in the article can also be referred to as the number of sample samples.

Among them, the calculation steps of sampling data for a similar product include:

C1) According to the product sentiment score of each online product link of the same product, combined with brand data, calculate the prior probability of each online product link of each brand of the same product, and the priori probability of each brand of the same product. Test probability:

Among them, the prior probability of the online product link b _{ij of the brand B i is:}

P(b _ij )=(Max(x)-x _ij )/(Max(x)-Min(x))×100%

Among them, x _ij is the product sentiment score of the online sales product link b _ij , and Max(x) and Min(x) are the maximum and minimum product sentiment scores of all the online sales product links under the _{brand B i in the similar product.}

The worse the product user evaluation under each online product link, the lower the product sentiment score of the online product link, and the greater the prior probability. Conversely, the better the product user evaluation under the online product link, the higher the emotional score of the corresponding emotional product, and the lower the prior probability.

Which a priori probability brand B _i is:

Wherein, w _j is the sales of goods proportion brand B _i in the net sales product link b _ij in the same product (for brand under the similar goods B _i, net sales product link b _ij representing this same product of the brand B _i of Proportion of sales of all online merchandise links),

n _{i is} the number of links to all online _{merchandise of the brand B i} under the same kind of merchandise.

C2) Combining the brand data of the similar product to determine the sampling probability of each brand of the similar product:

As shown in Figure 8, this plan adopts stratified sampling, which is mainly divided into two levels: the first level is to determine the brand selected under the same product and the sampling probability of each brand; the second level is to determine the sale of the same product The online sales stores of each brand and the sampling probability of each brand under the same product in each online sales store.

Because the sum of the prior probabilities of each brand under the same product G calculated according to the previous steps is not necessarily equal to 1, it is necessary to further calculate the probability of each brand and each online store under the same product G when sampling, that is, Probability of sampling.

Among them, the sampling probability _{of brand B i} under similar product G is:

Among them, P(G|B _i ) is the proportion of the sales volume _{of the brand B i} under the similar product G _{, and n b} is the number of all brands under the similar product G. The classification of the similar product G in the article can be large or small according to actual application needs. For example, the similar product G is positioned as an air conditioner. Of course, the similar product G can also be positioned as a vertical air conditioner or a wall-mounted air conditioner. However, products that do not belong to the same category, that is, products of different categories, and products corresponding to links of different categories of online merchandise cannot be mixed together for calculation.

At the same time, considering that quality inspection and random inspection is a continuous work, for example, relevant national administrative departments will conduct regular and irregular supervision and random inspection of product quality, and publish the results, and the information of these quality inspection historical data is very important. Carrying out later quality supervision or sampling inspection has important reference significance or guiding function. Therefore, in some preferred embodiments, the present invention further combines the historical data of quality inspection to provide a method of tightening treatment for unqualified products and/or brands in the historical data of quality inspection to increase the probability of sampling.

The specific strategy for tightening is as follows: Combining the historical data of quality inspections, if the brands B _i , B _i+1 ,..., B _{i+k of the} same product G under the historical quality inspection are unqualified in the previous year's supervision and random inspection, they will be given priority the ratio of the number of R5 is selected brand B _{i, B i + 1, ...,} B i + _{k is} the sample probability _{P (B i | G),} P (B i + 1 | G), ..., P (B i _+k |G) to tighten. After the stricter treatment, the sampling probability corresponding to the unqualified brand selected according to the quality inspection historical data will be relatively increased, that is, the probability of being selected will be increased, while the sampling probability of other brands will be relatively reduced.

Therefore, in some preferred embodiments, after the step C2) determines the sampling probability of each brand of the same product, the step of tightening and normalization is further included, which includes the following steps:

C2P1) Combining the historical data of quality inspection, use the priority number R5 as the ratio to compare the selected brands B _i ,B _i+1 ,…,B _i+h under the similar product G (that is, the unqualified brand selected based on the historical data of quality inspection The sampling probability of) is tightened, and then the sampling probability of all brands under the same product G is normalized:

in,

Here, n _b is the number of all brands under the same product G;

C2P2) Update the sampling probability of all brands under similar product G as:

P(B _k |G)=P′(B _k |G)

Among them, k∈[1,n _b ].

In this embodiment, the priority number

As a ratio, increase the sampling probability of unqualified brands in the historical data of quality inspection. In practice, the priority number can be adjusted according to the actual needs of the application.

C3) Combining the online store data of the similar product and the sampling probability of each brand under the same product, determine the sampling probability of each brand under the same product in each online sales store:

Among them, the sampling probability _{of the brand B i} under the similar product G in the online store T _{k is:}

_{Wherein, P (B i | T k} ) in the network for the brand B _i T _k pin proportion of sales in shops similar products G, n _t is the number of net sales shops selling the same brand of product B _i G.

C4) Combining the total number of products to be sampled for the same product, determine the sample number of each brand of the same product:

Among them, the sample number of products _{of the brand B i} under the similar product G to be inspected is:

Among them, M is the total number of products to be sampled of similar products G to be inspected, where the symbol

Indicates that the calculated number in this symbol is rounded down.

C5) According to the total number of products to be sampled for the same product, determine the sample number of products of each brand under the same product in each online store:

Among them, the sample number of _{products of the brand B i} in the online sales store T _j under the similar product G to be inspected is:

Where the symbol

Indicates that the calculated number in this symbol is rounded down.

Taking the similar product G as an air conditioner as an example, some steps of the above embodiment are given experimental examples to illustrate: the online sales data in this experimental example mainly comes from Tmall and JD. Through the above step B3), the calculation of this similar product (air conditioner) ) The emotional score of each online product link; calculate the emotional score of each brand of the similar product (air conditioner) through the above step B4); calculate the emotional score of each brand of the similar product (air conditioner) through the above step C1) The priori probability of each online-sold product link and the priori probability of each brand of the similar product (air conditioner); through the above step C2) determine the sampling probability of each brand under the same product (air conditioner); through the above Step C3) Determine the sampling probability of each brand of the same type of product (air conditioner) in each online store; then use the above step C4) to determine the sample quantity of each brand of the same type of product (air conditioner); and then pass the above Step C5) Determine the sample quantity of each brand in each online store for the same type of product (air conditioner).

Figure 9 shows the comparison of the emotional scores of each brand (ie the brand emotional score and the emotional total score in the figure) and the prior probability of each brand (the abscissa is each brand) under the same product (air-conditioning category). Figure 9 shows that the lower the emotional score of each brand in the overall trend, the higher the prior probability of the brand.

Figure 10 shows the comparison chart of the emotional score of each brand, the prior probability of each brand, and the sampling probability of each brand (without strict normalization processing) under the same product (air-conditioning category) (the abscissa is each brand) ), it can be seen from Figure 10 that the lower the emotional score of each brand in the overall trend, the higher the prior probability of the brand. At the same time, the sampling probability of each brand is subject to the double of the prior probability and the sales volume of each brand. Impact, some brand user reviews are very good, the prior probability will be relatively low, but if the sales volume is high (such as Oaks, Gree, Midea, etc. in the figure), the corresponding sampling probability will also increase. That is to say, if the user reviews are poor, or the users buy a lot, it needs to be spot-checked. In addition, the sampling probability in each figure is the sampling probability, which can also be called the posterior probability.

In order to illustrate the stricter treatment strategy, for similar products (air conditioners), this example combines the quality inspection historical data published in the 2018 Shanghai household air conditioner product quality supervision and random inspection results (data source: Shanghai Quality and Technical Supervision Official Website-Information Center- Bulletin Board-Spot Check Report-"2018 Shanghai Household Air Conditioner Product Quality Supervision and Spot Check Results", website link: http://shzj.scjgj.sh.gov.cn/art/2018/9/4/art_358_1325245.html) , The publicity data shows that the results of the product spot check under the product brand MBO are unqualified (here only for illustration) (another note: the basic online marketing data in this example relates to the brand MBO in the publicity data, not Involving the other two unqualified brands in the public data), the brand MBO is listed as the selected brand and its sampling probability will be tightened, and then the sampling probability of all brands under the same product (air-conditioning category) will be normalized After the processing of steps C2P1) and C2P2), the adjustment of sampling data for strict processing of substandard products and/or brands in the quality inspection history data is realized, making the subsequent quality inspection sampling more scientific and reasonable And with continuity.

Figure 11 shows the comparison chart of the sample probability of each brand under the same product (air conditioner category) before and after the strict normalization treatment (the abscissa is each brand), as can be seen from Figure 11, after the strict normalization treatment , The sampling probability of the selected brand MBO under this similar product (air conditioner category) has increased compared to before, but at the same time, since the sales volume of MBO brand air conditioners is not high, the change in sampling probability is not abrupt or obvious. Reasonable adjustment after various factors are balanced.

The above are only the preferred embodiments of the present invention. It should be pointed out that the above implementation list does not limit the present invention. Various changes and modifications made by relevant staff within the scope of the technical idea of the present invention fall under the protection of the present invention. Within range.

Claims (10)

1. A big data-based inspection and sampling method for online sales of goods, characterized in that: the method includes an emotion score calculation step and a sampling data calculation step;

   The steps for calculating the emotional score for a similar product include:

   B1) Collect the online merchandise links of the same kind of merchandise on the web platform and their corresponding data information including brand data, review data, and sales volume data;

   B2) Through the sentiment analysis method based on the comment analysis dictionary, perform sentiment analysis on each comment of each online product link collected in step B1), and calculate the sentiment score of each comment under each online product link;

   B3) Based on the emotional score of each comment under each online product link, calculate the product emotional score of each online product link;

   Among them, the calculation steps of sampling data for a similar product include:

   C1) According to the product sentiment score of each online product link of the same product, combined with brand data, calculate the prior probability of each online product link of each brand of the same product, and the priori probability of each brand of the same product. Probability

   C2) Combining the brand data under the similar product to determine the sampling probability of each brand under the similar product;

   C4) Combining the total number of products to be sampled for the same type of product, determine the number of products to be sampled for each brand of the same type of product.
2. The big data-based inspection and sampling method for online sales of goods according to claim 1, characterized in that: said step B1) is: collecting online sales of goods belonging to the same kind of goods on the online platform and the corresponding links including brand data, Data information including online store data, review data, and sales volume data;

   After the step C2), it also includes step C3): combining the online store data of the similar product and the sampling probability of each brand under the similar product to determine the sample entry of each brand under the similar product in each online sales store Probability

   After the step C4), it also includes step C5): according to the total number of products to be sampled for the similar product, determine the sample quantity of each brand of the same product in each online sales store.
3. The big data-based online commodity inspection sampling method according to claim 1, characterized in that: the method further includes an initialization step; wherein the initialization step includes: A0) Based on the user’s multi-source comment data on the commodity on the network platform, construct And/or the steps to update the product review analysis dictionary;

   The comment analysis dictionary includes an emotional word dictionary, a negative word dictionary, a degree word dictionary, and/or a stop word dictionary;

   The emotional word dictionary includes a number of emotional words and the corresponding emotional word score of each emotional word;

   The negative word dictionary includes several negative words;

   The degree word dictionary includes several degree words and the corresponding degree word scores of each degree word;

   The stop word dictionary includes several stop words.
4. The big data-based inspection and sampling method for online sales of goods according to claim 1, characterized in that: said B2) through the sentiment analysis method based on the comment analysis dictionary, for each of the links of the online sales collected in step B1) Sentiment analysis is performed on each comment, and in the step of calculating the sentiment score of each comment under each online product link, sentiment analysis is performed on a certain comment under _{a certain online product link b ij based on the emotional tendency analysis method of the comment analysis dictionary} Calculating the sentiment score of the comment includes the following steps:

   B21) Clause segmentation: According to punctuation, the comment text c corresponding to a comment is divided into several clauses

   

   B22) Modification relationship analysis: According to the comment analysis dictionary, for each clause, identify the emotional words (a ₁ ,a ₂ ,...), degree words (d ₁ ,d ₂ ,...), negative words (h ₁ ,h ₂ ,...) and stop words, and record their positions; combine the stop words to determine the target sentiment words modified by each degree word and negative word, and analyze the corresponding degree word scores and sentiment words in the dictionary with comments The score and the number of negative words determine the modification relationship between each degree word, negative word and target emotional word in the clause;

   B23) Calculation of the emotional score of each clause: Determine the emotional score of each clause according to the obtained modification relationship, where the emotional score of the clause c _i is:

   

   Among them, |H| represents the number of times the negative word appears, D represents the score of the degree word,

   

   Represents the emotional word score of the emotional word w _{k , n w} represents the number of occurrences of the emotional word in the _{clause c i} ; among them, the emotional score of the clause c _{i where s i} is a positive value uses the clause positive emotion score

   

   Indicates that _{the emotional score of the clause c i whose s i} is negative is used for the negative emotional score of the clause

   

   Express;

   B24) Calculation of the sentiment score of the comment: For the comment text c corresponding to the comment, add up the positive sentiment scores of all clauses to obtain the positive sentiment score s ^{+ of} the comment, and all The negative sentiment scores of the clauses in the clause are accumulated to obtain the negative sentiment score s ^-of the comment:

   

   

   Among them, m _c represents the number of clauses of comment c.
5. The online product inspection sampling method based on big data according to claim 4, characterized in that: the step B3) is: based on the sentiment score of each comment under each online product link, combined with standardization and entropy-based weighting Processing method, calculate the emotional score of each online merchandise link;

   Among them, the product sentiment score of a certain online product link b _{ij is:}

   

   in,

   

   with

   

   These are the positive and negative emotional standard scores of the online merchandise link b _ij:

   

   

   in,

   n _c is the total number of comments under the online merchandise link b _ij;

   

   with

   

   Respectively are the positive and negative sentiment scores of the kth comment c _k of the online merchandise link b _ij;

   

   with

   

   Are the average values of the positive and negative sentiment scores of all comments on the online merchandise link b _ij;

   

   with

   

   Are the standard deviations of the positive and negative sentiment scores of all comments on the online merchandise link b _ij;

   Among them, α ⁺ and α ^- are the positive weight and the negative weight, respectively.
6. The big data-based inspection and sampling method for online sales of goods according to claim 5, characterized in that: for the online sales of goods link b _ij , the positive weight α ⁺ and the negative weight α ^- are obtained by the following steps:

   K1) Perform min-max standardization on the positive and negative sentiment scores of each comment under the online merchandise link b _{ij to map the results to the [0,1] interval, including:}

   Convert the positive sentiment scores of all comments under the online merchandise link b _ij into a positive indicator, where the online merchandise link b _{ij is} the positive indicator of the uth comment

   

   for:

   

   And, transform the negative sentiment scores of all comments under the online marketing product link b _ij into a negative indicator, where the online marketing product link b _{ij is} the negative indicator of the uth comment

   

   for:

   

   Among them, u=1,2,...n _c ;

   

   Is the positive sentiment score of the uth comment of the online product link b _ij ^{; Max(S +} ) and Min(S ⁺ ) are the maximum and minimum values of the positive sentiment score of all reviews of the online product link b _ij;

   

   Score of net sales product link b _ij u review article negative ^{emotions; Max (S -), Min} (S -) were linked network of goods sold b _ij review all negative emotion score maximum and minimum;

   Positive, negative gravity to index K2) calculated network link b _ij of each pin product reviews, wherein the web link pin product u b _ij article reviews the positive, negative gravity indicator are:

   

   

   K3) positive indicators calculated net sales in all product link b _ij comment entropy e ⁺ and negative indicators of entropy e ^-:

   

   

   K4) the calculation of net sales product link b _ij difference coefficient under all positive indicators of the comments of g ⁺ and negative indicators of differences in coefficient g ^-:

   g ⁺ ＝1-e ⁺

   g ^- = 1-e ^-

   K5) Calculate the positive weight a ⁺ and the negative weight a ^{-of the} online merchandise link b _ij :

   

   
7. The big data-based inspection and sampling method for online sales of goods according to claim 5, characterized in that: said step C1) according to the product sentiment score of each online sales product link of the same type of goods, combined with brand data, calculates the same type of product. Among the prior probability of each online product link of each brand under the product, and the prior probability of each brand under the same product,

   The prior probability of the online product link b _ij of the brand B _{i is:}

   P(b _ij )=(Max(x)-x _ij )/(Max(x)-Min(x))×100%

   Among them, x _ij is the product sentiment score of the online product link b _ij , and Max(x) and Min(x) are the maximum and minimum product sentiment scores of all the online product links under the _{brand B i in the similar product;}

   Priori probability brand B _i is:

   

   Among them, w _j is the proportion of the product sales of the online product link b _ij of the brand B _{i under the similar product,}

   

   n _{i is} the number of links to all online _{merchandise of the brand B i} under the same kind of merchandise.
8. The big data-based inspection and sampling method for online sales of goods according to claim 7, wherein the step C2) determines the sampling probability of each brand of the same type of goods,

   The sampling probability _{of brand B i} under similar product G is:

   

   Among them, P(G|B _i ) is the proportion of the sales volume _{of the brand B i} under the same product G _{, and n b} is the number of all brands under the same product G;

   Said step C4) combined with the total number of products to be sampled of the same kind of commodities, determine the sample quantity of each brand of the same kind of commodities,

   The sample quantity of products _{of the brand B i} under the similar product G to be inspected is:

   

   Among them, M is the total number of products to be sampled of similar products G to be inspected, where the symbol

   

   Indicates that the calculated number in this symbol is rounded down.
9. The big data-based inspection and sampling method for online sales of goods according to claim 8, characterized in that: said step C3) combines the online store data of the same type of goods and the sampling probability of each brand of the same type of goods to determine Among the sampling probabilities of each brand under the same product in each online store,

   The sampling probability _{of the brand B i} under the similar product G in the online store T _{k is:}

   

   _{Wherein, P (B i | T k} ) in the network for the brand B _i T _k pin proportion of sales in shops similar products G, n _t is the number of net sales shops selling the same brand of product G to B _i;

   Said step C5) according to the total number of products to be sampled for the same product, determine the sample number of products of each brand under the same product in each online sales store,

   The sample quantity of _{products of the brand B i} in the online sales store T _j under the similar product G to be inspected is:

   

   Where the symbol

   

   Indicates that the calculated number in this symbol is rounded down.
10. The big data-based inspection and sampling method for online sales of goods according to claim 8, characterized in that: after the sampling probability of each brand of the same type of goods is determined through the step C2), the method further comprises the following steps:

    C2P1) Combining the historical data of quality inspection, use the priority number R5 as the ratio to strictly deal with the sampling probability _{of selected brands B i} , B _i+1 ,..., B _{i+h under similar product G, and then treat similar products} The sampling probability of all brands under G is normalized:

    

    in,

    

    Here, n _b is the number of all brands under the same product G;

    C2P2) Update the sampling probability of all brands under similar product G as:

    P(B _k |G)=P′(B _k |G)

    Among them, k∈[1,n _b ].

Images