Artificial intelligence and automatic recognition application in B2C e-commerce platform consumer behavior recognition

In recent years, the e-commerce industry has developed rapidly. This is one aspect of the wide application of science and technology in business. Artificial intelligence technology is now inseparable from human society, and artificial technology and human society are each other to promote progress. Voice is the carrier of language, and voice recognition is actually the conversion of voice signals into symbols that can be recognized by the system. Accuracy can be greatly improved, but also converts a variety of different languages into a universal symbol so that it can be recognized by a computer and its meaning can be understood. The development of the e-commerce industry did not explode suddenly but went through a long process. At the beginning of its development, it was mainly through the Internet to communicate with the e-commerce industry under the universal application of artificial intelligence technology. E-commerce is also combined with new technologies to continuously enrich the development model. Under this development trend, B2C e-commerce has also been greatly promoted. However, there are still many B2C e-commerce which retain the traditional business model and cannot make good use of big data. Therefore, this paper analyzes the current situation of B2C e-commerce platform, introduces artificial intelligence technology into consumer behavior recognition and analyze the urgent needs of consumer behavior recognition. For B2C e-commerce companies, in order to improve management efficiency and reduce operating costs, it is necessary to classify users in a targeted manner and find the target users they need. Combine data mining algorithms to build data models, this article effectively suggests a set of consumer behavior recognition models to help companies find consumers quickly.


Introduction
Artificial intelligence has integrated the knowledge of many disciplines, so it is very comprehensive and can be applied to various fields (Kokina and Davenport 2017). At present, the development of artificial intelligence has broken away from the stage of pure pursuit of breadth and is gradually moving to a deeper level in various fields, which makes artificial intelligence promote the production of many new technologies and new technologies (Chatila et al. 2017). Artificial intelligence is not omnipotent and perfect. The development of artificial intelligence is a double-edged sword for human society. While AI brings technological advances and improves the convenience of life, it also raises many ethical issues (Lazzeretti et al. 2022). Therefore, when researching artificial intelligence issues, to maximize the effectiveness of artificial intelligence, use artificial intelligence to promote the development of human society and weaken the harm of artificial intelligence rather than using the powerful artificial intelligence to wantonly invade others' lives and bring threats to social development (Kulkarni et al. 2022). Cloud recognition technology is also a new scientific technology in recent years. The current speech recognition technology mainly includes a large number of vocabulary continuous speech recognition systems and miniaturized and portable cloud speech recognition systems. The service targets of these two different recognition systems are different: large vocabulary continuous speech recognition & Tian Xie xietian@hngm.edu.cn systems are mainly used in the Internet, telephone networks or some large-scale voice query service systems, and these systems are often complex in structure and diverse in levels; and a miniaturized and convenient voice recognition system can be applied to mobile terminals (Johnson et al. 2014;Juang and Rabiner 2005). Now, the ways people communicate with machines are becoming more and more diversified, but how to choose the most convenient way from them is a problem we need to consider (MacArthur and Cavalier 2004). In order to solve this problem, speech recognition systems have begun to move from ordinary computer platforms to portable platforms such as mobile terminals. At present, there are still some shortcomings in the interaction between B2C e-commerce enterprises and customers. To solve these shortcomings, this paper proposes a rough set-based data mining method to identify consumer behavior (Cao 2021). This method is realized under the premise of exponential increase in data volume and rapid development of data mining. In this way, we can help companies find potential customers and effectively establish consumer behavior recognition models.

Related work
According to the current status of artificial intelligence development, the development of artificial intelligence has pressed the accelerator button for the development process of human society. The rapid development of artificial intelligence has brought great changes to human society. At the same time, the application of artificial intelligence technology can effectively improve the efficiency of various industries and reduce costs. Artificial intelligence is gradually becoming a key technology among countries competing (Duvnjak et al. 2020). However, artificial intelligence is a science and technology. He will also bring certain harm to human society. The power of artificial intelligence can convey information as quickly as possible through data mining and achieve data processing efficiency beyond human reach through machine learning (Tanoli et al. 2021). If this ability is used effectively, it can accelerate the development of information and social progress, but if it is used recklessly, it will bring great adverse effects to the society. People should fully avoid the risks faced by artificial intelligence while developing artificial intelligence so that the advantages outweigh the disadvantages (Zubatiuk and Isayev 2021). Speech is the carrier of language, and speech recognition is actually to convert speech signals into symbols that can be recognized by the system. This not only improves the efficiency and accuracy of recognition but also converts a variety of different languages into a universal symbol so that it can be recognized by a computer and its meaning can be understood. This kind of universal symbol can be accurately recognized by the computer and can ensure the accuracy of the information (Alhawiti 2015). In the information age, science and technology have made people's ways of communicating more and more abundant; natural language is the most diversified delivery method. A convenient way of interaction in this context, the literature shows that the application of speech recognition is becoming more and more widespread, and it has now been widely used in mobile terminals (Panda 2017). B2C e-commerce is an emerging e-commerce operation method in recent years. With the popularization of online shopping, more and more traditional industries have begun to transform. The literature discusses the current problems and deficiencies of B2C e-commerce companies, summarizes its development status, and conducts this research based on rough-level data mining methods (Gong-min 2010). Literature introduced what is data mining technology and explained its related theories, then described neural network and genetic algorithm according to the k-means algorithm and verified the effect of the combined model through calculation. Finally, it provides theoretical support for enterprises by analyzing consumer behavior. Literature describes the recognition system in detail based on actual applications. Finally, according to the purchase behavior of target users, data mining technology can be used to analyze data to provide theoretical support for enterprises (Hariguna 2020).
3 Distributed algorithm and robust speech recognition

Robust speech recognition
The core problem of speech recognition is to solve the problem of mismatch between model design and recognition environment and to improve the degree of matching. This article provides an in-depth analysis of why the match is not high. The accuracy of speech recognition technology extraction is promoted. Although the causes of the three types of noise are different, the principle of eliminating these three types of noise is noise robust technology. Through noise robust technology, noise can be effectively reduced to match the training model. The environment in which speech is converted from sound to data is called the acoustic environment. In this acoustic environment, there are two interference sources that affect speech recognition: additive noise and channel noise. Additive noise is generally common noise that we can hear in daily life, while channel noise mainly refers to noise that is difficult for people to hear, such as current tones or voice coding.
Additive noise can be divided into two types: stationary noise and non-stationary noise. As the name implies, steady noise means that the frequency of the sound is stable and unchanging within a certain period of time. The frequency of non-stationary noise changes regularly over time.
In order to measure the ''how much'' or ''content'' of the noise in the speech signal with additive noise, the signal-tonoise ratio is introduced, which is defined as: From the above formula, we can conclude that the larger the signal-to-noise ratio, the less the noise ''content'' in the signal. In order to be able to calculate the noise content in the speech signal more accurately, formula 4 is generally rewritten as: In the formula, k is the control factor, which can adjust the ''content'' of the noise in the noisy speech signal, and the relationship between it and the signal-to-noise ratio can be obtained according to formula 4, namely: Therefore, when a pure speech signal, background noise signal, and signal-to-noise ratio are given, we can calculate the noisy speech signal according to formula 5 and formula 7.
Speech enhancement is the most common technique for noise reduction. Speech enhancement can shield most of the noise in the speech signal and extract important information, which can effectively reduce noise and eliminate noise interference. However, the change of noise frequency makes it difficult for us to completely eliminate the interference of noise. In theory, we cannot extract completely pure voice information from the voice signal. The way that voice enhancement reduces the impact of noise is not to eliminate noise directly but to reduce the impact of noise by enhancing the quality of pure voice. Voice enhancement can firstly improve voice quality, reduce the impact of noise, and make voice information clearer, which increases the listener's acceptance of voice information to a certain extent, which can be used for human hearing perception; in addition, the increase in voice can also enable listeners to listen. Understanding the content of voice, this effect is more objective than the previous one, as shown in Fig. 1.
However, in most cases, these two goals are often not compatible.
The technology for performing speech enhancement algorithms in the frequency domain is a short-term analysis technology. When the short-term analysis technology is applied, the noise in the speech signal will be processed in frames, and then, the noise will be transferred to the frequency domain through FFT transformation. The noise will be processed in the frequency domain, and the estimated value of the noise will be increased so as to achieve the extraction of the pure signal frequency purpose, then perform the inverse fourier transform to obtain the enhanced speech signal, as shown in Fig. 2.
Suppose the noisy speech yt[n] of the t-th frame after windowing and framing is expressed as Among them, St[n] and dt[n] are the pure speech and noise of the t-th frame, respectively, and N is the speech frame length. Perform short-time Fourier transform on Yt[n] to get   If the noise suppression function G (wk), also known as the gain function, has been obtained at this time, the estimated spectrum of the pure speech can be estimated by the following formula: Spectral subtraction is one of the earliest methods used in speech enhancement. The basic idea of this method is to directly subtract the average frequency spectrum of noise from the frequency spectrum of noisy speech signals. Find the power spectrum of the 9 square If it is assumed that the speech signal and the noise signal are both zero mean, and they are not statistically correlated, that is to say, the cross term is 0, then Eq. 11 can be approximated as: In Eq. 12, we can estimate by taking the average of the previous frames of the noisy speech (assuming that the previous frames of the noisy speech segment are all noise), namely: Therefore, a more intuitive spectral subtraction can be obtained according to Eq. 12 and Eq. 13, namely: If the posterior signal-to-noise ratio Ut(wt) related to frequency is defined as: However, the posterior signal-to-noise ratio does not mean the signal-to-noise ratio. The posterior signal-tonoise ratio is a special signal-to-noise ratio. It is not only related to time but also to the frequency of the noise, so the gain function of the spectral subtraction can be obtained.
Since the power spectrum of speech is non-negative, the posterior signal-to-noise ratio must satisfy Ut(wt) [ 1. However, because the frequency of noise is difficult to determine, there will be a certain deviation. In order to reduce the influence of the deviation on the conclusion, some documents set a lower limit for it, and Eq. 16 can be rewritten as: When VAD detects that the t-th frame is a non-speech frame, use the following formula to re-estimate the noise: Smoothing between frequency components can effectively suppress noise while ensuring the quality of voice information, and enhance the auditory effect. Similarly, smoothing in time can also reduce distortion, namely: Smooth the previous multiple frames: The gain function of Eq. 17 can also be changed to any order, namely: where s[n] is the estimated pure speech signal, and there will be errors between it and the pure speech signal s[n] we expect: Obviously, e[n] is a random variable, and the basic idea of Wiener filter is to minimize the expected error square, namely: Use the minimum mean square error criterion to solve Eq. 21 and Eq. 22 and transform to the frequency domain, and finally get Then the prior signal-to-noise ratio is: Figure 3 shows the MFCC of pure speech and noisy speech. The second two-dimensional feature will change with time. The image is shown in Fig. 3. In addition, we can also see that when pure speech is disturbed by noise, there is a change in the image. When processed by CMS, the mean value of the features of both pure speech and noise is 0, but the range of values is different. For example, in (c), the output features of pure speech, it can be seen that their value ranges are roughly the same.
If the three different feature maps in Fig. 3 are drawn together, Fig. 4 can be obtained. From Fig. 4, we can see that the features of pure speech and noisy speech show similarities after special processing. Therefore, it can be inferred intuitively that CMVN is more robust than CMS and MFCC.

Distributed algorithm
If a state is randomly selected and calculated where the system is running, the convergence of the transmission efficiency of the system under different network scales can be seen from Fig. 5. As the number of iterations increases, the transmission rate of the system is also increasing. This is because the configuration of the system has been upgraded during the calculation process. The network system shown in Fig. 5 is of three different types: small, medium, and large. These three different types of network systems require different iteration times, which show that the iteration times will increase as the network scale increases.
As can be seen in Fig. 6, the speed of the algorithm is related to the network scale. As the network scale increases, the speed of the algorithm is getting faster and faster. Therefore, the results calculated by algorithms in largescale networks are more realistic and reliable.
In this section, the dynamic changes of the number of users in the environment will be simulated. In the simulation process, the unit time is cut into several identical time slices, and the dynamic correlation algorithm is run in the following two setting environments: Set I : k ¼ 3 and u ¼ 6: Set II : k ¼ 5 and u ¼ 3: Figure 7 shows the operating results of two different settings under different network scales. It can be seen from Fig. 7 that the algorithm changes the total transmission rate of the system according to different APs. 4 Online shopping consumer identification

Analysis of consumption patterns of online shoppers
The consumption patterns of online shoppers can be analyzed based on the consumption records of online shoppers and related data. By analyzing these data, basic information and characteristics of consumers can be obtained. After this information, the behavior patterns of consumers can be analyzed and make inferences. According to the promotion model of consumers, we can find that there are certain laws to follow in the consumption behavior of online shoppers. According to the consumption characteristics of online shoppers, it is possible to speculate on potential online shoppers to find potential customers for the company and increase the company's turnover. In other words, it is possible to dig out the products or services that online shoppers prefer from the consumption situation of online shoppers over the years so as to provide better services to these consumers and more similar consumers.

Data collation
The data in this article are all derived from the UCI database. The original data set is stored in this database. There are eight attributes related to online shopping consumers in all the original data sets. Due to space problems, these eight attributes will not be carried out in this article. To elaborate, we extracted a total of 4422 pieces of valid data from these data sets, and each piece of data extracted contains important information about online shopping consumers. Sampling survey is to select a part of the samples from all samples for research. From this part of the samples, the characteristics of the overall sample can be roughly inferred. This method has high practicability and simple operation, which is more suitable for this kind of research experiment with more overall samples. The sample survey is a part of the whole. Although it can roughly reflect the characteristics of the whole, it is still a non-comprehensive survey. When investigating, it is necessary to select an appropriate algorithm to calculate the selected sample, explore the characteristics of the sample through the data, and extend the characteristics to the whole.
Although comprehensive features can provide a more comprehensive overview of the characteristics of the samples, they are not applicable in this article. The main reasons for the method of sampling survey in this article are: (1) When conducting surveys by this method, all samples are randomly selected, so the uniformity of the selected samples can be guaranteed and errors can be reduced.
(2) The drawn samples can represent the whole to a certain extent. Although there are slight differences from the whole, these differences will not affect the final result.
(3) The number and attributes of the selected survey samples meet the requirements of this experiment, (1)12degrees The optimal value Algorithm 4.1 GAS RAS   I  IV  II  III  V   I  IV  II III V Fig. 6 Comparison of comparison algorithms under different network scales Fig. 7 Dynamic algorithm performance analysis Artificial intelligence and automatic recognition application in B2C e-commerce platform consumer… 7633 and the errors of the data are within the allowable range of the experiment, so the accuracy of the experimental results can be guaranteed. (4) The error of the sampling survey can be calculated, so the error of the sampling survey is small, and the accuracy is high.
In summary, we can say that the sampling survey method is the most suitable scientific method for this experiment.
In order for the data set to be used by the model, the category attributes of the data set need to be quantified. The quantitative comparison table is shown in Table 1.
The quantified sample data are shown in Table 2.

Data clustering
We divided these 226 pieces of data into 3 groups with very significant behavioral characteristics. The center points after clustering are shown in Table 3. Table 4 shows the clustering results of bank data. According to the characteristics of the average index of variables corresponding to each group combined with business analysis, the characteristics of each category obtained by subdivision can be described and named. The advantage of naming is to establish an intuitive concept of the group. Table 5 below explains the naming of the actual group after this clustering.
Through clustering and classification, the different groups are divided. The main basis of the division is the characteristics of each group, and targeted and personalized marketing strategies and methods are proposed. This method effectively transforms information into capital and becomes the tangible wealth of the enterprise. The following is an application discussion of the classification results: 1. Grouping information Quantity: 150; ratio: male: female = 1:1; average age: 29;

Analysis of online shoppers
(1) This group is more inclined to manage money than consumption; (2) The group trusts the loan and has already made the loan, so it can provide the group with credit products at a suitable price according to its real situation; (3) This group does not know much about financial management, but the income is more objective and trusts domestic products. There are very few financial matters involved in life, and the enthusiasm for financial management is not high; (4) The needs of this group for real life are much higher than those for virtual life. Compared with consumption and investment on the Internet, they prefer to focus on their own business and family steadily. They are not interested in wealth management products. They prefer practical and powerful products when shopping.

Marketing strategy
(1) The main members of this group are women, children, and the elderly; compared with the abovementioned types of people, this group of people has less demand for life, and they rarely come into contact with online shopping or financial management. They can conduct lectures or distribute leaflets. Promote products to this group of people; (2) Attract people's interest through promotion and lottery so that people can learn about the product; (3) It is not only necessary to have a head office for distribution but also to expand the scope of distribution through authorization, franchising, etc., and increase sales agency points to increase turnover.

Perform attribute reduction according to the generated categories and train the neural network
The optimized BP neural network can realize the training of classified data: input the relevant attributes of the online shopping user, and then the online shopping user can be classified and output the category of the online shopping user. The parameters of this process need to be saved for the next use. In addition, according to the classification results, the optimized BP neural network can be used to predict the group of consumers and classify consumers. Through the analysis of the same category and different categories of consumers, consumer behavior information can be obtained so as to provide companies with reliable information about consumers so that companies can understand the main factors that affect consumer behavior so as to increase consumers' Loyalty may enable companies to change their business model to make it more in line with consumer needs. In addition, by investigating the businesses that online shopping consumers use when they consume, it can help companies understand the business preferences that affect online shopping consumers and launch businesses that can better meet consumer needs. Therefore, it is necessary to classify users according to their attributes and classify consumers of the same category into the same Excel table. Relevant staff can formulate personalized marketing methods for different types of users according to the content of the form so that marketing is more in line with the needs of consumers. Since the consumption pattern prediction model is essentially from the perspective of the nature of consumption, it can also classify different online shopping consumers. This classification has been widely used in various fields, which can effectively distinguish consumer behaviors and improve enterprises. Table 6 shows Bank sample set after attribute reduction.

Conclusion
From the perspective of the enterprise, consumer behavior recognition can effectively mine customer data, thereby helping companies find potential customers. In practical applications, we find that data mining will vary depending on the recognition system. For the banking industry, Table 2 Quantified sample table of bank data set   A  B  C  D  E  F  G  H  I  J  K   30  11  2  3  1  1787  1  1  1  19  10   5 9  2  2  1  1  0  2  1  3  5  5 whether a customer purchases a bank's financial products involves data mining technology and consumer behavior recognition. In the current background of fierce international competition, the key point for the country to develop science and technology is to develop artificial intelligence technology. This is the best era and the worst era. In this era, we are faced with various development opportunities, but at the same time, we are also faced with many problems. How to effectively solve these problems is the key to restricting development. Speech is the carrier of language. Speech recognition is to convert speech into symbols that can be recognized by the computer. This kind of symbol is a general-purpose symbol that can transmit and recognize information in the computer, which can greatly improve the recognition efficiency and classification efficiency of the computer. It can also effectively enhance humancomputer interaction. Through this symbol, the computer can easily recognize voice information and convert the feedback information into voice to communicate with people, so voice recognition technology has now been widely used. The current era is the era of the Internet. The Internet drives the common development of other industries. B2C e-commerce is a new development model. As the number of online shoppers continues to increase, many traditional industries are slowly undergoing transformation, and the Internet industry is one of the main goals of the transformation. Currently, B2C e-commerce companies and customers are faced with problems such as lack of information and are facing many challenges. In order to solve these problems, this paper proposes the application of rough-based data mining methods in consumer behavior recognition. This method is carried out in the Internet where the amount of data has increased greatly, and data mining technology needs to be used to identify consumers' consumption behavior. This method can help companies find potential customers, increase their turnover,  This group of users has been paying attention to or buying products from the bank through the client, with a high degree of education, no loans, and zero high-quality users 2 Overall low-end consumer group This group of users has no chance of having bank client behaviors, and their education is generally low, and their overall level is not high 3 Overall mid-end consumer group This group of users has had clientside behaviors, but some people rejected the latest product. They have a lot of concerns in their consumption. Yes, their consumption fluctuates greatly Table 6 Bank sample set after attribute reduction s3 s5 s12 S14 S16 S17