Research and Application of User Behavior Data Analysis Technology for E-commerce

E-commerce uses computer network information to facilitate commodity exchanges [11, 12]. It relies on platforms like the Internet, Intranet, and VAN for trading and services, known as electronic and networked traditional business [13]. Unlike traditional methods, e-commerce allows non-face-to-face transactions with flexible delivery times and recorded information, enhancing contractual security.

While definitions vary by country, e-commerce's core is a business model using electronic equipment and network technology. Its market is expanding, encompassing sales, purchases, logistics, bill inquiries, and product push services [14]. E-commerce spans supply chain management, marketing strategies, electronic money exchange, electronic trading markets, network marketing, and electronic data interchange (EDI). As a prominent new industry, e-commerce leverages advanced Internet, database, and mobile technologies to broaden and deepen its industry.

Before the network communication technology developed to a certain height, the e-commerce system did not have the function of electronic payment, and the payment methods were often through telephone, mail, and bank transfer [15]. With the development of network and electronic information technology, e-commerce has broken through the limitations of traditional transaction modes and gradually formed a cross-regional, safe, and efficient electronic payment method.

The theoretical basis of user behavior analysis is the basis of studying the behavior patterns and characteristics of users on e-commerce platform. Through an in-depth study of user behavior theory, we can better understand the psychological and behavioral characteristics of users in the shopping process so as to provide effective decision support for enterprises [16, 17].

The user behavior model is a theoretical model that describes the behavior characteristics of users on the e-commerce platform. Common user behavior models include the AIDA model, DISC model, etc. [18]. These models explain and analyze user behavior from the aspects of user needs, motivation, and behavior. Understanding the user behavior model helps enterprises to better grasp user needs and provide targeted products and services. Data mining technology is one of the core technologies of user behavior analysis. Data mining technology can mine valuable information from massive data and reveal the laws and characteristics of user behavior [19, 20]. Common data mining techniques include association rule mining, cluster analysis, classification analysis, etc. Mastering data mining technology helps enterprises to extract valuable information from user behavior data and provide support for decision-making. The machine learning algorithm is the key technology to realize user behavior prediction and personalized recommendations. By training the machine learning model, the future behavior of users can be predicted, and personalized recommendations can be realized [21]. Common machine-learning algorithms include linear regression, decision trees, support vector machines, neural networks, etc. Mastering machine learning algorithms help enterprises build high-accuracy user behavior prediction models and improve the effect of personalized recommendations. The application of user behavior analysis in e-commerce mainly includes commodity recommendation, marketing strategy optimization, user satisfaction improvement, and so on. By analyzing user behavior data, enterprises can accurately position the market, formulate effective marketing strategies, and improve user experience. For example, by analyzing users' purchase history and browsing records, enterprises can provide users with personalized product recommendations. By analyzing user evaluation and feedback, enterprises can optimize product quality and improve user satisfaction.

The theoretical basis of user behavior analysis provides theoretical guidance for studying e-commerce user behavior [22, 23]. Through in-depth study of theoretical foundations such as user behavior models, data mining technologies, and machine learning algorithms, enterprises can better grasp user needs and provide targeted products and services, thereby enhancing competitiveness. At the same time, understanding the application of user behavior analysis in e-commerce will help enterprises to achieve accurate positioning of the market, formulate effective marketing strategies, and improve user experience.

Constructing features in machine learning is sometimes even more important than selecting good models and iteratively optimizing parameters because feature engineering determines the best limit that each index can reach, and better models and parameters are only closer to this limit, so it is important to dig out various effective features of users and commodities [24]. The purpose of feature engineering is to abstract multi-dimensional features that can represent data characteristics from the original disordered data heap to the maximum extent and serve as the input data of the data mining system to provide a data basis for training model parameters.

In the usual machine learning process, hundreds or even higher-dimensional features can be generated by manual extraction of features and automatic mutual fusion between features [25]. However, every time a one-dimensional feature is constructed, it is necessary to verify the validity of the feature from all aspects, that is, whether the dimensional feature is related to the result to be predicted. The higher the correlation, the more important the feature is, and it should be selected first. The importance of features can be checked from many aspects. For example, if the variance of a dimensional feature itself is approximately equal to 0, it means that all values of the dimensional feature are basically the same, and there is no discrimination between the sample wood, so the feature can be removed.

The feature engineering of recommendation systems is mainly carried out from three main dimensions: user dimension, commodity dimension, and user-commodity interaction behavior dimension [26, 27]. The main problem solved by the user dimension is that during the scoring process, the behavior habits of different users will affect the final scoring results, which are used to eliminate the differences between users. The problem solved by the commodity dimension is that in the scoring process, the attributes and characteristics of commodities will have an impact on the final satisfaction of users, which is used to eliminate the differences between commodities. The interactive behavior between user-commodity pairs shows the interest of specific users in specific commodities. The recommendation process of the recommendation system is to find out the products for which users have no interactive behavior and have high interest.

In practical applications, in order to display recommended products to users, record users' reactions to recommended products, and update recommendation results in real-time, the application architecture of the recommendation system usually includes three parts: front-end display module, background log storage module, and recommendation algorithm module. The algorithm architecture is shown in Figure 1.

Figure 1.

Algorithm architecture of recommendation system

The recommendation system is a key application in the e-commerce industry to connect users and commodities, in which the recommendation algorithm is the core. Recommendation algorithms are mainly divided into three categories: one is to recommend similar new products based on users' historical preferences; The second is to recommend untouched products according to the preferences of other users; The third is to make recommendations by mining the potential hidden factors of users and items. The regression method involves feature engineering to extract features, linear regression is used for continuous value prediction, logistic regression is used for classification prediction, and the latter realizes binary classification through logistic function.

The equation of the linear regression model is as follows (1), θ is the parameter to be learned by the model, where Θ is the parameter to be learned by the model, X (x₁, x₂,..., x_n) is the feature extracted from the user's historical behavior data set, n is the dimension of the feature, and Θ^T is the transposition of the parameter vector Θ, as shown in equation (1): 1 f(x)=ΘTX=θ0x0+θ1x1+…+θnxn

The equation of the logical function sigmoid mapping is as follows (2): 2 g(z)=11+e−z g (z) is the equation of logistic function sigmoid mapping. The algorithm of logistic regression uses logistic function for mapping on the basis of regression, as follows: 3 hθ(x)=g(ΘTX)=11+e−ΘTX h_θ(x) represents the output, and a logistic regression recommendation system is often used to estimate the click-through rate of advertisements. The key steps include extracting user historical data and item features because logistic regression is a linear model, and nonlinear relationships need to be mined through feature engineering. Firstly, features are extracted based on experience, and then a multivariate combination is carried out, and the feature combination is automatically discovered by GBDT. In e-commerce, the logistic regression system extracts user behavior, commodity attributes, and interaction as features, obtains feature weights through model training, and uses a sigmoid function to convert linear scores into binary classification results to predict user purchasing behavior. The logistic regression model is simple but requires complex feature engineering, which is suitable for researchers with industry experience. Nearest-neighbor recommendation technology is the foundation of the recommendation system, including two nearest-neighbor recommendation methods, user-based and item-based.

The user-based nearest neighbor recommendation algorithm firstly calculates the set of users who have common interests with the users to be recommended and then extracts the set of items N (i) that the users like in the user set, finds out the items i that the target users are interested in without producing the behavior to be predicted, and recommends them to the target users. Assume that the similarity between user μ and user ν is ω_μν, the interest degree of user ν in item i is γ_νi, the set size of similar users is K, and S (μ, K) represents the K users most similar to the interest points of user μ. Then, the formula (4) for calculating the interest degree of the user μ and a specific new item in the algorithm based on the user (UserCF) is as follows: 4 Iu,i=∑μ∈S(μ,K),i∈N(i)ωμvγvi

In general, the value of γ is custom-defined, and there are different numerical definition methods depending on the characteristics of the data set feedback used. The historical behavior feedback of users about items collected in the system is mainly divided into three kinds, namely, multivariate feedback, binary feedback, and univariate feedback. For example, on some e-commerce websites, users are allowed to give a 5-star rating on the purchased goods. If they are very satisfied, they can give 5 points, and if they are very dissatisfied, they can give 1 point or 0 points. This kind of feedback belongs to multiple feedback. Some video websites allow users to give binary feedback on the videos they watch. Mornay feedback is some implicit feedback of users' interest in commodities, which is manifested in users' online visit to a commodity, browsing time, and adding shopping cart behavior on e-commerce websites. If only one purchase behavior is used to define the preference degree of all users for items, γ_νi = 1 can be made, and if other unary feedback information is also used, the corresponding γ_νi can be taken as a specific weight value. The improved cosine formula (5) for the calculation of the similarity ω_μν between users is as follows: 5 ωμv=∑i∈N(μ)∩N(v)1log(1+|N(i)|)|N(μ)|*|N(v)|

For example, in the above formula, the numerator uses the reciprocal of the log function as a penalty for popular items to prevent popular items from contributing too much to similarity, which reduces the proportion of popular items in the recommendation results and enhances the ability to mine long-tail items.

The item-based neighbor recommendation algorithm is to find K similar items of the items that the target user is interested in in history, calculate the weighted sum of similar items in all the sets of items of interest, screen out the items with the highest ranking, and push them to the user. Its formulas (6)-(7) are as follows: 6 Iμ,i=∑i∈S(j,K),j∈N(μ)ωijγμi 7 ωij=∑u∈N(i)∩​N(j)1log(1+|N(μ)|)|N(i)|*|N(j)|

The log function is to punish active users and prevent active users from contributing too much to similarity. Another way to completely solve the problem of active users is to clean the corresponding data of users who are too active in the data preprocessing stage so as to avoid the unpredictable connection caused by such users with unrelated items.

Recommendation systems based on matrix factorization, as an application of the implicit semantic model, have attracted more and more attention because they can reveal implicit factors in user and item scoring and have higher accuracy. This paper will introduce several recommendation models based on factorization. Among them, the singular value decomposition technique (SVD) is the basic model, which reveals the hidden characteristics of users and items by decomposing the user's scoring matrix of items. SVD technology mines the hidden interests of users and the hidden categories of items through double dimensionality reduction and discovers hidden factors without the explicit participation of users, thus providing a better user experience. Assuming that the hidden factor dimension of the user and the item is k-dimensional, the k-dimensional hidden factor related to the item is the k-dimensional vector q_k ∈ R^k, the hidden factor related to the user is the k-dimensional vector p_k ∈ R^k, and the value of each dimension of the vector represents the importance of the factor, then the prediction score γ_μi of the new item i by the user μ can be obtained by the following formula (8): 8 γμi=qiTpμ

Among them, T stands for transposition, and the recommendation system is mainly used to predict the interaction results between users and new items. However, due to the differences between users and items, this natural and interaction-independent deviation should be eliminated in the predicted results as much as possible. Assuming that the average score of all items is v, the deviation between the score of user μ and the average score is bμ, and the deviation between the score of item i and the average score is bi, the improved predicted score result is the following formula (9): 9 γui=v+bu+bi+qiTpu

However, in practical applications, users often only rate a small number of items, so the user-item rating matrix is very sparse. When the matrix is sparse, the decomposition of SVD cannot be carried out, and the missing value of the matrix needs to be filled based on a certain method. However, the amount of filled matrix data increases greatly, which increases the difficulty of data decomposition and processing, which leads to more time-consuming data updating and upgrading, which is not conducive to the application of real-time recommendation technology. Inaccurate filling values will also make the final prediction result too biased and reduce the accuracy.

Therefore, the research focus of factorization models in recent years has focused on algorithms that do not need to fill the matrix, directly model based on observed scores and behaviors, improve the generalization ability of the model, and avoid overfitting by adding regularization terms. SVD + + belongs to this kind of algorithm, and it has been widely used recently with good effect. SVD + + algorithm is not limited to explicit feedback but considers implicit feedback information in the algorithm model because sometimes the cost of obtaining explicit feedback is too high or the amount of data is insufficient, resulting in poor prediction effect. Generally, it has a higher prediction accuracy than the SVD method. For example, we can add a factor vector of commodity i to the model. These new factor vectors can reveal the hidden characteristics of users' preferences for commodities according to users' implicit feedback information about commodities.

Equation (10) mainly considers the hidden preference degree of the number of goods rated by the user, where R (u) shows the number of all goods rated by the user. In the SVD algorithm, the user's preference is only modeled as p_u; In SVD + +, it is modeled as pu+|R(u)|12∑R(u)yj implicit feedback is explicitly modeled into the model by adding the latter term, which is the reason why this algorithm is superior to algorithms that only consider explicit feedback.

The entire model of the hierarchical recommendation system consists of an offline subsystem and a real-time subsystem. The offline subsystem is mainly used to produce basic data of users, commodities, relationships between users and commodities, etc., and is also responsible for producing basic scores of users. It can Use parallelization technology to produce background output within specified time intervals. After the offline subsystem, the real-time subsystem directly uses the results of offline calculation to personalize users and predict scores. The clustering-based hierarchical recommendation system model is shown in Figure 2.

Figure 2.

Hierarchical recommendation system model

The mouth of the grading recommendation system is used to score and predict user-commodities, and its core function is disassembled into two parts. The first part is how to generate prior knowledge, such as features and basic scores for a given application. It extracts useful features and information from users' historical behavior data in an offline environment as the first step to understanding users and forms users' prior knowledge. This part of the function is completed by the offline subsystem. The second part is how to predict users' final scores according to users' characteristics, basic scores, and other prior knowledge. Firstly, it directly uses the prior results of the offline output to mine users' personalized prosperity in real-time, predict the correlation between users and goods, and make score predictions. Then, the scoring result is weighted with the user benchmark score calculated offline to obtain the final scoring result.

In the aspect of the offline subsystem, using the historical behavior characteristics of users such as purchasing and browsing, the K-means + + clustering algorithm is used to calculate similarity degree among users, and similar user groups are clustered into small candidate set user lists, and a small candidate set list of similar users is produced. On the other hand, by extracting many important features of new users when registering, the benchmark score of users is calculated, which is used to prevent users from cold starting and supplement the scoring results of the factorization machine algorithm by insufficient prior knowledge.