Keywords

1 Introduction

The analysis of users’ network behavior characteristics is the design basis of many Internet products. Through in-depth analysis of user behavior, personalized recommendation can bring users a better application experience. In the field of market driven software engineering, user behavior analysis also provides new ideas and improvement directions for application development to meet the requirements of the new situation.

User network behavior can be divided into explicit feedback behavior and implicit feedback behavior. At present, a relatively stable and unified view has been formed on the definition, characteristics, differences and types of the two types of behavior. The display feedback behavior data can accurately express the user’s intention, but it interferes with the user’s normal interaction process in the network, increases the cognitive burden and reduces the user experience, so it is difficult to obtain the data. On the contrary, for the implicit feedback behavior data of users, it is much less difficult to obtain, and the information abundance is large. Therefore, although such information has low accuracy, large data noise, large context sensitivity, this research field is still getting more and more attention.

The research on recommendation methods based on user implicit feedback behavior has made some progress in recent years. Such research relies on user browsing, attention, purchase, transaction and other key intention behaviors to complete commodity recommendation, without fully considering the context of implicit feedback behavior. At the same time, some recommendation systems also explore the direct application of context information, especially time and location context, to recommendation systems, and have made some progress. In addition, by mining the interaction data of network applications in different context, collecting user network activity logs and questionnaires, some research results have been accumulated in understanding user network behavior, and some of them have been applied to the field of software design and human-computer interaction, However, such achievements have not been well extended to the field of personalized recommendation. In this work, we take context implicit feedback behavior personalized recommendation as a whole to supplement the previous research work.

Users’ implicit feedback network behavior is easily affected by the context of time, environment, user attributes, application content, interactive terminal, personality and emotional state. Especially for mobile intelligent terminals, the context sensitivity of implicit feedback network behavior is more prominent due to the scattered use time period, changeable environment, diverse crowd attributes and different device terminals. When using the implicit feedback behavior of mobile intelligent terminals for content recommendation, the recommendation results also show a certain sensitivity to the context. Therefore, it is more necessary to discuss the impact of context differences on the implicit feedback behavior applied to personalized recommendation.

2 Related Work

With the rapid development of social networks and e-commerce, the number of Internet users has greatly increased, and the demand for personalized recommendation services is also increasing. Accurately and effectively deal with the massive multi-source heterogeneous data generated by users browsing the mobile Internet is the focus and difficulty of the current research.

The original personalized recommendation service is mainly for PC based users. The relevant research is mainly divided into the following four aspects: Research on a certain application scenario, research on a certain class or technology, research on evaluation methods of recommendation system, and research on a certain kind of common problems in the recommendation system.

The study of user network behavior was initially applied in the field of information retrieval, which significantly improves the performance of information filtering compared to other feedback, and quickly filters from massive information sets, providing the retrieval set [1] with the highest correlation with their interest preferences. By comparing the results of user browsing time preference analysis with user explicit ratings, Morita [2] found the fact that users spend more energy and much longer time reading the preferring tidings on newspaper than regular tidings, representing user browsing time is a available information showing the user’s interest preferences. Konstan [3] applied Usenet News with browsing time-based collaborative filtering methods in 1997. Moreover, Oard and Kim validated the behavior when browsing a website like bookmarking, printing and saving could show user interest preferences and could be used to compensate for insufficient explicit feedback score data. While the Internet develop rapidly, the increase of the number of users, the data overload problem get significant. And the stability of the recommending results accuracy of relying merely on the behaviors which are called explicit feedback decreases, and the significance and requirement of the behaviors which are called implicit feedback, for example, exploring the website behavior in personal recommending models increase. When lots of scientists invest in implicit recommendation study, there are also ordinary solutions in manufacturing. Moreover, the behavior estimation from user website exploring in the recommending system with implicit cues is the most significant one of its core. In the Oard and Kim [4] and Kelly [5] opinions, who research on the website exploring behaviors, there are three groups about the user browsing behavior. They are saving behaviors [6], operational behaviors, and repetitive behaviors. a) the first behavior type- save: it includes download behavior, collection, printing, subscribe to, and bookmarks adding or deleting; b) the second behavior type- operation: it includes mouse clicking, searching information, browsing time on one web page, scroll bar dragging, page size adjusting, and copy data behavior; c) the third behavior type - repeat: it includes accessing a website or web page repeatedly, purchasing goods repeatedly, click on a item repeatedly.

Anyway, insufficient researches about the behaviors of website exploring that indicates user’s favorite. While users change their interaction devices in particular, their website exploring behaviors in the mobile network environment may be different. Therefore, carrying on studies about micro network behaviors with implicit attributes is essential.

By analyzing these behavioral data, we can obtain the behavioral habits of mobile users, which are helpful to enhance the servicing character and users’ enjoyment. Depending on the users’ website or web page exploring behaviors in mobile condition, paper [7] studied personal recommending method. In addition, group recommending method and the mining algorithm of uncertain attribute were also considered. The results were good. Relevant research focused on the direction of recommendation system. The website exploring behaviors in mobile condition is not deeply studied. Literature [8] combines users’ website exploring behaviors including mobile location data to analyze the influence from scenes and studied the users’ website exploring behaviors in the dimension of space and time. Not only it concerned users’ web page exploring behavior, but also it pays attention to the users’ mobile behavior. The researches about implicit behaviors are hot [9,10,11]. According to the statement above, this paper has finished the following work: 1) investigation of user micro network implicit feedback behavior for mobile intelligent terminal. 2) The influence of user attribute context on the implicit feedback behavior of user micro network.

3 Problem Description and Correlation Analysis

3.1 Problem Description

Users’ network implicit behavior contains their preference information, but it is generally not clearly expressed, so it is difficult to correctly judge their preferences. Researchers have done more work in this regard. At present, there are many researches on macro network implicit behavior, such as behavior sequence analysis or item recommendation based on browsing, adding shopping cart, shopping and so on. For the implicit feedback behavior of user micro network, there are few relevant studies and conclusions due to the problems of small data scale, few data categories and low data dimension. This paper intends to analyze the implicit feedback behavior of users in micro networks, focusing on the relationship between user attribute context and micro network implicit behavior.

3.2 Users’ Micro Implicit Behavior

Acquiring approach of users’ micro implicit behavior includes two ways. The first one is direct acquiring way, which is conducted by running some software in background. The other is indirect way, generally speaking, which is acquired by questionnaire. In direct acquisition, there are some problems such as sparse data, few categories and low dimensions, which is not conducive to subsequent analysis and deterministic conclusions. This paper analyzes the micro implicit feedback behavior by using the data obtained indirectly. Based on the questionnaire in literature [12], some survey contents (Q4–Q15) are extracted from the questionnaire, in addition, matched to users’ micro implicit behavior, which is demonstrated as below in Table 1.

Table 1. Micro implicit behavior.
Full size table

For the sake of easing the correlation analysis about influenced factors and users’ implicit behavior, according to the questionnaire data in literature [12], this paper divides users’ micro implicit feedback behavior into two categories: 1) mutually exclusive type, and 2) non-mutually exclusive type. In Table 1, IFB1-IFB3 are commonly clustered into the mutually exclusive type, which means every behavior exists once. For example, there are selecting only one application market category, determining a certain frequency of access and attention to element. IFB4-IFB11 belongs to non-mutually exclusive type. Ever person could select multiple behavior. For example, the frequency of inquiring items, while the person is discouraged or bored, or desires to accomplish a duty, etc.

3.3 User Attribute Context

To study the relationship between user attribute context and micro network implicit behavior, it is necessary to determine the content of user attributes. Based on the questionnaire in literature [12], the determined user attributes are shown in Table 2.

Table 2. User attributes.
Full size table

3.4 Correlation Analysis Between User Attribute Context and Implicit Behavior

This paper researches on the relations of users’ characteristics and micro implicit behaviors. That is, in the view of users’ characteristics, impact on users’ micro implicit behaviors is discussed. In addition, big impact factors are chosen. As statements earlier about users’ micro implicit behavior and users’ characteristics data, this paper selects IFB1-IFB11 as the dependent variable and user attributes Q17-Q27 as the independent variable, and uses logistic regression to complete the correlation analysis between users’ characteristics background and implicit behaviors.

Multiple Logistic Regression Analysis.

Through the observation of dataset, the type of IFB1, IFB2 and IFB3 is multi-classified micro implicit behavior, in which IFB1 is a disordered variable and IFB2 and IFB3 are ordered ones. Multiple logistic regression analyzing method is used to study the impact on micro implicit behaviors from users’ attributes.

Binary Logistic Regression Analysis.

Based on the observation of the data, IFB4-IFB11 is consist of multiple subsets. Moreover, this type of behaviors is described as binary. Therefore, binary logistic regression analyzing method to study impact on micro implicit behavior from users’ attributes is used in this paper.

4 Results and Discussion

4.1 User Attributes and Influencing Factors of IFBn

According to the significance index of model fitting, shown in Table 3, the fitting models of IFB1 and IFB3 are statistically significant and pass the test. The Pearson Chi-square significance of IFB1 model is 1. The model fitting status, as described in the column, to initial data passes the test. However, its pseudo r square value is flat, and the fitting degree is not actually distinguished.

In accord with the significance of likelihood ratio test in Table 4, for the micro implicit behavior IFB1, there exists results as below: eight user attribute influencing factors such as age, marital status, current country of residence, first language, years of education, physical barrier, current employment status and occupation all contribute significantly to model configurations, which is the crucial component effecting IFB1.

Table 3. Fitting information and forecast percentage (IFB1-IFB3).
Full size table
Table 4. Likelihood ratio test significance (IFB1-IFB3).
Full size table

In agreement with the exhaustive test dataset of model factors in Table 5, for the type of IFB4, the fitting mode of these micro implicit behaviors is commonly essential. Meanwhile, goodness of fit test and prediction correct percentage information show that, considering the IFB4 subgroup, the model fitting goodness of IFB4-1, IFB4-3 and IFB4-6 behavior subset is higher and the fitting model is better.

Table 5. Model sparsity test, goodness of fit and prediction percentage (IFB4).
Full size table
Table 6. Variable significance (IFB4).
Full size table

According to the significance index of each variable in Table 6 (in which the gray shadow part commonly shows the significance index >0.05), the micro implicit feedback behavior of item query frequency (IFB4) as a whole, age, current country of residence and current employment status are the main influencing factors of user attributes. Specifically, for the behavior subset IFB4-1 of micro implicit feedback behavior IFB4, four user attribute influencing factors such as age, marital status, current country of residence and current employment status contribute significantly to the model configurations and are the important factors impacting IFB4-1. Given the type of IFB4-3 behavior subset of micro implicit feedback behavior IFB4, age, current country of residence and years of education are the main factors affecting IFB4-3. For the behavior subset IFB4-6 of micro implicit feedback behavior IFB4, two user attribute influencing factors, nationality and current employment status, contribute significantly to the model configurations and are the important factors impacting IFB4-6. Analysis about user attributes and influencing factors of IFBn (n = 5–11) is similar as above.

4.2 Influence Ranking of User Attributes

Through the above analysis of user attribute influencing factors that make a significant contribution to user micro implicit feedback behavior IFBn, the ranking of influencing factors is obtained, as shown in Table 7. It can be seen that the user’s age attribute has a great impact on the micro implicit feedback behavior. The user attributes such as the current country of residence, the first language and the current employment status also affect the user behavior to a certain extent.

Table 7. User attribute impact.
Full size table

5 Conclusion

This paper analyzes the user micro network implicit feedback behavior of mobile intelligent terminal, and studies the influence of user attribute context on the user micro network implicit feedback behavior. The results reveal that users’ age attributes, regional attributes and professional attributes will have an impact on users’ behavior. The outcomes above establish a groundwork for future researches around users’ micro implicit behavior data in recommendation area.