1.

INTRODUCTION

With the popularization of mobile phones, it is convenient for users to obtain and share various data about themselves and their surroundings anytime and anywhere, and obtain various intelligent services, which has promoted the formation of LBSN. Mobile phones are widely used in daily lives and works.

How to obtain user feature is important to analyse user characteristics and provide intelligent services for them. However, in LBSN, the acquisition of user feature faces huge challenges, mainly in:

Based on the analysis above, a data driven user feature construction and analysis method is proposed in this study. The contributions in this paper including

In this manuscript, Section 2 gives the related works. Section 3 described the proposed method. Section 4 introduces the experimental evaluation, and Section 5 draws a conclusion.

2.

RELATED WORKS

User feature construction is an important research direction in LBSNs. These construction methods including two categories: explicit or implicit methods.

Explicit methods construct user features through direct communication with users⁵, which is intuitive and with a high accuracy. However, when the environment around the users is complex, it is difficult for user to describe their features clearly.

Implicit methods use multiple analyses methods to construct user features automatically. These methods including four categories: content-based, geographical-based, temporal-based, and social-based feature extraction methods⁶.

The first one constructs user features through related content of users, such as position or comments⁷. Geographical-based method aims to reveal users’ feature related to distance. Such as naive Bayesian network is used to depict the relationship between the venue that a user has been visited and the distance⁸. Temporal-based method aims to reveal user’s temporal related feature using data mining or machine learning methods⁹. Social-based methods assuming that users’ feature may be similar to their friend. Therefore, a set of the features shared by their friends are used to infer current user’s features¹⁰.

Of course, more than one of these influence factors may combined to describe user features¹¹.

Although many works devote to construct users’ features, few of these works pay attention to the applicable of feature models to a user. In this study, in the basis of the construction of multiple user feature models, an algorithm is proposed to distinguish the user set suit to a specific feature model, and a multi-channel CNN is designed to describe the suitable of different users to multiple feature models.

3.

METHOD

3.1.

Definitions

Definition 1: UCS

USC (User Context Set) is an attributes set to describe the context related to user activities.

Definition 2: VS

VS (View Set) is a perspectives set that analysts are interested in.

Definition 3: User context view feature set (UFS)

UFS (User context view Feature Set) is a feature set to describe user’s characteristic generated from user’s check-in data.

and:

3.2.

Overview of the method

The overall framework of the method is shown in Figure 1. The method is based on user check-in data. First, the set of view and user context is determined by the analyst, and the effect of the impact of context to the observer’s view is quantify analysed.

Figure 1.

Overview of the method.

Secondly, vectored user features are constructed from user check-in data. Thirdly, applicability analysis is conducted through difference value. Finally, a muti-channel CNN is designed to characterize the applicability of users to different features and used to predict user features.

3.3.

Information entropy gain-based influence analysis

The cardinal number of the influence combination |ES| is the product of |UCS| and |VS|. When |UCS| and |VS| increases, |ES| will increase rapidly, which is disadvantageous to storage and analysis. And in real life, not all user context set have a significant impact on each view. Therefore, the quantitative analysis of the impact of user context on the view is of great significance.

The concept of entropy can be used to describe the degree of chaos within the data, as described in equation (1), n describe number of categories and p_i means the probability of belonging to category i.

Entropy gain can be used to measure the effectiveness of data partitioning. The information gain and the gain rate are calculated as shown in equations (2) and (3), where A represents the attribute that divides the sample set S, Values represents the value set of the attribute A, v represents a value in the Values set, and Sv represents the sample set corresponding to the value v after the division.

Based on the analysis above, this study uses information gain rate to find the UCⁱ that has a significant impact on V^j based on the data set.

3.6.

Unified model—muti-channel CNN

Multichannel neural networks can effectively describe the local salience features of data, identify and analyse them, and then stack these different channels using a deep structure to support the fusion of multiple salient features. This feature is suitable for describing the user’s adaptability to muti-perspective features; therefore, this study designs a muti-channel convolutional neural network to analyse the user’s personalized features. The basic network structure is illustrated in Figure 2.

Figure 2.

Muti-channel CNN structure diagram.

The network contains |ES| channels, and the input is user set US_UCVF^ij, which is suitable for the UCVF^ij model and the matrix UCVF^ij for these users. After learning the user and the user’s matrix through different channels, the network can predict the user’s activities during a given context.

4.

EXPERIMENT

3.8.

Research question and evaluation plan

RQ 1: Is the assumption established that user is applicable to a feature model when the difference value in UCVF^ij is small?

RQ 2: Does applicability analysis help to improve the accuracy to predict user behaviours.

RQ 3: Does the unified model improve the prediction accuracy? How does it compare with existing methods?

Accuracy of top K is used to evaluate the effectiveness of the method, which is shown in equation (4):

where u refers to user, l refers to location, t refers to time, a refers to an activity, and TS stands for test set.

4.

EXPERIMENT RESULTS AND DISCUSSION

In this experiment, UCS = {UC^t, UC^d}, where t represents time and d represents distance. |UC^t|=24 because the time is depicted in hours. |UC^d|=4 because the distance is divided into four levels, which are within 1 kilometre, between 1 and 10 kilometres, between 10 and 30 kilometres, and more than 30 kilometres, so.

Based on the elements in the data, VS = {V^r, V^c}. The parameter r is a representation of root category, and the parameter c is a representation of category. | V^r | =9 and | V^c | =65 because the number of root categories and categories of POI in the experiment is 9 and 65. Therefore, UFS = {UCVF^{time-root category}, UCVF^{time-category}, UCVF^{distance-root category}, UCVF^{distance-category}}, which correspond to 24*9 matrix, 24*65 matrix, 4*9 matrix, and 4*65 matrix.

After UFS is constructed, the user set suit for each feature is analyzed, as shown in Table 2.

Table 2.

Users suit to different UCVFij

A multi-channel CNN is designed to depict the adaptability of a UCVF^ij to a user, and the network is used to predict user activities. Experimental results were presented and discussed as follows.

RQ 1: Is the assumption established that user is applicable to a feature model when the difference value in UCVF^ij is small?

The difference value was divided at intervals of 10, and the TOP K accuracy rates was analysed. The results are shown in Figures 3 and 4.

Figure 3.

Accuracy rate of dataset1.

Figure 4.

Accuracy rate of dataset2.

These figures show that in both data sets, when the difference value decrease, the accuracy increases, which means the assumption is established.

RQ 2: Does applicability analysis help to improve the accuracy to predict user behaviours.

To answer this question, UCVF^{time-root category}, UCVF^{time-category}, UCVF^{distance-root category} and UCVF^{distance-category} were used separately to predict user’s activity, and the accuracy rate is generated. After that, uses the suit UCVF^ij to describe corresponding users, and the accuracy rate is generated using equation (4), results are depicted in Figure 5.

Figure 5.

Effectiveness of applicability analysis.

The result shows that in both data sets, the predication of user’s suited UCVF^ij is of higher to using each UVCF^ij separately.

RQ 3: Does the unified model improve the prediction accuracy? How does it compare with existing methods?

• Baseline

User activity prediction can be divided into the next prediction and any time prediction in terms of timeliness, coarsegrained and fine-grained predictions in terms of the granularity¹⁴.

This research belongs to any time and coarse prediction in terms of timeliness and granularity. We adopt the HOSVD method, PFR method, PCLR method and STAP method as baselines for comparison^{12, 15-17}.

• Results

The proposed method was compared with baseline methods according to Top-K accuracy. The value of K in the Top-K accuracy rate formula is 1 and 5. the results are as follows in Figures 6 and 7.

Figure 6.

Accuracy comparison of Top 1.

Figure 7.

Accuracy comparison of Top 5.

From the graph, we can see that the adaptive analysis method and the muti-channel CNN method outperforms the baseline methods in both dataset1 and dataset2.

5.

CONCLUSIONS

Multiple user feature was constructed, and whether these features are suit to user is quantitative analysed. Furthermore, a multi-channel CNN is designed to depict user’s applicability to the feature models. The experiments result shows the effectiveness of the method.

Although the method is effectiveness in the experiment, it should be further evaluated in a more realistically data and scenarios. In addition, the feature model constructed now is data-sensitive, how to generated a more stable feature of the users in an abstract level is an interesting research direction.