Fuzzy trust based collaborative filtering analysis for mobile user preferences

Abstract

With the rapid development of mobile terminal devices, mobile user activities can be carried out anytime and anywhere through various mobile terminals. The current research on mobile communication network is mainly focused on extracting useful and interesting information for mobile user from massive and disordered information. However, the sparsity of scoring data matrix results in low quality of recommendation algorithm. In order to overcome this drawback, the traditional collaborative filtering algorithm is improved. First, the user-interest matrix and item-feature matrix were obtained by analyzing mobile user behavior and item attributes. Fuzzy trust based model is utilized for collaborative filtering analysis for mobile user preferences. Then, the similarity between different mobile users was calculated by weighted calculation. With this method, mobile user preference can be predicted effectively, making it possible to recommend rational resource and waste less time in extracting resources out of the massive information. Experimental results show that the proposed algorithm reduces the mean absolute error (MAE) and the impact of sparse scoring matrix data compared with the traditional collaborative filtering algorithm, and improves the recommendation effect to a certain extent.

Keywords

Collaborative filtering AI mobile user user interest similarity calculation fuzzy trust

1 Introduction

The continuous technology development of mobile terminal devices is a mixed blessing: it brings a variety of mobile network information to users, and produces serious information overload. As one of the most viable solutions to information overload, the mobile user preference analysis has become a focused topic in the mobile user behaviour recommendation [1]. Compared to the traditional network, the mobile communication network has its own characteristics, such as limited resources and restricted input and output. In order to provide better personalized recommendation, it is imperative to pinpoint the information that truly interests mobile users among the mass information.

The information overload can also be resolved by personalized recommendation, which is gradually concerned by scholars. The key to personalized recommendation lies in accurate identification of user preferences, for the information needs to be recommended to current mobile users based on historical behaviour preferences of users or those around them [2 –4].

Collaborative filtering is one of the most popular recommendation algorithms for user behaviour preference analysis. Despite the ability to disclose the potential interest of the target user, the collaborative filtering algorithm faces some intrinsic problems, e.g. the data sparsity caused by zero elements in scoring matrix. The trust model in a mobile environment is hard to handle due to involving of uncertainties. The theory involving fuzzy logic extends the explanation of mathematical research to be compound which leverages quantity and qualities, and which includes certain fuzziness. Introduction fuzzy logic into the trust management by associating the collaborative filtering. The fuzzy logic model is a methodology which involves the relationship in association with fuzzy concepts and logics using fuzzy model.

To overcome the above defects and get more useful and accurate mobile user behaviour preferenszxytcf@126.comces, this paper improves the traditional collaboration filtering algorithms for predicting the behaviour preference of mobile users. First, a user-interest matrix was constructed by analysing the explicit attributes and implicit attributes, and an item-feature matrix was obtained after investigating item attributes and the user’s evaluation for item. Then, similarity of mobile users was calculated respectively based on user interest and item feature. Finally, the similarity of mobile users was obtained by weighted calculation. Roan et al. [5] proposed a novel Intuitionistic fuzzy system for application in decision making. Devi et al. [6] proposed clustering approach for cluster selection based on Fuzzy C means. Roopa et al. [7] utilized information fuzzy network for analysis of ECG.

2 Theory of collaborative filtering

Proposed by Goldberg et al., the concept of collaborative filtering recommends resources to the target user through following steps: First, the target user’s historical preference is acquired explicitly or implicitly, and compared to the preferences of the other users; the nearest neighbours are identified by looking for those with similar preferences to the target user; the weighted evaluation value of the resource by the nearest neighbours is taken as the target user’s evaluation of the resource. So far, much research has been done on collaborative filtering, leaving a lot of valuable information for reference [8 –10].

The collaborative filtering is a three-stage process [11, 12]:

User information representation. The user information can be accessed in an explicit or implicit manner. In the explicit method, the recommendation system can provide a list of items for users to evaluate according to their interests; in the implicit method, the user interests are analysed based on his/her browsing behaviour, including browsing time, purchase records, clicks, and other information. The user information is usually represented in the form of score vector. Each vector reflects user’s score on an item. Score vectors of all users constitute the user-item scoring matrix.

Selection of the nearest neighbours. The most important step of collaborative filtering is to determine neighbours of the target user. The neighbours refer to those users with very similar interests to the target user. Similarity value sim(a,b) indicates the similarity between two users. For target user u, the nearest neighbours are selected to form a neighbour set, N_u = N₁, N₂,…, N_m, where sim (u, N₁) > sim (u, N₂) > …> sim (u, N_m).

Generation of recommendation results. The nearest neighbours of target user u is expressed as N_u =N₁, N₂... N_m, which is possible to generate the set of the potential scores given by the target user to all non-scored items S_u = S_u,1, S_u,2,…, S_u,k, and the first N items in S_u should be recommended to the user as the Top-N recommendation set.

3 Mobile user preference predicting based on collaborative filtering

In this paper, the algorithm flow of mobile user preference predicting is as shown in Fig. 1.

Fig.1

The algorithm flow of mobile user preference predicting.

3.1 Establishment of mobile user information

The mobile user data model was established according to the use records and evaluations of relevant mobile apps by mobile users. The data source is expressed as D = (U, I, R), where U represents the set of basic users, I is a set of items, and R is the m*n scoring matrix of the target user for all items. The number of mobile users and the number of items is denoted as m and n, respectively. The element r_ij shows the score of item j given by the mobile user i in the set U.

3.2 Calculation of similarity between mobile users

3.2.1 Similarity calculation based on mobile user interest

Interests of mobile user are mainly acquired based on user attributes [13 –15]. The user attributes mainly fall into two categories: explicit attribute and implicit attribute. The explicit attribute is relatively easy to obtain from the registration information, including the age, gender and hobbies of the mobile user, while the implicit attribute is calculated according to the number of visits, browsing time, and the user’s bookmarking, scoring and evaluation.

Definition 1. Number of mobile user visits is expressed as $B_{u, i} = \frac{B_{i}}{B_{u}}$ , where B_u is the total number of visits that the mobile user u has paid to items, and B_i is the number of visits that mobile user u has paid to the item i. Since B_u is basically constant for some time, the greater the value of B_i, the more the item is favoured by the mobile user.

Definition 2. The browsing time of mobile user is expressed as $T_{u, i} = \frac{T_{i}}{T_{u}}$ , where T_u is the browsing time of the mobile user u, and T_i is the browsing time that the mobile user u spent on the item i.

Depending whether the mobile user bookmarks an item, the interest degree of mobile user on the item is expressed as C_u = 1, or C_u = 0. Score of mobile user on the item can be obtained through the user-item scoring matrix.

By analysing implicit attribute, it is possible to obtain the user-interest matrix, which reflects the mobile user’s favourite types of items. The similarity formula for the mobile user a and the mobile user b is calculated as follows [13]: ${sim}_{I} (a, b) = \frac{\sum_{t \in T} (r_{a, t} - \bar{r_{a}}) \cdot (r_{b, t} - \bar{r_{b}})}{\sqrt{\sum_{t \in T} {(r_{a, t} - \bar{r_{a}})}^{2} \sum_{t \in T} {(r_{b, t} - \bar{r_{b}})}^{2}}}$ (1)

Where T is the user’s attribute set, including explicit and implicit attributes; r_a,t and r_b,t are values of user attribute t given by the mobile user a and the mobile user b, respectively; $\bar{r_{a}}$ and $\bar{r_{b}}$ are average values of all user attributes given by the mobile user a and the mobile user b, respectively.

3.2.2 Similarity calculation based on item attributes

Two different mobile users are deemed to bear high resemblance in interests and hobbies if they have graded the same score on a certain type of item in the user-item scoring matrix.

Assuming that a set of item attributes I = C₁,C₂, ... ,C_m contains the attributes of each item, the set can be rewritten as a matrix called the item-feature matrix $M = [\begin{matrix} m_{11} & m_{12} & \dots & m_{1 t} \\ m_{21} & m_{21} & \dots & m_{2 t} \\ ⋮ & ⋮ & ⋮ & ⋮ \\ m_{n 1} & m_{n 1} & \dots & m_{nt} \end{matrix}]$ , where m_ij is the relationship between item I_i and attribute C_j. Item I_i has the attribute C_j if m_ij = 1, and does not have the attribute If m_ij = 0.

Thus, the preference degree P_u,i of mobile user u for attribute C_i in the item set I can be defined as: $P_{u, i} = \frac{\bar{S_{u, i}}}{\bar{S_{u}}}$ (2)

Where $\bar{S_{u, 1}}$ is the average score of item given by the mobile user u, and $\bar{S_{u}}$ is the average score of all items given by the mobile user u.

Assuming that the items fall into N categories, the preference degree of each mobile user for each type of item can be calculated by formula (2). The result can be expressed as D_u = (P_u,1, P_u,2, ... , P_u,n,). Based on the mobile user’s interest for item attributes, the similarity between the mobile user a and the mobile user b is calculated as follows: ${sim}_{P} (a, b) = \frac{\sum_{i = 1}^{n} (P_{a, i} - \bar{P_{a}}) \cdot (P_{b, i} - \bar{P_{b}})}{\sqrt{\sum_{i = 1}^{n} (P_{a, i} - \bar{P_{a}})^{2} \cdot \sum_{i = 1}^{n} (P_{b, i} - \bar{P_{b}})^{2}}}$ (3)

Where P_a,i and P_b,i are the preference degree of the mobile user a and the mobile user b, respectively, for attribute C_i in item attribute set I; $\bar{P_{a}}$ and $\bar{P_{b}}$ are the preference degree of the mobile user a and the mobile user b, respectively, for all items.

Owing to the sparsity of mobile user’s evaluation and scoring in real life, the weighted calculation method was introduced to reduce the noise brought by sparse data on the basis of the user interest and item attribute. After the removal of noise, the similarity of different mobile users is defined as follows: $sim (a, b) = λ \times {sim}_{I} (a, b) + (1 - λ) \times {sim}_{P} (a, b)$ (4)

Where sim_I(a,b) is the similarity based on mobile users’ interest, and sim_P(a,b) is the similarity based on mobile users’ preference for items.

3.3 Prediction of mobile user preference

Through the formula (4), we can get the similarity between the target mobile user u and other mobile users. The other mobile users can form the neighbor set N, in which all members are ranked in descending order by similarity. Before recommending unscored items to the target mobile user, it is necessary to predict the scores of non-scored items. The value that target mobile user u may give to non-scored item i can be predicted by the formula below. $S_{u, i} = \bar{r_{u}} + \frac{\sum_{v \in N} sim (u, v) \times (r_{u, v} - \bar{r_{v}})}{\sum_{v \in N} sim (u, v)}$ (5)

Where $\bar{r_{u}}$ represents average score of evaluated items given by the target mobile user u; r_u,v represents score of the item i given by the neighbor v; $\bar{r_{v}}$ represents average score of evaluated items given by the neighbor v.

The score the prediction formula (5) makes it possible to calculate the score of each unscored item that the target mobile user may give to each item. In the end, the final recommendation list is formed according to the scores.

3.4 Prediction algorithm of mobile user preference

Input: user item scoring table, user interest table, item attribute table, parameter, target mobile user.

Output: predicted scores given by the target mobile user.

User-item scoring matrix, user-interest matrix and item-feature matrix are calculated according to the user item scoring table, user interest table, item attribute table, respectively.

The similarity sim_I(a,b) based on user interest is calculated according to the formula (2).

The similarity sim_P(a,b) based on item attribute is calculated according to the formula (3).

The final similarity sim(a,b) of mobile users is calculated according to the formula (4).

On the basis of the similarity between the target mobile user and the nearest neighbour set, the predicted scores given by the target mobile user can be obtained by the formula (5). Then, the predicted scores of the items are arranged in descending order, and the top N items are recommended to the target mobile user.

4 Fuzzy trust

We collaborating a trapezoid membership function (TMF) that make us to specify a variety for a provide trust phase instead of providing it a certain discrete range. A TMF depicted in Fig. 2 is mentioned as (x₁, x₂, x₃, x₄), here x₁ ⩽ x₂ ⩽ x₃ ⩽ x₄. Mentioned TMF represents as μ (a) is illustrated as given below in Equation 6;

Fig.2

Trapezoid membership function.

$μ (a) = {\begin{matrix} 1 b_{2} ⩽ a ⩽ b_{3} \\ 0 a = b_{1} or a = b_{4} \\ \frac{a - b_{1}}{b_{2} - b_{1}} b_{1} < a < b_{2} \\ \frac{a - b_{4}}{b_{3} - b_{4}} b_{3} < a < b_{4} \end{matrix}$ (6)

In case to utilize fuzzy inference to estimate the trust, we must define fuzzy inference trust using variables with a linking group of fuzzy trust phases that quantifies the behavior for node in mobile inference system. As mentioned let a_j (j = 1,2, ... n) be variable given as input. The rule based inference is depicted as below: $R_{1} if a_{1} is A_{i}^{1}, a_{2} is A_{i}^{2}, \dots . a_{2} is A_{i}^{2}, \dots . a_{n} is A_{i}^{n}$ (7)

Here $A_{i}^{n}$ is fuzzy set of ith rule mentioned in ith input variable. Let $a_{i}^{0}$ mentions the input value to reasoning model. Moreover, the degree of membership d of rule is estimated as: $d_{i} = \prod_{j = 1}^{n} μ A_{j}^{i} (a_{i}^{0})$ (8)

Fuzzy logic based decision forming utilized to find the various available network. Triggers and components are utilized to estimate the input which is given to fuzzy expert model.

Foe each elements a in the group of A, there exist the mapping a → μ (a), in which μ (a) ∈ [0, 1], μ (a) is mentioned as the membership function for each a. A function of membership mentions the degree to which a variable of fuzzy is a function of a group. Full membership is defined by 1 and does not have membership by 0.

5 Experiments and analysis

5.1 Experimental data

The experiment employed a dataset on mobile user behaviors provided by Massachusetts Institute of Technology (MIT). The dataset contains the data on mobile phone using behaviors of 106 lab workers in the MIT between June 2004 and September 2005. The data were originally collected by the MIT to study social network and social perception in the mobile communication environment.

The problem is the MIT dataset only provides the users’ use records of relevant mobile apps, failing to record the users’ evaluation of such apps. Hence, the use records should be collected and organized, and then mapped to the scoring table. The data in the MIT dataset were divided into training set and test set at the rate of 4:1.

5.2 Evaluation indicators

The mean absolute error (MAE) is the most common and simple performance evaluation standard in recommendation system [16 –18]. By this standard, the algorithm performance is measured by the difference between predicted score and real score. The MAE value is negatively correlated with the recommendation quality. The MAE formula is as follows:

$MAE = \frac{\sum_{i = 1}^{N} | P_{u, i} - R_{u, i} |}{N}$ (9)

Where P_u,i is predicted score of the item i given by the mobile user u; R_u,i is real score of the item i given by the mobile user u. N is the number of items.

In addition, this paper also selects the precision rate and recall rate as the evaluation index. The formula is as follows:

$P = \frac{| T_{1} \cap T_{2} |}{N}$ (10) $R = \frac{| T_{1} \cap T_{2} |}{T_{1}}$ (11)

Where T₁ represents the number of items in training set, T₂ represents the number of items which are recommended to mobile user, N represents the number of items.

5.3 Experimental analysis

5.3.1 Determination of the weight coefficient λ

This paper introduces the weight coefficient λ to adjust the similarities sim_I(u,v) and sim_P(u,v) in the proposed algorithm. The weight coefficient λ must fall in the range of [0, 1]. Otherwise, it will drag down the recommendation accuracy. The relationship between MAE and the value of λ is shown in Fig. 3.

Fig.3

The relationship between the MAE and λ.

Large λ means the recommendation system attaches more importance to the effect of user similarity sim_I(u,v) on the recommendation result, and small λ means the system emphasizes more on the effect of user similarity sim_P(u,v) on the recommendation result. As can be seen from Fig. 3, The MAE reaches the minimum value at the weight coefficient of 0.6. At this moment, the system has the best recommendation quality. Therefore, we set λ= 0.6 in subsequent experiments.

5.3.2 Comparison of different mobile user preference prediction methods

The proposed algorithm, the Pearson correlation similarity method and the Cosine similarity method were used to predict the mobile user preference on the MIT dataset. During the experiment, the λ was set to 0.6, and the number of neighbors was changed from 5 to 40 at the step of 5. The final experimental result on MAE is illustrated in Fig. 4.

Fig.4

The MAE of different preference prediction methods.

As shown in Fig. 4, the MAE of all three methods decreases with the increase of the number of neighbors. The proposed algorithm boasts smaller MAE than the other two methods. However, the MAE of these methods starts to rise again, when the number of neighbors increases to a certain number.

The experimental results of the three different experimental methods on the precision rate and recall rate are shown in Fig. 4 below.

From Fig. 5, we can see that the method proposed in this paper is higher and more accurate in precision rate than the other two methods, and the three methods have little change in recall rate.

Fig.5

The precision rate and recall rate of different preference prediction methods.

6 Conclusion

In order to solve the accuracy problem caused by data sparsity in traditional collaborative filtering algorithms, this paper proposes a weighted similarity method based on user interest and item attribute and applied the method in prediction of mobile user’s behavior preference. The performance of the proposed algorithm was discussed in details during the design of the experiment plan. Specifically, the author introduced the design idea and algorithm steps of mobile user preference analysis on the basis of collaborative filtering, and determined some parameters in the algorithm; then, the proposed algorithm was compared with traditional collaborative filtering algorithms. The algorithm analysis and experimental results show that the proposed algorithm outshines the other two algorithms in mobile user preference prediction. Whereas users’ choice of items may change with location, time, environment and other factors, the future research will take account of more context factors in the calculation of mobile user similarity, seeking to get a more accurate similarity measurement method.

Footnotes

Acknowledgments

This work was supported by Key Projects of basic demonstration teaching and research section of Anhui Province(No.2018jyssf113) and Big data and new technology teaching team project of Suzhou University (No. 2019XJSN06) and Professional leader project of Suzhou University (No. 2019XJZY23, 2019XJZY24).

References

Abdulsahib

G.M.

and Khalaf

O.I.

, Comparison and Evaluation of Cloud Processing Models in Cloud-Based Networks, International Journal of Simulation–Systems, Science & Technology 19(5) (2018).

Bobadilla

, Ortega

, Gutiárrez

and Alonso

, Classification-based Deep Neural Network Architecture for Collaborative Filtering Recommender Systems, International Journal of Interactive Multimedia and Artificial Intelligence 6(Special Issue on Soft Computing) (2020), 68–77.

Chatti

M.A.

, Dakova

and Thus

, Tag-based collaborative filtering recommendation in personal learning environments, IEEE Transactions on Learning Technologies 6(4) (2013), 337–349.

Devi

S.S.

, Singh

N.H.

and Laskar

R.H.

, Fuzzy C-Means Clustering with Histogram based Cluster Selection for Skin Lesion Segmentation using Non-Dermoscopic Images, International Journal of Interactive Multimedia and Artificial Intelligence 6(Special Issue on Soft Computing) (2020), 26–31.

Goldberg

, Nichols

, Oki

B.M.

and Terry

, Using collaborative filtering to weave an information tapestry, Communications of the ACM 35(12) (1992), 61–70.

Jia

C.X.

and Liu

R.R.

, Improve the algorithmic performance of collaborative filtering by using the interevent time distribution of human behaviors, Physica A (2015), 236–245.

Jiang

C.Q.

, Duan

and Jain

H.K.

, Hybrid collaborative filtering for high-involvement products: A solution to opinion sparsity and dynamics, Decision Support Systems (2015), 195–208.

Kenneth

K.F.

and Liu

X.Q.

, A collaborative filtering method for personalized preference-based service recommendation, IEEE International Conference on Web Services (2015), 400–408.

Kim

, Lee

and Chung

K.Y.

, Item recommendation based on context-ware model for personalized u-healthcare service, Multimedia Tools and Applications 71(7) (2014), 855–872.

10.

Lakiotaki

, Matsatsinis

N.F.

and Tsoukias

, Multicriteria user modeling in recommender systems, Intelligent Systems 26(2) (2011), 64–76.

11.

Park

, Park

and Jung

, Reversed CF: A fast collaborative filtering algorithm using a k-nearest neighbor graph, Expert Systems with Applications 42(8) (2015), 4022–4028.

12.

Patra

B.K.

, Launonen

and Ollikainen

, A new similarity measure using Bhattacharyya coefficient for collaborative filtering in sparse data, Knowledge-Based Systems (2015), 163–177.

13.

Ngan

R.T.

, Ali

, Fujita

and Giang

N.L.

, Gunasekaran Manogaran, MK Priyan, A New Representation of Intuitionistic Fuzzy Systems and Their Applications in Critical Decision Making, IEEE Intelligent Systems.

14.

Roopa

C.K.

and Harish

B.S.

, Automated ECG Analysis for Localizing Thrombus in Culprit Artery Using Rule Based Information Fuzzy Network, International Journal of Interactive Multimedia and Artificial Intelligence 6(Special Issue on Soft Computing) (2020), 16–25.

15.

Shi

, Larson

and Hanjalic

, Collaborative filtering beyond the user-item matrix: a survey of the state of the art and future challenges, ACM Comput Surv 47(1) (2014), 3–15.

16.

, Yeh

and Yu

, Music recommendation using content and context information mining, Intelligent Systems 25(1) (2010), 16–26.

17.

Jegadeesan

, Azees

, Kumar

P.M.

, Manogaran

, Chilamkurti

, Varatharajan

and Hsu

C.-H.

, An efficient anonymous mutual authentication technique for providing secure communication in mobile cloud computing for smart city applications, Sustainable Cities and Society 49 101522.

18.

Zhang

, Tao

and Teng

P.Q.

, An improved collaborative filtering algorithm based on user interst, Journal of Software 9(4) (2014), 999–1106.