A Novel Method Of Combining Trust

There is data boom going on currently in the world. We are living in the world of data, where data has changed the landscape of doing business. According to a research there are 2.5 quintillion bytes of data is being generated every day. In the last five years, more than 1 billion users have joined the internet. Mobile phones are reshaping the way we are connected to each other. More and more websites are being created every day. About 90 percent of the data has been generated in the last 2 years.

About half of the world population is online engaging over the social media channels and sharing pictures of their pets. All these users have made internet the biggest community on Earth. The importance of this data cannot be neglected.

Countries are now shifting towards smart cities with the help of valuable information this data provides. Every activity online is being recorded in one way or another. There is such vast amount of data available that future predictions are being made on its basis in many areas including business, science and technology.

Get quality help now

WriterBelle

Verified writer

Proficient in: Trust

4.7 (657)

“ Really polite, and a great writer! Task done as described and better, responded to all my questions promptly too! ”

+84 relevant experts are online

Hire writer

E-commerce businesses are booming and people are online through their mobile phones and PC searching and buying different products for their different needs. Behavior of these internet users are helping online businesses to conduct operations in a more streamline and organized fashion. Recommendation systems are being designed by business to engage more and more people and provide them with a more robust experience. The recommendations provided by these systems are helping people in making decisions and opening them a new variety of options they never thought of before.

Statement of Problem

More and more data is exploding on this internet every day. While this data provides a lot of useful information, it also confuses an average user. Users are given plantar of resources and options to choose from. Moreover, this information explosion have caused a challenge for users to choose from resulting in information overload. In order to overcome this problem faced by majority of users, recommendation systems were designed and built. These systems not only helped people in their journey online but also helped businesses to generate more revenue. Now a days users are getting personalized recommendations based on their past behavior. Collaborative filtering is used in order to provide these kind of recommendations. However these collaborative filtering techniques also suffer from different type of problems including data sparsity issues and cold start scenarios. Data sparsity problems occur when the system is unable to find many similar users or they are just not enough to provide recommendations .Cold start refers to a problems when a user have just started their online buying journey and have rated an very few items. Both of these problems makes it harder to recommend due to lack of past data and their behavior information. In the research we propose a novel and innovative method of making recommendations which overcomes these problems and provide high quality recommendations by combining trust with traditional collaborative filtering approaches. Trust networks help make find similar and reliable users to overcome fore stated challenges.

Research Significance

This research will help in improvement of recommendations being provided with the help of trust network construction and user similarity based on their past behavior. With the help of trust propagation a new range of users will be addressed and taken into the trust network.

This research will make important and helpful contributions in the online business community and techniques like combining trust with collaborative filtering will help business get more coverage and allow them to provide high quality recommendations to their users.

Methodology of Research

Many research approaches have been used in the past like case studies, surveys and experimentation in order to find desired results. My research will be supported by experimental results and on the basis of these results different evaluations for different scenarios will be made.

Organization of Thesis

This Thesis is organized in the following way

• Chapter 1: This chapter provides information for the research area and also some past information on the building of these recommender systems.
• Chapter 2: This chapter provides information of the significant and relevant work that has been done in this research area and different techniques used by researchers to overcome problems.
• Chapter 3: This chapter gives more detailed information of the proposed trust network construction and merging techniques to be used alongside collaborative filtering.
• Chapter 4: This chapter presents the implementation of the proposed methodology and gives experimental results to be evaluated.
• Chapter 5: This chapter concludes my thesis highlighting this research contribution and opening new research areas for future research.

Literature Review

In this chapter we have discussed in detail of the past research that have been done in the field of recommendation systems, CF and Trust based systems.

Recommendation Systems

Research on recommendation systems have been going on for a decade now and have gone through many phases. These RS allow users to get more suggestions based on the items and products that have liked or bought before. Online marketplaces like Amazon have stated that RS comprises of 30% of their total revenue. Moreover, social media channels like Facebook, Twitter and YouTube depends mostly on these systems to provide content to their users. Behavior and interests of users heavily impact the recommendations that they are getting and the predictions depends on them entirely. There are many types of approaches used for recommendations including content based, collaborative filtering based and hybrid approaches by combining these two methods.

Collaborative Filtering

Although CF techniques are used by many business online but e-commerce and online retailers have a very important core use of them. Many techniques have been introduced in the past to overcome data sparsity and cold start scenarios. There are two main kind of CF approaches: model based and memory based. Although work has been done on both of these approaches model-based techniques tend to give more robust predictions. To overcome foretold problems, made use of tags and friendships to establish similarity among users. Using social annotations is a plus but friends might not have same interests in products and items. However, people who have joined groups or communities online might have same interests that why they are easy to recommend.

Trust Network

Trust is an important factor in today's online communities. This factor if used correctly is a treat for businesses and online retailers. Many approaches have been proposed in order to increase accuracy of traditional RS. There are two main kind of trust factors that can be used in RS: explicit trust and implicit trust. The former is the degree of trust that a user have explicitly stated or given to another user while the latter is the degree of trust calculated indirectly using other means of trust information. Research on explicit trust have been done in the past, TrustWalker techniques is based on randomly selecting trusted neighbors and calculating their trust degree and combining it with item rating in order to give a recommendation.

Combining Trust and Collaborating Filtering

The Combining Process

In this chapter we will present our proposed method of combining trust based rating with collaborative filtering approaches. Firstly, the most trusted and nearest neighbors of the user are identified and then trust propagation is used to identify more trusted neighbors, which are useful for cold start scenarios. Secondly, a single value is formed by merging all of the trust rating from trusted neighbors. Every merged rating is calculated for a single item which must be rated by at least one trusted neighbor. In the end, similar users are identified based on the merged rating compiled before. With the help of these ratings recommendations are generated just like any other CF techniques. This whole process is further explained in the following sections.

Finding trusted neighbors

Those users who haven't rated any items or less than five items are known as cold users. Cold users are not that much active in the online system, that's why they might have low number of trusted neighbors. Social similarity is used to overcome this issue and provide with a better trust network. Trust propagation is used in order to find more trusted neighbors and overcome low number of trusted neighbors. There are two main kinds of trust propagation algorithms used including MoleTrust and TidalTrust. In our research we use MoleTrust to find indirectly connected neighbors. Trust value is set to binary, i.e. 0 or 1, 0 means no trust relation while 1 means trusted neighbor. Also in order to include range of the trust relation we also included distance.

T_(a,b)=1/d*T'_(a,b) (1)

Here d denotes the minimum distance between user a & b. If the distance between these users is increases then more number of trusted users can be inferred. As stated before, the trust values that user a will give to the user b will be in the binary domain. We will be using maximum distance of 3 in order to minimize extra searching steps and results shows that our method works best suing this limit. After the trusted neighbors has been identified then trust threshold will be established.

TN_a={b|T_(a,b)> _T,bU} (2)

Here trust threshold is denoted by _T.

Finding a merged rating

After finding all the trusted neighbors, it's time to find the items that has been rated by these users.

(I_a ) ={i|R_(b,i)R,bTN_a,i I} (3)

Here in this equation I_a contain at least one item that been rated by the users in the neighborhood. After that three separate ratings will be generated including rating similarity, trust and social similarity.

(R_(a,j ) )  = (TN_a Sim_(a,b).R_(b,j))/(TN_a |W_(a,b) |) (4)

(R_(a,j ) )  = (TN_a T_(a,b).R_(b,j))/(TN_a |W_(a,b) |) (5)

(R_(a,j ) )  = (TN_a J_(a,b).R_(b,j))/(TN_a |W_(a,b) |) (6)

These three separate merged ratings will then be combined to form a single rating.

W_(a,b)=.Sim_((a,b) )+.T_(a,b)+. (J_(a,b) (7)

Here Sim_((a,b) represents user similarity between user a & b. T_(a,b) represents trust between these users and (J_(a,b) represents social similarity between these users.

Now in order to find user similarity we used Pearson correlation coefficient.

Also to calculate social similarity between users we used Jaccard Index;

J(a,b) = |Ja*Jb| / |Ja*Jb| (9)

Finding the significance of merged ratings

We calculated these combined ratings with the help of Eq. (6) based on the similarity, trust and social similarity between users. Now in order to find the importance of this recently calculated rating we need to check it with our specially designed method of Significance. An item receive many ratings including negative and positive ratings, In order to find the number of these negative and positive ratings and balance them we will compute significance using Eq. (10)

Combining merged rating with collaborative filtering

After finding importance of the ratings, with will be incorporating that rating to find similarity. In order to do that, we will modify the existing PCC formula and integrate significance with it.

After finding significance based similarity, a set of similar uses will be selected and all the ratings of these similar users will be accumulated to give a prediction.

Evaluation

Experiments have been conducted in order to find the effectiveness of our method. Our main objective is to find performance of system compared to other proposed systems.

Data Collection

We have used two real world datasets namely Epinions and FilmTrust. These datasets contains user-item ratings also explicit trust data. On FilmTrust, people visit to review movies. The dataset ratings include 1508 users, 2071 movies and 35,497 ratings.

While Epinion is an online community where users can rate different products and provide trust information to other users. This dataset contains 6, 64,823 ratings, 1, 39,738 items and 40,163 users.

Experimentation and Results

We will be evaluating our method in terms of both accuracy and coverage. After prediction of rating from our system, error between actual rating and our predicted rating will be calculated. Here are some of the evaluation metrics

MAE = (12)

Here TR represents total number of test ratings. The smaller value of MAE indicates that our prediction is relatively good. Also RC is used for Ratings Coverage.

It accumulates of two parts, Number of ratings predicted and all the test ratings. In order to measure overall performance of the system we use f1 measure.

This equations actually represents the robustness of our systems and balance between coverage and accuracy.

The performance that we have measured can be seen in the table below

Table 1 : Experimental Results

Conclusion

This research have proposed a novel method of combining trust, social similarity with Collaborative filtering in order to overcome issues like data sparsity and cold start scenarios from which the traditional approaches suffered from. Trust network is constructed using MoleTrust, then Trust, Similarity and Social Similarity are combined into three separate merged ratings. Then an average is taken of all these ratings and then one single merged rating is formulated. Importance of that merged rating is measured by over novel method of Significance. Our experimentation showed that our method not only outperforms all other traditional approaches but also give high quality recommendations.

The present work opens new doorways for research purposes in these hybrid based approaches. We used explicit and implicit trust, but some people have reacted negatively to certain products and don't want to receive recommendation about these products. Negative trust also have an impact on the recommendations so for future work we intent to incorporate negative trust also in our process.