Enhancement of Resource Recommendation System with Graph Based Using som sandeep Pradhan Mr. Sandeep Gonnade



Download 59.37 Kb.
Date conversion13.06.2017
Size59.37 Kb.
Enhancement of Resource Recommendation System with Graph Based Using SOM

Sandeep Pradhan Mr. Sandeep Gonnade

Computer Sc. & Eng Asst. Professor, Computer Sc. & Engg.

MATS University MATS University

Raipur, India Raipur, India

Sandeeppradhan22@gmail.com sandeepg@matsuniversity.ac.in

Abstract: - This is a paper about recommender system, and the specific techniques that recommendation system uses to remedy the Information overload problem. Its center of attention lies on two most common approaches they are collaborative filtering and content based filtering. Both of these methods have their weaknesses and strengths. To beat the weaknesses of each and every process, quite a lot of hybrid filters had been developed. It will concentrate on Kohonen’s self-organizing map algorithm, used to identify the neighborhood of the current user so that relevant resources are suggested. The implementation of this algorithm shall be used alongside collaborative filtering process in the effort of designing a recommender system for books. A graph based recommender system makes useful recommendations by exploiting the significant content available. For calculating the ratings of the resources a weight based approach is used. This system is known as a web based client side application which makes recommendations by constructing user-resource graph and ranking the resources by a new method designed similar to that of search algorithms.

Keywords: - Recommendation system, graph based recommender system, Neighborhood, Weight based and Self-Organizing map.


  1. INTRODUCTION

Recommender systems are a subclass of information filtering system that obtain to predict the 'rating' or 'preference' that a user would give to an item [3]. Recommender systems became very common in recent years, and applied throughout a method of applications. The foremost fashionable ones are in all probability movies, music, news, books, analysis articles, search queries, social tags, and product usually.

The most common techniques used in today’s recommendation systems fall into two distinct categories: content-based methods and collaborative filtering methods [4,5]. And collaborative filtering has been known to be the most successful recommendation techniques. Collaborative methods recommend items based on aggregated user ratings of those items and these techniques don’t rely on the availability of textual descriptions. They share the common goal of assisting in the user’s search for items of interest, and thus attempt to address one of the key research problems in the area of the information age: locating needles in a haystack that is growing exponentially. Collaborative filtering systems can deal with massive number of individual and with many alternative items. However there’s a problem that the set of ratings is sparse, such that any two users will most likely have only a few co-rated items. The high dimensional sparsity of the user-item rating matrix and the problem of scalability result in low quality recommendations.


Personalized recommendation methods operate upon user ratings on observed items or item features making predictions concerning users’ interest on unobserved items [6,7]. In the most cases significantly  in real-world applications, the number of ratings obtained from users is sometimes terribly small compared to the number of ratings that must be predicted. And this problem is termed the Sparsity which significantly affects recommendation methods reducing the accuracy of prediction. The sparsity of ratings problem is specifically important in domains with a large number of items as well as a large number of users. Different solutions are required and different prediction techniques must be employed to resolve the problems.
Artificial Neural Networks are classification algorithms impressed by the functioning of the neurons in the human brain. It consists of layers of the neurons, with each layer depending on the next layer. Their versatility has been of great advantage in solving in many problems in the learning sphere and outside. There exist many structures, of which Feedforward is the most popular [8]. Method of trial and error is adapted to randomly initialized weights in the network to converge to an optimum value

The structured of the paper is as follows: Section 2 focuses on the various issues faced by recommender systems. Section 3 describes the proposed work. Sections 4 describe our system model and its various modules. Section 5 depicts the results that we got after experimentation with data and also the efficiency measure based upon performance metrics. Our work is concluded in section 6.




  1. ISSUES FACED BY RECOMMENDER SYSTEMS

Content based recommender systems face the disadvantage of training from a restricted data regarding items [9]. The illustration solely captures options concerning the information regarding the item, not the users’ expertise [10]. Therefore, if solely this content is alone chosen to act as input for the system, we'd be restricted with a recommendation of solely similar items, irrespective of their popularity, which can vary beneath numerous factors.

Collaborative filtering recommender systems to have their own share of issues, related to data sparsity [11] are the most prominent one. It’s highly likely that a user would rate solely a few items in an item domain and thus the extremely sparse user-item ratings matrix would cause to a downgraded performance.




Rating Calculator

Resource identifier
PROPOSED WORK


Resource graph

Self-Organizing map



Resource ranking

User Profile Builder



Query



Resources

Database (book-rating)




Fig.1 Architecture of the proposed recommendation system

We proposed a new hybrid system which combines the techniques used in graph based resource recommendation system. The proposed system use the ratings of resources which is provided by graph based resource recommendation system [1] and by applying self-organizing map algorithm we can easily make clusters of similar resources by reducing the redundancy and data sparsity.



Proposed SOM Algorithm to Generate Cluster

  1. Assign random values to weight vector of neurons and determine the topology of the SOM.

  2. Initialize the weight vector Vi(0) for i=1,2,…….k randomly.

  3. Provide an input vector to the network.

  4. Traverse each node in the map

  • Use the Euclidean distance formula to find the similarity between the input vector and the map's node's weight vector

  • Track the node that produces the smallest distance (this node is the best matching unit, BMU)

  1. Update the nodes in the neighborhood of the BMU (including the BMU itself) by pulling them closer to the input vector

Vi(t+1) = Vi(t)+t*α(t){D(t) - Vi (t)}

Where


t denotes current iteration.

Vi denotes the current weight vector.

D denotes the target input.

α(t) denotes learning rate due to time.

6. Increment t and repeat from step 2 while t >

λ. Where λ is the limit on time iteration

Self organizing feature maps (SOFM) also known as Kohonen feature maps belong to the class of competitive learning based cluster. The nodes (neurons) of these become specifically tuned to varied input patterns. It consists of two layers of neurons, an input layer and a so-called competition layer. When an input pattern is presented to the network, that vegetative cell within the competition layer is determined, the reference vector of that is nearest to the input pattern. This neuron is termed the winner neuron. The learning is such just one node (neuron) at the output becomes active corresponding to a specific input pattern. the learning in SOFM, however, differs from the other competitive based.

nc (t1)d:\project\thesis work\project thesis\materials\somnet.jpg

nc (t2)

nc (t3)

nc (t4)

Fig.2 Topological Neighborhood of SOM net

clustering methods in that not only the weight of the winning neuron is going to change but also those in its neighborhood Nc (t) (with t1 < t2< t3 < t4) which is defined in terms of some proximity relation. This neighborhood relation is usually represented as a (usually two-dimensional) grid as shown in Fig:2, the vertices of which are the neurons. The grid is most often rectangular or hexagonal.



  1. IMPLEMENTATION

This research work is implemented with the help of MATLAB R2015a. In this research work four different data clustering techniques are compared, they are K-means clustering algorithm, K-medoids clustering algorithm, FCM and SOM algorithm. The system designed is used for recommending books-novels to the active user. The dataset contains about 150 user profiles, 200 books, more than 4000 ratings. The profile of the user contains three fields namely id, age, preference. Book details namely Book id, name, genre are present in the dataset where the age group is categorized in between 12 to 60. Science fiction, fairy tales, romance, Thriller and humour are the various genres of books considered.

The steps for recommending books to the user are explained below.



  • Step1: From the user input, the profile based on his preferences and age is built.

  • Step2: Generate the cluster by using SOM algorithm.

  • Step3: Find the cluster in which the user is present by calculating the Euclidean Distance with each centroid and assign the centroid with the minimum distance to be the neighborhood.

  • Step4: Identify the books which are more popular in the neighborhood. Calculate the ratings for the resources.

  • Step5: Construct the user–book graph and recommend the books based on the ranking method.



  1. RESULTS AND DISCUSSION

The system was implemented by forming the neighborhood in based on user’s age and using K-means, K-medoids, FCM, and SOM clustering technique and calculate the number of relevant books recommended to relevant age group.

Table 1: Number of relevant book recommendations

Book Profile

Clustering Methods

K-means

K-medoids

FCM

SOM

Fairy

12

15

17

23

Humour

16

14

18

25

Murder

8

7

6

14

Romance

14

17

12

19

Science Fiction

9

6

11

17

Series

13

16

22

27

Short Stories

8

12

11

14

Thriller

19

15

21

28

Graphic Novels

14

10

15

18

The above table contains the total number of relevant book count recommended by all four algorithm K-means, K-medoids, FCM and SOM algorithm for a relevant user’s age group. The calculation is done for 50 iteration for a particular age group. For example a user with 14 years old more likely to read fairy genre books and likely to rate high to such books. The dataset contains such records with rating given by all age groups user. If we run the proposed recommendation system then it recommends the fairy genre books which have highly rated by past users.

K-means, K-medoids, FCM and SOM clustering algorithm are compared by plotting a graph which is shown in Fig.3. The graph shows that by using SOM algorithm the number of relevant books recommended are more than other clustering technique. From the graph it is evident that using the SOM clustering algorithm does actually increase the correctness in recommendations. Xaxis consists of different genres of books while Yaxis consists of number of resources.



d:\project\thesis work\project thesis\output screens\comparison graph1.jpgFig.3 Comparison Graph

  1. CONCLUSION

Recommender systems provide an insight to the users regarding their interests and facilitate better searching capabilities. The technique proposed uses SOM clustering technique to find the user’s neighborhood,

Which provide more accurate prediction to the users.



Algorithmically we would like to investigate different SOM techniques like hierarchical SOMs and growing SOMs to see if they can contribute to making the interface even more interesting without losing any of its current capabilities. We are also interested in researching the possibilities of using a spherical SOM to increase the intuitively of the maps.

REFERENCES

  1. Dr. P. Pabitha, G. Amirthavalli, C. Vasuki, J. Mridhula “Graph based resource recommender system”. IEEE International conference on Recent Trends in Information Technology.2014.

  2. Dr. P. Pabitha, G. Amirthavalli, M. Rajaram “Weight based approach for resource recommender system” Medwell journals, International journal of soft computing. 2015.

  3. Francesco Ricci and Lior Rokach and Bracha Shapira, “Introduction to Recommender Systems Handbook, Springer, 2011, pp. 1-35

  4. M.G. Vozalis, K.G. Margaritis, Using SVD and demographic data for the enhancement of generalized Collaborative Filtering, Information Sciences 177 (2007) 3017–3037.

  5. Yu Li, Liu Lu, Li Xuefeng, “A hybrid collaborative filtering method for multiple-interests and multiple–content recommendation in E-Commerce” , Expert Systems with Applications 28 (2005) 67–77.

  6. George Lekakos, George M. Giaglis, A hybrid approach for improving predictive accuracy of collaborative filtering algorithms, User Model User-Adap Inter (2007) 17:5–40.

  7. Breese J, Hecherman D, Kadie C. Empirical analysis of predictive algorithms for collaborative filtering. In: Proceedings of the 14th Conference on Uncertainty in Artificial Intelligence (UAI’98). 1998. 43~52.

  8. Razavi, S., & Tolson, B. A. “A new formulation for feedforward neural networks,” Neural Networks, IEEE Transactions on, 22(10), 1588-1598, 2011.

  9. Gupta, Anant, and Kuldeep Singh. “Location based personalized restaurant recommendation system for mobile environments,” Advances in Computing, Communications and Informatics (ICACCI), 2013 International Conference on. IEEE.



  1. P. Lops, M. de Gemmis, G. Semeraro, F. Narducci, C. Musto. “Leveraging the linkedin social network data for extracting content-based user profiles.” In Proceedings of the 2011 ACM Conference on Recommender Systems, RecSys, Chicago, IL, USA, 2011

  2. Su, X., & Khoshgoftaar, T. M. “A survey of collaborative filtering techniques,” Advances in artificial intelligence, 2009.

  3. M. Balabanovic, Y. Shoham “Fab: content –based collaborative recommendation” Comm. ACM 1997.

  4. B. Fritzke “A growing neural gas network learns topologies” Advances in Neural information processing systems 1995.

  5. T. Martinetz, K. Schulten “A neural gas learns topologies”. Artificial neural networks, Elsevier science Publishers 1991.

  6. B. M. Sarwar, G. Karypis, J. A. Konstan, and J. Riedl, “Recommender systems for large-scale E-commerce: scalable neighborhood formation using clustering,” In Proceedings of the 5th International Conference on Computer and Information Technology 2002.

  7. M. Lee, P. Choi, Y. Woo, “A Hybrid recommender system combining collaborative filtering with neural network”. Springer 2002.

  8. A. Gupta, Dr. B. K. Tripathy, “A Generic Hybrid Recommender system based on neural networks”. IEEE 2014.




The database is protected by copyright ©ininet.org 2016
send message

    Main page