Recommendation Engine: A Best Way for Providing Recommendation of Any Items on the Internet

Recommendation Engine: A Best Way for Providing Recommendation of Any Items on the Internet

Er. Meenakshi

Phd (Scholar), BMU Rohtak

Dr. Satpal

Ex. Professor, BMU Rohtak

Abstract:- Recommendation engine has been shown as a significant resolution for the issues of data or information overload, by giving progressively practical and customized info services to the clients. Recommendation engine provides recommendation about the items, services or information that the consumer needed to know. It is a smart app to help the client in the process of decision-making where they require deciding single product between the conceivably overpowering set of different services or items. It is likewise a standout amongst the various apps having a considerable effect on the execution of Ecommerce websites and the subdivisions commonly. Recommendation engine has been subjugated for suggesting financial services,, books, electronics, news, movies, travels, CDs, and several different services and items. Recommendation engine turns to be very prevalent even in simple Ecommerce websites additionally. Recommendation engine is utilized in numerous different Ecommerce websites to customize the data for their clients. In spite of the fact that the recommendation engine recommends products that are determined on the individual's taste, they can likewise be utilized an increasingly broad manner to make every site more client driven

Keywords: Recommendation System ,long tail ,website ,CB,CF etc.

I. INTRODUCTION: RECOMMENDER SYSTEMS BASED ON LONG TAIL

KEYWORDS:

The idea has discovered some basic for experiments, investigation and application. The term utilized in E- business, user-driven, microfinance, mass-media novelty and social-network instruments financial models, promoting a frequency-distribution with a long-tail has been considered by analysts since in any event year 1946.The term has likewise been utilized in the insurance &finance business for a long time.

In recommender system, long tail items are considered to be particularly valuable. Many clustering algorithms based on CF designed only to tackle the long tail items, while others trade off the overall recommendation accuracy and long tail performance.

Our approach is based on utilizing the item popularity information. We demonstrate that long tail recommendation can be inferred precisely by balanced item popularity of each cluster.

We propose a novel popularity-sensitive clustering method. In terms of long tail and overall accuracy, our method outperforms previous ones via experiments on Movie-Lens, citeUlike, and MobileApp

A recommendation framework alludes to framework that is fit for anticipating the future inclination of a lot of things for a client, and suggests the great products. One main motivation behind why we require a recommendation framework in present society is that individuals have a lot of alternatives to utilize from because of the commonness of Internet. Before, individuals used to shop in a physical store, in which the things accessible are constrained.

For example, the no. of films that can be set in a Blockbuster store relies upon the store size.

Frequently called as Recommendation Engines, they are basic algos which objective to give the most important exact things to the client by separating valuable stuff from of immense pool of data -base. Recommender finds data- patterns in the dataset by learning clients picks and gives the results that co-identifies with their interests needs.

1. Recommender Systems: What Long-Tail tells?

   Long-tail graph shows the distribution of ratings or popularity among items or products in marketplace. On the X-column items are ordered by their popularity or rating frequencies, whereas y-column shows the popularity in terms of ratings, demand etc. This graph basically points 3 important facts for recommender systems:

   Fig:2. Facts for recommender systems

   The recommendation engine [108] or interactive decision aids [46] can be considered as one form of personalization to facilitate in helping the users making purchase decisions. Main differences between the traditional brick-and-mortar stores and e-commerce websites are the infinite shelf-space on the Web. Unlike the traditional stores which have limited storage, the E-commerce websites provide the consumers a wide variety of options, alternatives and product information. The diversity of product choices and the abundance of messages on an e-commerce site have led to the problem of overloading. To overcome this problem demand of web personalization and real-time adaptation

   catering to the users need has arise. Reason being shopping experience can be overwhelming especially when there is no assistance available in deciding what products to purchase. In addition, the effort and time spent on searching aimlessly may lead to poor quality of decision and dissatisfaction of the consumers [17]. Therefore, to find the ideal products in mind effectively and efficiently, online customers not only look for the suggestions from their peers, and editorial picks [12] but also heavily count on the real-time recommendation systems featured on the e-commerce websites.

   Recent advancement in web technology helps internet organizations to acquire different client's data progressively. In view of obtained data, they construct nitty gritty profiles and offer customized administrations. In this manner online shops presently have the view to increment their execution by concentrating on different consumer inclinations and necessities, along these lines expanding fulfillment, improve faithfulness, and making one-to-one connections.

   A recommendation engine is application or system that encourages the client to choose an appropriate thing or finding pertinent data amongst a lot of candidates utilizing a knowledge-base that can any of be hand coded by specialists or gained from practices of the clients. Regularly, a recommender makes 3 of functions:

   • Information Collection: The system gathers all the usable data for the forecast undertaking including the clients' properties, practices, or the content of the assets the client gets to.

   • Learning: It applies a learning algo for filtering and exploiting the features of clients from the gathered data.

   • Prediction: It infers the sort of resources the consumer might incline toward area that point made either legitimately dependent on the dataset gathered in the phase of data collection (memory- based predictions) or with a model gained from it (model-based predictions).

     Recommendation engines are commonly utilized by electronic commerce websites for recommending different items to their customers and likewise to give them with data to enable them to choose which items to buy. The items recommended can be picked dependent on smash hits on a site, on the area of the shopper, or on premise of the past purchasing behavior of the purchaser as a forecast for future purchasing conduct. Different types of suggestions made up of recommending items to the customer, introducing custom-made item data and giving community assessments. Commonly, these recommendation strategies are piece of customization on a site since they enable the site to adapt itself to different consumer.

     Recommendation engines are equivalent to, yet not at all like from, DSS (decision support systems), supply chain systems and marketing frameworks. Though marketing frameworks helps the advertiser in settling on choices about how to elevate items to customers, for the most part by partitionig a wide target market into subsections of customers who have ordinary requirements and gathering the items in classifications that can be related with the

     marketing sections. Well along on advertising promotion would then be able to be raced to additionally energize buyers in different portions to buy items from groups chosen by the advertiser. On the other hand, recommendation engines openly communicate with customers, helping them to pick products they would like to buy. Recommendation engines commonly comprise of procedures that are carried physically, for example, making cross-sell lists and processes that are achieved to a great extent by PC, as CF. Recommendation engines increment electronic commerce sales in 3 different ways:

   • Convincing Browsers into Buyers: Website visitors frequently stay around the website without buying all. Recommendation engines can influence customers to purchase products they like to purchase.

   • Increasing Cross-sell: Recommendation engines improve sales by suggesting additional products for the consumer to purchase. By any chance that the recommendations are good, it will increment the average order size. For instance, a website may suggest further items dependent on items as of now in the shopping cart throughout procedure of checkout.

   • Building Loyalty: Globally, where a websites adversaries are just some clicks away, gaining customer loyalties is a basic business methodology. Recommendation engines create loyalty by making a value-added connection between the website and the buyer. Websites regularly make investment in learning about their shoppers; utilize recommendation engines to learn customer conduct, and to create incredible personalized interfaces that match customer requirements. Customers reimburse these websites by visiting back those that finest match their needs. The more a shopper required with the recommender engine teaching it what he requires the more faithful buyer turns to that website.

B) WORK FLOW OF RECOMMENDATION SYSTEM: Frequently called as Recommendation Engines, they are basic algos which objective to give the most important exact things to the client by separating valuable stuff from of immense pool of data -base. Recommender finds data- patterns in the dataset by learning clients picks and gives the results that co-identifies with their interests & needs.

Fig: 2. Work Flow Of Recommendation System

1. TYPES OF RECOMMENDATION SYSTEM

   Types of recommender-engines can be characterized dependent on methods utilized for recommendation

   1. CONTENT-BASED FILTERING (CBF)

      The CBF makes suggestion dependent on the relationship among different assets. In content-based recommender- engines, resources are defined as a vector of attributes. The framework at that point learns profile of the clients interests dependent on the highlights introduced in the articles that client has evaluated. When making a forecast on the clients' inclinations, the framework investigates the connection among the itemse valuated by the clients and different items by figuring the comparability among their attribute vectors. The kind of user profile inferred by a CBR is subject to the learning technique utilized. Neural nets, Decision trees and vector-based illustrations have all been utilized.

      A essential issue in CBR-frameworks is the require to recognize a adequately large set of significant attributes. At the point set is excessively small, there is lacking data to become familiar with the client profile. In this way, CBR- frameworks can't be utilized for new clients who bought just once, likely clients who visit the website however have not made any buy, and clients who need to purchase an item that isn't regularly bought.

      This is a traditional method which is utilized when the data overload issues need to be managed. This filtering method suggests products for the consumer dependent on the descriptions of formerly analyzed products for the consumer. They have been broadly utilized in creating recommendations of products information. A Recommendation engine utilizes this method and suggests products which are like the one favored by the consumer previously. Products are characterized by their related characteristics. Consumer inclinations shown up in view of those related characteristics in things previously rated by consumers. This recommender suggests products dependent on the products content as opposed to different consumers ratings.

      The CBF (Content-Based Filtering) makes recommendation dependent on the relationship among various resources. In CBR method, resources are portrayed as a vector of properties. The recommender then learns profiles of the consumers interests dependent on the characteristics obtainable in the objects that consumer has rated. When making a forecast on the users likings, the framework examines the association among the things rated by the consumers and additional items by computing the similitude among their property vectors. The kind of user- profile determined by a CB recommendation system relied upon the learning technique employed. Neural Networks (NN), Decision-Trees, and Vector-Based depictions have all been utilized. A fundamental issue in CBframework is the requirement to detect a adequately huge set of key elements. When the set is excessively minor, there is inadequate data to learn the users profile. Accordingly, CB recommender framework cant be utilized for novel

      consumers who bought just once, possible consumers who visit the website however havent made any buy, and consumers who need to purchase a item that isnt normally bought.

      The most important things for implementing this method are:

      1. Collecting the content information about the product.

      2. There should be an easiest method for individuals to characterize the ratings.

      3. To accumulate a users profile utilizing the content info taken out in a and rating data provided in b.

      4. An algorithm to match un-rated contents with the profile of user collected in the 3rd step and allocating rating to the products dependent on the match quality.

         As per their scores, the products are presented and ranked for the consumer in the arrangement. Nevertheless, CBF methods have the subsequent flaws in suggesting great products.

         1. A consumers choice is frequently relied upon individual characteristics whereas CBF methods are dependent on the objective characteristics (for instance, products description) about the products.

         2. For certain areas, either there is no content info is present, or the content is difficult to examine.

         3. Formulating quality and taste isnt so easy to do.

         4. Such recommenders can recommend only products whose content matches with the profile of consumer.

      If the consumer has taste that isnt signified in her/his profile, products talking to the un-represented taste will not be suggested.

   2. COLLABORATIVE FILTERING

      CF is broadly used and most recognized of the information filtering innovations. CBR systems gather products rating or recommendation; recognize likenesses between clients dependent on their ratings, and make novel recommendations dependent on inter-user assessments. A great user profile in a collaborative system is dependent on vector of products in addition to their ratings, continually altered as the user interrelates with the framework over time.

      CF-algos are categorized into two categories: ModelBased Algo and MemoryBased Algo. Memory-based algos work over the entire customer database to make forecasts. Prominent memory-based models are established with respect to the idea of closest neighbors, utilizing determination of distance measures. Model-based frameworks are established with respect to a conservative model reasoned from the information, which have utilized a variety of learning procedures comprising of NN, LSI and Bayesin networks. CF-strategies work well for composite objects for example music or films, where difference in

      taste is responsible for a significant part of the adjustment in preferring.

      The real contrast among content and collaborative based filtering framework is that the collaborative systems pursue previous activities of a group of customers to make a suggestion for specific individual of the gathering. Utilizing this technique, clients might currently be able to get suggestion for items that are distinctive in content to those they have formerly rated, given that other comparable mindset customers demonstrated their interests in these items.

      The CFperceives shoppers with comparable interests to those of a given buyer, and suggests enjoyed items by the given client. Notwithstanding, as most prevailing CF-algos acutely depend on the client's evaluations on items to make proposals, their execution perish seriously when the client rates some products in the database, which is known as cold start issue in the CF research.

      The CF is generally used and most well-known among the other data filtering techniques. Collaborative recommendation engine gather suggestion or rating of objects; recognize likenesses between clients dependent on their evaluations, and produce new suggestions dependent on internet user correlations. A classic consumer-profile in a CF is made up of vector of products in addition to their ratings, continually modified as the consumer cooperates with the system with time. Collaborative Filtering (CF) algos are categorized into two modules: MemoryBased and ModelBased. Memory-Based models work over the full consumer database for making predictions. Prevalent memory-based algos are initiated on the thought of nearby neighbors, utilizing choice of distance measures. Model- based algos are initiated on a compacted model gathered from the information, which have utilized a variety of learning procedures comprising of NN, bayesian networks and latent semantic indexing. CF (Collaborative Filtering) functions well for composite things for example music or movies, where variation in taste is accountable for a great part of the difference in liking. The foremost distinction among CF and CB frameworks is that the CF systems pursue previous activities of a user-group for making a recommendation for specific user of the group. By means of this technique, consumers possibly will be capable of getting suggestions for items that are diverse in content to those they have formerly rated, given that additional likewise mind set of users presented their inclination in these items. The CF recognizes buyers with likewise interests to those of a specified buyer, and suggests liked items by the given buyer. Though, as most prevailing CF algos distinctly depend on the consumers ratings on items for making suggestions, their exhibitions decease extremely when the buyer rates certain products in the database, which is known as cold start issue in the Content Filtering study.

      A Corresponding method generally utilized is CF (collaborative filtering). The general thought behind CF is individuals suggesting product to each other. This method fundamentally automates the procedures of word-of- mouth suggestions. In this method goods are suggested to a buyer relied upon values given by different buyers with

      likewise interests. In this method, buyer expresses her/his likings by rating products that the recommender shows them. After that these ratings against evaluations serves as an estimated demonstration of buyers taste in that system. The framework, after that, coordinates these rating against ratings given by other buyers of the system. The outcome is the set of buyers nearest neighbors, this sanctifies the idea of buyers with similar likings. Thus CF discovers method for matching buyers with similar likings and then creating recommendations on the basis of this. Three foremost problems in implementing this method are; (a) several buyers should take an interest (b) there should be an simple way for buyers to show their likings to the recommender and (c) to create an algo which will match buyers with similar likings.

      Nevertheless, CF techniques likewise have few disadvantages, for example; (a) extensive number of individuals should take an interest in order to enhance the possibility that any one individual will find different individuals with similar likings. It is hard to make a huge number of individuals to take an interest and henceforth it is costly. (b) It might provide poor suggestions, when individuals share smaller amount of info with the recommender. (c) When a novel products get into the database, there would not be any additional choice apart from suggesting it to a buyer until additional info is gotten through other buyer either evaluating it or indicating which different products is similar to.

   3. KNOWLEDGE-BASED RECOMMENDATION

      The knowledge dependent suggestions attempts to offer objects relied upon implications about clients liking and requirements. Such strategies are diverse in that they have practical knowledge: they have information about how a specific product meets a specific client requirement, and can along these lines question about the connection among a required and a conceivable suggestion. The client profile coukd be any information structure that keeps up this implication. In the typical case it might be a straightforward query that the client has written while in different, it might be an increasingly thorough portrayal of the client's requirements.

      This recommender technique endeavors to recommend items dependent on implications about consumers interests and requirements. Such techniques are dissimilar in that they have functional information: they have information about how a exact thing fulfills a certain buyer requirement, and can thus query about the association among a possible suggestion and requirement. The buyer profile could be any knowledge-based structure that keeps up this implication. In the ordinary case, it might be a straightforward question that the buyer has written while in others, it might be progressively thorough demonstration of the buyers requirement.

      This recommender generally depends on an obvious depiction of knowledge, as groups of statements, or different types of rule systems. Implementing this filtering method, the recommender suggests products dependent on implications from buyers likings and requirements. The

      buyer profile comprises of functional-information that is organized and taken as per the recommendation system. If an app needs inference or reasoning, deciding the knowledge-dependent method enables the developers to advantage from the software elements, knowledge illustration and rules planned for the recommender all in all. Knowledge based recommender dependent on a specific rule system. Information connects item and buyer is obviously expressed frequently as groups of statements. This method is extremely adaptable and has a great performance as well. Approach implemented in the rule- based recommender is that such systems function with already defined rules stated by the administrator or developer. Particularly this functions great for unsuitable items. Conversely, there is rule breaking down which considers inter dependencies of rules and makes novel rules dependent on present knowledge.

   4. UTILITY-BASED RECOMMENDATION

      The utility-based recommender doesnt endeavors to make long-term representation about their buyer, yet rather utilize their suggestions on a calculation of the match among the buyers requirements and the collection of choices on hand. These recommenders provide recommendations by taking out the usefulness of every item to the buyer. The benefit of this recommendation-engine is that it can evaluate non- product features, for example availability of product or seller consistency, into the usefulness calculation, therefore making it probable, to tradeoff cost against schedule of delivery for a buyer with an instantaneou requirement.

   5. DEMOGRAPHIC RECOMMENDATION

      The demographic recommender aims to classify the client based upon individual attributes and then give suggestions relied upon demographical classes. The users responses are compared with a collection of manually collected user stereotypes. The depiction of demographical data in a user model can contrast significantly. Demographic methods form people-to-people association such as collaborative ones, yet utilized istinct information. The major benefits of demographic technique are that it might not need a previous record ofclient ratings of the kind as required by content and collaborative dependent methods.

      These recommender systems intend to classify the consumers dependent on individual properties and then provide recommendations dependent on demographic groups. The buyers feedbacks are evaluated with a set of physically gathered buyers generalizations. The demonstration of demographic data in a user-model may differ significantly. Demographic methods form user-to- user relationship such as collaborative ones, yet utilize diverse information. The foremost benefit of this recommendation technique is that it might not need buyers previous ratings of the type as required by content-based and CF methods.

      However, demographic recommendation method is dependent on buyers individual characteristics. This method utilizes buyers descriptions (for example

      occupation, gender and age) to know about the association among a solitary product and the kinds of buyers who love it. The recommender method is applied with an already defined collection of generalizations, and provides suggestion dependent on which generalization well suited the buyer. Nevertheless, this technique has two primary limitations. To start with, it builds buyers profiles by categorizing the individuals utilizing descriptions of generalizations. Consequently, the same products are suggested to individuals with comparable demographic profiles. Though, in different cases, the generalizations are so common to make fine recommendations. Secondly, if the buyers interests changes with passage of time, then demographic filtering won't adjust their profiles. For such causes, demographic filtering is once in while used autonomously of the additional filtering methods.

   6. HYBRID SYSTEMS

These recommendation systems can combine any grouping of above mentioned metrics and techniques. These systems would evaluate ratings from a huge number of internal algos, beforehand grouping these in a solitary metric to enable consist ranking. Dependent on the keywords and on the individuals communications, this method joins the benefits of a detailed explanation of the association among the items. This permits relevant suggestions from the starting with a constant enhancement with time by grouping and utilizing additional individuals information. In a hybrid recommender, both subjective and objective features of a product are considered in anticipating its quality to suggest this product to the recommendation seeker. While hybrid recommender conquers the shortcomings of pure collaborative and content-based system relating to the subjective and objective qualities of suggestions, they do as such in an inflexible and prearranged way.

In particular, such recommenders endeavor to utilize one of the suggestion features (either objective or subjective) as accompaniment for the shortcomings of the other, at the time the last doesnt work viably. Though, there is no automatic method for finding in what conditions which sorts of properties (subjective, objective or both) are appropriate to a specific buyer in their present situation. Interestingly, by utilizing the market to reward successful recommendation systems (regardless of whether they use objective or subjective techniques, or a blend of the both); our recommenders progressively adjusts the comparative significance of the techniques as indicated by the feedback got from the individuals.

Fig: 3 General Frame work of Recommender System

1. INPUTS FOR RECOMMENDER SYSTEM

   With the aim of creating a working recommendation engine, a significant amount of information must be gathered. The most easiest and simple technique to gather a considerable amount of information is to get the client's undertaken viewpoints that ought to be their selves. Every recommender systems ask its customer to rate items which they bought or experienced. This is the easiest method of building database for a recommendation engine. Most basic rating systems include giving rating to a product of 5 or 10 (or even 100) or assigning a recommendations percentage to an product or for few characteristic of the product. Anyway it is additionally conceivable to utilize a binary rating system which essentially utilizes two ratings viz., 0 or 1 or dislike and like. The foremost disadvantage of a binary rating system is that it diminishes the quantity of information accessible to the framework. Demographics are additional decent type of input. Demographical data alludes to essential data taken out from the users profiles. Information for example, profession, gender, age and many more can be taken from profile of individuals. This info can be then utilized for categorizing the clients. The statistical techniques can likewise be utilized for making speculations. For instance a common statement Every College students like watching movies and Cricket is frequently untrue. Preferably a recommendation engine is to be completely automatic; with the purpose to categorize indiviuals without the requirement of direction. The metadata of the products might be drawnupon as a last wellspring of data. The metadata determines features of a product, comprising release date, genre, artist, and so on. In the rare case, the complete contents of a product (for example, the text written on a book) can be assumed for collecting the products. Such information might be utilized for making speculations dependent on classes from the product profile, for instance considering that if a user likes some product then the recommender would suggest that he would like every product in that one type which he loved.

2. OUTPUTS

   Recommendation engine ordinarily creates two type of output: recommendation and prediction. Prediction signifies a speculation: how a consumer would rate a product for which no assessment would be done. This needs a great numerical methodology and, as such, the techniques that apply finest to make predictions are the statistical strategies. Though, in most Ecommerce environments creating difficult numerical forecasts is unnecessary. Rather, what is need is a best-N list. The knowledge of a top-N list creates a list of a specific size that comprises the consumers likely most loved products and likewise it tends to be given to the consumer as a suggestion list. If it is proficient to make predictions then it is not difficult to create a top-N list selecting and sorting the N-highest prediction. As soon as the product similarity technique is viewed at as in Amazon, even this necessity can be loose by and by. The Item-Item similarity strategy doesnt endeavor to restore the optimum outcomes; however it is quick and great. The objective of a recommendation engine is to make suggestions about novel products or forecast the usage of a particular product for a certain consumer. Throughout past decade, various different recommendation methods have been established. These strategies are principally founded on 3 types of filtering methods that are Demographic, Collaborative and Content-based. These methods are as explained thoroughly.

   Utilizing the terms presented, these detested or novel things have its place to the LongTail of the product dispersion. Succeeding soul of broad investigation on the LongTail phenomena, these sorts of things ought not be disposed of or overlooked but rather profitably used in proposal strategies.

3. CONCLUSION:

In the era of Internet, mst serious issue for an individual who needs to purchase something on the web isn't just how to obtain sufficient data to make a choice yet additionally how to take a correct choice with that massive data or info. These days, individuals dependably look through the Internet to discover the best possible services and items that they require. Deliberately or unknowingly, they rely upon the recommendation engine to beat the problem of data overload. Recommendation engine has been shown as a significant resolution for the issues of data or information overload, by giving progressively practical and customized info services to the clients. Recommendation engine provides recommendation about the items, services or information that the consumer needed to know

REFERENCES:

1. Oestreicher-Singer, G., & Sundararajan, A. (2012). Recommendation networks and the long tail of electronic commerce. Mis quarterly, 65-83.

2. Szpektor, I., Gionis, A., & Maarek, Y. (2011, March). Improving recommendation for long-tail queries via templates. In Proceedings of the 20th international conference on World wide web (pp. 47-56). ACM.

3. Huang, H., Zhao, Y., Yan, J., & Yang, L. (2015). Improve the long tailrecommendation through popularity-sensitive clustering.

4. Shi, K., & Ali, K. (2012, August). Getjar mobile application recommendations with very sparse datasets. In Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 204-212). ACM.

5. Shi, L. (2013, October). Trading-off among accuracy, similarity, diversity, and long-tail: a graph-based recommendation approach. In Proceedings of the 7th ACM conference on Recommender systems (pp. 57-64). ACM.

6. Huang, Z., Cautis, B., Cheng, R., Zheng, Y., Mamoulis, N., & Yan, J. (2018). Entity-Based Query Recommendation for Long- Tail Queries. ACM Transactions on Knowledge Discovery from Data (TKDD), 12(6), 64.

7. Lee, K., Yeo, W. S., & Lee, K. (2010). Music Recommendation in the Personal Long Tail: Using a Social-based Analysis of a Users Long-Tailed Listening Behavior.

8. Yin, H., Cui, B., Li, J., Yao, J., & Chen, C. (2012). Challenging the long tail recommendation. Proceedings of the VLDB Endowment, 5(9), 896-907.

9. Zhou, W., Li, J., Zhang, M., Wang, Y., & Shah, F. (2018). Deep Learning Modeling for Top-N Recommendation With Interests Exploring. IEEE Access, 6, 51440-51455.

10. Hui, S., Pengyu, L., & Kai, Z. (2011, August). Improving item- based collaborative filtering recommendation system with tag. In Artificial Intelligence, Management Science and Electronic Commerce (AIMSEC), 2011 2nd International Conference on(pp. 2142-2145). IEEE.