Размещено на http://www.allbest.ru/

FEDERAL STATE EDUCATIONAL INSTITUTION

OF HIGHER EDUCATION

NATIONAL RESEARCH UNIVERSITY

HIGHER SCHOOL OF ECONOMICS

Saint Petersburg School of Economics and Management

Effects of product network relationships on demand in Russian ecommerce

Master's thesis

In the field 38.04.02 `Management'

Educational programme `MANAGEMENT AND ANALYTICS FOR BUSINESS'

product recommendation network сommerce

Kochnev Mikhail, Mendoza Ricardo

Supervisor

Docent, Candidate of Sciences (PhD)

E. Pokryshevskaya

Saint Petersburg 2020

Table of content

Introduction

Literature review

Recommendation systems. Concepts

Additional factors. Recommendation network, eWOM

Factors related to recommendation systems

Summary of researches

Research design and methodology

Hypothesis

Conceptual model

Choosing a category and a suitable layout of a product

Collecting the data

Scraping the data

Operationalizing variables

Data clustering: An additional process

Results and findings

An overview about OZON's performance

Descriptive statistics

Reviews

Discount

Ratings

Sales

Correlation between variables

A regression model for the TV sections of OZON

Driving factors of the amount of ratings

Driving factors of the amount of reviews

Driving factors of the amount of sales

Clustering analysis of OZON

Linear regression for clusters of OZON

Results of hypothesis

Conclusion and discussion

Conclusions 66

Managerial implications

Optimization of factors to increase sales

Limitations and Further Research

References

Appendix A

Appendix B



Abstract

The product recommendation network plays an important role in the e-commerce industry; they help the users to better understand what they might want to acquire. In addition, these referral systems allow an e-commerce site to analyze the customer behavior to optimize the store to procure sales. This research analyzes the relationship and influence of the values contained in the Product Recommendation Network, and how they impact on an e-commerce's demand. We carried out an empirical analysis of the TV category in a major e-commerce from Russia - Ozon.ru. We used the information from the recommended products of all products in such category; moreover, we added the information of the co-purchased products to our dataset. As a result, we found that part of what influences a sale of a product is due to the number of reviews, discount, and ratings of the neighborhood similar products and co-purchased products. A relevant finding was that the recommendation network has a different influence depending on the characteristics of the main product. To give a deeper exploration on this, we divided the category by clusters to provide an explanatory model for each type of product. The results suggest that by optimizing some variables from the Recommendation Network System the probability to sell a product would be higher.



Introduction

With the increasing development of modern technology, people began to save their time on daily tasks. Tools such voice assistants (Siri, Alice), smart homes, 3D modeling and online shopping apps are familiar technologies nowadays. Our work will be associated with one of the most growing areas in popularity - e-commerce. According to a study by Yandex.Market and the agency GfK Rus (Bakharev, 2019) the share of Russians ordering goods online has already exceeded 42% (more than 60 million people).

Customer-oriented e-commerce sites count with recommendation systems that can significantly reduce the client's search time and help to serve more accurately her needs. This topic is an increasing interest on the academic field. Currently, there exist several researches that addresses the importance of a Product Recommendation Network of online stores. The most notable, and closer to our work, researches are about the effects of Recommendation Networks on e-commerce (Pan et. al, 2019); the value that a network brings to the e-commerce (Oestreicher et. al, 2013); the effect on demand from the diversity of Product Recommendation (Lin et. al, 2017) and the like. These researches contribute significantly to understand how complex a recommendation system can be and the increasing usefulness they are for online customers. Following this line, we aim to provide additional insights to this specific topic and give information about the effects of product network relationships on demand in a Russian ecommerce. We want to discover at what stage the information contained in the recommended products affect the demand in an online store.

The research questions we want to pursue for this research are:

· What is the relationship between the Product Recommendation Network and the demand of a main product in the TV category of Ozon.ru?

· What are the most important factors of a Recommendation Network that might influence the demand of an-ecommerce?

In our work, we analyze the impact of one of the most popular Russian e-commerce sites (ozon.ru) to assess the relationship of recommendation systems on key business indicators (sales, reviews, product ratings). The aim of our work is to analyze and identify the relationship and impact of recommendation systems on the demand of e-commerce sites using the Ozon.ru marketplace as an example.

The main objectives to achieve the goal of our research are:

· To scrap the available information from the TV section of Ozon.ru to storage in datasets and analyze them.

· To provide an overview of one of the major e-commerce in Russia to better understand how they generate traffic.

· To carry out different explanatory models that show how the Recommendation Network system influences on Sales.

· To apply such explanatory models in different clusters of the same category to get a deeper analysis of this relationship.

· To provide practical actions that a retailer can perform to the online store to optimize the demand.

For this analysis we created several hypotheses that might help us to explain the overall of our work. Such hypotheses are based in previous researches. It is important to mention that our conceptual model looks to explain those additional factors that trigger demand. Based on that explanatory model, we will provide practical actions so a retailer can consider optimizing the additional variables that trigger sales. Nonetheless, this actions and findings are followed by several limitations that need to be addressed in further researches.



Literature review

Recommendation systems. Concepts

The recommender systems are a vast field with several algorithms and models. Its main objective is to provide personalized information to the user based on its interests and preferences. According to Linyuаn Luа et. al (2012), the most important concept to learn is to define the meaning of a network. The authors defined it as a set of elements with connections between them. To get an image from this concept we can see it as nodes connected with edges. Another important concept defined is the recommender system, which is a system that uses the input of data to obtain a prediction of potential interests of a user. The authors classified different types of recommendation systems as follows:

Content-based recommendations: Items related because of its similar content of the items selected previously by the user. Collaborative recommendations: Items based on past reviews of a large group of users. The recommendation can be based on a group with similar interest to the user (memory-based) or based on trained models to identify patterns (model-based). Hybrid recommendation: An approach that combines the content with collaborative methods. 

One method that has been presented as one of the most successful for recommendation is the Similarity-Based method. This approach has found popularity to be applied in e-commerce. The basic concept of this method is that people who take a decision in a past evaluation tend to take the same decision in future evaluations. For this method presented the following algorithms:

User similarity: Here the main objective is to create a prediction of user preferences by analyzing selected items from many other users with similar evaluations to the present user. In this algorithm only users with the most similarity is considered to evaluate a prediction. 

Item similarity: This approach has the characteristic that tends to be more static that the User similarity, this allows to compute the values offline as well.

Slope one predictor: This algorithm takes into consideration the users who rated the same product and other products rated by the same user. Linyuаn Lua et. al (2012) claimed that this approach can reduce the storage requirements and latency of the recommended systems. 

The work of Professor Gоmzin A.G. (2012) provides an overview of the basic algorithms used in recommendation systems. Collaborative filtering methods, methods analyzing the contents of objects, and methods using knowledge bases are considered. All methods considered have their weaknesses. The work considered in what ways it is possible to combine the methods for constructing recommendations to get rid of these shortcomings. Bearing in mind the results of the study and the considered algorithms, which will be considered in this paper, it is possible to filter out the most suitable objects quickly and efficiently for a particular user. The main task of the recommendation system is to have a list of objects that are most interesting to a specific user, having data about objects and users. Depending on what data is used to calculate recommendations, systems are divided into three large classes: 1) Collaborative filtering methods; 2) Methods that analyze the contents of objects; 3) Knowledge-based methods.

Collaborative filtering is a recommendation method in which only the reaction of users to objects is analyzed: the ratings that users put to objects. Ratings can be both explicit (the user explicitly indicates how many "stars" he estimates the object), and implicit (for example, the number of views of one video). The more grades collected, the more accurate the recommendations will be. It turns out that users help each other in filtering objects. Therefore, this method is also called collaborative filtering.

Methods of the second class, on the contrary, use the contents of objects to obtain recommendations. These methods work when the contents of objects are presented in the form of texts. They are well suited for recommending books. They can also be used to compare titles, descriptions, and other textual information available from films, songs, goods, etc. For example, a film was made - a continuation of a film. It is added to the system. The question arises, to whom to recommend it. Collaborative filtering will not give an answer due to the “cold start” problem. But it makes sense to recommend a new film to those who watched the first part of the film. 

Knowledge-based methods require the user to describe their requirements for the objects he needs. The idea of ??the method is as follows: users formulate their requirements for the product, the system tries to find the right product. In the first case, only those objects that exactly match all the user's requirements are searched and recommended. In the second case, objects with characteristics close to the requirements are searched (using some proximity measures).

To draw the conclusion, one can say that this work examined rather deeply all modern methods of recommendation systems from both an empirical and conceptual point of view. In any case, the more input data is available, the more accurate the recommendation system can be developed using various hybrid recommendation methods (Gomzin, 2012).

A similar study in this area is the work of Gal Oеstreicher-Singеr et. al., (2013), here they presented an approach to obtain the estimation of the true value of a product in the environment of the product networks. In their research they observed that there might be an overestimation of best sellers' products (in this case books) because they are recommended more frequently. They realized that such recommendations are received from books with higher sales, with a high conversion rate of their links. This suggests that the intrinsic value of such best sellers does not compensate for the value of the recommendation links from products with high volume of sales, leading it to an overestimation, while the long tail products might be underestimated.

Gal Oеstreicher-Singеr et. al., in 2013 also emphasized about the intrinsic value of a product vs the external value it receives such as the network value. In their research they concluded that the book's incoming value (or external) is lower than its intrinsic value. This means that the product itself generates more value than the external it receives from the recommendation networks. Nonetheless, this value may vary depending on the revenue tiers, for example, in lower seller products the incoming value is the lowest, while the best sellers get a high ratio, which means that the incoming value is higher for best sellers. 

In the same research Gаl et. al., in 2013 got results about the outgoing value of a product high sellers provide the highest outgoing as it was expected. However, the ratio between the outgoing value and revenue is higher for lower sellers. This means that less popular products generate more revenue than providing external value to other products. 

According to Huаng et. al., in 2019 the most popular and generally accepted Recommender systems techniques are collaborative filtering, data-mining techniques, content-based filters and techniques for context awareness. With these systems it is possible to reduce consumers search cost and increase the certainty of information for the user.

This study considers many concepts related to product networks. The authors labeled the concept of product networks as the collection of websites where a link represents a bound between two pages. In the same concept they mention that e-commerce products are not isolated, and many recommender systems connect the products by different algorithms. In other words, products become the nods and the hyperlinks that connect the products are the lines of the network. 

Huаng's study of product similarity network (2019) found that similar product reviews of connected products have a negative impact on customers' purchase behavior. This is explained by the fact that similar reviews present a good substitute for the same product. In addition, they discovered that products become substitutable when the product description might appear similar to the customer. Furthermore, in the same study, they empirically discover that recommendations can help to increase the sales of those recommended products by well connecting a page and then increase the visibility. 

An interesting implication of Huаng and colleagues is that the connection of the products must be balanced between powerful sellers and weak sellers, this will influence the performance of the total environment in e-commerce. A proper distribution might help to have a stronger effect on those similar products between high sellers and the lower ones.

It is important to consider the relationship between the demand of a product and how evenly distributed are the most demanded products in the product recommendations. Recent evidence Oestreicher-Singеr in 2012 suggests that there is an influence of a product position on ecommerce in its hyperlinked network of recommendations. They found that increasing the average network influence on a set of product categories is related to an increase of revenue for the less popular products. On the other hand, for the most popular products. The effect of these results is enhanced when there is higher assortative mixing and the clustering is lower.

The research of Oestreichеr-Singer et. al., (2012) is particularly interesting because they offer an approach of linking recommendation networks to the long tail products and the effect on relative demand and revenue based on its position. The authors point out that those products that are more highly and evenly influenced by visible networks tend to present a flatter demand than the rest. This means that connecting those visible products to the long tail might considerably increase the relative revenue for the less popular products. 

In the same research the authors found that the relationship of the influence of the recommendation network and flatter revenue distribution is improved when the influence is shared evenly across the products (in this case books) in the same category rather than concentrated on a few products. Moreover, they claim that, when recommendations start and finish in the same category the redistribution of interest is more even inside the same category rather than redirecting to a popular product (book) in a different category.

Digitalization allows us to receive a huge amount of data from consumers. So, professors Cagаn Urkup and Burcin Bozkаya in 2018 based on navigation data and mobile payments (transactions) analyzed data in the financial sector, namely a commercial bank, to assess the possibility of using one of the bank's products. The main research product was Individual Consumer Loan (ICL). During the study, several dozen data sets were generated, hypotheses were tested, and the best classification algorithm was determined. This work is a clear confirmation that the data on navigation and financial transactions are key for developing a recommendation network for consumer loans and for banks, respectively. Timeliness and relevance of the customer behavior data provided increases the accuracy of the product recommendation system. This study allows you to determine the potential calculation of the forecast use of any financial product with high accuracy.

Buying and choosing a product on an e-commerce website without analyzing reviews and ratings is like going to a store with closed eyes. Today there is a wide selection of recommendation systems that can reduce search time and increase brand loyalty. One of the most effective recommendation algorithms was developed by Professor Karthik, it is called Feature Based Product Ranking and Recommendation Algorithm (FBPRRA). This algorithm analyzes products on e-commerce websites and social networks and independently distributed recommended products. The algorithm was developed after working with the Amazon dataset and using the RapidMiner tool. The main difference from standard recommendation systems is that the algorithm processes and cleans data (removing stop words and tokenization), then segmenting users by age, location, gender, or other functional differences. All initial information about the product is stored in the database together with the functional features of a user.

After the creation of FBPRRA, a separate comparative analysis was carried out with other underlying technologies, such as Naive Bayes, KNN, Random Forest and Decision Tree. The accuracy and efficiency of the new FBPRRA algorithm surpassed its analogues by 22%. The new algorithm we have presented proves that digitalization and the e-commerce market are a place for new discoveries that can significantly facilitate and improve the lives of customers. Today the accuracy of this method is about 82%, which indicates possible improvements to the recommendation systems in the future, at least another 18% (Karthik et. al. 2018). 

With the development of e-commerce sites, the more relevant issue of the movement and presentation of goods on the site, the range of goods, the availability of a network of recommendations becomes more relevant. Professor Zhijiе and colleagues in 2017 examined the effectiveness of existing recommendation networks in the direction of diversity and stability of recommendation networks. First, this work serves as a good theoretical basis for any research on the topic of the recommendation network. The influence of customer demand on the development and diversity of the recommendation network stands out especially. One of the key issues that has been resolved through this study is the analysis of two types of recommendation networks. The first type is a joint viewing of goods, where a network of recommendations analyzes the behavior of a client on an e-commerce site for various goods. The second type is a joint purchase of goods, where a network of recommendations analyzes the purchased goods and forms “pairs” of goods that are most suitable for the future customer. 

As a result of the analysis, a comparative scheme of the data of the two networks was presented, and it was also revealed that the influence of the recommendation system on demand has the greatest influence when buying expensive goods, where a high degree of customer involvement is required.

An interesting approach is the work of Zhijiе Lin et. al, (2017). They studied the effects a Recommender system has on the demand in certain categories of e commerce. In their study they wanted to answer if the demand of a product is influenced by the characteristics of a product network and if this demand is also influenced by both incoming and outgoing recommendations. They found that the diversity in the recommended categories have a positive effect on the demand of a product. Also, they statistically proved that the recommendation related to co-purchase has a higher effect on demand than the recommendation related to co-view. This is an important contribution to the economic impact that a Recommender System can have on a purchaser when buying online, and the differences of the impact between the co purchases and co views of a product. 

The study of Zhijiе et. al., (2017) on the demand effect of recommender systems suggest that an increase in the diversity of incoming co purchase recommendation systems lead to an increase of the product demand. The most important finding of their study is about the effect of both incoming and outgoing networks, they implied that outgoing recommendation links in an ecommerce have more influence than the incoming recommendation links. A practical implication from this study is for ecommerce owners to realize the different effect on demand a certain recommendation system can make, retailers can set up a more diverse category when it comes to incoming recommendation links. Also, this allows the consumer to discover new products and avoid limitations. 

The book by Diеtmar Jannаch and colleagues in 2012 was originally written for students and researchers in the subject of the e-commerce market for a detailed immersion in the world of recommendation systems. This study reflects the relevance of creating retrospective (autonomous) recommendation systems, the creation of knowledge-based recommendation systems and the use of hybrid recommendations for e-commerce sites. The authors draw special attention to the problem of confidentiality and data security, due to the large amount of plagiarism in recommendation systems. For us, this work has become a real guide to the data analysis action for the analysis of the recommendation model.

Additional factors. Recommendation network, eWOM

Music is an integral part of the life of a modern person. We listen to it in our free time, at work, at the wheel of a car and before an important life event. That is why with the creation of modern gadgets and systems, the question arose of an effective and accurate musical recommendation system that will analyze our preferences and include the right song in the playlist. This question was studied by Professоr Jа-Hwung Suа and others, to analyze existing systems and offer the market one of the newest, most innovative and breakthrough music recommendation systems.

After analyzing about 15 new available music recommendation systems, the professor and colleagues took data from the latest Last.fm music site and developed a new music recommendation system, the Multi-modal Music Recommender System (MMR). This system allows you to consider client preferences at a completely different level, due to the integration of additional social and collaborative information within the system. MMR solves the problems of modern music recommendation systems such as the lack of music ratings and the difficult way to integrate tag information into the recommendation system.

The experimental results are confirmed by the best indicators compared to other systems: RMSE (standard error) and NDCG (normalized accumulated discounting). To further improve the recommendation systems for music, it is worth considering the influence of the human environment, he is in the office, car or at home near the stove - this will help to better understand the behavior of customers and their intentions before including the coveted music playlist (Suа el. al., 2017). 

Cristiano Ronaldo, Ariana Grande, Dwayne Johnson (The Rock), Selena Gomez, Kylie Jenner, Kim Kardashian West and Leo Messi - there are people whose Instagram accounts have the most subscribers in the world (in millions). Today, Instagram has more than 1 billion users and it is one of the most popular social networks in the world, because inside it you can edit and share photos and videos from your life to your community. For example, at the end of 2019, Instagram registered 500 million active users with “Stories” function, which allows you to share photos or video materials that are available for 1 day (Clеment, 2020).

The main issue of the study, conducted by Professor Nаrges Delаfrooz and others in 2019, assess the impact of Instagram users and the electronic word of mouth (eWOM) brand. Users daily show their attitude to the brand, give feedback, likes, comments and save products and even buy directly on Instagram, so creating an effective brand management action plan will significantly predict the behavior and reaction of customers. According to the study, the use of social networks is most often necessary for communication with peers, meeting the need for recognition in society, as well as self-realization. When users start following brands, brands could create a positive opinion and increase customer loyalty, as well as increase their eWOM. Today, with the right positioning and change of eWOM, marketers must build effective relationships with the end user. One of the key results is the expansion of the available literature in the field of assessing the impact of eWOM on the e-commerce market. In addition, it was found that the quality of interaction with brands affects the growth of eWOM, as well as an increase in activity in the profile and activity with other brands leads to a large influence on the relationship with eWOM. All three confirmed hypotheses are a useful complement to information about user behavior on Instagram and the impact on eWOM.

The results of the study are a useful instruction for marketers and SMM-managers to understand users' behavior, their activity in relation to brands and the quality of interaction with brands, which affects eWOM. The main recommendations for brand management are: search for effective bloggers with a high eWOM rating (people with a loyal audience) who can increase brand loyalty; take into account the algorithms of the social network and publish only high-quality content that will allow your brand to grow organically; always understand customer behavior, trends and foresights, which will increase your brand's eWOM and anticipate customer requirements (Delаfrooz el. al. 2019). 

Business success depends entirely on customer satisfaction. We must create all auspicious conditions, one of it - supporting service functionality (SSF). An empirical analysis of Professor Ronаld T. Sеnfetеlli in 2008 revealed that SSF improves service quality, customer satisfaction, and other business-relevant metrics. In addition to this pattern, DVDs, books, and other products were reviewed. SSF helps make strategic forecasts, develop a website in the B2C market. The developed model combines marketing and technology to understand customer behavior and the necessary components for their use. The development of SSF can be developed in e-commerce. 

A large study on electronic word of mouth (eWOM) was done by Professor Anа Bаbiс Rosаrio and colleagues who did a large literary analysis on eWOM consisting of more than 1000 scientific articles published in the 20th and 21st centuries. The key problems that were considered and solved by this work are the big problems with knowledge in the evolution of eWOM as an indicator. A pattern was identified in which IT and technology developments will have a significant impact on 3 eWOM factors: creation, change, and evaluation. So, Bose virtual reality sunglasses will change eWOM in audio form. Automation and increased manageability of eWOM contributed to the creation of artificial intelligence. In today's world, the ability to manage and use the electronic word of mouth indicator is a key advantage in the e-commerce market.

Product search can become an overwhelming task due to the increasing amount of information available online. The availability of product linked networks can ease the task of finding the right product (Gоldenberg et al, 2012).

In an analysis about dual-network structure, Gоldenbеrg found that user pages have unique structural properties that can have function as content brokers and might show relevant information to their users. This study was made using YouTube-based content. In the same study they found that dual network provides more interesting content during the process of searching for information as well as helps users to find information faster. In addition, in the same study made by authors (Goldеnberg et al, 2012), they concluded that users exposed to dual network are more engaged and tend to spend more time on the website before leaving. 

The authors also highlight that the quality of content referred to their study might be related to the likelihood of content that a person perceives as favorite. This means that when a user is exposed to dual networks, she is exploring content that has been previously filtered as a quality content and therefore an opportunity to find something interesting. So, they came to the conclusion that user-generated links are better at presenting a variety of content.

One of the promising areas of recommender systems is Linked open data (LOD) that help solve the problems of recommendations for users based on location or geo mark. To a large extent, the development of such systems today is more relevant in the field of tourism. This technology is used by such world giants as Booking.com, TripAdvisor, Maps.me, Yelp and others. The work of Professor Phаtricha Yоchum and others is aimed at the classification and literature review of existing studies over the past 17 years, the systematizations of research results and a grouping of related data sources for further practical application.

The results of the study showed that the expansion of existing recommendation systems with LOD improves the accuracy and quality of recommendation systems. Visual examples of the operation of this method confirm that the associated LOD is an effective tool for transferring information about the user's movements and location, which allows us to understand the behavior of the client, especially in the field of tourism (Yоchum el. at. 2019). 

A study by Professor Wаyne Xin Zhаo proved that not only in the development of recommended networks, but also in social networks. An example 3 years ago on the site of commercial advertising, an authorization mechanism began to appear through social networks, as well as the placement of recently purchased goods with links to e-commerce sites. Thus, a new recommendation network is created that stimulates the sale of goods. The main problem is that data on social networks must be approved for the cold launch and sale of goods. As a solution, a model was created where user profiles from social networks and websites using recurrent neural networks were used. Attributes of users that have been extracted from social networks. Successful experimental results, which were based on data from the Chinese social network “Sina weibo” and the Chinese e-commerce site “Jingdong”, showed the effectiveness of this structure (Xin el. at. 2016). 

Professor Gаl Oestrеicher-Singеr has many works on the topic of recommendation networks, one of which clearly demonstrates the impact of demand for sales of goods. In 2007, actress D. McCаrthy during a television show discussed the book “Louder than Words: A Mother's Journey in Healing Autism” about how she struggles with her son's autism in real life. After the airing of the book, the book broke all Amazon.com's book sales records (up 200 times). Research by Professor Gаl Oestreichеr-Singеr and colleagues in 2017 proved that the demand for the main book significantly changed the demand for additional products that were in the recommendation network (an increase of 25 times). Therefore, the demand for goods within an e-commerce site can be compared with contagious infections or groups that spread from person to person through the air.

Researchers have come up with significant results that should be reflected in our work. Demand for goods within the recommendation network not only affects 2-3 nearby products, it extends to the entire goods network of the e-commerce site.

It was also revealed a pattern in which goods that recommend each other fall into the so-called "echo chamber", where demand for goods and sales increase many times over. An e-commerce site benefits more from a recommendation than from an advertised product. This means that the effect of "D. McCаrthy" should be calculated for the network, and not for a specific product of the grocery network. This study provides an understanding for researchers and marketers that advertising campaigns should be improved with the effect of "D. McCarthy", as well as a thorough study of the recommendation network, will allow us to develop innovative and innovative marketing strategies for promoting products (Oеstreicher-Singеr el. at. 2017). 

The information load on users of e-commerce websites without the presence of recommendation networks would lead to rejection of purchases by customers. That is why the efficiency and accuracy of the recommendation network model is the main issue that scientists are working on today. One of them is Kyоung-jae Kimа, who suggested using the basic algorithm for building recommendation systems - collaborative filtering and increasing the efficiency of the model using additional analysis of social networks. After analyzing social networks, according to an empirical study, we received a data set with the most influential people on social networks (bloggers) and less influential. For the high-quality creation of a recommendation network, a cluster analysis is required, which will increase the level of accuracy of forecasting the requested information from the client. In practical applications, the difficulties of this model were found, when it is used by many users, so the question of the scalability of this model is the question of the next study of Professor Kyоung-jae Kimа. It is worth noting that this algorithm can significantly reduce data processing time, increase the accuracy and efficiency of recommendation networks when analyzing such large social networks as Facebook, Twitter, Vkontanke.

Early examples of research into Recommender Systems include the work of (Kyоung-jae Kimа et. al, 2017) where they analyzed the interactions between the collaborative filtering, clustering, and Social Network Analysis to produce accurate recommendation results. They got an empirical validation that states a SNA can enhance the prediction accuracy of collaborative filtering. This can be achieved by using the information from the most influential people in a certain Social Media as the main center in the clusters. Their model presented several recommendations using the cluster collaborative filtering, this experiment indicates that their model gives a more accurate prediction than the conventional collaborative filtering. Even though their sample is small, the experiment opens the doors to further analysis of how the SNA can lead to better recommendation results. 

There are several measurements and evaluations of recommendation systems. (Sоng Chеn, et al. 2013) suggested some few and common methods to evaluate a recommendation system, some of the most notable are: Mean Absolute Error, Classification Accuracy Metrics and Prediction-Rating correlation. In addition, (Sоng Chеn, et al. 2013) proposed some common limitations when it comes to recommendation systems and they labeled in several categories:

· The new user: A recommendation system can find difficulties in proposing items to new users. 

· The issue with Sparsity: There will be many items that will not be rated, and therefore will not have contact with users that might be interested in.

· Over specialization: This issue might appear when a system can suggest a product that the customer already got in touch with or with a high score. This might limit the recommendation to users.

Another area of knowledge that has not been fully studied is recommendation systems based on reviews from tourists to choose the most suitable hotel or hostel. To find a place where people spend their days off, you need to spend a lot of time evaluating the place, infrastructure, conditions, and feedback from previous customers, and do it manually. Professor Omаmah Alnоgaithan and colleagues in 2019 have created an effective recommendation model that analyzes visitor feedback, assessment and sentiment and can significantly meet customer expectations.

As a result of data analysis, separate Europe (more than 520 thousand hotels), a prototype recommendation system was created that allows you to receive output on the basis of incoming data - hotel rating by location, positive visualization (words that were most often found in the review), hotel rating and the review itself . The implementation of the recommendation system in the field of tourism expands the field of application of recommendation systems in the e-commerce market and allows us to hypothesize that these systems can adapt to many sectors of the life of a modern person.

The recommendation systems can be applied also in the field of tourism. (Alnоgaithan, and Algazlаn, 2019) listed a series of already existing systems for this industry: 

·  To avoid data duplication.

· Grouping pictures and keywords to provide the user a new suggestion.

· Personalized, that recommends based on interest of the users.

· System based on web recommendation. 

With the advent of the e-commerce market, the commercial interaction on the Internet has grown many times, which uses recommendation systems in its work. These systems form bundles of products with each other effectively using popular products as a “bait”. In a study by Gаl Oеstreicher-Singеr and colleagues in 2012, key factors and demand levels of recommendation networks are identified. For empirical analysis, a dataset of more than 250 thousand books, which are sold on Amazon.com, was taken. Each book on the site has its own position in the recommendation network, also on the page of the book are recommended books, that is, there are "virtual shelves" that link to each other and affect the final structure of the client.

As a result of data analysis, it was revealed that an effective built-up recommendation system with the possibility of joint acquisition leads to an increase in the influence of consumer demand by 3 times. Also, the most significant variable for e-commerce sites is “store design,” which allows you to get at least customer attention or purchase a product. The visibility of recommendation networks has a positive effect on the demand for additional goods, and the buyer, as a rule, makes a joint purchase. This study is different from many others due to the identification of a specific relationship, the influence of the visibility of the recommendation network on changing customer demand, therefore, for specialists in marketing, management and analysis of big data, this study will be an excellent instruction that structurally explains why the development strategy of recommendation systems is better than the development strategy viral marketing or social networks (Oestrеicher-Singеr et. al. 2012). 

A practical approach in the studied topic is the work of (Xuе Pan et. al, 2019). They studied the effect product distance can make on the eWOM (Word of Mouth) in recommendation networks. The main objective was to statistically prove if the reviews of a product are positively related to its nearest connected recommended products. And also, to prove if products with the nearest connectivity have similar levels of reviews. To empirically prove such effect, they divided the types of distance of the recommended products in two categories: The first one called “neighborhood to product level” which are those products connected by up to three clicks away from the main product. And the second was called “product to product level”. The results support that those products with near levels of proximity have an influence in the number of reviews of a product, the closer the proximity the more influence has on the recommendation from customers. Moreover, their study concludes that both direct and indirect connections between products can lead to a similar type of review. 

The impact of the research of (Xuе Pаn et. al, 2019) can have on product recommendation can be useful for ecommerce operators. For example, it can be possible to use the influence of positive eWOM to drive demand by connecting those positive reviews to the objective. In addition, retailers need to consider how to build and distribute the recommended systems to help the user to get relevant information.

Factors related to recommendation systems

Customer lifetime value (CLV) is one of the key business indicators which we can predict net income, as well as future relations between the company and the client. The accuracy and sophistication of the model depends on many factors that Professor Pavеl Jasеk considered in his work. Using data analysis of 35 thousand customers, he proposed measures to improve traditional CLV models. Improved CLV models will allow focusing on business indicators, which will more accurately determine customer behavior and effectively build a marketing campaign. A key factor in the new model is data clustering, which brings a significant amount of ideas for targeting the target audience.

According to Professor Randаll Lеwis (2015), online display advertising affects not only sales development in the e-commerce market, but also competitors' sales.

Surprisingly, the amount of additional search for information about competitors is sometimes 2-8 times more. Consumers manually compose their recommendation system based on display advertising, which certainly affects search queries, the eWOM rating and the e-commerce market as a whole. Additional searches determine the future positive or negative effects of brand display in the future. The hypotheses were confirmed after AB-testing for users of the website www.yahoo.com, on the main page of which three types of advertisements were placed to evaluate the further behavior of customers. It is noteworthy that after advertising “Progressive” car insurance, about 4,000 users started searching and became interested in the brand, on the other hand, clarification of the insurance conditions for competitors amounted to 23,000 requests. That is why, summing up the results, Professor Randall Lewis notes that at the moment, brand advertisers in the e-commerce market cannot fully get all the advantages of their advertising campaign, without analyzing the adverse effects, competitive effects of advertising and the operation of recommendation networks in general.

Professor Bо Xiaо conducted an in-depth study of referral agents that significantly better define customer behavior in the e-commerce market and provide the necessary guidance. During the study, the mechanism of the work of the recommendation agents was described, their frequency of use and characteristics. Empirical analysis has shown that there are explicit and implicit methods for identifying customer preferences. Moreover, explicit methods are highly labor intensive, which leads to high quality decision making and efficiency. Implicit methods are performed with less labor and average decision-making quality. It is proposed to create a model in which both methods will be integrated, which will allow to catch the balance between the complexity of the search for goods and the quality of the search. This work made it possible to understand the key knowledge gaps that exist today in the topic of recommendation agents, as well as to understand in which direction it is necessary to move in order to create a customer-oriented service in the e-commerce market.

An interesting study by Bо Xiaо et. al, (2007) proposed the use and impact of Recommendation Agents. The authors carried out the analysis due to increasing growth of ecommerce and the increased use of customers with the interaction of products. The defined a conceptual model that is looking to explain the effect that RAs have on customers.

Bо Xiaо proposed that RAs have influence on the customers' decisions. This means that RAs reduce the time of searching a product by reducing the number of alternatives that a customer might be interested in. In addition, these agents help to find alternative products related to the option that a customer is looking for. The authors also proposed that the use of Recommendation agents improves the quality of decision of customers, this can be measured as a calculated variant.

Another interesting approach by Bо Xiaо et. al, (2007) in terms of influence of RAs is that including product attributes might increase the customers preference. These attributes serve as an additional help to customers to select the most appropriate product for them. Moreover, the recommendation and reviews also influence the customer's choice, when a rating or review of a product is presented, the customer is likely to be influenced in her selections. It is important to mention that also the format in what the recommendation is presented affects the user's choice and efforts of selection.

Recommendation systems are becoming essential to consumers to discover new items (Kаrtik Hosanаgar et. al, 2013). An interesting study in this area is the work of these authors that suggest that users tend to have more in common after recommendations. And this can be explained by two reasons:

· Users aim their purchases to similar items due to the product mix.

· When users are clustered their distance shrinks after the recommendation.

The study made by Hоsanagar (2013) about the characteristics of consumer fragmentation can be useful when it comes to targeted marketing. Based on this study, this study of segmentation can be applied by e commerce depending on their categories to increase their sales.

One of the most interesting studies in customer behavior was conducted by Professor Еls Breugеlmans and colleagues (2011), they analyzed the effectiveness of advertising in two types of stores: online and offline. In addition, key factors for the success of advertising in online stores were highlighted. The results showed that today e-commerce sites show a positive trend, capping sales (up to 106%) using various tools. These tools have been studied in detail and include: increasing visual attention at the time of purchase, displaying highlighted products, systematically changing the layout of the advertising campaign, “smart” placement of goods on the site in a certain order, as well as the presence of a recommendation system that can reduce time and volume to search for the desired product. 

One of the main conclusions of this work is a review of existing mechanisms for increasing sales on an e-commerce site, as well as a selection of the right elements for implementing a successful product promotion strategy (Breugеlmans et. al. 2011).

Every year, universities create a large flow of scientific research, which is published in foreign and Russian publications. For a high-quality and effective writing of literary analysis in this work, we needed to spend hours working in the library to find the necessary and relevant publication that fits the topic of recommendation systems in the e-commerce market.

Professor Pеngfei Zhаo and others in 2018 took up a relevant topic for research - analysis of systems for recommending the academic environment through knowledge sharing. This analysis made it possible to create a recommendation system in which the relevance of the topic, the quality of the study, the number of downloads, the Hirsch index are combined, and the machine learning method is used to analyze the content of the work. One of the key limitations of the model is its development for only 1 type of scientific research - scientific articles, testing on other types of scientific research is the next step of the team of Professor Pеngfei Zhаo. The main result of the work was a recommendation model, which can greatly facilitate the search for academic information from the point of view of authors, students, and teachers. 

(Xiаojing Shеng, et. al. 2014) conducted a research in which they explained the usefulness of Recommendation Agents. Their research model started by measuring the intention of reusing the Recommendation Agent based on some variables like ease of use, risk involved, perception of usefulness and the like. This research was a survey of about 200 people asked how they perceived the usefulness of a Recommendation Agent. Their main objective was to have data about the consumer role using the RAs. 

(Xiаojing Shеng, et. al. 2014) demonstrated that, based on their survey, consumers who participate in using RA tend to perceive enjoyment of using it. However, the more participation in using RAs is perceived as a low ease of use of these systems. They led to the conclusion that consumers who participate using RA should not be overburdened with excessive interactions. An interesting managerial implication the authors proposed based on this research is that marketers and e commerce should employ more enjoyable experience of RA for those customers who have a low-income level and the risk of purchase is higher. By doing this the authors foresee a reuse of the RAs by the customers. 

...