Tag Archives: Recommendation system

Know Yourself And Know Your Enemy

‘If you know yourself and know your enemy, you need not fear the result of a hundred battles’ (Sun Tzu, Art of War)

Online retailers currently implement and leverage a variety of sales support tools. This article provides insight in consumer (users) behaviour on those e-commerce websites (Adomavicius and Tuzhilin 2005). It looks at different factors that might affect consumer behaviour (specifically consumers’ decisions whether or not to buy a product) (Hennig-Thurau et al. 2012). In this research the effect of two things on sales of a product is examined. One is recommendation systems (generated by the firm) and the other is online review systems (generated by customers). Previous literature has primarily focused on both loose effects; this study particularly looks at how these two factors might combine together to affect consumers’ behaviour.

Recommendation systems are where each product will give you recommendations to other (similar) products (Oestreicher-Singer and Sundararajan 2012). This study argues that we should see recommendations systems as networks with a number of different products linked to each other. This study seeks to analyse the whole network of related product referrals. The author predicts that the position of all the different products within a network might affect the sales of a product. Any product near to the centre of the recommendation network will get more attention. So, if the Product A is near to the centre of the network it will get more sales. If the competing product (or normally products) is near to the centre these will get more attention, taking attention away from the Product A, and so Product A will get less sales.
Online review systems (eWOM: electronic word-of-mouth) are where users of this site write a review of a product. This study observes when the consumers is shopping on the site, they can see both reviews of the product they are considering and also all reviews of other recommended products. The author predicts that the fact that they can see these other reviews creates competition, which makes the customer less likely to buy the particularly product.

The empirical analysis are performed with collected data from Amazon.com on 1.740 randomly selected books within four categories (programming, business, health and guide books) over a period from two years. This empirical analysis yields three major findings. The research has shown that recommendation systems intensify the competition between products. The authors state that the products which are linked to recommendation systems generate more sales if they have a central place in the referral recommendation network. There have been extensive findings that these sales gains are impeded by improvements in the reviews (eWOM) of competing products. This indicates that a positive eWOM received by a competing book worsen the rank of the focal book.

The main limitation of this study is that they used data at aggregate level, and not at individual consumer level. Therefore, the collected data does not track the actual activity of reviews/recommendations. This is a missing link for virtually all eWOM studies. The solution to this limitation could may be collecting click-stream data to better connect behavior and action.



Adomavicius, G., and Tuzhilin, A. 2005. “Towards the Next Generation of Recommender Systems: A Survey of the State-of- the-Art and Possible Extensions,” IEEE Transactions on Knowl- edge and Data Engineering (17:6), pp. 734-749.

Hennig-Thurau, T., Marchand, A., and Marx, P. 2012. “Can Auto- mated Group Recommender Systems Help Consumers Make Better Choices?,” Journal of Marketing (76:5), pp. 89-109.

Oestreicher-Singer, G., and Sundararajan, A. 2012. “Recommen- dation Networks and the Long Tail of Electronic Commerce,” MIS Quarterly (36:1), pp. 65-83



Will the global village fracture into tribes? Recommender systems and their effects on consumer fragmentation

Netflix, Spotify and Amazon and many other companies recommend personalized content to best suit the customers’ preferences and increase consumption or sales (De et al., 2007). Recommender systems are used for different types of media such as movies, books, music, news and television. Moreover, they strongly affect what customers view and buying behavior. Many positive effects of personalization are known such as reduction of information overload, increasing relevance and loyalty (Tsekouras, lecture 2017).

However, personalization may also have drawbacks. As the internet becomes more and more specific to our interests, this “ hyperspecification” may fragment users, reduce shared experience and narrow media consumption. For example, fully personalizing news sites may mean that we no longer see the same news articles. Therefore one could argue that recommender systems will create fragmentation and will cause users to have less in common.
On the other hand, recommenders may have the opposite effect as they share information among customers who otherwise would not have communicated. The paper presents empirical evidence on whether recommender systems fragment or homogenize customers.

In an observational two-group experiment the behavior of the participants before and after recommendations is compared. Recommendations are provided in iTunes and the user experience is personalized through mostly content-based recommendations and a small part collaborative data.

Interestingly, contrary to the author’s expectations, recommendation systems increase purchase similarity (homogeneity among customers) and therefore do not increase fragmentation. Customers tend to purchase more after being exposed to personalized recommendations due to the volume effect, which in turn increases the chance of customers purchasing the same product.
Furthermore, customers buy a more similar mix of products after recommendations due to the product-mix effect. Consequently, customer networks become denser and smaller. However, it is important to note that the recommender system does not recommend the same items to many users but that the diversity of items consumed increases.

Within business, the findings of the paper confirm that recommendation tools can help marketers to increase sales but also alter the product mix customers buy. Off course, these are some of the the reason why numerous retailers (e.g. Netflix, Amazon, Spotify) use these recommendation systems.
However, what I find more interesting about the paper is the application of recommendation systems in the context of society. As we increase to spend time online and recommendation systems become more advanced, I think this is an interesting application. Similarly, I think it would be interesting to research to what extent the proposed homogenization of customers has implications for society and (business) cultures.
Fragmentation may have negative implications for society as when people share little information filter bubbles and echo chambers may exist. However, in business context, I think there are situations in which fragmentation of customers may provide better results than homogeinity . For example in the context of innovation, crowdsourcing and generating new business ideas, a variety of views and opinions may yield to better and more refreshing ideas from different anchors. In this case, it would not be desirable if every customer listens to the same music and reads the same news articles and books, as different point of views are needed. I think for companies who want to generate new ideas from their customers, the homogenization of their customer base might be something to take into account.


De P, Hu YJ, Rahman MS (2010) Technology usage and online sales: An empirical study. Management Sci. 56(11):1930–1945.

Hosanagar, K., Fleder, D., Lee, D., & Buja, A. (2013). Will the global village fracture into tribes? Recommender systems and their effects on consumer fragmentation. Management Science60(4), 805-823.

Tsekouras, D. (10 February 2017), Lecture Customer Centric Digital Commerce, “Personalization & Product Recommendations.

“Buy a present for my wife” said Jan to the phone

This year St. Valentine’s Day caught millions of men by surprise, again, leaving them wondering what present to buy for their partners. What if somebody or something could perform this burdensome task in a timely manner? There might be a solution…


Viv is an intelligent personal assistant introduced to the market on May 9, 216 and acquired by Samsung in October 2016. Similar products such as Siri, Google Now, Microsoft’s Cortana and Amazon’s Alexa can perform some basic tasks but nothing beyond the tasks they’ve been programmed to do. Due to artificial intelligence, Viv can generate code by itself and learn about the world as it gets exposed to more user requests and new information.

This makes it by no means a universal product. Viv is expected to learn and store information about every user, and learn with time how to serve him or her personally. For example, if the owner asks: “I need to buy a present for my life for St. Valentine’s Day”, Viv should be able to predict what a suitable present would be or perhaps book a table for two at a fancy restaurant downtown.

Continue reading “Buy a present for my wife” said Jan to the phone

Utility-based recommendation systems

Decisions are hard, and with access to more and more options, we as customers encounter our limitations when it comes to making purchase choices. Luckily, companies provide us with recommendation systems that help the customer choose. They suggest similar products, link it to previous behaviour or show us what other customers have chosen. But recommendation systems do not only help the customers.  With the collection of all this, mostly personalized, data, it could also provide the company with more detailed information about their customers. This information can be used to come up with personal offers or adjustment of the prices based on the popularity of products.

Continue reading Utility-based recommendation systems

Help(l)ing Anyone

A dream came true for the start-up Helpling. They had a brilliant but simple idea and after only 1 year they are offering their service in more than 12 countries and 200 cities. But how did they get so successful? What is so special about Helpling?
There is a huge black market out there, a huge market of housekeepers that serves millions of households and a lot of these families have the same questions: Is it legal? Is this for me the cheapest and safest way? How can I find the perfect match? Everyone recognizes the issues. Millions of people need it, but it is difficult to find: a housekeeper that fits your budget, lives around the corner and you can trust. A difficult combination but Helpling has found a way to make everything more legal, transparent and they promise every customer to find the perfect housekeeper in 60 seconds.

Helpling is started in Germany when two guys presented a simple business concept to Rocket Internet. Rocket Internet liked it and one week after the meeting they already started to develop the concept. On March 29th 2014 Helpling is officially launched.
But why is this simple idea so brilliant? Partly probably because of the contemporary business model: co-creation. Nowadays customers want to contribute and support each other more and more. Helpling is making use of this trend.

Helpling is an online platform that matches customers with independent cleaning providers (called ‘Helplings’). Helplings register online, go through a registration process, choose their availability and after that they are able to receive bookings. Households who want to book a housekeeper have to make a booking online by entering their address and making an appointment. Then Helpling is matching a customer with a housekeeper based on location and number of positive reviews. So if a housekeeper is doing a great job, he gets higher rankings and he will receive more jobs since Helpling is offering jobs first to the best-reviewed housekeepers.

Through this way Helpling is supporting the Helpling community to review each other and the Helpling community improves the overall service by doing this. This recommendation-system supports the quality of Helpling.
Because Helpling is matching, also the perceived effort for the customer is becoming less and Tsekouras & Li (2015) found out that this means that the overall perceived quality would be higher. A customer has to put in a minimum effort and gets great results because of transparency and the Helpling’s review-based selection. Besides, because Helpling is helping customers to find the best-reviewed housekeeper, customers are more likely to do something back and are willing to write these valuable reviews as well.

Of course there are also some challenges of co-creation and the Helpling business model. It could be risky to fully rely on the individual housekeepers. They might seem reliable but housekeepers can also ruin the brand-image by misleading honest customers. Therefore, the most difficult part for Helpling is to recruit the cleaners, since one single Helpling can damage the brand, and they are now thinking about how they can safely scale this task. It is important to explore the risks and opportunities of this co-created platform.

“We would never do anything that harms Helpling as a brand. It only works if we can provide a good service. Otherwise you create hype and everyone tries it, and tries it exactly once, and you are in a really bad position.” (Benedikt Franke, 2015)

Luckily for now the quality remains good and there are increasingly more housekeepers and customers that are making use of this simple and helpful platform. If Helpling can live up to the ultimate dream to serve the world with their service? Give them one more year and we will see where they get.


Which one do you think I should pick ?

Today is Saturday, it is sunny and warm outside. Perfect day to go shopping ! You are obviously going to spend this afternoon with your favorite side-kick. Your mission is to find the perfect outfit that will fit your shape, your personality, and evidently match your preferences.

You finally find these really cool jeans, but you are still hesitant about the top : the blue and black, or the white and gold one ? The retailer comes to you and says that the jeans and the blue and black top would be the perfect match. On the other hand, your friend who has similar tastes as you – so similar you have a couple of clothes in common – votes for the second option.

What would you do ? Who would you trust ? How to make sure which one is the right choice ?

As a matter of fact, when you are browsing the internet, you might encounter the same problem-resolving process.

When a Youtube user surfs on Youtube, they first type the title of the video. On the right side of the video, a section displays other videos. In this section, two types of recommendations are presented :

  • Featured / Related videos :

They are based on the Youtube’s product network ; that is to say that they are based on the site recommendation algorithm only.

  • “Recommended for you” :

These are based on the Youtube’s social network. In fact, another user marked the video you just watched as favoriteas well as another one. Therefore, the second video will be “recommended for you”.

Thus, which of these videos matches users’ preferences the best ?

In a study, participants were asked to watch videos on Youtube, and rate each of them from one star – Poor – to 5 stars – Awesome. One group had access to the product network only – Related-Featured videos – the second group to both product network and the social network – Dual Network – and the third group to user-generated links only – Recommended for you.

What emerges from the study is illustrated in the graph below :

The fact that the second group curve is the lowest shows that finding a liked video takes less time when using the dual network than any of the other networks. Thus, rather than proposing either one or the other, offering both at the same time gives the user more possibilities, more choice, and therefore, more opportunities to reach the right video.

Therefore, the most efficient way for Youtube to satisfy its users, is to offer them as many choices as possible using different methods : the product network as well as the social network.

So, next time you will go shopping and do not know what choice to make, try the clothes on in the fitting rooms ! – or buy both.

References :

Jacob Goldenberg, Gal Oestreicher-Singer, Shachar Reichman, The Quest for Content : How User-Generated Links Can Facilitate Online Exploration, Journal of Marketing Research, August 2012, 462-468, 17p.

Yang Sok Kim, Ashesh Mahidadia, Paul Compton, Alfred Krzywicki, Wayne Wobcke, Xiongcai Cai, Michael Bain,People-to-People Recommendation Using Multiple Compatible Subgroups, AI 2012 : Advances in Artificial Intelligence, 2012.

Stitch Fix’s value co-creation

In the past it was easy to make a choice due to the lack of options among products. However, this has changed drastically. According to the NY times consumers see around 3,000 to 20,000 marketing messages a day. Moreover, we are bombarded with choices. Choice overload, which happens offline as well as online, exists for all kinds of products and services that we need.

Choice overload also exists when we shop for new clothes. Offline overload could exist when we go into a store and are presented with too many choices. Online overload happens when we browse a web store and experience choice overload. Source overload happens when too many websites and or platforms exist that makes it difficult to choose where to shop.

In 2011 Katrina Lake launched Stitch Fix. Stitch Fix is a fashion retailer that combines expert styling and technology to deliver a shopping experience that is personalized for the customer. It does so by having you first fill out a survey so that your personal stylist knows your preferences. Then it will send you five clothing items or accessories from various brands tailored unique to your taste. You can try all the items at home. You can buy what you want and then return the rest. Shipping and returning items are for free. Hence, Stitch Fix partially solves the choice overload that customers experience when shopping, by filtering the endless amount of choices into five choices.

Online recommendation systems used by companies such as Amazon.com learn from your preferences by tracking your searches and clicks in order to provide you with products that you might want. However, depending whether the algorithm and machine learning works well, this could lead to undesirable outcomes. For example, you could end up getting unwanted recommendations after you buy a gift for a friend or your mother who have different tastes and preferences in products than you have.

What makes Stitch Fix different than other online shopping experiences is that there is a human personal stylist involved in the item recommendation process. Your personal stylist handpicks five items just for you, based on your set preferences. Stich fix charges $20 styling fee for each item that you buy. But if you decide to buy all five items, you will get a 25% reduction on the total price of the shipment. The company gathers data and your feedback on the items that you decide to buy and the items that you decide to return to update your preferences. This will be then used to make better recommendations in the future. Hence, as stated by Stitch Fix, the more you make use of the service the better your personal recommendations will be and greater your satisfaction.

The success of Stich Fix proves that there is indeed a need for personalized recommendations and active customer participation. Furthermore, it shows that customers are willing to pay a premium on top of store prices for services that makes things convenient and or saves time.