Recommendation Engines, also statistical Recommmender Systems,are automated recommendation technologies. They filter information based on statistical data in order to find matching elements (product) according to another element (product). This enables them to recommend products matching to other elements (products or user profiles) out of a wide range of products.
Statistical Recommender systemes are based on two technics:
- Content-Based Filtering (CBF)
provides recommendations depending on the correlation and similarities between two elements (e.g. two books of the same genre)
Collaborative Filtering (CLF) provides recommendations that focus on the similarity between the user and privious users. It analyses the user's profil and recommends the user products, which other previous users with the same or similar profil have chosen.
The goal of statistical Recommender Systems
The most important goal of a Recommendation Engine is to reduce the insecurity of users and thus to lead them to a certain buying decision. This is achieved by providing information about similar products, alternative products, or products that have been bought by users in a similar situation.
Statistical recommenders use historic data, provided by previous visitors, to calculate recommendations. Matching Engines on the other hand calculate recommendations on the basis of data that the active provided by the user or their user profile.
The advantages of well configurated statistical Recommender Systems:
- automatic calculation of product recommendations (with sufficient data of usage and/or profile)
- it does not require any manual knowledge modelling (often also not possible)
- recommendations can be calculated even if the element cannot be classified by attributes. (e.g. music or movies)
Disadvantages of statistical Recommender Systems:
- Recommenders need a huge amount of data in order to learn which recommendations are bought and useful. Therefore, they require a lot of traffic on the website, in the specific product categories and for the product itself.
- Recommenders do not provide useful recommendations for the Long Tail:
The Long Tail of products, so all the products of a sortiment which are bought rather rarely, is dominated by outliers of usage data.
- for that reason statistical recommender systems can also offer products that do not match or are, objectively speaking, useless. (e.g. a user interested in a speical and rearely bought notebook receives a recommendtion of a purple G-string due to the fact that the privious user bought this combination)
- This causes difficulties especially for new products or quickly changing product ranges due to the fact that the Recommender System does not receive sufficient usage data during the lifetime of the product. This way statistical Recommender Systems can not learn effectively which recommendations are meaningful and which are useless.
- another significant disadvantage of statistical Recommender Systems is that they do not guide customers actively to a certain buying decision and can not follow a specific business strategy like a Guided Selling System. They simply recommend and combine products which previous users have bought and do not consider preferences of online shops.
- statistical Recommenders do not provide a reasoning for their recommendation because they do not identify and understand product attributes. However, a strong and objective reasoning is essential to lead customers to a certain buying decision.
A Recommender System is not an alternative to free text search, Matching Engines, Guided-Selling Systems or filter searches. In some cases, if objective product recommendations are not required (e.g. music, videos) or recommendations are made accross several caterogies, recommender Systems can serve as effective addition.