INSIGHTS ABOUT PREDICTION
Finding the right balance among different methodologies to deliver the best recommendations to your customers.
How it’s done
In general, the main benefit of a recommendation engine is that it drives incremental sales (e.g. sales that otherwise would not have happened). We want to continuously increase customer engagement with our shop. The concept of recommendations is not new, and indeed, draws much of the same conclusions as targeted ads on the web. There are clear correlations between content relevance and increased engagement of the individual. Success is determined by how we use the information customers provide, and of course, what’s running under the hood of our finely tuned machine. A recommendation engine is like a funnel, with many different inputs going in.
Pure content-based techniques were often lacking in helping users find the documents they wanted. Keyword-based representations could do an adequate job of describing the content of documents, but could do little to help users understand the application of the keywords or the quality of those documents. Hence, a keyword search for “Chicago Rocks” might yield not only scholarly articles by the Chicago Rocks and Minerals Society but also “shallower” postings such as comments to a music message board regarding one visitor’s opinion of the 1970s rock band . Algorithms of this kind first build up profiles according to basic personal information and user interest fields, then check the degree of matching between user profile and the would-be predicted commodity, then finally select the commodity with larger similarities to make the best possible recommendation.
- Can be done offline
- Very predictable result (safe bet), as recommending the same color jackets or same brand shoes in this way, could mean the site loses incremental sales.
- Absence of personalized recommendations
- Synonymous handling
Collaborative recommendation has been successfully applied to various e-commerce recommendation systems. As one of the most successful approaches to building recommender systems, collaborative filtering uses the known preferences of a group of users to make recommendations or predictions of the unknown preferences for other users. This technology brings together the opinions of large interconnected communities on the web, supporting filtering of substantial quantities of data.
- Unpredicted recommendations — this is an advantage, as collaborative filtering recommends products based on customer interaction only, which means an individual can’t predict why this shirt is linked to other shirt and the answer is simply because other people like them, or click them together.
- Dynamic recommendation, recommended products changes as we get more customers’ opinions
- Cold start: new products added to the shop, or new customer registers (not enough actionable data)
- Matrix Sparsity: We will have a lot of gaps in the matrix (customer — product, customer — customer)
- Scalability: It’s difficult to scale collaborative-based recommendation when you have millions of customers and millions of products
- Shilling Attacks: for example, some Apple devotees tend to rate Samsung products poorly and without reason
Having one balanced recommendation engine is the result of mixing the collaborative based and content based filtering into one space which we call a Hybrid Based Recommendation Engine.
Is it easy? Absolutely not, merging two different spaces of recommendation with different attributes, conditions and values is a big challenge.
Everyone wants to have a recommendation engine based on user interaction, behavior, and preferences, whereas having that alone might put your business at risk because of the limitation of the collaborative model.
- Combine both content and collaborative models into one system, therefore you can gain the strength of both.
- Hard to implement (linking two different spaces in a sensible way).
Why Recommendations in fashion are so different:
When it comes to fashion, there’s a large community of influencers who impact opinions and decisions of would-be-buyers. One might see a movie and enjoy it — that’s straightforward, but when we talk about a t-shirt, it can be difficult to judge it based on specific attributes or customer behaviors around it. You might find that this t-shirt is not a good fit for everyday wearing, but it may be perfect when it’s part of a complete outfit.
With fashion, one must be responsive to seasonality, trends, idols, brands, and many other factors that might control your recommendations. Cost has a sizeable impact on the types of recommendations delivered. It opens the door to defining price points for the individual and also using that as an input for finding similarities between customers.
Unlike movies or books, defining attributes for fashion products can be a difficult task, as there are many sellers of the same products and each have their own methods of defining those attributes. Size for instance, varies widely between brands, and sometimes even within the same brands.
When giving products specific details and attributes, you try to ensure they are unique, which therefore makes it easier to recommend other products alongside.
Sometimes it’s hard to define all the attributes that relate to a single product, because one product may have hidden ones that appeal to customers, such as color and material. Thus, you need to apply certain advanced techniques to better identify the attributes. One useful technique is image recognition, which scans for and recognizes the features available from a set of images of the same product.
With our recommendation engine, we aim to deliver quality suggestions to customers, based on his or her individual preferences. It’s very much like peering into the future with a crystal ball. This is an intelligent way to open new channels for the customer to see more types of products they have already indicated having a preference for. In addition, it blurs the lines between a regular online shopping experience and moves more toward the realm of discovery shopping, meaning: we help our customers discover new products that they may otherwise have not seen through manual product searches / navigation. This has proven successful within the e-commerce model, therefore we can highly recommend it! ;)
J. Ben Schafer, D. F. (2007). Collaborative Filtering Recommender Systems. (p. 34). University of Northern Iowa.
code.talks Commerce Special 2017 Exclusive with Miriam Neubauer
ABOUT YOU TECH
code.talks Commerce Special 2017 Exclusive with Boris Lokschin
ABOUT YOU TECH