There are already quite a few recommendation engines, services and products that have recommendation features. Each product takes their own swing at a strategy for recommending new products and services that users will like. Today’s recommendation atmosphere does not seem to contain any perfect solutions, but they’re each utilizing strategies that help narrow down their catalogs and databases of information to content that is more customized to someone’s preferences.
Common Techniques for Delivering Recommendations
- Word of Mouth
- Expert Ratings
- User Preferences
- Collaborative Filtering
Less Common but More Effective Techniques for Delivering Recommendations
- Taste/Attribute Filtering
- Taste Profile Filtering (Expert Ratings + User Preferences + Collaborative Filtering + Taste Profile Filtering)
Word of Mouth
Almost everyone has asked their friends and family for suggestions of where to eat, what movie to see, what book to read, or what consumer electronics to buy. Word of mouth is often an easy way to get a quick head start in new topic areas, as the people you are close to might have already spent some time there. You will consider taking a look at something is someone you trust offers you a suggestion or product referral.
If you do not share specific commonalities with these people, the chances that you will like the recommendation are slim. In order to get good suggestions, you would need to be socially connected with people who have a similar set of preferences and interests (such as a western movie club, or alternative rock concert troop).
Famous critics and reviewers evaluate products and services, and meticulously analyze it based on their own scale. People over time begin to appreciate particular critics whom they value the reviews, ratings, and recommendations from. Experts tend to offer a critical look at particular topic areas (movies, books, music and wine). They’re unable to test and trial every possible product in a topic area, so they often focus on niches (scary movies, historical books) or incredibly broad (pop).
Experts typically create a rating scale (1 to 4 stars, thumbs up or down, or 0-100 points) offering users a quick way to see what the critic thought.
The problem with only relying on expert ratings is that people do not tend to agree with everything that a critic says. It’s often quite hit or miss. Frequent misses causes a person to reassess whether or not the critic is useful to them anymore, requiring them to discover another critic that might more closely match. Edge case users — people with more unusually eclectic taste — will often have no other critics to turn to once they’ve found a critic that offers at least a few hits.
User Preferences offer a way to understand a user’s general preferences for particular groups and categorical clusters. When a user contributes a particular preference, you can shut out or include entire categories of content. Take, for example, a user who prefers classical music. Knowing that you can more confidently recommend a CD of music by Bach over a CD of music by Korn. If the user further provides that they don’t like dance or rap music, you have a more specific understanding of that user’s interests.
But unfortunately, not every user is black and white about their preferences. More often than not, people like a few hiphop songs here, a few rock songs there, and a country song or two. Gross categories could have excluded most of those, even though the user might actually like them.
Collaborative Filtering utilizes collective intelligence to determine recommendations. This is one of the currently most popular ways to recommend. It is often based on data contributed by a graph of other users that are most similar to a person, although there are quite a few varieties of collaborative filtering.
- Historical Purchases — what a user has purchased in the past determines other items they may prefer, using either a social graph (what others also bought who bought items similar to you) or an item correlation map (relationships between items purchased).
- Social Graph Ratings — recommendations are determined based on the user’s rating similarities, placing them into a neighborhood of similar users and the highest rated items are returned that the user has not yet rated.
If items have been mapped with a structure of meaningful attributes and then ranked for each attribute, a user can provide the level of each attribute that he prefers. For example, a user might specify that they like suspenseful elements and extreme character development as qualities in books they prefer to read. Assuming a collection of books have been mapped as having these attributes, they can be recommended to the user.
Mapping a user’s taste attributes to the attributes specifically defined in various topic areas helps to define an understanding of that person’s taste more precisely that the other methods defined above.
The first challenge to this technique is that it requires a lot of upfront modeling and attribute research (preferably from a panel of experts) in order to identify the attributes for a topic area, and then the attribute rankings for the products or items. This could be costly.
The second challenge is that users would need to provide a reasonably amount of information about themselves, perhaps via a survey or quick questions over a longer period of time. Until users can feel confident that the system will generate accurate recommendations, they will be hesitant to contribute personal information about their taste.
These two challenges is likely the main reason why most products have yet to utilize this strategy yet. However, it offers the most precise results out of the strategies I have presented in this article. This is one of our primary focus areas at WeLike.
Taste Profile Filtering (Expert Ratings + User Preferences + Collaborative Filtering + Taste Profile Filtering)
Slightly more experimental, and also another area that I have focused on at WeLike, is combining all of the strategies together. I call this Taste Profile Filtering. Taste specific recommendations can be provided using a combination of the strategies together, while also weighting them by a topic area model, or the user’s desired order of importance. This allows maximum flexibility so that your recommender system is not tuned to a particular group of people. For example, it still allows people that do not wish to tune their taste attributes, but want to rely on a social graph neighborhood, for more popular recommendations. Then it also allows people who are more eclectic and on the edge to more closely tune their actual taste.
I believe this is where the future of recommender systems is; this is where it will be headed over the next 5 to 10 years. As the number of choices continue to grow over the coming years, sifting through and finding what is most meaningful will be necessary, and accurate recommender systems become ever more important and vital to end users.
If this is something that interests you, I’d love to chat. I am looking to find an additional engineering co-founder to add to the team at WeLike. Send me an email at email@example.com.