Saturday, August 24, 2013

Recommendation - Its all about the possibilities

Recommendation engine plays a crucial role in consumer websites/apps. Some applications like pandora are driven entirely based on the recommendation.





If we  take an example of consumer website like eCommerce, news website., there are various factors that can be taken into account for recommendation engine development.

I'll list down these possibilities for a better understanding and share some tips on implementation...

Collaborative Filtering and ML

Recommendation Engine can leverage Machine Learning frameworks like Mahout  for the below types of recommendations.

User Based - Recommendation engine can give recommendation by mapping the similar user and their preference.

Item Similarity - Similarity between items are calculated and recommended

User based and Item Similarity recommendations comes under Collaborative Filtering technique.

Content Based - We can compare the attributes between items and users to provide similarity. It is especially useful when we don't have sufficient data to perform collaborative filtering. We can use some similarity algorithm from ML library like Mahout or we can use search system like SOLR for this.

Clustering - Clustering techniques can be used to group the similar items

Other Possibilities:

Below I'm listing other factors that can be considered for recommendation. I believe these things are more intuitive and we can identify them when we see from an audience view point.

Out of user boundary / Opposing Views -   We can recommend something fresh out of an user boundary.  Something user never explored before. It can also be a content from others with opposing views.

History - Recommend an item by tracking user's viewing history. Views can also be tracked by a heat map or attention to particular area of web page by the user.

Top selling  Items - Recommending Top selling items, Top 10 items .etc

Trending Items - Trending/Most shared item in social media like facebook, youtube, twitter

New Releases -  New release may trigger user interest. This is highly relevant in fashion world.

Friend's actions - Recommend an item based on the friends actions by leveraging social media. But this is not always relevant. I may not want to buy the same shirt my friend wear but I like to read the news he commented.   In the section  - "On Implementation" - Uniqueness and Weightage talks about this

Location based - Some purchase decisions may be location based. When a user located in Mumbai and if it is a monsoon, we can recommend rain accessories applicable to that location.

Map user personality - We can map user personality  and recommend an item. We can pull information  from social media sites like likes, favorites etc and map the user personality. We can also map a personality by mapping user's past actions with the product. A hybrid of both approach also works.

Personalities can be like this - One who try new things, One who spend on premium/expensive stuff,  One who always purchase on deals .etc

Time based - Decision making may be different from starting of the week, end of the week, month starting. etc.

Screens - Recommendation may be different on various screens like laptop, mobile, tablet. For example, if your customer using your mobile app and you able to access his location, you can recommend an nearest store deals, distance to nearest store .etc

Still there  may be many  things to factor, more techniques &algorithms that can be useful to give a better recommendation to the user.

On Implementation


How to factor all these possibilities on the implementation?. I'm sharing some tips on the implementation front...

Unique Approach -  Unique approach for each usecase. If we take this recommendation example, recommending an article in news website is different from recommending a shirt in an online apparel store.

Hybrid Models-  Combining the various approaches/possibilities may work. Here we can combine multiple ways of recommendation.  It can be a combination of user based  +location based +something else.
Its not always Yes or No, try to find alternative paths.

Weightage - Weigh the features, have a strategy and incrementally build and release the features. For instance, you can build the basic recommendation engine using item based or content based and incrementally add the features based upon user testing and feedback.

Right Metrics - Have a right metrics.  For the above example, the metrics would be how a consumer reacts to the recommendation, what is the change of traffic pattern .etc.

Place for Experimentation - Have a place for experimentation. Check and measure with labeled or test data. You you can also experiment with A/B testing to test which approach actually works better.


No comments:

Post a Comment