Friday, August 31, 2012

Designing The Next-generation Review And Recommendation System

It's unfortunate that despite of the popularity of social networks and plenty of other services that leverage network effects, the review and recommendation systems that are supposed to help users make the right decisions haven't changed much.

Thumbs-up and thumbs-down or likes and unlikes signal two things: popularity and polarization. If a YouTube video has 400 thumbs-up and 500 thumbs-down it means that the video is popular as well as polarized, but it doesn't tell me whether I will like it or not. The star review system also signals two things - on average how good something is and whether it's significant or not. There are multiple problems with this approach. An item with 8 reviews, all 5 stars, could be really bad compared to an item that has 300 reviews with 3.5 stars. Star ratings alone, without associated descriptive reviews, wouldn't make much sense if there aren't enough people who have reviewed the item. Also, relying on an average rating alone could also be problematic since it lacks the polarization element. On top of it, the review and likes could be gamed.

Pandora's as well as Netflix's recommendations are a good example of using collaborative filtering to fine tune recommendations based on user preferences. The system aggregates the overall likes and dislikes and combines that with your taste profile and a few killer algorithms to recommend what you might like. If designed well and if it has large user population, it does work. But, the challenges with such system are missing descriptive reviews and lack of ability to perform any analysis on it. If I dislike a song on Pandora, it doesn't mean the song is bad in the absolute sense. It simply means it doesn't match my taste profile. This isn't entirely true if I dislike a blender. In this case, a descriptive context is more meaningful such as I don't like this blender because it doesn't crush spinach well. People who care to make smoothies and crush ice may not care about this issue. But, these consumers have to wade through large number of reviews to determine the product fit.

E-commerce sites review systems use the same descriptive as well as non-descriptive review systems, commonly used at all places on the internet, without any significant modifications, even if the expected investment of a user is much higher on their site. If I don't like a song, I can skip it. If I don't like a YouTube video, I can stop watching it and now if I don't like a movie I can stop streaming it. This does not apply in the traditional world of e-commerce. I absolutely need to make sure that I buy something that I like. Returning an item is a far more involved process than stop watching a movie. It's an exception, not a norm.

Word of mouth and passive buying

People shop in two ways: 1) they look for a specific product, research for it, and buy it. 2) they come across a product while not looking for it, like it, and buy it.

The second way of shopping, passive buying, is as important as active buying. There are many companies with a business model built around this impulse or "serendipitous commerce", but they don't leverage collaborative filtering. I would happily read reviews of products written by my friends and people that I trust regardless of whether I'm looking for those products or not. Think of it as Disqus-style aggregated reviews by people that I trust in my social graph. This is like an online version of a cocktail party conversation where someone is raving about a new phone that he just bought. I'm not looking for a phone, but I might, in a few days. This could create new interest or expedite my decision process. This isn't done well in the online world.

The word of mouth is still by far the best system for following recommendations. I invariably watch movies that my brother recommends to me and one of my friends will read all the books that I recommend to her. I have non-transactional relationship with my friends and family.

Contextualized long tail 

One of my favorite things, when I travel (leisure or business), is to try out at least one or two recommended Indian restaurants to see how Indian food compares from city to city and country to country (so far my vote for the best Indian food outside of India goes to London). While researching for a restaurant, I typically read all the reviews that I can find. Some reviewers are Indians and some are not. Also, for the reviews written by non-Indians, some are new to Indian food and some are not. In most cases people don't identify who they are and I end up guessing based on their username, description etc. These reviews, positive or negative, don't help me much to narrow down which restaurant I should try out.

I have always found the best food at the most unusual places. All sophisticated recommendation systems would fall short of helping me find such an unusual place. These places are not the hits. They are the long tail. Getting to this long tail isn't an easy process - a lot of asking around, digging for reviews, trying out a few awful places etc.

Privacy concerns and connected identities

As the debate between anonymity and identity continues, there has been a little or no effort to get to the middle-ground, a connected identity. As a marketer I don't care who Jane is in its absolute sense but I am interested in what she likes and dislikes based on her collective and aggregated behavior across the Internet and beyond. This is not an easy system to build and consumers won't sign up for this unless there's a significant value for them. The popularity of social networks is an example where even if users are arguably upset about their privacy they still use it since the value that they receive far outweighs their concern. And remember the social networks follow the power laws. As more and more people use it the network becomes more and more valuable to the users.

Why not design review and recommendation systems that are based on connected identities? Users don't want ads, the marketers do. If companies can focus on building good products, incentivize users to write reviews, and rely on great recommendation systems to connect the right users with right products they wouldn't need ads. The marketers are chasing the illusion of targeting the right users but the inconvenient truth is that it's incredibly hard to find those users and if they do find them, they don't really want ads. What they really want is value for their money. That is the inherent conflict between the marketers and end users.

Using connected identities beyond reviews and recommendations

Connected identities are also useful beyond reviews and recommendation systems. Comcast support is one of those examples where using connected identities could greatly improve their customer support.

Comcast started using Twitter early on to respond to customers' support issues. It was a novel concept in the beginning and they really understood Twitter as an effective social media channel, but lately that model has turned out to be as bad as their phone customer support. When I tweet to @comcastcares someones gets back to me asking who I am and what issues I have. You follow me, I follow you, you DM me, I DM you my info, and after few minutes, we are nowhere close to resolving the issue. What if Comcast allowed me to attach my Twitter account to my Comcast profile? I will OAuth that, for sure. When I tweet, they exactly know who I am, what problem I am experiencing, and how they might be able to help me. This is an example of using a connected identity without compromising privacy. Comcast knows their customer's billing information; it's transactional information. But they attempt to use Twitter to communicate with you without connecting these two identities.

I don't want to "like" Comcast or "follow" Comcast to be a victim of their spam and indifference. Comcast is easy to pick on, but there are plenty of other examples where connected identities could be useful.

Users don't like to be sold at, but they do want to buy. Let's build the next-generation review and recommendation system to help them.

Monday, August 20, 2012

Applying Moneyball To Cricket

What if a cricket team gets two batsmen to replace Sachin Tendulkar and still collectively get 100 runs out of them or have two not-so-great bowlers to replace Shane Warne and still get the other side out? If an eventual goal is to score, say 300+ runs in an ODI match does it matter how the runs are scored? What if you could find four players scoring 50 runs each instead of counting on Sehwag and Tendulkar types to score a century and lose miserably when they don't?

This is not how people think when it comes to cricket. That's also not how people used to think when it came to baseball until Billy Beane applied radical thinking to baseball, sabermetrics, now popularly known as Moneyball. On Base Percentage (OBP) became one of the most important metrics since then.

As yet another provocative aspect of Moneyball suggests, only thing that matters is whether a hitter puts a ball in play or not. Once the ball is in play the hitter does not control the outcome of that play. In cricket, when a fielder drops a catch could it be because the ball came too quickly to him, he was at the wrong position, or he was just too lame to catch it. Is there a difference between a batsman getting caught near the boundary as opposed to getting bowled? Currently, none. But, based on Moneyball, if a batsman gets caught, at least that batsman put the "ball in play." A little more practice and precision and that could have been a four or a six.

I want the cricket team selectors and captains (an equivalent of baseball general managers) to apply some of the Moneyball concepts to cricket, a sport older and more popular than baseball. In cricket, even though it's a team that wins or loses, there's typically more emphasis on the ability of an individual as opposed to measuring individuals in the capacity of how they help the team.

Bowling and batting powerplays are relatively a new concept in cricket. Skippers on either side don't have access to deep analysis of current situation and performance of opposite players in deciding when to take a powerplay. They make such crucial decisions based on their gut feeling and opinion of key players on the field. This is where data can do wonders. In baseball, managers keep a tab on an extensive set of data to make dynamic decisions such as which bullpen pitcher has a better track record against the current hitter, success of a hitter to get walks as opposed to hits etc. Most recent example is of Tampa Bay Rays aggressively using field shifting against powerful lefties, a practice that most baseball franchises still don't use or approve of.

In cricket, right handed bowlers switch from over the wicket to round the wicket mostly when whatever they are trying is not working. These decisions are not necessarily based on any historic data. In this case, it could be as simple as gathering and analyzing data about which batsmen have poor performance when bowled round the wicket as opposed to over the wicket. In baseball, using a left-handed pitcher against a left-handed hitter and using a right-handed pitcher against a right-handed hitter have proven to work well in most cases (with some exceptions). That's why there are switch hitters in baseball to take this advantage away from a pitcher. Why are there no switch hitters in cricket?

Why can't there be a dedicated bowler to finish the last over of the cricket match just like a closer in baseball? Imagine a precision bowler — a batsman who is trained as a "closer" — whose job is to throw six deliveries, accurately at a spot, fast or slow.  The regular bowlers are trained to bowl up to 10 overs, 6-8 at once, with a variety of deliveries (pitches) and a mission to stop batsmen from scoring runs and getting them out. A closer would only have one goal: stop batsmen from scoring. Historically, there have been a very few good all-rounders in cricket. It's incredibly difficult to be a great batsman as well as a great bowler, but there's a middle ground - to be a a great batsman and a closer. Some batsmen such as Sachin Tendulkar have been good at bowling off and on when the regular bowlers get in trouble (an equivalent of a reliever in baseball), but invariably their task becomes getting a wicket to break the partnership. Even if wickets are important, in most cases, it's the ability to stop the opposite team from scoring in the last couple of overs brings team a victory.

There is just one baseball, but there's no one cricket. The game of cricket differs so much from a test match to one day international (ODI) to Twenty20. But, a fresh look at data and analysis on what really matters and courage to implement those changes could do wonders.