I retired from personal blogging in July 2008.
But you can find me over at http://blog.xero.com.
Posted by Rod in TechBiz at 2:48 pm on Sunday, 23 March 2008
A more positive example of a New Zealand eCommerce site is DVD rental site Fatso.co.nz.
DVD rentals online is semi complex online and real world logistics orchestration. Fatso is slick. They have great promotions, great communications and it all just works.
One of the best New Zealand online experiences I’ve had.
They have a free trial so give it a go. I’ll never go to a movie rental store again.
Hopefully we’ll see more Blu-Ray movies available soon. I’ve already seen most of the ones I’d want to watch.
blog
Been a Fatso customer for a couple of years now and you’re right - we haven’t been back to Video Ezy or Civic since. It’s a great service.
Don’t forget either - if you refer three friends who then sign up as customers, you’ll get an IPod Shuffle.
I personally used MovieShack.
http://www.movieshack.co.nz/?_hello
Same idea, EXCELLENT communication etc. Highly recommended.
Amazon do it over here in the UK. I imagine there are a few others too. I’ve only been past one video shop in London - a video and internet cafe type thing in Clapham. Never seen another one.
For me, however, it’s itunes all the way :)
Just been browsing thru the Fatso site and I see that they’re using some form of recommendation engine. That’s excellent. I don’t see many e-commerce sites in New Zealand using online product recommendation engine. If you click on a movie item on Fatso’s list, it comes with a recommendation of 4 other items:
Customers who like this also liked these titles
I am surprised that the Herald IT journos haven’t given a product overview of Fatso. Herald IT section from time to time have lamented the lack of any local ecommerce site that adopts an Amazon-type technology (Amazon automated product recommendation is what they’re well-known for). Perhaps, Chris Barton from the Herald might want to arrange a visit to Fatso and interview them for a featured article on the Herald instead of his usual anti-Telecom useless piece that fills the Herald business section.
One of the fundamental goal in the development & deployment of automated online product recommendation is the accuracy, ie, retrieve & recommend items that are of interest to the user and not recommend items that are completely irrelevant to the query (the item the user has selected in his/her query). An engine with low classification error has more robust than one with high classification error. Suppose an engine recommends an(some) irrelevant item(s) once for every 10 queries, then the classification error is 10%. An engine that gets the recommendation wrong (irrelevant recommendation) say once for every 20 queries has a classification error of 5%. The engine with 5% capability has a better performance compared to the one that has 10% classification error. When Amazon first deployed their recommendation engine, they monitored the rate of its adoption by online users. The classification errors were high that they had to improve it, because they found out that some users had bad experiences with irrelevant recommendation of the engine, that they never used it (book recommendation) in their subsequent visits. When they improved their engine’s capability (low classification error algorithm), they noted that users who had experienced the bad or irrelevant recommendation have started using the recommendation again, since users are recommended items that they perceive as relevant to what they were looking for.
Last year (2007) I was alerted (by a Java architect from Wellington) to the Netflix’s (an online DVD rental service) competition for a one million US dollar prize for the best system with algorithm that has the lowest classification error for online movie recommendation which they detailed it here. This Java SOA architect proposed to me if we could form a small NZ team (perhaps 4) for this competition. I could contribute in the recommendation algorithm development while other members concentrate on web development & its architecture. I was interested first, but then I saw that Netflix had advertised the competition here at KDnuggets and I immediately knew that if we participated , there would have been no chance at all. KDnuggets is a news repository for machine learning & data-mining researchers and those who are involve in this community are top of the top in this sorts of recommendation algorithm development. I had some algorithms in mind to explore which one has the lowest error, but these are from publicly available literatures and there was a chance that some team from somewhere were thinking the same algorithms as I was thinking. Machine learning researchers who have participated would implement something that wasn’t available in the literatures (they would have invent one that perhaps better than algorithms available from the literatures) and this was the reason that if we participated, we had no chance at all. There are many algorithms for online recommendations, but they varied in speed & accuracy (classification error).
I will alert Fatso about the author of this paper , where he had made his matlab code of the Maximum Margin Matrix Factorization (MMMF) algorithm available to be downloaded from here. In this paper, it is reported that MMMF has lower error rate (ie, more robust), than similar existing algorithms for movie recommendation engines. Perhaps if Fatso is interested to benchmark MMMF against their current system, then I am happy to help them.
The author of MMMF algorithm (Matrix Margin Matrix Factorisation) has blogged about the performance of MMMF compared to other movie recommendation algorithms here, where he quoted the following:
We compared versus the nine different algorithms (including an improved variant of the SVD approach) which Ben Marlin tested in his Master’s thesis and found that MMMF substantially outperformed all of them, on both data sets, according to both evaluation criteria that Marlin proposed.
I am pretty much sure that whatever recommendation algorithm that Fatso is using, it is not MMMF, but if they’re keen to improve the accuracy of their movie recommendation engine, then perhaps they can’t ignore the superior performance of MMMF. The MMMF code is freely available.
Blockbuster here is the US is also doing something interesting to combine bricks and mortar with online. Same as Fatso, Netflix etc… but they also allow you to drop into a store and swap out movies. So if you get stuck without a fresh movie for a quiet Sunday on the couch you can pop in and swap out a movie you got via mail.
Netflix are also expanding their downloadable catalog of movies. You can get subscriptions for say 2 physical DVDs and 5 downloaded ones a month for example.
-CJ.
I’ve used Fatso in the past and I’m currently using Dvdunlimited. The problem I’ve found with these services is that you have to keep a substantially longer list of movies than the minimum of 20 they recommend, in order to always have say 3 items when you’re on the 3 item plan. They simply don’t have enough stock of the movies to meet the demand. I eventually cancelled my Fatso account for this reason. Unfortunately Dvdunlimited is going the same way and I will probably be cancelling them soon too. At Dvdunlimited I only have HD-DVD, Blu-ray and Games on my list and have at least 24 selected and they cannot ensure that I always have 3 items on my 3 item plan.
The other part that I don’t enjoy is the unpredictability of the item that you eventually receive. Going to a physical store at least mean you can choose exactly what you want to view at that particular time.
Video on demand is what I really want, but that will still be too far off - especially locally.
I see there’s a session at this years JavaOne conference about Project Aura - an open source recommendation engine:
http://blogs.sun.com/plamere/entry/project_aura_at_javaone
Some other links:
http://blogs.sun.com/plamere/entry/recommendation_for_the_rest_of
http://research.sun.com/projects/caroline/