Recommendation Engine


The purpose of a recommendation engine is to increase sales or traction (for example web page visits, or ad click-through). In order to achieve this goal, Recommendation systems typically produce a list of relevant recommendations. They rank the various options in the order to increase the chances of achieving the desired objectives.


Characteristics of Recommendation Engines

There are several level of sophistication to achieve the desired objective. The basic level consists of presenting to the user a sub-set of products/options from the entire list of viable choices. The next level would be to customize the offering to the customer needs, so the list of products/options is optimized and personalized for every single individual. The third level of sophistication consists in understanding the customer “sensitivities”, their “hot buttons”, what makes them “buy” or act. This consists primarily in capturing or inferring customer tradeoffs and preferences. How much they value each attribute of a product/solution/option. Finally the utmost level of sophistication consists in customizing the product/option based on the consumer preferences. i.e. this could be generating on the fly a webpage that meets more closely the customer profile, or packaging a product (including price) to better meet the customer need. For example if a customer expresses a strong preference for more performing computer the goal would be to beef up the CPU and Memory so that the configured computer meets closely the consumer individual tradeoffs. By understanding consumer tradeoffs, you realize how much they are willing to pay for each feature, such as a roof top, or customer service, or a certain color. Thus you can configure the offering in a way that meets their needs. Imagine a website you build on the fly based on the latest visits of the user, or a car that is configured and priced based on the tradeoffs expressed by a customer. With the exception of Tradeoff technology, no Recommendation engines vendor has pushed the envelope to that extent.


Why Tradeoffs Engines

Why did Gartner pick Auguri as its “Selected Vendor for Recommendation Engines”? Why does Auguri win hands down every head to head benchmark it has entered against other solutions? What does the Tradeoff based Recommendation engine deliver that others do not:

1) Provides personalized recommendations irrespective of whether there is a history of behavior or not. As it is based on a dynamic online interaction.

2) The personalization is not based on the others behavior. It is based on the individual preferences and tradeoffs. Preferences reflects the appreciation of various colors. Tradeoffs reflects the importance attached to color in comparison to the other criteria.

3) The sellers and manufacturers can understand why an individual or a group have made a given purchase or selection. Tradeoff systems can even go further and measure the tradeoff that each individual associates with each attribute. i.e. how much would an individual pay for an additional horsepower or mile per gallon or sunroof.


Recommendation Engine Comparison

Before the advent of Tradeoffs, there were four main approaches to Recommendations:

Parametric selection consists in setting some constraints to narrow down the choices. Products that meet these constraints are selected and the others are eliminated. It requires the users to be intimately familiar with the metrics of the data stored. i.e. to select a car using curb weight, requires that the user be versed in the art of curb weight, including a solid knowledge of the metrics used in the system. In addition, it takes usually several trial and error to obtain a reasonable number of results. It simply does not compare to tradeoffs. The concept of tradeoffs is to add the element of importance to the criteria and this takes into account what matters for EACH individual user. Thus yielding superior results EVERY time. For example a Parametric Filtering system would bring the same cars for 2 users that select Price < $25K and Performance > 300HP, irrespective of the fact that the car buff really care about Performance while the one who is more price sensitive is willing to compromise on the performance for a less expensive vehicle.  This is where tradeoffs wins EVERY time as it will yield more relevant cars depending on the tradeoffs of each individual. In addition to yielding superior results, the tradeoff approach can shield the necessity to be an expert in the domain. As the system capture default knowledge imparted from the experts or learned from the users.

Rule Based Systems work by applying a set of rules that affect some factors. These factors may relate to the users or the product. There are some cases where Rule based systems are the best options. They typically occur when the user has no choice. For example if a vendor sells components to build your own table, some table tops can work with square legs only while other work with round legs only. In these cases, irrespective of the user preference, the leg selection is limited by the forced feasibility. In general Rule based are the best solution when the constraints are not optional. Whenever the configuration of a system is deterministic, there is no way around rule based systems. Although they have been extensively been used to allocate users to buckets (e.g. Product), these systems are not meant to and are not really good at matching customer needs with product features. They are even worse at understanding customer preferences. These systems are typically complementary. Tradeoff system will win any time on recommending solutions but cannot work well in deterministic situations.

Collaborative Filtering predicts a user’s behavior based on closest neighbor’s past behavior. It has a couple of advantages: (i) it does not require an understanding of the underlying item being recommended and can work with basic behavior data. (ii) The technique takes implicitly into statistical account all factors. However, that fact that collaborative Filtering is based on the premise that all statistical neighbors behave the same way is the source of several problems: (i) the mere premise defeats the purpose of a recommendation engine that is to differentiate and personalize the recommendation (ii) Have you ever driven in a neighborhood where everyone drives the same Jeep Cherokee? Therefore the premise that the entire system is based upon is at best questionable. Does require a large amount of data. (iii) it requires collecting a sizable amount of data (iv) it provides no insight with respect to the reasons for which a given product is selected, much less which attribute are more favored than others. (v) the system cannot adapt to new products that have no previous traction. In the absence of interaction, it is a decent way to guess the user behavior. It does not compare favorably to tradeoffs when web interaction is available. It does not have the same level of personalization. Most importantly it provides no insight as to why a product was selected over another. There is no way to infer the value assessed by customer for each attribute as it is the case with tradeoff based solutions. Tradeoff based solution have won every head to head benchmark.

Content Discovery utilizes Bayesian and other statistical techniques on keywords describing a series of discrete characteristics of an item. A widely used algorithm is the tf-if representation (also called vector space representation). The main advantage of this approach is that it is well suited to leverage web sentiments, trends, buzz and consumer opinions expressed in blogs and other social networks. However, keywords are not a good technique to analyze and compare products. Tradeoffs can take into account the user opinion and buzz on products as one of the various factor while adding to that product specification and allow the user to determine the importance of these various factors.


The best analogy is to think of a customer walking into a Blockbuster store. If the store used a content discovery it would lookup his friends and neighbors and see their sentiment about the various movies using keywords statistics to recommend a movie. If it used collaborative filtering, it would lookup the age, race, address, income, and other factors about the customer and find based on that a group of similar customers and recommend the most popular movie in that group. if it used rule based system it would rely on a set of rules that the smart engineers put together such as “if customer has betamax eliminate movies in VHS and CD formats”. If it used parametric search the sales person would ask the customer about the actor, the genre, the length of the movie, the producer, take everything into account equally and find all the movies (if any) that meet these requirements. If it used the tradeoff system, the customer would be asked about the mood s/he is is in? what s/he is looking for? what matters most, the genre, the actor, the feedback of her/his friends, the critiques rating, or eh novelty? By capturing the relative importance of these criteria the system would be able to recommend a ranked list of product that are significantly more relevant than with the other techniques, thus yielding a better chance of renting and a higher customer satisfaction, thus a repeat business. In addition, by feeding back the customer tradeoffs to the movie maker, these would be able to create movies that are better suited for every audience. Thus increasing sales on the longer term. If movies were configurable and could dynamically evolve depending on the interaction with the viewer, it is conceivable that eventually movies would be produced in a way that they could evolve differently depending on the mood, and tradeoffs of the viewer at the time of viewing. i.e. on Monday afternoon on a rainy day the movie would  be a drama with more “grey” scenes, while on a sunny day at 2pm the same movie would evolve with more sunshine, humor and a happy ending, to satisfy the mood of the interactive viewer. This would not be possible without a solid understanding of the preference and tradeoffs of the customer.

Conclusion: It’s all about Conversion

The bottom line is that it all boils down to conversion. The Recommendation Engine that yields the highest conversion rate (converting visitors into acting customers) will be the one to offer the best ROI. The more sophisticated and interactive the engine, the higher the chances of leading to a transaction. Tradeoff technology is as a result the technology that will offer the best odds to increase the top line.