Resources - Blog
Technology Series: The Benefits of Machine-Learning Algorithms
The technology series focuses on Feedvisor’s AI platform, data-driven intelligence, and machine-learning algorithms. The series illustrates various ways in which our technology assists our customers in their e-commerce activities and challenges, with Amazon at the center. This post is led by Feedvisor’s VP of Research, Yehonatan Elon.
While for a long period of time artificial intelligence provided an unfulfilled promise to a better — or at least a more efficient — future, it seems that at last the AI revolution is here.
Intelligent machines and the technology behind AI, known as machine-learning, are of tremendous value not only in providing human-like services (such as virtual assistants or autonomous drivers) but most commonly in making complicated decisions in front of big data.
This blog is one of the first of its kind, intended to give you a taste of Feedvisor’s AI technology. In this post, I am going to review some of the advantages of applying machine-learning tools as part of Feedvisor’s methodology.
Rule-Based vs. Machine-Learning Algorithms
The common ‘old world’ practice of automated systems is based on rule-based algorithms, i.e. the system is fed with deterministic rules defining what output it should generate for each valid input. Rule-based systems are not necessarily simple. On the contrary, in order to control a complicated task, many irregular behaviors must be treated explicitly by the system to avoid failures.
On learning systems, instead of instructing the system with the solution for the task, you define the system to your goal (such as maximizing the user profit), and the system tries to find the optimal solution by itself. In a sense, if a rule-based algorithm is baking a cake using a recipe, a machine-learning algorithm is more like attending a konditorei school.
While the design and implementation of machine-learning algorithms may sound more demanding, it bears inherited advantages when applied to sophisticated arenas, such as the Amazon marketplace.
A critical feature of software products is the automated tests that validate that the system is doing what it is expected to do. Typically, automated tests verify that the system behaves correctly. For example, they will alert me if I publish a price for a product in Euros if I am selling on a U.S. marketplace, or if I try to advertise a product using an invalid keyword.
With rule-based systems, it is much harder to apply integral tests that examine how well the system performs with respect to its goals.
On the contrary, machine-learning algorithms are familiar with the system KPIs almost by definition. Hence, tracing the system’s actual performance — and the impact of the algorithm on the performance — can be an integral part of the system.
Staying Relevant in a Dynamic Market
The strategies one has to receive, understand, and implement in the market depends on the market dynamics. As an example, a pricing strategy for a product with a mild amount of competition will obviously have to change if an aggressive competitor joins the competition, or if Amazon has modified their Buy Box winning algorithm.
While a rule-based system could in principle refer to several scenarios, it is impossible to cover all the relevant options. Moreover, changes in the marketplace may make the applied strategy become completely irrelevant.
Machine-learning algorithms, on the other hand, are inherently adaptive: when the game rules change, the algorithm will simply learn a new strategy which enables it to adapt to the new situation.
Even the most routine decisions we make as humans are based on a surprisingly complex set of signals and relations. As an example, when we walk, each step involves feedback from our nervous system, as well as visual and sensory inputs. If our brain limited its response to a set of orders to our right and left legs, we would probably stumble and hurt ourselves on a daily basis.
The training process of machine-learning algorithms is similar: as the algorithms converge to become a strategy, they examine the situations in which the strategy fails, look for indications that could have predicted the failure, and modify the strategy accordingly. By repeating the process, the algorithms learn how to protect themselves from failures, reducing the chances of a real-time catastrophe. In the next post in this series, we will discuss how to properly train a machine-learning algorithm to help our clients achieve their business goals.