All you need to know about automated, scalable brand protection technology, and why it has become a key service for brands looking to optimise their online sales strategies.
The Internet is a safer place today than it was 20 years ago thanks to advances in security and regulation. This has helped it to become a vital tool in all areas of our personal and professional lives.
It is this ubiquity that has made it a very attractive and lucrative environment for counterfeiting through the illicit use of intellectual property to profit from the sale of fake products. This inflicts great damage to the brands that are targeted, not only to their sales, but also their brand image can be seriously impacted. It is estimated that around 80% of brands experience this situation, making it important to detect counterfeit products and the introduction of brand protection a necessity.
Various forms of brand protection technology have been developed to give visibility of the digital ecosystem, and in turn the ability to detect counterfeit products and eliminate such infringements.
Counterfeiting and brand protection on the internet may seem like easy problems to solve at first. For a human, it could be as simple as searching for the name of the brand or product in a search engine such as Google and reviewing a few links to see if they are infringing based on certain factors which might include price, country of seller, or stock. Once discovered the next step would be to submit a deindexation request to the search engine in question.
This brand protection process may seem simple and intuitive, however, it requires a lot of time and resources if it is to be carried out with adequate quality, coverage and regularity. That entails the constant monitoring of the various distribution channels where a product may be sold to ensure that infringements are detected and taken down quickly and efficiently. With this in mind, it is clearly not a scalable job for a human if we want to see positive results.
However, thanks to the technological advances that have occurred over recent years in information technology such as high-level programming languages, communication interfaces between services, big data and artificial intelligence (AI), this brand protection process can be carried out by a machine under human supervision. Specifically, we are talking about tools that can help the aforementioned processes which include:
These processes based on these technologies can be configured and executed recurrently and automatically for each of the brands or products in our portfolio, requiring the presence of a human only in the final stages of the process in order to verify the actions taken by the machine. Therefore, it guarantees that the most tedious and routine tasks of the process are executed in an automated, flexible, and recurring manner, and the human can focus on providing the greatest possible value to the brand protection process.
At Smart Protection we have a platform that executes this process in a manner similar to that described above, which allows us to offer a quality and efficient anti counterfeiting and brand protection service to a wide range of clients from different business verticals.
Over the coming months we will release a series of articles in which we will go into detail about each of the previously described technologies and we will see how you can detect counterfeit products through automated online brand protection.
In this first technical post of the series we aim to provide a cross-sectional view of the type of problems that can be solved using Machine Learning techniques and associate some of these with what online brand protection entails today.
Machine Learning (ML) is a discipline of Artificial Intelligence (AI) that allows learning the behaviour of a process from examples in which that process has been carried out (dataset) and roughly characterises this process in a model mathematician (algorithm). In other words, an algorithm observes the data and "learns" a behaviour that explains that data, in such a way that in the face of new events the algorithm can apply the learned logic to estimate its result.
Depending on the information that we provide in the data for the model to learn, we distinguish two types of approach to solving the problem:
As anticipated in the previous paragraph, in supervised learning problems, the final result (commonly called target or label) is provided within the training data for each of the observations fed into the model. Why? Because during the training process (learning) what is tried to minimise is the difference between the model prediction (y_pred) and the real data (y).
Depending on the type of objective that we seek to predict, we are faced with two main types of problems: Classification and regression.
We speak of classification when the objective is to assign each observation a category from among a limited number of mutually exclusive possibilities.
In the example that we illustrated earlier, the objective was to classify if it rains on a day (class 1) or if it does not rain (class 2). We would also be talking about classification if we try to predict if one day the sky will be sunny, partly cloudy, very cloudy, or raining.
Within classification, we can distinguish between binary classification (we only have two classes to predict) or multiclass classification (we have more than two classes to predict).
However, when the objective to estimate for each observation is a numerical value, we speak of a regression problem.
An example of regression, once again applied to time, would be the prediction of the litres of rainfall per square metre in Madrid on a specific day, in which we no longer expect to know if it is going to rain or not, but rather if it is going to rain. to drop 8.2L/m2 (a lot) or 0.23L/m2 (a little).
Unlike supervised learning, in unsupervised learning problems the final value of the result is not available for each observation. The most typical type of unsupervised learning problem is clustering.
In a clustering problem, the objective is to identify a number (known or unknown) of groups (clusters) in which the observations are concentrated based on common characteristics. When the clusters converge (they are grouped) it is possible, for example, to identify in a less biassed way potential categories to use for a later classification problem.
In the example that we highlighted earlier, the identification of topics that are covered in a series of news would be an example of clustering, in which the most similar news would be grouped to form the topics.
So, what kind of problems within the field of brand protection can be addressed using Machine Learning technology? Here are some examples:
At Smart Protection we use Machine Learning techniques to detect counterfeit products, monitor online sales, and provide the most effective brand protection to our clients. Our goal is clear: to satisfy customer needs by providing the highest degree of automation.
Co-authors: Miguel Saenz and Alberto Polidura (Smart Protection Data Science Team)