Draft:Artificial intelligence in the digital marketing

Currently, the processes associated with artificial intelligence (AI) have a vital impact on marketing and advertising on the internet, as they facilitate market research for brands. This has made it possible to deliver advertising information appropriate to the characteristics and interests of users, a process also known as behavioural targeting. AI, then, serves to study, define and segment users in order to create discourses and strategies that respond to the demands and attributes of their audience. To this end, AI also contemplates different forms of tracking and information collection such as tracking cookies or data capture behind different free platforms that feed their databases. Thus, depending on the quantity and quality of the information, and the objectives and timing of the companies, there is a certain type of AI machine learning that is optimal for the task at hand. However, the generation of "solutions" and results is a repetitive and relentless process as companies always seek to anticipate reality.

Data collection
Data collection is the set of processes by which it is possible to obtain the information to create or enrich a database. It can take place in an analogue or virtual manner. In the non-virtual world, people share their information when they, for example, fill out an application, register to vote, register a product for a warranty, purchase a driver's licence or enter a raffle. Sensitive data such as transactional data may also become known when people use their credit card or pay a bill with a cheque. Virtual information, on the other hand, is mainly favoured by the internet, where cookies record every click, browser searches and people's interaction on social networks are recorded, mobile phones record their owners if they say "Hello Siri" or "Ok Google", street cameras keep records in databases, and so on. Furthermore, in the future, when the internet is further developed and integrated into everyday life, the amount and type of information collected will be much more detailed.

The following is the list of categories of data that an organisation can collect:


 * Demographics: Name, Gender, Age, Race, Address, Phone, Fingerprint, Heart Rate, Weight, Device, Government ID.
 * History: Education, Career, Criminal Record, Press Exposure, Publications, Awards, Association Memberships, Credit Score, Legal Matters, Divorces, Travel, Loans.
 * Preferences: Settings, Idea Promotion, Political Party, Social Groups, Social Likes, Entertainment, Hobbies, News Feeds, Browser History, Brand Affinity.
 * Possessions: Income, Household, Automobiles, Devices, Clothing, Jewellery, Investments, Subscriptions, Collections, Social relationships.
 * Activities: Keystrokes, Gestures, Eye tracking, Part of the day, Location, IP address, Social posts, Eating out, Watching TV, Heart rate over time.
 * Personality: Religion, Values, Giving, Political Party, Scepticism / Altruism, Introvert / Extrovert, Generous / Greedy, Adaptable / Inflexible, Aggressive / Passive, Opinion, Mood.

Data hygiene
To feed the data into AI, it is necessary to know how it was collected, cleaned, sampled, aggregated, segmented and what transformation is required before combining it with other data streams. This process is paramount to ensure that the result of the analysis serves the desired purpose and can influence the outside world: for example finding the perfect headline to induce a group of people to buy. For this reason a data expert is recommended to decide which bits (information) should be included and which rectified.

See also Data Mining

Data characteristics
For a marketing expert, it is necessary to know the typology of the data with which artificial intelligence works. At this point, there are two key concepts: cardinality and dimensionality. The first refers to the uniqueness of the elements in a database column. For example, an email has high cardinality because it should be unique while living in "Paris" has low cardinality because more than one shares that characteristic. As far as dimensionality is concerned, it is recognised as the number of attributes obtained about an individual; when you have information about more than one individual you generate a database where each attribute becomes a dimension. AI is key in the treatment of multidimensional databases because through its artificial neural network it is possible to find statistically based connections and patterns. This multidimensional data can be mapped and studied by means of support vector machines that use algorithms to predict the category of a new piece of data.

See also Entity-relationship model

Types of learning in Artificial Intelligence
In general there are 3 types or levels of learning in artificial intelligence: supervised, unsupervised and reinforcement learning, of which, depending on the specific need of the company, one or the other may be applied.

Supervised learning
Main article Supervised learning

The idea is to teach the machine certain rules (training data) so that it creates a profile and recognises the results that meet those inputs. For example, if it is taught to identify cats through a set of images, it should be able to identify them on its own in the future. Or, if a brand has already defined its best user type, you can use those characteristics to locate all of them.

Unsupervised learning
Main article Unsupervised learning

The machine makes associations and draws new conclusions from the information it already contains: If it already identifies cats then it can study their context and recognise that they are found on chairs and sofas as a trend. Or, for example, it may find that the person who searched for "Sony DSC W830 20.1 Megapixel digital camera" after having searched for "digital camera", "digital camera reviews" and "wifi cameras" is 50% more likely to buy than the person who only searched for "digital camera", "digital camera reviews" and "digital cameras for sale".

Associations
By association rules, machines can infer, for example, whether a person is likely to buy something by the logic of "those who bought this also bought that". Data analysis by association has two key concepts: support and confidence. The former would refer to the number of times an item has appeared in the shopping bag and the latter relates the number of times two items have been purchased together. For example, if a person bought toothpaste 400 times and dental floss 300 times, and 300 times they bought the products together it means that the confidence is ¾ or 75%, but the association between the two is 100%.

Anomalies
Contrary to patterns are anomalies, to which special attention should be paid as they are unexpected changes that need to be explained in order to take action. For example, fraud can be detected if a purchase appears to have been made in a place that does not match the person's actual location. However, there may also be beneficial anomalies for marketing decisions such as being trending on Twitter and being able to leverage fame to induce a purchase.

Reinforcement learning
The machine, from its own learning process, generates outputs and conclusions that it tests in order to learn and improve. Reinforcement learning differs from supervised learning in that in supervised learning the human must indicate when the machine is wrong, whereas "reinforcement learning" is where the machine creates its own mental model of the world in which, for example, it decides which poster is the most impressive to a certain group of people.

How unsupervised learning works
This learning system is composed of neural networks that function like a brain where each neuron transmits information to other neurons to generate a result. Each artificial neuron has its limits because at the individual level it has certain inputs and outputs. However, if there is a situation with high support and high confidence, it sends a message to others, because given the situation, it makes them prone to disseminate the signal. Considering the situation of a trip to the cinema, the inputs would be the factors that can influence whether it happens or not: weather, effort, cost. These inputs are not binary but operate on a grey scale (as the feeling about the weather or the effort do not have a positive or negative answer; for this reason the outputs are a percentage (e.g. 65% probability of going to the cinema).

These many layers and factors create different decision layers that become deep learning. This learning combines many layers of information (e.g. level of education, probability of buying toothpaste and dental floss etc.) to enrich each neural unit and thus provide new outputs and conclusions. In this way, it is no longer the human who establishes these relationships, but from the data that the machine creates a learning process. The strengthening and robustness of this learning becomes what was previously called "reinforcement learning".

Google
Google, through its "Google display network" consisting of different websites (also called publishers) supports its service/programme named "Google ads". Here Google receives ads from advertisers and then selects the websites (publishers) associated with the ad depending on criteria such as the relevance of the content, the bid price and the revenue it would earn. Thus, in Google's targeted advertising model, publishers are used to track users as they browse the internet (via the DoubleClick cookie whose domain belongs to Google) and at the same time to profile users when they visit their pages. For example, if a user frequently visits a football website, they will be tagged in the "sport" category and in the "football" subcategory. This combined with the demographic information that Google has (such as age, gender, location) creates a user profile that will be used to display advertising (the behavioural targeting method). According to research, 88% of tags/categories that profile an individual (such as "football") receive targeted ads that are directly associated with the keywords that define them ("sports"). These defining tags/categories are updated in the range of 1 and 2 minutes and can be observed on the Google Ads preferences page.

Facebook
Facebook, like other platforms such as Amazon, makes use of the personalized marketing method where it uses: the history of pages visited, information collected by data brokerage firms (such as Experian, Acxiom and Epsilon that are dedicated to digital data profiling) and the data and interactions of users on the platform to profile and advertise to them according to their interests. Specifically, Facebook changed their targeted ad technology when in May 6, 2015 it partnered with IBM in order to give its users a more personalised and relevant experience. In practice, Facebook's personalised ads have made the platform an indispensable tool for advertisers (92% of marketers use it) as it is less expensive than other mediums and has a huge reach (at least 1.39 billion active users per month).

Concerns about the use of AI in Marketing
Business models such as Google's where personal information acquires a monetary value raise major concerns about users' privacy. Sensitive categories such as sexual orientation, health, religion and political ideology are being used to display targeted advertising even though in many places the use of such information is prohibited. According to research, between 10% and 40% of ads shown to people profiled with such sensitive conditions were ads that appealed to these characteristics.