Beyond data driven
Posted: Sun Jan 19, 2025 3:20 am
Unlike Shapley-based models, Markov chain-based models do not take into account the effect of eliminating one of the players from the game to establish what their contribution was, but rather, as their name indicates, they analyze the complete chains that end in conversion and the impact that the fact that the user has been impacted has on the final conversion so that the final path ends in conversion.
One way to understand this is to think of it in terms of states . When the user initiates a first impact via email, for example, he enters the email state, which generates its corresponding chain. Models based on Markov chains focus more on complete chains than on the incidence of each individual impact. One advantage of this is that it is easier to scale than models based on Shapley. On the other hand, Shapley is easier to implement, especially in models with few channels, which do not suffer from scalability problems.
The big problem with data-driven models based on algorithms such as Shapley or Markov is that they tend to be black boxes. We have to make a leap of faith at the level of the particular implementation in each case and we cannot know exactly how it has been implemented.
As a result of this and the popularization of machine learning, bosnia and herzegovina business email database we are increasingly finding cases of models tailored to each client. Instead of applying a pre-built model with rules hidden from us, we work with raw data and build an algorithm from scratch, 100% developed according to the particular case of each client.
Another factor that has led us at Labelium to work with our own models is the possibility of incorporating user behavior into the equation. We no longer only take into account the origin of each visit, but also what actions the user has performed during each visit. We can thus be much more precise when assessing the exact contribution of each channel, campaign and visit. Additionally, this way of understanding the problem also has great advantages for identifying fraud.
Broadly speaking, a behaviour-based model estimates the conversion probability of a user's specific session at the beginning of the session. It measures all the actions that the user performs during that session and recalculates the conversion probability again at the end of the session. This delta difference between the conversion probability at the beginning of the session and at the end of the session will be what is attributed to the channel/impact that brought the visit. Once all the sessions of all the users are accounted for, we will have a complete attribution map, with the contribution of each channel and campaign.
One way to understand this is to think of it in terms of states . When the user initiates a first impact via email, for example, he enters the email state, which generates its corresponding chain. Models based on Markov chains focus more on complete chains than on the incidence of each individual impact. One advantage of this is that it is easier to scale than models based on Shapley. On the other hand, Shapley is easier to implement, especially in models with few channels, which do not suffer from scalability problems.
The big problem with data-driven models based on algorithms such as Shapley or Markov is that they tend to be black boxes. We have to make a leap of faith at the level of the particular implementation in each case and we cannot know exactly how it has been implemented.
As a result of this and the popularization of machine learning, bosnia and herzegovina business email database we are increasingly finding cases of models tailored to each client. Instead of applying a pre-built model with rules hidden from us, we work with raw data and build an algorithm from scratch, 100% developed according to the particular case of each client.
Another factor that has led us at Labelium to work with our own models is the possibility of incorporating user behavior into the equation. We no longer only take into account the origin of each visit, but also what actions the user has performed during each visit. We can thus be much more precise when assessing the exact contribution of each channel, campaign and visit. Additionally, this way of understanding the problem also has great advantages for identifying fraud.
Broadly speaking, a behaviour-based model estimates the conversion probability of a user's specific session at the beginning of the session. It measures all the actions that the user performs during that session and recalculates the conversion probability again at the end of the session. This delta difference between the conversion probability at the beginning of the session and at the end of the session will be what is attributed to the channel/impact that brought the visit. Once all the sessions of all the users are accounted for, we will have a complete attribution map, with the contribution of each channel and campaign.