Performance Analysis of Machine Learning Algorithms on Multi-Touch Attribution Model

Business owners are growingly making use of digital channels to advertise on their products / services. Most sellers advertise on multiple channels and consumers view these before making a buying decision. Multi-touch attribution (MTA) is an advertising measuring technique that scores the value of each touch point (viewing an advertisement) leading to conversion (sale of the product).We used two models to solve two different challenges in this research. The first model is the bi-directional LSTM attention model which assigns a weight to each channel based on how much money a company will spend on advertising. According to the Attention model, channel 1 accounts for 45 percent of conversions, channel 2 accounts for 20%, and channel 3 accounts for approximately 35% which is closer accuracy as per the given data. The second model uses a combination of machine learning and deep learning techniques to predict whether or not an advertisement sequence will be converted. The score is evaluated based on the sequence of touches leading a sale conversion. We observe that Random Forest algorithm, at 99.01%, performs best for our dataset. Additionally, we observe that conventional algorithms such as Decision Tree, Logistic regression, SVM perform better than LSTM with attention modeling.