Factorization Machines from scratch
Real time bidding is a field that concerns itself with the way of publishers selling their ad impressions to the advertisers.
One of the specific features of real time bidding (RTB) is that it is highly automated. RTB is run via programmatic auctions which determine who wins an impression on specific page.
This is all happening in the background while the users are visiting websites.
The way ads are served on many websites today roughly follows the following steps:
1) when a user visits the site of a publishes, the latter then sends a bid request with additional metadata (about visiting user, the ad, etc.) to hundreds of thousands of potential advertisers than are able to bid on this impression by using so-called ad exchanges,
2) in the next step each of these hundreds of thousands of advertisers quantitatively evaluates ad opportunity based on received information, by using complex mostly machine learning models,
3) a bid price is then submitted to the auction and the ad exchange then select the advertiser where the winner of the auction commonly pays the second highest bid.
The whole bidding process outlined above is usually done in less than 100 ms. The ad from the advertiser that won is then displayed on the publishers website, resulting in potential user response e.g. click on ad.
So-called Demand Side Platforms are most often bidding on behalf of advertisers on these ad exchanges. Typical paying models for publishers are cost per impressions (CPM), cost-per-click (CPC) and cost-per-action (CPA) for conversions.
To predict CTRs, RTB industry employs many different models, mostly from machine learning field. They can include sentiment analysis API, text classification, automated product categorization API and others. One of the more advanced models are so-called Factorization Machines (FM) and I will explore them in more detail in this series of articles.
Although efficient libraries are available for factorization machines such as libFM, let us start from scratch by following the paper:
https://www.csie.ntu.edu.tw/~b97053/paper/Factorization%20Machines%20with%20libFM.pdf
CTR-prediction using SGD and hashing trick
In second part of our article we will concern ourselves with specific problem in RTB -> CTR prediction.
One mostly uses machine learning models for this task, which enable us to predict whether an ad will be clicked by the user or not.
A great data set to experiment this on is Avazu data set which consists of around 40 million rows:
https://www.kaggle.com/c/avazu-ctr-prediction.
This machine learning problem should be considered as an out-of-core problem, which means that we are not going to load the entire training data set into RAM.
What we usually do instead is thus take each row at a time, process it and then input it to ML training, which updates the weights of ML model one row at a time.
A great algorithm to achieve this is SGD - Stochastic Gradient Descent.
Stochastic Gradient Descent
Stochastic Gradient Descent method is a great general method to determine optimal weights of a machine learning model.
Scikit-learn is an excellent choice for implementation of SGD, we will be using SGDClassifier class.
SGDClassifier has a complexity of O(knp), where k is the number of epochs, p is the average number of non-zero features per sample and n is the number of data instances. This means that this model's time complexity increases linearly with the size of training data, which is favorable. It also has a fixed RAM size, which makes it easier to scale.
We have used out-of-core learning SGD for many projects, it is very efficient in the setting of AI lead generation.
One of the tricks of using this class but in the out-of-core setting is to not use method fit, but the method partial_fit.
However, if we want to use it, we must check the whole dataset once because we need to determine the classes that we want to target for prediction.
They need to be passed to the method partial_fit.

URL classification API for online stores

Another interesting application of machine learning models is for building URL classification API that can be then used to categorize online stores.
URL classification API is usually built by first gathering enough training data in form of texts and their assigned classes. Then a machine learning model is trained on this data. Machine learning models for text classifications can be many: logistic regression, support vector machines, convolutional neural networks, recurrent neural networks.
Once the machine learning model is built one needs to deploy it. A useful service for this is e.g. flask: https://flask.palletsprojects.com/en/2.0.x/
But there are many other frameworks for deploying machine learning models in production.

Top shopify stores - use case

There are many possible uses of such URL classification API. One is e.g. to categorize all the shopify stores and then offer a platform for finding top shopify stores in united states.
This helps owners of online stores to find their top competitors. For suppliers to shopify stores, the platform can be used to find shopify stores that sell similar products as their existing clients.
STEM education Machine Learning Consulting company
Math and physics education has become important in recent years, as the AI and data science in general has gained in importance for the society. A lot of people are thus investing a lot of time in improving their natural science skills. In Germany, there has been consequently a rise in popularity of Physik Nachhilfe online, e.g. in München.
In the last few years, machine learning consultancy has also become a highly sought after service, with ever increasing amount of companies deciding to step in and introduce AI solution to their products and services.
Machine Learning consulting company usually works on many different and overlapping field, from general ones, like machine learning, deep learning to specialized ones like computer vision, which is focused on dealing with images to natural language processing which is focused on text analysis and text models.

Items