Based on the assumption that models often fail to produce predictions when dealing with complex phenomena, we identify recurring patterns in the data – independent of any particular model – that occur locally. We have developed a predictive bump hunting algorithm for phenomena that are by nature undetectable using a global model, and more importantly difficult to interpret without taking vast amounts of variables into consideration. Beyond its mathematical advantages, the methodology we propose is designed to allow many different applications, in particular industrial. Our goal is to accelerate the understanding of complex issues by simulating the acquisition of experience through computing. We differentiate ourselves through our strong methodological and technological choices.
Bump Hunting Optimization
Searching for “bumps” in data – areas with unusual local densities within multidimensional data spaces – is a simple concept, but its combinatorial complexity can be explosive. The time and hardware resources needed to perform the calculations required to detect these signals quickly become unmanageable once the databases grow to more than a few thousand items of data. Our challenge is to offer an excellent ratio between strategic added value and the cost for obtaining new knowledge – in addition to the information obtained through more conventional Tools.
Association rule learning, which is designed to predict specific processes for a random variable by exploring existing data, runs into execution constraints that quickly become unmanageable on standard computer platforms, even if clustered, because the combinatorial analysis is so explosive. Our innovation is to have implemented the core of our algorithm on reprogrammable logic circuits such as FPGA, while obtaining performances that are far superior to those provided by current microprocessors – at much lower cost and with much less energy consumption.
Any sufficiently advanced technology is indistinguishable from magic.