An increasing amount of people book their travels today online. Whether it is a hotel, plane ticket or rental car, internet bookings have skyrocketed the past years. This is facilitated by websites who package offers and answer your search queries of your perfect journey. This new way of booking travels has paved way for research into drivers of click and bookings to optimise search query results and therefore conversion.
A website in the travel industry did not have thorough analysis in place what drove their customer’s needs in terms of search queries and resulting clicks and bookings. The website’s data consisted of customer data on Amazon AWS, where different customer traits were documented, such as amount of rooms, time of booking, destination, prior website visite and competitor availability.
By leveraging the data available, we developed a search algorithm, that ensures the highest relevance per customer based on a search query and available user data. This led us by designing an algorithm, that would optimize the relevance score per search query, trained on a dataset of 800 000 different search queries, and indication whether that led to a click or a booking on the booking page. We used LambdaMART, which consists of a combination of the LambdaRank algorithm, and MART, which stands for Multiple Additive Regression Trees.
It leverages the power of gradient boosted regression trees and a ranking criterion to employ machine learning and generate impact for the travel website. The appropriate criterion on which the algorithm is trained, is called the NDCG, which stands for Normalized Discounted Cumulative Gain, and measures the ranking quality of the search query results by maximizing the relevance which is done by swapping the order of the results and learning from the change in relevance.
By tuning the LambdaMART search algorithm based on the customer data at hand, conversion-to-click raised by 11% and conversion-to-book by 14% on a 6 – month basis. Time from the initial analysis to final deployment, was approximately 3 months, so were very happy with the positive result for the travel website within such a short period! We can look back on a great co-operation, and are still involved with the team, since more data, means there are even more opportunities for the algorithm to generate better results, since more data means more value!
Sustainability is the future. Humanity needs to look for alternate resources, in terms of renewable energy. Therein, wind energy plays a key role. Impact Analytics optimized the position of the rotor relative to the wind direction by smart adaptation on weather conditions.
The wind direction itself is typically measured using a simple wind vane installed at the top of the wind turbine. Unfortunately, gushes of wind are often turbulent. As such the measured wind direction is often noisy and an attempt to align the rotor in reaction to any gush of wind coming sideways would render the wind turbine less efficient. Instead, observation-driven filters such as the one below are used to filter the wind direction carefully and determine the optimal position of the rotor to increase generated energy.
The fundamental challenge was to filter out noise (wind gushes) from wind data to find the true wind direction. The turbine is less efficient when it `sees’ the wind arriving at an angle. Therefore we used a function describing the relative effficiency of a wind turbine. Also a measure for the wind direction offset was set up, describing the discrepancy between the wind direction and the position the rotor. A zero offset is optimal and occurs when the rotor’s position is perfectly aligned with the incoming wind’s direction. The final model was dependant on estimates for last observations of wind direction and wind offset.
The final filter was capable of separating the noise component from the fundamental slow shift in wind direction. Using a smart observation-driven filter, the relative efficiency increased from an average of less than 89% to 94%. An average onshore wind turbine can produce more than 6 million kWh in a year – enough to supply 1,500 average EU households with electricity. By simply updating the software chip in the wind turbine, it is able to supply an additional 84 average EU households.