Company Description
Smart Host (https://www.smart-host.com) was founded in 2017 and is now one of the leading CRM systems for hotels in Europe. Based in Berlin, Germany we provide a SaaS solution to help hotels maximise their revenue and at the same time become better hosts by gathering valuable information about their guests and their individual interests. Currently we serve more than 250 hotels in 4 countries.
Our core product is a campaign system which allows hotels to target guests who already had touch-points with the hotel and may or may not have spent their holidays there. These touch points include marketing campaigns, chats, surveys, reviews and individual offers. With us a hotel can easily convert one-time guests into returning, loyal guests.
What differentiates us from our competition is that we do not only offer a campaign toolkit for sending out email campaigns and tracking their success but we also provide meaningful campaign suggestions with pre-calculated audiences and start dates. While our campaigns are currently limited to email we will add other channels like Whatsapp in the near future.
Project Description
Our current campaign targeting is based on the assumption that people should be addressed at the right time with the right offer. In order to achieve that we are creating cohorts and use statistical methods to create campaign plans based on past guest behaviour. While this works well so far, we would like to take our campaign targeting to the next level by not addressing cohorts but individual guests at the perfect time for each guest in order to maximise conversions for our hotels.
The aim of this project is to develop a framework that starts by process mining models reflecting the behaviour of different cohorts, and then applies novel ML models that utilize the process models to predict when to send a campaign / offer to a particular guest based on the available data and existing touch-points of this guest. Solutions from existing customer journey optimization applications that utilise process mining like [1, 2, 3, 4, 5] can be used as a starting point for end-to-end solutions. Hereby the main success indicator (or key performance indicator; KPI) should be the conversion rate, i.e. the addressed guests makes a request or a booking.
Data Description
In order to train prediction models like the mentioned-above one you can make use of the anonymised data of more than 250 individual 4 to 5 star leisure hotels mainly based in central Europe. For training about 18 million guest records, 12 million reservation records and more than 30 million sent campaign emails can be used. Some of our clients have reservation data dating back more than 10 years.
You will start from an existing proof of concept pipeline using Google’s temporal fusion transformer for time series forecasting [6]. From what we saw, this pipeline produces promising booking forecasts, even without much optimisation. This only serves as a starting point for the project since the question of better suited prediction models as well as the comparison between such models for booking data remains unanswered so far.
Deployment
As a very important part of this project we aim to test developed models in the real world. Together with one or more of our clients as project participants we would like to start campaigns which behave according to the optimized models, ideally applying an A/B testing for highlighting the effectiveness of the designed framework. Based on the project’s outcome we aim to implement the newly developed ML targeting approach into our application, making it available for all of our clients.
There are no requirements to be physically present in Berlin, as our company is more than 50% remote anyways. For kick-off and end-of-project evaluations, in-person-meetings will be beneficial though.
Contact
For more information, contact dr. ing. Marwan Hassani.
References
- Alessandro Terragni, Marwan Hassani: Optimizing customer journey using process mining and sequence-aware recommendation. SAC 2019: 57-65 (https://dl.acm.org/doi/10.1145/3297280.3297288)
- Marwan Hassani, Stefan Habets: Predicting Next Touch Point In A Customer Journey: A Use Case In Telecommunication. ECMS 2021: 48-54 (https://www.scs-europe.net/dlib/2021/2021-0048.htm)
- Yorick Spenrath, Marwan Hassani, Boudewijn F. van Dongen: Online Prediction of Aggregated Retailer Consumer Behaviour. ICPM Workshops 2021: 211-223 (https://link.springer.com/chapter/10.1007/978-3-030-98581-3_16)
- Sophie van den Berg, Marwan Hassani: On Inferring a Meaningful Similarity Metric for Customer Behaviour. ECML/PKDD (5) 2021: 234-250 (https://link.springer.com/chapter/10.1007/978-3-030-86517-7_15)
- Yorick Spenrath, Marwan Hassani, Boudewijn F. van Dongen, Haseeb Tariq: Why Did My Consumer Shop? Learning an Efficient Distance Metric for Retailer Transaction Data.ECML/PKDD (5) 2020: 323-338 (https://link.springer.com/chapter/10.1007/978-3-030-67670-4_20)
- Bryan Lim, Sercan Ö. Arık, Nicolas Loeff, Tomas Pfister Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting. International Journal of Forecasting, Volume 37, Issue 4, 2021, Pages 1748-1764 (https://doi.org/10.1016/j.ijforecast.2021.03.012)