Location, Location, Location.   It may sound crazy that in today’s digitized world, physical location still matters a great deal to organizations. Think about any real estate business (hotels, restaurants, movie theaters, etc., you name it!)—finding a good location still remains somebody’s full-time job.   DID YOU KNOW? An array of surveys (from HubSpot, TripAdvisor, etc.) has revealed that: For restaurants, 1 in 3 customers won’t eat at a restaurant rated only 3 stars. For hospitality, 32% eliminated from consideration those properties with a rating below 4 stars. A one-star increase on Yelp will lead to a 9% increase in restaurant revenue. 96% of travelers consider reviews important when researching a hotel. As you can see, customers leave their interaction footprint in all shapes and forms across the media. And yes, this is how your customers behave today!   Facing the increasing number of reviews online, it seems inevitable that businesses should consider customer reviews when making any location decisions. However, the question is: HOW should they do so? Text data is unstructured, massive and messy. Collecting, crunching and extracting meaningful insights from text data, then embedding this information into location models, is not an easy task. The good news: it is possible.   Challenge: Strategically locate a new hotel   During the AI Hackathon week in which SAS’ Customer Advisory Academy 5 participated, our group's task was to create a feasible and scalable solution for businesses using SAS’ Artificial Intelligence (AI) capabilities, including text analytics, network analytics, computer vision, recommendation engine, and optimization. Inspired by the potential of text analytics, we decided to take on the challenge. To make our solution more relevant for tangible business value, we picked a fictitious hotel chain as our target client. The business problem: Find a strategic location in the US for the hotel chain to put a brand-new hotel. The hotel’s management also wanted to leverage the growing segment of young, middle-class customers in this expansion.   Our approach: Text analytics with online reviews   To approach this business problem, our group identified that the most important pain point for the client was the need to secure its first-mover advantage. Identifying uptrend locations ahead of the hype was essential to the client’s success. As mentioned above, reviews strongly influence service providers. Therefore, we incorporated review data into our analysis and analyzed how sentiments expressed online affect strategic location decisions. Given the time limit (24 hours), we limited our scope to two locations: Raleigh and Portland. As a result, we ended up with 2,297 reviews across different entertainment types (restaurants, bars and pubs, shopping malls, gyms, theme parks, etc.), which we felt would be appealing to the client’s target customers (young, middle-class Millennials).   We started off by using SAS® Visual Text Analytics (VTA) to identify concepts and topics Millennials most care about and discuss online. Next, we scored customer sentiments on the scale of 1 to 3(Positive, Negative and Neutral). We also used VTA to identify common topics among customer reviews. Below are a few snapshots of our quick VTA effort.   Concept Node: extracts predefined concepts or create additional custom concepts that you can discover in a document or set of documents. Visual Text Analytics model pipeline:SAS Visual Text Analytics provides a number of text analysis pipeline nodes, arranged in a sequence that you control. Diagnostic Metrics for Automatically Generated Categories: shows the F-Measure, Precision, and Recall values for each automatically generated category. After calculating sentiment scores for customer reviews both in Portland and Raleigh, we surfaced the insights back to SAS® Visual Analytics to put things into perspective. Via SAS Visual Analytics, we were able to achieve a global overview of the trendiness of each area. Below is our overview dashboard.     The report was interactive and drill-able, allowing us to look deeper into each specific area of focus.       As shown in the graph, when we drilled down into Raleigh, shopping malls seemed to have lower sentiment scores compared to restaurants, fitness centers and cafes. From a business standpoint, putting a hotel near a shopping mall may not be ideal as customer sentiments toward these areas are likely to be more negative, which in turn may affect hotel reviews as a spill-over effect.   We then used distance data (latitude and longitude) to cluster reviews on a geographical scale. Next, we used sentiments to assign weights to the cluster centroids, which are the middle points of each cluster. These weights were calculated by multiplying each cluster’s sentiment score by the average star rating of the corresponding properties and the total number of reviews. We used these weights to adjust the centroids correspondingly. The below visualization shows how the centroids were shifted.     Two pieces of feedback that we received from the judging panel during the AI Hackathon were: 1) Sentiment scores and average star ratings might be correlated, and 2) It might be simpler to use sentiments directly in the clustering model, rather than the weights. We believe that these were excellent ideas, and we would love to explore them more to improve our approach. One possible alternative we may explore is to use sentiment score directly in the clustering model, which would allow us to easily measure the effect of inserting the sentiment scores into a traditional geo-clustering algorithm. The good news is that this can be done smoothly and effortlessly within SAS® Visual Analytics.   So, why don’t you give it a try and let us know if you find something interesting?   Authors   Collaborators with @yennguyen_sas on this app and article include: @jmowlai, @TylerGillson, and @XinRu.   ... View more