The Restaurant Recommender

February 10, 2017

340 views
149 downloads


Report Abuse
This experiment demonstrates how to build a recommender system in AML to recommend the top three favorable restaurants to a customer.
In this experiment we are going to build a recommender system that recommend a customer three restaurant based on his eating habits. We chose to recommend top three restaurants to any given customer. The top three restaurants number could be easily modified by modifying the **Maximum number of items to recommend to a user**. We optimized the performance from of this model by modifying the "Fraction of training only users" and "Fraction of test user ratings for training" parameters to get the best recommendation to a customer. Below table shows various parameter values and the performance of the model for the training data set we used. ![model performance][1] ![overall experiment][2] ### Join To Show Restaurant Names The purpose of the "Join Data" modules in this experiment is to show a cleaned up output by providing restaurant names recommendations. This section is optional, The experiment shows the restaurant names rather than ids. ![Model Joins][3] ### Three Restaurant Recommendation Output The model output shows restaurant recommendations for **29** users. You can view this by clicking on "Visualize" option on the matchbox recommender module. ![Recommender Output][4] You can notice that some users (only three users) the model was not able to recommend three restaurants due to their eating habits didn't match at least two other users in the data set. To view restaurant names rather than ids, click on "Visualize" option from "Select Columns" module that provide cleaned up output. ![Recomender Final Output][5] You will notice that the cleaned up output contains recommendation for **24** users only due to the "inner join" modules we used to present the restaurant names for the top three recommendations. ### Recommender Performance The best performance we achieved for this model is **0.972299** after the tweaking work in the model to get the best recommendation out of this public restaurant ratings data set. ![Recommender Model Performance][6] [1]: https://github.com/melzoghbi/DataCamp/raw/master/Images/RestRecommender/NDCGs.PNG [2]: https://github.com/melzoghbi/DataCamp/raw/master/Images/RestRecommender/OverviewExp.PNG [3]: https://github.com/melzoghbi/DataCamp/raw/master/Images/RestRecommender/JoinRestNames.PNG [4]: https://github.com/melzoghbi/DataCamp/raw/master/Images/RestRecommender/RecommenderOutput.PNG [5]: https://github.com/melzoghbi/DataCamp/raw/master/Images/RestRecommender/TopThreeRestOutput.PNG [6]: https://github.com/melzoghbi/DataCamp/raw/master/Images/RestRecommender/RecPerf.PNG