보름달 🌝 #2.2 Machine Learning - Analgesic Recommendation (ENG)

1. Introduction

    The easiest way to relieve menstrual pain is to take analgesic, especially over-the-counter general pharmaceuticals. Though hundred kinds of painkillers are available on the market, consumers habitually choose familiar ones because of lack of information.

    Of course, the basic effects are the same since they are all alagesic. However, the symptoms are various, and the effects also vary depending on the ingredients of the painkiller. Proper dosage will help the user to relieve pain quickly. We aim to recommend appropriate kind of painkiller in accordance with symptoms using symptoms that user put in the calendar detail page.

2. Datasets

   Since there is no accessible public data set, we made fake dataset with 1521 cases after researching for the relationship between symptoms and painkillers.

    Analgesic is divided into 2 broad groups, acetaminophen and nonsteroidal. Acetaminophen family can be a burden on the liver, and the ibuprofen, aspirin and naphroxen in the nonsteroid family can be a burden on the stomach. Referring to 「알고 먹는 약 모르고 먹는 약, 김정환」 and Mint Hospital Blog, we established rules to recommend painkillers that prioritize user's liver and stomach conditions.

    We classified Pill Numbers considering back pain, heartburn, severe pain, menstrual irregularity, swelling, mild fever, headache, abdominal pain, convulsions, psychological symptoms, diarrhea, liver condition, and stomach condition. In the symptoms stored in the databse entered through mobile application and nugu speakers are in the boolean format. Since it is the fake dataset, pill recommendation is stored in Pill Number without feature engineering process.

Pill Number	Pill Recommendation
1	스피드펜 나노, 자이날, 이지엔
2	이지엔6 이브
3	그날엔
4	이즈펜
5	우먼스 타이레놀
6	이지엔6 이브, 이브큐 레이디
7	싸이베린
8	게보린, 펜잘큐, 사리돈 에이
9	탁센
10	부스코판 플러스
11	그날엔 Q
12	자이날

3. Methodology

Multiclass Classification Algorithm
One of Softmax Regression which is for choosing 1 out of more than 3 classes.

One-Hot Encoding
Encoding which represents True for 1 feature and False for the last. For example, 0 is encoded in [1,0,0,0,0] and 4 is [0,0,0,0,1]. It is done to Apply various models, inprove accuracy, and for ease of operation in classification problem.

    What we should do in machine learing is to predict 12 Pill Number using 14 features. So we judged the multiclass classification algorithm would be right. Also, we did one-hot encoding to change Pill Number which is from 1 to 12 to 0 to 11.

    We used tensorflow.js for the machine learning. We divided dataset 70:30 traing set and test set, trained the model and tested.

4. Evaluation & Analysis

    Since we've done prior investigation to make only related features into the dataset, we decided not to test what features would affect the result. We did training assuming that all features are important.

   The graph above shows the training performance of the multiclass classification model using tensorflow.

    In the tensorflow, epoch is the number of training, batch size is the size of data at one training, and iteration is the number of traing batch at one epoch. Since our data set is limitied, one epoch is not enough. However, when the epoch is too much, there could be an overfitting. So we draw the graph to show the loss value in accordance with epoch and batch size.

    We achieved a loss value around 0,03 on both training and testing, using 400 epochs and 32 a batch size of 32

5. Related Work

• tfjs-vis: A library to show how tensorflow work on the browser. We used tfjs-vis to draw graph at 4. vealuation & Analysis.
• Machine Learning in JavaScript with Tensorflow.js

6. Conclusion

    We were able to get the results we aimed for. Full Moon Service is designed to help users to recommend painkillers using the information entered. Users can expect faster effects by taking them. If there's any missing information, NUGU can get it through the slot filling function.
   One thing that's unfortunate is that the current data set is a random situation. We will be able to progress as collecting the real user's information and feedback.