**1. Data Info **
We were provided a year SNS crawling data which were mainly about vitamin reviews SNS involved Facebook, Instagram, Naver Blog, Naver Cafe, Daum Blog, Daum Cafe, and Youtube
Data Count per Brands
- Aronamin Gold : 3,238 | Aronamin C+ : 1,118건 | Urusa : 12,330 | Impactamin : 3,099 | Centrum : 11,359
- Vitamins : 29,736, Nutrient Supplements : 56,959
Data Sample
| SNS | Title | Contents | Address | Date |
|---|---|---|---|---|
| naver_blog | 찬찬약사.. | 아로나민 골드, 아로나민 씨플러스 많이 들어보셨죠?? ... | https://blog.naver.com/... | 2017#### |
2. Preprocessing
We removed Youtube because the data only contained URL and title of video, which are insufficient to analyze solid reviews.
We put the most priority on the subjectivity of reviews. We classified reviews based on subjectivity of contents.
1) Official Advertisement
Contents that the compnay voluntarily advertise their products through Facebook or official blogs
2) Unofficial Advertisement
Contents that certain customer or middleman to advertise
Removed ambiguous ones by visiting direct URL and reviewing real contents
ex) 'Hot-deal', 'Coupang', 'Group purchase'
3) Duplicates
Instagram Repost or Twitter Retweet
2-1. Simple analysis of brand impages after preprocessing We summarized the analysis below (in Korean) : https://docs.google.com/document/d/1j9fo8MiQO1yc5b-gLkDuwctuijfbbv4xHG0MZNV4y2g/edit?usp=sharing
3. Analysis
- Frequency analysis
1-1) Wordcloud: visualize based on simple token counts
1-2) TFIDF
- Association Rules
2-1) Apriori Association Rules
utilized _nims jupyter_ and _Comoran Tokenizer_ on preprocessed data to tokenize nouns held minimum threshold of _support_ as 0.05 held minimum thresholdl of _lift_ as 0.08
- Similarity Analysis
3-1) word2vec: word embedding
analyzed similarity between keywords of product name, important features (ex) side effects, smells) and advertisement actors
- Sentiment Analysis
4-1) KOSAC sentiment dictionary
utilized _Polarity_ dictionary of KOSAC sentiment dictionary
4. Results
4-1. Identify brand images and brand awareness
produced price score and image score
if several brands are correlated each other, we grouped them and analyzed differences
4-2. Identify customers' reactions on past advertisements
extracted the data that are related to advertisements
conducted similarity analysis by using word2vec on several keywords
4-3. Strategy of customization
Categorized prospective customer into 6 groups (ex) Kids, Pregnant women, Students...)
There were differences of highly-related nutrients between groups, which are important factors when selecting products.
Also, we examined familiar keywords of each group such as multivitamin and general vitamins
Based on these results, we suggested marketing strategies to a target company