Full report and the presentation slides along with the codes can be accessed here.
- Customer reviews not only help buyers to make better purchase decision but also enable companies to improve their products.
- Companies need to find an automated way to understand customers' sentiments based on the customer reviews
- Companies would also like to utlize customer reviews to know what product features are good and what need improvement
- A LSTM network to predict customers' attitudes based on the customer reviews
- An algorithm to extract product features and customers' opinions on those features
Here I included results of Samsung and LG for illustration purpose. Results for other brands can be accessed here Feature_Extraction_Results
-
Feature extraction results of Samsung products
-
Feature extraction results of LG products
-
Predictive Model Results
| Model | Accuracy | Precision (Weighted) | Recall (Weighted) | F1 (Weighted) |
|---|---|---|---|---|
| Random Forest | 0.815 | 0.838 | 0.815 | 0.822 |
| KNN | 0.862 | 0.846 | 0.852 | 0.847 |
| Naive Bayes | 0.674 | 0.628 | 0.674 | 0.610 |
| LSTM | 0.877 | 0.895 | 0.852 | 0.874 |
Customer reviews not only help buyers to make better purchase decision but also enable companies to improve their products. In this project, we plan to train a LSTM network on the Amazon Reviews: Unlocked Mobile Phone from Kaggle data set to classify the correct customer sentiment from given product reviews. Besides, we plan to perform characteristic extraction on the given reviews and summarize the customers’ opinions on those characteristics and give business insights to the companies. The data set we used in this project contains 413,840 data entries and six features including product name, rating, brand name, price, reviews and review votes. Through our analysis, we found out that customer satisfaction toward the product is positively related to the product price and negatively related to the length of the review. To be more specific, more expensive phones tend to get higher ratings, and more satisfied customers tend to leave shorter comments. We trained three models as our baselines, which are Random Forest, K-Nearest Neighbors (KNN), and Naive Bayes. Both the Random Forest and KNN models produced good results, which obtained weighted-F1 scores of 0.822 and 0.847 respectively. Naive Bayes performed worse than the previous two models which obtain an weighted-F1 score of 0.610.