User behavior analysis, modelling and prediction by text classification and multimodal dataset from job ads dataset.
- Objective 1: To predict the behavior of user based on the text description and location for multimodal model.
- Objective 2: To create a recommender system based on the text desciption and job classification.
- 2 sources of dataset
- ads-50k-events.csv
- ads-50k.json
- Model deployed: Model 1.1 (Bi-LSTM, 20 epochs)
| Model idx | Model name | Features col | Target col | Epochs | Accuracy (%) |
|---|---|---|---|---|---|
| 0 | Naive-Bayes | title,abstract,content | kind | - | 56.39 |
| 1 | Bi-LSTM | title,abstract,content | kind | 10 | 56.88 |
| 1.1 | Bi-LSTM (multimodal) | title,abstract,content,location | kind | 10 | 57.39 |
| 1.1 | Bi-LSTM (multimodal) | title,abstract,content,location | kind | 20 | 57.81 |
| 2 | Bi-LSTM | title,abstract,content | classification | 10 | 8.50 |
- Deep learning modelling sucessfully exceed the accuracy of base model.
- Hypotethically, by increasing the epochs of training, the model accuracy will improve too.
- However, there is a limitation in terms of computation power due to model is trained by using Google Colab free GPU (Tesla T4). Kernel has down for some reasons.
- Possible further experimentation:
- Case 1:
- Training for 100 epochs
- Using pre-trained embedding (USE) and Huggingface transformer model.
- Using different columns for multi-modal model.
- Using useful callbacks such as ReduceLROnPlateau and EarlyStopping.
- Case 2:
- Building similarity score algorithm based on embedding vector.
- Case 1: