Files for tweet embedding model:
- data.py - Loads data from files and prepares it for input to the model.
- data_helpers.py - Additional functionality for loading and preprocessing.
- evaluation.py - Functionality for evaluating model performance (can be run as a standalone program).
- model.py - Code for tweet embedding model framework.
- train.py - Trains a tweet embedding model.
Files for hashtag recommendation:
- feature_extraction.py - Code for generating text and image embeddings from a trained tweet embedding model for input to hashtag recommendation baseline (see below).
- hashtag_recommendation.py - Tests retrieval-based hashtag recommendation using a trained tweet embedding model.
Files for hashtag recommendation baseline:
- coa/co_attention.py - Trains and tests hashtag recommendation baseline.
- coa/data.py - Loads data from files and prepares it for input to the model.
- coa/selfDef.py - Code for hashtag recommendation baseline co-attention network and loss function.
To train tweet embedding model:
train.py [options]
To test retrieval-based hashtag recommendation:
train.py [options]
hashtag_recommendation.py
To test hashtag recommendation baseline:
coa/co_attention.py
To test hashtag recommendation baseline with image and text vectors from tweet embedding model:
train.py [options]
feature_extraction.py
coa/co_attention.py --precomp_features
Latest results:
Retrieval-based hashtag recommendation:
train.py --precomp_images --max_lt_tweets 50 --components image,text,hashtags,user,location --k_vals 1,5,10 --max_violation --max_similarity --target average
hashtag_recommendation.py
| k | Accuracy | Precision | Recall | F1 |
|---|---|---|---|---|
| 1 | 14.14 | 14.14 | 10.15 | 11.15 |
| 5 | 20.1 | 4.58 | 15.04 | 6.61 |
| 10 | 21.74 | 2.56 | 16.32 | 4.23 |
Retrieval-based hashtag recommendation:
train.py --precomp_images --max_lt_tweets 50 --components image,text,hashtags,user,location --k_vals 1,5,10 --max_violation --max_similarity --target hashtags
hashtag_recommendation.py
| k | Accuracy | Precision | Recall | F1 |
|---|---|---|---|---|
| 1 | 21.78 | 21.78 | 15.32 | 16.97 |
| 5 | 27.02 | 6.22 | 20.08 | 8.94 |
| 10 | 29.75 | 3.7 | 22.68 | 6.09 |
Baseline with tweet embedding vectors (target hashtags):
train.py --precomp_images --max_lt_tweets 50 --components image,text,hashtags,user,location --k_vals 1,5,10 --max_violation --max_similarity --target hashtags
feature_extraction.py
coa/co_attention.py --precomp_features
| k | Accuracy | Precision | Recall | F1 |
|---|---|---|---|---|
| 1 | 0.44 | 0.44 | 0.13 | 0.2 |
| 5 | 8.86 | 1.77 | 6.86 | 2.68 |
| 10 | 9.68 | 1.03 | 7.67 | 1.76 |
Baseline:
coa/co_attention.py
| k | Accuracy | Precision | Recall | F1 |
|---|---|---|---|---|
| 1 | 10.28 | 10.28 | 7.88 | 8.52 |
| 5 | 14.38 | 3.68 | 11.84 | 5.28 |
| 10 | 16.36 | 2.11 | 13.44 | 3.48 |