So as to realize a better understanding of the effectiveness of our proposed models, we perform ablation examine and current the leads to Table II. We display the quantitative outcomes of our experiments in Table I. The performance of different models is evaluated by four metrics: precision, facebook password change recall, F-measure, and accuracy. In Table II, E-attn is the abbreviation for facebook password change express consideration and that i-attn is the abbreviation for implicit consideration. For instance, in the primary and fifth plots (from upper left to bottom proper), the attention value of emoji «star», «four-leaf clover» and «heart» are much greater than the other words in the same captions. After filtering out the unrelated texts, we exhibit the incidence number of typical texts in Figure 8. We manually separate the phrases into three levels: prime 25%, median 50%, and backside 25%. In contrast, we divide the emojis into two levels: prime 50% and backside 50%. The reason is that almost all emojis appear in common posts and there will not be a transparent boundary between unpopular emojis and the other emojis. We manually divide the captions into two classes: words and emojis. During our coaching process, Adam optimization is used with a studying charge of 0.001 for the first two epochs and with a studying rate of 0.0001 for the following epochs.
We first examine the baselines with Explicit Attention mannequin. As is known to all, it is meaningless and unfair to check traditional algorithms like SVR, FM with neural community methods immediately on uncooked knowledge. Because the consumer setting calculation doesn’t rely on any extra info, (our user environments are extracted from users’ photos and captions), we evaluate Dual-consideration mannequin along with the other models. Env signifies the model that concatenates user environment directly by a hierarchical construction. To further enhance the model, we include user atmosphere and implicit attention mannequin to construct Dual-attention model. The dataset we assemble incorporates 441 customers and their 60785 picture-caption pairs, along with the corresponding number of «likes». Besides, the performance of mannequin improves as the variety of additive constructions will increase. For the reason that input of their model is a big matrix (the three-dimensional tensor) which already accommodates obvious and complete information of users, they apply a modified Factorization Machine (FM) to generate prediction outcomes.
Single Visual. The enter of Visual mannequin only includes pictures. However, on this paper, the twin-attention mannequin takes the raw image-caption pairs as input. Ultimately, there are 21874 picture-caption pairs for training, 2430 image-caption pairs for validation, and 6064 image-caption pairs for testing. We assume that only the picture-caption pairs can be found because we intention at predicting the publish recognition for specific customers. In Figure eight (a), words from «love» to «day» are selected from the highest 25%. Based on statistic results, words that describe time («year», «day», «time»), attribute («amazing», «beautiful») and correlated with vacation («festival», «weekend», «selfie») are very seemingly show up amongst the highest 25%. Plainly the posts embody these phrases have higher tendency to receive «like» from other users. Similar to cross-promotion, brands can publish person-generated content material to indicate appreciation for existing clients and generate social proof at the identical time. In another phrase, the ultimate prediction rating can be thought-about as the average value of visible prediction score and textual prediction rating. Based on this result, we infer that caption info plays a extra reliable position than picture info on submit recognition prediction activity.
