In this paper, we present a semi-supervised approach to modeling social and agentic characteristics of happiness. For this, we build four one-class autoencoder models, respectivley trained with 1) only social, 2) non-social, 3) agentic, and 4) non-agentic happiness. Then, we extract data from unlabeled data that are likely to belong to a prescribed type, as determined by the models. This paper presents the performance of predicting agency and social class with and without the extracted data. Our evaluation shows that the results are promising.
Emotion is a crucial factor for the understanding of humans and human
activities. As we are not purely logical, emotions do a ect our actions, responses,
and thoughts. A great deal of human artifacts (e.g., ne arts, novels,
architectures) might not be created without emotions, whether they are positive (e.g.,
happiness, love) or negative (e.g., sadness, anger, fear). Emotion researchers
often employ two di erent types of emotion models to describe various aspects of
emotions - either categorical or dimensional. In the categorical emotion model,
several discrete emotions are labeled distinctively as six basic emotions (joy,
sadness, anger, fear, disgust, and surprise) [
In the dimensional model, di erent types of emotions can be mapped into
either a two-dimensional (valence-arousal) [
Happiness is an abstract term that locates in an emotional state with
positive valence. Its de nition includes various metaphors and metonomies [
As Internet and smartphones are widely used, it is getting easier to gather a
variety of (unlabelled) data. Labelling or annotation of the data, however, is still
time-consuming and expensive. Thus interests and demand in semi-supervised
learning is increasing, as it can classify or cluster a large amount of unlabelled
data using a small set of labelled data. [
In this paper we present a deep-learning based semi-supervised approach to
the prediction of possible labels relating to happiness using the Happy DB data
set [
We test if two factors (socialness and agency) of happiness can be indicated by the textual description of the happy moment with a one-class autoencoder deep learning model. For this, we rst convert the data into one-hot encoding representation for the deep learning model, and carry out an initial evaluation of predicting social and agency classes. Next, we build four autoencoders, each modeled using only a certain social and agency label. Those models are employed to label subsets from the unlabeled data for further training. Lastly, the expanded train data are used to build the nal classi cation models for comparison. 2.1
The preprocessing step encodes a moment in the HappyDB [
While autoencoders are often used for dimensionality reduction and generation
tasks [
The basic idea of our approach is to build an autoencoder that reconstructs the given input as close as possible to the learned pattern. We hypothesize that the model trained with social happiness only will try to recover any input moment as close as possible to the description of a socially happy moment. Therefore, if a given input belongs to a non-social happiness, for example, it is likely that the recovered output is signi cantly di erent from the input. We can then determine whether the data that result in small di erences (e.g., Mean Square Error, Cross Entropy Error, L1) between the input and the output belongs to the class which the model is trained with. The next section describes our experiments to test this hypothesis.
We built our autoencoder model using the Keras framework [
We converted the moment texts into 1024 dimensional vectors as described previously in the Preprocessing section. The training data were prepared by selecting randomly 80% from the labeled data in the train set. The rest 20% were used as the test set. We further split the training data into 2 groups (SY, SN ) based on the social feature value (see Table 1), where SY refers to the data whose social values are `yes', and SN denotes the data whose social values are `no'; SY+ and SN+ are extended data sets that include the extracted accounts from the unlabeled data in the train set classi ed as social and non-social respectively using our approach. 3.2
We applied three representative machine learning algorithms as the classi ers: SVM, Xgboost, and KNN. The results show that the SVM algorithm (with Linear/RBF Kernel) performs best in predicting the social and non-social classes from the textual description of happy moments. We obtained the best performance 0.91 of F1-score when SVM with linear kernel is used as the classi er (see Table 2). This section describes the process of extracting additional data from the unlabeled set. As mentioned above, we hypothesize that the model trained with social happiness only would recover any input moment as close as possible to the description of socially happy moment. Likewise, the model trained with non-social happiness will try to recover the input close to the non-socially happy pattern. We compute the error as the di erence between the input and output. If the error is small, we can determine that the input account belongs to the class which the model is trained with.
We applied the autoencoder model to the unlabeled data in the train set. The model employs SGD as the optimizer and Categorical Cross Entropy as the loss function. The hidden layers use LeakyRelu as the activation function and the output layer uses sigmoid as the activation function. Each node is initalized using Xavier normal initializer. The batch size is 32, and the epoch is 32 as determined via numerous experimentations. We did not use dropout nor batch normalization, because the application of these normalization techniques degrades the performance, especially when predicting the non-social happiness. The di erence between the input and the output was computed with the L1 function. Our evaluation shows that the precision is very high with the top 3% of the output that displayed the smallest error or the di erence compared to the input. For example, when the model extracts accounts which generate errors that are smaller than the top 3% threshold of the output, the precision of the social happiness reached 0.9 and that of the non-social was 1. The precision of the social happiness went up to 0.95 when the top 1% was set.
We set the threshold of top 3% of the output data to obtain a larger size data. In other words, if an account in the unlabeled data displays an error smaller than the top 3% threshold set above, it can be interpreted that the unlabeled data is labeled as social happiness with the model trained with the SY set. Non-social happiness can be classi ed in the same fashion. A total of 2,169 moments each are added to make the SY+ set and SN+ set respectively (Table 1).
Table 2 shows that SVM and Xgboost perform worse, though with a trivial margin, when the unlabeled data were added to the train set. When KNN was used, on the other hand, adding the unlabeled data to the train set improves the performance marginally. The best F1-sore was 0.71 (k=3) for social happiness prediction when trained with extended data set and 0.75 (k=5) for non-social happiness prediction with original data. 4 4.1
We initially prepared the data sets for agency class prediction using the same training data that were prepared for the social class prediction. However, it turns out that the agency class was imbalanced: 6,209 accounts of yes and 2,239 accounts of no in the train set, and 1,587 accounts of yes and 525 accounts of no. When these data sets were used, the best performance of F1-score was about 0.87 for the agency class prediction regardless of the algorithms used. Adding extra data from the unlabeled data set lowered the best performance down to 0.58. On the other hand, the highest F1-scores for non-agency class prediction was about 0.49. The use of extra unlabeled data also degraded the performance down to 0.35. We attribute this bad performance to class imbalance, hence we prepared another data set to balance the yes and no classes.
We randomly under-sampled the agency data from the train and the test data, to balance the two classes (see Table 3). Since the number of moments in the train set are less than that in the social train data, we only chose 1.5% from the unlabeled data and obtained additional 1,084 moments. For agency prediction, the best model consists of one 512 dimensional input layer, 5 hidden fully connected layers each with (256,128,64,128,256) nodes and one 512 dimensional output layer to match the input size. 4.2
As Table 4 shows, the best performances were obtained when SVM with RBF kernel was used. The highest prediction of the happiness with agency was obtained when the model trained with the original labeled data (F1-score = 0.75). Yet, it is noted that the prediction of the happiness without agency was greatest when the extracted unlabeled data were added to the train set (F1-score=0.73). Figure 2 illustrates the performances of SVM algorithms in terms of F1-score when di erent train sets were used.
Overall, our approach shows higher performance for predicting the social class than for predicting the agency class. It also shows that training the SVM model with the labeled data in the train set generally results in a slightly better performance than training the model with a combined data set. However, training with the combined data enhanced the model non-trivially for the prediction of the non-agency happiness. We also noted that the social class prediction performs best when the linear kernel was used, while the agency class prediction performs best when the RBF kernel was used. This may indicate that the problem of agency class prediction is non-linear. 5
This paper presented a deep learning based semi-supervised approach to characterize the description of happy moments in terms of socialness and agency. We suggested a novel method of using an autoencoder model to extract relevant data for training from unlabeled data. We experimented with SVM, Xgboost, and KNN trained with using the textual description of the original train set with and without the data extracted from the unlabeled set. We obtained the best performance 0.91 of F1-score when SVM was trained with the original train set for the social class prediction. For the agency class prediction, F1-score of 0.75 was obtained for the prediction of the happiness with agency when SVM was trained with the under-sampled labeled data. Yet, the highest prediction of the happiness without agency was (F1-score=0.73) obtained when the model was trained with the combined data. We found that the adding extra data enhances the prediction performance for the non-agentic happiness without harming the prediction performance for the other classes.
Our evaluation is not conclusive due to the following problems. First, since our input vector is large and the autoencoder model is complex, a larger data set is necessary to train the model. Second, we ran the tests with only several model structures and a limited number of hyper-parameter settings.
Our autoencoder-based approach for predicting labels from the unlabeled data o ers two main advantages. First, our model can extract data with con dence. Therefore, when the available data is large, we can choose the accounts with a very high con dence as extra training data. Second, our approach is very useful especially when the dataset is extremely imbalanced. When most data belong to a certain class, the autoencoder trained with such data without labeling successfully models the class. Therefore, a moment that induces high di erence between the input and the output is likely to not belong to the class in question.
As future work we will enhance the model varying the autoencoder model structure and hyper-parameters. We also plan to test various word embeddings. Finally, we plan to investigate deeper into a general model of happiness considering previous studies on subjective well-being (e.g., features such as levels of negative a ect, satisfaction in particular domains, etc.) in addition to a proposed autoencoder learning model.
This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(2016R1D1A1B03933002). This research was supported by the MIST (Ministry of Science and ICT), Korea, under the National Program for Excellence in SW supervised by the IITP(Institute for Information & communications Technology Promotion)(2015-0-00914). This work was also supported in part by Institute for Information & communications Technology Promotion(IITP) grant funded by the Korea government(MSIT) (No.2017-0-01772 , Development of QA systems for Video Story Understanding to pass the Video Turing Test).