Image sharing in social networks has increased exponentially in the past years. For example, according to official Instagram stats, there are 600 million Instagrammers uploading around 100 million photos and videos per day. Although the analysis of trends, topics and brands in social networks is mostly based solely on texts, the analysis of such a vast amount of images is starting to play an important role for understanding and predicting human decision making, while becoming essential for digital marketing, and customer understanding, among others. Previous works have proven the relation between text and the personality of the authors [7, 14], and recent studies have also shown that some image features can be related to the personality of users in social networks .
The main hypothesis of this work is that the relation between text and personality observed by researchers like Yarkoni  translates well into a relation between images and personality when we consider the images conditioned on specific word use. In his work, Yarkoni proved that there exist words that correlate with different personality traits with statistical evidence, see Table 1. For example, a neurotic personality trait correlates positively with negative emotion words such as ’awful’ or ’terrible’, whereas an extroverted person correlates positively with words reflecting social settings or experiences like ’bar’, ’drinking’ and ’crowd’. Considering this proven relation between text and personality, and the fact that posted images have a relation with their accompanying texts, we propose a methodology which, taking advantage of such existing text-personality correlation, exploits the relation between texts and images in social networks to determine those images most correlated with personality traits. The final aim is to use this set of images to train a personality model with similar performances than previous work based on texts or images alone.
|O||High||culture, films, folk, humans, literature, moon, narrative, novel, poet, poetry, sky|
|Low||anniversary, detest, diaper, hate, hatred, hubby, implore, loves, prayers, thankful, thanks|
|C||High||achieved, adventure, challenging, determined, discipline, persistence, recovery, routine, snack, vegetables, visit|
|Low||bang, bloody, boring, deny, drunk, fool, protest, soldier, stupid, swear, vain|
|E||High||bar, concert, crowd, dancing, drinking, friends, girls, grandfather, party, pool, restaurant|
|Low||blankets, books, cats, computer, enough, interest, knitting, lazy, minor, pages, winter|
|A||High||afternoon, beautiful, feelings, gifts, hug, joy, spring, summer, together, walked, wonderful|
|Low||asshole, bin, cost, drugs, excuse, harm, idiot, porn, sexual, stupid, violence|
|N||High||annoying, ashamed, awful, horrible, lazy, sick, stress, stressful, terrible, upset, worse|
|Low||completed, county, ground, later, mountain, oldest, poem, road, southern, sunset, thirty|
In the computer vision community, the relationship between language and images has been exploited to automatically generate textual descriptions from pictures [37, 18, 38]. Indeed automatic captioning can be understood as a sampling process of words from a text distribution given an image , or . In this paper we aim at the opposite: we want to determine , that means, those images mostly related with a specific word that is strongly associated to a particular personality trait. Once the set of images is defined, the potential relation between personality and images will be modeled using a state-of-the-art deep neural network. Classification results will suggest whether there is psychological content in certain images, like it has been previously observed for certain texts and image features.
In our work, the human personality characterization called the Big Five model of personality is considered [4, 1, 8]. The Big Five model is a well-researched personality description, which has been proven to be consistent across age, gender, language and culture [23, 29]. In essence, this model distinguishes five main different human personality dimensions: Openness to experience (O), Conscientiousness (C), Extraversion (E), Agreeableness (A) and Neuroticism (N), hence it is often referred as OCEAN. Personality traits are characterized in the OCEAN model by the following features:
Openness: Appreciate art and ideas, imaginative, aware of feelings. People with this trait tend to have artistic interests and have a certain level of intellectuality.
Conscientiousness: Disciplined, dutiful, persistent, compulsive and perfectionist as opposite to spontaneous and impulsive. People with this trait tend to strive for something, and to be hard-workers and organized.
Extraversion: Warm, assertive, action-oriented, and thrill-seeking. Individuals with high levels of extraversion tend to be friendly, sociable, cheerful and fond of being in company with other people.
Agreeableness: Compassionate, cooperative, considerate. Agreeable people tend to be trusty, modest and optimistic.
Neuroticism: Emotional instability, anxious, hostile, prone to depression. Neurotics tend to be frustrated, anxious and experience negative emotions.
The five personality traits have already been related to text  and images  uploaded by users. Therefore, personality might be an important factor in the underlying distribution of the user’s public posts in social media, and thus, it is possible to infer some degree of personality knowledge from such data. In this work we go a step beyond the works in , and , showing that personality remains invariant to changes from the text domain to the image domain. Concretely, our contributions are:
A new framework for estimating users personality from images, called MindPics, accessed using a refined set of the tags proposed in .
A personality inference model that uses MindPics, as a proof that personality remains invariant across textual and visual domains.
2 Related work
The increasing growth and significance of the social media in our lives has attracted the attention of researchers, which use this data in order to infer about the personality, interests, and behavior of the users. Regarding personality, its inference has mainly been based on (i) text uploaded by users, and (ii) uploaded images.
2.1 Text-based personality inference
The relationship between language and personality has been studied extensively. As commented before, Yarkoni et al.  performed a large-scale analysis of personality and word use in a large sample of blogs whose authors answered questionnaires to assess their personality. This way, by analyzing the text written by users whose personality is known, the author could investigate the relation between word use and personality. The results of the analysis concluded that the usage of some specific words are correlated with the personality of the blogs’ authors.
Iacobelli et al.  used a large corpus of blogs to perform personality classification based on the text of the blogs. They proved that both the structure of the text and the words used are relevant features to estimate the personality from text. Also, Oberlander et al. 
studied if the personality of blog-authors could be inferred from their personal posts. To do so, the personality profile of 71 different users was collected through a questionnaire. By using word n-grams as features and SVM classifiers they successfully managed to infer the personality of the authors of the post.
Besides blogs, where long texts are common, personality studies based on language have also focused in much shorter texts such as the shared in the social network Twitter, where the text shared by users is limited to a maximum of 140 characters. In a similar way, Golbeck et al.  showed that the personality of users from Twitter could be estimated from their tweets taking into account also other information as the number of followers, mentions or words per tweet.
2.2 Image-based personality inference
An early attempt to model personality from images was presented in Steele et al. . This work explored the characteristics of profile pictures and the relationship with the impression that other users had of the owners of these pictures. Their findings show that users better agreed with the targets’ self-reported personalities when the profile picture was a photo of a human, taken outdoors instead of indoors, and the face appears smiling.
Cristiani et al.  proved that there are visual patterns that correlate with the personality traits of 300 Flickr users and thus, that personality traits of those users could be inferred from the images they tagged as favorite. To do so, two categories of visual features are extracted from the images: ”Aesthetic” and ”Content” features. Aesthetic features focus on aesthetic information of the images, aiming to represent the user preferences. This type of features include the amount of certain colors, the number of edges in the image, the entropy and the level of detail among others. Content features focus on describing what appears in the image by counting the number of faces and identifying the different objects using the GIST descriptor .
Guntuku et al. 
improved the low level features used in previous work by changing the usual Features-to-Personality (F2P) approach to a two step approach: Features-to-Answers (F2A) + Answers-to-Personality (A2P). Instead of building a model that directly maps features extracted from an image to a personality, with this approach the features are first mapped to the answers of the questionnaire BFI-10 for personality assessment. Then, the answers are mapped to a personality. Besides this two-step approach, they also add new semantic features to extract from the images, like Black & White vs. Color image, Gender identification and Scene recognition.
Later, Segalin et al.  proposed a new set of features that better encode the information of the image used to infer the personality of the user who favourited it. They proposed to describe each image with 82 different features, divided in four major categories: Color, Composition, Textural Properties and Faces. These groups of features are similar to the ones proposed by , but excluding the content
group and using instead the number of faces as a feature of the image. The reason behind is that faces are very common in the images and that the human brain is specifically wired to perform accurate face detection and recognition. Their method proved to be suitable to map an image to a personality trait, but it worked better for attributed personality traits rather than self-assessed personality.
The study of the personality conveyed by images has not only been used to infer the personality of users, but also to analyze how brands express and shape their identity through social networks. Ginsberg et al.  analyzed the pictures posted in Instagram by the leading food brands and classified them into different categories: product, person and product, people and product, humor and product, world events, recipes, campaign with no products, user-generated, celebrity, and video. Then, the analysis of the types of images posted by the brands was used to interpret the identity of each brand along five dimensions of personality: sincerity, excitement, competence, sophistication, and ruggedness.
|Word use only||Image use only|
|Golbeck ||Iacobelli ||Segalin ||Guntuku |
2.3 Integrating text and image for personality inference
Table 2 contains the recognition accuracy of each of the OCEAN traits for 5 different methods based on text or images. In this paper, we want to determine whether the correlations found between personality and texts or images separately, also holds when combining image and word use.
Indeed all previous visual-based approaches take advantage of the many ways in which users interact with images in social networks, such as posting an image, liking it or commenting on it. Specifically, most of the works described above consist on assessing the personality of users based on the images they have liked. For example, the main difference between the  and our approach is that in such paper, the personality is inferred based on which images have been tagged as favorite, thus becoming a study on the relation between aesthetic preferences and personality. In contrast, in our case, we explore directly the images shared by users based on accompanying texts strongly related to personality traits, so here the relationship between images and personality arises from the mere act of posting a picture in a social network as a process of communication with others. The whole procedure is detailed next.
As proven by Yarkoni et al.  there exists a relationship between the personality of people and the language they use. In other words, the language that we use can reveal our personality traits, so there is statistical evidence that the use of specific words correlate with the personality of online users. Based on that, we design a set of images conditioned by those words most related to specific personality traits. This can be seen as sampling images from a distribution of images conditioned on text , or .
For each trait of the Big Five personality traits defined before we have selected the list of words that correlate most positively or negatively with the trait, as suggested by . The positively correlated words have been used to identify those images most associated to the strong presence of a trait, and the negatively correlated words are used to determine those images most associated to its absence.
From this set of images we can train a deep learning model that learns to extract a personality representation from a picture, and use it to automatically infer the personality that the picture conveys.
3.1 Finding MindPics in Social Networks
Based on the aforementioned relationship between text and personality, the proposed personality model is built considering a large quantity of images, called Mindpics, tagged with specific personality-related words. These words are the most correlated ones with each personality trait, as presented in . In Table 1, we showed the words mostly related to each personality trait, which have been also used to identify that set of images for each trait used for training the neural network. So each image related to a tag will correspond to one of the five personality traits, and within the trait it will represent the high or the low presence of such trait.
As the aim of this paper is to evaluate whether there is any of the author’s personality information embedded in real world images, we have considered images posted in Instagram [5, 13, 33] . In this social network the users take and share images by posting them together with words or hashatgs. To build the Mindpics dataset, we first crawl a large collection of publicly-shared photos using the Instagram API.
The reason of choosing Instagram as the source of selecting the images for training the Neural Network is threefold: (i) images are the main content of Instagram instead of text. Contrary to more text-based social networks such as Twitter, most of the content and information shared by the Instagram user is conveyed in the image, so it is reasonable to think that such images embed personality, up to some extent. (ii) The fact that the images can be accompanied by text allows to easily identify those images that appear together with specific words. (iii) By considering in our experiments the public pictures posted by hundreds of users, these users are not aware that they are being part of a psychological experiment.
Instagram images are very interesting, because there are no boundaries in the kind of pictures that users use to communicate with their followers. Thus, different kinds of users will post different kinds of pictures, as proven by Hu et al. . In their work, they show that the pictures posted in Instagram can be classified into eight main categories, and the users can be divided in five different groups, depending on what kind of pictures they post. These eight main picture categories are: friends, food, gadget, captioned pictures, pets, activity, selfies, and fashion. The difference on the type of images posted can be influenced by the city of the users,  or their age .
In order to determine the MindPics set of images, we used the words from Table 1 to query images. For each personality trait, 22 words were used, 11 for each component, and about 1,100 images were selected for each word. The total number of MindPics images used for training the personality model is . Because we used the same number of words per trait and the same number of images per word, the number of training images results balanced within the 5 personality traits and also within the used tags, thus each trait being trained from around images.
In Figure 1 we show some 10 random samples of each personality trait obtained with the procedure described above. From left to right the classes are: high Openness, low Openness, high Conscientiousness, low Conscientiousness, high Extraversion, low Extraversion, high Agreeableness, low Agreeableness, high Neuroticism, low Neuroticism. By observing these pictures we can see how despite the challenging variability of the data we can identify that some visual patterns emerge, like pictures containing more people and crowds for High Extraversion than for Low Extraversion, or that Low Neuroticism correlates well with images depicting more nature and landscape patterns than in High Neuroticism.
As it can be seen, images of the same class present a lot of variability, see for example how High Openness contains images of different objects like books, faces, text, drawings, and landscapes among others. Also, we can see how images of people appear in several classes. This huge variability is a very challenging problem, because it is harder to recognize unique patterns in the images of each class.
Summarizing, despite the huge intra-class and inter-class variabilities of the images associated to each of the personality traits, we can show that there are consistence between the images of the same class, whose hashtags are related to the words suggested by .
3.2 Building the Personality model
Once we have defined the procedure to determine which set of images is most related to each personality trait, we next describe how we can model this relationship between MindPics and personality.
In this work we have used a neural network model that maps an input image to a desired output by learning a set of parameters that produce a good representation of the input. This model is hierarchical, i.e. it consists of several layers of feature detectors that build a hierarchical representation of the input, from local edge features, to abstract concepts. The final layer consists on a linear classifier, which projects the last layer features into the label space. Let be an input image and a parametric function that maps this input to an output, where are the parameters. Because a neural network model is a hierarchical combination of computation layers, the output is a composition of non-linear functions:
where is the number of layers in the model, and each computation layer corresponds to a non-linear function  with its own parameters . We find by empirical risk minimization:
is the cross-entropy loss function,is the output of the model defined as and is the correct output for input
. This function is minimized iteratively by means of Stochastic Gradient Descent (SGD), and the process of minimizing the loss function over a set of imagesis referred as training the model. There are different types of Deep Learning architectures, in this work we use a Convolutional Neural Network (CNN) model , 
, since CNNs are especially suited for 2D data. After training, the output of the CNN for an image is a vector of scores for each of the personality traits.
For predicting the high and low scores for each of the personality traits, we propose two different strategies: one model for each trait, and a all-in-one model.
3.2.1 One model for each trait
Having one model for each trait consists on training five CNNs, using five different sets of parameters , each one to predict one of the Big Five personality traits. Namely, each CNN receives an image as input and produces a vector of two dimensions as output. This vector is used as input to the softmax function, that produces
, a score vector indicating the probability of an individual personality trait to be inferred from the image. For an output vectorof two components the softmax function for each component is defined:
Note that . Because the task to solve is a classification task, the Multinomial Logistic Loss is used as the loss function to minimize. The loss for one sample is defined as:
where is the probability assigned by the CNN model of the input belonging to the true class .
Note that, although training one model for each trait might seem conceptually easier, it increases the computations by a factor of five, and does not benefit from feature reusing, which is important for the model to generalize. To address these shortcomings, we propose the all-in-one model.
3.2.2 All-in-one model
We propose to use the same CNN for all five traits, but using five different classifiers on the last layer, one for each of the Big Five traits. Each output layer is independent of the others, and consists on a binary classifier like the described before, which has its own loss function. In Figure 2 it is depicted an scheme of this model.
The loss used for each of the 5 independent binary classifiers is the Multinomial Logistic Loss from Equation 4
as in the single-classifier model, with one minor modification. For the multi-classifier setup we must consider that we only want to backpropagate errors made by the classifier responsible for the ground truth label. For instance, if the true label of an image is”high Openness” the only error backpropagated should be the error produced by the ”Openness” classifier. Hence, we modify Equation 4 so that the loss for each classifier is zero if the classifier does not have to be considered:
This equation defines the loss function computed for one of the independent classifiers . Each classifier is independent from the others, so each classifier will compute its own loss function and backpropagate the error. The reason for Equation 5 is that by setting the loss to 0 for the classifiers that do not have to be considered for the input image, the gradients backpropagated to the previous layers by those classifiers will be also 0. Hence, the only classifier that will take part in updating the network parameters is the desired one.
In this section we describe the different models that we have tested to infer personality traits from images and explain the details of the training process. We finish the section by exposing the quantitative results of our approach and with a qualitative analysis of the model trained for personality inference.
CNN models. There are some common CNN architectures [19, 32, 36, 11] that are well established for computer vision tasks. In our experiments, we have used two of these CNN models. The first model is the same as proposed by Krizhevsky et al.  with the only difference that we replace the output layer of 1000 dimensions by the two configurations explained before. This architecture receives an input image of pixels which is ran through five consecutive convolutional layers. Then, the output of these convolutional layers is ran through two Fully Connected layers before feeding the output layer. From now on this model will be referred as Alexnet for brevity. The other CNN model we have used is a Residual Network presented in , abbreviated to ResNet
. In this case we also change only the output layer to adequate the architecture to our task. The main difference between these two models is that the latter uses residual connections between the network layers, allowing for an easier optimization process of deep networks, and thus allowing to increase their depth. Another difference is that the ResNet architecture uses Batch Normalization after every convolutional layer. This technique reduces the internal covariate shift in the distribution of the model activations by normalizing the inputs of each layer. In the experiments we use a ResNet of 50 layers, whereas the Alexnet model consists of 8 layers. For both models we test the two output configurations explained previously.
Finetunning. We also test two different options for initializing the network’s weights: the first option is to train the model from scratch starting from random weights and the other option is the finetunning approach, which consists in initializing the network with the weights learned for another task. In this case, we initialized the network with the weights of a model trained on ImageNet . The finetunning approach has been proved  to be useful to train neural networks in small datasets with superior performance. The idea behind this approach is that the network first learns how to extract good visual features in a large dataset and then uses this features to learn a classifier in a smaller dataset.
To train the models, we used the deep learning framework Caffe and a GPU NVIDIA GTX 770 with 4GB of memory. The optimization algorithm used to train the network is Stochastic Gradient Descent with momentum of 0.9. For the Alexnet model we use a batch size of 128, whereas for the ResNet network we set the batch size to 32. The learning rate is set to when training a network from scratch and to when finetunning, increasing the learning rate by a factor of at the new layers. Dropout  is used in the Fully Connected layers with a probability . regularization of the weights is also used with a factor of . Additionally, during the training stage we apply data augmentation to the input images. Namely, we randomly apply horizontal mirroring to the images and crop a random patch of pixels of the original images. The cropped patch is used as the input to the network. The only pre-processing of the images is mean subtraction. The mean image to be subtracted is computed on the training set. A random split of the Mindpics dataset is used to divide the images in non-overlapping training and testing sets, of the images are used for training and for testing.
4.1 Quantitative Results
In Table 3 the accuracies on personality recognition for each trait obtained by the different models and configurations we tested are shown. The first two rows show the performance of the Alexnet model trained from scratch with random weights initialization. The first row shows the results obtained by training an independent model for each trait, and the second row the results obtained by training just one model that shares most of the computation and that has an independent output layer for each trait. As it can be seen in the table, the all-in-one model obtains better accuracies in all the traits except for Agreeableness, and the average accuracy is 0.5 % better compared to training five independent models.
The rest of the results correspond to the finetunning of Alexnet and ResNet models starting from the weights learned on the ImageNet dataset. The results show that pre-training the network in a larger dataset in order to learn to extract better features increases performance in all the personality traits, increasing the average accuracy by . It can also be seen that the all-in-one model performs better in this scenario too, both for Alexnet and ResNet neural networks. In these cases however, the increase of performance is smaller, the average accuracy increases by in both cases.
The all-in-one configuration learns better features because the personality classifiers share the feature extraction layers and each classifier contributes to the learning, whereas a model that only has one classifier does not see as many different images so it learns worse features. However, when pre-training on the Imagenet, both networks start already with good feature extractors, so this effect is not as important as when the networks are trained from scratch. Besides the increase in accuracy performance, the all-in-one network is also much more efficient, because it shares most of the image processing steps for all traits, thus reduces the amount of computation by a factor of five.
Finally, the results also show that the ResNet network is significaly better than the shallower network Alexnet, achieving up to more in average accuracy and consistently increasing the performance for all traits. This increase in performance can be explained from the increase in depth of the model, which allow the models to learn better representations. To further explore the performance results, in Figure 3 we show the Receiver Operating Characteristics (ROC) and Precision-recall (PR) curves for each trait obtained by best model, along with the area under the ROC curve (AUC) and Average Precision (AP). The predictions used to plot the curves are obtained from the ResNet all-in-one CNN. As it can be seen, the Extraversion trait obtains the best scores, followed by Conscientiousness. The dashed line in the ROC curves figure represents a hypothetical random classifier with and AUC of 0.5. As it can be seen, the ROC curves for each personality trait are far from being random, indicating that the relation between text and images exists and that the model is able to learn a mapping from image to personality.
4.2 Qualitative Results
In this section we analyze the qualitative results obtained with our deep personality model based on MindPics. As previously seen, the best models are those pre-tained in ImageNet, which learn good features in the pre-training process, and the all-in-one models that learn to extract better features by sharing the feature extractor layers between different personality classifiers. The relationship between the quality of the extracted features and the final classification performance has already been explored in the literature. For instance,  showed that for personality recognition improving the hand-crafted features extracted from images lead to an improvement of the results obtained with more basic features .
Therefore, we can assume that the increase in performance of the models that get better results is caused because the network is able to extract a more suitable representation of the image in the feature space, a representation that allows the classifier to better discriminate the images. To get a better insight of what the new deep features are detecting, we visualize and analyze the images that maximally activate the output of the network for each personality trait.
Namely, in order to know which images better represent each personality trait, we find those pictures that maximally activate a specific output of our model. Similarly to 
, we input all the images to our model, inspect the activation values of a specific neuron, and look for the images that produce these maximum activations.
In our case, we inspect the output units associated to each personality trait. For example, by looking at the output of our Extraversion classifier for each of the MindPics, we can know which ones are classified as High Extraversion with more confidence. In Figures 4, 5, 6, 7 and 8 the most representative MindPics of the High and Low scores for each trait are shown.
Among the MindPics that maximally activate the High Openness trait we can see pictures of books, the moon and sky, while for Low Openness the most relevant pictures are love related. For High Conscientiousness most of the images are photographs of food, specially healthy food, whereas for Low Conscientiousness we mostly see pictures of people. In Extraversion it can be seen a clearly distinction between the images that maximally activate the High and Low outputs. The High Extraversion output is mostly activated by pictures of a lot of people, whereas the Low Extraversion output reacts to cats, books, and knitting images. In High Agreeablenes we mostly see flower pictures, whereas the Low score responds to pictures with text and naked torsos. Lastly, in the Neuroticism trait we observe that the High score is maximally activated by pets, whereas for Low score we see pictures of landscapes and sunsets.
In order to further analyze the behavior of our model, we study the image representations that the model produces. These representations are obtained by extracting the activations of the last feature detector of our model (the 50-layer ResNet), a high-level feature vector that describes the input image. In our case, the last layer is a Pooling layer which produces a 2048-dimensional vector. Given that these extracted features reside in a high-dimensional manifold, we can not directly visualize the images in feature space. By projecting the high-dimensional features to a 2D space, we are able to observe the underlying structure of the images representations.
To compute this projection we use the t-Distributed Stochastic Neighbor Embedding (t-SNE) 
technique for visualizing high-dimensional data. This method retains local structure of the data and global structure of the data, so similar images appear together in the 2D space. It works by translating similarities between data to joint probabilities, and minimizes the Kullback-Leibler divergence between the probability distribution over high-dimensional points and the probability distribution over the low-dimensional data points.
. In the t-SNE figures we show with an orange border the pictures belonging to the High score and with a blue border the images belonging to the Low score of each trait. For the t-SNE data visualization we have used a perplexity value ofand a random initialization of the embedding. The visualization of the image features projected onto a 2D plane shows that for all traits, the pictures belonging to the High and Low scores are in general well separated, with only a few errors. We must remark that t-SNE does not use the labels of the images in the projection, so these clusters arise from the image representations made by the model.
Besides these intra-class separation, we also observe some inter-class clusters within the projected images. For example, in the Openness t-SNE projection, Figure 3(c), we can see three different types of images of High Openness. In the right hand of the figure, we see a cluster of images of books and writings, in the middle a small cluster of pictures of the moon and in the left cluster we observe more landscapes and sky images. We also project the MindPics that maximally activate each trait together in Figure 9. We can see how even that the model is not explicitly trained to distinguish images of different traits, the representations are discriminative enough so that images of the same trait cluster together in the feature space.
Social media is the product of the expression of the users by means of text, images, and speech. This is a great opportunity for companies and researchers not only to know about the content of the images that are being shared, but to know about the users that create and interact with those images themselves. One interesting trait to learn from the users is personality. In fact, different methods have already been proposed to extract the users’ personality from social media text and favorited images. This indicates that personality might be an underlying factor in the distribution of the users’ data. In this work we made a step towards confirming this hypothesis, directly focusing on the authors of the pictures and showing that personality remains invariant when moving from the text to the image domain.
In particular, we showed that given the images uploaded by those users who used words correlated to the different personality traits , it is possible to retrieve their personality from the images. Thus, images and text are correlated, which can be explained if both depend on a third variable, which is personality. In this study, we do not recover the full spectrum of personality traits of one user, but we infer a specific personality trait that a single image conveys. The underlying hypothesis is that when a user posts a picture in a social network, the picture is not expressing everything the author has in mind, only the specific message intended by the author. In the same way, a picture does not describe the whole personality of the user, but a portion of it. So, in order to get an estimation of the whole personality profile of the users, one could analyze all the different images posted by them, because each image conveys only partial information of their personalities.
Quantitative and qualitative results were provided, showing that a neural network is indeed able to successfully learn to map MindPics to personality traits. Moreover, the images retrieved with our model directly reflect the description of the Big Five traits, for instance, highly conscientiousness, which is described as: ”Conscientiousness is a tendency to display self-discipline, act dutifully, and strive for achievement against measures or outside expectations…” in [4, 1, 8], highly correlates to images of vegetables, and people exercising. Another example of our findings is that we have seen how images of groups of people, concerts, and social settings like a bar or a restaurant are associated with high confidence to high extraversion, in contrast to pictures related to activities that one tend to do alone like reading or knitting, that are associated with low extraversion.
Lastly, in Figure 10 we show some random samples of the Openness trait obtained with the procedure described in this paper, the left side showing high scoring images and on the right side low scoring ones. As it can be seen, images of the same class present a lot of variability, for example High Openness contains images of different objects like books, faces, text, drawings, and landscapes among others. This is due to the fact that some tags indeed correspond to quite abstract concepts, that leads to very different content in those related images.
Besides the challenge of having this huge inter-class variability, we found that even images from different classes can be visually similar. For example, in Figure 10 we can see pictures of text and people in both High and Low scores of Openness. This is caused by the huge variability in real world images, and by the fact there are a lot of different images that are tagged with the same word.
A new framework for user personality inference has been proposed. Differently from previous approaches, the proposed framework was used to directly infer the personality of the authors of social media images retrieved according to the tags proposed in , which have been shown highly correlated to each of the Big Five personality traits. We then showed that it is possible to model such personality traits assigned to the retrieved images, using different strategies based on deep learning. Quantitative and qualitative results are shown, suggesting that personality is an underlying factor of the social media data distribution of users. These results open new directions of research for improving the proposed personality model by considering more words and specific images, evaluated with the supervision of psychology experts.
Authors acknowledge the funding received by the European Union’s H2020 SME Instrument project under grant agreement 728633, the Spanish project TIN2015-65464-R (MINECO/FEDER), the 2016FI_B 01163 and the 2017-SGR-1669 grants by the CERCA Programme/Generalitat de Catalunya, and the COST Action IC1307 iV&L Net (European Network on Integrating Vision and Language) supported by COST (European Cooperation in Science and Technology). Authors would like to specially thank Ms. Daniela Rochelle Kent and Mr. Alvaro Granados Villodre for their invaluable help with the ontology of words and the images used in the experiments, respectively. We lastly gratefully acknowledge the support of NVIDIA Corporation with the donation of a Tesla K40 GPU and a GTX TITAN GPU, used for this research.
- Barrick and Mount  M. R. Barrick and M. K. Mount. The big five personality dimensions and job performance: a meta-analysis. Personnel psychology, 44(1):1–26, 1991.
- Cristani et al.  M. Cristani, A. Vinciarelli, C. Segalin, and A. Perina. Unveiling the multimedia unconscious: Implicit cognitive processes and multimedia content analysis. In Proceedings of the 21st ACM international conference on Multimedia, pages 213–222. ACM, 2013.
- Deng et al.  J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009.
- Digman  J. M. Digman. Personality structure: Emergence of the five-factor model. Annual review of psychology, 41(1):417–440, 1990.
- Ginsberg  K. Ginsberg. Instabranding: Shaping the personalities of the top food brands on instagram. Elon Journal of Undergraduate Research in Communications, 6(1), 2015.
Girshick et al. 
R. Girshick, J. Donahue, T. Darrell, and J. Malik.
Rich feature hierarchies for accurate object detection and semantic
Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580–587, 2014.
- Golbeck et al.  J. Golbeck, C. Robles, M. Edmondson, and K. Turner. Predicting personality from twitter. In Privacy, Security, Risk and Trust (PASSAT) and 2011 IEEE Third Inernational Conference on Social Computing (SocialCom), 2011 IEEE Third International Conference on, pages 149–156. IEEE, 2011.
- Goldberg  L. R. Goldberg. An alternative” description of personality”: the big-five factor structure. Journal of personality and social psychology, 59(6):1216, 1990.
- Goodfellow et al.  I. Goodfellow, Y. Bengio, and A. Courville. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org.
- Guntuku et al.  S. C. Guntuku, S. Roy, and W. Lin. Personality modeling based image recommendation. In MMM (2), pages 171–182, 2015.
- He et al.  K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016.
- Hochman and Schwartz  N. Hochman and R. Schwartz. Visualizing instagram: Tracing cultural visual rhythms. In Proceedings of the workshop on Social Media Visualization (SocMedVis) in conjunction with the sixth international AAAI conference on Weblogs and Social Media (ICWSM–12), pages 6–9, 2012.
- Hu et al.  Y. Hu, L. Manikonda, S. Kambhampati, et al. What we instagram: A first analysis of instagram photo content and user types. In ICWSM, 2014.
- Iacobelli et al.  F. Iacobelli, A. J. Gill, S. Nowson, and J. Oberlander. Large scale personality classification of bloggers. In Affective computing and intelligent interaction, pages 568–577. Springer, 2011.
- Ioffe and Szegedy  S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
- Jang et al.  J. Y. Jang, K. Han, P. C. Shih, and D. Lee. Generation like: comparative characteristics in instagram. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, pages 4039–4042. ACM, 2015.
- Jia et al.  Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093, 2014.
- Karpathy and Fei-Fei  A. Karpathy and L. Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3128–3137, 2015.
- Krizhevsky et al.  A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105, 2012.
- LeCun et al.  Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
Maaten and Hinton 
L. v. d. Maaten and G. Hinton.
Visualizing data using t-sne.
Journal of Machine Learning Research, 9(Nov):2579–2605, 2008.
- Machajdik and Hanbury  J. Machajdik and A. Hanbury. Affective image classification using features inspired by psychology and art theory. In Proceedings of the 18th ACM international conference on Multimedia, pages 83–92. ACM, 2010.
- McCrae and John  R. R. McCrae and O. P. John. An introduction to the five-factor model and its applications. Journal of personality, 60(2):175–215, 1992.
- Oberlander and Nowson  J. Oberlander and S. Nowson. Whose thumb is it anyway?: classifying author personality from weblog text. In Proceedings of the COLING/ACL on Main conference poster sessions, pages 627–634. Association for Computational Linguistics, 2006.
- Oliva and Torralba  A. Oliva and A. Torralba. Modeling the shape of the scene: A holistic representation of the spatial envelope. International journal of computer vision, 42(3):145–175, 2001.
- Oquab et al.  M. Oquab, L. Bottou, I. Laptev, and J. Sivic. Learning and transferring mid-level image representations using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1717–1724, 2014.
- Rammstedt and John  B. Rammstedt and O. P. John. Measuring personality in one minute or less: A 10-item short version of the big five inventory in english and german. Journal of research in Personality, 41(1):203–212, 2007.
- Russakovsky et al.  O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015.
- Schmitt et al.  D. P. Schmitt, J. Allik, R. R. McCrae, and V. Benet-Martínez. The geographic distribution of big five personality traits: Patterns and profiles of human self-description across 56 nations. Journal of cross-cultural psychology, 38(2):173–212, 2007.
- Segalin et al. [2016a] C. Segalin, D. S. Cheng, and M. Cristani. Social profiling through image understanding: Personality inference using convolutional neural networks. Computer Vision and Image Understanding, 2016a.
- Segalin et al. [2016b] C. Segalin, A. Perina, M. Cristani, and A. Vinciarelli. The pictures we like are our image: continuous mapping of favorite pictures into self-assessed and attributed personality traits. IEEE Transactions on Affective Computing, 2016b.
- Simonyan and Zisserman  K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
- Souza et al.  F. Souza, D. de Las Casas, V. Flores, S. Youn, M. Cha, D. Quercia, and V. Almeida. Dawn of the selfie era: The whos, wheres, and hows of selfies on instagram. In Proceedings of the 2015 ACM on conference on online social networks, pages 221–231. ACM, 2015.
- Srivastava et al.  N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):1929–1958, 2014.
- Steele et al.  F. Steele, D. C. Evans, and R. K. Green. Is your profile picture worth 1000 words? photo characteristics associated with personality impression agreement. 2009.
- Szegedy et al.  C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1–9, 2015.
- Vinyals et al.  O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3156–3164, 2015.
- Vinyals et al.  O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: Lessons learned from the 2015 mscoco image captioning challenge. IEEE transactions on pattern analysis and machine intelligence, 2016.
- Yarkoni  T. Yarkoni. Personality in 100,000 words: A large-scale analysis of personality and word use among bloggers. Journal of research in personality, 44(3):363–373, 2010.