Concept building.

Our main goal was to extract quantitative features related to diverse characteristics from documents (see Fig 2). We considered the feature value as the degree of relevance for a feature. In detail, the feature value represents the extent to which a document has a particular characteristic. For example, sentiment analysis concerns one such quantitative feature, or the extent to which a document is positive or negative. We generalised this idea to various other user-defined characteristics. Examples of such characteristics include the extent to which a document is related to finance, immigration, and family issues. In particular, we built a lexicon, i.e. a set of keywords, relevant to the characteristics and utilised it to assign a feature value to a document by computing the degree to which the document contains those characteristics defined in the lexicon and other potentially relevant keywords. In this study, we considered a characteristic to be a concept describing a particular phenomenon or object, and defined a concept by constructing a set of keywords, whose meanings were relevant.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Concept for building the workflow.

https://doi.org/10.1371/journal.pone.0177630.g002

Concepts can play an important role in document analysis in diverse fields. That is, one can build useful domain-specific concepts in economics, politics, and social sciences and define the characteristics of documents with respect to these concepts. In the case of a spam-filtering task on documents and comments, for example, we can actively employ a ‘spam’ concept consisting of suspicious terms that usually appear in spam mails to measure the likelihood of the comment being unsolicited mail.

Here, the concept building process was composed of two steps: (1) the initial construction of a relevant keyword set, followed by its (2) user-interactive expansion. In order to facilitate the first step, we provided a user with the initial sets of coherent keywords obtained with two different techniques. The first technique we used was topic modelling, which algorithmically computes those representative keywords emerging from a document corpus. The user can then select some of them as an initial word set for their own concepts. As the other method to provide initial keywords, we computed the representative keywords from the centroid vectors obtained by k-means clustering on word embedding vectors[37].

Once a user formed an initial, small-sized lexicon for a particular concept, the second step was to interactively expand it by using a recently proposed visual analytics system named ConceptVector. Based on the initial lexicon given as user inputs, ConceptVector recommended potentially relevant keywords to enable users to easily add a subset of them to the lexicon. As the lexicon expanded, ConceptVector adjusted the recommended keywords that match the semantic meaning of the concept.

The foregoing procedure is discussed further below.

Topic modelling for initial lexicon building.

The topic modelling approach we used to extract representative keywords emerging from a document corpus is non-negative matrix factorisation, where the non-negativity allows users to interpret the value from factor matrices as the relevance score of a word or a document to a particular topic as mentioned above.

In particular, we constructed a document-term matrix A from the 17,381 forum articles and 627,122 user comments collected from the Bitcoin forum (See Table 1). Each article contains five attributes, ‘content’, ‘topic’, ‘comments’, ‘date’, and ‘views’, whereas each comment contains ‘content’ and ‘date’ features. Using the `date’ field, we split the document-term matrix per day for our analysis. We then applied the topic modelling to each so as to extract the different topic sets and their representative keywords across different dates.

The mathematical details of this process are as follows. Given a document-term matrix where m is the number of articles and n is the dictionary size, Non-negative Matrix Factorization(NMF) approximately factorises it into two matrices and , where d represents the number of topics (50 in our study), e.g. (1)

The columns in the resulting matrix W correspond to different topics and the keywords corresponding to the dimensions of the k largest value in each column function as the representative keywords of the topic.

Expanding the lexicon via word recommendation.

We proposed two types of concepts in the system. A unipolar concept represents exactly one concept such as crude oil and immigration. A bipolar concept has two polarities that oppose each other, e.g. positive vs. negative, progressivism vs. conservatism. In the case of building a concept, the system has positive, negative, and irrelevant word sets. When a user provides a word as an input, the system provides 50 recommended words that are potentially relevant to the seed word. We then automatically sorted the recommended words into five clusters, using the k-means clustering, to gather closely related terms into one group.

Once the lexicon of a concept is created by user interactions, the document rating process utilises the concept built in the process above. Because of the lack of expression resulting from the limited number of words a person could manage, we applied the kernel density estimation (KDE) in the word rearranging phase.

Computation of document relevance to concept.

Prior to the KDE, the concept had a limited number of descriptive terms for a characteristic, which resulted in a lack of expression and description. Therefore, the KDE served for the probabilistic smoothing over every word. This smoothing process is the most important procedure for document analysis since the score rating process cannot consider synonyms or closely related words that also represent a specific concept. Based on the assumption that the input terms describe the concept sufficiently well, we constructed a kernel that exerts influence on the entire vocabulary. ConceptVector adopts a Gaussian kernel as described below.

For the class y ∈ {positive,negative,irrelevant}, the conditional probability for each class can be calculated by the distance function d that represents the distance between a word in the word set in each class and the kernel k that ensures a proper balance between the given word and the others. The conditional probability of a keyword z for a class c can be computed as below: (2) which can also be seen as the relevance score to each class.

Since our final goal was to obtain scores by taking all classes into consideration, we rated a concept in view of all classes. For instance, ‘happy’, in the case of a bipolar concept, was rated for the positive, negative, and irrelevant classes. We calculated the bipolar rating as below: (3) (4)

The range of the bipolar score is [-1, 1] because the max value of p(y = positive,z) and p(y = negative,z) are 1.

Granger causality test.

The Granger causality test is based on the supposition that if a variable X causes Y, then any change in X will methodically happen before any change in Y[17, 22, 38]. As shown in past research, slacked estimations of X display a measurably noteworthy connection with Y[17, 22, 38]. Nevertheless, connection does not imply causation. We test whether the time arrangement of a discussion of conclusions contains any prescient data with respect to vacillations in the Bitcoin transaction and price.

Our time arrangement at the Bitcoin transaction count and price, indicated by S_t, reflects day-to-day change in the Bitcoin transaction count and price. To test whether the idea of gathering feelings in the time arrangement could forecast the change in the vacillation in terms of the Bitcoin transaction and price, we considered the difference clarified by two linear models as in (5) and (6) below. The first model uses just n slacked estimations of S_t for the forecast. However, the second model uses the n slacked estimations of both S_t and the time series of a concept of forum opinions, meant by X_t−1,⋯,X_t−n. We completed the Granger causality test as indicated by the models in (5) and (6).

(5)

(6)

In view of the consequences of the Granger causality test, we can reject the null hypothesis, whereby the time series of a concept of forum opinions does not predict fluctuations in the Bitcoin transaction count and price with a high level of confidence. The Granger causality test was performed on the Bitcoin transaction count and price for a time lag of 1 to 12 days.

Deep learning model.

Using the gathered data and the analysed and rated comment data, we built a model for predicting the fluctuation in the Bitcoin price and transaction through deep learning. Deep learning is widely used for addressing diverse challenges[8, 39]. Despite the quantitative and qualitative increases in Bitcoin-related formal and informal data following the broadening applicability of Bitcoin, deep learning has rarely been used to explore Bitcoin price trends and to address other Bitcoin-related challenges. We created a setting to apply deep learning to the data spanning a period of 2.8 years.

As the first step, we standardised the data to improve its applicability to the learning model. An example of applicable input data is provided in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Example of deep learning data set.

The z-score (, where and represent the mean and standard deviation for every date, respectively) of data for the previous 12 days (t = 12) was used as the values.

https://doi.org/10.1371/journal.pone.0177630.t002

Subsequently, to use the input data for prediction, we set up a deep learning model. Multiple hidden layers were accumulated for learning to identify deep data structures. Specifically, 1, 2, 3, and 5 hidden layers were constructed to select the layer structure that returned the best possible prediction result. The number of neurons that were allocated to each hidden layer was 1,024.

As for the input layers, based on the input data provided in Table 2, 15 input data points were represented as serial vectors to allocate neurons based on the cumulative number of days spent on learning, i.e. 45, 75, 105, and 180 neurons were allocated to cumulative 3, 5, 7, and 12 days. As for the output layer, two neurons were allocated while the probability of rise/fall was represented with the softmax function. The prediction model was built using Google Tensorflow[7], and GPU operation (nVIDIA CUDA) was used to accelerate the deep learning process.