All datasets were processed by collecting the yearly/monthly traffic of occurrence of its most frequent words. By choosing a time evolution pattern you can specify how the time evolution of the words to be searched for should look like. In every case there are 3 patterns, which are the different variations of the same concepts. The concepts are: increasing, decreasing, A-shaped (increasing and afterwards decreasing), V-shaped (decreasing and afterwards increasing). Not in all datasets there can be found matching words for every pattern (e.g. there are no V-shaped words in the Stack Overflow dataset).
After choosing a dataset-pattern pair, the list of all words appear in the second table, ordered by the matching of the pattern. At the same time, the efficiency of a prediction based on the first half of the timeline can be evaluated. The word-pattern match contains the proportion of the points which fit the selected pattern (the best fit of the three patterns described above are chosen). A word is defined as matching the pattern if this proportion is at least 80%.
The prediction is based on the 20 most similar words ("neighbors"), which are often mentioned together with the originally chosen word. The assumption is that the trend of these words can predict a turn of the original one. The diffseq shows for every timeframe if the neighbors have a higher value (+) or the original word (-). Both curves are normalized by the sum of its values. In case of a 180° turn, the first half of the diffseq is expected to consist only of pluses, the second only of minuses (or the opposite, based on which turn do we want to predict). The diffseq-pattern match is calculated by the proportion of signs which are the same as expected. If this proportion is at least 80%, a turn is predicted.
The first plot shows the selected word in the first half of the timeframe. It is compared visually with its first 20 neighbors (see 2.). The second plot shows the selected word on the full timeframe. Guidelines are inserted in order to help the user to see the proportion of the points falling into the desired area according to the selected pattern (see 1.).