The data were kindly provided. Please help protect the privacy of individuals who contributed to the data set by _not_ downloading or sharing the data, and carry out all the analysis on the platform.
\nBecause of the wide range of backgrounds of your peers, you can only use the methods discussed in the book in Chapters 3 of the ISLR book. Any use of any other methods will automatically disqualify your submission.
\n
All the available data may be used. Additional data sources (e.g., specialized word lists) may also be included. For extracting text features from the transcript ideas may be obtained from anywhere (i.e., from papers published on the topic, etc.), but should be limited to extracting substrings from the text (i.e., words, word pairs, character sequences, punctuation, etc.).
\n
You work in teams of 4 or 5 members (6 only if absolutely necessary). Each team submits their work as a team. For this competition each team can enter 2 submissions.
\nYour last submission should be made right before the closing deadline; that is, only after your presentations about your first submission.
\nTake advantage of the first round of presentations to improve your teams position on the leaderboard. That is, learn from the other teams presentations and 'steal' their good ideas!
\n\n\ndplyr with pipes (`%>%`) where possible (this generalizes much more easily to large data bases such as SQL data bases, makes it much easier to understand processing steps)\nYou cannot sign up to Kaggle from multiple accounts and therefore you cannot submit from multiple accounts.
\n\n Privately sharing code or data outside of teams is not permitted.\n It's okay to share code if made available to all participants on the forums.\n
\nTeam mergers are not allowed in this competition.
\n\nThe maximum team size is 6 members.
\nYou may submit a maximum of 2 entries per day.
\nYou may select up to 2 final submissions for judging.
\nStart Date: TBA
\nMerger Deadline: None
\nEntry Deadline: None
\nEnd Date: 25/09/2018 14:00 CET
\n\t\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "data-description", "content": "The YouTube Personality Dataset\n-------------------------------\n\nThe YouTube personality dataset consists of a collection of behavorial features, speech transcriptions, and personality impression scores for a set of 404 YouTube vloggers that explicitly show themselves in front of the a webcam talking about a variety of topics including personal issues, politics, movies, books, etc. There is no content-related restriction and the language used in the videos is natural and diverse. \n\nDetailed description\n--------------------\nSee https://www.idiap.ch/dataset/youtube-personality for complete description.\n\n**Behavioral features** \n\nNonverbal cues were automatically extracted from the conversational excerpts of vlogs and aggregated at the video level. \n\n- Nonverbal Features from Audio:\n\nAudio cues were computed from the audio channel the toolbox developed by the [Human Dynamics group at MIT Media Lab ](http://groupmedia.media.mit.edu/data.php) and include both speaking activity and prosody cues. This is a total of 21 features.\n\n- Nonverbal Features from Video:\n\nThe overall motion of the vlogger is an indicator of vloggers’ excitement and kinetic expressiveness. We computed the overall visual activity of the vlogger with a modified version of motion energy images called ”Weighted Motion Energy Images” (wMEI)]. The normalized wMEI describes the motion throughout a video as a gray-scale image, where the intensity of each pixel indicates the visual activity in it. From the normalized wMEIs, we extract statistical features as descriptors of the vlogger’s body activity such as the entropy, median, and center of gravity (in horizontal and vertical dimensions). This is a total of 4 features.\n\n**Manual transcripts**\n\nWe asked a professional company to manually transcribe the audio from vlogs. The transcription corresponds to the full video duration. In total, around 28h of video were annotated. \nWe manually anonymized the data by substituting with a ‘XXXX’ any name entities referenced by the speaker that could be use to identify the video or the vlogger, as well as references to other social media accounts. The transcriptions are provided in raw text and contain a total of ~10K unique words and ~240K word tokens.\n\n**Personality Impression Scores**\n\nThe personality impressions consist of Big-Five personality scores that were collected using [Amazon Mechanical Turk](https://www.mturk.com/mturk/) and the [Ten-Item Personality Inventory ](http://homepage.psy.utexas.edu/HomePage/Faculty/Gosling/tipi%20site/tipi.htm). MTurk annotators watched one-minute slices of each vlog, and rated impressions using a personality questionnaire. The aggregated Big-Five scores are reliable with the following intra-class correlations (ICC(1,k), k=5): Extraversion (ICC = .76), Agreeableness (ICC = .64), Conscientiousness (ICC = .45), Emotional Stability (ICC = .42), Openness to Experience (ICC = .47), all significant with p < 10^-3.\n\nThe personality scores are only available for 324 of the participants; the missing 80 constitute the test set. \n\n**Gender**\n\nThe collection is mostly mostly balanced in gender, with 194 males (48%%) and 210 females (52%). The gender labels are also provided.\n\n\nAcknowledgments\n-----------------------\nThe authors", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Description", "content": "The YouTube personality dataset consists of a collection of behavorial features, speech transcriptions, and personality impression scores for a set of 404 YouTube vloggers that explicitly show themselves in front of the a webcam talking about a variety of topics including personal issues, politics, movies, books, etc. There is no content-related restriction and the language used in the videos is natural and diverse.
\nYour team should employ any of the methods discussed in some detail in Chapter 3 of ISLR (James, Witten, Hastie & Tibshirani, 2017), in combination with creative feature engineering and feature selection, to obtain the best numerical predictions for all the personality axes.
\nFor 80 of the 404 vloggers the personality impression scores are missing. Your method should be used to predict these missing personality scores.
\nNo other methods than the ones mentioned in ISLR chapters 1 through 3 may be used. Doing so disqualifies your team.
\nPlease note that the file format in which you need to upload your team's predictions is different from the file format of the data set.
\n\n## Detailed data structure\n\nThe data is stored in a .zip file. How to extract files from this zip file on the Kaggle platform is demonstrated in the Quick Start Notebook that you can fork on the Kernels tab. \n\nOnce extracted, in the working directory you will find a directory called 'youtube-personality' which contains the following files and folders\n\n- 'YouTube-Personality-audiovisual_features.csv'\n- 'YouTube-Personality-gender.csv'\n- 'YouTube-Personality-Personality_impression_scores_train.csv'\n\nThe 'transcripts' in the zip file contains all the transcripts transcribed from the videos in plain text files. These need to be extracted in a special way. See the Quick Start notebook.\n\nData corresponding to a single vlogger are identified by the `vlogId`.\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Evaluation", "content": "The evaluation of your team's predictions accuracy is measured the percentage of correct predictions as aggregated across all test set authors across all personality axis. That is, the performance in computed as\n\\[\n\\sqrt{{1\\over 5N}\\sum_{a \\in Axes} \\sum_{i\\in\\text{Test}} (y_{ia} - \\hat y_{ia})^2}.\n\\]\nHere \n\\[\n\\begin{align}\nAxes &\\qquad \\text{is the set $\\{Extr, Emot, Agr, Cons, Open\\}$}, \\\\\nTest &\\qquad \\text{is the set of vlogger IDs in the test set}, \\\\\ny_{ia} &\\qquad \\text{is the observed score for personality axis $a$ for the $i$-th vlogger in the test set, and} \\\\\n\\hat y_{ia} &\\qquad \\text{is the corresponding predicted score.}\n\\end{align}\n\\]\nN is the size of the test set.
For every author in the dataset, submission files should contain two columns: Id and value. The Id column identifies the test set author and personality axis, by means of concatenating the vlogId string and the axis label (Extr, Agr, Cons, Emot or Open) together with an underscore (_) in between. The value column should contain the predicted value. For example, if for `vlogId` VLOG14 your team's prediction is that this person has scores Extr = 4.518818, Emot = 2.445342, Agr = 3.005654, Cons = 1.221221, and Open = 4.005012, then the submission file should contain the rows
⋮\nVLOG15_Extr,4.518818\nVLOG15_Emot,2.445342\nVLOG15_Agr,3.005654\nVLOG15_Cons,1.221221\nVLOG15_Open,4.005012\n⋮\n\n
The file should contain a header and have the following format:
\nId,Expected\nVLOG8_Extr,3.518818\nVLOG8_Emot,2.946352\nVLOG8_Agr,1.205654\nVLOG8_Cons,3.221221\nVLOG8_Open,4.702312\n⋮\netc.\n⋮\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false } ] }