2 years, 1 month ago

First semester project for the Interactive Machine Learning 2021/2022 course

This is the first semester project for the Interactive Machine Learning 2021/2022 course. The task is to choose optimal batches of queries for training a multi-label prediction model.


The goal of this competition is to choose three subsets of samples from the data pool such that they bring the largest improvement to a prediction model. 

Competition rules are given in Terms and Conditions.

The description of the task, data, and evaluation metric is in the Task description section. 

Terms & Conditions
Final results
Rank Team Name Is Report Preliminary Score Final Score Submissions
True 0.0738 0.075700 2
True 0.0948 0.059800 6
True 0.0538 0.058100 3
True 0.0946 0.055800 21
True 0.1036 0.055300 11
Polska Gurom
True 0.0964 0.049000 15
150 g masła 200 g mąki ziemniaczanej 100 g mąki pszennej 3 jajka 150 g drobnego cukru do wypieków lub więcej 15 płaskiej łyżeczki proszku do pieczenia
True 0.0591 0.045100 16
True 0.0822 0.044700 28
True 0.0861 0.043700 20
True 0.0813 0.042500 44
True 0.0620 0.042300 13
Michał Siennicki
True 0.0717 0.040400 18
smacznej kawusi
True 0.0451 0.038100 17
True 0.0480 0.037600 20
True 0.0713 0.037000 24
True 0.0587 0.035200 3
True 0.0751 0.029200 21

The task in this project is to choose three subsets of samples from the data pool such that they bring the largest improvement to a prediction model. The size of subsets should be 50, 200, 500, respectively. 

The initial batch of samples, and the data pool are given as sparse matrices in the MatrixMarket format. Files in this format are simple text files with a header and three columns. The header consists of two lines of which the first one is irrelevant and the second one contains the dimensions of the sparse matrix - i.e., three integers giving the number of rows, the number of columns, and the total number of non-zero values in the matrix. After the header, there are three data columns giving the actual matrix values as EAV triples. In particular, in each line of the file, there are exactly three numbers: number of a row, number of a column, and the value. Keep in mind that the rows and columns are indexed from 1 (not 0).

Each row in the data corresponds to a preprocessed textual document, represented as a vector of TF-IDF values of its terms.

Additionally, there is a file with labels for the documents from the initial data batch. Each document can be labeled with one or more labels from the set A, B, C, D, E, F, G, H, I, J, K.

Format of submissions: solutions should be submitted as text files with three lines. The first line should contain exactly 50 integers - the indices of samples from the data pool (samples are indexed starting from 1), separated by commas. The second and the third line should contain analogous indices for the second and the third set of samples, with sizes 200, and 500, respectively.

Evaluation: the evaluation of submitted solutions will be done using an XGBoost model, trained independently on the three sets of samples added to the initial data which was made batch available to participants. Each model will be evaluated on a separate test set (hidden from participants). The quality metric used for the evaluation will be the average F1score. From each score, we will subtract the average F1score obtained by a model trained only on the initial data batch and the results for three sets will be summed with weights 10, 2.5, and 1 for the subset size 50, 200, and 500, respectively.

During the challenge, your solutions will be evaluated on a small fraction of the test set, and your best preliminary score will be displayed on the public Leaderboard. After the end of the competition, the selected solutions will be evaluated on the remaining part of the test data and this result will be used for the evaluation of the project.

In order to download competition files you need to be enrolled.
This forum is for all users to discuss matters related to the competition. Good manners apply!
There is no topics in this competition.