mldc2020

KO | EN

gitlite — search

#javascript #hacktoberfest #typescript #llm #golang #android #rust #deep-learning #nodejs #cli #kubernetes #vue #security #ios

(★ 8)

Code to fit my Mercado Libre Data Challenge 2020 submission (7th place).

mldc2020 Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# mldc2020

Code to fit my [Mercado Libre Data Challenge](https://ml-challenge.mercadolibre.com/) 
2020 submission (7th place).

## Run

_Tested on a t2.xlarge AWS EC2 instance._

Create a new conda environment with:

```shell
conda create -n r-mldc2020 r-base=3.6.1 r-essentials=3.6.0
```

Then activate the environment and run:

```shell
Rscript download.R && Rscript main.R
```

Submission will be written to `data/submission.csv`.

## About

_This section is a bit superficial. As of now, you can understand the model 
better through the `infer` function in the `functions/infer.R` script._

The model was made from scratch, without any common recsys or classification 
algorithms. It's basically a weighted average followed by some sorting rules, 
and inference is very fast (and it could easily be faster with parallelization). 
The model roughly described by the pseudocode below:

```
for each user history do:
  for each distinct item do:
    item frequency = normalized count of item views;
    item recency = normalized rank of items ordered by ascending date;
    item popularity = normalized count of item sales in other histories;
    item score = weighted mean of features;
  end
  top category = category of highest scoring item;
  relevant items = subset of items with at least one sale in other histories;
  recommendations = relevant items ranked by score;
  if number of recommendations is less than 10 do:
    recommendations = concatenate recommendations with top category items ranked by number of sales in other histories;
  end
  return recommendations
end
```

During pre-processing, fuzzy string matching is used to associate searches to 
item ids.

## Tune

There are only three parameters: weights for the weighted mean of frequency, recency and popularity.
By default, the `main.R` script fits the model with the parameters in `data/best_params.csv`. To tune them, first run:

```r
# This installation is separated from the others because this package 
# imports too many dependencies. It might be a good idea to substitute it
# in a future iteration.
remotes::install_version('ParBayesianOptimization', '1.2.1')
```

Then change the `train` variable in the `main.R` script (line 30).

```r
train <- TRUE # FALSE is the default
```

Afterwards, run the script. It will test new parameters and overwrite the `data/best_params.csv` file with the best set. Then, change the `train` variable back to `FALSE` and rerun the script to generate a submission with the newly tuned parameters.

mldc2020

mldc2020 Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

mldc2020

mldc2020 Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README