{ "id": 316903, "title": "Recommender System using turicreate library", "url": "/competitions/h-and-m-personalized-fashion-recommendations/discussion/316903", "author_name": "Andrada", "post_date": "2022-04-04T14:58:47.941000", "votes": 21, "comment_count": 0, "views": 0, "content": "

As this was my first time handling Recommender Systems, I naturally went to Google to see a few approaches that I might learn and apply on this competition. However most of them were used for rating-based data.

\n

These examples work something like: if you have customer A and B that are very similar in their preferences towards movies, you can use what customer A has watched new to also recommend to customer B.

\n

\n

However in this competition we are working with something different - we are trying to recommend what a customer will buy based on previous purchases, which at quantity level are different than ratings (which are within a strict interval of [1, 5] or [1, 10]).

\n

After some digging I have found this medium post that helped me get started with this competition.

\n

Turicreate library

\n

Turicreate has very easy functions (you can check out my notebook or the article mentioned above for examples) that can be used to create a recommender system.

\n
\n

Note: to install turicreate in your notebook use !pip install turicreate --user instead of the simple !pip install turicreate - as the simple version has dependency issues.

\n
\n

Popularity Recommender

\n

The easiest fastest way is to create a popularity reccomendation - meaning what are the items that are sold the most across all customers (I have found this model to be very low in performance).

\n

Example code: tc.popularity_recommender.create(train_data, user_id, item_id, target)

\n

Similarity Recommender

\n

This methodology looks a lot in my opinion like Market Basket Analysis - where if you see that customers A buys milk, eggs, butter and flour, then customer D with milk, eggs, butter in the cart might also need flour (because, you guessed it, they're making pancakes).

\n

\n

There are multiple algorithms that calculate how similar is a one product to another. The idea is to compute for all milk, eggs, butter and flour how many times they were bought together and individually for each customer. Then, 4 vectors for all 4 items are computed in space.

\n

The way to compute how close together (similar) are these 2 vectors can be done in various ways, but within the article and my notebook 2 methodologies are used: cosine distance and pearson distance.

\n

cosine : tc.item_similarity_recommender.create(train_data, user_id, item_id, target, similarity_type='cosine')

\n

pearson : tc.item_similarity_recommender.create(train_data, user_id, item_id, target, similarity_type='pearson')

\n

After we have assessed that, for example, these vectors are VERY similat to eachother we can compute how likely is that Jane will buy the flour.

\n

Cosine Similarity Formula

\n

\n

Pearson Similarity Formula

\n

", "messages": [ { "id": 1745059, "postDate": "2022-04-04T14:58:47.940Z", "content": "

As this was my first time handling Recommender Systems, I naturally went to Google to see a few approaches that I might learn and apply on this competition. However most of them were used for rating-based data.

\n

These examples work something like: if you have customer A and B that are very similar in their preferences towards movies, you can use what customer A has watched new to also recommend to customer B.

\n

\n

However in this competition we are working with something different - we are trying to recommend what a customer will buy based on previous purchases, which at quantity level are different than ratings (which are within a strict interval of [1, 5] or [1, 10]).

\n

After some digging I have found this medium post that helped me get started with this competition.

\n

Turicreate library

\n

Turicreate has very easy functions (you can check out my notebook or the article mentioned above for examples) that can be used to create a recommender system.

\n
\n

Note: to install turicreate in your notebook use !pip install turicreate --user instead of the simple !pip install turicreate - as the simple version has dependency issues.

\n
\n

Popularity Recommender

\n

The easiest fastest way is to create a popularity reccomendation - meaning what are the items that are sold the most across all customers (I have found this model to be very low in performance).

\n

Example code: tc.popularity_recommender.create(train_data, user_id, item_id, target)

\n

Similarity Recommender

\n

This methodology looks a lot in my opinion like Market Basket Analysis - where if you see that customers A buys milk, eggs, butter and flour, then customer D with milk, eggs, butter in the cart might also need flour (because, you guessed it, they're making pancakes).

\n

\n

There are multiple algorithms that calculate how similar is a one product to another. The idea is to compute for all milk, eggs, butter and flour how many times they were bought together and individually for each customer. Then, 4 vectors for all 4 items are computed in space.

\n

The way to compute how close together (similar) are these 2 vectors can be done in various ways, but within the article and my notebook 2 methodologies are used: cosine distance and pearson distance.

\n

cosine : tc.item_similarity_recommender.create(train_data, user_id, item_id, target, similarity_type='cosine')

\n

pearson : tc.item_similarity_recommender.create(train_data, user_id, item_id, target, similarity_type='pearson')

\n

After we have assessed that, for example, these vectors are VERY similat to eachother we can compute how likely is that Jane will buy the flour.

\n

Cosine Similarity Formula

\n

\n

Pearson Similarity Formula

\n

", "rawMarkdown": "As this was my first time handling Recommender Systems, I naturally went to Google to see a few approaches that I might learn and apply on this competition. However most of them were used for rating-based data. \n\nThese examples work something like: if you have customer A and B that are very similar in their preferences towards movies, you can use what customer A has watched new to also recommend to customer B.\n\n\n\nHowever in this competition we are working with something different - we are trying to recommend what a customer will buy based on previous purchases, which at quantity level are different than ratings (which are within a strict interval of [1, 5] or [1, 10]).\n\nAfter some digging I have found [this medium post](https://medium.datadriveninvestor.com/how-to-build-a-recommendation-system-for-purchase-data-step-by-step-d6d7a78800b6) that helped me get started with this competition.\n\n# Turicreate library\n\nTuricreate has very easy functions ([you can check out my notebook](https://www.kaggle.com/code/andradaolteanu/h-m-eda-rapids-and-market-basket-analysis) or the article mentioned above for examples) that can be used to create a recommender system.\n\n> **Note**: to install `turicreate` in your notebook use `!pip install turicreate --user` instead of the simple `!pip install turicreate` - as the simple version has dependency issues.\n\n## Popularity Recommender\n\nThe easiest fastest way is to create a popularity reccomendation - meaning what are the items that are sold the most across all customers (I have found this model to be very low in performance).\n\nExample code: `tc.popularity_recommender.create(train_data, user_id, item_id, target)`\n\n## Similarity Recommender\n\nThis methodology looks a lot in my opinion like Market Basket Analysis - where if you see that customers A buys milk, eggs, butter and flour, then customer D with milk, eggs, butter in the cart might also need flour (because, you guessed it, they're making pancakes).\n\n\n\nThere are multiple algorithms that calculate *how similar is a one product to another*. The idea is to compute for all milk, eggs, butter and flour how many times they were bought together and individually for each customer. Then, 4 vectors for all 4 items are computed in space.\n\nThe way to compute how close together (similar) are these 2 vectors can be done in various ways, but within the article and my notebook 2 methodologies are used: `cosine` distance and `pearson` distance.\n\n`cosine` : `tc.item_similarity_recommender.create(train_data, user_id, item_id, target, similarity_type='cosine')`\n\n`pearson` : `tc.item_similarity_recommender.create(train_data, user_id, item_id, target, similarity_type='pearson')`\n\nAfter we have assessed that, for example, these vectors are VERY similat to eachother we can compute how likely is that Jane will buy the flour.\n\n### Cosine Similarity Formula\n\n
\n\n### Pearson Similarity Formula\n\n
", "votes": 21 } ], "comments": [], "raw_markdown_by_id": { "1745059": "As this was my first time handling Recommender Systems, I naturally went to Google to see a few approaches that I might learn and apply on this competition. However most of them were used for rating-based data. \n\nThese examples work something like: if you have customer A and B that are very similar in their preferences towards movies, you can use what customer A has watched new to also recommend to customer B.\n\n\n\nHowever in this competition we are working with something different - we are trying to recommend what a customer will buy based on previous purchases, which at quantity level are different than ratings (which are within a strict interval of [1, 5] or [1, 10]).\n\nAfter some digging I have found [this medium post](https://medium.datadriveninvestor.com/how-to-build-a-recommendation-system-for-purchase-data-step-by-step-d6d7a78800b6) that helped me get started with this competition.\n\n# Turicreate library\n\nTuricreate has very easy functions ([you can check out my notebook](https://www.kaggle.com/code/andradaolteanu/h-m-eda-rapids-and-market-basket-analysis) or the article mentioned above for examples) that can be used to create a recommender system.\n\n> **Note**: to install `turicreate` in your notebook use `!pip install turicreate --user` instead of the simple `!pip install turicreate` - as the simple version has dependency issues.\n\n## Popularity Recommender\n\nThe easiest fastest way is to create a popularity reccomendation - meaning what are the items that are sold the most across all customers (I have found this model to be very low in performance).\n\nExample code: `tc.popularity_recommender.create(train_data, user_id, item_id, target)`\n\n## Similarity Recommender\n\nThis methodology looks a lot in my opinion like Market Basket Analysis - where if you see that customers A buys milk, eggs, butter and flour, then customer D with milk, eggs, butter in the cart might also need flour (because, you guessed it, they're making pancakes).\n\n\n\nThere are multiple algorithms that calculate *how similar is a one product to another*. The idea is to compute for all milk, eggs, butter and flour how many times they were bought together and individually for each customer. Then, 4 vectors for all 4 items are computed in space.\n\nThe way to compute how close together (similar) are these 2 vectors can be done in various ways, but within the article and my notebook 2 methodologies are used: `cosine` distance and `pearson` distance.\n\n`cosine` : `tc.item_similarity_recommender.create(train_data, user_id, item_id, target, similarity_type='cosine')`\n\n`pearson` : `tc.item_similarity_recommender.create(train_data, user_id, item_id, target, similarity_type='pearson')`\n\nAfter we have assessed that, for example, these vectors are VERY similat to eachother we can compute how likely is that Jane will buy the flour.\n\n### Cosine Similarity Formula\n\n
\n\n### Pearson Similarity Formula\n\n
" } }