Congrats to everyone in this competition. I notice that the rule states that:
\r\nAnd some teams have present their solutions in small version of this competition in foruma. I here present our solution to the big version of this competition in this topic.
\r\nWe use naive bayes as our algorithm in this competition. The features we used is:
\r\nA. We want to know the probability π(πβπ) that user click sku π in context π. We use naive bayes to predict this probability. Then select 5 item with highest predicted probability as prediction in context π. Here context π is query π.
\r\nThe first context we used is query. By naive bayes, we have:
\r\nCongrats to everyone in this competition. I notice that the rule states that:
\r\nAnd some teams have present their solutions in small version of this competition in foruma. I here present our solution to the big version of this competition in this topic.
\r\nWe use naive bayes as our algorithm in this competition. The features we used is:
\r\nA. We want to know the probability π(πβπ) that user click sku π in context π. We use naive bayes to predict this probability. Then select 5 item with highest predicted probability as prediction in context π. Here context π is query π.
\r\nThe first context we used is query. By naive bayes, we have:
\r\nMajor techniques used in my solution are summarized as following:
\r\n
\r\n1. Query/keyword normalization, such as case lowering, non-alphanumeric-character removal, stopword removal, and word ordering.
\r\n
\r\n2. Obtain a rule of user information: 98~99% people don't click items multiple times.
\r\n
\r\n3. For head queries that have sufficient samples, the feature tuple is (query, category, week number of query time) for ML. The reason of using week number is that besbuy usually changes promotions and categories weekly
\r\n
\r\n4. For tail queries that have few or no samples, kNN is used to find matched queries in a query set in which every query at least results in clicking the given category. The distance is defined by common keyword # / max (keyword #). I believe using TF-IDF\r\n weighted distance could improve performance further. I don't optimize the algorithm of kNN since it takes less than one hour and data will be cached.
\r\n
\r\nThe code is written in Python. Redis is used for DB.
\r\n
\r\nRAM requirement: the code makes a Redis DB of 400MB. A linux PC with 2G RAM should be able to run the code. BTW, I don't index users and queries. If do so, the DB size could be reduced by half.
Hello,
\r\n\r\n
It seems that you have only used train.csv for your algorithms.Is Given Product data is of no use . Or that can be used to improve the performance.
\r\n\r\n
Please Reply ASAP
\r\n
Would it be possible to get the SKUs for the test set? It would be nice to be able to compare apples-to-apples when testing new algorithms and seeing how they perform against the solutions presented here.
", "rawMarkdown": "", "votes": null }, { "id": "161680", "postDate": "02/15/2017 07:05:38", "content": "Is there any way I can submit my predictions now, to compare my results?
", "rawMarkdown": "Is there any way I can submit my predictions now, to compare my results?", "votes": null } ], "comments": [ { "id": 15490, "author_name": "songgc", "author_url": "", "post_date": "10/12/2012 21:35:48", "content": "Major techniques used in my solution are summarized as following:
\r\n
\r\n1. Query/keyword normalization, such as case lowering, non-alphanumeric-character removal, stopword removal, and word ordering.
\r\n
\r\n2. Obtain a rule of user information: 98~99% people don't click items multiple times.
\r\n
\r\n3. For head queries that have sufficient samples, the feature tuple is (query, category, week number of query time) for ML. The reason of using week number is that besbuy usually changes promotions and categories weekly
\r\n
\r\n4. For tail queries that have few or no samples, kNN is used to find matched queries in a query set in which every query at least results in clicking the given category. The distance is defined by common keyword # / max (keyword #). I believe using TF-IDF\r\n weighted distance could improve performance further. I don't optimize the algorithm of kNN since it takes less than one hour and data will be cached.
\r\n
\r\nThe code is written in Python. Redis is used for DB.
\r\n
\r\nRAM requirement: the code makes a Redis DB of 400MB. A linux PC with 2G RAM should be able to run the code. BTW, I don't index users and queries. If do so, the DB size could be reduced by half.
Hello,
\r\n\r\n
It seems that you have only used train.csv for your algorithms.Is Given Product data is of no use . Or that can be used to improve the performance.
\r\n\r\n
Please Reply ASAP
\r\n
Would it be possible to get the SKUs for the test set? It would be nice to be able to compare apples-to-apples when testing new algorithms and seeing how they perform against the solutions presented here.
", "votes": null, "replies": [] }, { "id": 161680, "author_name": "shantokhoka", "author_url": "", "post_date": "02/15/2017 07:05:38", "content": "Is there any way I can submit my predictions now, to compare my results?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "15468": "", "15490": "", "22656": "", "31020": "", "161680": "Is there any way I can submit my predictions now, to compare my results?" }, "source": "existing_discussion", "complete": true, "competition_id": 3064, "competition_slug": "acm-sf-chapter-hackathon-big", "team_id": 22346, "solution_url": "/competitions/acm-sf-chapter-hackathon-big/writeups/rapstar-solution-to-this-competition" }