So far, most matrix factorization and \"classic\" recommendation techniques like ALS give a public LB score less than simple heuristics such as this notebook (@hengzheng) or this one by @cdeotte .
\nThis shows that simple heuristics such as recommending recently repeat purchase and products that were popular last week are (at least for now) better than techniques such as ALS.
\nI have been able to play a bit with the previous notebooks and reach 0.0215 on public LB (almost same on CV), however I was curious whether it's possible to use this type of heuristics and reach higher scores such as 0.024 or more with no ML/DL ?
", "messages": [ { "id": "1706764", "postDate": "02/27/2022 19:06:56", "content": "So far, most matrix factorization and \"classic\" recommendation techniques like ALS give a public LB score less than simple heuristics such as this notebook (@hengzheng) or this one by @cdeotte .
\nThis shows that simple heuristics such as recommending recently repeat purchase and products that were popular last week are (at least for now) better than techniques such as ALS.
\nI have been able to play a bit with the previous notebooks and reach 0.0215 on public LB (almost same on CV), however I was curious whether it's possible to use this type of heuristics and reach higher scores such as 0.024 or more with no ML/DL ?
", "rawMarkdown": "So far, most matrix factorization and \"classic\" recommendation techniques like ALS give a public LB score less than simple heuristics such as [this notebook](https://www.kaggle.com/hengzheng/time-is-our-best-friend-v2) (@hengzheng) or [this one](https://www.kaggle.com/cdeotte/recommend-items-purchased-together-0-021) by @cdeotte .\n\nThis shows that simple heuristics such as recommending recently repeat purchase and products that were popular last week are (at least for now) better than techniques such as ALS. \n\nI have been able to play a bit with the previous notebooks and reach 0.0215 on public LB (almost same on CV), however I was curious whether it's possible to use this type of heuristics and reach higher scores such as 0.024 or more with no ML/DL ?", "votes": null }, { "id": "1707700", "postDate": "02/28/2022 17:40:01", "content": "I tried LightFM which is another classic recommendation technique but still getting low CV scores around 0.001-0.005, I think maybe we need to use other features about customers and articles!
", "rawMarkdown": "I tried LightFM which is another classic recommendation technique but still getting low CV scores around 0.001-0.005, I think maybe we need to use other features about customers and articles!", "votes": null }, { "id": "1712048", "postDate": "03/04/2022 15:21:18", "content": "Possibly a good playground to do classic customer segmentation approach, based on few product attributes. E.g. buyer of basic products (a lot of socks and basic jeans as top seller!), buyer of ladieswear only, buyer of family items, etc. I tried extracting some customer attributes to enrich the customer table, feel free to use further from there if you're thinking of using customer attributes / segmentation approach. Dataset link
", "rawMarkdown": "Possibly a good playground to do classic customer segmentation approach, based on few product attributes. E.g. buyer of basic products (a lot of socks and basic jeans as top seller!), buyer of ladieswear only, buyer of family items, etc. I tried extracting some customer attributes to enrich the customer table, feel free to use further from there if you're thinking of using customer attributes / segmentation approach. [Dataset link](https://www.kaggle.com/aussie84/customer-product-attributes-aggregation-for-hm)", "votes": null }, { "id": "1712953", "postDate": "03/05/2022 14:17:35", "content": "I think recommending recently repeat purchase and products that were popular last week were initial attempts which gave success. As time goes on, multiple developments may happen. Heuristics algorithms can be further refined and become more involved. Also, some attempt at combining heuristics features like mentioned by @aussie84 into ALS and other ML algorithms may give better score.
\nAlso, note what is not shared publicly, is also out there.
I tried LightFM which is another classic recommendation technique but still getting low CV scores around 0.001-0.005, I think maybe we need to use other features about customers and articles!
", "votes": null, "replies": [] }, { "id": 1712048, "author_name": "aussie84", "author_url": "", "post_date": "03/04/2022 15:21:18", "content": "Possibly a good playground to do classic customer segmentation approach, based on few product attributes. E.g. buyer of basic products (a lot of socks and basic jeans as top seller!), buyer of ladieswear only, buyer of family items, etc. I tried extracting some customer attributes to enrich the customer table, feel free to use further from there if you're thinking of using customer attributes / segmentation approach. Dataset link
", "votes": null, "replies": [] }, { "id": 1712953, "author_name": "atulverma", "author_url": "", "post_date": "03/05/2022 14:17:35", "content": "I think recommending recently repeat purchase and products that were popular last week were initial attempts which gave success. As time goes on, multiple developments may happen. Heuristics algorithms can be further refined and become more involved. Also, some attempt at combining heuristics features like mentioned by @aussie84 into ALS and other ML algorithms may give better score.
\nAlso, note what is not shared publicly, is also out there.