PROJECT - Personalized Fashion Recommendations based on previous purchases.
Dataset: H&M Group Recommendation Dataset (Kaggle Competition)
-
Collect data and conduct data analysis. Visual reports.
-
Evaluate Classification Models (Decision Trees, Random Forests, Linear Regression) with Deep Learning Models (CNN, RNN etc) and based on the outcome Design and implement one or more deep learning systems, experiment with various algorithms to maximize the learning capability. Evaluate the performance and document findings.
We are asked to create a product recommendation for 7 products for each customer using the datasets given from us.
By grouping the characteristics of people who buy products sold in a store, we can assume that people in that group will be inclined to buy that product.
We will continue this approach in our solution process, take the characteristics of the people who buy each product, classify our products and make an estimate for each customer.
We will make these estimations by finding which group the customer belongs to from the products purchased in the past, and by giving the most purchased products in that group.
- A DF will be created including the characteristics of the users who bought each product, and classification will be made with this information.
- Then, each customer's information will be given to this model and its class will be found.
- The top 7 of the products are chosen for the cluster in which the prediction lies in.
There are 3 metadata .csv files and 1 image file:
- 📸 images - images of every article_id
- 🙋 articles - detailed metadata of every article_id (105,542 data points)
- 👔 customers - detailed metadata of every customer_id (1,371,980 data points)
- 🧾 transactions_train - file containing the customer_id, the article that was bought and at what price (31,788,324 data points)
article_id : A unique identifier of every article.
product_code, prod_name : A unique identifier of every product and its name (not the same).
product_type, product_type_name : The group of product_code and its name
graphical_appearance_no, graphical_appearance_name :The group of graphics and its name
colour_group_code, colour_group_name: The group of color and its name
perceived_colour_value_id, perceived_colour_value_name, perceived_colour_master_id, perceived_colour_master_name: The added color info
department_no, department_name : A unique identifier of every dep and its name
index_code, index_name: : A unique identifier of every index and its name
index_group_no, index_group_name : A group of indices and its name
section_no, section_name : A unique identifier of every section and its name
garment_group_no, garment_group_name : A unique identifier of every garment and its name
detail_desc : Details
customer_id : A unique identifier of every customer
FN : 1 or missed
Active : 1 or missed
club_member_status : Status in club
fashion_news_frequency : How often H&M may send news to customer
age : The current age
postal_code : Postal code of customer
t_dat : A unique identifier of every customer
customer_id : A unique identifier of every customer (in customers table)
article_id : A unique identifier of every article (in articles table)
price : Price of purchase
sales_channel_id : 1 or 2s
- Ladieswear accounts for a significant part of all dresses. Sportswear has the least portion.
- The garments grouped by index: Jersey fancy is the most frequent garment, especially for women and children. The next by number is accessories, many various accessories with low price.
- 70% of products are either ladieswear or children's wear.
- Most sold product is the Dragonfly dress.
- Ladieswear and Children/Baby have subgroups.
- Accessories are really various, the most numerous: bags, earrings and hats. However, trousers prevail.
- Trousers are the most sold product type followed by dress and sweater
- Total Frequent product Appearances is solid. And the most frequent color is black in the products bought.
- The most common age is about 21-23
- Status in H&M club. Almost every customer has an active club status, some of them begin to activate it (pre-create). A tiny part of customers abandoned the club.
- Customers prefer not to get any messages about the current news.
- Denims, Trousers and Undergarments are sold the most.
- The prices are altered, with the highest one being 0.59 and the lowest being 0.0000169.
- The most expensive items are leather garments.
- The average order has around 23 units and costs ~0.649.
- The units/order is directly correlated with the price/order: as the units increase, the price within the order increases too.
A dataset is prepared from articles, customers and transactions tables that includes the characteristics of the users who bought each product. A lot of preprocessing is done to make this dataset combining all the features and hot encoding a lot of them. In the end we get a dataset with 529 features.
We use PCA to reduce the features. I determined that we can get 95% variability with 135 features so I am going to reduce them to 135 features using principal component analysis.
I will then sort these products into clusters based on these characteristics.
Determine the best value for clusters using the elbow method. I have now grouped the products into 27 clusters. We will then divide the data into train and test and compare performance of different models on this dataset.
These results obtained for different model are
All the models are performing very well. It is mainly because of our good data preprocessing and feature engineering. Our model predicts the cluster in which the product lies. There are 26 clusters based on the characteristics of customer, product etc. Highest accuracy is given by Feed Forward Neural Network which is 0.984.
Furthermore, some further preprocessing is done to account for customer details in prediction and are inserted in the dataset of 529 features to make it more than that. We make some precautions to deal with missing values, combine all and then predict with our models. The below result is from the prediction of LGBMClassifier.
Now I will use these 7 predictions to recommend similar items.
I choose a product and preprocess the data for that image to convert it into a form that is passable by the model I pass the product to the model and it predicts the cluster in which the product lies in.
Note so the main thing that is helping are 2 things to make the model prediction:
- The trained model lgbm which has been trained on 5000 products to cluster them.
- Second is the dataset of predictions, after the model is trained on 5000 products.
The training make the model more accurate to give the correct cluster.
Secondly,the already predicted products that lie in different clusters are similar products. For each cluster there are 200-300 items which we can use to recommend since we have 200-300 similar products in each cluster.
I see the cluster in which my product lies in and choose the top 7 products from that cluster.
I combined another approach as well. It is I prepared a similarity search dataset.For each product, I have 10 similar products which i prepared using faiss similarity search. Each of the 115000 products have 10 recommendation in this one. So Now I have 70 predicted products.
For Customer Recommendation, the only thing different is that i get the latest product that the customer has bought use the model ready in put data for it and then use it to make prediction. Then pass the 7 recommendation through the similarity search dataset and get 70 recommendations in total.
Finally, you can play around it in Google colab with this link: H&M Predictions Colab Notebook
- https://www.kaggle.com/code/datamaniactr/h-m-recommendation-classification-eng/notebook
- https://www.kaggle.com/code/andradaolteanu/h-m-eda-rapids-and-similarity-recommenders#6.-Prepare-Datasets
This Notebook has been released under the Apache 2.0 open source license.