0
$\begingroup$

I have a dataset representing the stock of a shop over several days. For each day, I have hourly inventories of the objects in the shop. Some products are sold, and others might temporarily disappear (e.g., moved to the backstore). Therefore, the number of items in the inventory fluctuates hourly. Each item also has several attributes, such as its position, detection quality, and others.

Example dataset:

example datset

I have ground truth data indicating the actual stock present. My main goals are:

Predict Inventory: Every hour, I want to predict whether an object is in the inventory or not and make this predicted inventory as accurate as possible compared to the real inventory.

Classify Objects: Based on their behavior, I want to classify objects.

Challenges:
Pattern Detection and Prediction:

  • I observe a repeating pattern in the stock levels over 24 hours.
  • I considered training an LSTM to predict future behavior, but this is complex because: The model struggles to learn due to the large variability in behavior among thousands of objects.
  • The number of objects is not fixed at each time step.
  • Training a separate LSTM for each object is not feasible due to the large number of objects.

Behavior Classification:

  • I tried using K-means and similar clustering methods, but these only classify objects at a specific hour, which is not helpful for my temporal data.
  • I need to classify objects based on their behavior over 24 hours.

Questions: How can I generalize the LSTM or any other method to effectively learn from the data, given the variability and large number of objects? What would be a suitable approach to classify objects based on their behavior over 24 hours? Any guidance or suggestions on how to tackle these problems would be greatly appreciated. Thank you in advance!

Improve this question
$\endgroup$
2
  • $\begingroup$ This is always my advice, but have you considered any simpler methods before diving right into LSTM etc. Perhaps a multiple linear regression model. Chuck in all your data, and hour-of-the-day one hot (dummy) variables and see what happens? What you learn from that could help inform the direction you should take methodologically. $\endgroup$
    – Jamie
    Commented Aug 7 at 16:47
  • $\begingroup$ @Jamie as per my understanding, a multiple linear regression would work on independent samples data but it would not capture temporal relationships. I'm not sure about how can it help. In order to make a correct prediction should I train a different model for each hour of the day for example by gathering the data of the same hour across multiple days for example? $\endgroup$
    – Danielakaws
    Commented Aug 12 at 8:14

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.