How to input historical data?

Question

Neural networks

I have never written neural networks in my life, but the task has arisen and we need to solve it.
I need the neural network to be able to predict what quantity of goods and on what day a certain client will order, based on his past orders. That is, analyze the intervals between past orders and the quantity ordered.

Implemented as follows:
For training, I take 10 orders, sorted by date, and 1 next as the correct answer. (in total 5400+ such bundles) I
bring the date to UNIX timestep and divide it by 10bn to get something like 0.1526995496.
I divide the number by 10000 (I have not yet figured out how to normalize the data more correctly).
I give an array as an input, where the date and quantity for each order alternate

[0.1526995496, 0.2, 0.1526995637, 0.4, 0.1526995809, 1.0...]

At the output, an array with the date and number [0.1536585814, 0.0001]

Activator for the first ReLu layer, for the output Linear.
Adam optimizer, MAE loss
(tried everything)

model = Sequential()
model.add(Dense(units=20, activation=act.relu))
model.add(Dense(units=2, activation=act.linear))
model.compile(optimizer=opt.Adam(), loss=ls.mae, metrics=['accuracy'])
history = model.fit(np_input, np_output, batch_size=1000, epochs=500, validation_split=0.2)

The training graph gives out complete nonsense like this (blue is history['loss'], yellow is accuracy)


The result is also not satisfactory

Previous orders:
2019-12-06 12:14:27 - 4.0
2019-12-10 11:25:44 - 2.0
2019-12-16 14:24:47 - 4.0
2020-01-03 13:16:28 - 4.0
2020-01-17 13:04:57 - 10.0
2020-03-27 14:25:57 - 2.0
2020-10-30 14:57:39 - 2.0
2020-11-09 15:00:24 - 3.0
2020-11-20 11:30:43 - 2.0
2021-01-06 11:20:15 - 1.0

Answer:
Date: 2020-01-02 14:58:58.516617
Amount:  13.0

This is the most appropriate answer. But the date is a thing of the past, some kind of anomalous number.
Sometimes says "Order -16000 items in 2851".

In general, the question is what. How can I correctly submit historical data to the input of the network? And am I doing everything right?

Answer 1

I would try to submit the input data by day (week and so on, how deep is your history). Conditionally on an input of 30 numbers.
1) 31 days before day X, bought 0 / 1000
2) 30 days before day X, bought 1 / 1000
....
30) yesterday bought 50 / 1000
Exit interpreted as softmax by options + quantity of goods
1) will call today , probability 0.1
2) will call tomorrow, probability 0.6
3) this week - 0.1
4) went to competitors and is not going to buy anything - 0.2
A) will order goods 8 / 1000
P.S. And that your network is so primitive, it doesn’t smell like deep learning, well, add two more hidden layers of 10-20 neurons or something.
P.P.S. And a batch for 1/5 of the entire sample is such a thing.