TRAIN LABEL

[marzooq-unbxd]

In inference for a new user,what is the right way to input item events for the user?
Suppose our seq_max_len =49
and we have a list of 50 events = user_item_events
shouldnt input of user history be user_item_events[1:] instead of user_item_events[:-1]?Since we train on max 49 events?

taking any example such as

DeepMatch/examples/run_dssm_inbatchsoftmax.py

Line 81 in 5dab795

test_user_model_input = test_model_input

While training/validation I understand ,out trarin seq should be [:-1], and that we should try to make the model learn the 50th item.
In inference for a new user , shouldnt we use latest data[1:]?

---

The train label should relate to train_iid, the last item in the users's event history.We try to learn this through events upto that item.
The label we are using is basically [1].What does this mean, linked below

DeepMatch/examples/preprocess.py

Line 61 in 5dab795

(reviewerID, pos_list[i], 1, hist[::-1][:seq_short_len][::-1],

used as

DeepMatch/examples/preprocess.py

Line 104 in 5dab795

train_label = np.array([line[2] for line in train_set])

Are we using [1] for the train_iid in the last layer, and [0] for all other items?

--

2.While generating train, test data sets,
We are only adding one test data sample no matter how many interactions user has.
Should we split each user's events to maybe 100 events, and generate multiple test_data?

--

3.specific to just SDM,
the rating list specified in

DeepMatch/examples/preprocess.py

Line 53 in 5dab795

rating_list = hist['rating'].tolist()

is not used anywhere

Thanks for your help