I'm training every day XGBoost binary classifier on 150k of records.
every day, approximately 500 records added to the train set.
The test set (inference) 10M records without labels.
I noticed that two models trained in consecutive days produce fundamentally different predictions on the same records (up to 10% of the inference data).
What I have tried for now:
set same seeds and parameters.
different classifiers.
My concern is that the model is not generalize enough so there are significant changes in the daily predictions.
I would love to hear about possible ways to address this issue
Related
I'm currently training a random forest on some data I have and I'm finding that the model performs better on the validation set, and even better on the test set, than on the train set. Here are some details of what I'm doing - please let me know if I've missed any important information and I will add it in.
My question
Am I doing anything obviously wrong and do you have any advice for how I should improve my approach because I just can't believe that I'm doing it right when my model predicts significantly better on unseen data than training data!
Data
My underlying data consists of tables of features describing customer behaviour and a binary target (so this is a binary classification problem). Technically I have one such table per month and I tend to use several months of data to train and then a different month to predict (e.g. Train on Apr, May and Predict on Jun)
Generally this means I end up with a training dataset of about 100k rows and 20 features (I've previously looked into feature selection and found a set of 7 features which seem to perform best, so have been using these lately). My prediction set generally has around 50k rows.
My dataset is heavily unbalanced (approximately 2% incidence of target feature), so I'm using oversampling techniques - more on that below.
Method
I've searched around online quite a lot and this has led me to the following approach:
Take scaleable (continuous) features in the training data and standardise them (currently using sklearn StandardScaler)
Take categorical features and encode them into separate binary columns (one-hot) using Pandas get_dummies function
Remove 10% of the training data to form a validation set (I'm currently using a random seed in this process for comparability whilst I vary different things such as hyperparameters in the model)
Take the remaining 90% of training data and perform a grid search across a few parameters of the RandomForestClassifier() (currently min_samples_split, max_depth, n_estimators and max_features)
Within each hyperparameter combination from the grid I perform kfold validation with 5 folds and using a random state
Within each fold I oversample my minority class for training data only (sometimes using imbalanced-learn's RandomOverSampler() and sometimes using SMOTE() from the same package), train the model on the training data and then apply the model to the kth fold and record performance metrics (precision, recall, F1 and AUC)
Once I've been through 5 folds on each hyperparameter combination I find the best F1 score (and best precision if two combinations are tied on F1 score) and retrain a random forest on the entire 90% training data using those hyperparameters. During this step I use the same oversampling technique as I did in the kfold process
I then use this model to make predictions on the 10% of training data that I put aside earlier as a validation set, evaluating the same metrics as above
Finally I have a test set, which is actually based on data from another month, which I apply the already trained model to and evaluate the same metrics
Outcome
At the moment I'm finding that my training set achieves an F1 score of around 30%, the validation set is consistently slightly higher than this at around 36% (mostly driven by a much better precision than the training data e.g. 60% vs. 30%) and then the testing set is getting an F1 score of between 45% and 50% which is again driven by a better precision (around 65%)
Notes
Please do ask about any details I haven't mentioned; I've had my stuck in this for weeks and so have doubtless omitted some details
I've had a brief look (not a systematic analysis) of the stability of metrics between folds in the kfold validation and it seems that they aren't varying very much, so I'm fairly happy with the stability of the model here
I'm actually performing the grid search manually rather than using a Python pipeline because try as I might I couldn't get imbalanced-learn's Pipeline function to work with the oversampling functions and so I run a loop with combinations of hyperparameters, but I'm confident that this isn't impacting the results I've talked about above in an adverse way
When I apply the final model to the prediction data (and get an F1 score around 45%) I also apply it back to the training data itself out of interest and get F1 scores around 90% - 100%. I suppose this is to be expected as the model is trained and predicts on almost exactly the same data (except the 10% holdout validation set)
I'm working on a project to predict demand for a product based on past historical data for multiple stores. I have data from multiple stores over a 5 year period. I split the 5-year time series into overlapping subsequences and use the last 18 months to predict the next 3 and I'm able to make predictions. However, I've run into a problem in choosing a cross-validation method.
I want to have a holdout test split, and use some sort of cross-validation for training my model and tuning parameters. However, the last year of the data was a recession where almost all demand suffered. When I use the last 20% (time-wise) of the data as a holdout set, my test score is very low compared to my OOF cross-validation scores, even though I am using a timeseriessplit CV. This is very likely to be caused by this recession being new behavior, and the model can't predict these strong downswings since it has never seen them before.
The solution I'm thinking of is using a random 20% of the data as a holdout, and a shuffled Kfold as cross-validation. Since I am not feeding any information about when the sequence started into the model except the starting month (1 to 12) of the sequence (to help the model explain seasonality), my theory is that the model should not overfit this data based on that. If all types of economy are present in the data, the results of the model should extrapolate to new data too.
I would like a second opinion on this, do you think my assumptions are correct? Is there a different way to solve this problem?
Your overall assumption is correct in that you can probably take random chunks of time to form your training and testing set. However, when doing it this way, you need to be careful. Rather than predicting the raw values of the next 3 months from the prior 18 months, I would predict the relative increase/decrease of sales in the next 3 months vs. the mean of the past 18 months.
(see here)
http://people.stern.nyu.edu/churvich/Forecasting/Handouts/CourantTalk2.pdf
Otherwise, the correlation between the next 3 months with your prior 18 months data might give you a misleading impression about the accuracy of your model
Given that I have a deep learning model(handover from former colleague). For some reason, the train/dev set was missing.
In my situation, I want to classify my dataset into 100 categories. The dataset is extremely imbalanced. The dataset size is about tens of millions
First of all, I run the model and got the prediction on the whole dataset.
Then, I sample 100 records per category(according to the prediction) and got a 10,000 test set.
Next, I labeled the ground truth of each record for the test set and calculate the precision, recall, f1 for each category and got F1-micro and F1-macro.
How to estimate the accuracy or other metrics on the whole dataset? Is it correct that I use the weighted sum of each category's precision(the weight is the proportion of prediction on the whole) to estimate?
Since the distribution of prediction category is not same as the distribution of real category, I guess the weighted approach does not work. Any one can explain it?
The issue if you take a weighted average is that if your classifier performs well on the majority class, but poorly on minority classes (which is the typical scenario), it will not be reflected in the score.
One of the recommended approaches is rather to use the balanced accuracy score (see here for the scikit learn implementation). Basically, it is an average of all recall scores: for each observation in a class, it looks at how many of were correctly classified, and averages this across all classes. This will give you a sensible overall score to report.
I'm trying to build a classifier to predict stock prices. I generated extra features using some of the well-known technical indicators and feed these values, as well as values at past points to the machine learning algorithm. I have about 45k samples, each representing an hour of ohlcv data.
The problem is actually a 3-class classification problem: with buy, sell and hold signals. I've built these 3 classes as my targets based on the (%) change at each time point. That is: I've classified only the largest positive (%) changes as buy signals, the opposite for sell signals and the rest as hold signals.
However, presenting this 3-class target to the algorithm has resulted in poor accuracy for the buy & sell classifiers. To improve this, I chose to manually assign classes based on the probabilities of each sample. That is, I set the targets as 1 or 0 based on whether there was a price increase or decrease.
The algorithm then returns a probability between 0 and 1(usually between 0.45 and 0.55) for its confidence on which class each sample belongs to. I then select some probability bound for each class within those probabilities. For example: I select p > 0.53 to be classified as a buy signal, p < 0.48 to be classified as a sell signal and anything in between as a hold signal.
This method has drastically improved the classification accuracy, at some points to above 65%. However, I'm failing to come up with a method to select these probability bounds without a large validation set. I've tried finding the best probability values within a validation set of 3000 and this has improved the classification accuracy, yet the larger the validation set becomes, it is clear that the prediction accuracy in the test set is decreasing.
So, what I'm looking for is any method by which I could discern what the specific decision probabilities for each training set should be, without large validation sets. I would also welcome any other ideas as to how to improve this process. Thanks for the help!
What you are experiencing is called non-stationary process. The market movement depends on time of the event.
One way I used to deal with it is to build your model with data in different time chunks.
For example, use data from day 1 to day 10 for training, and day 11 for testing/validation, then move up one day, day 2 to day 11 for training, and day 12 for testing/validation.
you can save all your testing results together to compute an overall score for your model. this way you have lots of test data and a model that will adapt to time.
and you get 3 more parameters to tune, #1 how much data to use for train, #2 how much data for test, # per how many days/hours/data points you retrain your data.
I am using Jason Brownlee's tutorial (mirror) to apply LSTM network on some syslog/network log data. He's a master!
I have syslog data(a specific event) for each day for last 1 year and so I am using LSTM network for time series analysis. I am using LSTM from Keras deep learning library.
As I understand -
About Batch_size
A batch of data is a fixed-sized number of rows from the training
dataset that defines how many patterns to process before updating
the weights of the network. Based on the batch_size the Model
takes random samples from the data for the analysis. For time series
this is not desirable, hence the batch_size should always be 1.
About setting value for shuffle value
By default, the samples within an epoch are shuffled prior to being exposed to the network. This is undesirable for the LSTM
because we want the network to build up state as it learns across
the sequence of observations. We can disable the shuffling of
samples by setting “shuffle” to “False“.
Scenario1 -
Using above two rules/guidelines - I ran several trials with different number of neurons, epoch size and different layers and got better results from the baseline model(persistence model).
Scenario2-
Without using above guidelines/rules - I ran several trials with different number of neurons, epoch size and different layers and got even better results than Scenario 1.
Query - Setting shuffle to True and Batch_size values to 1 for time series. Is this a rule or a guideline?
It seems logical reading the tutorial that the data for time series should not be shuffled as we do not want to change the sequence of data, but for my data the results are better if I let the data be shuffled.
At the end what I think, what matters is how I get better predictions with my runs.
I think I should try and put away "theory" over concrete evidence, such as metrics, elbows, RMSEs,etc.
Kindly enlighten.
It depends a lot on the size of your data, also in the number of variables, decreasing batch size in my experience gives better results since the update is more frequent but in huge datasets it is very expensive. And you have to play with this trade-off (training time vs result).
About your shuffle it may be the case that your data is not that correlated with the past, if that is the case shuffling the data helps the network to learn and be able to generalize (like ordered by label) check reason 7 of the following 37 reasons your neural network not working
Batch size the larger the difficult it is to generalize (reason 11). When data clearly depends on the past you can declare your LSTM in Keras to stateful, this means: "that the states computed for the samples in one batch will be reused as initial states for the samples in the next batch" according to Keras API. Hope this helps.