按小时重采样分类数据的方法及序列预测需求咨询

Hey there! Let's tackle your problem step by step—first, exploring alternative ways to resample your hourly categorical data, then diving into methods to predict the subsequent type sequence.

You already mentioned using the mode (most frequent value) for hourly aggregation, which is a solid default. Here are other practical, context-aware options depending on your business needs:

• Last Observation Carried Forward (LOCF)：Take the last type recorded in each hour. This makes sense if you care about the final state of the hour—for example, if each entry represents a state change, the last entry reflects the hour's ending status. For your sample data, the 09:00 hour's last entry is D, so you'd mark that hour as D.
• First Observation of the Hour：Conversely, take the first type of the hour. This works if the initial event sets the tone for the entire hour (e.g., a morning workflow that starts with a specific operation). For your 11:00 hour, the first entry is D, so you'd use that.
• Weighted by Recency：Assign higher weight to entries that occur later in the hour. For example, an entry at 11:50 gets more weight than one at 11:10. You can calculate a weighted mode where each entry's weight is proportional to how close it is to the end of the hour. This balances frequency and recency.
• Priority-Based Selection：Define business rules to prioritize certain type values over others. For instance, if D represents a critical operation (like deletion), you might flag any hour containing a D as D, regardless of other entries. This is great for use cases where specific events take precedence.
• Threshold-Based Classification：Only assign a type to the hour if it makes up a minimum percentage of entries (e.g., 50% or more). If no type meets the threshold, mark the hour as "Mixed" or use a placeholder. This avoids forcing a classification when the hour's data is evenly split.

Given a pre-sequence like RRCD, here are actionable methods to forecast future hours:

• Markov Chain Model：This is perfect for short-term sequence prediction with discrete categorical data. First, build a transition probability matrix from your historical data: count how often each type follows another (e.g., how many times R is followed by R, C, or D). Then, use the last element of your pre-sequence (D) to look up the most probable next type. For multi-step predictions, chain these probabilities together (predict the next type, then use that as the new starting point for the next prediction).
• N-Gram Frequency Analysis：Look at the frequency of consecutive n-length sequences in your historical data. For example, with a 3-gram model, check how often the sequence RCD (the last 3 elements of RRCD) is followed by each type. The most frequent follower becomes your next prediction. This works well if your sequences have short-term patterns.
• Baseline Frequency Model：If there's no strong sequential pattern, fall back to the overall frequency of each type in your historical hourly data. For example, if R appears in 40% of hours, C in 35%, and D in 25%, predict R for all subsequent hours. This is a simple baseline to compare more complex models against.
• LSTM/GRU Neural Networks：If you have a large enough dataset (dozens or hundreds of hourly entries), use a recurrent neural network (RNN) like LSTM to capture long-term sequential dependencies. Convert each type to a one-hot encoded vector, train the model on your historical sequence, and then generate predictions for future hours. This is ideal if your data has complex, non-obvious patterns.

内容的提问来源于stack exchange，提问作者Aditya C S