Making precise decisions for high-stakes applications such as finance, health, and self-driving is critical for increasing the economy of an entity or the quality of life. In most scenarios, decision quickness is also as essential as accuracy. This is particularly true in the case of event detection problems, where late detection can cause financial or physical damage. While recent work focuses on combining fast unsupervised AI decision systems and precise human decisions to solve this problem, the quality of this cooperation remains questionable. A human can generate ground-truth labels for the AI decision systems for future improvements. However, having noisy ground truth can worsen the performance. To address this challenge, this paper proposes FLAL (Feedback Loop Active Learner), a novel bridge system between the AI decision system and human/s, designed to understand human expertise and interest using a recommender mechanism and improve AI system performance using an active learning mechanism. FLAL is able to identify human behavior and makes entity recommendations to users who can generate better ground-truth labels for these entities. Our experiments show that FLAL performs better than competing baselines and converges fast.
Accuracy is one of the critical evaluation metrics in decision systems, especially for high-stake
applications [
Certain high-stakes decisions, such as detecting anomalies in operating server machines wherein missing them would cause financial loss, can be easily given by AI decision systems. In order to make such decisions, AI Decision systems are created using either historical experience (previous anomaly patterns, i.e., learned features) or algorithmic design (certain behaviors are anomalies, i.e., expert-designed features). However, historical experiences are not always available because of the label scarcity problem in AI for high-stakes applications. Therefore, the anomaly decision systems generally are designed as unsupervised classification models, which afects generalizability and generates multiple misclassifications. A human decision maker may solve this problem. However, human decision making is very time-consuming and not ideal where a fast decision is necessary. For instance, if a server machine fails, AI could detect this quickly compared to a human however human expertise is needed to check and confirm the detection as well as understand its root cause. In such a context, the human required should be an expert in understanding the problem. This scenario is not limited to high-stakes decisions. Credit card approval systems or insurance acceptance systems are other examples that AI may need the help of human decisions.
Feedback loop (Human-in-the-loop) systems have been studied [
In this paper, we are looking at specific scenarios of multiple independent entities. Each entity has temporal multidimensional features, and the AI system makes a decision for each entity and time (e.g., anomaly/failure decision). Entities will be ranked by their relevance score to the humans and queried to them to learn their expertise and interests with a pre-defined budget. It helps the framework generate accurate ground-truth labels and labeling operation will not be challenging or boring for a human since they are interested in answering.
We propose FLAL–a novel Feedback Loop Active Learner for better ground-truth labeling– which aims to learn the expertise and interest of a human before querying the entities to them using the active learning mechanism. Figure 1 illustrates a summary of FLAL bridging between an AI decision system and a human. FLAL collects decision for entities from the AI system, ranks entities based on their relevance score to human/s, and send queries based on the budget. The human/s answers these queries and sends them back to a FLAL, which learns their behavior towards these entities as well as stores the answers as ground truths. These ground truths will be used to improve AI decision systems in the future. Our main contributions can be summarized as follows: • We highlight the limitation of current AI decision systems, human decisions, and their cooperation to generate data labels. AI decision system makes many mistakes, human decisions are slow, and cooperation may be limited because of the lack of expertise or interest from humans. • We propose FLAL, a novel feedback loop active learning framework, for better groundtruth generation and understanding of human behavior. It uses active learning to train the framework based on human feedback and stores generated data labels to improve AI decision systems in the future. • We conduct experiments to verify the efectiveness of FLAL. We show that our framework performs better than competing baselines: random forest active learner, AI decisionbased, and random recommendations. FLAL not only has the best performance but also converges fast.
Human-in-the-Loop
Human-in-the-loop, in other words, feedback-loop, mechanisms are studied in the literature to
enhance AI performance by label annotation [
Recommendation systems are one of the solutions for understanding the behavior of
individuals. They are designed to infer interests and recommend items to humans based on their
historical experience or user interactions [
Temporal Embeddings
Representation learning on time-series data has become a popular technique to reduce the
dimension of the temporal data while keeping the representation (meaning) of it intact [
We formulate our problem as a self-supervised time-series classification. Given an entity set ℰ = {1, 2, . . . , } where each entity represents a multivariate time-series data (i.e., = [1, 2, . . . , ]), an unsupervised AI decision system which generates decision probability for each entity and timestamp (i.e, () = ), a user set = {1, 2, . . . , }, and interest labels ∈ {0, 1}× × for each user to certain timestamp and entity; our goal is to learn a function ^ : ℰ , → that approximates the expertise of users.
We introduce FLAL, a feedback loop active learner that generates ground-truth labels for the AI decision system while learning human expertise and/or interest. FLAL performs user mapping and feature extraction that optimizes the human expert’s predictions. As a result, it obtains better ground-truth labels for the AI decision system.
Figure 2 describes the steps of FLAL in detail. FLAL finds a global embedding space using pre-trained time-series embedders and translates the global embedding space into personalized embedding space. To overcome the cold-start problem, FLAL extracts features from user embedding space and incorporates AI decisions into this feature set. Finally, it calculates relevance scores for each entity and sends the top to human experts for evaluation. The human classifies each query which in turn generates ground truth information and adds feedback (explicit or implicit) about their expertise or interest in a certain query. FLAL trains its framework using this feedback via an active learning mechanism and stores the ground truth information to improve AI decision systems in the future when necessary.
To increase the expressiveness and compactness of our data, we use an unsupervised multivariate
time-series embedder to represent the entity’s time series. This part of our algorithm is
pretrained with the data, which is not used in our experiments. We used [
ℎ = ([(− ); . . . ; ]) where ℎ is an embedding of , is an unsupervised multivariate time-series embedder, is multivariate time-series data for , and [· ; · ] is the row concatenation operator. (1)
Global embedding space may not be as representative as user embedding space, where one can understand the expertise of users. Therefore, we have a user embedding mapper that maps generated embeddings from the stream data embedder to the user embedding space. This allows them to distinguish between relevant and irrelevant entities to the user. Figure 4 shows an example of the usefulness of user embedding mapper. For an anomaly detection problem, multiple reasons can cause an anomaly. But the users are often experts on a certain subset of those anomalies and the user embedding mapper will map these types close to each other. As a result, they can be separated from the other anomaly types or normal ones. The user embeddings are generated as follows: ℎ = (ℎ) (2) where ℎ is a user embedding for User A, and is the user embedding mapper. can be designed as any function, such as the identity or neural networks.
Since we do not know any information about the users and cannot conduct an initial survey as most recommenders do, we need to extract features from the user embedding space. Features can be learned or designed for a specific application scenario. An example of an expert-designed feature for anomaly applications can be the average distance from one item to others, which will likely be higher for anomaly cases. However, learned features are shown as more expressive than expert-designed features because it is hard to design or engineer all useful features. Feature extractor can also be seen as a function layer on top of the user embedding mapper. So it will learn new features from ℎ:
ℎ′ = (ℎ) where ℎ′ has smaller dimension than ℎ, and is a feature extractor function. can also represent a set of learned and expert-designed functions. In that case, ℎ′ will be a concatenation of extracted features.
In order to find better queries for specific users, we calculate each entity’s relevance score to a user. Note that the AI decision system already calculates decision probability for entity . Even though this probability may not be fully correct, we can incorporate it into relevance score calculation to ease the cold-start problem. Furthermore, we will also use the extracted features from the feature extractor. The relevance score of for a user calculation will be as the following:
= 1 × + ∑︁ ⊙ ℎ′ where 1 and are learned weights. Also, these weights may tell us a story about the importance of AI decision systems and extracted features for diferent users. Once relevance scores are calculated for all entities, the recommender will send the top relevant entities to the user to get feedback.
The users will have a list of queries to be checked and answered. The user will respond to each query with two answers: (1) what should be the decision of the AI system? (2) what is their expertise/interest in this query? The first answer is stored to update the AI decision system if necessary, while the second is used to train FLAL. Note that our main algorithm is not controlling the answering part done by the user. If a query has no response, it means no decision (i.e., do not use to improve the AI decision system) and no expertise (improve FLAL with the information that the user is not an expert). Furthermore, an interaction system can be designed to understand the expertise or interest of the user in queries by looking at their click count or other related metrics. However, this is out of the scope of this project. The expertise information will be stored in ∈ {0, 1}, and used to update FLAL. 1–15 (3) (4)
We train our framework based on active learning principles since the problem requires learning the behavior of users in real-time to collect better ground truth labels for AI decision systems. The collected ground truth information will be stored to update the AI decision system if necessary. Our active learning mechanism focuses on updating the recommender using feedback from expertise information. For each timestamp t, we train our objective which aims to sort relevance scores of entities, = [1, . . . , ] based on the expertise information array = [1, . . . , ]. More specifically we use a contrastive loss for our active learning training which contains three diferent terms as follows: max ALL(, ) = 1 * where WIDEN widens the gap between correctly ranked positive and negative samples, NARROW narrows the gap between wrongly ranked positive and negative samples, and RECOVERY recovers unrecommended by narrowing the gap between wrongly-recommended samples and unrecommended samples (which may contain useful recommendations). We use 1,2,3 as a tunable hyperparameter to value each term respectively. They can be optimized based on the scenario and the need of an application.
Let = [1.5, 1.2, 1.1, 0.9, 0.3, 0.2], = 4, = [1, 0, 0, 1, ?, ?], 1 = 0.50, 2 = 0.75, and 3 = 0.25, then our objective will be calculated as follows: ALL(, ) = 0.50 * ( (1.5 − 1.2) + (1.5 − 1.1)) → WIDEN + 0.75 * ( (0.9 − 1.2) + (0.9 − 1.1)) → NARROW + 0.25 * ( (0.3 − 1.2) + (0.2 − 1.2) + (0.3 − 1.1) + (0.2 − 1.1)) → RECOVER
To see the efectiveness of our feedback-loop part, we need to simulate the user answers. One way to do this is by using ground truth information for the AI decision system if it is available (simulating that the user’s expertise/interest is the ground truth). We apply this strategy in this paper. However, this would allow only one user available in the system. To extend the number of users to more than one, we propose a new way of simulating users. Each user is represented as a Gaussian latent space in the entity embedding space. The user space is assigned randomly. If a query entity is in this space, the user is considered an expert. The user embedding mapper will map related entities into this space. Figure 5 shows the mapping example.
We illustrate the empirical verification of FLAL compared to three baselines on a dataset. First, we compare method performance with precision at Section 4.3. Later on, we conduct two ablation studies: the efect of user embedding mapper (Section 4.4) and objective function weights (Section 4.5).
We use a public Server Machine [
We used three baselines to compare our method.
Random: It makes random recommendations in the recommender step to the user.
AI System Decisions: It only uses AI decision system probability to recommend entities to the user.
Random Forest Active Learner: It combines uncertainty and confidence scores for each entity and recommends them to the user. The model is trained using active learning with a random forest as the estimator and the same settings as FLAL.
Model selection: We select user embedding mapper as a linear layer as a result of ablation study (See Section 4.4), feature extractor generates 15 learned features with linear layer, and recommender is also a linear layer to generate a relevance score for an entity. To simulate user feedback, we use the ground truth information of the dataset.
Hyperparameters: We tune hyperparameters of FLAL using grid search. We optimize our model using Adam optimizer with a learning rate of 0.0001, L2 normalization weight on model weights 0.001. We select loss function weights 1, 2, and 3 from {0.0, 0.5, 1.0} (See Section 4.5). In our experiments, is set to 127 (the length of the multivariate time series data becomes 128 with the current snapshot) and the embedding size is 128. We set to 10 and 20 for diferent runs. The number of recommended items is also set to Q.
Evaluation metrics: Since the real evaluation can only consider feedback from the recommended entities, we compare precision metrics in our experiments. At each round, we calculate precision@Q and average precision@Q.
Figure 6 shows precision@Q and average precision@Q, where Q is 10 and 20. We calculate a cumulative average of precision performance at each step. Note that this performance will reflect improved AI decision system performance since we use the user’s interest/expertise label as a ground truth decision. Our method, Feedback Loop Active Learner, outperforms the competing baselines at all reported metrics, especially after 10-20 steps. Another essential requirement of learning human interest is convergence speed. FLAL converges in around 50 steps, faster than the best baseline Random Forest Active Learner.
Notice that the precision performance of the AI Decision System is not enough. However, active learner mechanisms improve the performance of the decision system drastically, while random decisions are still worse than the original AI decision system. Another notable diference between = 10 and = 20 is in precision@Q performance. When = 20, the performance drops below 0.6. However, this essentially could happen because the number of anomalies in the data is hardly more than 12 (i.e., 20 × 0.6) at a certain time. This shows us that we should optimize the number of recommended entities based on the number of anomalies at every step instead of using the same values as the budget of . The average precision is less vulnerable to this issue since the leading zeros do not afect the result. For both , the performance is close to each other. Random
AI Decision System
Random Forest Active Learner
Feedback Loop Active Learner 0.1 0 50 100 150 250 300 350 0 50 100 150 250 300 350
1.0 0 10.8 = Q @ iion0.6 s c e r P0.4 e g a r ve0.2 A 0.0 1.0 0 20.8 = Q @ n io0.6 s i c e r P e0.4 g a r e v A0.2 0.1 Identity Linear NonLinear Non-Linear2 0 50 100 150 250 300 350 0 50 100 150 250 300 350 0.1 0.8 0.7 1.0 0.8 0.7 0.6 0.5 0.4 0.3
Figure 8 shows a sensitivity analysis on objective function terms for the Machine dataset. Each block represents the last step cumulative average of precision@10 scores from FLAL parameterized by diferent 1,2,3 values. All terms contribute to the performance, while the absence of the first term (i.e., 1 = 0) makes the model performance very bad. This behavior is expected since the first term is the core part of the contrastive loss. The second term is not efective as the first term for the Machine dataset, but still increasing the value of 2 makes the performance better. The third term has a similar efect as the second term. They both aim to narrow the positive and negative sample rankings. The best performance is achieved when hyperparameters are equal to 1, or 1 = 1, 2 = 0.5, and 3 = 0.5.
We investigate better and more efective ground truth generation by incorporating recommendation systems into AI decision systems and human collaboration for entity-based time-series data. We propose FLAL understanding the expertise and interest of a human over queries to make feedback more eligible and accurate using active learning. FLAL trains a personalized embedding mapper; uses features extraction and AI system decisions to solve the cold-start problem of recommender systems. FLAL performs better than competing baselines: random forest active learners, AI decision-based, and random recommenders; and it converges fast. Furthermore, our ablation studies show that the linear user embedding mapper is learning enough information and each term in the objective function contributes to the result.
In future work, we want to investigate this problem using diferent datasets and our proposed user simulation setting. We also desire to conduct human experiments to show the efectiveness of FLAL in real settings. Furthermore, we will optimize the number of recommendations instead of using it as the budget . 1–15
The work was done when Mert Kosan was an intern at Visa Research.