2024arXiv surveyEvaluation metrics
A Comprehensive Survey of Evaluation Techniques for Recommendation Systems
Jadon & Patil (Juniper Networks).
A catalogue, not a new method: the survey groups recommender-system metrics into five families. Similarity metrics (cosine, Jaccard, Pearson, Manhattan, …) compare users or items to each other. Candidate-generation metrics (novelty, diversity, serendipity, catalog and distributional coverage) ask whether the system explores broadly rather than always returning the same popular items. Predictive metrics (RMSE, MAE, R², …) score rating-prediction accuracy. Ranking-based metrics (Precision@k, Recall@k, F1@k, MRR, nDCG@k, MAP, ARHR@k) score how well a list is ordered, not just which items appear in it. Business metrics (CTR, conversion, revenue, retention) check whether any of the above moves a real-world outcome. The paper’s central argument is that no single metric captures system quality — each family trades off against the others.
Why it is on my list: the TeleNotArzt assistant is, structurally, a recommendation system — it returns ranked shortlists of diagnoses, treatments, and measurements — so the ranking-based and candidate-generation families describe exactly the evaluation problem I face, and this survey gives a vocabulary of alternatives beyond the three metrics (recall@k, macro/weighted AUC, coverage ratio) I used in that paper.
When to use each family, for a system like mine:
- Precision@k vs. Recall@k — use precision@k when a noisy top-k list is the main risk (too many wrong suggestions to sift through); use recall@k when missing a correct item is the costlier failure. I chose recall@k for TeleNotArzt because in emergency care an omitted diagnosis is worse than a slightly noisy shortlist.
- F1@k — use when you need one headline number instead of two curves, i.e. for a quick comparison across model versions rather than a full precision/recall trade-off analysis.
- MRR — use when only the position of the first correct answer matters, e.g. how far down the list the clinician’s eventual diagnosis sat, rather than how many correct items are in the top-k overall.
- nDCG@k — use when relevance is graded rather than binary. Some wrong suggestions are more dangerous than others (a near-miss diagnosis vs. an implausible one); nDCG can weight results by that graded relevance in a way plain AUC or recall@k cannot — a natural next metric for this work.
- MAP — use for a single ranking score averaged across many cases/patients, when you want one system-level summary rather than per-case recall@k figures.
- Catalog / distributional coverage — this is close to where my own coverage ratio sits, generalised from “did this case get every relevant recommendation” to “across all cases, is the full range of diagnoses and treatments ever suggested, or does the model default to the common ones.” Worth adding to check the assistant isn’t quietly ignoring rare conditions.
- Novelty / serendipity — not something I use today, but relevant if the goal shifts toward surfacing rare-but-plausible diagnoses instead of only the statistically common ones, i.e. checking the model hasn’t taken a majority-class shortcut.
- Business metrics (CTR, adoption, conversion) — don’t map onto a clinical tool the way they do onto e-commerce, but the underlying instinct — pair offline ranking metrics with a real-world outcome, such as clinician acceptance rate or time-to-treatment — is exactly how this kind of assistant should eventually be validated beyond offline numbers.
Use Cases
The survey’s tables spell out, metric by metric, when to reach for each one. The first four cover the similarity, candidate-generation, predictive, and ranking families (verbatim from the paper):
Table 1 — Summary of Similarity Metrics
| Metric |
Use Cases |
| Cosine Similarity |
Use cosine similarity when evaluating the orientation, not magnitude, of user or item vectors in recommendation systems, making it ideal for text-based or attribute-rich data where the pattern of interest, rather than the absolute value, is indicative of user preferences. |
| Euclidean Distance |
Use Euclidean Distance in evaluating recommendation systems when comparing profiles in a feature space with numerical attributes to gauge similarity based on the ‘straight-line’ distance between points (user-item pairs). |
| Jaccard Index |
Use the Jaccard Index when evaluating the similarity between users’ or items’ sets of preferences or choices, particularly when the data is binary and the size of the intersection relative to the union of datasets is of interest. |
| Hamming Distance |
Use Hamming Distance in evaluating recommendation systems when comparing binary vectors of user preferences or item attributes, typically for assessing similarity or diversity in content-based filtering or collaborative filtering with binary datasets. |
| Manhattan Distance |
Use Manhattan Distance in recommendation systems when dealing with high-dimensional, sparse data sets, as it can be more robust to outliers than Euclidean distance and better at capturing differences when multiple dimensions contribute to the dissimilarity. |
| Chebyshev Distance |
Use Chebyshev Distance in recommendation systems when capturing the maximum disparity across dimensions is crucial, especially in high-dimensional spaces where the most significant difference between items is the most informative for recommendations. |
| Adjusted Cosine Similarity |
Use Adjusted Cosine Similarity when evaluating item-based collaborative filtering systems to account for varying user ratings and reduce bias by subtracting the user’s average rating from each of their ratings before computing similarity. |
| Pearson Correlation Coefficient |
Use Pearson Correlation in evaluating recommendation systems when assessing the linear relationship between users’ ratings, especially when the scale of ratings is important and you assume a normal distribution of the underlying data. |
| Spearman Rank Order Correlation Coefficient |
Use Spearman Rank Order Correlation Coefficient in recommendation systems when assessing the strength and direction of a monotonic relationship between ranked variables, particularly useful in scenarios with non-parametric, ordinal data or when evaluating the ranking quality of recommendations. |
Table 2 — Summary of Candidate Generation Metrics
| Metric |
Use Cases |
| Novelty |
Use novelty in evaluating recommendation systems when assessing the system’s ability to suggest unexpected, lesser-known, or new items, thereby enhancing user experience by introducing diversity and reducing the filter bubble effect prevalent in personalized recommendations. |
| Diversity |
Use Diversity in recommendation systems when the goal is to broaden user exposure beyond familiar items, avoid echo chambers, and enhance user engagement by presenting a varied range of options, thereby catering to a wider spectrum of user interests and preferences. |
| Serendipity |
Use serendipity in evaluating recommendation systems when assessing their ability to offer novel, unexpected, yet relevant items, enhancing user experience by introducing diversity beyond typical or predictable suggestions. |
| Catalog Coverage |
Use Catalog Coverage to evaluate a recommendation system when assessing its ability to suggest a wide variety of items, particularly important in scenarios where exposing users to a broader selection of the catalog is crucial for enhancing user experience and business objectives. |
| Distributional Coverage |
Distributional Coverage is used when evaluating the diversity of recommendations in a system, ensuring a wide range of items are suggested, rather than repeatedly offering popular or similar items, thus enhancing user discovery and experience. |
Table 3 — Summary of Predictive Metrics
| Metric |
Use Cases |
| Root Mean Squared Error (RMSE) |
Use Root Mean Squared Error (RMSE) to measure the magnitude of prediction errors, giving more weight to larger errors, which is ideal for scenarios where large deviations from actual values are particularly undesirable. |
| Mean Absolute Error (MAE) |
Use Mean Absolute Error (MAE) to measure the average magnitude of errors in predictions, prioritizing equal weighting to all deviations regardless of direction, useful for scenarios where all errors are equally important. |
| Mean Squared Error (MSE) |
Use Mean Squared Error (MSE) when evaluating a recommendation system’s predictive accuracy for continuous output, as it penalizes larger errors more severely, ensuring the model’s robustness by emphasizing the minimization of large prediction errors. |
| Mean Absolute Percentage Error (MAPE) |
Use Mean Absolute Percentage Error (MAPE) to measure forecast accuracy as a percentage, which is useful when you need to understand error magnitude relative to actual values, particularly in inventory or capacity planning where proportional errors are more meaningful. |
| R² |
Use R² (coefficient of determination) in evaluating recommendation systems when assessing the proportion of variance in the observed outcomes (e.g., ratings) that can be predicted from the input variables, providing a measure of how well unseen samples are likely to be predicted. |
| Explained Variance |
Use Explained Variance in evaluating recommendation systems when you aim to measure the proportion of user preference or rating variance that the model successfully captures, indicating the system’s overall effectiveness in predicting user behavior accurately. |
Table 4 — Summary of Ranking Metrics
| Metric |
Use Cases |
| Mean Reciprocal Rank (MRR) |
Use Mean Reciprocal Rank (MRR) when evaluating the effectiveness of a recommendation system at returning the first relevant item in a ranked list of results, emphasizing the importance of the top-most recommendation. |
| ARHR@k (Average Reciprocal Hit Rank at k) |
Use ARHR@k (Average Reciprocal Hit Rank at k) to measure the average quality of recommendations by considering the reciprocal of the rank at which relevant items are found in the top k positions. It’s useful for scenarios where ranking accuracy and the order of recommendations are crucial factors. |
| Normalized Discounted Cumulative Gain (nDCG) |
Use nDCG when ranking relevance matters, and you need to evaluate the quality of the recommendations in a list, considering the position of highly relevant items. |
| Precision@k |
Use Precision@k in recommendation systems when evaluating the proportion of relevant items among the top-k recommendations, particularly when the cost of false positives is high and the focus is on the quality of the top few recommendations. |
| Recall@k |
Use Recall@k in recommendation systems when prioritizing the model’s ability to capture all relevant items in the top-k suggestions is more important than avoiding irrelevant ones, especially when the cost of missing a relevant recommendation is high. |
| F1@K |
Use F1@K (F1 score at K) for scenarios where the focus is on the precision of the top K recommendations, such as content curation or ranking tasks, to ensure that the most relevant items are presented to users within a limited set. |
| Average Recall@K |
Use Average Recall@K when you need to assess the proportion of relevant items that a recommendation system successfully retrieves within the top K recommendations, prioritizing comprehensive coverage of relevant content over precision. |
| Average Precision@K |
Use Average Precision@k when evaluating the effectiveness of a recommendation system for scenarios where the order of recommended items matters, such as search engine result rankings or content playlists, as it quantifies the precision of the top k recommendations. |
| Mean Average Precision (MAP) |
Use Mean Average Precision (MAP) when evaluating the quality of a ranked list of items in recommendation systems particularly when relevance varies across items, as it provides a comprehensive measure of precision at different points in the ranking, emphasizing the importance of higher-ranked items. |
Reference: Jadon, A., & Patil, A. (2024). A comprehensive survey of evaluation techniques for recommendation systems. arXiv:2312.16015.
arXiv:2312.16015 →
2024Explainability
A Review of Multimodal Explainable Artificial Intelligence: Past, Present and Future
Shilin Sun, Wenbin An, Feng Tian, Fang Nan, Qidong Liu, Jun Liu, Nazaraf Shah, Ping Chen.
Artificial Intelligence has evolved rapidly over the last two decades. We have moved from simple machine learning models that relied on hand-crafted features to powerful multimodal systems such as GPT-4V, Gemini, and LLaVA that can simultaneously understand images, text, audio, and video. While these models achieve impressive performance, they introduce a critical challenge:
How do we understand and trust their decisions?
This question is at the heart of Multimodal Explainable Artificial Intelligence (MXAI), a research field that aims to make multimodal AI systems transparent and interpretable. The paper “A Review of Multimodal Explainable Artificial Intelligence: Past, Present and Future” provides a comprehensive survey of the field, tracing its evolution, current state, and future directions.
Why Multimodal Explainability Matters
Traditional Explainable AI (XAI) focuses on explaining decisions made from a single modality, such as an image classifier highlighting important image regions or a tabular model identifying influential features.
Modern AI systems, however, often process multiple modalities simultaneously:
- Images
- Text
- Audio
- Video
- Sensor data
For example, a medical AI system may analyze a chest X-ray, clinical notes, and patient history before producing a diagnosis. An explanation is no longer just about highlighting image pixels — it must also explain how textual information, visual evidence, and other data sources interact to produce the final decision.
This need has given rise to Multimodal Explainable AI (MXAI).
The Evolution of Explainability
One of the most interesting contributions of the survey is its historical perspective. The authors divide the development of explainable AI into four major eras.
1. Traditional Machine Learning Era (2000–2009)
During this period, AI models were relatively simple and inherently interpretable.
Common models included:
- Decision Trees
- Random Forests
- Logistic Regression
- Support Vector Machines
Because these models were smaller and more structured, explanations often came naturally. For example, a decision tree could directly reveal the sequence of rules that led to a prediction.
Strengths
- High transparency
- Easy-to-understand decision processes
- Straightforward feature importance analysis
Limitations
- Limited predictive power
- Difficulty handling complex multimodal data
- Poor scalability
2. Deep Learning Era (2010–2017)
The rise of deep learning dramatically improved performance across computer vision, speech recognition, and natural language processing.
Popular architectures included:
- Convolutional Neural Networks (CNNs)
- Recurrent Neural Networks (RNNs)
- LSTMs
However, interpretability suffered. As networks became deeper and more complex, understanding their internal reasoning became increasingly difficult. Researchers responded with explainability techniques such as:
- Saliency Maps
- Grad-CAM
- Activation Visualization
- Attention Mechanisms
Strengths
- Significant performance improvements
- Better multimodal feature learning
- Automatic representation learning
Limitations
- Black-box behavior
- Unstable explanations
- Limited understanding of internal reasoning
This period marked the birth of modern Explainable AI research.
3. Foundation Model Era (2017–2022)
The introduction of Transformers fundamentally changed AI. Models such as:
- BERT
- Vision Transformer (ViT)
- CLIP
- BLIP
learned rich representations from enormous datasets and could align multiple modalities within a shared semantic space. For example, CLIP learns connections between images and text by placing them into the same embedding space.
Explainability Methods
- Cross-attention visualization
- Similarity analysis
- Token attribution
- Concept alignment
Strengths
- Powerful multimodal understanding
- Strong transfer learning capabilities
- General-purpose representations
Limitations
- Massive model size
- Difficult internal interpretation
- Attention does not always correspond to reasoning
This era laid the foundation for today’s multimodal AI systems.
4. Large Language and Multimodal Model Era (2022–Present)
The latest generation of AI systems includes GPT-4V, Gemini, Claude, LLaVA, and Flamingo. These models can generate text, answer questions about images, perform reasoning tasks, and interact with users conversationally. While impressive, they introduce a new explainability challenge.
The Problem of Hallucinated Explanations
Large models can generate explanations that sound convincing but may not accurately reflect the actual reasoning process used internally. For example, a model may provide a detailed explanation for a medical diagnosis while internally relying on entirely different features. This creates a gap between:
- What the model says it did
- What the model actually did
The survey identifies this as one of the most important open problems in explainability research.
Understanding the Three Levels of Explainability
The survey organizes explainability methods into three broad categories.
1. Data Explainability
This focuses on understanding the data itself. Typical questions include which features are important, which modality contributes most, and what information drives predictions. Methods include feature selection, dimensionality reduction, and correlation analysis.
2. Model Explainability
This focuses on understanding the internal behavior of models — hidden representations, attention mechanisms, feature interactions, and modality fusion strategies. The goal is to understand how information flows through the model.
3. Post-hoc Explainability
Post-hoc methods generate explanations after predictions have already been made. Popular approaches include LIME, SHAP, Grad-CAM, and counterfactual explanations. These methods do not change the model but attempt to explain its behavior from the outside.
Datasets Commonly Used in MXAI Research
The survey reviews many datasets used to evaluate multimodal explainability systems.
Vision-Language Datasets
VQA, VQA-X, e-SNLI-VE, Flickr30k, and MS COCO — often used to study visual reasoning and explanation generation.
Medical Datasets
MIMIC-CXR and ChestX-ray14 — important because explainability is essential in healthcare applications.
Audio-Visual Datasets
AudioCaps and AVSD — support explainability research involving both speech and visual information.
Multimodal Reasoning Datasets
ScienceQA and MMMU — evaluate complex reasoning across multiple modalities.
How Do Researchers Evaluate Explanations?
Evaluating explanations is significantly harder than evaluating predictions. For predictions, metrics such as accuracy, precision, recall, F1 score, and AUC are well established. For explanations, researchers often use:
Text-Based Metrics
BLEU, ROUGE, METEOR, CIDEr.
Vision-Based Metrics
Intersection over Union (IoU), Pointing Game Accuracy, Localization Accuracy.
Multimodal Metrics
CLIP Score, Cross-modal Similarity, Human Evaluation.
A key conclusion of the survey is that there is still no universally accepted benchmark for evaluating explanations.
Real-World Applications of MXAI
Healthcare
Medical applications require high levels of trust and transparency — disease diagnosis, medical image analysis, clinical decision support. Doctors must understand why a system reaches a particular conclusion before relying on it.
Autonomous Driving
Self-driving systems combine information from cameras, LiDAR, radar, and GPS. Explainability is crucial for understanding driving decisions and investigating failures.
Security and Surveillance
Explainable systems can help analysts understand threat detection, fraud identification, and anomaly detection.
Education
Educational AI systems increasingly provide personalized learning experiences. Explainability can help students and educators understand recommendations and assessments.
Human-AI Collaboration
Future AI assistants will increasingly work alongside humans. Transparent explanations can improve trust, usability, and decision quality.
Major Challenges Facing MXAI
Despite rapid progress, several challenges remain.
Cross-Modal Complexity
Different modalities represent information in fundamentally different ways. Understanding how text, images, audio, and other data sources interact remains difficult.
Hallucinated Explanations
Large multimodal models may generate plausible but inaccurate explanations. Ensuring faithfulness between explanations and actual model reasoning is an active research area.
Lack of Ground-Truth Explanations
In many domains, there is no single correct explanation. This makes evaluation particularly challenging.
Scalability
Modern foundation models contain billions of parameters. Many existing interpretability techniques struggle to scale to these enormous architectures.
Human-Centered Evaluation
Most current evaluation frameworks rely heavily on automated metrics. However, explanations ultimately exist for people. The survey highlights the need for more user studies involving domain experts, end users, and decision-makers.
Future Directions
The authors identify several promising research directions.
Concept-Based Explainability
Rather than focusing on individual neurons, future systems may explain decisions using human-understandable concepts — “dog,” “road sign,” “lung opacity” — instead of abstract internal activations.
Causal Explainability
Researchers increasingly want explanations that capture cause-and-effect relationships rather than simple correlations.
Mechanistic Interpretability
This area aims to understand the actual computations occurring inside large neural networks.
Explainable Multimodal Agents
Future AI systems will not only perceive information but also act on it. Understanding their actions will become increasingly important.
Explainable Large Multimodal Models
As models such as GPT-4V and Gemini continue to evolve, explainability methods must evolve alongside them.
Final Thoughts
This survey serves as an excellent roadmap for understanding the rapidly growing field of Multimodal Explainable AI. Rather than proposing a new algorithm, the paper organizes the entire research landscape — from interpretable machine learning models to modern large multimodal systems — and highlights the key challenges that remain unsolved. The central message is clear:
As AI systems become more capable and more deeply integrated into real-world decision-making, explainability will become just as important as accuracy.
For researchers entering the field, this paper provides a valuable foundation for understanding where MXAI has been, where it stands today, and where it is likely headed in the future.
Why it is on my list: the disease-classification and TeleNotArzt work fuses tabular, time-series, and free-text data into a single clinical prediction, and a clinician won't act on a shortlist without knowing why it was produced. The survey's three-level split (data / model / post-hoc) is a useful map of where to add explainability to a fusion model, and its warning about hallucinated explanations is a direct risk for any explanation layer built on top of an LLM-based text encoder.
Reference: Shilin Sun, Wenbin An, Feng Tian, Fang Nan, Qidong Liu, Jun Liu, Nazaraf Shah, Ping Chen (2024). A Review of Multimodal Explainable Artificial Intelligence: Past, Present and Future. arXiv:2412.14056.
arXiv:2412.14056 →